Introduction to Econometric

Production Analysis with R

(Draft Version)
Arne Henningsen
Department of Food and Resource Economics
University of Copenhagen
January 15, 2014

Foreword
This is an incomplete collection of my lecture notes for various courses in the field of econometric
production analysis. These lecture notes are still incomplete and may contain many typos, errors,
and inconsistencies. Please report any problems to [email protected]. I am grateful
to my former students who helped me to improve my teaching and these notes through their
questions, suggestions, and comments. Finally, I thank the R community for providing so many
excellent tools for econometric production analysis.
January 15, 2014
Arne Henningsen

Contents
1 Introduction

10

1.1

Objectives of the course and the lecture notes . . . . . . . . . . . . . . . . . . . . . 10

1.2

An extremely short introduction to R . . . . . . . . . . . . . . . . . . . . . . . . . . 10

1.2.1

Some commands for simple calculations . . . . . . . . . . . . . . . . . . . . 11

1.2.2

Creating objects and assigning values . . . . . . . . . . . . . . . . . . . . . 12

1.2.3

Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

1.2.4

Simple functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

1.2.5

Comparing values and boolean values . . . . . . . . . . . . . . . . . . . . . 15

1.2.6

Data sets (data frames) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

1.2.7

Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

1.2.8

Other useful comands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

1.2.9

Extension packages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

1.2.10 Reading data into R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

1.3

Data sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.3.1

1.3.2

1.4

French apple producers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

1.3.1.1

Description of the data set . . . . . . . . . . . . . . . . . . . . . . 20

1.3.1.2

Abbreviating name of data set . . . . . . . . . . . . . . . . . . . . 21

1.3.1.3

Calculation of input quantities . . . . . . . . . . . . . . . . . . . . 21

1.3.1.4

Calculation of total costs and variable costs . . . . . . . . . . . . . 21

1.3.1.5

Calculation of profit and gross margin . . . . . . . . . . . . . . . . 22

Rice producers on the Philippines

. . . . . . . . . . . . . . . . . . . . . . . 22

1.3.2.1

Description of the data set . . . . . . . . . . . . . . . . . . . . . . 22

1.3.2.2

Mean-scaling Quantities . . . . . . . . . . . . . . . . . . . . . . . . 23

1.3.2.3

Logarithmic Mean-scaled Quantities . . . . . . . . . . . . . . . . . 23

1.3.2.4

Mean-adjusting the Time Trend . . . . . . . . . . . . . . . . . . . 24

1.3.2.5

Specifying Panel Structure . . . . . . . . . . . . . . . . . . . . . . 24

Mathematical and statistical methods . . . . . . . . . . . . . . . . . . . . . . . . . 24

1.4.1

Aggregating quantities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

1.4.2

Quasiconcavity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

1.4.3

Delta method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

Contents
2 Primal Approach: Production Function
2.1

28

Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.1.1

Production function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

2.1.2

Average Products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

2.1.3

Total Factor Productivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

2.1.4

Marginal Products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

2.1.5

Output elasticities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

2.1.6

Elasticity of scale . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

2.1.7

Marginal rates of technical substitution . . . . . . . . . . . . . . . . . . . . 30

2.1.8

Relative marginal rates of technical substitution . . . . . . . . . . . . . . . 30

2.1.9

Elasticities of substitution . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.1.9.1

Direct Elasticities of Substitution . . . . . . . . . . . . . . . . . . 30

2.1.9.2

Allen Elasticities of Substitution . . . . . . . . . . . . . . . . . . . 31

2.1.9.3

Morishima Elasticities of Substitution . . . . . . . . . . . . . . . . 32

2.1.10 Profit Maximization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

2.1.11 Cost Minimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.1.12 Derived Input Demand Functions and Output Supply Functions . . . . . . 34
2.1.12.1 Derived from profit maximization . . . . . . . . . . . . . . . . . . 34
2.1.12.2 Derived from cost minimization . . . . . . . . . . . . . . . . . . . 35
2.2

2.3

Productivity Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.2.1

Average Products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

2.2.2

Total Factor Productivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

Linear Production Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

2.3.1

Specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

2.3.2

Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

2.3.3

Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

2.3.4

Predicted output quantities . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

2.3.5

Marginal Products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

2.3.6

Output Elasticities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

2.3.7

Elasticity of Scale . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

2.3.8

Marginal rates of technical substitution . . . . . . . . . . . . . . . . . . . . 48

2.3.9

Relative marginal rates of technical substitution . . . . . . . . . . . . . . . 49

2.3.10 First-order conditions for profit maximisation . . . . . . . . . . . . . . . . . 49

2.3.11 First-order conditions for cost minimization . . . . . . . . . . . . . . . . . . 52
2.3.12 Derived Input Demand Functions and Output Supply Functions . . . . . . 54
2.4

Cobb-Douglas production function . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

2.4.1

Specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

2.4.2

Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

2.4.3

Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

Contents
2.4.4

Predicted output quantities . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

2.4.5

Output elasticities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

2.4.6

Marginal products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

2.4.7

Elasticity of Scale . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

2.4.8

Marginal Rates of Technical Substitution . . . . . . . . . . . . . . . . . . . 60

2.4.9

Relative Marginal Rates of Technical Substitution . . . . . . . . . . . . . . 61

2.4.10 First and second partial derivatives . . . . . . . . . . . . . . . . . . . . . . . 62

2.4.11 Elasticities of substitution . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
2.4.11.1 Direct Elasticities of Substitution . . . . . . . . . . . . . . . . . . 64
2.4.11.2 Allen Elasticities of Substitution . . . . . . . . . . . . . . . . . . . 66
2.4.11.3 Morishima Elasticities of Substitution . . . . . . . . . . . . . . . . 67
2.4.12 Quasiconcavity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
2.4.13 First-order conditions for profit maximisation . . . . . . . . . . . . . . . . . 69
2.4.14 First-order conditions for cost minimization . . . . . . . . . . . . . . . . . . 71
2.4.15 Derived Input Demand Functions and Output Supply Functions . . . . . . 73
2.4.16 Derived Input Demand Elasticities . . . . . . . . . . . . . . . . . . . . . . . 76
2.5

Quadratic Production Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

2.5.1

Specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

2.5.2

Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

2.5.3

Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

2.5.4

Predicted output quantities . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

2.5.5

Marginal Products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

2.5.6

Output Elasticities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84

2.5.7

Elasticity of Scale . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84

2.5.8

Marginal Rates of Technical Substitution . . . . . . . . . . . . . . . . . . . 85

2.5.9

Relative Marginal Rates of Technical Substitution . . . . . . . . . . . . . . 87

2.5.10 Elasticities of Substitution . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

2.5.11 Quasiconcavity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
2.5.12 First-order conditions for profit maximisation . . . . . . . . . . . . . . . . . 92
2.5.13 First-order conditions for cost minimization . . . . . . . . . . . . . . . . . . 93
2.6

Translog Production Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96

2.6.1

Specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96

2.6.2

Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96

2.6.3

Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98

2.6.4

Predicted Output Quantities . . . . . . . . . . . . . . . . . . . . . . . . . . 99

2.6.5

Output Elasticities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99

2.6.6

Marginal Products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

2.6.7

Elasticity of Scale . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

2.6.8

Marginal Rates of Technical Substitution . . . . . . . . . . . . . . . . . . . 103

Contents
2.6.9

Relative Marginal Rates of Technical Substitution . . . . . . . . . . . . . . 105

2.6.10 Second partial derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

2.6.11 Elasticities of Substitution . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
2.6.12 Quasiconcavity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
2.6.13 First-order conditions for profit maximisation . . . . . . . . . . . . . . . . . 111
2.6.14 First-order conditions for cost minimization . . . . . . . . . . . . . . . . . . 113
2.6.15 Mean-scaled quantities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
2.7

2.8

Evaluation of Different Functional Forms

2.7.1

Goodness of Fit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118

2.7.2

Theoretical Consistency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119

2.7.3

Plausible Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120

2.7.4

Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121

Non-parametric production function . . . . . . . . . . . . . . . . . . . . . . . . . . 121

3 Dual Approach: Cost Functions

3.1

3.2

. . . . . . . . . . . . . . . . . . . . . . . 118

127

Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
3.1.1

Cost function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127

3.1.2

Cost flexibility and elasticity of size . . . . . . . . . . . . . . . . . . . . . . 127

3.1.3

Short-Run Cost Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127

Cobb-Douglas Cost Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128

3.2.1

Specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128

3.2.2

Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128

3.2.3

Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129

3.2.4

Estimation with linear homogeneity in input prices imposed . . . . . . . . . 130

3.2.5

Checking Concavity in Input Prices . . . . . . . . . . . . . . . . . . . . . . 132

3.2.6

Optimal Cost Shares . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137

3.2.7

Derived Input Demand Functions . . . . . . . . . . . . . . . . . . . . . . . . 138

3.2.8

Derived Input Demand Elasticities . . . . . . . . . . . . . . . . . . . . . . . 141

3.2.9

Cost flexibility and elasticity of size . . . . . . . . . . . . . . . . . . . . . . 143

3.2.10 Marginal Costs, Average Costs, and Total Costs . . . . . . . . . . . . . . . 143

3.3

3.4

Cobb-Douglas Short-Run Cost Function . . . . . . . . . . . . . . . . . . . . . . . . 146

3.3.1

Specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146

3.3.2

Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147

3.3.3

Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147

3.3.4

Estimation with linear homogeneity in input prices imposed . . . . . . . . . 148

Translog cost function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149

3.4.1

Specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149

3.4.2

Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150

3.4.3

Linear homogeneity in input prices . . . . . . . . . . . . . . . . . . . . . . . 151

3.4.4

Estimation with linear homogeneity in input prices imposed . . . . . . . . . 153

Contents
3.4.5

Cost Flexibility and Elasticity of Size . . . . . . . . . . . . . . . . . . . . . 157

3.4.6

Marginal Costs and Average Costs . . . . . . . . . . . . . . . . . . . . . . . 158

3.4.7

Derived Input Demand Functions . . . . . . . . . . . . . . . . . . . . . . . . 162

3.4.8

Derived input demand elasticities . . . . . . . . . . . . . . . . . . . . . . . . 164

3.4.9

Theoretical consistency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167

4 Dual Approach: Profit Function

4.1

171

Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
4.1.1

Profit functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171

4.1.2

Short-run profit functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171

4.2

Graphical illustration of profit and gross margin . . . . . . . . . . . . . . . . . . . 171

4.3

Cobb-Douglas Profit Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173

4.4

4.3.1

Specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173

4.3.2

Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174

4.3.3

Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174

4.3.4

Estimation with linear homogeneity in all prices imposed . . . . . . . . . . 175

4.3.5

Checking Convexity in all prices . . . . . . . . . . . . . . . . . . . . . . . . 177

4.3.6

Predicted profit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182

4.3.7

Optimal Profit Shares . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183

4.3.8

Derived Output Supply Input Demand Functions . . . . . . . . . . . . . . . 185

4.3.9

Derived Output Supply and Input Demand Elasticities . . . . . . . . . . . . 185

Cobb-Douglas Short-Run Profit Function . . . . . . . . . . . . . . . . . . . . . . . 187

4.4.1

Specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187

4.4.2

Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188

4.4.3

Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188

4.4.4

Estimation with linear homogeneity in all prices imposed . . . . . . . . . . 189

4.4.5

Returns to scale . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190

4.4.6

Shadow prices of quasi-fixed inputs . . . . . . . . . . . . . . . . . . . . . . . 190

5 Stochastic Frontier Analysis

5.1

192

Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
5.1.1

Different Efficiency Measures . . . . . . . . . . . . . . . . . . . . . . . . . . 192

5.1.1.1

Output-Oriented Technical Efficiency with One Output . . . . . . 192

5.1.1.2

Input-Oriented Technical Efficiency with One Input . . . . . . . . 192

5.1.1.3

Output-Oriented Technical Efficiency with Two or More Outputs 193

5.1.1.4

Input-Oriented Technical Efficiency with Two or More Inputs . . 193

5.1.1.5

Output-Oriented Allocative Efficiency and Revenue Efficiency . . 193

5.1.1.6

Input-Oriented Allocative Efficiency and Cost Efficiency . . . . . . 194

5.1.1.7

Profit Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194

5.1.1.8

Scale efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195

Contents
5.2

Stochastic Production Frontiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195

5.2.1

Specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
5.2.1.1

5.3

5.4

Marginal products and output elasticities in SFA models . . . . . 197

5.2.2

Skewness of residuals from OLS estimations . . . . . . . . . . . . . . . . . . 197

5.2.3

Cobb-Douglas Stochastic Production Frontier . . . . . . . . . . . . . . . . . 198

5.2.4

Translog Production Frontier . . . . . . . . . . . . . . . . . . . . . . . . . . 203

5.2.5

Translog Production Frontier with Mean-Scaled Variables . . . . . . . . . . 206

Stochastic Cost Frontiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208

5.3.1

Specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208

5.3.2

Skewness of residuals from OLS estimations . . . . . . . . . . . . . . . . . . 208

5.3.3

Estimation of a Cobb-Douglas stochastic cost frontier . . . . . . . . . . . . 210

Analyzing the Effects of z Variables . . . . . . . . . . . . . . . . . . . . . . . . . . 212

5.4.1

Production Functions with z Variables . . . . . . . . . . . . . . . . . . . . . 212

5.4.2

Production Frontiers with z Variables . . . . . . . . . . . . . . . . . . . . . 213

5.4.3

Efficiency Effects Production Frontiers . . . . . . . . . . . . . . . . . . . . . 216

6 Panel Data and Technological Change

6.1

221

Average Production Functions with Technological Change . . . . . . . . . . . . . . 221

6.1.1

Cobb-Douglas Production Function with Technological Change . . . . . . . 221

6.1.1.1

Pooled estimation of the Cobb-Douglas Production Function with

Technological Change . . . . . . . . . . . . . . . . . . . . . . . . . 222

6.1.1.2

Panel data estimations of the Cobb-Douglas Production Function

with Technological Change . . . . . . . . . . . . . . . . . . . . . . 223

6.1.2

Translog Production Function with Constant and Neutral Technological

Change . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
6.1.2.1

Pooled estimation of the Translog Production Function with Constant and Neutral Technological Change . . . . . . . . . . . . . . . 228

6.1.2.2

Panel-data estimations of the Translog Production Function with

Constant and Neutral Technological Change . . . . . . . . . . . . 229

6.1.3

Translog Production Function with Non-Constant and Non-Neutral Technological Change . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
6.1.3.1

Pooled Estimation of a Translog Production Function with NonConstant and Non-Neutral Technological Change . . . . . . . . . . 235

6.1.3.2

Panel-data estimations of a Translog Production Function with

Non-Constant and Non-Neutral Technological Change . . . . . . . 241

6.2

Frontier Production Functions with Technological Change . . . . . . . . . . . . . . 247

6.2.1

Cobb-Douglas Production Frontier with Technological Change . . . . . . . 247

6.2.1.1

Time-invariant Individual Efficiencies . . . . . . . . . . . . . . . . 247

6.2.1.2

Time-variant Individual Efficiencies . . . . . . . . . . . . . . . . . 250

6.2.1.3

Observation-specific efficiencies . . . . . . . . . . . . . . . . . . . . 255

Contents
6.2.2

Translog Production Frontier with Constant and Neutral Technological

Change . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
6.2.2.1

6.2.3

Observation-Specific Efficiencies . . . . . . . . . . . . . . . . . . . 257

Translog Production Frontier with Non-Constant and Non-Neutral Technological Change . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260
6.2.3.1

6.2.4
6.3

Observation-Specific Efficiencies . . . . . . . . . . . . . . . . . . . 260

Decomposition of Productivity Growth . . . . . . . . . . . . . . . . . . . . . 263

Data Envelopment Analysis (DEA) . . . . . . . . . . . . . . . . . . . . . . . . . . . 264

1 Introduction
1.1 Objectives of the course and the lecture notes
Knowledge about production technologies and producer behavior is important for politicians,
business organizations, government administrations, financial institutions, the EU, and other national and international organizations who desire to know how contemplated policies and market
conditions can affect production, prices, income, and resource utilization in agriculture as well as
in other industries. The same knowledge is relevant in consultancy of single firms who also want
to compare themselves with other firms and their technology with the best practice technology.
The participants of my courses in the field of econometric production analysis will obtain
relevant theoretical knowledge and practical skills so that they can contribute to the knowledge
about production technologies and producer behavior. After completing my courses in the field
of econometric production analysis, the students should be able to:
use econometric production analysis and efficiency analysis to analyze various real-world

questions,
interpret the results of econometric production analyses and efficiency analyses,
choose a relevant approach for econometric production and efficiency analysis, and
critically evaluate the appropriateness of a specific econometric production analysis or efficiency analysis for analyzing a specific real-world question.
These lecture notes focus on practical applications of econometrics and microeconomic production theory. Hence, they complement textbooks in microeconomic production theory (rather
than substituting them).

1.2 An extremely short introduction to R

Many tutorials for learning R are freely available on-line, e.g. the official Introduction to R
(http://cran.r-project.org/doc/manuals/r-release/R-intro.pdf) or the many tutorials
listed in the categoryContributed Documentation(http://cran.r-project.org/other-docs.
html). Furthermore, many good books are available, e.g. A Beginners Guide to R (Zuur, Ieno,
and Meesters, 2009), R Cookbook (Teetor, 2011), or Applied Econometrics with R (Kleiber
and Zeileis, 2008).

10

1 Introduction

1.2.1 Some commands for simple calculations

R is my favourite pocket calculator. . .
> 2 + 3
[1] 5
> 2 - 3
[1] -1
> 2 * 3
[1] 6
> 2 / 3
[1] 0.6666667
> 2^3
[1] 8
R uses the standard order of evaluation (as in mathematics). One can use parenthesis (round
brackets) to change the order of evaluation.
> 2 + 3 * 4^2
[1] 50
> 2 + ( 3 * ( 4^2 ) )
[1] 50
> ( ( 2 + 3 ) * 4 )^2
[1] 400
In R, the hash symbol (#) can be used to add comments to the code, because the hash symbol
and all following characters in the same line are ignored by R.
> sqrt(2)

# square root

[1] 1.414214
> 2^(1/2)

# the same

11

1 Introduction
[1] 1.414214
> 2^0.5

# also the same

[1] 1.414214
> log(3)

# natural logarithm

[1] 1.098612
> exp(3)

# exponential function

[1] 20.08554
The commands can span multiple lines. They are executed as soon as the command can be
considered as complete.
> 2 +
+

[1] 5
> ( 2
+

3 )

[1] 5

1.2.2 Creating objects and assigning values

> a <- 2
> a
[1] 2
> b <- 3
> b
[1] 3
> a * b
[1] 6
Initially, the arrow symbol (<-, consistent of a smaller than sign and a dash) was used to assign
values to objects. However, in recent versions of R, also the equality sign (=) can be used for this.

12

1 Introduction
> a = 4
> a
[1] 4
> b = 5
> b
[1] 5
> a * b
[1] 20
In these lecture notes, I stick to the traditional assignment operator, i.e. the arrow symbol (<-).
Please note that R is case-sensitive, i.e. R distinguishes between upper-case and lower-case
letters. Therefore, the following commands return error messages:
> A

# NOT the same as "a"

> B

# NOT the same as "b"

> Log(3)

# NOT the same as "log(3)"

> LOG(3)

# NOT the same as "log(3)"

1.2.3 Vectors
> v <- 1:4

# create a vector with 4 elements: 1, 2, 3, and 4

> v
[1] 1 2 3 4
> 2 + v

# adding 2 to each element

[1] 3 4 5 6
> 2 * v

# multiplying each element by 2

[1] 2 4 6 8
> log( v )

# the natural logarithm of each element

[1] 0.0000000 0.6931472 1.0986123 1.3862944

> w <- c( 2, 4, 8, 16 )

# concatenate 4 numbers to a vector

> w

13

1 Introduction
[1]

> v + w
[1]

# element-wise addition
6 11 20

> v * w
[1]

8 16

# element-wise multiplication
8 24 64

> v %*% w

# scalar product (inner product)

[,1]
[1,]

98

> w[2]

# select the second element

[1] 4
> w[c(1,3)]

# select the first and the third element

[1] 2 8
> w[2:4]
[1]

> w[-2]
[1]

# select the second, third, and fourth element

8 16
# select all but the second element
8 16

> length( w )
[1] 4

1.2.4 Simple functions

> sum( w )
[1] 30
> mean( w )
[1] 7.5
> median( w )
[1] 6
> min( w )
[1] 2
> max( w )
[1] 16

14

1 Introduction

1.2.5 Comparing values and boolean values

> a == 2
[1] FALSE
> a != 2
[1] TRUE
> a > 4
[1] FALSE
> a >= 4
[1] TRUE
> w > 3
[1] FALSE

TRUE

TRUE

TRUE

> w == 2^(1:4)
[1] TRUE TRUE TRUE TRUE
> all.equal( w, 2^(1:4) )
[1] TRUE
> w > 3 & w < 6
[1] FALSE

# ampersand = and

TRUE FALSE FALSE

> w < 3 | w > 6

# vertical line = or

[1]

TRUE

TRUE FALSE

TRUE

1.2.6 Data sets (data frames)

The data set women is included in R.
> data( "women" )

# load the data set into the workspace

> women

15

1 Introduction
height weight
1

58

115

59

117

60

120

61

123

62

126

63

129

64

132

65

135

66

139

10

67

142

11

68

146

12

69

150

13

70

154

14

71

159

15

72

164

> names( women )

# display the variable names

[1] "height" "weight"

> dim( women )
[1] 15

# dimension of the data set (rows and columns)

> nrow( women )

# number of rows (observations)

[1] 15
> ncol( women )

# number of columns (variables)

[1] 2
> women[[ "height" ]]

# display the values of variable "height"

[1] 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72
> women$height

# short-cut for the previous command

[1] 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72
> women$height[ 3 ]

# height of the third observation

[1] 60

16

1 Introduction
> women[ 3, "height" ]

# the same

[1] 60
> women[ 3, 1 ]

# also the same

[1] 60
> women[ 1:3, 1 ]

# height of the first three observations

[1] 58 59 60
> women[ 1:3, ]

# all variables of the first three observations

height weight
1

58

115

59

117

60

120

> women$cmHeight <- 2.54 * women$height

# new variable: height in cm

> women$kgWeight <- women$weight / 2.205

# new variable: weight in kg

> women$bmi <- women$kgWeight / ( women$cmHeight / 100 )^2

> women
height weight cmHeight kgWeight

bmi

58

115

147.32 52.15420 24.03067

59

117

149.86 53.06122 23.62685

60

120

152.40 54.42177 23.43164

61

123

154.94 55.78231 23.23643

62

126

157.48 57.14286 23.04152

63

129

160.02 58.50340 22.84718

64

132

162.56 59.86395 22.65364

65

135

165.10 61.22449 22.46110

66

139

167.64 63.03855 22.43112

10

67

142

170.18 64.39909 22.23631

11

68

146

172.72 66.21315 22.19520

12

69

150

175.26 68.02721 22.14711

13

70

154

177.80 69.84127 22.09269

14

71

159

180.34 72.10884 22.17198

15

72

164

182.88 74.37642 22.23836

17

# new variable: BMI

1 Introduction

1.2.7 Functions
In order to execute a function in R, the function name has to be followed by a pair of parenthesis
(round brackets). The documentation of a function (if available) can be obtained by, e.g., typing
at the R prompt a question mark followed by the name of the function.
> ?log
One can read in the documentation of the function log, e.g., that this function has a second
optional argument base, which can be used to specify the base of the logarithm. By default, the
base is equal to the Euler number (e, exp(1)). A different base can be chosen by adding a second
argument, either with or without specifying the name of the argument.
> log( 100, base = 10 )
[1] 2
> log( 100, 10 )
[1] 2

1.2.8 Other useful comands

> class( a )
[1] "numeric"
> class( women )
[1] "data.frame"
> class( women$height )
[1] "numeric"
> ls()
[1] "a"
> rm(w)

# list all objects in the workspace

"b"

"v"

"w"

"women"

# remove an object

> ls()
[1] "a"

"b"

"v"

"women"

18

1 Introduction

1.2.9 Extension packages

Currently (June 12, 2013, 2pm GMT), 4611 extension packages for R are available on CRAN
(Comprehensive R Archive Network, http://cran.r-project.org). When an extension package
is installed, it can be loaded with the command library. The following command loads the R
package foreign that includes function for reading data in various formats.
> library( "foreign" )
Please note that you should cite scientific software packages in your publications if you used them
for obtaining your results (as any other scientific works). You can use the command citation
to find out how an R package should be cited, e.g.:
> citation( "frontier" )
To cite package 'frontier' in publications use:
Tim Coelli and Arne Henningsen (2013). frontier: Stochastic Frontier
Analysis. R package version 1.1-1.
http://CRAN.R-Project.org/package=frontier.
A BibTeX entry for LaTeX users is
@Manual{,
title = {frontier: Stochastic Frontier Analysis},
author = {Tim Coelli and Arne Henningsen},
year = {2013},
note = {R package version 1.1-1},
url = {http://CRAN.R-Project.org/package=frontier},
}

1.2.10 Reading data into R

R can read and import data from many different file formats. This is described in the official
R manual R Data Import/Export (http://cran.r-project.org/doc/manuals/r-release/
R-data.pdf). I usually read my data into R from files in CSV (comma separated values) format.
This can be done by the function read.csv. The command read.csv2 can read files in the
European CSV format (values separated by semicolons, comma as decimal separator). The
functions read.dta, read.spss, and read.xport (all in package foreign) can read STATA binary
files, SPSS data files, and SAS XPORT files, respectively. Functions for reading MS-Excel files
are available, e.g., in the packages XLConnect and xlsx.

19

1 Introduction

1.3 Data sets

In my courses in the field of econometric production analysis, I usually use two data sets: one
cross-sectional data set of French apple producers and a panel data set of rice producers on the
Philippines.

1.3.1 French apple producers

1.3.1.1 Description of the data set
In this course, we will predominantly use a cross-sectional production data set of 140 French
apple producers from the year 1986. These data are extracted from a panel data set that has
been used in an article published by Ivaldi et al. (1996) in the Journal of Applied Econometrics.
The full panel data set is available in the journals data archive: http://www.econ.queensu.
ca/jae/1996-v11.6/ivaldi-ladoux-ossard-simioni/.1
The cross-sectional data set that we will predominantly use in the course is available in the R
package micEcon. It has the name appleProdFr86 and can be loaded by the command:
> data( "appleProdFr86", package = "micEcon" )
The names of the variables in the data set can be obtained by the command names:
> names( appleProdFr86 )
[1] "vCap"

"vLab"

"vMat"

"qApples"

"qOtherOut" "qOut"

[7] "pCap"

"pLab"

"pMat"

"pOut"

"adv"

The data set includes following variables:2

vCap costs of capital (including land)
vLab costs of labor (including remuneration of unpaid family labor)
vMat costs of intermediate materials (e.g. seedlings, fertilizer, pesticides, fuel)
qOut quantity index of all outputs (apples and other outputs)
pCap price index of capital goods
pLab price index of labor
pMat price index of materials
pOut price index of the aggregate output
adv use of advisory service
Please note that variables indicated by

are not in the original data set but are artificially

generated in order to be able to conduct some further analyses with this data set. Variable names
starting with v indicate volumes (values), variable names starting with q indicate quantities, and
variable names starting with p indicate prices.
1

In order to focus on the microeconomic analysis rather than on econometric issues in panel data analysis, we
only use a single year from this panel data set.
This information is also available in the documentation of this data set, which can be obtained by the command:
help( "appleProdFr86", package = "micEcon" ).

20

1 Introduction
1.3.1.2 Abbreviating name of data set
In order to avoid too much typing, give the data set a much shorter name (dat) by creating a
copy of the data set and removing the original data set:
> dat <- appleProdFr86
> rm( appleProdFr86 )
1.3.1.3 Calculation of input quantities
Our data set does not contain input quantities but prices and costs (volumes) of the inputs.
As we will need to know input quantities for many of our analyses, we calculate input quantity
indices based on following identity:
v i = x i wi ,

(1.1)

where wi is the price, xi is the quantity and vi is the volume of the ith input. In R, we can
calculate the input quantities with the following commands:
> dat$qCap <- dat$vCap / dat$pCap
> dat$qLab <- dat$vLab / dat$pLab
> dat$qMat <- dat$vMat / dat$pMat
1.3.1.4 Calculation of total costs and variable costs
Total costs are defined as:
c=

N
X

wi x i ,

(1.2)

i=1

where N denotes the number of inputs. We can calculate the apple producers total costs by
following command:
> dat$cost <- with( dat, vCap + vLab + vMat )
Alternatively, we can calculate the costs by summing up the products of the quantities and the
corresponding prices over all inputs:
> all.equal( dat$cost, with( dat, pCap * qCap + pLab * qLab + pMat * qMat ) )
[1] TRUE
Variable costs are defined as:
cv =

wi xi ,

(1.3)

iN 1

where N 1 is a vector of the indices of the variable inputs. If capital is a quasi-fixed input and
labor and materials are variable inputs, the apple producers variable costs can be calculated by
following command:
> dat$vCost <- with( dat, vLab + vMat )

21

1 Introduction
1.3.1.5 Calculation of profit and gross margin
Profit is defined as:
=py

N
X

wi xi = p y c,

(1.4)

i=1

where all variables are defined as above. We can calculate the apple producers profits by:
> dat$profit <- with( dat, pOut * qOut - cost )
Alternatively, we can calculate the profit by subtracting the products of the quantities and the
corresponding prices of all inputs from the revenues:
> all.equal( dat$cost, with( dat, pCap * qCap + pLab * qLab + pMat * qMat ) )
[1] TRUE
The gross margin (variable profit) is defined as:
v = p y

wi xi = p y cv ,

(1.5)

iN 1

where all variables are defined as above. If capital is a quasi-fixed input and labor and materials
are variable inputs, the apple producers gross margins can be calculated by following command:
> dat$vProfit <- with( dat, pOut * qOut - vLab - vMat )

1.3.2 Rice producers on the Philippines

1.3.2.1 Description of the data set
In the last part of this course, we will use a balanced panel data set of annual data collected
from 43 smallholder rice producers in the Tarlac region of the Philippines between 1990 and 1997.
This data set has the name riceProdPhil and is available in the R package frontier. Detailed
information about these data is available in the documentation of this data set. We can load this
data set with following command:
> data( "riceProdPhil", package = "frontier" )
The names of the variables in the data set can be obtained by the command names:
> names( riceProdPhil )
[1] "YEARDUM"

"FMERCODE" "PROD"

"AREA"

"LABOR"

"NPK"

[7] "OTHER"

"PRICE"

"AREAP"

"LABORP"

"NPKP"

"OTHERP"

"EDYRS"

"HHSIZE"

"NADULT"

"BANRAT"

[13] "AGE"

The following variables are of particular importance for our analysis:

22

1 Introduction
PROD output (tonnes of freshly threshed rice)
AREA area planted (hectares).
LABOR labor used (man-days of family and hired labor)
NPK fertilizer used (kg of active ingredients)
YEARDUM time period (1 = 1990, . . . , 8 = 1997)
In our analysis of the production technology of the rice producers we will use variable PROD as
output quantity and variables AREA, LABOR, and NPK as input quantities.
1.3.2.2 Mean-scaling Quantities
In some model specifications, it is an advantage to use mean-scaled quantities. Therefore, we
create new variables with mean-scaled input and output quantities:
> riceProdPhil$area

<- riceProdPhil$AREA / mean( riceProdPhil$AREA )

> riceProdPhil$labor <- riceProdPhil$LABOR / mean( riceProdPhil$LABOR )

> riceProdPhil$npk

<- riceProdPhil$NPK / mean( riceProdPhil$NPK )

> riceProdPhil$prod

<- riceProdPhil$PROD / mean( riceProdPhil$PROD )

As expected, the sample means of the mean-scaled variables are all one so that their logarithms
are all zero (except for negligible very small rounding errors):
> colMeans( riceProdPhil[ , c( "prod", "area", "labor", "npk" ) ] )
prod
1

area labor
1

npk

> log( colMeans( riceProdPhil[ , c( "prod", "area", "labor", "npk" ) ] ) )

prod

area

labor

npk

0.000000e+00 -1.110223e-16

0.000000e+00

0.000000e+00

1.3.2.3 Logarithmic Mean-scaled Quantities

As we use logarithmic input and output quantities in the Cobb-Douglas and Translog specifications, we can reduce our typing work by creating variables with logarithmic (mean-scaled) input
and output quantities:
> riceProdPhil$lArea

<- log( riceProdPhil$area )

> riceProdPhil$lLabor <- log( riceProdPhil$labor )

> riceProdPhil$lNpk

<- log( riceProdPhil$npk )

> riceProdPhil$lProd

<- log( riceProdPhil$prod )

Please note that the (arithmetic) mean values of the logarithmic mean-scaled variables are not
equal to zero:

23

1 Introduction
> colMeans( riceProdPhil[ , c( "lProd", "lArea", "lLabor", "lNpk" ) ] )
lProd

lArea

lLabor

lNpk

-0.3263075 -0.2718549 -0.2772354 -0.4078492

1.3.2.4 Mean-adjusting the Time Trend
In some model specifications, it is an advantage to have a time trend variable that is zero at the
sample mean. If we subtract the sample mean from our time trend variable, the sample mean of
the adjusted time trend is zero:
> riceProdPhil$mYear

<- riceProdPhil$YEARDUM - mean( riceProdPhil$YEARDUM )

> mean( riceProdPhil$mYear )

[1] 0
1.3.2.5 Specifying Panel Structure
This data set does not include any information about its panel structure. Hence, R would ignore
the panel structure and treat this data set as cross-sectional data collected from 352 different
producers. The command plm.data of the plm package (Croissant and Millo, 2008) can be used
to create data sets that include the information on its panel structure. The following commands
creates a new data set of the rice producers from the Philippines that includes information on the
panel structure, i.e. variable FMERCODE indicates the individual (farmer), and variable YEARDUM
indicated the time period (year):3
> library( "plm" )
> pdat <- plm.data( riceProdPhil, c( "FMERCODE", "YEARDUM" ) )

1.4 Mathematical and statistical methods

1.4.1 Aggregating quantities
Sometimes, it is desirable to aggregate quantities of different goods to an aggregate quantity.
This can be done by a quantity index, e.g. the Laspeyres or Paasche quantity index
xij pi0
x
i i0 pi0

xij pij
,
x
i i0 pij

XjL = Pi

XjP = Pi

(1.6)

where subscript i indicates the good, subscript j indicates the observation, xi0 is the base
quantity, and pi0 is the base price of the ith good, e.g. the sample means.
3

Please note that the specification of variable YEARDUM as the time dimension in the panel data set pdat converts
this variable to a categorical variable. If a numeric time variable is needed, it can be created, e.g., by the
command pdat$year <- as.numeric( pdat$YEARDUM ).

24

1 Introduction
The Paasche and Laspeyres quantity indices of all three inputs in the data set of French apple
producers can be calculated by:
> dat$XP <- with( dat,
+

( vCap + vLab + vMat ) /

( mean( qCap ) * pCap + mean( qLab ) * pLab + mean( qMat ) * pMat ) )

> dat$XL <- with( dat,

+

( qCap * mean( pCap ) + qLab * mean( pLab ) + qMat * mean( pMat ) ) /

( mean( qCap ) * mean( pCap ) + mean( qLab ) * mean( pLab ) +

mean( qMat ) * mean( pMat ) ) )

In many cases, the choice of the formula for calculating quantity indices does not have a major
influence on the result. We demonstrate this with two scatter plots, where we set argument log
of the second plot command to the character string "xy" so that both axes are measured in
logarithmic terms and the dots (firms) are more equally spread:
> plot( dat$XP, dat$XL )

5.0

> plot( dat$XP, dat$XL, log = "xy" )

2.0

0.5

1.0

XL

3
1

XL

0.5

XP

1.0

2.0

5.0

XP

Figure 1.1: Comparison of Paasche and Laspeyres quantity indices

The resulting scatter plots are shown in figure 1.1.
As a compromise, one can use the Fisher quantity index, which is the geometric mean of the
Paasche quantity index and the Laspeyres quantity index:
> dat$X <- sqrt( dat$XP * dat$XL )
We can can also use function quantityIndex from the micEcon package to calculate the quantity index:
> library( "micEcon" )
> dat$XP2 <- quantityIndex( c( "pCap", "pLab", "pMat" ),

25

1 Introduction
+

c( "qCap", "qLab", "qMat" ), data = dat, method = "Paasche" )

> all.equal( dat$XP, dat$XP2, check.attributes = FALSE )

[1] TRUE
> dat$XL2 <- quantityIndex( c( "pCap", "pLab", "pMat" ),
+

c( "qCap", "qLab", "qMat" ), data = dat, method = "Laspeyres" )

> all.equal( dat$XL, dat$XL2, check.attributes = FALSE )

[1] TRUE
> dat$X2 <- quantityIndex( c( "pCap", "pLab", "pMat" ),
+

c( "qCap", "qLab", "qMat" ), data = dat, method = "Fisher" )

> all.equal( dat$X, dat$X2, check.attributes = FALSE )

[1] TRUE

1.4.2 Quasiconcavity
A function f (x) : RN R is quasiconcave if its level plots (isoquants) are convex. This is the
case if
f (xl + (1 )xu ) min(f (xl ), f (xu ))

(1.7)

for any combination of xl , xu , and with 0 1 (Chambers, 1988, p. 311).

If f (x) is a continuous and twice-continuously differentiable function, a necessary condition for
quasiconcavity is |B1 | 0, |B2 | 0, |B3 | 0, . . . , (1)N |BN | 0, where |Bi | is the ith principal
minor of the bordered Hessian

f1

f2

...

fN

f1

B=
f2
.
..

f11

f12

...

f12
..
.

f22
..
.

...
..
.

f1N

fN

f1N

f2N

. . . fN N

f2N
..
.

(1.8)

fi denotes the partial derivative of f (x) with respect to xi , fij denotes the second partial derivative
of f (x) with respect to xi and xj , |B1 | is the determinant of the upper left 2 2 sub-matrix of B,
|B2 | is the determinant of the upper left 3 3 sub-matrix of B, . . . , and |BN | is the determinant
of B (Chambers, 1988, p. 312; Chiang, 1984, p. 393f).

1.4.3 Delta method

If we have estimated a parameter vector and its variance covariance matrix V ar() and we
calculate a vector of measures (e.g. elasticities) based on the estimated parameters by z = g(),

26

1 Introduction
we can calculate the approximate variance covariance matrix of z by:


V ar(z)

g()

>

V ar()

g()
,

(1.9)

where g()/ is the Jacobian matrix of z = g() with respect to and the superscript > is
the transpose operator.

27

2 Primal Approach: Production Function

2.1 Theory
2.1.1 Production function
The production function
y = f (x)

(2.1)

indicates the maximum quantity of a single output (y) that can be obtained with a vector of
given input quantities (x). It is usually assumed that production functions fulfill some properties
(see Chambers, 1988, p. 9).

2.1.2 Average Products

Very simple measures to compare the (partial) productivities of different firms are the inputs
average products. The average product of the ith input is defined as:
APi =

f (x)
y
=
xi
xi

(2.2)

The more output one firm produces per unit of input, the more productive is this firm and
the higher is the corresponding average product. If two firms use identical input quantities,
the firm with the larger output quantity is more productive (has a higher average product).
And if two firms produce the same output quantity, the firm with the smaller input quantity is
more productive (has a higher average product). However, if these two firms use different input
combinations, one firm could be more productive regarding the average product of one input,
while the other firm could be more productive regarding the average product of another input.

2.1.3 Total Factor Productivity

As average products measure just partial productivities, it is often desirable to calculate total
factor productivities (TFP):
TFP =

y
,
X

where X is a quantity index of all inputs (see section 1.4.1).

28

(2.3)

2 Primal Approach: Production Function

2.1.4 Marginal Products

The marginal productivities of the inputs can be measured by their marginal products. The
marginal product of the ith input is defined as:
M Pi =

f (x)
xi

(2.4)

2.1.5 Output elasticities

The marginal productivities of the inputs can also be measured by their output elasticities. The
output elasticity of the ith input is defined as:
i =

f (x) xi
MP
=
xi f (x)
AP

(2.5)

In contrast to the marginal products, the changes of the input and output quantities are measured
in relative terms so that output elasticities are independent of the units of measurement. Output
elasticities are sometimes also called partial output elasticities or partial production elasticities.

2.1.6 Elasticity of scale

The returns of scale of the technology can be measured by the elasticity of scale:
=

(2.6)

If the technology has increasing returns to scale ( > 1), total factor productivity increases
when all input quantities are proportionally increased, because the relative increase of the output
quantity y is larger than the relative increase of the aggregate input quantity X in equation (2.3).
If the technology has decreasing returns to scale ( < 1), total factor productivity decreases when
all input quantities are proportionally increased, because the relative increase of the output
quantity y is less than the relative increase of the aggregate input quantity X. If the technology
has constant returns to scale ( = 1), total factor productivity remains constant when all input
quantities change proportionally, because the relative change of the output quantity y is equal to
the relative change of the aggregate input quantity X.
If the elasticity of scale (monotonically) decreases with firm size, the firm has the most productive scale size at the point, where the elasticity of scale is one.

29

2 Primal Approach: Production Function

2.1.7 Marginal rates of technical substitution

The marginal rate of technical substitution between input i and input j is (Chambers, 1988,
p. 29):
y
xi
M Pj
xj
=
=
=
y
xj
M Pi
xi

M RT Si,j

(2.7)

2.1.8 Relative marginal rates of technical substitution

The relative marginal rate of technical substitution between input i and input j is:
y
xi xj
xj
=
=
y
xj xi
xi

RM RT Si,j

xj
j
y
xi =
i
y

(2.8)

2.1.9 Elasticities of substitution

The elasticity of substitution measures the substitutability between two inputs. It is defined as:

ij =

xi
M Pj
x
 j  MxPi
i
MP
d M Pji
xj

d
=

xi
xj

M RT Sij

d (M RT Sij )

xi
xj

d
=

xi
xj

d M RT Sij

M RT Sij
xi
xj

(2.9)

Thus, if input i is substituted for input j so that the input ratio xi /xj increases by ij %, the
marginal rate of technical substitution between input i and input j will increase by 1%.
2.1.9.1 Direct Elasticities of Substitution
The direct elasticity of substitution can be calculated by:
D
ij
=

fi xi + fj xj Fij
,
xi xj
F

(2.10)

where F is the determinant of the bordered Hessian matrix B with

f1

f2

...

fN

f1

B=
f2
.
..

f11

f12

...

f12
..
.

f22
..
.

...
..
.

f1N

fN

f1N

f2N

. . . fN N

30

f2N
..
.

(2.11)

2 Primal Approach: Production Function

Fij is the co-factor of fij , i.e.1












Fij = (1)i+j 











f1

f2

...

fj1

fj+1

...

f1

f11

f12

...

f1,j1

f1,j+1

...

f2
..
.

f12
..
.

f22
..
.

...
..
.

f2,j1
..
.

f2,j+1
..
.

...
..
.

fi1 f1,i1 f2,i1 . . . fi1,j1 fi1,j+1 . . .

fi+1 f1,i+1 f2,i+1 . . . fi+1,j1 fi+1,j+1 . . .
..
..
..
..
..
..
..
.
.
.
.
.
.
.
fN

f1N

f2N

...

fj1,N

fj+1,N

...





f1N 

f2N 

..

.

,
fi1,N 

fi+1,N 

..

.


fN N 

fN

(2.12)

fi is the partial derivative of the production function f with respect to the ith input quantity
(xi ), and fij is the second partial derivative of the production function f with respect to the ith
and jth input quantity (xi , xj ).
As the bordered Hessian matrix is symmetric, the co-factors are also symmetric (Fij = Fji ) so
D = D ).
that also the direct elasticities of substitution are symmetric (ij
ji

2.1.9.2 Allen Elasticities of Substitution

The Allen elasticity of substitution is another measure of the substitutability between two inputs.
It can be calculated by:
P

k fk

xk Fij
,
xi xj F

ij =

(2.13)

where Fij and F are defined as above.

As with the direct elasticities of substitution, also the Allen elasticities of substitution are
symmetric (ij = ji ).
The Allen elasticities of substitution are related to the direct elasticities of substitution in the
following way:
D
ij

fi xi + fj xj
= P
k fk xk

fi xi + fj xj
k fk xk Fij
= P
ij
xi xj F
k fk xk

(2.14)

As the input quantities and the marginal products should always be positive, the direct elasticities
of substitution and the Allen elasticities of substitution always have the same sign and the direct
elasticities of substitution are always smaller than the Allen elasticities of substitution in absolute
D | | |.
terms, i.e. |ij
ij

Following condition holds for Allen elasticities of substitution:

X
i

fi xi
Ki ij = 0 with Ki = P
k fk xk

(2.15)

The exponent of (1) usually is the sum of the number of the deleted row (i + 1) and the number of the deleted
column (j + 1), i.e. i + j + 2. In our case, we can simplify this to i + j, because (1)i+j+2 = (1)i+j (1)2 =
(1)i+j .

31

2 Primal Approach: Production Function

(see Chambers, 1988, p. 35).
2.1.9.3 Morishima Elasticities of Substitution
The Morishima elasticity of substitution is a third measure of the substitutability between two
inputs. It can be calculated by:
M
ij
=

fj Fij
fj Fjj

,
xi F
xj F

(2.16)

where Fij and F are defined as above. In contrast to the direct elasticity of substitution and the
Allen elasticity of substitution, the Morishima elasticity of substitution is usually not symmetric
M 6= M ).
(ij
ji

From the above definition of the Morishima elasticities of substitution (2.16), we can derive
the relationship between the Morishima elasticities of substitution and the Allen elasticities of
substitution:
M
ij

fj xj
fj xj
k fk xk Fij
=P
P
f
x
x
x
F
i j
k k k
k fk xk
fj xj
fj xj
ij P
jj
=P
f
x
k k k
k fk xk
fj xj
(ij jj ) ,
=P
k fk xk

k fk
x2j

xk Fjj
F

(2.17)
(2.18)
(2.19)

where jj can be calculated as the Allen elasticities of substitution with equation (2.13), but does
not have an economic meaning.

2.1.10 Profit Maximization

We assume that the firms maximize their profit. The firms profit is given by
=py

wi xi ,

(2.20)

where p is the price of the output and wi is the price of the ith input. If the firm faces output
price p and input prices wi , we can calculate the maximum profit that can be obtained by the
firm by solving following optimization problem:
max p y
y,x

wi xi , s.t. y = f (x)

(2.21)

This restricted maximization can be transformed into an unrestricted optimization by replacing

y by the production function:
max p f (x)
x

X
i

32

wi x i

(2.22)

2 Primal Approach: Production Function

Hence, the first-order conditions are:
f (x)

=p
wi = p M Pi wi = 0
xi
xi

(2.23)

wi = p M Pi = M V Pi

(2.24)

so that we get

where M V Pi = p (y/xi ) is the marginal value product of the ith input.

2.1.11 Cost Minimization

Now, we assume that the firms take total output as given (e.g. because production is restricted
by a quota) and try to produce this output quantity with minimal costs. The total cost is given
by
c=

wi x i ,

(2.25)

where wi is the price of the ith input.

If the firm faces input prices wi and wants to produce y units of output, the minimum costs
can be obtained by
min

wi xi , s.t. y = f (x)

(2.26)

This restricted minimization can be solved by using the Lagrangian approach:

L=

wi xi + (y f (x))

(2.27)

So that the first-order conditions are:

L
f (x)
= wi
= wi M Pi = 0
xi
xi
L
= y f (x) = 0

(2.28)
(2.29)

From the first-order conditions (2.28), we get:

and

wi = M Pi

(2.30)

wi
M Pi
M Pi
=
=
= M RT Sji
wj
M Pj
M Pj

(2.31)

As profit maximization implies producing the optimal output quantity with minimum costs,
the first-order conditions for the optimal input combinations (2.31) can be obtained not only

33

2 Primal Approach: Production Function

from cost minimization but also from the first-order conditions for profit maximization (2.24):
wi
M V Pi
p M Pi
M Pi
=
=
=
= M RT Sji
wj
M V Pj
p M Pj
M Pj

(2.32)

2.1.12 Derived Input Demand Functions and Output Supply Functions

In this section, we will analyze how profit maximizing or cost minimizing firms react on changing
prices and on changing output quantities.
2.1.12.1 Derived from profit maximization
If we replace the marginal products in the first-order conditions for profit maximization (2.24)
by the equations for calculating these marginal products and then solve this system of equations
for the input quantities, we get the input demand functions:
xi = xi (p, w),

(2.33)

where w = [wi ] is the vector of all input prices. The input demand functions indicate the optimal
input quantities (xi ) given the output price (p) and all input prices (w). We can obtain the
output supply function from the production function by replacing all input quantities by the
corresponding input demand functions:
y = f (x(p, w)) = y(p, w),

(2.34)

where x(p, w) = [xi (p, w)] is the set of all input demand functions. The output supply function
indicates the optimal output quantity (y) given the output price (p) and all input prices (w).
Hence, the input demand and output supply functions can be used to analyze the effects of prices
on the (optimal) input use and output supply. In economics, the effects of price changes are
usually measured in terms of price elasticities. These price elasticities can measure the effects of
the input prices on the input quantities:
ij (p, w) =

xi (p, w)
wj
,
wj
xi (p, w)

(2.35)

the effects of the input prices on the output quantity (expected to be non-positive):
yj (p, w) =

y(p, w) wj
,
wj
y(p, w)

(2.36)

the effects of the output price on the input quantities (expected to be non-negative):
ip (p, w) =

xi (p, w)
p
,
p
xi (p, w)

34

(2.37)

2 Primal Approach: Production Function

and the effect of the output price on the output quantity (expected to be non-negative):
yp (p, w) =

p
y(p, w)
.
p
y(p, w)

(2.38)

The effect of an input price on the optimal quantity of the same input is expected to be nonpositive (ii (p, w) 0). If the cross-price elasticities between two inputs i and j are positive
(ij (p, w) 0, ji (p, w) 0), they are considered as gross substitutes. If the cross-price elasticities
between two inputs i and j are negative (ij (p, w) 0, ji (p, w) 0), they are considered as gross
complements.
2.1.12.2 Derived from cost minimization
If we replace the marginal products in the first-order conditions for cost minimization (2.30) by
the equations for calculating these marginal products and the solve this system of equations for
the input quantities, we get the conditional input demand functions:
xi = xi (w, y)

(2.39)

These input demand functions are called conditional, because they indicate the optimal input
quantities (xi ) given all input prices (w) and conditional on the fixed output quantity (y). The
conditional input demand functions can be used to analyze the effects of input prices on the
(optimal) input use if the output quantity is given. The effects of price changes on the optimal
input quantities can be measured by conditional price elasticities:
ij (w, y) =

xi (w, y)
wj
wj
xi (w, y)

(2.40)

The effect of the output quantity on the optimal input quantities can also be measured in terms
of elasticities (expected to be positive):
iy (w, y) =

xi (w, y)
y
.
y
xi (w, y)

(2.41)

The conditional effect of an input price on the optimal quantity of the same input is expected
to be non-positive (ii (w, y) 0). If the conditional cross-price elasticities between two inputs i
and j are positive (ij (w, y) 0, ji (w, y) 0), they are considered as net substitutes. If
the conditional cross-price elasticities between two inputs i and j are negative (ij (w, y) 0,
ji (w, y) 0), they are considered as net complements.

35

2 Primal Approach: Production Function

2.2 Productivity Measures

2.2.1 Average Products
We calculate the average products of the three inputs for each firm in the data set by equation 2.2:
> dat$apCap <- dat$qOut / dat$qCap
> dat$apLab <- dat$qOut / dat$qLab
> dat$apMat <- dat$qOut / dat$qMat
We can visualize these average products with histograms that can be created with the command
hist.
> hist( dat$apCap )
> hist( dat$apLab )

50

100

150

50
40
30
10
0

10
0
0

20

Frequency

15
5

10

Frequency

40
30
20

Frequency

50

20

60

> hist( dat$apMat )

10

apCap

15

20

25

apLab

50

150

250

350

apMat

Figure 2.1: Average products

The resulting graphs are shown in figure 2.1. These graphs show that average products (partial
productivities) vary considerably between firms. Most firms in our data set produce on average
between 0 and 40 units of output per unit of capital, between 2 and 16 units of output per unit
of labor, and between 0 and 100 units of output per unit of materials. Looking at each average
product separately, There are usually many firms with medium to low productivity and only a
few firms with high productivity.
The relationships between the average products can be visualized by scatter plots:
> plot( dat$apCap, dat$apLab )
> plot( dat$apCap, dat$apMat )
> plot( dat$apLab, dat$apMat )
The resulting graphs are shown in figure 2.2. They show that the average products of the three
inputs are positively correlated.

36

50

100

150

100

150

50

dat$apCap

dat$apCap

200

100

50

50

15
10
0

dat$apMat

300

200

20

300

dat$apLab

100

dat$apMat

25

2 Primal Approach: Production Function

10

15

20

25

dat$apLab

Figure 2.2: Relationships between average products

As the units of measurements of the input and output quantities in our data set cannot be
interpreted in practical terms, the interpretation of the size of the average products is practically
not useful. However, they can be used to make comparisons between firms. For instance, the
interrelation between average products and firm size can be analyzed. A possible (although not
perfect) measure of size of the firms in our data set is the total output.
> plot( dat$qOut, dat$apCap, log = "x" )
> plot( dat$qOut, dat$apLab, log = "x" )

1e+05

5e+05

5e+06

1e+05

5e+05

qOut

5e+06

200

100

50

20
15

10

apLab

100

50

apCap

300

apMat

25

150

> plot( dat$qOut, dat$apMat, log = "x" )

1e+05

qOut

5e+05

5e+06
qOut

Figure 2.3: Average products for different firm sizes

The resulting graphs are shown in figure 2.3. These graphs show that the larger firms (i.e. firms
with larger output quantities) produce also a larger output quantity per unit of each input. This
is not really surprising, because the output quantity is in the numerator of equation (2.2) so that
the average products are necessarily positively related to the output quantity for a given input
quantity.

37

2 Primal Approach: Production Function

2.2.2 Total Factor Productivity

After calculating a quantity index of all inputs (see section 1.4.1), we can use equation 2.3 to
calculate the total factor productivity, where we arbitrarily choose the Fisher quantity index:
> dat$tfp <- dat$qOut / dat$X
The variation of the total factor productivities can be visualized as before in a histogram:

0e+00

0
0e+00 2e+06 4e+06 6e+06

1e+05

5e+05

TFP

5e+06
qOut

4e+06

2e+06

6e+06

TFP

0e+00

2e+06

4e+06

TFP

30
20
10

Frequency

40

6e+06

> hist( dat$tfp )

0.5

1.0

2.0

5.0

Figure 2.4: Total factor productivities

The resulting histogram is shown in the left panel of figure 2.4. It indicates that also total factor
productivity varies considerably between firms.
Where do these large differences in (total factor) productivity come from? We can check the
relation between total factor productivity and firm size with a scatter plot. We use two different
measures of firm size, i.e. total output and aggregate input. The following commands produce
scatter plots, where we set argument log of the plot command to the character string "x" so
that the horizontal axis is measured in logarithmic terms and the dots (firms) are more equally
spread:
> plot( dat$qOut, dat$tfp, log = "x" )
> plot( dat$X, dat$tfp, log = "x" )
The resulting scatter plots are shown in the middle and right panel of figure 2.4. This graph clearly
shows that the firms with larger output quantities also have a larger total factor productivity.
This is not really surprising, because the output quantity is in the numerator of equation (2.3) so
that the total factor productivity is necessarily positively related to the output quantity for given
input quantities. The total factor productivity is only slightly positively related to the measure
of aggregate input use.
We can also analyze whether the firms that use an advisory service have a higher total factor
productivity than firms that do not use an advisory service. We can visualize and compare the

38

2 Primal Approach: Production Function

total factor productivities of the two different groups of firms (with and without advisory service)
using boxplot diagrams:
> boxplot( tfp ~ adv, data = dat )
> boxplot( log(qOut) ~ adv, data = dat )

6e+06

1.5

17

7e+06

> boxplot( log(X) ~ adv, data = dat )

1e+06

0.5
0.5

0.0

log(X)

13

14

3e+06

log(qOut)

4e+06

15

2e+06

12

1.0

TFP

5e+06

1.0

16

0e+00

no advisory

advisory

no advisory

advisory

no advisory

advisory

Figure 2.5: Total factor productivities and advisory service

The resulting boxplot graphic is shown on the left panel of figure 2.5. It suggests that the firms
that use advisory service are slightly more productive than firms that do not use advisory service
(at least when looking at the 25th percentile and the median).
However, these boxplots can only indicate a relationship between using advisory service and
total factor productivity but they cannot indicate whether using an advisory service increases
productivity (i.e. a causal effect). For instance, if larger firms are more likely to use an advisory
service than smaller firms and larger firms have a higher total factor productivity than smaller
firms, we expect that firms that use an advisory service have a higher productivity than smaller
firms even if using an advisory service does not affect total factor productivity. However, this
is not the case in our data set, because farms with and without advisory service use rather
similar input quantities (see right panel of figure 2.5). As farms that use advisory service use
similar input quantities but have a higher total factor productivity than farms without advisory
service (see left panel of figure 2.5), they also have larger output quantities than corresponding
farms without advisory service (see middle panel of figure 2.5). Furthermore, the causal effect
of advisory service on total factor productivity might not be equal to the productivity difference
between farms with and without advisory service, because it might be that the firms that anyway
were the most productive were more (or less) likely to use advisory service than the firms that
anyway were the least productive.

39

2 Primal Approach: Production Function

2.3 Linear Production Function

2.3.1 Specification
A linear production function with N inputs is defined as:
y = 0 +

N
X

i xi

(2.42)

i=1

2.3.2 Estimation
We can add a stochastic error term to this linear production function and estimate it for our data
set using the command lm:
> prodLin <- lm( qOut ~ qCap + qLab + qMat, data = dat )
> summary( prodLin )
Call:
lm(formula = qOut ~ qCap + qLab + qMat, data = dat)
Residuals:
Min

1Q

Median

3Q

Max

-3888955

-773002

86119

769073

7091521

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) -1.616e+06

2.318e+05

-6.972 1.23e-10 ***

qCap

1.788e+00

1.995e+00

0.896

qLab

1.183e+01

1.272e+00

9.300 3.15e-16 ***

qMat

4.667e+01

1.123e+01

4.154 5.74e-05 ***

0.372

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Residual standard error: 1541000 on 136 degrees of freedom

Multiple R-squared:

0.7868,

F-statistic: 167.3 on 3 and 136 DF,

Adjusted R-squared:

0.7821

p-value: < 2.2e-16

2.3.3 Properties
As the coefficients of all three input quantities are positive, the monotonicity condition is (globally) fulfilled. However, the coefficient of the capital quantity is statistically not significantly
different from zero. Therefore, we cannot be sure that the capital quantity has a positive effect
on the output quantity.

40

2 Primal Approach: Production Function

As every linear function is concave (and convex), also our estimated linear production is concave
and hence, also quasi-concave. As the isoquants of linear productions functions are linear, the
input requirement sets are always convex (and concave).
Our estimated linear production function does not fulfill the weak essentiality assumption,
because the intercept is different from zero. The production technology described by a linear
production function with more than one (relevant) input never shows strict essentiality.
The input requirement sets derived from linear production functions are always closed and
non-empty for y > 0 if weak essentiality is fulfilled (0 = 0) and strict monotonicity is fulfilled
for at least one input ( i {1, . . . , N } : i > 0), as the input quantities must be non-negative
(xi 0 i).
The linear production function always returns finite, real, and single values for all non-negative
and finite x. However, as the intercept of our estimated production function is negative, the nonnegativity assumption is not fulfilled. A linear production function would return non-negative
values for all non-negative and finite x if 0 0 and the monotonicity condition is fulfilled
(i 0 i = 1, . . . , N ).
All linear production functions are continuous and twice-continuously differentiable.

2.3.4 Predicted output quantities

We can calculate the predicted (fitted) output quantities manually by taking the linear production function (2.42), the observed input quantities, and the estimated parameters, but it is
easier to use the fitted method to obtain the predicted values of the dependent variable from
an estimated model:
> dat$qOutLin <- fitted( prodLin )
> all.equal( dat$qOutLin, coef( prodLin )[ "(Intercept)" ] +
+

coef( prodLin )[ "qCap" ] * dat$qCap +

coef( prodLin )[ "qLab" ] * dat$qLab +

coef( prodLin )[ "qMat" ] * dat$qMat )

[1] TRUE
We can evaluate the fit of the model by comparing the observed with the fitted output
quantities:
> compPlot( dat$qOut, dat$qOutLin )
> compPlot( dat$qOut[ dat$qOutLin > 0 ], dat$qOutLin[ dat$qOutLin > 0 ],
+

log = "xy" )

The resulting graphs are shown in figure 2.6. While the graph in the left panel uses a linear
scale for the axes, the graph in the right panel uses a logarithmic scale for both axes. Hence, the
deviations from the 45-line illustrate the absolute deviations in the left panel and the relative

41

0.0e+00

2e+05

1.0e+07

2e+04

2e+06

fitted

1.0e+07
0.0e+00

fitted

2.0e+07

2e+07

2 Primal Approach: Production Function

2.0e+07

2e+04

2e+05

observed

2e+06

2e+07

observed

Figure 2.6: Linear production function: fit of the model

deviations in the right panel. As the logarithm of non-positive values is undefined, we have to
exclude observations with non-positive predicted output quantities in the graphs with logarithmic
axes. The fit of the model looks okay in both scatter plots.
As negative output quantities would render the corresponding output elasticities useless, we
have carefully check the sign of the predicted output quantities:
> sum( dat$qOutLin < 0 )
[1] 1
One predicted output quantity is negative.

2.3.5 Marginal Products

In the linear production function, the marginal products are equal to coefficients of the corresponding input quantities.
M Pi =

y
= i
xi

(2.43)

Hence, if a firm increases capital input by one unit, the output will increase by 1.79 units; if
a firm increases labor input by one unit, the output will increase by 11.83 units; and if a firm
increases materials input by one unit, the output will increase by 46.67 units.

2.3.6 Output Elasticities

As we do not know the units of measurements of the input and output quantities, the interpretation of the marginal products is practically not very useful. Therefore, we calculate the output
elasticities (partial production elasticities) of the three inputs.
i =

y xi
xi
M Pi
= M Pi
=
xi y
y
APi

42

(2.44)

2 Primal Approach: Production Function

As the output elasticities depend on the input and output quantities and these quantities generally
differ between firms, also the output elasticities differ between firms. Hence, we can calculate
them for each firm in the sample:
> dat$eCap <- coef(prodLin)["qCap"] * dat$qCap / dat$qOut
> dat$eLab <- coef(prodLin)["qLab"] * dat$qLab / dat$qOut
> dat$eMat <- coef(prodLin)["qMat"] * dat$qMat / dat$qOut
We can obtain their mean values by:
> colMeans( subset( dat, , c( "eCap", "eLab", "eMat" ) ) )
eCap

eLab

eMat

0.1202721 2.0734793 0.8631936

However, these mean values are distorted by outliers (see figure 2.7). Therefore, we calculate the
median values of the the output elasticities:
> library( "miscTools" )
> colMedians( subset( dat, , c( "eCap", "eLab", "eMat" ) ) )
eCap

eLab

eMat

0.08063406 1.28627208 0.58741460

Hence, if a firm increases capital input by one percent, the output will usually increase by around
0.08 percent; if the firm increases labor input by one percent, the output will often increase by
around 1.29 percent; and if the firm increases materials input by one percent, the output will
often increase by around 0.59 percent.
We can visualize (the variation of) these output elasticities with histograms. The user can
modify the desired number of bars in the histogram by adding an integer number as additional
argument:
> hist( dat$eCap )
> hist( dat$eLab, 20 )
> hist( dat$eMat, 20 )
The resulting graphs are shown in figure 2.7. If the firms increase capital input by one percent,
the output of most firms will increase by between 0 and 0.2 percent; if the firms increase labor
input by one percent, the output of most firms will increase by between 0.5 and 3 percent;
and if the firms increase materials input by one percent, the output of most firms will increase
by between 0.2 and 1.2 percent. While the marginal effect of capital on the output is rather
small for most firms, there are many firms with implausibly high output elasticities of labor and
materials (i > 1). This might indicate that the true production technology cannot be reasonably
approximated by a linear production function.

43

0.4

0.8

1.2

10
0

10
0

0
0.0

20

Frequency

30

40
30
20

Frequency

60
40
20

Frequency

80

2 Primal Approach: Production Function

eCap

10 12 14

eLab

eMat

Figure 2.7: Linear production function: output elasticities

In contrast to a pure theoretical microeconomic model, our empirically estimated model includes a stochastic error term so that the observed output quantities (y) are not necessarily equal
to the output quantities that are predicted by the model (
y = f (x)). This error term comes
from, e.g., measurement errors, omitted explanatory variables, (good or bad) luck, or unusual(ly)
(good or bad) weather conditions. The better the fit of our model, i.e. the higher the R2 value,
the smaller is the difference between the observed and the predicted output quantities. If we
believe in our estimated model, it would be more consistent with microeconomic theory, if we
use the predicted output quantities and disregard the stochastic error term.
We can calculate the output elasticities based on the predicted output quantities (see section 2.3.4) rather than the observed output quantities:
> dat$eCapFit <- coef(prodLin)["qCap"] * dat$qCap / dat$qOutLin
> dat$eLabFit <- coef(prodLin)["qLab"] * dat$qLab / dat$qOutLin
> dat$eMatFit <- coef(prodLin)["qMat"] * dat$qMat / dat$qOutLin
> colMeans( subset( dat, , c( "eCapFit", "eLabFit", "eMatFit" ) ) )
eCapFit

eLabFit

eMatFit

0.1421941 2.2142092 0.9784056

> colMedians( subset( dat, , c( "eCapFit", "eLabFit", "eMatFit" ) ) )
eCapFit

eLabFit

eMatFit

0.07407719 1.21044421 0.58821500

> hist( dat$eCapFit, 20 )
> hist( dat$eLabFit, 20 )
> hist( dat$eMatFit, 20 )
The resulting graphs are shown in figure 2.8. While the choice of the variable for the output
quantity (observed vs. predicted) only has a minor effect on the mean and median values of the

44

80 100
60
20
0

20
0

0
0.0 0.5 1.0 1.5 2.0 2.5 3.0

40

Frequency

120
80
60
40

Frequency

80 100
60
40
20

Frequency

2 Primal Approach: Production Function

10

10

eCapFit

20

30

40

50

eLabFit

10

15

20

25

eMatFit

Figure 2.8: Linear production function: output elasticities based on predicted output quantities
output elasticities, the ranges of the output elasticities that are calculated from the predicted
output quantities are much larger than the ranges of the output elasticities that are calculated
from the observed output quantities. Due to 1 negative predicted output quantity, the output
elasticities of this observation are also negative.

2.3.7 Elasticity of Scale

The elasticity of scale is the sum of all output elasticities
=

i

(2.45)

Hence, the elasticities of scale of all firms in the sample can be calculated by:
> dat$eScale <- with( dat, eCap + eLab + eMat )
> dat$eScaleFit <- with( dat, eCapFit + eLabFit + eMatFit )
The mean and median values of the elasticities of scale can be calculated by
> colMeans( subset( dat, , c( "eScale", "eScaleFit" ) ) )
eScale eScaleFit
3.056945

3.334809

> colMedians( subset( dat, , c( "eScale", "eScaleFit" ) ) )

eScale eScaleFit
1.941536

1.864253

Hence, if a firm increases all input quantities by one percent, the output quantity will usually
increase by around 1.9 percent. This means that most firms have increasing returns to scale and
hence, the firms could increase productivity by increasing the firm size (i.e. increasing all input
quantities).
The (variation of the) elasticities of scale can be visualized with histograms:

45

2 Primal Approach: Production Function

> hist( dat$eScale, 30 )
> hist( dat$eScaleFit, 50 )

10

20

Frequency

30

60
20

40

Frequency

20
15
10
0

10

15

Frequency

25

40

30

> hist( dat$eScaleFit[ dat$eScaleFit > 0 & dat$eScaleFit < 15 ], 30 )

eScale

20

40

60

80

eScaleFit

10

12

14

0 < eScaleFit < 15

Figure 2.9: Linear production function: elasticities of scale

The resulting graphs are shown in figure 2.9. As the predicted output quantity of 1 firm is negative, the elasticity of scale of this observation also is negative, if the predicted output quantities
are used for the calculation. However, all remaining elasticities of scale that are based on the
predicted output quantities are larger than one, which indicates increasing returns to scale. In
contrast, 15 (out of 140) elasticities of scale that are calculated with the observed output quantities indicate decreasing returns to scale. However, both approaches indicate that most firms have
an elasticity of scale between one and two. Hence, if these firms increase all input quantities by
one percent, the output of most firms will increase by between 1 and 2 percent. Some firms even
have an elasticity of scale larger than five, which is very implausible and might indicate that the
true production technology cannot be reasonably approximated by a linear production function.
Information on the optimal firm size can be obtained by analyzing the interrelationship between
firm size and the elasticity of scale:
> plot( dat$qOut, dat$eScale, log = "x" )
> abline( 1, 0 )
> plot( dat$X, dat$eScale, log = "x" )
> abline( 1, 0 )
> plot( dat$qOut, dat$eScaleFit, log = "x", ylim = c( 0, 15 ) )
> abline( 1, 0 )
> plot( dat$X, dat$eScaleFit, log = "x", ylim = c( 0, 15 ) )
> abline( 1, 0 )
The resulting graphs are shown in figure 2.10. They indicate that very small firms could enormously gain from increasing their size, while the benefits from increasing firm size decrease with
size. Only a few elasticities of scale that are calculated with the observed output quantities indicate decreasing returns to scale so that productivity would decline when these firms increase their

46

2 Primal Approach: Production Function

15

15

10

eScale

10

eScale

0.5

1.0

2.0

5.0

1e+05

5e+05

2e+06

eScaleFit

15

10

2e+07

qOut

eScaleFit

10

15

5e+06

0.5

1.0

2.0

5.0

1e+05

5e+05

2e+06

5e+06

qOut

Figure 2.10: Linear production function: elasticities of scale for different firm sizes

47

2e+07

2 Primal Approach: Production Function

size. For all firms that use at least 2.1 times the input quantities of the average firm or produces
more than 6,000,000 quantity units (approximately 6,000,000 Euros), the elasticities of scale that
are based on the observed input quantities are very close to one. From this observation we could
conclude that firms have their optimal size when they use at least 2.1 times the input quantities
of the average firm or produce at least 6,000,000 quantity units (approximately 6,000,000 Euros
turn over). In contrast, the elasticities of scale that are based on the predicted output quantities
are larger one even for the largest firms in the data set. From this observation, we could conclude
that the even the largest firms in the sample would gain from growing in size and thus, the most
productive scale size is lager than the size of the largest firms in the sample.
The high elasticities of scale explain why we found much higher partial productivities (average
products) and total factor productivities for larger firms than for smaller firms.

2.3.8 Marginal rates of technical substitution

As the marginal products based on a linear production function are equal to the coefficients, we
can calculate the MRTS (2.7) as follows:
> mrtsCapLab <- - coef(prodLin)["qLab"] / coef(prodLin)["qCap"]
qLab
-6.615934
> mrtsLabCap <- - coef(prodLin)["qCap"] / coef(prodLin)["qLab"]
qCap
-0.1511502
> mrtsCapMat <- - coef(prodLin)["qMat"] / coef(prodLin)["qCap"]
qMat
-26.09666
> mrtsMatCap <- - coef(prodLin)["qCap"] / coef(prodLin)["qMat"]
qCap
-0.03831908
> mrtsLabMat <- - coef(prodLin)["qMat"] / coef(prodLin)["qLab"]
qMat
-3.944516
> mrtsMatLab <- - coef(prodLin)["qLab"] / coef(prodLin)["qMat"]

48

2 Primal Approach: Production Function

qLab
-0.2535165
Hence, if a firm wants to reduce the use of labor by one unit, he/she has to use 6.62 additional
units of capital in order to produce the same output as before. Alternatively, the firm can replace
the unit of labor by using 0.25 additional units of materials. If the firm increases the use of labor
by one unit, he/she can reduce capital by 6.62 units whilst still producing the same output as
before. Alternatively, the firm can reduce materials by 0.25 units.

2.3.9 Relative marginal rates of technical substitution

We can calculate the RMRTS (2.8) derived from the linear production function as follows:
> dat$rmrtsCapLab <- - dat$eLab / dat$eCap
> dat$rmrtsLabCap <- - dat$eCap / dat$eLab
> dat$rmrtsCapMat <- - dat$eMat / dat$eCap
> dat$rmrtsMatCap <- - dat$eCap / dat$eMat
> dat$rmrtsLabMat <- - dat$eMat / dat$eLab
> dat$rmrtsMatLab <- - dat$eLab / dat$eMat
We can visualize (the variation of) these RMRTSs with histograms:
> hist( dat$rmrtsCapLab, 20 )
> hist( dat$rmrtsLabCap )
> hist( dat$rmrtsCapMat )
> hist( dat$rmrtsMatCap )
> hist( dat$rmrtsLabMat )
> hist( dat$rmrtsMatLab )
The resulting graphs are shown in figure 2.11. According to the RMRTS based on the linear
production function, most firms need between 20% more capital or around 2% more materials to
compensate a 1% reduction of labor.

2.3.10 First-order conditions for profit maximisation

In this section, we will check to what extent the first-order conditions for profit maximization
(2.24) are fulfilled, i.e. to what extent the firms use the optimal input quantities. We do this by
comparing the marginal value products of the inputs with the corresponding input prices. We
can calculate the marginal value products by multiplying the marginal products by the output
price:
> dat$mvpCap <- dat$pOut * coef(prodLin)["qCap"]
> dat$mvpLab <- dat$pOut * coef(prodLin)["qLab"]
> dat$mvpMat <- dat$pOut * coef(prodLin)["qMat"]

49

40
30
10

0.20

0.00

60

50
Frequency

Figure 2.11: Linear production function:

(RMRTS)

10
0

10
0
rmrtsMatCap

40

50
30
20

Frequency
0.0

20

rmrtsCapMat

40

30
20
10
0
0.5 0.4 0.3 0.2 0.1

40

rmrtsLabCap

40

rmrtsCapLab

0.10

30

50

20

100

0
150

Frequency

20

Frequency

30

10

10

20

Frequency

40
30
20

Frequency

50

50

60

2 Primal Approach: Production Function

1.5

1.0

0.5

rmrtsLabMat

0.0

rmrtsMatLab

relative marginal rates of technical substitution

50

2 Primal Approach: Production Function

The command compPlot (package miscTools) can be used to compare the marginal value products
with the corresponding input prices:
> compPlot( dat$pCap, dat$mvpCap )
> compPlot( dat$pLab, dat$mvpLab )
> compPlot( dat$pMat, dat$mvpMat )
> compPlot( dat$pCap, dat$mvpCap, log = "xy" )
> compPlot( dat$pLab, dat$mvpLab, log = "xy" )

10

10 15 20 25 30 35

20

40

60

10.0

120

1.0

10

0.5

0.2

80

w Mat

2.0

5.0

2.0
1.0

0.5

MVP Cap

w Lab

MVP Lab

5.0

w Cap

0
0

100

50

20

MVP Mat

80

100

60

40

20

30
25
MVP Lab

4
3
2

20

MVP Cap

15

MVP Mat

140

35

> compPlot( dat$pMat, dat$mvpMat, log = "xy" )

0.2

0.5

1.0

2.0

5.0

0.5 1.0 2.0

w Cap

5.0

20.0

10

w Lab

20

50

100

w Mat

Figure 2.12: Marginal value products and corresponding input prices

The resulting graphs are shown in figure 2.12. The graphs on the left side indicate that the
marginal value products of capital are sometimes lower but more often higher than the capital
prices. The four other graphs indicate that the marginal value products of labor and materials
are always higher than the labor prices and the materials prices, respectively. This indicates that
some firms could increase their profit by using more capital and all firms could increase their
profit by using more labor and more materials. Given that most firms operate under increasing
returns to scale, it is not surprising that most firms would gain from increasing mostor even

51

2 Primal Approach: Production Function

allinput quantities. Therefore, the question arises why the firms in the sample did not do this.
There are many possible reasons for not increasing the input quantities until the predicted optimal input levels, e.g. legal restrictions, environmental regulations, market imperfections, credit
(liquidity) constraints, and/or risk aversion. Furthermore, market imperfections might cause that
the (observed) average prices are lower than the marginal costs of obtaining these inputs (e.g.
Henning and Henningsen, 2007), particularly for labor and capital.

2.3.11 First-order conditions for cost minimization

As the marginal rates of technical substitution are constant for linear production functions, we
compare the input price ratios with the negative inverse marginal rates of technical substitution
by creating a histogram for each input price ratio and drawing a vertical line at the corresponding
negative marginal rate of technical substitution:
> hist( dat$pCap / dat$pLab )
> lines( rep( - mrtsLabCap, 2), c( 0, 100 ), lwd = 3

> hist( dat$pCap / dat$pMat )

> lines( rep( - mrtsMatCap, 2), c( 0, 100 ), lwd = 3

> hist( dat$pLab / dat$pMat )

> lines( rep( - mrtsMatLab, 2), c( 0, 100 ), lwd = 3

> hist( dat$pLab / dat$pCap )

> lines( rep( - mrtsCapLab, 2), c( 0, 100 ), lwd = 3

> hist( dat$pMat / dat$pCap )

> lines( rep( - mrtsCapMat, 2), c( 0, 100 ), lwd = 3

> hist( dat$pMat / dat$pLab )

> lines( rep( - mrtsLabMat, 2), c( 0, 100 ), lwd = 3

The resulting graphs are shown in figure 2.13. The upper left graph shows that the ratio between
the capital price and the labor price is larger than the absolute value of the marginal rate of
technical substitution between labor and capital (0.151) for the most firms in the sample:
wcap
M Pcap
> M RT Slab,cap =
wlab
M Plab

(2.46)

Or taken the other way round, the lower left graph shows that the ratio between the labor
price and the capital price is smaller than the absolute value of the marginal rate of technical
substitution between capital and labor (6.616) for the most firms in the sample:
M Plab
wlab
< M RT Scap,lab =
wcap
M Pcap

(2.47)

Hence, the firm can get closer to the minimum of the costs by substituting labor for capital,
because this will decrease the marginal product of labor and increase the marginal product of

52

30
10
0

10
0

0
0

20

Frequency

30
20

Frequency

20
10

Frequency

30

40

40

40

50

2 Primal Approach: Production Function

0.0

0.2

0.6

0.8

0.1

w Cap / w Mat

0.2

0.3

0.4

w Lab / w Mat
40

4
w Lab / w Cap

30
10
0

10
0
0

20

Frequency

40
30
20

Frequency

40
20
0

Frequency

60

50

80

60

w Cap / w Lab

0.4

10

20

30

40

50

60

w Mat / w Cap

Figure 2.13: First-order conditions for costs minimization

53

10

15

w Mat / w Lab

20

2 Primal Approach: Production Function

capital so that the absolute value of the MRTS between labor and capital increases, the absolute
value of the MRTS between capital and labor decreases, and both of the MRTS get closer to the
corresponding input price ratios. Similarly, the graphs in the middle column indicate that almost
all firms should substitute materials for capital and the graphs on the right indicate that most of
the firms should substitute labor for materials. Hence, the firms could reduce production costs
particularly by using less capital and more labor.

2.3.12 Derived Input Demand Functions and Output Supply Functions

Given a linear production function (2.42), the input quantities chosen by a profit maximizing
producer are either zero, indeterminate, or infinity:

xi (p, w) =

if M V Pi < wi

indeterminate

if M V Pi = wi

(2.48)

if M V Pi > wi

If all input quantities are zero, the output quantity is equal to the intercept, which is zero in case
of weak essentiality. Otherwise, the output quantity is indeterminate or infinity:

y(p, w) =

if M V Pi < wi i

if M V Pi > wi i

indeterminate

(2.49)

otherwise

A cost minimizing producer will use only a single input, i.e. the input with the lowest cost
per unit of produced output (wi /M Pi ). If the lowest cost per unit of produced output can be
obtained by two or more inputs, these input quantities are indeterminate.

xi (w, y) =

if

y0

indeterminate

if

i
wi
i
wi

<
>

j
wj
j
wj

j
j 6= i

(2.50)

otherwise

Given that the unconditional and conditional input demand functions and the output supply
functions based on the linear production function are non-continuous and often return either zero
or infinite values, it does not make much sense to use this functional form to predict the effects
of price changes when the true technology implies that firms always use non-zero finite input
quantities.

54

2 Primal Approach: Production Function

2.4 Cobb-Douglas production function

2.4.1 Specification
A Cobb-Douglas production function with N inputs is defined as:
y=A

N
Y
xi i .

(2.51)

i=1

This function can be linearized by taking the (natural) logarithm on both sides:
ln y = 0 +

N
X

i ln xi ,

(2.52)

i=1

where 0 is equal to ln A.

2.4.2 Estimation
We can estimate this Cobb-Douglas production function for our data set using the command lm:
> prodCD <- lm( log( qOut ) ~ log( qCap ) + log( qLab ) + log( qMat ),
+

data = dat )

> summary( prodCD )

Call:
lm(formula = log(qOut) ~ log(qCap) + log(qLab) + log(qMat), data = dat)
Residuals:
Min

1Q

Median

3Q

Max

-1.67239 -0.28024

0.00667

0.47834

1.30115

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) -2.06377

1.31259

-1.572

0.1182

log(qCap)

0.16303

0.08721

1.869

log(qLab)

0.67622

0.15430

4.383 2.33e-05 ***

log(qMat)

0.62720

0.12587

4.983 1.87e-06 ***

0.0637 .

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Residual standard error: 0.656 on 136 degrees of freedom

Multiple R-squared:

0.5943,

F-statistic: 66.41 on 3 and 136 DF,

Adjusted R-squared:
p-value: < 2.2e-16

55

0.5854

2 Primal Approach: Production Function

2.4.3 Properties
The monotonicity condition is (globally) fulfilled, as the estimated coefficients of all three (logarithmic) input quantities are positive and the output quantity as well as all input quantities are
non-negative (see equation 2.54). However, the coefficient of the (logarithmic) capital quantity is
only statistically significantly different from zero at the 10% level. Therefore, we cannot be sure
that the capital quantity has a positive effect on the output quantity.
The quasi-concavity of our estimated Cobb-Douglas production function is checked in section 2.4.12.
The production technology described by a Cobb-Douglas production function always shows
weak and strict essentiality, because the output quantity becomes zero, as soon as a single input
quantity becomes zero (see equation 2.51).
The input requirement sets derived from Cobb-Douglas production functions are always closed
and non-empty for y > 0 if strict monotonicity is fulfilled for at least one input ( i {1, . . . , N } :
i > 0), as the input quantities must be non-negative (xi 0 i).
The Cobb-Douglas production function always returns finite, real, and single values if the input
quantities are non-negative and finite. The predicted output quantity is non-negative as long as
A and the input quantities are non-negative, where A = exp(0 ) is positive even if 0 is negative.
All Cobb-Douglas production functions are continuous and twice-continuously differentiable.

2.4.4 Predicted output quantities

We can calculate the predicted (fitted) output quantities manually by taking the Cobb-Douglas
function (2.51), the observed input quantities, and the estimated parameters, but it is easier
to use the fitted method to obtain the predicted values of the dependent variable from an
estimated model. As we estimated the Cobb-Douglas function in logarithms, we have to use the
exponential function to obtain the predicted values in levels (non-logarithms):
> dat$qOutCD <- exp( fitted( prodCD ) )
> all.equal( dat$qOutCD,
+

with( dat, exp( coef( prodCD )[ "(Intercept)" ] ) *

qCap^coef( prodCD )[ "log(qCap)" ] *

qLab^coef( prodCD )[ "log(qLab)" ] *

qMat^coef( prodCD )[ "log(qMat)" ] ) )

[1] TRUE
We can evaluate the fit of the Cobb-Douglas production function by comparing the observed
with the fitted output quantities:
> compPlot( dat$qOut, dat$qOutCD )
> compPlot( dat$qOut, dat$qOutCD, log = "xy" )

56

2 Primal Approach: Production Function

1e+07

0.0e+00

1e+05

5e+05 2e+06

fitted

1.0e+07
0.0e+00

fitted

2.0e+07

1.0e+07

2.0e+07

1e+05

observed

5e+05

5e+06

observed

Figure 2.14: Cobb-Douglas production function: fit of the model

The resulting graphs are shown in figure 2.14. While the graph in the left panel uses a linear
scale for the axes, the graph in the right panel uses a logarithmic scale for both axes. Hence, the
deviations from the 45-line illustrate the absolute deviations in the left panel and the relative
deviations in the right panel. The fit of the model looks okay in the scatter plot on the left-hand
side, but if we use a logarithmic scale on both axes (as in the graph on the right-hand side), we
can see that the output quantity is generally over-estimated if the the observed output quantity
is small.

2.4.5 Output elasticities

In the Cobb-Douglas function, the output elasticities of the inputs are equal to the corresponding
coefficients.
i =

y xi
ln y
=
= i
xi y
ln xi

(2.53)

Hence, if a firm increases capital input by one percent, the output will increase by 0.16 percent;
if a firm increases labor input by one percent, the output will increase by 0.68 percent; and if
a firm increases materials input by one percent, the output will increase by 0.63 percent. The
output elasticity of capital is somewhat larger and the output elasticity of labor is considerably
smaller when estimated by a Cobb-Douglas production function than when estimated by a linear
production function. Indeed, the output elasticities of all three inputs are in the reasonable range,
i.e. between zero one one, now.

2.4.6 Marginal products

In the Cobb-Douglas function, the marginal products of the inputs can be calculated by following
formula:

y
ln y y
y
=
= i = i APi
xi
ln xi xi
xi

57

(2.54)

2 Primal Approach: Production Function

As the marginal products depend on the input and output quantities and these quantities generally differ between firms, the marginal products based on Cobb-Douglas also differ between firms.
Hence, we can calculate them for each firm in the sample:
> dat$mpCapCD <- coef(prodCD)["log(qCap)"] * dat$apCap
> dat$mpLabCD <- coef(prodCD)["log(qLab)"] * dat$apLab
> dat$mpMatCD <- coef(prodCD)["log(qMat)"] * dat$apMat
We can visualize (the variation of) these marginal products with histograms:
> hist( dat$mpCapCD )
> hist( dat$mpLabCD )

30
0

10

15

20

25

30
0

10

20

Frequency

20
15
10

Frequency

30
20
10

Frequency

40

25

50

> hist( dat$mpMatCD )

MP Cap

10

15

50

MP Lab

100

150

200

MP Mat

Figure 2.15: Cobb-Douglas production function: marginal products

The resulting graphs are shown in figure 2.15. If the firms increase capital input by one unit, the
output of most firms will increase by between 0 and 8 units; if the firms increase labor input by
one unit, the output of most firms will increase by between 2 and 12 units; and if the firms increase
materials input by one unit, the output of most firms will increase by between 20 and 80 units.
Not surprisingly, a comparison of these marginal effects with the marginal effects from the linear
production function confirms the results from the comparison based on the output elasticities:
the marginal products of capital are generally larger than the marginal product estimated by
the linear production function and the marginal products of labor are generally smaller than the
marginal product estimated by the linear production function, while the marginal products of
fuel are (on average) rather similar to the marginal product estimated by the linear production
function.

2.4.7 Elasticity of Scale

As the elasticity of scale is the sum of all output elasticities (see equation 2.45), we can calculate
it simply by summing up all coefficients except for the intercept:

58

2 Primal Approach: Production Function

> sum( coef( prodCD )[ -1 ] )
[1] 1.466442
Hence, if the firm increases all input quantities by one percent, output will increase by 1.47
percent. This means that the technology has strong increasing returns to scale. However, in
contrast to the results of the linear production function, the elasticity of scale based on the CobbDouglas production function is (globally) constant. Hence, it does not decrease (or increase), e.g.,
with the size of the firm. This means that the optimal firm size would be infinity.
We can use the delta method (see section 1.4.3) to calculate the variance and the standard error
of the elasticity of scale. Given that the first derivatives of the elasticity of scale with respect to
the estimated coefficients are /0 = 0 and /Cap = /Lab = /M at = 1, we can
do this by following commands:
> ESCD <- sum( coef(prodCD)[-1] )
[1] 1.466442
> dESCD <- c( 0, 1, 1, 1 )
[1] 0 1 1 1
> varESCD <- t(dESCD) %*% vcov(prodCD) %*% dESCD
[,1]
[1,] 0.0118237
> seESCD <- sqrt( varESCD )
[,1]
[1,] 0.1087369
Now, we can apply a t test to test whether the elasticity of scale significantly differs from one.
The following commands calculate the t value and the critical value for a two-sided t test based
on a 5% significance level:
> tESCD <- (ESCD - 1) / seESCD
[,1]
[1,] 4.289645
> cvESCD <- qt( 0.975, 136 )
[1] 1.977561

59

2 Primal Approach: Production Function

Given that the t value is larger than the critical value, we can reject the null hypothesis of
constant returns to scale and conclude that the technology has significantly increasing returns to
scale. The P value for this two-sided t test is:
> pESCD <- 2 * ( 1 - pt( tESCD, 136 ) )
[,1]
[1,] 3.372264e-05
Given that the P value is close to zero, we can be very sure that the technology has increasing
returns to scale. The 95% confidence interval for the elasticity of scale is:
> c( ESCD - cvESCD * seESCD, ESCD + cvESCD * seESCD )
[1] 1.251409 1.681476

2.4.8 Marginal Rates of Technical Substitution

The MRTS based on the Cobb-Douglas production function differ between firms. They can be
calculated as follows:
> dat$mrtsCapLabCD <- - dat$mpLabCD / dat$mpCapCD
> dat$mrtsLabCapCD <- - dat$mpCapCD / dat$mpLabCD
> dat$mrtsCapMatCD <- - dat$mpMatCD / dat$mpCapCD
> dat$mrtsMatCapCD <- - dat$mpCapCD / dat$mpMatCD
> dat$mrtsLabMatCD <- - dat$mpMatCD / dat$mpLabCD
> dat$mrtsMatLabCD <- - dat$mpLabCD / dat$mpMatCD
We can visualize (the variation of) these MRTSs with histograms:
> hist( dat$mrtsCapLabCD )
> hist( dat$mrtsLabCapCD )
> hist( dat$mrtsCapMatCD )
> hist( dat$mrtsMatCapCD )
> hist( dat$mrtsLabMatCD )
> hist( dat$mrtsMatLabCD )
The resulting graphs are shown in figure 2.16. According to the MRTS based on the CobbDouglas production function, most firms only need between 0.5 and 2 additional units of capital
or between 0.05 and 0.15 additional units of materials to replace one unit of labor.

60

30
10
6

50

mrtsMatCapCD

0.0

10

50
Frequency

10
0

10
0
0.2

20

40

60
40
30
20

Frequency
0.4

30

mrtsCapMatCD

50

50
40
30
20
10
0
0.6

40

mrtsLabCapCD

60

mrtsCapLabCD

30

20

10
0

5
0
6

Frequency

20

Frequency

40
30
20

Frequency

20
15
10

Frequency

25

50

30

60

35

2 Primal Approach: Production Function

35

25

15

mrtsLabMatCD

0.4

0.3

0.2

0.1

0.0

mrtsMatLabCD

Figure 2.16: Cobb-Douglas production function: marginal rates of technical substitution (MRTS)

2.4.9 Relative Marginal Rates of Technical Substitution

As we do not know the units of measurements of the input quantities, the interpretation of
the MRTSs is practically not very useful. To overcome this problem, we calculate the relative
marginal rates of technical substitution (RMRTS) by equation (2.8). As the output elasticities
based on a Cobb-Douglas production function are equal to the coefficients, we can calculate the
RMRTS as follows:
> rmrtsCapLabCD <- - coef(prodCD)["log(qLab)"] / coef(prodCD)["log(qCap)"]
log(qLab)
-4.147897
> rmrtsLabCapCD <- - coef(prodCD)["log(qCap)"] / coef(prodCD)["log(qLab)"]
log(qCap)
-0.241086
> rmrtsCapMatCD <- - coef(prodCD)["log(qMat)"] / coef(prodCD)["log(qCap)"]
log(qMat)
-3.847203

61

2 Primal Approach: Production Function

> rmrtsMatCapCD <- - coef(prodCD)["log(qCap)"] / coef(prodCD)["log(qMat)"]
log(qCap)
-0.2599291
> rmrtsLabMatCD <- - coef(prodCD)["log(qMat)"] / coef(prodCD)["log(qLab)"]
log(qMat)
-0.9275069
> rmrtsMatLabCD <- - coef(prodCD)["log(qLab)"] / coef(prodCD)["log(qMat)"]
log(qLab)
-1.078159
Hence, if a firm wants to reduce the use of labor by one percent, it has to use 4.15 percent more
capital in order to produce the same output as before. Alternatively, the firm can replace one
percent of labor by using 1.08 percent more materials. If the firm increases the use of labor by one
percent, it can reduce capital by 4.15 percent whilst still producing the same output as before.
Alternatively, the firm can reduce materials by 1.08 percent.

2.4.10 First and second partial derivatives

For the Cobb-Douglas production function with three inputs (2.51), the first derivatives (marginal
products) are
y
y
= 1 A x1 1 1 x2 2 x3 3 = 1
x1
x1
y
y
f2 =
= 2 A x1 1 x2 2 1 x3 3 = 2
x2
x2
y
y
f3 =
= 3 A x1 1 x2 2 x3 3 1 = 3
x3
x3

f1 =

(2.55)
(2.56)
(2.57)

and the second derivatives are

f11 =
f22 =
f33 =
f12 =
f13 =

f1
x1
f2
x2
f3
x3
f1
x2
f1
x3

f1
x1
f2
= 2
x2
f3
= 3
x3
f2
= 1
x1
f3
= 1
x1
= 1

y
=
x21
y
2 2 =
x2
y
3 2 =
x3
f1 f2
=
y
f1 f3
=
y
1

62

f12
f1

y
x1
2
f2
f2

y
x2
2
f3
f3

y
x3

(2.58)
(2.59)
(2.60)
(2.61)
(2.62)

2 Primal Approach: Production Function

f23 =

f3
f2 f3
f2
= 2
=
.
x3
x2
y

(2.63)

Generally, for an N -input Cobb-Douglas function, the first and second derivatives are
y
xi
fi fj
fi
fij =
ij ,
y
xi
fi = i

(2.64)
(2.65)

where ij denotes Kroneckers delta with

1 if i = j
ij =
0 if i =
6 j

(2.66)

In the calculations of the partial derivatives (fi ), we have simplified the formulas by replacing
the right-hand side of the Cobb-Douglas function (2.51) by the output quantities. When we
calculated the marginal products (partial derivatives) of the Cobb-Douglas function in in section 2.4.6, we have used the observed output quantities for y. However, as the fit (R2 value)
of our model is not 100 %, the observed output quantities are generally not equal to the output
quantities predicted by our model, i.e. the right-hand side of the Cobb-Douglas function (2.51)
using the estimated parameters. The better the fit of our model, the smaller is the difference
between the observed and the predicted output quantities. If we believe in our estimated model,
it would be more consistent with microeconomic theory, if we use the predicted output quantities and disregard the stochastic error term (difference between observed and predicted output
quantities) that is caused, e.g., by measurement errors, (good or bad) luck, or unusual(ly) (good
or bad) weather conditions.
We can calculate the first derivatives (marginal products) with the predicted output quantities
(see section 2.4.4):
> dat$fCap <- coef(prodCD)["log(qCap)"] * dat$qOutCD / dat$qCap
> dat$fLab <- coef(prodCD)["log(qLab)"] * dat$qOutCD / dat$qLab
> dat$fMat <- coef(prodCD)["log(qMat)"] * dat$qOutCD / dat$qMat
Based on these first derivatives, we can also calculate the second derivatives:
> dat$fCapCap <- with( dat, fCap^2 / qOutCD - fCap / qCap )
> dat$fLabLab <- with( dat, fLab^2 / qOutCD - fLab / qLab )
> dat$fMatMat <- with( dat, fMat^2 / qOutCD - fMat / qMat )
> dat$fCapLab <- with( dat, fCap * fLab / qOutCD )
> dat$fCapMat <- with( dat, fCap * fMat / qOutCD )
> dat$fLabMat <- with( dat, fLab * fMat / qOutCD )

63

2 Primal Approach: Production Function

2.4.11 Elasticities of substitution

2.4.11.1 Direct Elasticities of Substitution
In order to calculate the elasticities of substitution, we need to construct the bordered Hessian
matrix. As the first and second derivatives of the Cobb-Douglas function differ between observations, also the bordered Hessian matrix differs between observations. As a starting point, we
construct the bordered Hessian Matrix just for the first observation:
> bhm <- matrix( 0, nrow = 4, ncol = 4 )
> bhm[ 1, 2 ] <- bhm[ 2, 1 ] <- dat$fCap[ 1 ]
> bhm[ 1, 3 ] <- bhm[ 3, 1 ] <- dat$fLab[ 1 ]
> bhm[ 1, 4 ] <- bhm[ 4, 1 ] <- dat$fMat[ 1 ]
> bhm[ 2, 2 ] <- dat$fCapCap[ 1 ]
> bhm[ 3, 3 ] <- dat$fLabLab[ 1 ]
> bhm[ 4, 4 ] <- dat$fMatMat[ 1 ]
> bhm[ 2, 3 ] <- bhm[ 3, 2 ] <- dat$fCapLab[ 1 ]
> bhm[ 2, 4 ] <- bhm[ 4, 2 ] <- dat$fCapMat[ 1 ]
> bhm[ 3, 4 ] <- bhm[ 4, 3 ] <- dat$fLabMat[ 1 ]
> print(bhm)
[,1]

[,2]

[1,]

0.000000

6.229014e+00

[2,]

6.229014 -6.202845e-05

[3,]

6.031225

[4,] 59.090913

[,3]

[,4]

6.031225e+00 59.0909133861
1.169835e-05

0.0001146146

1.169835e-05 -5.423455e-06

0.0001109752

1.146146e-04

1.109752e-04 -0.0006462733

Based on this bordered Hessian matrix, we can calculate the co-factors Fij :
> FCapLab <- - det( bhm[ -2, -3 ] )
[1] -0.06512713
> FCapMat <- det( bhm[ -2, -4 ] )
[1] -0.006165438
> FLabMat <- - det( bhm[ -3, -4 ] )
[1] -0.02641227
So that we can calculate the direct elasticities of substitution (of the first observation):
> esdCapLab <- with( dat[1,], ( qCap * fCap + qLab * fLab ) /
+

( qCap * qLab ) * FCapLab / det( bhm ) )

64

2 Primal Approach: Production Function

[1] 0.5723001
> esdCapMat <- with( dat[ 1, ], ( qCap * fCap + qMat * fMat ) /
+

( qCap * qMat ) * FCapMat / det( bhm ) )

[1] 0.5388715
> esdLabMat <- with( dat[ 1, ], ( qLab * fLab + qMat * fMat ) /
+

( qLab * qMat ) * FLabMat / det( bhm ) )

[1] 0.8888284
As all elasticities of substitution are positive, we can conclude that all pairs of inputs are substitutes for each other and no pair of inputs is complementary. If the firm substitutes capital for labor
so that the ratio between the capital and labor quantity (xcap /xlab ) increases by 0.57 percent,
the (absolute value of the) MRTS between capital and labor (|dxcap /dxlab | = flab /fcap ) increases
by one percent. Or, the other way round, if the firm substitutes capital for labor so that the
absolute value of the MRTS between capital and labor (|dxcap /dxlab | = flab /fcap ) increases by
one percent, e.g. because the price ratio between labor and capital (wlab /wcap ) increases by one
percent, the ratio between the capital and labor quantity (xcap /xlab ) will increase by 0.57 percent.
We can calculate the elasticities of substitution for all firms by automatically repeating the
above commands for each observation using a for loop:2
> dat$esdCapLab <- NA
> dat$esdCapMat <- NA
> dat$esdLabMat <- NA
> for( obs in 1:nrow( dat ) ) {
+

bhmLoop <- matrix( 0, nrow = 4, ncol = 4 )

bhmLoop[ 1, 2 ] <- bhmLoop[ 2, 1 ] <- dat$fCap[ obs ]

bhmLoop[ 1, 3 ] <- bhmLoop[ 3, 1 ] <- dat$fLab[ obs ]

bhmLoop[ 1, 4 ] <- bhmLoop[ 4, 1 ] <- dat$fMat[ obs ]

bhmLoop[ 2, 2 ] <- dat$fCapCap[ obs ]

bhmLoop[ 3, 3 ] <- dat$fLabLab[ obs ]

bhmLoop[ 4, 4 ] <- dat$fMatMat[ obs ]

bhmLoop[ 2, 3 ] <- bhmLoop[ 3, 2 ] <- dat$fCapLab[ obs ]

bhmLoop[ 2, 4 ] <- bhmLoop[ 4, 2 ] <- dat$fCapMat[ obs ]

bhmLoop[ 3, 4 ] <- bhmLoop[ 4, 3 ] <- dat$fLabMat[ obs ]

FCapLabLoop <- - det( bhmLoop[ -2, -3 ] )

FCapMatLoop <- det( bhmLoop[ -2, -4 ] )

As I want to use the bordered Hessian matrix and some of its co-factors after the loop, I do not want to overwrite
the values in bhm, FCapLab, FCapMat, and FLabMat in the loop. Therefore, I use not the same variable names
for the bordered Hessian matrix and the co-factors in the loop.

65

2 Primal Approach: Production Function

+

FLabMatLoop <- - det( bhmLoop[ -3, -4 ] )

dat$esdCapLab[ obs ] <- with( dat[obs,],

( qCap * fCap + qLab * fLab ) /

( qCap * qLab ) * FCapLabLoop / det( bhmLoop ) )

dat$esdCapMat[ obs ] <- with( dat[ obs, ],

( qCap * fCap + qMat * fMat ) /

( qCap * qMat ) * FCapMatLoop / det( bhmLoop ) )

dat$esdLabMat[ obs ] <- with( dat[ obs, ],

( qLab * fLab + qMat * fMat ) /

( qLab * qMat ) * FLabMatLoop / det( bhmLoop ) )

+ }
> range( dat$esdCapLab )
[1] 0.5723001 0.5723001
> range( dat$esdCapMat )
[1] 0.5388715 0.5388715
> range( dat$esdLabMat )
[1] 0.8888284 0.8888284
The direct elasticities of substitution based on the Cobb-Douglas production function are the
same for all firms.
2.4.11.2 Allen Elasticities of Substitution
The calculation of the Allen elasticities of substitution is similar to the calculation of the direct
elasticities of substitution:
> numerator <- with( dat[1,], qCap * fCap + qLab * fLab + qMat * fMat )
> esaCapLab <- numerator /
+

( dat$qCap[ 1 ] * dat$qLab[ 1 ] ) *

FCapLab / det( bhm )

[1] 1
> esaCapMat <- numerator /
+

( dat$qCap[ 1 ] * dat$qMat[ 1 ] ) *

FCapMat / det( bhm )

[1] 1

66

2 Primal Approach: Production Function

> esaLabMat <- numerator /
+

( dat$qLab[ 1 ] * dat$qMat[ 1 ] ) *

FLabMat / det( bhm )

[1] 1
All elasticities of substitution are exactly one. This is no surprise and confirms that our calculations have been done correctly, because the Cobb-Douglas production function always has Allen
elasticities of substitution equal to one, irrespective of the input and output quantities and the
estimated parameters. Hence, the Cobb-Douglas function cannot be used to analyze the substitutability of the inputs, because it will always return Allen elasticities of substitution equal to
one, no matter if the true elasticities are close to zero or close to infinity.
Although it seemed that we got free estimates of the direct elasticities of substitution from
the Cobb-Douglas production function in section 2.4.11.1, they are indeed forced to be (fi xi +
fj xj )/(

k fk

xk ) = (i y + j y)/(

k k

y) = (i + j )/, where  is the elasticity of scale

(see equation 2.14). Hence, the Cobb-Douglas production function cannot be used to analyze
substitutability between inputs.
2.4.11.3 Morishima Elasticities of Substitution
In order to calculate the Morishima elasticities of substitution, we need to calculate the co-factors
of the diagonal elements of the bordered Hessian matrix:
> FCapCap <- det( bhm[ -2, -2 ] )
> FLabLab <- det( bhm[ -3, -3 ] )
> FMatMat <- det( bhm[ -4, -4 ] )
> esmCapLab <- with( dat[1,], ( fLab / qCap ) * FCapLab / det( bhm ) +

( fLab / qLab ) * FLabLab / det( bhm ) )

[1] 1
> esmLabCap <- with( dat[1,], ( fCap / qLab ) * FCapLab / det( bhm ) +

( fCap / qCap ) * FCapCap / det( bhm ) )

[1] 1
> esmCapMat <- with( dat[1,], ( fMat / qCap ) * FCapMat / det( bhm ) +

( fMat / qMat ) * FMatMat / det( bhm ) )

[1] 1
> esmMatCap <- with( dat[1,], ( fCap / qMat ) * FCapMat / det( bhm ) +

( fCap / qCap ) * FCapCap / det( bhm ) )

67

2 Primal Approach: Production Function

[1] 1
> esmLabMat <- with( dat[1,], ( fMat / qLab ) * FLabMat / det( bhm ) +

( fMat / qMat ) * FMatMat / det( bhm ) )

[1] 1
> esmMatLab <- with( dat[1,], ( fLab / qMat ) * FLabMat / det( bhm ) +

( fLab / qLab[ 1 ] ) * FLabLab / det( bhm ) )

[1] 1
As with the Allen elasticities of substitution, all Morishima elasticities of substitution based on
Cobb-Douglas functions are exactly one.
From the condition 2.15, we can show that all Morishima elasticities of substitution are always
M = 1 i 6= j), if all Allen elasticities of substitution are one ( = 1 i 6= j):
one (ij
ij
M
ij
= Kj ij Kj jj = Kj +

X
k6=j

Kk kj =

Kk = 1

(2.67)

2.4.12 Quasiconcavity
We start by checking whether our estimated Cobb-Douglas production function is quasiconcave
at the first observation:
> bhm
[,1]

[,2]

[1,]

0.000000

6.229014e+00

[2,]

6.229014 -6.202845e-05

[3,]

6.031225

[4,] 59.090913

[,3]

[,4]

6.031225e+00 59.0909133861
1.169835e-05

0.0001146146

1.169835e-05 -5.423455e-06

0.0001109752

1.146146e-04

1.109752e-04 -0.0006462733

> det( bhm[ 1:2, 1:2 ] )

[1] -38.80062
> det( bhm[ 1:3, 1:3 ] )
[1] 0.003345742
> det( bhm )
[1] -1.013458e-05

68

2 Primal Approach: Production Function

The first principal minor of the bordered Hessian matrix is negative, the second principal minor is
positive, and the third principal minor is negative. This means that our estimated Cobb-Douglas
production function is quasiconcave at the first observation.
Now we check quasiconcavity at all observations:
> dat$quasiConc <- NA
> for( obs in 1:nrow( dat ) ) {
+

bhmLoop <- matrix( 0, nrow = 4, ncol = 4 )

bhmLoop[ 1, 2 ] <- bhmLoop[ 2, 1 ] <- dat$fCap[ obs ]

bhmLoop[ 1, 3 ] <- bhmLoop[ 3, 1 ] <- dat$fLab[ obs ]

bhmLoop[ 1, 4 ] <- bhmLoop[ 4, 1 ] <- dat$fMat[ obs ]

bhmLoop[ 2, 2 ] <- dat$fCapCap[ obs ]

bhmLoop[ 3, 3 ] <- dat$fLabLab[ obs ]

bhmLoop[ 4, 4 ] <- dat$fMatMat[ obs ]

bhmLoop[ 2, 3 ] <- bhmLoop[ 3, 2 ] <- dat$fCapLab[ obs ]

bhmLoop[ 2, 4 ] <- bhmLoop[ 4, 2 ] <- dat$fCapMat[ obs ]

bhmLoop[ 3, 4 ] <- bhmLoop[ 4, 3 ] <- dat$fLabMat[ obs ]

dat$quasiConc[ obs ] <- det( bhmLoop[ 1:2, 1:2 ] ) < 0 &

det( bhmLoop[ 1:3, 1:3 ] ) > 0 & det( bhmLoop ) < 0

+ }
> sum( dat$quasiConc )
[1] 140
Our estimated Cobb-Douglas production function is quasiconcave at all of the 140 observations.
In fact, all Cobb-Douglas production functions are quasiconcave in inputs if A 0, 1 0,
. . . , N 0, while Cobb-Douglas production functions are concave in inputs if A 0, 1 0,
PN

. . . , N 0, and the technology has decreasing or constant returns to scale (

i=1 i

1).3

2.4.13 First-order conditions for profit maximisation

In this section, we will check to what extent the first-order conditions (2.24) for profit maximization are fulfilled, i.e. to what extent the firms use the optimal input quantities. We do this by
comparing the marginal value products of the inputs with the corresponding input prices. We
can calculate the marginal value products by multiplying the marginal products by the output
price:
> dat$mvpCapCd <- dat$pOut * dat$fCap
> dat$mvpLabCd <- dat$pOut * dat$fLab
> dat$mvpMatCd <- dat$pOut * dat$fMat
3

See, e.g., http://econren.weebly.com/uploads/9/0/1/5/9015734/lecture16.pdf or http://web.mit.edu/14.

102/www/ps/ps1sol.pdf.

69

2 Primal Approach: Production Function

The command compPlot (package miscTools) can be used to compare the marginal value products
with the corresponding input prices:
> compPlot( dat$pCap, dat$mvpCapCd )
> compPlot( dat$pLab, dat$mvpLabCd )
> compPlot( dat$pMat, dat$mvpMatCd )
> compPlot( dat$pCap, dat$mvpCapCd, log = "xy" )
> compPlot( dat$pLab, dat$mvpLabCd, log = "xy" )

250

30

> compPlot( dat$pMat, dat$mvpMatCd, log = "xy" )

30

200
50
10

20

25

30

50

100

100 200

200

250

0.5

0.2

1.0

2.0

5.0

10.0

150

w Mat

50

MVP Lab

20.0
5.0

15
w Lab

0.5

2.0

150

MVP Mat
5

w Cap

MVP Cap

0
0

20

20

10

10

100

25
10

MVP Mat

20

MVP Lab

20
0

15

10

MVP Cap

30

0.2

0.5

2.0

5.0

20.0

0.5

1.0

2.0

w Cap

5.0

20.0

10

w Lab

20

50

200

w Mat

Figure 2.17: Marginal value products and corresponding input prices

The resulting graphs are shown in figure 2.17. They indicate that the marginal value products
are always nearly equal to or higher than the corresponding input prices. This indicates that
(almost) all firms could increase their profit by using more of all inputs. Given that the estimated
Cobb-Douglas technology exhibits increasing returns to scale, it is not surprising that (almost)
all firms would gain from increasing all input quantities. Therefore, the question arises why the
firms in the sample did not do this. This questions has already been addressed in section 2.3.10.

70

2 Primal Approach: Production Function

2.4.14 First-order conditions for cost minimization

As the marginal rates of technical substitution differ between observations for the Cobb-Douglas
functional form, we use scatter plots for visualizing the comparison of the input price ratios with
the negative inverse marginal rates of technical substitution:
> compPlot( dat$pCap / dat$pLab, - dat$mrtsLabCapCD )
> compPlot( dat$pCap / dat$pMat, - dat$mrtsMatCapCD )
> compPlot( dat$pLab / dat$pMat, - dat$mrtsMatLabCD )
> compPlot( dat$pCap / dat$pLab, - dat$mrtsLabCapCD, log = "xy" )
> compPlot( dat$pCap / dat$pMat, - dat$mrtsMatCapCD, log = "xy" )
> compPlot( dat$pLab / dat$pMat, - dat$mrtsMatLabCD, log = "xy" )

0.4

0.8

0.0

0.2

0.4

0.3
0.2

0.6

0.8

0.1

w Cap / w Mat

0.2

0.5

1.0

2.0

0.3

0.4

5.0

0.50

0.20

0.10

0.05

0.05 0.10 0.20

0.2

w Lab / w Mat

0.02

MRTS Mat Cap

5.0
2.0

1.0
0.5

0.2

w Cap / w Lab

MRTS Lab Cap

MRTS Mat Lab

0.6

0.1

0.2

MRTS Mat Lab

0.0

0.4

MRTS Mat Cap

MRTS Lab Cap

0.02

w Cap / w Lab

0.05

0.20

0.50

0.05

w Cap / w Mat

0.10

0.20

w Lab / w Mat

Figure 2.18: First-order conditions for cost minimization

The resulting graphs are shown in figure 2.18.
Furthermore, we use histograms to visualize the (absolute and relative) differences between the
input price ratios and the corresponding negative inverse marginal rates of technical substitution:
> hist( - dat$mrtsLabCapCD - dat$pCap / dat$pLab )
> hist( - dat$mrtsMatCapCD - dat$pCap / dat$pMat )

71

2 Primal Approach: Production Function

> hist( - dat$mrtsMatLabCD - dat$pLab / dat$pMat )
> hist( log( - dat$mrtsLabCapCD / ( dat$pCap / dat$pLab ) ) )
> hist( log( - dat$mrtsMatCapCD / ( dat$pCap / dat$pMat ) ) )

30
10

20

Frequency

40
30
20

Frequency

60
40
4

10

20

Frequency

80

50

40

> hist( log( - dat$mrtsMatLabCD / ( dat$pLab / dat$pMat ) ) )

0.6

0.2

0.0

0.2

0.3 0.2 0.1

MrtsMatCap wCap / wMat

0.0

0.1

0.2

MrtsMatLab wLab / wMat

1.0

0.0

1.0

log(MrtsLabCap / (wCap / wLab))

15

Frequency

5
0

0
2.0

10

40
30
10

20

Frequency

40
30
20
0

10

Frequency

20

50

MrtsLabCap wCap / wLab

0.4

2.5

1.5

0.5

0.5

log(MrtsMatCap / (wCap / wMat))

1.5

0.5

0.0

0.5

1.0

log(MrtsMatLab / (wLab / wMat))

Figure 2.19: First-order conditions for costs minimization

The resulting graphs are shown in figure 2.19. The left graphs in figures 2.18 and 2.19 show
that the ratio between the capital price and the labor price is larger than the absolute value of
the marginal rate of technical substitution between labor and capital for the most firms in the
sample:
wcap
M Pcap
> M RT Slab,cap =
wlab
M Plab

(2.68)

Hence, most firms can get closer to the minimum of their production costs by substituting labor
for capital, because this will decrease the marginal product of labor and increase the marginal
product of capital so that the absolute value of the MRTS between labor and capital increases
and gets closer to the corresponding input price ratio. Similarly, the graphs in the middle column
indicate that most firms should substitute materials for capital and the graphs on the right
indicate that the majority of the firms should substitute materials for labor. Hence, the majority

72

2 Primal Approach: Production Function

of the firms could reduce production costs particularly by using less capital and more materials4
but there might be (legal) regulations that restrict the use of materials (e.g. fertilizers, pesticides).

2.4.15 Derived Input Demand Functions and Output Supply Functions

Given a Cobb-Douglas production function (2.51), the input quantities chosen by a profit maximizing producer are

xi (p, w) =

Y
i

P A

wi

j
wj

1
!j 1

if < 1

(2.69)

if = 1
if > 1

and the output quantity is

y(p, w) =

with =

A P

j
wj

1
!j 1

if < 1

if = 1

(2.70)

if > 1

j . Hence, if the Cobb-Douglas production function exhibits increasing returns

to scale ( = > 1), the optimal input and output quantities are infinity. As our estimated
Cobb-Douglas production function has increasing returns to scale, the optimal input quantities
are infinity. Therefore, we cannot evaluate the effect of prices on the optimal input quantities.
A cost minimizing producer would choose the following input quantities:

y Y i wj
xi (w, y) =
A j6=i j wi

!j 1

(2.71)

For our three-input Cobb-Douglas production function, we get following conditional input demand
functions

y
xcap (w, y) =
A

cap
wcap

y
A

wcap
cap

xlab (w, y) =
4

!lab +mat 

!cap 

lab
wlab

wlab
lab

lab 

wmat
mat

cap +mat 

mat

wmat
mat

1
cap +lab +mat

(2.72)

1
mat ! cap +lab
+mat

(2.73)

This generally confirms the results of the linear production function for the relationships between capital and
labor and the relationship between capital and materials. However, in contrast to the linear production function,
the results obtained by the Cobb-Douglas functional form indicate that most firms should substitute materials
for labor (rather than the other way round).

73

2 Primal Approach: Production Function

xmat (w, y) =

y
A

wcap
cap

!cap 

wlab
lab

lab 

mat
wmat

1
cap +lab ! cap +lab
+mat

(2.74)

We can use these formulas to calculate the cost-minimizing input quantities based on the observed
input prices and the predicted output quantities. Alternatively, we could calculate the costminimizing input quantities based on the observed input prices and the observed output quantities. However, in the latter case, the predicted output quantities based on the cost-minimizing
input quantities would differ from the predicted output quantities based on the observed input
quantities so that a comparison of the cost-minimizing input quantities with the observed input
quantities would be less useful.
As the coefficients of the Cobb-Douglas function repeatedly occur in the formulas for calculating
the cost-minimizing input quantities, it is convenient to define short-cuts for them:
> A <- exp( coef( prodCD )[ "(Intercept)" ] )
> aCap <- coef( prodCD )[ "log(qCap)" ]
> aLab <- coef( prodCD )[ "log(qLab)" ]
> aMat <- coef( prodCD )[ "log(qMat)" ]
Now, we can calculate the cost-minimizing input quantities:
> dat$qCapCD <- with( dat,
+

( ( qOutCD / A ) * ( aCap / pCap )^( aLab + aMat )

* ( pLab / aLab )^aLab * ( pMat / aMat )^aMat

)^(1/( aCap + aLab + aMat ) ) )

> dat$qLabCD <- with( dat,

+
+
+

( ( qOutCD / A ) * ( pCap / aCap )^aCap

* ( aLab / pLab )^( aCap + aMat ) * ( pMat / aMat )^aMat
)^(1/( aCap + aLab + aMat ) ) )

> dat$qMatCD <- with( dat,

+
+
+

( ( qOutCD / A ) * ( pCap / aCap )^aCap

* ( pLab / aLab )^aLab * ( aMat / pMat )^( aCap + aLab )
)^(1/( aCap + aLab + aMat ) ) )

Before we continue, we will check whether it is indeed possible to produce the predicted output
with the calculated cost-minimizing input quantities:
> dat$qOutTest <- with( dat,
+

A * qCapCD^aCap * qLabCD^aLab * qMatCD^aMat )

> all.equal( dat$qOutCD, dat$qOutTest )

[1] TRUE
Given that the output quantities predicted from the cost-minimizing input quantities are all equal
to the output quantities predicted from the observed input quantities, we can be pretty sure that

74

2 Primal Approach: Production Function

our calculations are correct. Now, we can use scatter plots to compare the cost-minimizing input
quantities with the observed input quantities:
> compPlot( dat$qCapCD, dat$qCap )
> compPlot( dat$qLabCD, dat$qLab )
> compPlot( dat$qMatCD, dat$qMat )
> compPlot( dat$qCapCD, dat$qCap, log = "xy" )
> compPlot( dat$qLabCD, dat$qLab, log = "xy" )

6e+05

> compPlot( dat$qMatCD, dat$qMat, log = "xy" )

4e+05

6e+05

200000

600000

100000

1000000

qMatCD

5e+04

5e+05

5e+04

5e+03

5e+04

2e+04

5e+05

1e+05

100000

2e+05

2e+04

1e+05

60000

qLab

2e+04

qLabCD

5e+03

20000

qMat

5e+05

qCapCD

5e+03

qMat

20000

60000

1000000

2e+05

200000

2e+05
0e+00

qLab

qCap

600000

4e+05

0e+00

qCap

2e+05

qCapCD

5e+05

5e+03

qLabCD

2e+04

5e+04

qMatCD

Figure 2.20: Optimal and observed input quantities

The resulting graphs are shown in figure 2.20. As we already found out in section 2.4.14, many
firms could reduce their costs by substituting materials for capital.
We can also evaluate the potential for cost reductions by comparing the observed costs with
the costs when using the cost-minimizing input quantities:
> dat$costProdCD <- with( dat,
+

pCap * qCapCD + pLab * qLabCD + pMat * qMatCD )

> mean( dat$costProdCD / dat$cost )

75

2 Primal Approach: Production Function

[1] 0.9308039
Our model predicts that the firms could reduce their costs on average by 7% by using costminimizing input quantities. The variation of the firms cost reduction potentials are shown by
a histogram:

15
0 5

Frequency

25

> hist( dat$costProdCD / dat$cost )

0.75

0.80

0.85

0.90

0.95

1.00

costProdCD / cost

Figure 2.21: Minimum total costs as share of actual total costs

The resulting graph is shown in figure 2.21. While many firms have a rather small potential for
reducing costs by reallocating input quantities, there are some firms that could save up to 25%
of their total costs by using the optimal combination of input quantities.
We can also compare the observed input quantities with the cost-minimizing input quantities
and the observed costs with the minimum costs for each single observation (e.g. when consulting
individual firms in the sample):
> round( subset( dat, , c("qCap", "qCapCD", "qLab", "qLabCD", "qMat", "qMatCD",
+

"cost", "costProdCD") ) )[1:5,]

qCap qCapCD

qLab qLabCD

qMat qMatCD

cost costProdCD

1 84050

33720 360066 405349 34087

38038 846329

790968

2 39663

18431 249769 334442 40819

36365 580545

545777

3 37051

14257 140286 135701 24219

32176 306040

281401

4 21222

13300

83427

69713 18893

25890 199634

191709

5 44675

28400

89223 108761 14424

13107 226578

221302

2.4.16 Derived Input Demand Elasticities

We can measure the effect of the input prices and the output quantity on the cost-minimizing
input quantities by calculating the conditional price elasticities based on the partial derivatives
of the conditional input demand functions (2.71) with respect to the input prices and the output

76

2 Primal Approach: Production Function

quantity. In case of two inputs, we can calculate the demand elasticities of the first input by:
1

y 1 w2 2
x1 (w, y) =
A 2 w1
w1
x1 (w, y)
11 (w, y) =
w1
x1 (w, y)


1
=

y
A

1
=

1
=

1 w2
2 w1

y
A

y
A

1 y
=
A
1 2
= x1
x1
x1 (w, y)
12 (w, y) =
w2


1
=

y
A

y
A

1
=

1
=

y
A

1 w2
2 w1

2  1 1

y
A

y
A

2 1 

1 w2
2 w1

2 1

1 w2
2 w1

 2

1 w2

2 w12

1 w2 2
2 w1 x 1

2
x1

1 w2
2 w1

2  1 1

1 w2
2 w1

2  1 1

1 w2
2 w1

2  1 1

1

w1
x1

(2.77)
(2.78)
(2.79)

y
1 w2
2
A
2 w1


y
A

y
A

(2.80)
(2.81)
(2.82)

2 1

1 w2
2 w1

2 1

1 w2
2 w1

 2

1 1 w2
2 w1 x 1

1 w2 2
2 w1 x 1

2
x1

2
x1

(2.83)
(2.84)
(2.85)
(2.86)
(2.87)
(2.88)

1 w2
2 w1

2  1 1

1 w2
2 w1

2  1 1

1

1 w2 2 2
2 w1
x1
2
1
1
=
=
=
1

w2
x1 (w, y)

1 y 1 w2
=
A 2 w1
1
1
1
= x1
=
x1

2  1 1

y
A

y
1 w2
2
A
2 w1

1 w2
2 w1

1 y 1 w2 2
=
A 2 w1
1 2
2
= x1
=
x1

x1 (w, y)
y
1y (w, y) =
y
x1 (w, y)


(2.76)

2  1 1

y
A

(2.75)

2  1

1
A

y
A

1 w2
2 w1

 2

1 w2
2 w1

 2

1
x1

y
x1

(2.89)

1
x1

(2.90)
(2.91)
(2.92)

and analogously the demand elasticities of the second input:



x2 (w, y) =

y
A

2 w1
1 w2

1  1

77

(2.93)

2 Primal Approach: Production Function

w2
x2 (w, y)
1
2
2
=
=
=
1
w2
x2 (w, y)

x2 (w, y)
w1
1
21 (w, y) =
=
w1
x2 (w, y)

y
1
x2 (w, y)
= .
2y (w, y) =
y
x2 (w, y)

22 (w, y) =

(2.94)
(2.95)
(2.96)

One can similarly derive the input demand elasticities for the general case of N inputs:
xi (w, y) wj
j
=
ij
wj xi (w, y)

y
xi (w, y)
1
iy (w, y) =
= ,
y
xi (w, y)

ij (w, y) =

(2.97)
(2.98)

where ij is (again) Kroneckers delta (2.66). We have calculated all these elasticities based on the
estimated coefficients of the Cobb-Douglas production function; these elasticities are presented in
table 2.1. If the price of capital increases by one percent, the cost-minimizing firm will decrease
the use of capital by 0.89% and increase the use of labor and materials by 0.11% each. If the
price of labor increases by one percent, the cost-minimizing firm will decrease the use of labor
by 0.54% and increase the use of capital and materials by 0.46% each. If the price of materials
increases by one percent, the cost-minimizing firm will decrease the use of materials by 0.57%
and increase the use of capital and labor by 0.43% each. If the cost-minimizing firm increases
the output quantity by one percent, (s)he will increase all input quantities by 0.68%.
Table 2.1: Conditional demand elasticities
wcap
xcap -0.89
xlab
0.11
xmat 0.11

derived from Cobb-Douglas production function

wlab wmat
y
0.46 0.43 0.68
-0.54 0.43 0.68
0.46 -0.57 0.68

2.5 Quadratic Production Function

2.5.1 Specification
A quadratic production function is defined as
y = 0 +

i xi +

1 XX
ij xi xj ,
2 i j

(2.99)

where the restriction ij = ji is required to identify all coefficients, because xi xj and xj xi are
the same regressors. Based on this general form, we can derive the specification of a quadratic

78

2 Primal Approach: Production Function

production function with three inputs:
1
1
1
y = 0 +1 x1 +2 x2 +3 x3 + 11 x21 + 22 x22 + 33 x23 +12 x1 x2 +13 x1 x3 +23 x2 x3 (2.100)
2
2
2

2.5.2 Estimation
We can estimate this quadratic production function with the command
> prodQuad <- lm( qOut ~ qCap + qLab + qMat
+

+ I( 0.5 * qCap^2 ) + I( 0.5 * qLab^2 ) + I( 0.5 * qMat^2 )

+ I( qCap * qLab ) + I( qCap * qMat ) + I( qLab * qMat ),

data = dat )

> summary( prodQuad )

Call:
lm(formula = qOut ~ qCap + qLab + qMat + I(0.5 * qCap^2) + I(0.5 *
qLab^2) + I(0.5 * qMat^2) + I(qCap * qLab) + I(qCap * qMat) +
I(qLab * qMat), data = dat)
Residuals:
Min

1Q

Median

3Q

Max

-3928802

-695518

-186123

545509

4474143

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept)

-2.911e+05

3.615e+05

-0.805 0.422072

qCap

5.270e+00

4.403e+00

1.197 0.233532

qLab

6.077e+00

3.185e+00

1.908 0.058581 .

qMat

1.430e+01

2.406e+01

0.595 0.553168

I(0.5 * qCap^2)

5.032e-05

3.699e-05

1.360 0.176039

I(0.5 * qLab^2) -3.084e-05

2.081e-05

-1.482 0.140671

I(0.5 * qMat^2) -1.896e-03

8.951e-04

-2.118 0.036106 *

I(qCap * qLab)

-3.097e-05

1.498e-05

-2.067 0.040763 *

I(qCap * qMat)

-4.160e-05

1.474e-04

-0.282 0.778206

I(qLab * qMat)

4.011e-04

1.112e-04

3.608 0.000439 ***

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Residual standard error: 1344000 on 130 degrees of freedom

Multiple R-squared:

0.8449,

F-statistic: 78.68 on 9 and 130 DF,

Adjusted R-squared:
p-value: < 2.2e-16

79

0.8342

2 Primal Approach: Production Function

Although many of the estimated coefficients are statistically not significantly different from zero,
the statistical significance of some quadratic and interaction terms indicates that the linear production function, which neither has quadratic terms not interaction terms, is not suitable to model
the true production technology. As the linear production function is nested in the quadratic
production function, we can apply a Wald test or a likelihood ratio test to check whether the
linear production function is rejected in favor of the quadratic production function. These tests
can be done by the functions waldtest and lrtest (package lmtest):
> library( "lmtest" )
> waldtest( prodLin, prodQuad )
Wald test
Model 1: qOut ~ qCap + qLab + qMat
Model 2: qOut ~ qCap + qLab + qMat + I(0.5 * qCap^2) + I(0.5 * qLab^2) +
I(0.5 * qMat^2) + I(qCap * qLab) + I(qCap * qMat) + I(qLab *
qMat)
Res.Df Df
1

136

130

Pr(>F)

6 8.1133 1.869e-07 ***

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

> lrtest( prodLin, prodQuad )

Likelihood ratio test
Model 1: qOut ~ qCap + qLab + qMat
Model 2: qOut ~ qCap + qLab + qMat + I(0.5 * qCap^2) + I(0.5 * qLab^2) +
I(0.5 * qMat^2) + I(qCap * qLab) + I(qCap * qMat) + I(qLab *
qMat)
#Df

LogLik Df

5 -2191.3

11 -2169.1

Chisq Pr(>Chisq)

6 44.529

5.806e-08 ***

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

These tests show that the linear production function is clearly inferior to the quadratic production
function and hence, should not be used for analyzing the production technology of the firms in
this data set.

80

2 Primal Approach: Production Function

2.5.3 Properties
We cannot see from the estimated coefficients whether the monotonicity condition is fulfilled.
Unless all coefficients are non-negative (but not necessarily the intercept), quadratic production
functions cannot be globally monotone, because there will always be a set of input quantities
that result in negative marginal products. We will check the monotonicity condition at each
observation in section 2.5.5.
Our estimated quadratic production function does not fulfill the weak essentiality assumption,
because the intercept is different from zero (but its deviation from zero is not statistically significant). The production technology described by a quadratic production function with more than
one (relevant) input never shows strict essentiality.
The input requirement sets derived from quadratic production functions are always closed and
non-empty.
The quadratic production function always returns finite, real, and single values but the nonnegativity assumption is only fulfilled, if all coefficients (including the intercept), are non-negative.
All quadratic production functions are continuous and twice-continuously differentiable.

2.5.4 Predicted output quantities

We can obtain the predicted output quantities with the fitted method:
> dat$qOutQuad <- fitted( prodQuad )
We can evaluate the fit of the model by comparing the observed with the fitted output
quantities:
> compPlot( dat$qOut, dat$qOutQuad )
> compPlot( dat$qOut, dat$qOutQuad, log = "xy" )

0.0e+00

1e+07
1e+05

5e+05 2e+06

fitted

1.0e+07
0.0e+00

fitted

2.0e+07

1.0e+07

2.0e+07

1e+05

observed

5e+05

5e+06

observed

Figure 2.22: Quadratic production function: fit of the model

81

2 Primal Approach: Production Function

The resulting graphs are shown in figure 2.22. While the graph in the left panel uses a linear
scale for the axes, the graph in the right panel uses a logarithmic scale for both axes. Hence, the
deviations from the 45-line illustrate the absolute deviations in the left panel and the relative
deviations in the right panel. The fit of the model looks okay in the scatter plot on the left-hand
side, but if we use a logarithmic scale on both axes (as in the graph on the right-hand side), we
can see that the output quantity is over-estimated if the the observed output quantity is small.
As negative output quantities would render the corresponding output elasticities useless, we
have carefully check the sign of the predicted output quantities:
> sum( dat$qOutQuad < 0 )
[1] 0
Fortunately, not a single predicted output quantity is negative.

2.5.5 Marginal Products

In case of a quadratic production function, the marginal products are
M Pi = i +

ij xj

(2.101)

We can simplify the code for computing the marginal products and some other figures by using
short names for the coefficients:
> b1 <- coef( prodQuad )[ "qCap" ]
> b2 <- coef( prodQuad )[ "qLab" ]
> b3 <- coef( prodQuad )[ "qMat" ]
> b11 <- coef( prodQuad )[ "I(0.5 * qCap^2)" ]
> b22 <- coef( prodQuad )[ "I(0.5 * qLab^2)" ]
> b33 <- coef( prodQuad )[ "I(0.5 * qMat^2)" ]
> b12 <- b21 <- coef( prodQuad )[ "I(qCap * qLab)" ]
> b13 <- b31 <- coef( prodQuad )[ "I(qCap * qMat)" ]
> b23 <- b32 <- coef( prodQuad )[ "I(qLab * qMat)" ]
Now, we can use the following commands to calculate the marginal products in R:
> dat$mpCapQuad <- with( dat,
+

b1 + b11 * qCap + b12 * qLab + b13 * qMat )

> dat$mpLabQuad <- with( dat,

+

b2 + b21 * qCap + b22 * qLab + b23 * qMat )

> dat$mpMatQuad <- with( dat,

+

b3 + b31 * qCap + b32 * qLab + b33 * qMat )

82

2 Primal Approach: Production Function

We can visualize (the variation of) these marginal products with histograms:
> hist( dat$mpCapQuad, 15 )
> hist( dat$mpLabQuad, 15 )

10

10

30
10
0

0
20

20

Frequency

30
10

20

Frequency

30
20
0

10

Frequency

40

40

> hist( dat$mpMatQuad, 15 )

10

MP Cap

15

20

25

30

50

50 100

MP Lab

200

MP Mat

Figure 2.23: Quadratic production function: marginal products

The resulting graphs are shown in figure 2.23. If the firms increase capital input by one unit,
the output of most firms will increase by around 2 units. If the firms increase labor input by
one unit, the output of most firms will increase by around 5 units. If the firms increase material
input by one unit, the output of most firms will increase by around 50 units. These graphs also
show that the monotonicity condition is not fulfilled for all observations:
> sum( dat$mpCapQuad < 0 )
[1] 28
> sum( dat$mpLabQuad < 0 )
[1] 5
> sum( dat$mpMatQuad < 0 )
[1] 8
> dat$monoQuad <- with( dat, mpCapQuad >= 0 & mpLabQuad >= 0 & mpMatQuad >= 0 )
> sum( !dat$monoQuad )
[1] 39
28 firms have a negative marginal product of capital, 5 firms have a negative marginal product
of labor, and 8 firms have a negative marginal product of materials. In total the monotonicity
condition is not fulfilled at 39 out of 140 observations. Although the monotonicity conditions are
still fulfilled for the largest part of firms in our data set, these frequent violations could indicate
a possible model misspecification.

83

2 Primal Approach: Production Function

2.5.6 Output Elasticities

We can obtain output elasticities based on the quadratic production function by the standard
formula for output elasticities:
i = M Pi

xi
y

(2.102)

As explained in section 2.4.11.1, we will use the predicted output quantities rather than the
observed output quantities. We can calculate the output elasticities with:
> dat$eCapQuad <- with( dat, mpCapQuad * qCap / qOutQuad )
> dat$eLabQuad <- with( dat, mpLabQuad * qLab / qOutQuad )
> dat$eMatQuad <- with( dat, mpMatQuad * qMat / qOutQuad )
We can visualize (the variation of) these output elasticities with histograms:
> hist( dat$eCapQuad, 15 )
> hist( dat$eLabQuad, 15 )

0.4

0.0

0.4

0.8

50
40
30
0

10

20

Frequency

30
10

20

Frequency

30
20
10

Frequency

40

50

60

> hist( dat$eMatQuad, 15 )

0.5 0.0

0.5

eCap

1.0

1.5

2.0

2.5

1.5

eLab

0.5

0.5 1.0 1.5

eMat

Figure 2.24: Quadratic production function: output elasticities

The resulting graphs are shown in figure 2.24. If the firms increase capital input by one percent,
the output of most firms will increase by around 0.05 percent. If the firms increase labor input
by one percent, the output of most firms will increase by around 0.7 percent. If the firms increase
material input by one percent, the output of most firms will increase by around 0.5 percent.

2.5.7 Elasticity of Scale

The elasticity of scale canas alwaysbe calculated as the sum of all output elasticities.
> dat$eScaleQuad <- dat$eCapQuad + dat$eLabQuad +
+

dat$eMatQuad

The (variation of the) elasticities of scale can be visualized with a histogram.

84

2 Primal Approach: Production Function

> hist( dat$eScaleQuad, 30 )

12
0.8

1.0

1.2

1.4

1.6

0 2 4 6 8

Frequency

25
15
0 5

Frequency

> hist( dat$eScaleQuad[ dat$monoQuad ], 30 )

1.1

eScaleQuad

1.3

1.5

1.7

eScaleQuad[ monoQuad ]

Figure 2.25: Quadratic production function: elasticities of scale

The resulting graphs are shown in figure 2.25. Only a very few firms (4 out of 140) experience
decreasing returns to scale. If we only consider the observations where all monotonicity conditions
are fulfilled, our results suggest that all firms have increasing returns to scale. Most firms have an
elasticity of scale around 1.3. Hence, if these firms increase all input quantities by one percent,
the output of most firms will increase by around 1.3 percent. These elasticities of scale are much
more realistic than the elasticities of scale based on the linear production function.
Information on the optimal firm size can be obtained by analyzing the interrelationship between
firm size and the elasticity of scale, where we can either use the observed output or the quantity
index of the inputs as proxies of the firm size:
> plot( dat$qOut, dat$eScaleQuad, log = "x" )
> abline( 1, 0 )
> plot( dat$X, dat$eScaleQuad, log = "x" )
> abline( 1, 0 )
> plot( dat$qOut[ dat$monoQuad ], dat$eScaleQuad[ dat$monoQuad ], log = "x" )
> plot( dat$X[ dat$monoQuad ], dat$eScaleQuad[ dat$monoQuad ], log = "x" )
The resulting graphs are shown in figure 2.26. They all indicate that there are increasing returns
to scale for all firm sizes in the sample. Hence, all firms in the sample would gain from increasing
their size and the optimal firm size seems to be larger than the largest firm in the sample.

2.5.8 Marginal Rates of Technical Substitution

We can calculate the marginal rates of technical substitution (MRTS) based on our estimated
quadratic production function by following commands:

85

2 Primal Approach: Production Function

0.8 1.0 1.2 1.4 1.6

eScaleQuad

0.8 1.0 1.2 1.4 1.6

eScaleQuad

5e+05

2e+06

1e+07

0.5

1e+05

5e+05

2e+06

1.7
1.5

2.0

5.0

1.3

eScaleQuad[ monoQuad ]

1.3

1.0

quantity index of inputs

1.1

1.7
1.5

1.1

eScaleQuad[ monoQuad ]

observed output

1e+05

5e+06

0.5

observed output

1.0

2.0

quantity index of inputs

Figure 2.26: Quadratic production function: elasticities of scale at different firm sizes
> dat$mrtsCapLabQuad <- with( dat, - mpLabQuad / mpCapQuad )
> dat$mrtsLabCapQuad <- with( dat, - mpCapQuad / mpLabQuad )
> dat$mrtsCapMatQuad <- with( dat, - mpMatQuad / mpCapQuad )
> dat$mrtsMatCapQuad <- with( dat, - mpCapQuad / mpMatQuad )
> dat$mrtsLabMatQuad <- with( dat, - mpMatQuad / mpLabQuad )
> dat$mrtsMatLabQuad <- with( dat, - mpLabQuad / mpMatQuad )
As the marginal rates of technical substitution (MRTS) are meaningless if the monotonicity
condition is not fulfilled, we visualize (the variation of) these MRTSs only for the observations,
where the monotonicity condition is fulfilled:
> hist( dat$mrtsCapLabQuad[ dat$monoQuad ], 30 )
> hist( dat$mrtsLabCapQuad[ dat$monoQuad ], 30 )
> hist( dat$mrtsCapMatQuad[ dat$monoQuad ], 30 )
> hist( dat$mrtsMatCapQuad[ dat$monoQuad ], 30 )
> hist( dat$mrtsLabMatQuad[ dat$monoQuad ], 30 )
> hist( dat$mrtsMatLabQuad[ dat$monoQuad ], 30 )
The resulting graphs are shown in figure 2.27. As some outliers hide the variation of the majority
of the RMRTS, we use function colMedians (package miscTools) to show the median values of
the MRTS:

86

40

20

80
60
20
0

10
0

0
60

40

Frequency

30
20

Frequency

30
20
10

Frequency

40

40

50

2 Primal Approach: Production Function

15

1000

600

mrtsLabCapQuad

200

mrtsCapMatQuad

10
0

5
0
6

20

Frequency

10

Frequency

40
0

20

Frequency

60

30

15

80

mrtsCapLabQuad

10

60

mrtsMatCapQuad

40

20

mrtsLabMatQuad

mrtsMatLabQuad

Figure 2.27: Quadratic production function: marginal rates of technical substitution (RMRTS)
> colMedians( subset( dat, monoQuad,
+
+

c( "mrtsCapLabQuad", "mrtsLabCapQuad", "mrtsCapMatQuad",

"mrtsMatCapQuad", "mrtsLabMatQuad", "mrtsMatLabQuad" ) ) )

mrtsCapLabQuad mrtsLabCapQuad mrtsCapMatQuad mrtsMatCapQuad mrtsLabMatQuad

-2.23505371

-0.44741654

-14.19802214

-0.07043235

-7.86423950

mrtsMatLabQuad
-0.12715788
Given that the median marginal rate of technical substitution between capital and labor is -2.24,
a typical firm that reduces the use of labor by one unit, has to use around 2.24 additional units
of capital in order to produce the same amount of output as before. Alternatively, the typical
firm can replace one unit of labor by using 0.13 additional units of materials.

2.5.9 Relative Marginal Rates of Technical Substitution

As we do not have a practical interpretation of the units of measurement of the input quantities,
the relative marginal rates of technical substitution (RMRTS) are practically more meaningful
than the MRTS. The following commands calculate the RMRTS:

87

2 Primal Approach: Production Function

> dat$rmrtsCapLabQuad <- with( dat, - eLabQuad / eCapQuad )
> dat$rmrtsLabCapQuad <- with( dat, - eCapQuad / eLabQuad )
> dat$rmrtsCapMatQuad <- with( dat, - eMatQuad / eCapQuad )
> dat$rmrtsMatCapQuad <- with( dat, - eCapQuad / eMatQuad )
> dat$rmrtsLabMatQuad <- with( dat, - eMatQuad / eLabQuad )
> dat$rmrtsMatLabQuad <- with( dat, - eLabQuad / eMatQuad )
As the (relative) marginal rates of technical substitution are meaningless if the monotonicity
condition is not fulfilled, we visualize (the variation of) these RMRTSs only for the observations,
where the monotonicity condition is fulfilled:
> hist( dat$rmrtsCapLabQuad[ dat$monoQuad ], 30 )
> hist( dat$rmrtsLabCapQuad[ dat$monoQuad ], 30 )
> hist( dat$rmrtsCapMatQuad[ dat$monoQuad ], 30 )
> hist( dat$rmrtsMatCapQuad[ dat$monoQuad ], 30 )
> hist( dat$rmrtsLabMatQuad[ dat$monoQuad ], 30 )

15

10

700

rmrtsLabCapQuad

Frequency

15

100

rmrtsCapMatQuad

10

10

Frequency

300

25

80
60
40
40

30

20

10

rmrtsMatCapQuad

20

Frequency

500

30

20

rmrtsCapLabQuad

60
20

20

35

200

20

400

15

600

0
800

40

Frequency

20

40

Frequency

40
0

20

Frequency

60

60

80

80

80

> hist( dat$rmrtsMatLabQuad[ dat$monoQuad ], 30 )

rmrtsLabMatQuad

15

10

rmrtsMatLabQuad

Figure 2.28: Quadratic production function: relative marginal rates of technical substitution
(RMRTS)
The resulting graphs are shown in figure 2.28. As some outliers hide the variation of the majority
of the RMRTS, we use function colMedians (package miscTools) to show the median values of

88

2 Primal Approach: Production Function

the RMRTS:
> colMedians( subset( dat, monoQuad,
+
+

c( "rmrtsCapLabQuad", "rmrtsLabCapQuad", "rmrtsCapMatQuad",

"rmrtsMatCapQuad", "rmrtsLabMatQuad", "rmrtsMatLabQuad" ) ) )

rmrtsCapLabQuad rmrtsLabCapQuad rmrtsCapMatQuad rmrtsMatCapQuad rmrtsLabMatQuad

-5.5741780

-0.1793986

-4.2567577

-0.2349206

-0.7745132

rmrtsMatLabQuad
-1.2911336
Given that the median relative marginal rate of technical substitution between capital and labor
is -5.57, a typical firm that reduces the use of labor by one percent, has to use around 5.57 percent
more capital in order to produce the same amount of output as before. Alternatively, the typical
firm can replace one percent of labor by using 1.29 percent more materials.

2.5.10 Elasticities of Substitution

In the following, we only calculate the Allen elasticities of substitution. The calculation of
the direct elasticities of substitution and the Morishima elasticities of substitution requires only
minimal changes of the code. In order to check whether our calculations are correct, we can use
equation (2.15) to derive the following conditions:
X

xi M Pi ij = 0 j

(2.103)

In order to check this condition, we need to calculate not only (normal) elasticities of substitution
(ij ; i 6= j) but also economically not meaningful elasticities of self-substitution (ii ):
> dat$esaCapLabQuad <- NA
> dat$esaCapMatQuad <- NA
> dat$esaLabMatQuad <- NA
> dat$esaCapCapQuad <- NA
> dat$esaLabLabQuad <- NA
> dat$esaMatMatQuad <- NA
> for( obs in 1:nrow( dat ) ) {
+

bhmLoop <- matrix( 0, nrow = 4, ncol = 4 )

bhmLoop[ 1, 2 ] <- bhmLoop[ 2, 1 ] <- dat$mpCapQuad[ obs ]

bhmLoop[ 1, 3 ] <- bhmLoop[ 3, 1 ] <- dat$mpLabQuad[ obs ]

bhmLoop[ 1, 4 ] <- bhmLoop[ 4, 1 ] <- dat$mpMatQuad[ obs ]

bhmLoop[ 2, 2 ] <- b11

bhmLoop[ 3, 3 ] <- b22

bhmLoop[ 4, 4 ] <- b33

89

2 Primal Approach: Production Function

+

bhmLoop[ 2, 3 ] <- bhmLoop[ 3, 2 ] <- b12

bhmLoop[ 2, 4 ] <- bhmLoop[ 4, 2 ] <- b13

bhmLoop[ 3, 4 ] <- bhmLoop[ 4, 3 ] <- b23

FCapLabLoop <- - det( bhmLoop[ -2, -3 ] )

FCapMatLoop <- det( bhmLoop[ -2, -4 ] )

FLabMatLoop <- - det( bhmLoop[ -3, -4 ] )

FCapCapLoop <- det( bhmLoop[ -2, -2 ] )

FLabLabLoop <- det( bhmLoop[ -3, -3 ] )

FMatMatLoop <- det( bhmLoop[ -4, -4 ] )

numerator <- with( dat[ obs, ],

+
+
+
+
+
+
+
+
+
+
+
+
+

qCap * mpCapQuad + qLab * mpLabQuad + qMat * mpMatQuad )

dat$esaCapLabQuad[ obs ] <- with( dat[obs,],
numerator / ( qCap * qLab ) * FCapLabLoop / det( bhmLoop ) )
dat$esaCapMatQuad[ obs ] <- with( dat[ obs, ],
numerator / ( qCap * qMat ) * FCapMatLoop / det( bhmLoop ) )
dat$esaLabMatQuad[ obs ] <- with( dat[ obs, ],
numerator / ( qLab * qMat ) * FLabMatLoop / det( bhmLoop ) )
dat$esaCapCapQuad[ obs ] <- with( dat[obs,],
numerator / ( qCap * qCap ) * FCapCapLoop / det( bhmLoop ) )
dat$esaLabLabQuad[ obs ] <- with( dat[ obs, ],
numerator / ( qLab * qLab ) * FLabLabLoop / det( bhmLoop ) )
dat$esaMatMatQuad[ obs ] <- with( dat[ obs, ],
numerator / ( qMat * qMat ) * FMatMatLoop / det( bhmLoop ) )

+ }
Before we take a look at and interpret the elasticities of substitution, we check whether the
conditions (2.103) are fulfilled:
> range( with( dat, qCap * mpCapQuad * esaCapCapQuad +
+

qLab * mpLabQuad * esaCapLabQuad + qMat * mpMatQuad * esaCapMatQuad ) )

[1] -3.725290e-08

1.043081e-07

> range( with( dat, qCap * mpCapQuad * esaCapLabQuad +

+

qLab * mpLabQuad * esaLabLabQuad + qMat * mpMatQuad * esaLabMatQuad ) )

[1] -3.725290e-09

2.095476e-09

> range( with( dat, qCap * mpCapQuad * esaCapMatQuad +

+

qLab * mpLabQuad * esaLabMatQuad + qMat * mpMatQuad * esaMatMatQuad ) )

[1] -1.117587e-08

1.396984e-09

90

2 Primal Approach: Production Function

The extremely small deviations from zero are most likely caused by rounding errors that are
unavoidable on digital computers. This test does not proof that all our calculations are done
correctly but if we had made a mistake, we would have discovered it with a very high probability.
Hence, we can be rather sure that our calculations are correct.
As the elasticities of substitution measure changes in the marginal rates of technical substitution
(MRTS) and the MRTS are meaningless if the monotonicity conditions are not fulfilled, also the
elasticities of substitution are meaningless if the monotonicity conditions are not fulfilled. Hence,
we visualize (the variation of) the Allen elasticities of substitution only for the observations,
where the monotonicity condition is fulfilled:
> hist( dat$esaCapLabQuad[ dat$monoQuad ], 30 )
> hist( dat$esaCapMatQuad[ dat$monoQuad ], 30 )

25
20
15
10

Frequency

15
5

10

Frequency

30
20

esaCapLabQuad

10

Frequency

40

50

20

> hist( dat$esaLabMatQuad[ dat$monoQuad ], 30 )

esaCapMatQuad

0.0

0.5

1.0

1.5

2.0

2.5

esaLabMatQuad

Figure 2.29: Quadratic production function: elasticities of substitution

The resulting graphs are shown in figure 2.29. The estimated elasticities of substitution suggest
that capital and labor are always complements, labor and materials are always substitutes, and
capital and materials are partly complements and partly substitutes. The estimated elasticity
of substitution between labor and materials lies for the most firms between the value of the
Leontief production function ( = 0) and the values of the Cobb-Douglas production function
( = 1). Hence, the substitutability between labor and materials seems to be between very low
and moderate. In fact, the elasticity of substitution between labor and materials is for a large
share of firms around 0.5. Hence, if labor is substituted for materials (or vice versa) so that the
MRTS between labor and materials increases (decreases) by one percent, the ratio between the
labor quantity and the quantity of materials increases (decreases) by 0.5 percent. If the firm is
minimizing costs and the price ratio between materials and labor increases by one percent, the
firm will substitute labor for materials so that ratio between the labor quantity and the quantity
of materials increases by 0.5 percent. Hence, the relative change of the quantity ratios is smaller
than the relative change of price ratios, which indicates a low substitutability between labor and
materials.

91

2 Primal Approach: Production Function

2.5.11 Quasiconcavity
We check whether our estimated quadratic production function is quasiconcave at each observation:
> dat$quasiConcQuad <- NA
> for( obs in 1:nrow( dat ) ) {
+

bhmLoop <- matrix( 0, nrow = 4, ncol = 4 )

bhmLoop[ 1, 2 ] <- bhmLoop[ 2, 1 ] <- dat$mpCapQuad[ obs ]

bhmLoop[ 1, 3 ] <- bhmLoop[ 3, 1 ] <- dat$mpLabQuad[ obs ]

bhmLoop[ 1, 4 ] <- bhmLoop[ 4, 1 ] <- dat$mpMatQuad[ obs ]

bhmLoop[ 2, 2 ] <- b11

bhmLoop[ 3, 3 ] <- b22

bhmLoop[ 4, 4 ] <- b33

bhmLoop[ 2, 3 ] <- bhmLoop[ 3, 2 ] <- b12

bhmLoop[ 2, 4 ] <- bhmLoop[ 4, 2 ] <- b13

bhmLoop[ 3, 4 ] <- bhmLoop[ 4, 3 ] <- b23

dat$quasiConcQuad[ obs ] <- det( bhmLoop[ 1:2, 1:2 ] ) < 0 &

det( bhmLoop[ 1:3, 1:3 ] ) > 0 & det( bhmLoop ) < 0

+ }
> sum( dat$quasiConcQuad )
[1] 0
Our estimated quadratic production function is quasiconcave at none of the 140 observations.

2.5.12 First-order conditions for profit maximisation

In this section, we will check to what extent the first-order conditions for profit maximization
(2.24) are fulfilled, i.e. to what extent the firms use the optimal input quantities. We do this by
comparing the marginal value products of the inputs with the corresponding input prices. We
can calculate the marginal value products by multiplying the marginal products by the output
price:
> dat$mvpCapQuad <- dat$pOut * dat$mpCapQuad
> dat$mvpLabQuad <- dat$pOut * dat$mpLabQuad
> dat$mvpMatQuad <- dat$pOut * dat$mpMatQuad
The command compPlot (package miscTools) can be used to compare the marginal value products
with the corresponding input prices. As the logarithm of a non-positive number is not defined,
we have to limit the comparisons on the logarithmic scale to observations with positive marginal
products:

92

2 Primal Approach: Production Function

> compPlot( dat$pCap, dat$mvpCapQuad )
> compPlot( dat$pLab, dat$mvpLabQuad )
> compPlot( dat$pMat, dat$mvpMatQuad )
> compPlot( dat$pCap[ dat$monoQuad ], dat$mvpCapQuad[ dat$monoQuad ], log = "xy" )
> compPlot( dat$pLab[ dat$monoQuad ], dat$mvpLabQuad[ dat$monoQuad ], log = "xy" )
> compPlot( dat$pMat[ dat$monoQuad ], dat$mvpMatQuad[ dat$monoQuad ], log = "xy" )

20

500
400

10

w Cap

300
200

40

0.1

1.0

0.5

2.0

100

300

5.0

500

w Mat
100
10

20

50

MVP Mat

5.0 10.0

0.5

0.1

30

2.0

5.0
2.0

0.5

MVP Cap

MVP Lab

20

w Lab

100

10
0

60

40

30

60

MVP Mat

20

20

MVP Lab

40

MVP Cap

40

0.5 1.0 2.0

w Cap

5.0

20.0

10

w Lab

20

50

100

w Mat

Figure 2.30: Marginal value products and corresponding input prices

The resulting graphs are shown in figure 2.30. They indicate that the marginal value products of
most firms are higher than the corresponding input prices. This indicates that most firms could
increase their profit by using more of all inputs. Given that the estimated quadratic function
shows that (almost) all firms operate under increasing returns to scale, it is not surprising that
most firms would gain from increasing all input quantities. Therefore, the question arises why the
firms in the sample did not do this. This questions has already been addressed in section 2.3.10.

2.5.13 First-order conditions for cost minimization

As the marginal rates of technical substitution differ between observations for the three other
functional forms, we use scatter plots for visualizing the comparison of the input price ratios

93

2 Primal Approach: Production Function

with the negative inverse marginal rates of technical substitution. As the marginal rates of
technical substitution are meaningless if the monotonicity condition is not fulfilled, we limit the
comparisons to the observations, where all monotonicity conditions are fulfilled:
> compPlot( ( dat$pCap / dat$pLab )[ dat$monoQuad ],
+

- dat$mrtsLabCapQuad[ dat$monoQuad ] )

> compPlot( ( dat$pCap / dat$pMat )[ dat$monoQuad ],

+

- dat$mrtsMatCapQuad[ dat$monoQuad ] )

> compPlot( ( dat$pLab / dat$pMat )[ dat$monoQuad ],

+

- dat$mrtsMatLabQuad[ dat$monoQuad ] )

> compPlot( ( dat$pCap / dat$pLab )[ dat$monoQuad ],

+

- dat$mrtsLabCapQuad[ dat$monoQuad ], log = "xy" )

> compPlot( ( dat$pCap / dat$pMat )[ dat$monoQuad ],

+

- dat$mrtsMatCapQuad[ dat$monoQuad ], log = "xy" )

> compPlot( ( dat$pLab / dat$pMat )[ dat$monoQuad ],

+

- dat$mrtsMatLabQuad[ dat$monoQuad ], log = "xy" )

15

10

15

10.00

2.00
0.010

w Cap / w Lab

0.100

1.000

0.20 0.50

0.02

Figure 2.31: First-order conditions for costs minimization

94

w Cap / w Mat

The resulting graphs are shown in figure 2.31.

w Lab / w Mat

0.02 0.05

1.000
0.100

0.001

0.50 2.00

0.001

MRTS Mat Cap

0.50
0.10

0.010

10.00
2.00

0.02

0.10

w Cap / w Mat

MRTS Lab Cap

w Cap / w Lab

0.02

MRTS Mat Lab

4
3
0

MRTS Mat Lab

MRTS Mat Cap

10
5
0

MRTS Lab Cap

0.10

0.50

w Lab / w Mat

2.00

2 Primal Approach: Production Function

Furthermore, we use histograms to visualize the (absolute and relative) differences between the
input price ratios and the corresponding negative inverse marginal rates of technical substitution:
> hist( ( - dat$mrtsLabCapQuad - dat$pCap / dat$pLab )[ dat$monoQuad ] )
> hist( ( - dat$mrtsMatCapQuad - dat$pCap / dat$pMat )[ dat$monoQuad ] )
> hist( ( - dat$mrtsMatLabQuad - dat$pLab / dat$pMat )[ dat$monoQuad ] )
> hist( log( - dat$mrtsLabCapQuad / ( dat$pCap / dat$pLab ) )[ dat$monoQuad ] )
> hist( log( - dat$mrtsMatCapQuad / (dat$pCap / dat$pMat ) )[ dat$monoQuad ] )

40
20
10

20
0

10

10
0
5

15

MrtsMatCap wCap / wMat

log(MrtsLabCap / (wCap / wLab))

40
10

20

Frequency

15
10

10

15

Frequency

20

30

20

25

MrtsMatLab wLab / wMat

25

30

MrtsLabCap wCap / wLab

Frequency

30

Frequency

40

Frequency

40
30
20

Frequency

50

50

60

60

60

70

> hist( log( - dat$mrtsMatLabQuad / (dat$pLab / dat$pMat ) )[ dat$monoQuad ] )

log(MrtsMatCap / (wCap / wMat))

3 2 1

log(MrtsMatLab / (Lab / wMat))

Figure 2.32: First-order conditions for costs minimization

The resulting graphs are shown in figure 2.32. The left graphs in figures 2.31 and 2.32 show that
the ratio between the capital price and the labor price is larger than the absolute value of the
marginal rate of technical substitution between labor and capital for a majority of the firms in
the sample:
wcap
M Pcap
> M RT Slab,cap =
wlab
M Plab

(2.104)

Hence, these firms can get closer to the minimum of their production costs by substituting labor
for capital, because this will decrease the marginal product of labor and increase the marginal

95

2 Primal Approach: Production Function

product of capital so that the absolute value of the MRTS between labor and capital increases
and gets closer to the corresponding input price ratio. Similarly, the graphs in the middle column
indicate that a majority of the firms should substitute materials for capital and the graphs on
the right indicate that a little more than half of the firms should substitute materials for labor.
Hence, the majority of the firms could reduce production costs particularly by using less capital
and using more labor or more materials.

2.6 Translog Production Function

2.6.1 Specification
The Translog function is a more flexible extension of the Cobb-Douglas function as the quadratic
function is a more flexible extension of the linear function. Hence, the Translog function can be
seen as a combination of the Cobb-Douglas function and the quadratic function. The Translog
production function has following specification:
ln y = 0 +

1 XX
ij ln xi ln xj
2 i j

i ln xi +

with ij = ji .

2.6.2 Estimation
We can estimate this Translog production function with the command
> prodTL <- lm( log( qOut ) ~ log( qCap ) + log( qLab ) + log( qMat )
+

+ I( 0.5 * log( qCap )^2 ) + I( 0.5 * log( qLab )^2 )

+ I( 0.5 * log( qMat )^2 ) + I( log( qCap ) * log( qLab ) )

+ I( log( qCap ) * log( qMat ) ) + I( log( qLab ) * log( qMat ) ),

data = dat )

> summary( prodTL )

Call:
lm(formula = log(qOut) ~ log(qCap) + log(qLab) + log(qMat) +
I(0.5 * log(qCap)^2) + I(0.5 * log(qLab)^2) + I(0.5 * log(qMat)^2) +
I(log(qCap) * log(qLab)) + I(log(qCap) * log(qMat)) + I(log(qLab) *
log(qMat)), data = dat)
Residuals:
Min

1Q

Median

3Q

Max

-1.68015 -0.36688

0.05389

0.44125

1.26560

This generally confirms the results of the Cobb-Douglas production function.

96

(2.105)

2 Primal Approach: Production Function

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept)

-4.14581

21.35945

-0.194

0.8464

log(qCap)

-2.30683

2.28829

-1.008

0.3153

log(qLab)

1.99328

4.56624

0.437

0.6632

log(qMat)

2.23170

3.76334

0.593

0.5542

I(0.5 * log(qCap)^2)

-0.02573

0.20834

-0.124

0.9019

I(0.5 * log(qLab)^2)

-1.16364

0.67943

-1.713

0.0892 .

I(0.5 * log(qMat)^2)

-0.50368

0.43498

-1.158

0.2490

0.56194

0.29120

1.930

0.0558 .

I(log(qCap) * log(qMat)) -0.40996

0.23534

-1.742

0.0839 .

I(log(qLab) * log(qMat))

0.42750

1.539

I(log(qCap) * log(qLab))

0.65793

0.1262

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Residual standard error: 0.6412 on 130 degrees of freedom

Multiple R-squared:

0.6296,

F-statistic: 24.55 on 9 and 130 DF,

Adjusted R-squared:

0.6039

p-value: < 2.2e-16

None of the estimated coefficients is statistically significantly different from zero at the 5% significance level and only three coefficients are statistically significant at the 10% level. As the
Cobb-Douglas production function is nested in the Translog production function, we can apply
a Wald test or likelihood ratio test to check whether the Cobb-Douglas production function is
rejected in favor of the Translog production function. This can be done by the functions waldtest
and lrtest (package lmtest):
> waldtest( prodCD, prodTL )
Wald test
Model 1: log(qOut) ~ log(qCap) + log(qLab) + log(qMat)
Model 2: log(qOut) ~ log(qCap) + log(qLab) + log(qMat) + I(0.5 * log(qCap)^2) +
I(0.5 * log(qLab)^2) + I(0.5 * log(qMat)^2) + I(log(qCap) *
log(qLab)) + I(log(qCap) * log(qMat)) + I(log(qLab) * log(qMat))
Res.Df Df
1

136

130

Pr(>F)

6 2.062 0.06202 .

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

> lrtest( prodCD, prodTL )

97

2 Primal Approach: Production Function

Likelihood ratio test
Model 1: log(qOut) ~ log(qCap) + log(qLab) + log(qMat)
Model 2: log(qOut) ~ log(qCap) + log(qLab) + log(qMat) + I(0.5 * log(qCap)^2) +
I(0.5 * log(qLab)^2) + I(0.5 * log(qMat)^2) + I(log(qCap) *
log(qLab)) + I(log(qCap) * log(qMat)) + I(log(qLab) * log(qMat))
#Df

LogLik Df

5 -137.61

11 -131.25

Chisq Pr(>Chisq)

6 12.727

0.04757 *

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

At the 5% significance level, the Cobb-Douglas production function is accepted by the Wald test
but rejected in favor of the Translog production function by the likelihood ratio test. In order
to reduce the chance of using a too restrictive functional form, we proceed with the Translog
production function.

2.6.3 Properties
We cannot see from the estimated coefficients whether the monotonicity condition is fulfilled. The
Translog production function cannot be globally monotone, because there will be always a set of
input quantities that result in negative marginal products.6 The Translog function would only be
globally monotone, if all first-order coefficients are positive and all second-order coefficients are
zero, which is equivalent to a Cobb-Douglas function. We will check the monotonicity condition
at each observation in section 2.6.5.
All Translog production functions fulfill the weak and the strong essentiality assumption, because as soon as a single input quantity approaches zero, the right-hand side of equation (2.105)
approaches minus infinity (if monotonicity is fulfilled), and thus, the output quantity y = exp(ln y)
approaches zero. Hence, if a data set includes observations with a positive output quantity but
at least one input quantity that is zero, strict essentiality cannot be fulfilled in the underlying
true production technology so that the Translog production function is not a suitable functional
form for analyzing this data set.
The input requirement sets derived from Translog production functions are always closed and
non-empty. The Translog production function always returns finite, real, non-negative, and single
values as long as all input quantities are strictly positive. All Translog production functions are
continuous and twice-continuously differentiable.

Please note that ln xj is a large negative number if xj is a very small positive number.

98

2 Primal Approach: Production Function

2.6.4 Predicted Output Quantities

As before, we can easily obtain the predicted output quantities with the fitted method. As we
used the logarithmic output quantity as dependent variable in our estimated model, we must use
the exponential function to obtain the output quantities measured in levels:
> dat$qOutTL <- exp( fitted( prodTL ) )
Now, we can evaluate the fit of the model by comparing the observed with the fitted output
quantities:
> compPlot( dat$qOut, dat$qOutTL )
> compPlot( dat$qOut, dat$qOutTL, log = "xy" )

1e+07

2.0e+07

0.0e+00

1.0e+07

1e+05

5e+05 2e+06

fitted

1.0e+07
0.0e+00

fitted

2.0e+07

1e+05

observed

5e+05

5e+06

observed

Figure 2.33: Translog production function: fit of the model

The resulting graphs are shown in figure 2.33. While the graph in the left panel uses a linear
scale for the axes, the graph in the right panel uses a logarithmic scale for both axes. Hence,
the deviations from the 45-line illustrate the absolute deviations in the left panel and the relative deviations in the right panel. The fit of the model looks rather okay, but there are some
observations, at which the predicted output quantity is not very close to the observed output
quantity.

2.6.5 Output Elasticities

The output elasticities calculated from a Translog production function are:
i =

X
ln y
= i +
ij ln xj
ln xi
j

(2.106)

We can simplify the code for computing these output elasticities by using short names for the
coefficients:

99

2 Primal Approach: Production Function

> a1 <- coef( prodTL )[ "log(qCap)" ]
> a2 <- coef( prodTL )[ "log(qLab)" ]
> a3 <- coef( prodTL )[ "log(qMat)" ]
> a11 <- coef( prodTL )[ "I(0.5 * log(qCap)^2)" ]
> a22 <- coef( prodTL )[ "I(0.5 * log(qLab)^2)" ]
> a33 <- coef( prodTL )[ "I(0.5 * log(qMat)^2)" ]
> a12 <- a21 <- coef( prodTL )[ "I(log(qCap) * log(qLab))" ]
> a13 <- a31 <- coef( prodTL )[ "I(log(qCap) * log(qMat))" ]
> a23 <- a32 <- coef( prodTL )[ "I(log(qLab) * log(qMat))" ]
Now, we can use the following commands to calculate the output elasticities in R:
> dat$eCapTL <- with( dat,
+

a1 + a11 * log(qCap) + a12 * log(qLab) + a13 * log(qMat) )

> dat$eLabTL <- with( dat,

+

a2 + a21 * log(qCap) + a22 * log(qLab) + a23 * log(qMat) )

> dat$eMatTL <- with( dat,

+

a3 + a31 * log(qCap) + a32 * log(qLab) + a33 * log(qMat) )

We can visualize (the variation of) these output elasticities with histograms:
> hist( dat$eCapTL, 15 )
> hist( dat$eLabTL, 15 )

Frequency

0.4

0.0

0.4

0.8

5
0

10

20

30

25
20
15
10

Frequency

15
10

Frequency

20

25

> hist( dat$eMatTL, 15 )

1.0

0.0 0.5 1.0 1.5 2.0

eCap

0.0

eLab

0.5

1.0

1.5

2.0

eMat

Figure 2.34: Translog production function: output elasticities

The resulting graphs are shown in figure 2.34. If the firms increase capital input by one percent,
the output of most firms will increase by around 0.2 percent. If the firms increase labor input by
one percent, the output of most firms will increase by around 0.5 percent. If the firms increase
material input by one percent, the output of most firms will increase by around 0.7 percent.
These graphs also show that the monotonicity condition is not fulfilled for all observations:

100

2 Primal Approach: Production Function

> sum( dat$eCapTL < 0 )
[1] 32
> sum( dat$eLabTL < 0 )
[1] 14
> sum( dat$eMatTL < 0 )
[1] 8
> dat$monoTL <- with( dat, eCapTL >= 0 & eLabTL >= 0 & eMatTL >= 0 )
> sum( !dat$monoTL )
[1] 48
32 firms have a negative output elasticity of capital, 14 firms have a negative output elasticity
of labor, and 8 firms have a negative output elasticity of materials. In total the monotonicity
condition is not fulfilled at 48 out of 140 observations. Although the monotonicity conditions
are fulfilled for a large part of firms in our data set, these frequent violations indicate a possible
model misspecification.

2.6.6 Marginal Products

The first derivatives (marginal products) of the Translog production function with respect to the
input quantities are:

X
y
y ln y
y
M Pi =
=
=
ij ln xj
i +
xi
xi ln xi
xi
j

(2.107)

We can calculate the marginal products based on the output elasticities that we have calculated
above. As argued in section 2.4.11.1, we use the predicted output quantities in this calculation:
> dat$mpCapTL <- with( dat, eCapTL * qOutTL / qCap )
> dat$mpLabTL <- with( dat, eLabTL * qOutTL / qLab )
> dat$mpMatTL <- with( dat, eMatTL * qOutTL / qMat )
We can visualize (the variation of) these marginal products with histograms:
> hist( dat$mpCapTL, 15 )
> hist( dat$mpLabTL, 15 )
> hist( dat$mpMatTL, 15 )
The resulting graphs are shown in figure 2.35. If the firms increase capital input by one unit,
the output of most firms will increase by around 4 units. If the firms increase labor input by
one unit, the output of most firms will increase by around 4 units. If the firms increase material
input by one unit, the output of most firms will increase by around 70 units.

101

10

10

20

15
0

10

Frequency

20

20
15
0

10

Frequency

15
10
5

Frequency

20

25

2 Primal Approach: Production Function

mpCapTL

10

15

20

25

50

mpLabTL

100

mpMatTL

Figure 2.35: Translog production function: marginal products

2.6.7 Elasticity of Scale

The elasticity of scale canas alwaysbe calculated as the sum of all output elasticities.
> dat$eScaleTL <- dat$eCapTL + dat$eLabTL +
+

dat$eMatTL

The (variation of the) elasticities of scale can be visualized with a histogram.

> hist( dat$eScaleTL, 30 )

8
6
0

Frequency

10
5

Frequency

15

> hist( dat$eScaleTL[ dat$monoTL ], 30 )

1.2 1.3 1.4 1.5 1.6 1.7

1.2

eScaleTL

1.3

1.4

1.5

1.6

1.7

eScaleTL[ monoTL ]

Figure 2.36: Translog production function: elasticities of scale

The resulting graphs are shown in figure 2.36. All firms experience increasing returns to scale
and most of them have an elasticity of scale around 1.45. Hence, if these firms increase all input
quantities by one percent, the output of most firms will increase by around 1.45 percent. These
elasticities of scale are realistic and on average close to the elasticity of scale obtained from the
Cobb-Douglas production function (1.47).

102

2 Primal Approach: Production Function

Information on the optimal firm size can be obtained by analyzing the relationship between
firm size and the elasticity of scale. We can either use the observed or the predicted output:
> plot( dat$qOut, dat$eScaleTL, log = "x" )
> plot( dat$X, dat$eScaleTL, log = "x" )
> plot( dat$qOut[ dat$monoTL ], dat$eScaleTL[ dat$monoTL ], log = "x" )
> plot( dat$X[ dat$monoTL ], dat$eScaleTL[ dat$monoTL ], log = "x" )

1e+05

5e+05

2e+06

1.6
1.4

1.2

eScaleTL

1.4
1.2

eScaleTL

1.6

1e+07

0.5

1e+05

5e+05

2e+06

1e+07

1.2 1.3 1.4 1.5 1.6 1.7

2.0

5.0

quantity index of inputs

eScaleTL[ monoTL ]

1.2 1.3 1.4 1.5 1.6 1.7

eScaleTL[ monoTL ]

observed output

1.0

0.5

observed output

1.0

2.0

5.0

quantity index of inputs

Figure 2.37: Translog production function: elasticities of scale at different firm sizes
The resulting graphs are shown in figure 2.37. Both of them indicate that the elasticity of scale
slightly decreases with firm size but there are considerable increasing returns to scale even for
the largest firms in the sample. Hence, all firms in the sample would gain from increasing their
size and the optimal firm size seems to be larger than the largest firm in the sample.

2.6.8 Marginal Rates of Technical Substitution

We can calculate the marginal rates of technical substitution (MRTS) based on our estimated
Translog production function by following commands:
> dat$mrtsCapLabTL <- with( dat, - mpLabTL / mpCapTL )
> dat$mrtsLabCapTL <- with( dat, - mpCapTL / mpLabTL )
> dat$mrtsCapMatTL <- with( dat, - mpMatTL / mpCapTL )

103

2 Primal Approach: Production Function

> dat$mrtsMatCapTL <- with( dat, - mpCapTL / mpMatTL )
> dat$mrtsLabMatTL <- with( dat, - mpMatTL / mpLabTL )
> dat$mrtsMatLabTL <- with( dat, - mpLabTL / mpMatTL )
As the marginal rates of technical substitution are meaningless if the monotonicity condition is
not fulfilled, we visualize (the variation of) these MRTS only for the observations, where the
monotonicity condition is fulfilled:
> hist( dat$mrtsCapLabTL[ dat$monoTL ], 30 )
> hist( dat$mrtsLabCapTL[ dat$monoTL ], 30 )
> hist( dat$mrtsCapMatTL[ dat$monoTL ], 30 )
> hist( dat$mrtsMatCapTL[ dat$monoTL ], 30 )
> hist( dat$mrtsLabMatTL[ dat$monoTL ], 30 )

150

100

50

10 20 30 40 50 60 70

Frequency

40
30
0

10

20

Frequency

40
20

Frequency

60

50

60

> hist( dat$mrtsMatLabTL[ dat$monoTL ], 30 )

60

20

800

mrtsLabCapTL

600

400

200

mrtsCapMatTL

0.6

0.4

0.2

0.0

30
0

10

20

Frequency

40

60
0

20

40

Frequency

10
5

Frequency

15

50

80

mrtsCapLabTL

40

1200

mrtsMatCapTL

800

400

mrtsLabMatTL

mrtsMatLabTL

Figure 2.38: Translog production function: marginal rates of technical substitution (MRTS)
The resulting graphs are shown in figure 2.39. As some outliers hide the variation of the majority
of the MRTS, we use function colMedians (package miscTools) to show the median values of the
MRTS:
> colMedians( subset( dat, monoTL,
+
+

c( "mrtsCapLabTL", "mrtsLabCapTL", "mrtsCapMatTL",

"mrtsMatCapTL", "mrtsLabMatTL", "mrtsMatLabTL" ) ) )

104

2 Primal Approach: Production Function

mrtsCapLabTL mrtsLabCapTL mrtsCapMatTL mrtsMatCapTL mrtsLabMatTL mrtsMatLabTL
-0.83929283

-1.19196521 -12.72554396

-0.07858435 -12.79850828

-0.07813810

Given that the median marginal rate of technical substitution between capital and labor is -0.84,
a typical firm that reduces the use of labor by one unit, has to use around 0.84 additional units
of capital in order to produce the same amount of output as before. Alternatively, the typical
firm can replace one unit of labor by using 0.08 additional units of materials.

2.6.9 Relative Marginal Rates of Technical Substitution

As we do not have a practical interpretation of the units of measurement of the input quantities,
the relative marginal rates of technical substitution (RMRTS) are practically more meaningful
than the MRTS. The following commands calculate the RMRTS:
> dat$rmrtsCapLabTL <- with( dat, - eLabTL / eCapTL )
> dat$rmrtsLabCapTL <- with( dat, - eCapTL / eLabTL )
> dat$rmrtsCapMatTL <- with( dat, - eMatTL / eCapTL )
> dat$rmrtsMatCapTL <- with( dat, - eCapTL / eMatTL )
> dat$rmrtsLabMatTL <- with( dat, - eMatTL / eLabTL )
> dat$rmrtsMatLabTL <- with( dat, - eLabTL / eMatTL )
As the (relative) marginal rates of technical substitution are meaningless if the monotonicity
condition is not fulfilled, we visualize (the variation of) these RMRTS only for the observations,
where the monotonicity condition is fulfilled:
> hist( dat$rmrtsCapLabTL[ dat$monoTL ], 30 )
> hist( dat$rmrtsLabCapTL[ dat$monoTL ], 30 )
> hist( dat$rmrtsCapMatTL[ dat$monoTL ], 30 )
> hist( dat$rmrtsMatCapTL[ dat$monoTL ], 30 )
> hist( dat$rmrtsLabMatTL[ dat$monoTL ], 30 )
> hist( dat$rmrtsMatLabTL[ dat$monoTL ], 30 )
The resulting graphs are shown in figure 2.39. As some outliers hide the variation of the majority
of the RMRTS, we use function colMedians (package miscTools) to show the median values of
the RMRTS:
> colMedians( subset( dat, monoTL,
+
+

c( "rmrtsCapLabTL", "rmrtsLabCapTL", "rmrtsCapMatTL",

"rmrtsMatCapTL", "rmrtsLabMatTL", "rmrtsMatLabTL" ) ) )

rmrtsCapLabTL rmrtsLabCapTL rmrtsCapMatTL rmrtsMatCapTL rmrtsLabMatTL

-2.8357239

-0.3539150

-3.0064237

rmrtsMatLabTL
-0.7439008

105

-0.3331325

-1.3444115

300

100

25

20

15

10

60

400

rmrtsLabCapTL

300

200

100

rmrtsCapMatTL

40
30
20

Frequency

40
30
20

3.0

2.0

1.0

rmrtsMatCapTL

0.0

10

10

10

Frequency

15

50

50

60

20

rmrtsCapLabTL

40
20
0

0
500

Frequency

Frequency

20

40

Frequency

60
40
0

20

Frequency

60

80

80

2 Primal Approach: Production Function

50 40 30 20 10
rmrtsLabMatTL

35

25

15

rmrtsMatLabTL

Figure 2.39: Translog production function: relative marginal rates of technical substitution
(RMRTS)

106

2 Primal Approach: Production Function

Given that the median relative marginal rate of technical substitution between capital and labor
is -2.84, a typical firm that reduces the use of labor by one percent, has to use around 2.84 percent
more capital in order to produce the same amount of output as before. Alternatively, the typical
firm can replace one percent of labor by using 0.74 percent more materials.

2.6.10 Second partial derivatives

In order to compute the elasticities of substitution, we need obtain the second derivatives of the
Translog function. We can calculate them as derivatives of the first derivatives of the Translog
function:
y
xi
xj

2y
=
xi xj

(i +
=

k ik ln xk )

y
xi

(2.108)

xj

ij y
i +
=
+
xj xi
ij y i +
=
+
xi xj

X
ik ln xk y
ij i +
ik ln xk
xi
xj
k

ik ln xk
xi

j +

X
k

jk ln xk

y
x2i

X
y
ij i +
ik ln xk
xj
k

(2.109)
!

y
x2i

(2.110)
ij y i j y
i y
=
+
ij 2
xi xj
xi xj
xi
y
=
(ij + i j ij i ) ,
xi xj

(2.111)
(2.112)

where ij is (again) Kroneckers delta (2.66). Alternatively, the second derivatives of the Translog
function can be expressed based on the marginal products (instead of the output elasticities):
ij y M Pi M Pj
M Pi
2y
=
+
ij
xi xj
xi xj
y
xi
Now, we can calculate the second derivatives for each observation in our data set:
> dat$fCapCapTL <- with( dat,
+

qOutTL / qCap^2 * ( a11 + eCapTL^2 - eCapTL ) )

> dat$fLabLabTL <- with( dat,

+

qOutTL / qLab^2 * ( a22 + eLabTL^2 - eLabTL ) )

> dat$fMatMatTL <- with( dat,

+

qOutTL / qMat^2 * ( a33 + eMatTL^2 - eMatTL ) )

> dat$fCapLabTL <- with( dat,

+

qOutTL / ( qCap * qLab ) * ( a12 + eCapTL * eLabTL ) )

> dat$fCapMatTL <- with( dat,

+

qOutTL / ( qCap * qMat ) * ( a13 + eCapTL * eMatTL ) )

> dat$fLabMatTL <- with( dat,

107

(2.113)

2 Primal Approach: Production Function

+

qOutTL / ( qLab * qMat ) * ( a23 + eLabTL * eMatTL ) )

2.6.11 Elasticities of Substitution

As for the quadratic production function, we only calculate the Allen elasticities of substitution.
The calculation of the direct elasticities of substitution and the Morishima elasticities of substitution requires only minimal changes of the code. In order to check whether our calculations are
correct, we willas beforecheck if the conditions (2.103) are fulfilled. In order to check these
conditions, we need to calculate not only (normal) elasticities of substitution (ij ; i 6= j) but also
economically not meaningful elasticities of self-substitution (ii ):
> dat$esaCapLabTL <- NA
> dat$esaCapMatTL <- NA
> dat$esaLabMatTL <- NA
> dat$esaCapCapTL <- NA
> dat$esaLabLabTL <- NA
> dat$esaMatMatTL <- NA
> for( obs in 1:nrow( dat ) ) {
+

bhmLoop <- matrix( 0, nrow = 4, ncol = 4 )

bhmLoop[ 1, 2 ] <- bhmLoop[ 2, 1 ] <- dat$mpCapTL[ obs ]

bhmLoop[ 1, 3 ] <- bhmLoop[ 3, 1 ] <- dat$mpLabTL[ obs ]

bhmLoop[ 1, 4 ] <- bhmLoop[ 4, 1 ] <- dat$mpMatTL[ obs ]

bhmLoop[ 2, 2 ] <- dat$fCapCapTL[ obs ]

bhmLoop[ 3, 3 ] <- dat$fLabLabTL[ obs ]

bhmLoop[ 4, 4 ] <- dat$fMatMatTL[ obs ]

bhmLoop[ 2, 3 ] <- bhmLoop[ 3, 2 ] <- dat$fCapLabTL[ obs ]

bhmLoop[ 2, 4 ] <- bhmLoop[ 4, 2 ] <- dat$fCapMatTL[ obs ]

bhmLoop[ 3, 4 ] <- bhmLoop[ 4, 3 ] <- dat$fLabMatTL[ obs ]

FCapLabLoop <- - det( bhmLoop[ -2, -3 ] )

FCapMatLoop <- det( bhmLoop[ -2, -4 ] )

FLabMatLoop <- - det( bhmLoop[ -3, -4 ] )

FCapCapLoop <- det( bhmLoop[ -2, -2 ] )

FLabLabLoop <- det( bhmLoop[ -3, -3 ] )

FMatMatLoop <- det( bhmLoop[ -4, -4 ] )

numerator <- with( dat[ obs, ],

+
+
+
+
+
+

qCap * mpCapTL + qLab * mpLabTL + qMat * mpMatTL )

dat$esaCapLabTL[ obs ] <- with( dat[obs,],
numerator / ( qCap * qLab ) * FCapLabLoop / det( bhmLoop ) )
dat$esaCapMatTL[ obs ] <- with( dat[ obs, ],
numerator / ( qCap * qMat ) * FCapMatLoop / det( bhmLoop ) )
dat$esaLabMatTL[ obs ] <- with( dat[ obs, ],

108

2 Primal Approach: Production Function

+

numerator / ( qLab * qMat ) * FLabMatLoop / det( bhmLoop ) )

dat$esaCapCapTL[ obs ] <- with( dat[obs,],

numerator / ( qCap * qCap ) * FCapCapLoop / det( bhmLoop ) )

dat$esaLabLabTL[ obs ] <- with( dat[ obs, ],

numerator / ( qLab * qLab ) * FLabLabLoop / det( bhmLoop ) )

dat$esaMatMatTL[ obs ] <- with( dat[ obs, ],

numerator / ( qMat * qMat ) * FMatMatLoop / det( bhmLoop ) )

+ }
Before we take a look at and interpret the elasticities of substitution, we check whether the
conditions (2.103) are fulfilled:
> range( with( dat, qCap * mpCapTL * esaCapCapTL +
+

qLab * mpLabTL * esaCapLabTL + qMat * mpMatTL * esaCapMatTL ) )

[1] -5.960464e-08

1.907349e-06

> range( with( dat, qCap * mpCapTL * esaCapLabTL +

+

qLab * mpLabTL * esaLabLabTL + qMat * mpMatTL * esaLabMatTL ) )

[1] -1.490116e-08

5.960464e-08

> range( with( dat, qCap * mpCapTL * esaCapMatTL +

+

qLab * mpLabTL * esaLabMatTL + qMat * mpMatTL * esaMatMatTL ) )

[1] -3.725290e-08

1.311302e-06

The extremely small deviations from zero are most likely caused by rounding errors that are
unavoidable on digital computers. This test does not proof that all of our calculations are done
correctly but if we had made a mistake, we probably would have discovered it. Hence, we can be
rather sure that our calculations are correct.
As the elasticities of substitution measure changes in the marginal rates of technical substitution
(MRTS) and the MRTS are meaningless if the monotonicity conditions are not fulfilled, also the
elasticities of substitution are meaningless if the monotonicity conditions are not fulfilled. Hence,
we visualize (the variation of) the Allen elasticities of substitution only for the observations,
where the monotonicity condition is fulfilled:
> hist( dat$esaCapLabTL[ dat$monoTL ], 30 )
> hist( dat$esaCapMatTL[ dat$monoTL ], 30 )
> hist( dat$esaLabMatTL[ dat$monoTL ], 30 )
> hist( dat$esaCapLabTL[ dat$monoTL & abs( dat$esaCapLabTL ) < 10 ], 30 )
> hist( dat$esaCapMatTL[ dat$monoTL & abs( dat$esaCapMatTL ) < 10 ], 30 )
> hist( dat$esaLabMatTL[ dat$monoTL & abs( dat$esaLabMatTL ) < 10 ], 30 )

109

200

500

1000

1500

50

150

esaLabMatTL

10

Frequency

15

20

12
10
8

10

Frequency

10
5

100

25

esaCapMatTL

15

esaCapLabTL

Frequency

30
10
0

10
0

0
400

20

Frequency

30
20

Frequency

30
20
10

Frequency

40

40

40

50

50

50

2 Primal Approach: Production Function

10

abs( esaCapLabTL ) < 10

abs( esaCapMatTL ) < 10

abs( esaLabMatTL ) < 10

Figure 2.40: Translog production function: elasticities of substitution

The resulting graphs are shown in figure 2.40. The estimated elasticities of substitution between
capital and labor suggest that capital and labor are substitutes for almost half of the firms but
complements for the majority of firms. In contrast, capital and materials as well as labor and
materials are substitutes for the majority of firms. As some outliers hide the variation of the
majority of the elasticities of substitution, we use function colMedians (package miscTools) to
obtain the median values of the Allen elasticities of substitution:
> colMedians( subset( dat, monoTL,
+

c( "esaCapLabTL", "esaCapMatTL", "esaLabMatTL" ) ) )

esaCapLabTL esaCapMatTL esaLabMatTL

-0.2130532

2.5436068

0.4193423

The median elasticity of substitution between labor and materials (0.42) lies between the elasticity
of substitution of the Leontief production function ( = 0) and the elasticity of substitution of
the Cobb-Douglas production function ( = 1). Hence, the substitutability between labor and
materials seems to be rather low. A typical firm who substitutes materials for labor (or vice versa)
so that the MRTS between materials and labor increases (decreases) by one percent, has increased
(decreased) the ratio between the quantity of materials and the labor quantity by 0.42 percent. If
the firm is maximizing profit or minimizing costs and the price ratio between labor and materials

110

2 Primal Approach: Production Function

increases by one percent, the firm will substitute materials for labor so that the ratio between
the quantity of materials and the labor quantity increases by 0.42 percent. Hence, the relative
change of the quantity ratio is smaller than the relative change of price ratio, which indicates a low
substitutability between labor and materials. In contrast, the median elasticity of substitution
between capital and materials is larger than one (2.54), which indicates that it is much easier to
substitute between capital and materials.

2.6.12 Quasiconcavity
We check whether our estimated Translog production function is quasiconcave at each observation:
> dat$quasiConcTL <- NA
> for( obs in 1:nrow( dat ) ) {
+

bhmLoop <- matrix( 0, nrow = 4, ncol = 4 )

bhmLoop[ 1, 2 ] <- bhmLoop[ 2, 1 ] <- dat$mpCapTL[ obs ]

bhmLoop[ 1, 3 ] <- bhmLoop[ 3, 1 ] <- dat$mpLabTL[ obs ]

bhmLoop[ 1, 4 ] <- bhmLoop[ 4, 1 ] <- dat$mpMatTL[ obs ]

bhmLoop[ 2, 2 ] <- dat$fCapCapTL[ obs ]

bhmLoop[ 3, 3 ] <- dat$fLabLabTL[ obs ]

bhmLoop[ 4, 4 ] <- dat$fMatMatTL[ obs ]

bhmLoop[ 2, 3 ] <- bhmLoop[ 3, 2 ] <- dat$fCapLabTL[ obs ]

bhmLoop[ 2, 4 ] <- bhmLoop[ 4, 2 ] <- dat$fCapMatTL[ obs ]

bhmLoop[ 3, 4 ] <- bhmLoop[ 4, 3 ] <- dat$fLabMatTL[ obs ]

dat$quasiConcTL[ obs ] <- det( bhmLoop[ 1:2, 1:2 ] ) < 0 &

det( bhmLoop[ 1:3, 1:3 ] ) > 0 & det( bhmLoop ) < 0

+ }
> sum( dat$quasiConcTL )
[1] 63
Our estimated Translog production function is quasiconcave at 63 of the 140 observations.

2.6.13 First-order conditions for profit maximisation

In this section, we will check to what extent the first-order conditions for profit maximisation
(2.24) are fulfilled, i.e. to what extent the firms use the optimal input quantities. We do this by
comparing the marginal value products of the inputs with the corresponding input prices. We
can calculate the marginal value products by multiplying the marginal products by the output
price:

111

2 Primal Approach: Production Function

> dat$mvpCapTL <- dat$pOut * dat$mpCapTL
> dat$mvpLabTL <- dat$pOut * dat$mpLabTL
> dat$mvpMatTL <- dat$pOut * dat$mpMatTL
The command compPlot (package miscTools) can be used to compare the marginal value products
with the corresponding input prices. As the logarithm of a non-positive number is not defined,
we have to limit the comparisons on the logarithmic scale to observations with positve marginal
products:
> compPlot( dat$pCap, dat$mvpCapTL )
> compPlot( dat$pLab, dat$mvpLabTL )
> compPlot( dat$pMat, dat$mvpMatTL )
> compPlot( dat$pCap[ dat$monoTL ], dat$mvpCapTL[ dat$monoTL ], log = "xy" )
> compPlot( dat$pLab[ dat$monoTL ], dat$mvpLabTL[ dat$monoTL ], log = "xy" )

30

> compPlot( dat$pMat[ dat$monoTL ], dat$mvpMatTL[ dat$monoTL ], log = "xy" )

60

150
10 15 20 25 30

100

5.00

100

150

5e02

0.20

MVP Lab

5e+00
5e01

MVP Cap

0.05

5e01

50

w Mat
200

20.00

5e02

w Lab

1.00

5e+01

w Cap

100

MVP Mat
5

50

40

20

20

10

MVP Mat

40

5
5

40

20

15

MVP Lab

40
20

50

25
20

MVP Cap

10

60

5e+00

5e+01

0.05

0.20

w Cap

1.00

5.00 20.00

10

w Lab

20

50

100

w Mat

Figure 2.41: Marginal value products and corresponding input prices

The resulting graphs are shown in figure 2.41. They indicate that the marginal value products of
most firms are higher than the corresponding input prices. This indicates that most firms could

112

2 Primal Approach: Production Function

increase their profit by using more of all inputs. Given that the estimated Translog function
shows that all firms operate under increasing returns to scale, it is not surprising that most firms
would gain from increasing all input quantities. Therefore, the question arises why the firms in
the sample did not do this. This questions has already been addressed in section 2.3.10.

2.6.14 First-order conditions for cost minimization

As the marginal rates of technical substitution differ between observations for the three other
functional forms, we use scatter plots for visualizing the comparison of the input price ratios
with the negative inverse marginal rates of technical substitution: As the marginal rates of
technical substitution are meaningless if the monotonicity condition is not fulfilled, we limit the
comparisons to the observations, where all monotonicity conditions are fulfilled:
> compPlot( ( dat$pCap / dat$pLab )[ dat$monoTL ],
+

- dat$mrtsLabCapTL[ dat$monoTL ] )

> compPlot( ( dat$pCap / dat$pMat )[ dat$monoTL ],

+

- dat$mrtsMatCapTL[ dat$monoTL ] )

> compPlot( ( dat$pLab / dat$pMat )[ dat$monoTL ],

+

- dat$mrtsMatLabTL[ dat$monoTL ] )

> compPlot( ( dat$pCap / dat$pLab )[ dat$monoTL ],

+

- dat$mrtsLabCapTL[ dat$monoTL ], log = "xy" )

> compPlot( ( dat$pCap / dat$pMat )[ dat$monoTL ],

+

- dat$mrtsMatCapTL[ dat$monoTL ], log = "xy" )

> compPlot( ( dat$pLab / dat$pMat )[ dat$monoTL ],

+

- dat$mrtsMatLabTL[ dat$monoTL ], log = "xy" )

The resulting graphs are shown in figure 2.42.

Furthermore, we use histograms to visualize the (absolute and relative) differences between the
input price ratios and the corresponding negative inverse marginal rates of technical substitution:
> hist( ( - dat$mrtsLabCapTL - dat$pCap / dat$pLab )[ dat$monoTL ] )
> hist( ( - dat$mrtsMatCapTL - dat$pCap / dat$pMat )[ dat$monoTL ] )
> hist( ( - dat$mrtsMatLabTL - dat$pLab / dat$pMat )[ dat$monoTL ] )
> hist( log( - dat$mrtsLabCapTL / ( dat$pCap / dat$pLab ) )[ dat$monoTL ] )
> hist( log( - dat$mrtsMatCapTL / ( dat$pCap / dat$pMat ) )[ dat$monoTL ] )
> hist( log( - dat$mrtsMatLabTL / ( dat$pLab / dat$pMat ) )[ dat$monoTL ] )
The resulting graphs are shown in figure 2.43. The graphs in the middle column of figures 2.42
and 2.43 show that the ratio between the capital price and the materials price is larger than the
absolute value of the marginal rate of technical substitution between materials and capital for a
majority of the firms in the sample:
wcap
M Pcap
> M RT Smat,cap =
wmat
M Pmat

113

(2.114)

0.6
60

1
0

1.000

0.100

0.001

0.001
1e+02

0.010

0.500
0.100
0.020

MRTS Mat Cap

w Lab / w Mat

0.005

1e+01
1e02

1e01

1e+00

w Cap / w Lab

w Cap / w Mat

1e+00

0.0 0.1 0.2 0.3 0.4 0.5 0.6

1e02

w Cap / w Lab

1e+02

0.4
0.3
0.2
0.1

40

0.0
20

MRTS Mat Lab

MRTS Lab Cap

MRTS Mat Cap

60
40
20
0

MRTS Lab Cap

0.5

MRTS Mat Lab

2 Primal Approach: Production Function

0.001

0.005

0.050

0.500

0.001

w Cap / w Mat

Figure 2.42: First-order conditions for costs minimization

114

0.010

0.100

w Lab / w Mat

1.000

60

80

10
0.6

0.2

0.4

0.6

MrtsMatCap wCap / wMat

MrtsMatLab wLab / wMat

20
15
10

Frequency

10

20

Frequency

15
10

log(MrtsLabCap / (wCap / wLab))

5
0

Frequency

20

30

25

25

40

MrtsLabCap wCap / wLab

0.2

30

40

5
0
20

40
20

10
10
0

30

Frequency

20
15

Frequency

30
20

Frequency

25

50

40

30

60

2 Primal Approach: Production Function

log(MrtsMatCap / (wCap / wMat))

log(MrtsMatLab / (wLab / wMat))

Figure 2.43: First-order conditions for costs minimization

115

2 Primal Approach: Production Function

Hence, these firms can get closer to the minimum of their production costs by substituting
materials for capital, because this will decrease the marginal product of materials and increase
the marginal product of capital so that the absolute value of the MRTS between materials and
capital increases and gets closer to the corresponding input price ratio. The graphs on the left
indicate that approximately half of the firms should substitute labor for capital, while the other
half should substitute capital for labor. The graphs on the right indicate that a majority of
the firms should substitute materials for labor. Hence, the majority of the firms could reduce
production costs particularly by using more materials and using less labor or less capital but
there might be (legal) regulations that restrict the use of materials (e.g. fertilizers, pesticides).

2.6.15 Mean-scaled quantities

The Translog function is often estimated with mean-scaled variables. The following commands
create variables with mean-scaled output and input quantities:
> dat$qmOut <- with( dat, qOut / mean( qOut ) )
> dat$qmCap <- with( dat, qCap / mean( qCap ) )
> dat$qmLab <- with( dat, qLab / mean( qLab ) )
> dat$qmMat <- with( dat, qMat / mean( qMat ) )
This implies that the logarithms of the mean values of these variables are zero (except for negligible very small rounding errors):
> log( colMeans( dat[ , c( "qmOut", "qmCap", "qmLab", "qmMat" ) ] ) )
qmOut

qmCap

qmLab

qmMat

-1.110223e-16 -1.110223e-16

0.000000e+00

0.000000e+00

Please note that mean-scaling does not imply that the mean values of the logarithmic variables
are zero:
> colMeans( log( dat[ , c( "qmOut", "qmCap", "qmLab", "qmMat" ) ] ) )
qmOut

qmCap

qmLab

qmMat

-0.4860021 -0.3212057 -0.1565112 -0.2128551

Now, we estimate the Translog production function with mean-scaled variables:
> prodTLm <- lm( log( qmOut ) ~ log( qmCap ) + log( qmLab ) + log( qmMat )
+

+ I( 0.5 * log( qmCap )^2 ) + I( 0.5 * log( qmLab )^2 )

+ I( 0.5 * log( qmMat )^2 ) + I( log( qmCap ) * log( qmLab ) )

+ I( log( qmCap ) * log( qmMat ) ) + I( log( qmLab ) * log( qmMat ) ),

data = dat )

> summary( prodTLm )

116

2 Primal Approach: Production Function

Call:
lm(formula = log(qmOut) ~ log(qmCap) + log(qmLab) + log(qmMat) +
I(0.5 * log(qmCap)^2) + I(0.5 * log(qmLab)^2) + I(0.5 * log(qmMat)^2) +
I(log(qmCap) * log(qmLab)) + I(log(qmCap) * log(qmMat)) +
I(log(qmLab) * log(qmMat)), data = dat)
Residuals:
Min

1Q

Median

3Q

Max

-1.68015 -0.36688

0.05389

0.44125

1.26560

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept)

-0.09392

0.08815

-1.065

0.28864

log(qmCap)

0.15004

0.11134

1.348

0.18013

log(qmLab)

0.79339

0.17477

4.540 1.27e-05 ***

log(qmMat)

0.50201

0.16608

3.023

I(0.5 * log(qmCap)^2)

-0.02573

0.20834

-0.124

0.90189

I(0.5 * log(qmLab)^2)

-1.16364

0.67943

-1.713

0.08916 .

I(0.5 * log(qmMat)^2)

-0.50368

0.43498

-1.158

0.24902

0.56194

0.29120

1.930

0.05582 .

I(log(qmCap) * log(qmMat)) -0.40996

0.23534

-1.742

0.08387 .

I(log(qmLab) * log(qmMat))

0.42750

1.539

I(log(qmCap) * log(qmLab))

0.65793

0.00302 **

0.12623

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Residual standard error: 0.6412 on 130 degrees of freedom

Multiple R-squared:

0.6296,

F-statistic: 24.55 on 9 and 130 DF,

Adjusted R-squared:

0.6039

p-value: < 2.2e-16

While the intercept and the first-order coefficients have adjusted to the new units of measurement,
the second-order coefficients of the Translog function remain unchanged (compare with estimates
in section 2.6.2):
> all.equal( coef(prodTL)[-c(1:4)], coef(prodTLm)[-c(1:4)],
+

check.attributes = FALSE )

[1] TRUE
In case of functional forms that are invariant to the units of measurement (e.g. linear, CobbDouglas, quadratic, Translog), mean-scaling does not change the relative indicators of the technology (e.g. output elasticities, elasticities of scale, relative marginal rates of technical substitution, elasticities of substitution). As the logarithms of the mean values of the mean-scaled input

117

2 Primal Approach: Production Function

quantities are zero, the first-order coefficients are equal to the output elasticities at the sample
mean (see equation 2.106), i.e. the output elasticity of capital is 0.15, the output elasticity of
labor is 0.793, the output elasticity of materials is 0.502, and the elasticity of scale is 1.445 at
the sample mean.

2.7 Evaluation of Different Functional Forms

In this section, we will discuss the appropriateness of the four different functional forms for
analyzing the production technology in our data set. If one functional form is nested in another
functional form, we can use standard statistical tests to compare these functional forms. We have
done this already in section 2.5 (linear production function vs. quadratic production function)
and in section 2.6 (Cobb-Douglas production function vs. Translog production function). The
tests clearly reject the linear production function in favor of the quadratic production function
but it is less clear whether the Cobb-Douglas production function is rejected in favor of the
Translog production function.
It is much less straight-forward to compare non-nested models such as the quadratic and the
Translog production function.

2.7.1 Goodness of Fit

As the quadratic and the Translog models use different dependent variables (y vs. ln y), we cannot
simply compare the R2 -values. However, we can calculate the hypothetical R2 -value regarding y
for the Translog production function and compare it with the R2 value of the quadratic production function. We can also calculate the hypothetical R2 -value regarding ln y for the quadratic
production function and compare it with the R2 value of the Translog production function. We
can calculate the (hypothetical) R2 values with function rSquared (package miscTools). The first
argument of this function must be a vector of the observed dependent variable and the second argument must be a vector of the residuals. We start by extracting the R2 value from the quadratic
model and calculate the hypothetical R2 -value regarding y for the Translog production function:
> summary(prodQuad)$r.squared
[1] 0.8448983
> rSquared( dat$qOut, dat$qOut - dat$qOutTL )
[,1]
[1,] 0.7696638
In this case, the R2 value regarding y is considerably higher for the quadratic function. Similarly,
we can extract the R2 value from the Translog model and calculate the hypothetical R2 -value
regarding ln y for the quadratic production function:

118

2 Primal Approach: Production Function

> summary(prodTL)$r.squared
[1] 0.6295696
> rSquared( log( dat$qOut ), log( dat$qOut ) - log( dat$qOutQuad ) )
[,1]
[1,] 0.5481309
In contrast to the R2 value regarding y, the R2 value regarding ln y is considerably higher for
the Translog function. Hence, in our case, the R2 values do not help much to select the most
suitable functional form. We could base our comparison on the unadjusted R2 values, because
the quadratic and the Translog function have the same number of coefficients. If the compared
models have different numbers of coefficients, the comparison must be based on adjusted R2
values.
Furthermore, we can visually compare the fit of the two models by looking at figures 2.22
and 2.33. The quadratic production function is clearly over-predicting the output of small firms
so that small firms have rather large relative error terms. On the other hand, the Translog
production function has rather large absolute error terms for large firms. In total, it seems that
the fit of the Translog function is slightly better.

2.7.2 Theoretical Consistency

Furthermore, we can compare the theoretical consistency of the two models. The total number
of monotonicity violations of the quadratic production function and the Translog production
function can be obtained by
> with( dat, sum( eCapQuad < 0 ) + sum( eLabQuad < 0 ) + sum( eMatQuad < 0 ) )
[1] 41
> with( dat, sum( eCapTL < 0 ) + sum( eLabTL < 0 ) + sum( eMatTL < 0 ) )
[1] 54
Alternatively, we could look at the number of observations, at which the monotonicity condition
is violated:
> sum( !dat$monoQuad )
[1] 39
> sum( !dat$monoTL )
[1] 48

119

2 Primal Approach: Production Function

Both measures show that the monotonicity condition is more often violated in the Translog
function.
While the Translog production function always returns a positive output quantity (as long
as all input quantities are strictly positive), this is not necessarily the case for the quadratic
production function. However, we have checked this in section 2.5.6 and found that all output
quantities predicted by our quadratic production function are positive. Hence, the non-negativity
condition is fulfilled for both functional forms.
Quasiconcavity is fulfilled at 63 out of 140 observations for the Translog production function
but at no observation for the quadratic production function. However, quasiconcavity is mainly
assumed to simplify the (further) economic analysis (e.g. to obtain continuous input demand
and output supply functions) and there can be found good reasons for why the true production
technology is not quasiconcave (e.g. indivisibility of inputs).

2.7.3 Plausible Estimates

While the elasticities of scale of some observations were implausibly large when estimated with
the linear production function, no elasticities of scale estimated by the quadratic and Translog
production function are in the implausible range:
> sum( dat$eScaleQuad > 2 | dat$eScaleQuad < 0.5 )
[1] 0
> sum( dat$eScaleTL > 2 | dat$eScaleTL < 0.5 )
[1] 0
However, some of the output elasticities are implausibly large:
> with( dat, sum( eCapQuad > 1 ) + sum( eLabQuad > 1 ) + sum( eMatQuad > 1 ) )
[1] 28
> with( dat, sum( eCapTL > 1 ) + sum( eLabTL > 1 ) + sum( eMatTL > 1 ) )
[1] 56
The Translog production function results in more implausible output elasticities than the quadratic
production function.
Regarding the elasticities of substitution, it seems to be rather implausible that capital and
labor are always complements as estimated with the quadratic production function.

120

2 Primal Approach: Production Function

2.7.4 Summary
The various criteria for assessing whether the quadratic or the Translog functional form is more
appropriate for analyzing the production technology in our data set are summarized in table 2.2.
While the quadratic production function results in less monotonicity violations and less implausible output elasticities, the Translog production function seems to give a better fit to the data
and results in slightly more plausible elasticities of substitution.
Table 2.2: Criteria for assessing functional forms
quadratic Translog
R2 of y
0.84
0.77
2
R of ln y
0.55
0.63
visual fit
()
ok
total monotonicity violations
41
54
observations with monotonicity violated
39
48
negative output quantities
0
0
observations with quasiconcavity violated
140
77
implausible elasticities of scale
0
0
implausible output elasticities
28
56
implausible elasticities of substitution
cap,lab

2.8 Non-parametric production function

In order to avoid the specification of a functional form of the production function, the production
technology can be analyzed by nonparametric regression. We will use a local-linear kernel regressor with an Epanechnikov kernel for the (continuous) regressors (see, e.g. Li and Racine, 2007;
Racine, 2008). One can think of this estimator as a set of weighted linear regressions, where a
weighted linear regression is performed at each observation and the weights of the other observations decrease with the distance from the respective observation. The weights are determined
by a kernel function and a set of bandwidths, where a bandwidth for each explanatory variable
must be specified. The smaller the bandwidth, the faster the weight decreases with the distance
from the respective observation. In our study, we make the frequently used assumption that the
bandwidths can differ between regressors but are constant over the domain of each regressor.
While the bandwidths were initially determined by using a rule of thumb, nowadays increased
computing power allows us to select the optimal bandwidths for a given model and data set according to the expected Kullback-Leibler cross-validation criterion (Hurvich, Simonoff, and Tsai,
1998). Hence, in nonparametric kernel regression, the overall shape of the relationship between
the inputs and the output is determined by the data and the (marginal) effects of the explanatory variables can differ between observations without being restricted by an arbitrarily chosen
functional form. (Czekaj and Henningsen, 2012). Given that the distributions of the output
quantity and the input quantities are strongly right-skewed in our data set (many firms with

121

2 Primal Approach: Production Function

small quantities, only a few firms with large quantities), we use the logarithms of the output and
input quantities in order to achieve more uniform distributions, which are preferable in case of
fixed bandwidths. Furthermore, this allows us to interpret the gradients of the dependent variable (logarithmic output quantity) with respect to the explanatory variables (logarithmic input
quantities) as output elasticities. The following commands load the R package np (Hayfield and
Racine, 2008), select the optimal bandwidths and estimate the model, and show summary results:
> library( "np" )
> prodNP <- npreg( log(qOut) ~ log(qCap) + log(qLab) + log(qMat), regtype = "ll",
+

bwmethod = "cv.aic", ckertype = "epanechnikov",

gradients = TRUE )

data = dat,

> summary( prodNP )

Regression Data: 140 training points, in 3 variable(s)
log(qCap) log(qLab) log(qMat)
Bandwidth(s): 0.9898614

641971.3 0.8432296

Kernel Regression Estimator: Local-Linear

Bandwidth Type: Fixed
Residual standard error: 0.6217269
R-squared: 0.6249665
Continuous Kernel Type: Second-Order Epanechnikov
No. Continuous Explanatory Vars.: 3
While the bandwidths of the logarithmic quantities of capital and materials are around one, the
bandwidth of the logarithmic labor quantity is rather large. These bandwidths indicate that the
logarithmic output quantity non-linearly changes with the logarithmic quantities of capital and
materials but it changes approximately linearly with the logarithmic labor quantity.
The estimated relationship between each explanatory variable and the dependent variable
(holding all other explanatory variables constant at their median values) can be visualized using
the plot method. We can use argument plot.errors.method to add confidence intervals:
> plot( prodNP, plot.errors.method = "bootstrap" )
The resulting graphs are shown in figure 2.44.
The estimated gradients of the dependent variable with respect to each explanatory variable
(holding all other explanatory variables constant at their median values) can be visualized using
the plot method with argument gradient set to TRUE:
> plot( prodNP, gradients = TRUE, plot.errors.method = "bootstrap" )

122

16.0
13.0

14.5

log(qOut)

16.0
14.5
13.0

log(qOut)

2 Primal Approach: Production Function

10

11

12

13

11.5

12.0

13.0

13.5

14.0

14.5

16.0

log(qLab)

13.0

log(qOut)

log(qCap)

12.5

9.0

9.5

10.0 10.5 11.0 11.5

log(qMat)

10

11

12

13

0.5
11.5

12.0

12.5

13.0

13.5

14.0

log(qLab)

0.5

log(qCap)

0.5

Gradient Component 2 of log(qOut)

0.5
0.5

0.5

Gradient Component 3 of log(qOut)

Gradient Component 1 of log(qOut)

Figure 2.44: Production technology estimated by non-parametric kernel regression

9.0

9.5

10.0 10.5 11.0 11.5

log(qMat)

Figure 2.45: Gradients (output elasticities) estimated by non-parametric kernel regression

123

2 Primal Approach: Production Function

The resulting graphs are shown in figure 2.45.
Function npsigtest can be used to obtain the statistical significance of the explanatory variables:
> npsigtest( prodNP )
Kernel Regression Significance Test
Type I Test with IID Bootstrap (399 replications)
Explanatory variables tested for significance:
log(qCap) (1), log(qLab) (2), log(qMat) (3)
log(qCap) log(qLab) log(qMat)
Bandwidth(s): 0.9898614

641971.3 0.8432296

Significance Tests
P Value:
log(qCap) 0.1478697
log(qLab) 0.0025063 **
log(qMat) 0.0025063 **
--Signif. codes:

0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

The results confirm the results from the parametric regressions that labor and materials have a
significant effect on the output while capital does not have a significant effect (at 10% significance
level).
The following commands plot histograms of the three output elasticities and the elasticity of
scale:
> hist( gradients( prodNP )[ ,1] )
> hist( gradients( prodNP )[ ,2] )
> hist( gradients( prodNP )[ ,3] )
> hist( rowSums( gradients( prodNP ) ) )
The resulting graphs are shown in figure 2.46. The monotonicity condition is fulfilled at almost
all observations, only 1 output elasticity of capital and 1 output elasticity of labor is negative.
All firms operate under increasing returns to scale with most farms having an elasticity of scale
around 1.4.
Finally, we visualize the relationship between firm size and the elasticity of scale based on our
non-parametric estimation results:
> plot( dat$qOut, rowSums( gradients( prodNP ) ), log = "x" )
> plot( dat$X, rowSums( gradients( prodNP ) ), log = "x" )

124

40
0

20

Frequency

40
20
0

Frequency

2 Primal Approach: Production Function

0.2

0.1

0.0

0.1

0.2

0.3

0.0

0.2

0.4

0.8

1.0

40
0

20

30

Frequency

labor

0 10

Frequency

capital

0.6

0.4

0.6

0.8

1.0

1.2

1.4

1.2

1.4

materials

1.6

1.8

scale

Figure 2.46: Output elasticities and elasticities of scale estimated by non-parametric kernel
regression

1.8

1e+05

5e+05

5e+06

1.2

elaScaleNP

1.6
1.4

1.6

1.8

1.2

elaScaleNP

1.4

0.5

qOut

1.0

2.0

5.0

Figure 2.47: Relationship between firm size and elasticities of scale estimated by non-parametric
kernel regression

125

2 Primal Approach: Production Function

The resulting graph is shown in figure 2.47. The smallest firms generally would gain most from
increasing their size. However, also the largest firms would still considerably gain from increasing
their sizeperhaps even more than medium-sized firms but there is probably insufficient evidence
to be sure about this.

126

3 Dual Approach: Cost Functions

3.1 Theory
3.1.1 Cost function
Total cost is defined as:
c=

wi x i

(3.1)

The cost function:

c(w, y) = min

wi xi , s.t. f (x) y

(3.2)

returns the minimal (total) cost that is required to produce at least the output quantity y given
input prices w.
It is important to distinguish the cost definition (3.1) from the cost function (3.2).

3.1.2 Cost flexibility and elasticity of size

The ratio between the relative change in total costs and the relative change in the output quantity
is called cost flexibility:
c(w, y)
y
y
c(w, y)

(3.3)

The inverse of the cost flexibility is called elasticity of size:

 (w, y) =

c(w, y)
y
c(w, y)
y

(3.4)

At the cost-minimizing points, the elasticity of size is equal to the elasticity of scale (Chambers,
1988, p. 7172). For homothetic production technologies such as the Cobb-Douglas production
technology, the elasticity of size is always equal to the elasticity of scale (Chambers, 1988, p. 72
74).1

3.1.3 Short-Run Cost Functions

As producers often cannot instantly adjust the quantity of the some inputs (e.g. buildings, land,
apple trees), estimating a short-run cost function with some quasi-fixed input quantities might
1

Further details about the relationship between the elasticity of size and the elasticity of scale are available, e.g.,
in McClelland, Wetzstein, and Musserwetz (1986).

127

3 Dual Approach: Cost Functions

be more appropriate than estimating a (long-run) cost function which assumes that all input
quantities quantities can be adjusted instantly.
In general, a short-run cost function is defined as
X

cv (w1 y, , x2 ) = min
x1

wi xi , s.t. f (x1 , x2 ) y

(3.5)

iN 1

where w1 denotes the vector of the prices of all variable inputs, x2 denotes the vector of the
quantities of all quasi-fixed inputs, cv denotes the variable costs defined in equation (1.3), and
N 1 is a vector of the indices of the variable inputs.

3.2 Cobb-Douglas Cost Function

3.2.1 Specification
We start with estimating a Cobb-Douglas cost function. It has the following specification:
!
Y
wi i
c=A

y y

(3.6)

i ln wi + y ln y

(3.7)

This function can be linearized to

ln c = 0 +

X
i

with 0 = ln A.

3.2.2 Estimation
The linearized Cobb-Douglas cost function can be estimated by OLS:
> costCD <- lm( log( cost ) ~ log( pCap ) + log( pLab ) + log( pMat ) + log( qOut ),
+

data = dat )

> summary( costCD )

Call:
lm(formula = log(cost) ~ log(pCap) + log(pLab) + log(pMat) +
log(qOut), data = dat)
Residuals:
Min

1Q

Median

3Q

Max

-0.77663 -0.23243 -0.00031

0.24439

0.74339

Coefficients:

128

3 Dual Approach: Cost Functions

Estimate Std. Error t value Pr(>|t|)
(Intercept)

6.75383

0.40673

16.605

< 2e-16 ***

log(pCap)

0.07437

0.04878

1.525

0.12969

log(pLab)

0.46486

0.14694

3.164

0.00193 **

log(pMat)

0.48642

0.08112

5.996 1.74e-08 ***

log(qOut)

0.37341

0.03072

12.154

< 2e-16 ***

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Residual standard error: 0.3395 on 135 degrees of freedom

Multiple R-squared:

0.6884,

Adjusted R-squared:

F-statistic: 74.56 on 4 and 135 DF,

0.6792

p-value: < 2.2e-16

3.2.3 Properties
As the coefficients of the (logarithmic) input prices are all non-negative, this cost function is
monotonically non-decreasing in input prices. Furthermore, the coefficient of the (logarithmic)
output quantity is non-negative so that this cost function is monotonically non-decreasing in
output quantities. The Cobb-Douglas cost function always implies no fixed costs, as the costs
are always zero if the output quantity is zero. Given that A = exp(0 ) is always positive,
all Cobb-Douglas cost functions that are based on its (estimated) linearized version fulfill the
non-negativity condition.
Finally, we check if the Cobb-Douglas cost function is positive linearly homogeneous in input
prices. This condition is fulfilled if
t c(w, y) = c(t w, y)
ln(t c) = 0 +

(3.8)
i ln(t wi ) + y ln y

(3.9)

(3.10)

ln t + ln c = 0 +

i ln t +

i ln wi + y ln y

ln c + ln t = 0 + ln t

i +

ln c + ln t = ln c + ln t

i ln wi + y ln y

(3.11)

(3.12)

ln t = ln t

(3.13)

1=

(3.14)

Hence, the homogeneity condition is only fulfilled if the coefficients of the (logarithmic) input
prices sum up to one. As they sum up to 1.03 the homogeneity condition is not fulfilled in our
estimated model.

129

3 Dual Approach: Cost Functions

3.2.4 Estimation with linear homogeneity in input prices imposed

In order to estimate a Cobb-Douglas cost function with linear homogeneity imposed, we re-arrange
the homogeneity condition to get
N = 1

N
1
X

(3.15)

i=1

and replace N in the cost function (3.7) by the right-hand side of the above equation:
ln c = 0 +
ln c = 0 +
ln c ln wN = 0 +
ln

c
= 0 +
wN

N
1
X
i=1
N
1
X
i=1
N
1
X
i=1
N
1
X

i ln wi + 1

N
1
X

i ln wN + y ln y

(3.16)

i=1

i (ln wi ln wN ) + ln wN + y ln y

(3.17)

i (ln wi ln wN ) + y ln y

(3.18)

i ln

i=1

wi
+ y ln y
wN

(3.19)

This Cobb-Douglas cost function with linear homogeneity in input prices imposed can be estimated by following command:
> costCDHom <- lm( log( cost / pMat ) ~ log( pCap / pMat ) + log( pLab / pMat ) +
+

log( qOut ), data = dat )

> summary( costCDHom )

Call:
lm(formula = log(cost/pMat) ~ log(pCap/pMat) + log(pLab/pMat) +
log(qOut), data = dat)
Residuals:
Min

1Q

Median

3Q

Max

-0.77096 -0.23022 -0.00154

0.24470

0.74688

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept)

6.75288

0.40522

16.665

log(pCap/pMat)

0.07241

0.04683

1.546

log(pLab/pMat)

0.44642

0.07949

5.616 1.06e-07 ***

log(qOut)

0.37415

0.03021

12.384

< 2e-16 ***

0.124
< 2e-16 ***

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

130

3 Dual Approach: Cost Functions

Residual standard error: 0.3383 on 136 degrees of freedom
Multiple R-squared:

0.5456,

Adjusted R-squared:

F-statistic: 54.42 on 3 and 136 DF,

0.5355

p-value: < 2.2e-16

The coefficient of the N th (logarithmic) input price can be obtained by the homogeneity condition
(3.15). Hence, the estimate of Mat is 0.4812 in our model.
As there is no theory that says which input price should be taken for the normalization/deflation,
it is desirable that the estimation results do not depend on the price that is used for the normalization/deflation. This desirable property is fulfilled for the Cobb-Douglas cost function and
we can verify this by re-estimating the cost function, while using a different input price for the
normalization/deflation, e.g. capital:
> costCDHomCap <- lm( log( cost / pCap ) ~ log( pLab / pCap ) + log( pMat / pCap ) +
+

log( qOut ), data = dat )

> summary( costCDHomCap )

Call:
lm(formula = log(cost/pCap) ~ log(pLab/pCap) + log(pMat/pCap) +
log(qOut), data = dat)
Residuals:
Min

1Q

Median

3Q

Max

-0.77096 -0.23022 -0.00154

0.24470

0.74688

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept)

6.75288

0.40522

16.665

log(pLab/pCap)

0.44642

0.07949

5.616 1.06e-07 ***

log(pMat/pCap)

0.48117

0.07285

6.604 8.26e-10 ***

log(qOut)

0.37415

0.03021

12.384

< 2e-16 ***

< 2e-16 ***

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Residual standard error: 0.3383 on 136 degrees of freedom

Multiple R-squared:

0.8168,

F-statistic: 202.2 on 3 and 136 DF,

Adjusted R-squared:

0.8128

p-value: < 2.2e-16

The results are identical to the results from the Cobb-Douglas cost function with the price of
materials used for the normalization/deflation. The coefficient of the (logarithmic) capital price
can be obtained by the homogeneity condition (3.15). Hence, the estimate of Cap is 0.0724 in
our model with the capital price as numeraire, which is identical to the corresponding estimate
from the model with the price of materials as numeraire. Both models have identical residuals:

131

3 Dual Approach: Cost Functions

> all.equal( residuals( costCDHom ), residuals( costCDHomCap ) )
[1] TRUE
However, as the two models have different dependent variables (c/pMat and c/pCap ), the R2 -values
differ between the two models.
We can test the restriction for imposing linear homogeneity in input prices, e.g. by a Wald
test or a likelihood ratio test. As the models without and with homogeneity imposed (costCD
and costCDHom) have different dependent variables (c and c/pMat ), we cannot use the function
waldtest for conducting the Wald test but we have to use the function linearHypothesis
(package car) and specify the homogeneity restriction manually:
> library( "car" )
> linearHypothesis( costCD, "log(pCap) + log(pLab) + log(pMat) = 1"

Linear hypothesis test

Hypothesis:
log(pCap)

+ log(pLab)

+ log(pMat) = 1

Model 1: restricted model

Model 2: log(cost) ~ log(pCap) + log(pLab) + log(pMat) + log(qOut)
Res.Df

RSS Df Sum of Sq

136 15.563

135 15.560

F Pr(>F)

1 0.0025751 0.0223 0.8814

> lrtest( costCDHom, costCD )

Likelihood ratio test
Model 1: log(cost/pMat) ~ log(pCap/pMat) + log(pLab/pMat) + log(qOut)
Model 2: log(cost) ~ log(pCap) + log(pLab) + log(pMat) + log(qOut)
#Df

LogLik Df

5 -44.878

6 -44.867

Chisq Pr(>Chisq)

1 0.0232

0.879

These tests clearly show that the data do not contradict linear homogeneity in input prices.

3.2.5 Checking Concavity in Input Prices

The last property that we have to check is the concavity in input prices. A continuous and twice
continuously differentiable function is concave, if its Hessian matrix is negative semidefinite. A

132

3 Dual Approach: Cost Functions

necessary condition for negative semidefiniteness is that all diagonal elements are non-positive,
while a sufficient condition is that the first principal minor is non-positive and all following
principal minors alternate in sign (e.g. Chiang, 1984). The first derivatives of the Cobb-Douglas
cost function with respect to the input prices are:
c
ln c c
c
=
= i
wi
ln wi wi
wi

(3.20)

Now, we can calculate the second derivatives as derivatives of the first derivatives (3.20):


i wci
c
2c
= wi =
wi wj
wj
wj
c
i c
=
ij i 2
wi wj
wi
i
c
c
=
j
ij i 2
wi wj
wi
c
,
= i (j ij )
wi wj

(3.21)
(3.22)
(3.23)
(3.24)

where ij (again) denotes Kroneckers delta (2.66). Alternative, the second derivatives of the
Cobb-Douglas cost function with respect to the input prices can be written as:
2c
fi
fi fj
ij ,
=
wi wj
c
wi

(3.25)

where fi = c/wi indicates the first derivative.

We start with checking concavity in input prices of the Cobb-Douglas cost function without
homogeneity imposed. As argued in section 2.4.11.1, we do the calculations with the predicted
dependent variables rather than with the observed dependent variables.2 We can use following
command to obtain the total costs which are predicted by the Cobb-Douglas cost function without
homogeneity imposed:
> dat$costCD <- exp( fitted( costCD ) )
To simplify the calculations, we define short-cuts for the coefficients:
> cCap <- coef( costCD )[ "log(pCap)" ]
> cLab <- coef( costCD )[ "log(pLab)" ]
> cMat <- coef( costCD )[ "log(pMat)" ]
Using these coefficients, we compute the second derivatives of our estimated Cobb-Douglas cost
function:
2

Please note that the selection of c has no effect on the test for concavity, because all elements of the Hessian
matrix include c as a multiplicative term and c is always positive so that the value of c does not change the
sign of the principal minors and the determinant, as |c M | = c |M |, where M denotes a quadratic matrix, c
denotes a scalar, and the two vertical bars denote the determinant function.

133

3 Dual Approach: Cost Functions

> hCapCap <- cCap * ( cCap - 1 ) * dat$costCD / dat$pCap^2
> hLabLab <- cLab * ( cLab - 1 ) * dat$costCD / dat$pLab^2
> hMatMat <- cMat * ( cMat - 1 ) * dat$costCD / dat$pMat^2
> hCapLab <- cCap * cLab * dat$costCD / ( dat$pCap * dat$pLab )
> hCapMat <- cCap * cMat * dat$costCD / ( dat$pCap * dat$pMat )
> hLabMat <- cLab * cMat * dat$costCD / ( dat$pLab * dat$pMat )
Now, we prepare the Hessian matrix for the first observation:
> hessian <- matrix( NA, nrow = 3, ncol = 3 )
> hessian[ 1, 1 ] <- hCapCap[1]
> hessian[ 2, 2 ] <- hLabLab[1]
> hessian[ 3, 3 ] <- hMatMat[1]
> hessian[ 1, 2 ] <- hessian[ 2, 1 ] <- hCapLab[1]
> hessian[ 1, 3 ] <- hessian[ 3, 1 ] <- hCapMat[1]
> hessian[ 2, 3 ] <- hessian[ 3, 2 ] <- hLabMat[1]
> print( hessian )
[,1]

[,2]

[,3]

[1,] -5031.9274

7323.804

775.3358

[2,]
[3,]

7323.8043 -152736.317 14046.1549

775.3358

14046.155 -1570.0447

As all diagonal elements of this Hessian matrix are negative, the necessary conditions for negative semidefiniteness are fulfilled. Now, we calculate the principal minors in order to check the
sufficient conditions for negative semidefiniteness:
> hessian[1,1]
[1] -5031.927
> det( hessian[1:2,1:2] )
[1] 714919939
> det( hessian )
[1] 121651514835
While the conditions for the first two principal minors are fulfilled, the third principal minor is
positive, while negative semidefiniteness requires a non-positive third principal minor. Hence, this
Hessian matrix is not negative semidefinite and consequently, the Cobb-Douglas cost function is
not concave at the first observation.3
3

Please note that this Hessian matrix is not positive semidefinite either, because the first principal minor is
negative. Hence, the Cobb-Douglas cost function is neither concave nor convex at the first observation.

134

3 Dual Approach: Cost Functions

We can check the semidefiniteness of a matrix more conveniently with the command semidefiniteness (package miscTools), which (by default) checks the signs of the principal minors and
returns a logical value indicating whether the sufficient conditions for negative or positive semidefiniteness are fulfilled:
> semidefiniteness( hessian, positive = FALSE )
[1] FALSE
In the following, we will check whether concavity in input prices is fulfilled at each observation
in the sample:
> dat$concaveCD <- NA
> for( obs in 1:nrow( dat ) ) {
+

hessianLoop <- matrix( NA, nrow = 3, ncol = 3 )

hessianLoop[ 1, 1 ] <- hCapCap[obs]

hessianLoop[ 2, 2 ] <- hLabLab[obs]

hessianLoop[ 3, 3 ] <- hMatMat[obs]

hessianLoop[ 1, 2 ] <- hessianLoop[ 2, 1 ] <- hCapLab[obs]

hessianLoop[ 1, 3 ] <- hessianLoop[ 3, 1 ] <- hCapMat[obs]

hessianLoop[ 2, 3 ] <- hessianLoop[ 3, 2 ] <- hLabMat[obs]

dat$concaveCD[obs] <- semidefiniteness( hessianLoop, positive = FALSE )

+ }
> sum( dat$concaveCD )
[1] 0
This shows that our Cobb-Douglas cost function without linear homogeneity imposed is concave
in input prices not at a single observation.
Now, we will check, whether our Cobb-Douglas cost function with linear homogeneity imposed
is concave in input prices. Again, we obtain the predicted total costs:
> dat$costCDHom <- exp( fitted( costCDHom ) ) * dat$pMat
We create short-cuts for the estimated coefficients:
> chCap <- coef( costCDHom )[ "log(pCap/pMat)" ]
> chLab <- coef( costCDHom )[ "log(pLab/pMat)" ]
> chMat <- 1 - chCap - chLab
We compute the second derivatives:
> hhCapCap <- chCap * ( chCap - 1 ) * dat$costCDHom / dat$pCap^2
> hhLabLab <- chLab * ( chLab - 1 ) * dat$costCDHom / dat$pLab^2

135

3 Dual Approach: Cost Functions

> hhMatMat <- chMat * ( chMat - 1 ) * dat$costCDHom / dat$pMat^2
> hhCapLab <- chCap * chLab * dat$costCDHom /
+

( dat$pCap * dat$pLab )

> hhCapMat <- chCap * chMat * dat$costCDHom /

+

( dat$pCap * dat$pMat )

> hhLabMat <- chLab * chMat * dat$costCDHom /

+

( dat$pLab * dat$pMat )

And we prepare the Hessian matrix for the first observation:

> hessianHom <- matrix( NA, nrow = 3, ncol = 3 )
> hessianHom[ 1, 1 ] <- hhCapCap[1]
> hessianHom[ 2, 2 ] <- hhLabLab[1]
> hessianHom[ 3, 3 ] <- hhMatMat[1]
> hessianHom[ 1, 2 ] <- hessianHom[ 2, 1 ] <- hhCapLab[1]
> hessianHom[ 1, 3 ] <- hessianHom[ 3, 1 ] <- hhCapMat[1]
> hessianHom[ 2, 3 ] <- hessianHom[ 3, 2 ] <- hhLabMat[1]
> print( hessianHom )
[,1]

[,2]

[,3]

[1,] -4901.0204

6835.826

745.4417

[2,]
[3,]

6835.8256 -151446.932 13318.1722

745.4417

13318.172 -1566.0312

As all diagonal elements of this Hessian matrix are negative, the necessary conditions for negative semidefiniteness are fulfilled. Now, we calculate the principal minors in order to check the
sufficient conditions for negative semidefiniteness:
> hessianHom[1,1]
[1] -4901.02
> det( hessianHom[1:2,1:2] )
[1] 695515989
> det( hessianHom )
[1] 0.0006325681
The conditions for the first two principal minors are fulfilled and the third principal minor is close
to zero, where it is negative on some computers but positive on other computers. As Hessian
matrices of linear homogeneous functions are always singular, it is expected that the determinant

136

3 Dual Approach: Cost Functions

of the Hessian matrix (the N th principal minor) is zero. However, the computed determinant of
our Hessian matrix is not exactly zero due to rounding errors, which are unavoidable on digital
computers. Given that the determinant of the Hessian matrix of our Cobb-Douglas cost function
with linear homogeneity imposed should always be zero, the N th sufficient condition for negative
semidefiniteness (sign of the determinant of the Hessian matrix) should always be fulfilled. Consequently, we can conclude that our Cobb-Douglas cost function with linear homogeneity imposed
is concave in input prices at the first observation. In order to avoid problems due to rounding
errors, we can just check the negative semidefiniteness of the first N 1 rows and columns of the
Hessian matrix:
> semidefiniteness( hessianHom[1:2,1:2], positive = FALSE )
[1] TRUE
In the following, we will check whether concavity in input prices is fulfilled at each observation
in the sample:
> dat$concaveCDHom <- NA
> for( obs in 1:nrow( dat ) ) {
+

hessianPart <- matrix( NA, nrow = 2, ncol = 2 )

hessianPart[ 1, 1 ] <- hhCapCap[obs]

hessianPart[ 2, 2 ] <- hhLabLab[obs]

hessianPart[ 1, 2 ] <- hessianPart[ 2, 1 ] <- hhCapLab[obs]

dat$concaveCDHom[obs] <-

semidefiniteness( hessianPart, positive = FALSE )

+ }
> sum( !dat$concaveCDHom )
[1] 0
This result indicates that the concavity condition is violated not at a single observation. Consequently, our Cobb-Douglas cost function with linear homogeneity imposed is concave in input
prices at all observations.
In fact, all Cobb-Douglas cost functions that are non-decreasing and linearly homogeneous in
all input prices are always concave (e.g. Coelli, 1995, p. 266).4

3.2.6 Optimal Cost Shares

Given Shepards Lemma, the optimal cost shares derived from a Cobb-Douglas cost function are
equal to the coefficients of the (logarithmic) input prices:
i =
4

ln c(w, y)
c(w, y) wi
wi
wi xi (w, y)
=
= xi (w, y)
=
= si (w, y),
ln wi
wi c(w, y)
c(w, y)
c(w, y)

(3.26)

A proof and further details are available at http://www.econ.ucsb.edu/~tedb/Courses/GraduateTheoryUCSB/

concavity.pdf.

137

3 Dual Approach: Cost Functions

where si = wi xi /c are the cost shares.
The following commands draw histograms of the observed cost shares and compare them to
the optimal cost shares, which are predicted by our Cobb-Douglas cost function with linear
homogeneity imposed:
> hist( dat$pCap * dat$qCap / dat$cost )
> lines( rep( chCap, 2), c( 0, 100 ), lwd = 3

> hist( dat$pLab * dat$qLab / dat$cost )

> lines( rep( chLab, 2), c( 0, 100 ), lwd = 3

> hist( dat$pMat * dat$qMat / dat$cost )

0.0

0.1

0.2

0.3

0.4

30
25
15
0

10

Frequency

20
15
0

10

Frequency

20

25

35
30
25
20
15
10

Frequency

30

> lines( rep( chMat, 2), c( 0, 100 ), lwd = 3

0.3

cost share Cap

0.4

0.5

0.6

0.7

0.8

0.1

0.2

cost share Lab

0.3

0.4

0.5

0.6

cost share Mat

Figure 3.1: Observed and optimal costs shares

The resulting graphs are shown in figure 3.1. These results confirm results based on the production
function: most firms should increase the use of materials and decrease the use of capital goods.

3.2.7 Derived Input Demand Functions

Shepards Lemma says that the partial derivatives of the cost functions with respect to the input
prices are the conditional input demand functions. Therefore, the input demand functions based
on a Cobb-Douglas cost function are equal to the right-hand side of equation (3.20):
xi (w, y) =

c(w, y)
c(w, y)
= i
wi
wi

(3.27)

Input demand functions should be homogeneous of degree zero in input prices:

xi (t w, y) = xi (w, y)

138

(3.28)

3 Dual Approach: Cost Functions

This condition is fulfilled for the input demand functions derived from any linearly homogeneous
Cobb-Douglas cost function:
xi (t w, y) = i

t c(w, y)
c(w, y)
c(t w, y)
= i
= i
= xi (w, y)
t wi
t wi
wi

(3.29)

Furthermore, input demand functions should be symmetric with respect to input prices:
xi (t w, y)
xj (t w, y)
=
wj
wi

(3.30)

This condition is fulfilled for the input demand functions derived from any Cobb-Douglas cost
function:
xi (w, y)
i c(w, y)
i
c(w, y)
i j
=
=
j
=
c(w, y) i 6= j
wj
wi wj
wi
wj
wi wj
xj (w, y)
j c(w, y)
j
c(w, y)
i j
=
=
i
=
c(w, y) i 6= j
wi
wj
wi
wj
wi
wi wj

(3.31)
(3.32)

Finally, input demand functions should fulfill the negativity condition:

xi (t w, y)
0
wi

(3.33)

This condition is fulfilled for the input demand functions derived from any linearly homogeneous
Cobb-Douglas cost function that is monotonically increasing in all input prices (as this implies
0 i 1):
i c(w, y)
c(w, y)
xi (w, y)
=
i
wi
wi wi
wi2
i
c(w, y)
c(w, y)
=
i
i
wi
wi
wi2
c(w, y)
= i
(i 1) 0
wi2

(3.34)
(3.35)
(3.36)

We can calculate the cost-minimizing input quantities that are predicted by a Cobb-Douglas
cost function by using equation (3.27). The following commands compare the observed input
quantities with the cost-minimizing input quantities that are predicted by our Cobb-Douglas
cost function with linear homogeneity imposed:
> compPlot( chCap * dat$costCDHom / dat$pCap, dat$qCap )
> compPlot( chLab * dat$costCDHom / dat$pLab, dat$qLab )
> compPlot( chMat * dat$costCDHom / dat$pMat, dat$qMat )
> compPlot( chCap * dat$costCDHom / dat$pCap, dat$qCap, log = "xy" )
> compPlot( chLab * dat$costCDHom / dat$pLab, dat$qLab, log = "xy" )
> compPlot( chMat * dat$costCDHom / dat$pMat, dat$qMat, log = "xy" )

139

1200000

3 Dual Approach: Cost Functions

20000

4e+05

400000

qMat optimal

qCap optimal

5e+05

5e+04

5e+05

5e+03

5e+04

2e+05

qLab observed

1e+05
2e+04

1e+05

100000

2e+04

60000

5e+03

20000

qLab optimal

1200000

2e+04

5e+05

800000

qMat observed

2e+05

qCap optimal

5e+03

qMat observed

800000

100000

400000

qLab observed

0e+00

qCap observed

60000

4e+05
2e+05

0e+00

qCap observed

5e+04

2e+05

5e+05

5e+03

qLab optimal

Figure 3.2: Observed and optimal input quantities

140

2e+04

5e+04

qMat optimal

3 Dual Approach: Cost Functions

The resulting graphs are shown in figure 3.2. These results confirm earlier results: most firms
should increase the use of materials and decrease the use of capital goods.

3.2.8 Derived Input Demand Elasticities

Based on the derived input demand functions (3.27), we can derive the conditional input demand
elasticities:
xi (w, y) wj
wj xi (w, y)
i c(w, y)
wj
wj
c(w, y)
=
ij i
2
wi wj xi (w, y)
xi (w, y)
wj

ij (w, y) =

c(w, y)
wj
c(w, y)
i
j
ij i
wi
wj xi (w, y)
wi xi (w, y)
c(w, y)
i
= i j
ij
wi xi (w, y)
si (w, y)
i j
i
=
ij
si (w, y)
si (w, y)
=

= j ij

(3.37)
(3.38)
(3.39)
(3.40)
(3.41)
(3.42)

y
xi (w, y)
y
xi (w, y)
c(w, y) i
y
=
y wi xi (w, y)
c(w, y) i
y
= y
y wi xi (w, y)
y
c(w, y)
= i y
y wi xi (w, y)
c(w, y)
= i y
wi xi (w, y)
i
= y
si (w, y)

iy (w, y) =

= y

(3.43)
(3.44)
(3.45)
(3.46)
(3.47)
(3.48)
(3.49)

All derived input demand elasticities based on our estimated Cobb-Douglas cost function with
linear homogeneity imposed are presented in table 3.1. If the price of capital increases by one
percent, the cost-minimizing firm will decrease the use of capital by 0.93% and increase the
use of labor and materials by 0.07% each. If the price of labor increases by one percent, the
cost-minimizing firm will decrease the use of labor by 0.55% and increase the use of capital and
materials by 0.45% each. If the price of materials increases by one percent, the cost-minimizing
firm will decrease the use of materials by 0.52% and increase the use of capital and labor by
0.48% each. If the cost-minimizing firm increases the output quantity by one percent, (s)he will

141

3 Dual Approach: Cost Functions

increase all input quantities by 0.37%. The price elasticities derived from the Cobb-Douglas cost
function with linear homogeneity imposed are rather similar to the price elasticities derived from
the Cobb-Douglas production function but the elasticities with respect to the output quantity are
rather dissimilar (compare Tables 2.1 and 3.1). In theory, elasticities derived from a cost function,
which corresponds to a specific production function, should be identical to elasticities which are
directly derived from the production function. However, although our production function and
cost function are supposed to model the same production technology, their elasticities are not
the same. These differences arise from different econometric assumptions (e.g. exogeneity of
explanatory variables) and the disturbance terms, which differ between both models so that the
production technology is fitted differently.
Table 3.1: Conditional demand elasticities derived from Cobb-Douglas cost function (with linear
homogeneity imposed)
wcap wlab wmat
y
xcap -0.93 0.45 0.48 0.37
xlab
0.07 -0.55 0.48 0.37
xmat 0.07 0.45 -0.52 0.37
Given Eulers theorem and the cost functions homogeneity in input prices, following condition
for the price elasticities can be obtained:
X

ij = 0 i

(3.50)

The input demand elasticities derived from any linearly homogeneous Cobb-Douglas cost function
fulfill the homogeneity condition:
X

ij (w, y) =

X
j

(j ij ) =

ij = 1 1 = 0 i

(3.51)

As we computed the elasticities in table 3.1 based on the Cobb-Douglas function with linear
homogeneity imposed, these conditions are fulfilled for these elasticities.
It follows from the necessary conditions for the concavity of the cost function that all own-price
elasticities are non-positive:
ii 0 i

(3.52)

The input demand elasticities derived from any linearly homogeneous Cobb-Douglas cost function
that is monotonically increasing in all input prices fulfill the negativity condition, because linear
P

homogeneity (

i i

= 1) and monotonicity (i 0 i) together imply that all i s (optimal cost

shares) are between zero and one (0 i 1 i):

ii = i 1 0 i

142

(3.53)

3 Dual Approach: Cost Functions

As our Cobb-Douglas function with linear homogeneity imposed fulfills the homogeneity, monotonicity, and concavity condition, the elasticities in table 3.1 fulfill the negativity conditions.
The symmetry condition for derived demand elasticities
si ij = sj ji i, j

(3.54)

is fulfilled for any Cobb-Douglas cost function:

si ij (w, y) = i j = j i = sj ji i 6= j i, j

(3.55)

Hence, the symmetry condition is also fulfilled for the elasticities in table 3.1, e.g. scap cap,lab =
cap cap,lab = 0.07 0.45 is equal to slab lab,cap = lab lab,cap = 0.45 0.07.

3.2.9 Cost flexibility and elasticity of size

The coefficient of the (logarithmic) output quantity is equal to the cost flexibility (3.3). A value
of 0.37 (as in our estimated Cobb-Douglas cost function with linear homogeneity in input prices
imposed) means that a 1% increase in the output quantity results in a cost increase of 0.37%. The
elasticity of size is the inverse of the cost flexibility (3.4). A value of 2.67 (as derived from our
estimated Cobb-Douglas cost function with linear homogeneity in input prices imposed) means
that if costs are increased by 1%, the output quantity increases by 2.67%.

3.2.10 Marginal Costs, Average Costs, and Total Costs

Marginal costs can be calculated by
c(w, y)
c(w, y)
= y
y
y

(3.56)

These marginal costs should be linearly homogeneous in input prices:

c(t w, y)
c(w, y)
=t
y
y

(3.57)

This condition is fulfilled for the marginal costs derived from a linearly homogeneous CobbDouglas cost function:
c(t w, y)
c(t w, y)
t c(w, y)
c(w, y)
c(w, y)
= y
= y
= t y
=t
y
y
y
y
y
We can compute the marginal costs by following command:
> chOut <- coef( costCDHom )[ "log(qOut)" ]
> dat$margCost <- chOut * dat$costCDHom / dat$qOut
We can visualize these marginal costs with a histogram.

143

(3.58)

3 Dual Approach: Cost Functions

0 5

15

Frequency

25

> hist( dat$margCost, 20 )

0.0

0.1

0.2

0.3

0.4

0.5

margCost

Figure 3.3: Marginal costs

The resulting graph is shown in figure 3.3. It indicates that producing one additional output unit
increases the costs of most firms by around 0.08 monetary units.
Furthermore, we can check if the marginal costs are equal to the output prices, which is a
first-order condition for profit maximization:
> compPlot( dat$pOut, dat$margCost )

0.0

0.5

1.0

1.5

2.0

2.5

3.0

0.50
0.10

0.02

margCost

2.00

0.0 0.5 1.0 1.5 2.0 2.5 3.0

margCost

> compPlot( dat$pOut, dat$margCost, log = "xy" )

0.02

0.10

pOut

0.50

2.00

pOut

Figure 3.4: Marginal costs and output prices

The resulting graphs are shown in figure 3.4. The marginal costs of all firms are considerably
smaller than their output prices. Hence, all firms would gain from increasing their output level.
This is not surprising for a technology with large economies of scale.
Now, we analyze, how the marginal costs depend on the output quantity:
> plot( dat$qOut, dat$margCost )
> plot( dat$qOut, dat$margCost, log = "xy" )

144

0.50

1.0e+07

0.20
0.05

0.10

0.0e+00

margCost

0.3
0.2
0.1

margCost

0.4

0.5

3 Dual Approach: Cost Functions

2.0e+07

1e+05

5e+05

qOut

5e+06
qOut

Figure 3.5: Marginal costs depending on output quantity and firm size
The resulting graphs are shown in figure 3.5. Due to the large economies of size, the marginal
costs are decreasing with the output quantity.
The relation between output quantity and marginal costs in a Cobb-Douglas cost function can
be analyzed by taking the first derivative of the marginal costs (3.56) with respect to the output
quantity:


y c(w,y)
M C
y
=
y
y
y c(w, y)
c(w, y)
=
y
y
y
y2
c(w, y)
y c(w, y)
y
y
=
y
y
y2
c
= y 2 (y 1)
y

(3.59)
(3.60)
(3.61)
(3.62)

As y , c, and y 2 should always be positive, the marginal costs are (globally) increasing in the
output quantity, if there are decreasing returns to size (i.e. y > 1) and the marginal costs are
(globally) decreasing in the output quantity, if there are increasing returns to size (i.e. y < 1).
Now, we illustrate our estimated model by drawing the total cost curve for output quantities
between 0 and the maximum output level in the sample, where we use the sample means of the
input prices. Furthermore, we draw the average cost curve and the marginal cost curve for the
above-mentioned output quantities and input prices:
> y <- seq( 0, max( dat$qOut ), length.out = 200 )
> chInt <- coef(costCDHom)[ "(Intercept)" ]
> costs <- exp( chInt + chCap * log( mean( dat$pCap ) ) +
+

chLab * log( mean( dat$pLab ) ) + chMat * log( mean( dat$pMat ) ) +

chOut * log( y ) )

145

3 Dual Approach: Cost Functions

> plot( y, costs, type = "l" )
> # average costs
> plot( y, costs/y, type = "l" )
> # marginal costs
> lines( y, chOut * costs / y, lty = 2 )
> legend( "right", lty = c( 1, 2 ),

0.0e+00

1.2
0.8
0.4

average costs
marginal costs

0.0

400000 800000

average costs, marginal costs

legend = c( "average costs", "marginal costs" ) )

total costs

1.0e+07

2.0e+07

0.0e+00

1.0e+07

2.0e+07

Figure 3.6: Total, marginal, and average costs

The resulting graphs are shown in figure 3.6. As the marginal costs are equal to the average costs
multiplied by a fixed factor, y (see equation 3.56), the average cost curve and the marginal cost
curve of a Cobb-Douglas cost function cannot intersect.

3.3 Cobb-Douglas Short-Run Cost Function

3.3.1 Specification
Given the general specification of a short-run cost function (3.5), a Cobb-Douglas short-run cost
function is

cv = A

wii

iN 1

xj j y y ,

(3.63)

jN 2

where cv denotes the variable costs as defined in (1.3), N 1 is a vector of the indices of the variable
inputs, and N 2 is a vector of the indices of the quasi-fixed inputs. The Cobb-Douglas short-run

146

3 Dual Approach: Cost Functions

cost function can be linearized to
ln cv = 0 +

i ln wi +

iN 1

j ln xj + y ln y

(3.64)

jN 2

with 0 = ln A.

3.3.2 Estimation
The following commands estimate a Cobb-Douglas short-run cost function with capital as a
quasi-fixed input and summarize the results:
> costCDSR <- lm( log( vCost ) ~ log( pLab ) + log( pMat ) + log( qCap ) + log( qOut ),
+

data = dat )

> summary( costCDSR )

Call:
lm(formula = log(vCost) ~ log(pLab) + log(pMat) + log(qCap) +
log(qOut), data = dat)
Residuals:
Min

1Q

Median

3Q

Max

-0.73935 -0.20934 -0.00571

0.20729

0.71633

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept)

5.66013

0.42523

13.311

< 2e-16 ***

log(pLab)

0.45683

0.13819

3.306

0.00121 **

log(pMat)

0.44144

0.07715

5.722 6.50e-08 ***

log(qCap)

0.19174

0.04034

4.754 5.05e-06 ***

log(qOut)

0.29127

0.03318

8.778 6.49e-15 ***

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Residual standard error: 0.3183 on 135 degrees of freedom

Multiple R-squared:

0.7265,

F-statistic: 89.63 on 4 and 135 DF,

Adjusted R-squared:

0.7183

p-value: < 2.2e-16

3.3.3 Properties
This short-run cost function is (significantly) increasing in the prices of the variable inputs (labor
and materials) as the coefficient of the labor price (0.457) and the coefficient of the materials

147

3 Dual Approach: Cost Functions

price (0.441) are both positive. However, this short-run cost function is not linearly homogeneous
in input prices, as the coefficient of the labor price and the coefficient of the materials price do not
sum up to one (0.457 + 0.441 = 0.898). The short-run cost function is increasing in the output
quantity with a short-run cost flexibility of 0.291, which corresponds to a short-run elasticity of
size of 3.433. However, this short-run cost function is increasing in the quantity of the fixed input
(capital), as the corresponding coefficient is (significantly) positive (0.192) which contradicts
microeconomic theory. This would mean that the apple producers could reduce variable costs
(costs from labor and materials) by reducing the capital input (e.g. by destroying their apple trees
and machinery), while still producing the same amount of apples. Producing the same output
level with less of all inputs is not plausible.

3.3.4 Estimation with linear homogeneity in input prices imposed

We can impose linear homogeneity in the prices of the variable inputs as we did with the (longrun) cost function (see equations 3.15 to 3.19):
ln

X
X
c
wi
= 0 +
i ln
+
j ln xj + y ln y
wk
wk
1
2
iN \k

(3.65)

jN

with k N 1 . We can estimate a Cobb-Douglas short-run cost function with capital as a quasifixed input and linear homogeneity in input prices imposed by the command:
> costCDSRHom <- lm( log( vCost / pMat ) ~ log( pLab / pMat ) +
+

log( qCap ) + log( qOut ), data = dat )

> summary( costCDSRHom )

Call:
lm(formula = log(vCost/pMat) ~ log(pLab/pMat) + log(qCap) + log(qOut),
data = dat)
Residuals:
Min

1Q

Median

3Q

Max

-0.78305 -0.20539 -0.00265

0.19533

0.71792

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept)

5.67882

0.42335

13.414

< 2e-16 ***

log(pLab/pMat)

0.53487

0.06781

7.888 9.00e-13 ***

log(qCap)

0.18774

0.03978

4.720 5.79e-06 ***

log(qOut)

0.29010

0.03306

8.775 6.33e-15 ***

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

148

3 Dual Approach: Cost Functions

Residual standard error: 0.3176 on 136 degrees of freedom

Multiple R-squared:

0.5963,

Adjusted R-squared:

F-statistic: 66.97 on 3 and 136 DF,

0.5874

p-value: < 2.2e-16

We can obtain the coefficient of the materials price from the homogeneity condition (3.15): 1
0.535 = 0.465. We can test the homogeneity restriction by a likelihood ratio test:
> lrtest( costCDSRHom, costCDSR )
Likelihood ratio test
Model 1: log(vCost/pMat) ~ log(pLab/pMat) + log(qCap) + log(qOut)
Model 2: log(vCost) ~ log(pLab) + log(pMat) + log(qCap) + log(qOut)
#Df

LogLik Df

5 -36.055

6 -35.838

Chisq Pr(>Chisq)

1 0.4356

0.5093

Given the large P -value, we can conclude that the data do not contradict the linear homogeneity
in the prices of the variable inputs.
While the linear homogeneity in the prices of all variable inputs is accepted and the short-run
cost function is still increasing in the output quantity and the prices of all variable inputs, the
estimated short-run cost function is still increasing in the capital quantity, which contradicts
microeconomic theory. Therefore, a further microeconomic analysis with this function is not
reasonable.

3.4 Translog cost function

3.4.1 Specification
The general specification of a Translog cost function is
ln c(w, y) = 0 +
+

1
2

N
X

i ln wi
i=1
N X
N
X

+ y ln y

1
ij ln wi ln wj + yy (ln y)2
2
i=1 j=1

N
X

iy ln wi ln y

i=1

with ij = ji i, j.

149

(3.66)

3 Dual Approach: Cost Functions

3.4.2 Estimation
The Translog cost function can be estimated by following command:
> costTL <- lm( log( cost ) ~ log( pCap ) + log( pLab ) + log( pMat ) +
+

log( qOut ) + I( 0.5 * log( pCap )^2 ) + I( 0.5 * log( pLab )^2 ) +

I( 0.5 * log( pMat )^2 ) + I( log( pCap ) * log( pLab ) ) +

I( log( pCap ) * log( pMat ) ) + I( log( pLab ) * log( pMat ) ) +

I( 0.5 * log( qOut )^2 ) + I( log( pCap ) * log( qOut ) ) +

I( log( pLab ) * log( qOut ) ) + I( log( pMat ) * log( qOut ) ),

data = dat )

> summary( costTL )

Call:
lm(formula = log(cost) ~ log(pCap) + log(pLab) + log(pMat) +
log(qOut) + I(0.5 * log(pCap)^2) + I(0.5 * log(pLab)^2) +
I(0.5 * log(pMat)^2) + I(log(pCap) * log(pLab)) + I(log(pCap) *
log(pMat)) + I(log(pLab) * log(pMat)) + I(0.5 * log(qOut)^2) +
I(log(pCap) * log(qOut)) + I(log(pLab) * log(qOut)) + I(log(pMat) *
log(qOut)), data = dat)
Residuals:
Min

1Q

Median

3Q

Max

-0.73251 -0.18718

0.02001

0.15447

0.82858

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept)

25.383429

3.511353

7.229 4.26e-11 ***

log(pCap)

0.198813

0.537885

0.370 0.712291

log(pLab)

-0.024792

2.232126

-0.011 0.991156

log(pMat)

-1.244914

1.201129

-1.036 0.301992

log(qOut)

-2.040079

0.510905

-3.993 0.000111 ***

I(0.5 * log(pCap)^2)

-0.095173

0.105158

-0.905 0.367182

I(0.5 * log(pLab)^2)

-0.503168

0.943390

-0.533 0.594730

I(0.5 * log(pMat)^2)

0.529021

0.337680

1.567 0.119728

I(log(pCap) * log(pLab)) -0.746199

0.244445

I(log(pCap) * log(pMat))

0.182268

0.130463

1.397 0.164865

I(log(pLab) * log(pMat))

0.139429

0.433408

0.322 0.748215

I(0.5 * log(qOut)^2)

0.164075

0.041078

3.994 0.000110 ***

I(log(pCap) * log(qOut)) -0.028090

0.042844

-0.656 0.513259

I(log(pLab) * log(qOut))

0.171134

0.044 0.964959

0.007533

150

-3.053 0.002772 **

3 Dual Approach: Cost Functions

I(log(pMat) * log(qOut))

0.048794

0.092266

0.529 0.597849

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Residual standard error: 0.3043 on 125 degrees of freedom

Multiple R-squared:

0.7682,

Adjusted R-squared:

F-statistic: 29.59 on 14 and 125 DF,

0.7423

p-value: < 2.2e-16

As the Cobb-Douglas cost function is nested in the Translog cost function, we can use a
statistical test to check whether the Cobb-Douglas cost function fits the data as good as the
Translog cost function:
> lrtest( costCD, costTL )
Likelihood ratio test
Model 1: log(cost) ~ log(pCap) + log(pLab) + log(pMat) + log(qOut)
Model 2: log(cost) ~ log(pCap) + log(pLab) + log(pMat) + log(qOut) + I(0.5 *
log(pCap)^2) + I(0.5 * log(pLab)^2) + I(0.5 * log(pMat)^2) +
I(log(pCap) * log(pLab)) + I(log(pCap) * log(pMat)) + I(log(pLab) *
log(pMat)) + I(0.5 * log(qOut)^2) + I(log(pCap) * log(qOut)) +
I(log(pLab) * log(qOut)) + I(log(pMat) * log(qOut))
#Df
1
2

LogLik Df

Chisq Pr(>Chisq)

6 -44.867
16 -24.149 10 41.435

9.448e-06 ***

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Given the very small P -value, we can conclude that the Cobb-Douglas cost function is not suitable
for analyzing the production technology in our data set.

3.4.3 Linear homogeneity in input prices

Linear homogeneity of a Translog cost function requires
ln(t c(w, y)) = ln c(t w, y)
ln t + ln c(w, y) = 0 +
+

1
2

N
X

i ln(t
i=1
N X
N
X

(3.67)
wi ) + y ln y

1
ij ln(t wi ) ln(t wj ) + yy (ln y)2
2
i=1 j=1

N
X

iy ln(t wi ) ln y

i=1

151

(3.68)

3 Dual Approach: Cost Functions

= 0 +

N
X

i ln(t) +

i=1

N
X

i ln(wi ) + y ln y

(3.69)

i=1

N X
N
1X

N X
N
1X
+
ij ln(t) ln(t) +
ij ln(t) ln(wj )
2 i=1 j=1
2 i=1 j=1

N X
N X
N
N
1X
1X
ij ln(wi ) ln(t) +
ij ln(wi ) ln(wj )
2 i=1 j=1
2 i=1 j=1

N
N
X
X
1
iy ln(t) ln y +
iy ln(wi ) ln y
+ yy (ln y)2 +
2
i=1
i=1

= 0 + ln(t)

N
X
i=1

i +

N
X

i ln(wi ) + y ln y

(3.70)

i=1
N X
N
X

N
N
X
X
1
1
ln(wj )
ij
+ ln(t) ln(t)
ij + ln(t)
2
2
i=1 j=1
j=1
i=1

N
N
N X
N
X
X
1
1X
ln(t)
ln(wi )
ij +
ij ln(wi ) ln(wj )
2
2 i=1 j=1
i=1
j=1

N
N
X
X
1
2
iy ln(wi ) ln y
iy +
+ yy (ln y) + ln(t) ln y
2
i=1
i=1

= ln c(w, y) + ln(t)

N
X

(3.71)

i=1
N
N X
X

N
N
X
X
1
1
+ ln(t) ln(t)
ij + ln(t)
ij
ln(wj )
2
2
i=1
i=1 j=1
j=1

N
N
N
X
X
X
1
ln(t)
ln(wi )
ij + ln(t) ln y
iy
2
i=1
j=1
i=1

N X
N
N
N
X
X
X
1
1
ij + ln(t)
ln(wj )
ij
ln t = ln(t)
i + ln(t) ln(t)
2
2
i=1 j=1
j=1
i=1
i=1
N
X

N
N
N
X
X
X
1
iy
ln(t)
ln(wi )
ij + ln(t) ln y
2
i=1
i=1
j=1

N
X

N X
N
N
N
X
X
1
1X
i + ln(t)
1=
ij +
ij
ln(wj )
2
2 j=1
i=1
i=1
i=1 j=1

(3.72)

(3.73)

N
N
N
X
X
1X
ln(wi )
ij + ln y
iy
2 i=1
j=1
i=1

Hence, the homogeneity condition is only globally fulfilled (i.e. no matter which values t, w, and
y have) if the following parameter restrictions hold:
N
X

i = 1

(3.74)

i=1

152

3 Dual Approach: Cost Functions

N
X

ij =ji

ij = 0 j

i=1
N
X

N
X

ij = 0 i

(3.75)

j=1

iy = 0

(3.76)

i=1

We can see from the estimates above that these conditions are not fulfilled in our Translog cost
function. For instance, according to condition (3.74), the first-order coefficients of the input
prices should sum up to one but our estimates sum up to 0.199 + (0.025) + (1.245) = 1.071.
Hence, the homogeneity condition is not fulfilled in our estimated Translog cost function.

3.4.4 Estimation with linear homogeneity in input prices imposed

In order to impose linear homogeneity in input prices, we can rearrange these restrictions to get
N = 1
N j =
iN =

N
1
X

i=1
N
1
X
i=1
N
1
X

(3.77)

ij j

(3.78)

ij i

(3.79)

iy

(3.80)

j=1

N y =

N
1
X
i=1

Replacing N , N y and all iN and jN in equation (3.66) by the right-hand sides of equations (3.77) to (3.80) and re-arranging, we get
ln

N
1
X
c(w, y)
wi
= 0 +
i ln
+ y ln y
wN
w
N
i=1

(3.81)

1
1 N
X
wi
1 NX
wj
1
+
ij ln
ln
+ yy (ln y)2
2 j=1 i=1
wN
wN
2

N
1
X
i=1

iy ln

wi
ln y.
wN

This Translog cost function with linear homogeneity imposed can be estimated by following
command:
> costTLHom <- lm( log( cost / pMat ) ~ log( pCap / pMat ) +
+

log( pLab / pMat ) + log( qOut ) +

I( 0.5 * log( pCap / pMat )^2 ) + I( 0.5 * log( pLab / pMat )^2 ) +

I( log( pCap / pMat ) * log( pLab / pMat ) ) +

I( 0.5 * log( qOut )^2 ) + I( log( pCap / pMat ) * log( qOut ) ) +

153

3 Dual Approach: Cost Functions

+

I( log( pLab / pMat ) * log( qOut ) ),

data = dat )

> summary( costTLHom )

Call:
lm(formula = log(cost/pMat) ~ log(pCap/pMat) + log(pLab/pMat) +
log(qOut) + I(0.5 * log(pCap/pMat)^2) + I(0.5 * log(pLab/pMat)^2) +
I(log(pCap/pMat) * log(pLab/pMat)) + I(0.5 * log(qOut)^2) +
I(log(pCap/pMat) * log(qOut)) + I(log(pLab/pMat) * log(qOut)),
data = dat)
Residuals:
Min

1Q

Median

3Q

Max

-0.6860 -0.2086

0.0192

0.1978

0.8281

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept)

23.714976

3.445289

6.883 2.24e-10 ***

log(pCap/pMat)

0.306159

0.525789

0.582 0.561383

log(pLab/pMat)

1.093860

1.169160

0.936 0.351216

-1.933605

0.501090

I(0.5 * log(pCap/pMat)^2)

0.025951

0.089977

0.288 0.773486

I(0.5 * log(pLab/pMat)^2)

0.716467

0.338049

2.119 0.035957 *

I(log(pCap/pMat) * log(pLab/pMat)) -0.292889

0.142710

-2.052 0.042144 *

I(0.5 * log(qOut)^2)

0.158662

0.039866

I(log(pCap/pMat) * log(qOut))

-0.048274

0.040025

-1.206 0.229964

I(log(pLab/pMat) * log(qOut))

0.008363

0.096490

0.087 0.931067

log(qOut)

-3.859 0.000179 ***

3.980 0.000114 ***

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Residual standard error: 0.3089 on 130 degrees of freedom

Multiple R-squared:

0.6377,

F-statistic: 25.43 on 9 and 130 DF,

Adjusted R-squared:

0.6126

p-value: < 2.2e-16

We can use a likelihood ratio test to compare this function with the unconstrained Translog cost
function (3.66):
> lrtest( costTL, costTLHom )
Likelihood ratio test

154

3 Dual Approach: Cost Functions

Model 1: log(cost) ~ log(pCap) + log(pLab) + log(pMat) + log(qOut) + I(0.5 *
log(pCap)^2) + I(0.5 * log(pLab)^2) + I(0.5 * log(pMat)^2) +
I(log(pCap) * log(pLab)) + I(log(pCap) * log(pMat)) + I(log(pLab) *
log(pMat)) + I(0.5 * log(qOut)^2) + I(log(pCap) * log(qOut)) +
I(log(pLab) * log(qOut)) + I(log(pMat) * log(qOut))
Model 2: log(cost/pMat) ~ log(pCap/pMat) + log(pLab/pMat) + log(qOut) +
I(0.5 * log(pCap/pMat)^2) + I(0.5 * log(pLab/pMat)^2) + I(log(pCap/pMat) *
log(pLab/pMat)) + I(0.5 * log(qOut)^2) + I(log(pCap/pMat) *
log(qOut)) + I(log(pLab/pMat) * log(qOut))
#Df

LogLik Df

Chisq Pr(>Chisq)

16 -24.149

11 -29.014 -5 9.7309

0.08323 .

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

The null hypothesis, linear homogeneity in input prices, is rejected at the 10% significance level
but not at the 5% level. Given the importance of microeconomic consistency and that 5% is the
standard significance level, we continue our analysis with the Translog cost function with linear
homogeneity in input prices imposed.
Furthermore, we can use a likelihood ratio test to compare this function with the Cobb-Douglas
cost function with homogeneity imposed (3.19):
> lrtest( costCDHom, costTLHom )
Likelihood ratio test
Model 1: log(cost/pMat) ~ log(pCap/pMat) + log(pLab/pMat) + log(qOut)
Model 2: log(cost/pMat) ~ log(pCap/pMat) + log(pLab/pMat) + log(qOut) +
I(0.5 * log(pCap/pMat)^2) + I(0.5 * log(pLab/pMat)^2) + I(log(pCap/pMat) *
log(pLab/pMat)) + I(0.5 * log(qOut)^2) + I(log(pCap/pMat) *
log(qOut)) + I(log(pLab/pMat) * log(qOut))
#Df

LogLik Df

5 -44.878

11 -29.014

Chisq Pr(>Chisq)

6 31.727

1.84e-05 ***

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Again, the Cobb-Douglas functional form is clearly rejected by the data in favor of the Translog
functional form.
Some parameters of the Translog cost function with linear homogeneity imposed (3.81) have
not been directly estimated (N , N y , all iN , all jN ) but they can be retrieved from the

155

3 Dual Approach: Cost Functions

(directly) estimated parameters and equations (3.77) to (3.80). Please note that the specification
in equation (3.81) is used for the econometric estimation only; after retrieving the non-estimated
parameters, we can do our analysis based on equation (3.66). To facilitate the further analysis,
we create short-cuts of all estimated parameters and obtain the parameters that have not been
directly estimated:
> ch0 <- coef( costTLHom )[ "(Intercept)" ]
> ch1 <- coef( costTLHom )[ "log(pCap/pMat)" ]
> ch2 <- coef( costTLHom )[ "log(pLab/pMat)" ]
> ch3 <- 1 - ch1 - ch2
> chy <- coef( costTLHom )[ "log(qOut)" ]
> ch11 <- coef( costTLHom )[ "I(0.5 * log(pCap/pMat)^2)" ]
> ch22 <- coef( costTLHom )[ "I(0.5 * log(pLab/pMat)^2)" ]
> chyy <- coef( costTLHom )[ "I(0.5 * log(qOut)^2)" ]
> ch12 <- ch21 <- coef( costTLHom )[ "I(log(pCap/pMat) * log(pLab/pMat))" ]
> ch13 <- ch31 <- 0 - ch11 - ch12
> ch23 <- ch32 <- 0 - ch12 - ch22
> ch33 <- 0 - ch13 - ch23
> ch1y <- coef( costTLHom )[ "I(log(pCap/pMat) * log(qOut))" ]
> ch2y <- coef( costTLHom )[ "I(log(pLab/pMat) * log(qOut))" ]
> ch3y <- 0 - ch1y - ch2y
Hence, our estimated Translog cost function has following parameters:
> # alpha_0, alpha_i, alpha_y
> unname( c( ch0, ch1, ch2, ch3, chy ) )
[1] 23.7149761

0.3061589

1.0938598 -0.4000187 -1.9336052

> # alpha_ij
> matrix( c( ch11, ch12, ch13, ch21, ch22, ch23, ch31, ch32, ch33 ), ncol=3 )
[,1]
[1,]

[,2]

[,3]

0.02595083 -0.2928892

0.2669384

[2,] -0.29288920
[3,]

0.7164670 -0.4235778

0.26693837 -0.4235778

0.1566394

> # alpha_iy, alpha_yy

> unname( c( ch1y, ch2y, ch3y, chyy ) )
[1] -0.048274484

0.008362717

0.039911768

156

0.158661757

3 Dual Approach: Cost Functions

3.4.5 Cost Flexibility and Elasticity of Size

Based on the estimated parameters, we can calculate the cost flexibilities and the elasticities of
size. The cost flexibilities derived from a Translog cost function (3.66) are
N
X
ln c(w, y)
= y +
iy ln wi + yy ln y
ln y
i=1

(3.82)

and the elasticities of size areas alwaystheir inverses:

ln y
=
ln c

ln c(w, y)
ln y

1

(3.83)

We can calculate the cost flexibilities and the elasticities of size with following commands:
> dat$costFlex <- with( dat, chy + ch1y * log( pCap ) +
+

ch2y * log( pLab ) +

ch3y * log( pMat ) + chyy * log( qOut ) )

> dat$elaSize <- 1 / dat$costFlex

Now, we can visualize these values using histograms:
> hist( dat$costFlex )
> hist( dat$elaSize )

0.2

0.4

0.6

cost flexibility

0.8

50
40
30
10
0

20
0

5
0

0.0

20

Frequency

80
60
40

Frequency

20
15
10

Frequency

25

30

120

60

35

> hist( dat$elaSize[ dat$elaSize > 0 & dat$elaSize < 10 ] )

20

20

40

60

elasticity of size

80

100

10

elasticity of size

Figure 3.7: Translog cost function: cost flexibility and elasticity of size
The resulting graphs are presented in figure 3.7. Only 1 out of 140 cost flexibilities is negative.
Hence, the estimated Translog cost function is to a very large extent increasing in the output
quantity. All cost flexibilities are lower than one, which indicates that all apple producers operate
under increasing returns to size. Most cost flexibilities are around 0.5, which corresponds to an
elasticity of size of 2. Hence, if the apple producers increase their output quantity by one percent,
the total costs of most producers increases by around 0.5 percent. Orthe other way roundif

157

3 Dual Approach: Cost Functions

the apple producers increase their input use so that their costs increase by one percent, the output
quantity of most producers would increase by around two percent.
With the following commands, we visualize the relationship between output quantity and
elasticity of size
> plot( dat$qOut, dat$elaSize )
> abline( 1, 0 )
> plot( dat$qOut, dat$elaSize, ylim = c( 0, 10 ) )
> abline( 1, 0 )
> plot( dat$qOut, dat$elaSize, ylim = c( 0, 10 ), log = "x" )

10

0.0e+00

1.0e+07

2.0e+07

dat$elaSize

dat$elaSize

60
40
20
20

dat$elaSize

10

80

100

> abline( 1, 0 )

0.0e+00

1.0e+07

qOut

2.0e+07

1e+05

qOut

5e+05

5e+06
qOut

Figure 3.8: Translog cost function: output quantity and elasticity of size
The resulting graphs are shown in figure 3.8. With increasing output quantity, the elasticity of
size approaches one (from above). Hence, small apple producers could gain a lot from increasing
their size, while large apple producers would gain much less from increasing their size. However,
even the largest producers still gain from increasing their size so that the optimal firm size is
larger than the largest firm in the sample.

3.4.6 Marginal Costs and Average Costs

Marginal costs derived from a Translog cost function are
c(w, y)
ln c(w, y) c(w, y)
=
=
y
ln y
y

y +

N
X

iy ln wi + yy ln y

i=1

c(w, y)
.
y

(3.84)

Hence, they areas alwaysequal to the cost flexibility multiplied by total costs and divided
by the output quantity. We can compute the total costs that are predicted by our estimated
Translog cost function by following command:
> dat$costTLHom <- exp( fitted( costTLHom ) ) * dat$pMat

158

3 Dual Approach: Cost Functions

Now, we can compute the marginal costs by:
> dat$margCostTL <- with( dat, costFlex * costTLHom / qOut )
We can visualize these marginal costs with a histogram.

40
20
0

Frequency

> hist( dat$margCostTL, 15 )

0.15

0.05

0.05

0.15

margCostTL

Figure 3.9: Translog cost function: Marginal costs

The resulting graph is shown in figure 3.9. It indicates that producing one additional output unit
increases the costs of most firms by around 0.09 monetary units.
Furthermore, we can check if the marginal costs are equal to the output prices, which is a
first-order condition for profit maximization:
> compPlot( dat$pOut, dat$margCostTL )
> compPlot( dat$pOut[ dat$margCostTL > 0 ],
dat$margCostTL[ dat$margCostTL > 0 ], log = "xy" )

0.0 0.5 1.0 1.5 2.0 2.5 3.0

0.50

2.00

0.10

0.02

margCostTL

2.0
1.0
0.0

margCostTL

3.0

0.02

0.10

pOut

0.50

2.00

pOut

Figure 3.10: Translog cost function: marginal costs and output prices
The resulting graphs are shown in figure 3.10. The marginal costs of all firms are considerably
smaller than their output prices. Hence, all firms would gain from increasing their output level.
This is not surprising for a technology with large economies of scale.

159

3 Dual Approach: Cost Functions

Now, we analyze, how the marginal costs depend on the output quantity:
> plot( dat$qOut, dat$margCostTL )
> plot( dat$qOut, dat$margCostTL, log = "x" )

1.0e+07

0.15

0.05

0.15

0.0e+00

0.05

margCost

0.05

0.05
0.15

margCost

0.15

2.0e+07

1e+05

qOut

5e+05

5e+06
qOut

Figure 3.11: Translog cost function: Marginal costs depending on output quantity
The resulting graphs are shown in figure 3.11. There is no clear relationship between marginal
costs and the output quantity.
Now, we illustrate our estimated model by drawing the average cost curve and the marginal
cost curve for output quantities between 0 and five times the maximum output level in the sample,
where we use the sample means of the input prices.
> y <- seq( 0, 5 * max( dat$qOut ), length.out = 200 )
> lpCap <- log( mean( dat$pCap ) )
> lpLab <- log( mean( dat$pLab ) )
> lpMat <- log( mean( dat$pMat ) )
> totalCost <- exp( ch0 + ch1 * lpCap + ch2 * lpLab + ch3 * lpMat +
+

chy * log( y ) + 0.5 * chyy * log( y )^2 +

0.5 * ch11 * lpCap^2 + 0.5 * ch22 * lpLab^2 + 0.5 * ch33 * lpMat^2 +

ch12 * lpCap * lpLab + ch13 * lpCap * lpMat + ch23 * lpLab * lpMat +

ch1y * lpCap * log( y ) + ch2y * lpLab * log( y ) + ch3y * lpMat * log( y ) )

> margCost <- ( chy + chyy * log( y ) +

+

ch1y * lpCap + ch2y * lpLab + ch3y * lpMat ) * totalCost / y

> # average costs

> plot( y, totalCost/y, type = "l" )
> # marginal costs
> lines( y, margCost, lty = 2 )

160

3 Dual Approach: Cost Functions

> # maximum output level in the sample
> lines( rep( max( dat$qOut ), 2 ), c( 0, 1 ) )
> legend( "topright", lty = c( 1, 2 ),
+

legend = c( "average costs", "marginal costs" ) )

> # average costs

> plot( y, totalCost/y, type = "l", ylim = c( 0.07, 0.10 ) )
> # marginal costs
> lines( y, margCost, lty = 2 )
> # maximum output level in the sample
> lines( rep( max( dat$qOut ), 2 ), c( 0, 1 ) )
> legend( "topright", lty = c( 1, 2 ),

0.0e+00

4.0e+07

8.0e+07

1.2e+08

0.100
0.080

0.090

average costs
marginal costs

0.070

0.1

0.2

0.3

0.4

average costs
marginal costs

average costs, marginal costs

0.5

legend = c( "average costs", "marginal costs" ) )

average costs, marginal costs

0.0e+00

4.0e+07

8.0e+07

1.2e+08

Figure 3.12: Translog cost function: marginal and average costs

The resulting graphs are shown in figure 3.12. The average costs are decreasing until an output
level of around 70,000,000 units (1 unit 1 Euro) and they are increasing for larger output
quantities. The average cost curve intersects the marginal cost curve (of course) at its minimum.
However, as the maximum output level in the sample (approx. 25,000,000 units) is considerably
lower than the minimum of the average cost curve (approx. 70,000,000 units), the estimated
minimum of the average cost curve cannot be reliably determined because there are no data in
this region.

161

3 Dual Approach: Cost Functions

3.4.7 Derived Input Demand Functions

We can derive the cost-minimizing input quantities from the Translog cost function using Shepards lemma:
c(w, y)
wi
ln c(w, y) c
=
ln wi wi

xi (w, y) =

= i +

N
X

(3.85)
(3.86)

ij ln wj + iy ln y

j=1

c
wi

(3.87)

And we can re-arrange these derived input demand functions in order to obtain the cost-minimizing
cost shares:
si (w, y)

N
X
wi xi (w, y)
= i +
ij ln wj + iy ln y
c
j=1

(3.88)

We can calculate the cost-minimizing cost shares based on our estimated Translog cost function
by following commands:
> dat$shCap <- with( dat, ch1 + ch11 * log( pCap ) +
+

ch12 * log( pLab ) + ch13 * log( pMat ) + ch1y * log( qOut ) )

> dat$shLab <- with( dat, ch2 + ch21 * log( pCap ) +

+

ch22 * log( pLab ) + ch23 * log( pMat ) + ch2y * log( qOut ) )

> dat$shMat <- with( dat, ch3 + ch31 * log( pCap ) +

+

ch32 * log( pLab ) + ch33 * log( pMat ) + ch3y * log( qOut ) )

We visualize the cost-minimizing cost shares with histograms:

> hist( dat$shCap )
> hist( dat$shLab )
> hist( dat$shMat )
The resulting graphs are shown in figure 3.13. As the signs of the derived optimal cost shares are
equal to the signs of the first derivatives of the cost function with respect to the input prices, we
can check whether the cost function is non-decreasing in input prices by checking if the derived
optimal cost shares are non-negative. Counting the negative derived optimal cost shares, we find
that our estimated cost function is decreasing in the capital price at 24 observations, decreasing
in the labor price at 10 observations, and decreasing in the materials price at 3 observations.
Given that out data set has 140 observations, our estimated cost function is to a large extent
non-decreasing in input prices.
As our estimated cost function is (forced to be) linearly homogeneous in all input prices, the
derived optimal cost shares always sum up to one:
> range( with( dat, shCap + shLab + shMat ) )

162

25
20
15
10

Frequency

0.1 0.0

0.1

0.2

0.3

0.4

10

20

Frequency

15
10

Frequency

20

30

30

25

3 Dual Approach: Cost Functions

0.0

0.5

shCap

1.0

0.2

0.2

shLab

0.4

0.6

0.8

1.0

shMat

Figure 3.13: Translog cost function: cost-minimizing cost shares

[1] 1 1
We can use the following commands to compare the observed cost shares with the derived
cost-minimizing cost shares:
> compPlot( dat$shCap, dat$vCap / dat$cost )
> compPlot( dat$shLab, dat$vLab / dat$cost )
> compPlot( dat$shMat, dat$vMat / dat$cost )

1.0

shMat

0.2

0.0

0.6

1.0

observed

0.5

observed

0.1 0.2 0.3 0.4

shLab

0.2

0.1

observed

shCap

0.1 0.0

0.1

0.2

0.3

0.4

0.0

optimal

0.5

1.0

optimal

0.2 0.0

0.2

0.4

0.6

0.8

1.0

optimal

Figure 3.14: Translog cost function: observed and cost-minimizing cost shares
The resulting graphs are shown in figure 3.14. Most firms use less than optimal materials, while
there is a tendency to use more than optimal capital and a very slight tendency to use more than
optimal labor.
Similarly, we can compare the observed input quantities with the cost-minimizing input quantities:
> compPlot( dat$shCap * dat$costTLHom / dat$pCap,
+

dat$vCap / dat$pCap )

163

3 Dual Approach: Cost Functions

> compPlot( dat$shLab * dat$costTLHom / dat$pLab,
+

dat$vLab / dat$pLab )

> compPlot( dat$shMat * dat$costTLHom / dat$pMat,

dat$vMat / dat$pMat )

qCap

qLab

qMat

1e+05

1e+05

3e+05

0e+00

5e+05

200000

100000

0e+00

5e+05

optimal

1e+06

optimal

5e+05

1e+06

observed

2e+05 4e+05
1e+05

observed

observed

50000

150000

250000

optimal

Figure 3.15: Translog cost function: observed and cost-minimizing input quantities
The resulting graphs are shown in figure 3.15. Of course, the conclusions derived from these
graphs are the same as conclusions derived from figure 3.14.

3.4.8 Derived input demand elasticities

Based on the derived input demand functions (3.87), we can derive the input demand elasticities
with respect to input prices:
ij (w, y) =

xi (w, y) wj
wj xi (w, y)



i +

(3.89)

PN

k=1 ik

ln wk + iy ln y

c
wi

wj
xi

wj
"

N
X
ij c
+ i +
ik ln wk + iy ln y
wj wi
k=1

ij i +

N
X

ik ln wk + iy ln y

k=1

"

xj
wi

(3.90)
(3.91)

c wj
wi2 xi

ij c
x i wi x j
x i wi c wj
=
+
ij
wi wj
c wi
c wi2 xi
ij c
wj x j
wj
+
ij
=
wi x i
c
wi
ij
=
+ sj ij ,
si

164

(3.92)
(3.93)
(3.94)

3 Dual Approach: Cost Functions

where ij (again) denotes Kroneckers delta (2.66), and the input demand elasticities with respect
to the output quantity:
iy (w, y) =

y
xi (w, y)
y
xi (w, y)



i +

(3.95)

PN

k=1 ik

ln wk + iy ln y

c
wi

y
"

N
X
iy c
+ i +
ik ln wk + iy ln y
y wi
k=1

iy c wi xi c 1 y
=
+
wi y
c y wi xi
iy c
c y
=
+
wi xi y c
iy
ln c
=
+
,
si
ln y


y
xi
!

(3.96)
#

c 1 y
y wi xi

(3.97)

(3.98)
(3.99)
(3.100)

where ln c/ ln y is the cost flexibility (see section 3.1.2).

With the following commands, we compute the input demand elasticities at the first observation:
> ela <- matrix( NA, nrow = 3, ncol = 4 )
> ela[ 1, 1 ] <- ch11 / dat$shCap[1] + dat$shCap[1] - 1
> ela[ 1, 2 ] <- ch12 / dat$shCap[1] + dat$shLab[1]
> ela[ 1, 3 ] <- ch13 / dat$shCap[1] + dat$shMat[1]
> ela[ 1, 4 ] <- ch1y / dat$shCap[1] + dat$costFlex[1]
> ela[ 2, 1 ] <- ch21 / dat$shLab[1] + dat$shCap[1]
> ela[ 2, 2 ] <- ch22 / dat$shLab[1] + dat$shLab[1] - 1
> ela[ 2, 3 ] <- ch23 / dat$shLab[1] + dat$shMat[1]
> ela[ 2, 4 ] <- ch2y / dat$shLab[1] + dat$costFlex[1]
> ela[ 3, 1 ] <- ch31 / dat$shMat[1] + dat$shCap[1]
> ela[ 3, 2 ] <- ch32 / dat$shMat[1] + dat$shLab[1]
> ela[ 3, 3 ] <- ch33 / dat$shMat[1] + dat$shMat[1] - 1
> ela[ 3, 4 ] <- ch3y / dat$shMat[1] + dat$costFlex[1]
> ela
[,1]
[1,] -0.6383107

[,2]

[,3]

[,4]

-1.1835104 1.821821136 0.1653394

[2,]

4.4484591 -11.3084023 6.859943173 0.2293691

[3,]

0.5938258

-0.5948896 0.001063746 0.3983336

These demand elasticities indicate that when the capital price increases by one percent, the
demand for capital decreases by 0.638 percent, the demand for labor increases by 4.448 percent,

165

3 Dual Approach: Cost Functions

and the demand for materials increases by 0.594 percent. When the labor price increases by one
percent, the elasticities indicate that the demand for all inputs decreases, which is not possible
when the output quantity should be maintained. Furthermore, the symmetry condition for the
elasticities (3.54) indicates that the cross-price elasticities of each input pair must have the same
sign. However, this is not the case for the pairs capitallabor and materialslabor. The reason
for this is the negative predicted input share of labor:
> dat[ 1, c( "shCap", "shLab", "shMat" ) ]
shCap

shLab

shMat

1 0.2630271 -0.06997825 0.8069511

Finally, the negativity constraint (3.52) is violated, because the own-price elasticity of materials
is positive (0.001).
When the output quantity is increased by one percent, the demand for capital increases by
0.165 percent, the demand for labor increases by 0.229 percent, and the demand for materials
increases by 0.398 percent.
Now, we create a three-dimensional array and compute the demand elasticities for all observations:
> elaAll <- array( NA, c( 3, 4, nrow( dat ) ) )
> elaAll[ 1, 1, ] <- ch11 / dat$shCap + dat$shCap - 1
> elaAll[ 1, 2, ] <- ch12 / dat$shCap + dat$shLab
> elaAll[ 1, 3, ] <- ch13 / dat$shCap + dat$shMat
> elaAll[ 1, 4, ] <- ch1y / dat$shCap + dat$costFlex
> elaAll[ 2, 1, ] <- ch21 / dat$shLab + dat$shCap
> elaAll[ 2, 2, ] <- ch22 / dat$shLab + dat$shLab - 1
> elaAll[ 2, 3, ] <- ch23 / dat$shLab + dat$shMat
> elaAll[ 2, 4, ] <- ch2y / dat$shLab + dat$costFlex
> elaAll[ 3, 1, ] <- ch31 / dat$shMat + dat$shCap
> elaAll[ 3, 2, ] <- ch32 / dat$shMat + dat$shLab
> elaAll[ 3, 3, ] <- ch33 / dat$shMat + dat$shMat - 1
> elaAll[ 3, 4, ] <- ch3y / dat$shMat + dat$costFlex
We can visualize the elasticities using histograms but we will include only observations, at
which the cost function is non-decreasing in all input prices so that the optimal input shares are
always positive.
> monoObs <- with( dat, shCap >= 0 & shLab >= 0 & shMat >= 0 )
> hist( elaAll[1,1,monoObs] )
> hist( elaAll[1,2,monoObs] )

166

3 Dual Approach: Cost Functions

> hist( elaAll[1,3,monoObs] )
> hist( elaAll[2,1,monoObs] )
> hist( elaAll[2,2,monoObs] )
> hist( elaAll[2,3,monoObs] )
> hist( elaAll[3,1,monoObs] )
> hist( elaAll[3,2,monoObs] )
> hist( elaAll[3,3,monoObs] )
> hist( elaAll[1,4,monoObs] )
> hist( elaAll[2,4,monoObs] )
> hist( elaAll[3,4,monoObs] )
The resulting graphs are shown in figure 3.16. While the conditional own-price elasticities of
capital and materials are negative at almost all observations, the conditional own-price elasticity
of labor is positive at almost all observations. These violations of the negativity constraint (3.52)
originate from the violation of the concavity condition. As all conditional elasticities of the capital
demand with respect to the materials price as well as all conditional elasticities of the materials
demand with respect to the capital price are positive, we can conclude that capital and materials
are net substitutes. In contrast, all cross-price elasticities between capital and labor as well as
between labor and materials are negative. This indicates that the two pairs capital and labor as
well as labor and materials are net complements.
When the output quantity is increased by one percent, most farms would increase both the
labor quantity and the materials quantity by around 0.5% and either increase or decrease the
capital quantity.

3.4.9 Theoretical consistency

The Translog cost function (3.66) is always continuous for positive input prices and a positive
output quantity.
The non-negativity is always fulfilled for the Translog cost function, because the predicted cost
is equal to the exponential function of the right-hand side of equation (3.66) and the exponential
function always returns a non-negative value (also when the right-hand side of equation (3.66) is
negative).
If the output quantity approaches zero (from above), the right-hand side of the Translog cost
functions (equation 3.66) approaches:

N
X

N X
N
N
X
1X
1
lim 0 +
i ln wi + y ln y +
ij ln wi ln wj + yy (ln y)2 +
iy ln wi ln y
y0+
2
2
i=1
i=1 j=1
i=1

(3.101)
= lim

y0+

N
X
1
y ln y + yy (ln y)2 +
iy ln wi ln y
2
i=1

167

(3.102)

15

20

100
200

20

10

80
60

80

100

70

10 0

80
Frequency

20
0
8

0.5

0.5

E mat lab

1.5

2.5

E mat mat

30

20

10

30
10
0

0
40

20

Frequency

10

20

Frequency

30
20
10
0

Frequency

30

40

40

40

50

E mat cap
50

30

100

100
60
20
0
4

50

E lab mat

40

Frequency
3

250

20
40

80

80
60
40

Frequency

20

200

0
20

E lab lab

150

60

Frequency
0

E lab cap

100

100

100
80
60
20
0
30

50

E cap mat

40

Frequency

80
60
40
0

20

Frequency

100

E cap lab

100

E cap cap

40

60
20
0

300

40

10

60

40

40

Frequency

80

80
60
0

20

40

Frequency

60
40
20

Frequency

80

100

3 Dual Approach: Cost Functions

0.0

0.5

E cap y

1.0

1.5

0.0

0.4

E lab y

Figure 3.16: Translog cost function: derived demand elasticities

168

0.8
E mat y

1.2

3 Dual Approach: Cost Functions

N
X
1
y + yy ln y +
iy ln wi ln y
2
i=1

= lim

y0+

N
X
1
y + yy ln y +
iy ln wi
2
i=1

= lim

y0+

(3.103)

lim ln y

(3.104)

y0+

= lim (yy ln y) lim ln y = (yy )() = yy

y0+

(3.105)

y0+

Hence, if coefficientt yy is negativ and the output quantity approaches zero (from above), the
predicted cost (exponential function of the right-hand side of equation 3.66) approaches zero so
that the no fixed costs property is asymptotically fulfilled.
Our estimated Translog cost function with linear homogeneity in input prices imposed (of
course) is linearly homogeneous in input prices. Hence, the linear homogeneity property is globally
fulfilled.
A cost function is non-decreasing in the output quantity if the cost flexibility and the elasticity
of size are non-negative. As we can see from figure 3.7, only a single cost flexibility and thus,
only a single elasticity of size is negative. Hence, our estimated Translog cost function with linear
homogeneity in input prices imposed violates the monotonicity condition regarding the output
quantity only at a single observation.
Given Shepards lemma, a cost function is non-decreasing in input prices if the derived costminimizing input quantities and the corresponding cost shares are non-negative. As we can see
from figure 3.13, our estimated Translog cost function with linear homogeneity in input prices
imposed predicts that 24 cost shares of capital, 10 cost shares of labor, and 3 cost shares of
materials are negative. In total, the monotonicity condition regarding the input prices is violated
at 36 observations:
> sum( dat$shCap < 0 | dat$shLab < 0 | dat$shMat < 0 )
[1] 36
Concavity in input prices of the cost function requires that the Hessian matrix of the cost
function with respect to the input prices is negative semidefinite. The elements of the Hessian
matrix are:
Hij =

2 c(w, y)
xi (w, y)
=
wi wj
wj
c
wi

N
X
ij c
=
+ i +
ik ln wk + iy ln y
wj wi
k=1



(3.106)

i +

PN

k=1 ik

ln wk + iy ln y

(3.107)

wj
N
X
xj
ij i +
ik ln wk + iy ln y
wi
k=1

c
wi2
(3.108)

ij c
x i wi x j
x i wi c
=
+
ij
wi wj
c wi
c wi2

(3.109)

169

3 Dual Approach: Cost Functions

=

ij c
xi xj
xi
+
ij ,
wi wj
c
wi

(3.110)

where ij (again) denotes Kroneckers delta (2.66). As the elements of the Hessian matrix have
the same sign as the corresponding elasticities (Hij = ij (w, y) xi /wj ), the positive own-price
elasticities of labor in figure 3.16 indicate that the element Hlab,lab is positive at all observations, where the monotonicity conditions regarding the input prices are fulfilled. As negative
semidefiniteness requires that all diagonal elements of the (Hessian) matrix are negative, we can
conclude that the estimated Translog cost function is concave at not a single observation where
the monotonicity conditions regarding the input prices are fulfilled.
This means that our estimated Translog cost function is inconsistent with microeconomic theory
at all observations.

170

4 Dual Approach: Profit Function

4.1 Theory
4.1.1 Profit functions
The profit function:
(p, w) = max p y
y,x

wi xi , s.t. y = f (x)

(4.1)

returns the maximum profit that is attainable given the output price p and input prices w.
It is important to distinguish the profit definition (1.4) from the profit function (4.1).

4.1.2 Short-run profit functions

As producers often cannot instantly adjust the quantity of the some inputs (e.g. capital, land,
apple trees), estimating a short-run profit function with some quasi-fixed input quantities might
be more appropriate than a (long-run) profit function which assumes that all input quantities
and output quantities can be adjusted instantly. Furthermore, a short-run profit function can
model technologies with increasing returns to scale, if the sum over the output elasticities of the
variable inputs is lower than one.
In general, a short-run profit function is defined as
n

v (p, w1 , x2 ) = max p y cs (w1 , y, x2 ) ,

y0

(4.2)

where w1 denotes the vector of the prices of all variable inputs, x2 denotes the vector of the
quantities of all quasi-fixed inputs, cs (w1 , y, x2 ) is the short-run cost function (see section 3.3),
v denotes the gross margin defined in equation (1.5), and N 1 is a vector of the indices of the
variable inputs.

4.2 Graphical illustration of profit and gross margin

We use the following commands to visualize the variation of the profits and the relationship
between profits and firm size:
> hist( dat$profit, 30 )
> plot( dat$X, dat$profit, log = "xy" )

171

4 Dual Approach: Profit Function

1e+07

2e+04

5e+05

profit

60
40
20

Frequency

80

0e+00

2e+07

4e+07

6e+07

0.5

profit

1.0

2.0

5.0

Figure 4.1: Profit

The resulting graphs are shown in figure 4.1. The histogram shows that 14 out of 140 apple
producers (10%) have (slightly) negative profits. Although this seems to be not unrealistic, this
contradicts the non-negativity condition of the profit function. However, the observed negative
profits might have been caused by deviations from the theoretical assumptions that we have made
to derive the profit function, e.g. that all inputs can be instantly adjusted and that there are no
unexpected events such as severe weather conditions or pests. We will deal with these deviations
from our assumptions later and for now just ignore the observations with negative profits in
our analyses with the profit function. The right part of figure 4.1 shows that the profit clearly
increases with firm size.
The following commands graphically illustrate the variation of the gross margins and their
relationship to the firm size and the quantity of the quasi-fixed input:
> hist( dat$vProfit, 30 )
> plot( dat$X, dat$vProfit, log = "xy" )
> plot( dat$qCap, dat$vProfit, log = "xy" )
The resulting graphs are shown in figure 4.2. The histogram on the left shows that 8 out of
140 apple producers (6%) have (slightly) negative gross margins. Although this does not seem
to be unrealistic, this contradicts the non-negativity condition of the short-run profit function.
However, the observed negative gross margins might have been caused by deviations from the
theoretical assumptions, e.g. that there are no unexpected events such as severe weather conditions or pests. The center part of figure 4.2 shows that the gross margin clearly increases with
the firm size (as expected). However, the right part of this figure shows that the gross margin is
only weakly positively correlated with the fixed input.

172

0e+00

2e+07

4e+07

6e+07

0.5

1.0

gross margin

2.0

5.0

5e+03 5e+04 5e+05 5e+06 5e+07

gross margin

5e+03 5e+04 5e+05 5e+06 5e+07

gross margin

60
40
0

20

Frequency

80

4 Dual Approach: Profit Function

5e+03

2e+04

1e+05

5e+05

qCap

Figure 4.2: Gross margins

Please note that according to microeconomic theory, the short-run total profit s in contrast
to the gross margin v might be negative due to fixed costs:
s (p, w, x2 ) = v (p, w1 , x2 )

wj x j ,

(4.3)

jN 2

where N 2 is a vector of the indices of the quasi-fixed inputs. However, in the long-run, profit
must be non-negative:
(p, w) = max s (p, w, x2 ) 0,
x2

(4.4)

as all costs are variable in the long run.

4.3 Cobb-Douglas Profit Function

4.3.1 Specification
The Cobb-Douglas profit function1 has the following specification:
!
p

= Ap

Y
w i
i

(4.5)

This function can be linearized to

ln = 0 + p ln p +

i ln wi

(4.6)

with 0 = ln A.

Please note that the Cobb-Douglas profit function is used as a simple example here but that it is much too
restrictive for most real empirical applications (Chand and Kaul, 1986).

173

4 Dual Approach: Profit Function

4.3.2 Estimation
The linearized Cobb-Douglas profit function can be estimated by OLS. As the logarithm of a
negative number is not defined and function lm automatically removes observations with missing
data, we do not have to remove the observations (apple producers) with negative profits manually.
> profitCD <- lm( log( profit ) ~ log( pOut ) + log( pCap ) + log( pLab ) +
+

log( pMat ), data = dat )

> summary( profitCD )

Call:
lm(formula = log(profit) ~ log(pOut) + log(pCap) + log(pLab) +
log(pMat), data = dat)
Residuals:
Min

1Q

Median

3Q

Max

-3.6183 -0.2778

0.1261

0.5986

2.0442

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept)

13.9380

0.4921

28.321

< 2e-16 ***

log(pOut)

2.7117

0.2340

11.590

< 2e-16 ***

log(pCap)

-0.7298

0.1752

-4.165 5.86e-05 ***

log(pLab)

-0.1940

0.4623

-0.420

0.676

log(pMat)

0.1612

0.2543

0.634

0.527

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Residual standard error: 0.9815 on 121 degrees of freedom

(14 observations deleted due to missingness)
Multiple R-squared:

0.5911,

F-statistic: 43.73 on 4 and 121 DF,

Adjusted R-squared:

0.5776

p-value: < 2.2e-16

As expected, lm reports that 14 observations have been removed due to missing data (logarithms
of negative numbers).

4.3.3 Properties
A Cobb-Douglas profit function is always continuous and twice continuously differentiable for all
p > 0 and wi > 0 i. Furthermore, a Cobb-Douglas profit function automatically fulfills the
non-negativity property, because the profit predicted by equation (4.5) is always positive as long
as coefficient A is positive (given that all input prices and the output price are positive). As A

174

4 Dual Approach: Profit Function

is usually obtained by applying the exponential function to the estimate of 0 , i.e. A = exp(0 ),
A and hence, also the predicted profit, are always positive (even if 0 is non-positive).
The estimated coefficients of the output price and the input prices indicate that profit is
increasing in the output price and decreasing in the capital and labor price but it is increasing in
the price of materials, which contradicts microeconomic theory. However, the positive coefficient
of the (logarithmic) price of materials is statistically not significantly different from zero.
The Cobb-Douglas profit function is linearly homogeneous in all prices (output price and all
input prices) if the following condition is fulfilled:
t (p, w) = (t p, t w)

(4.7)

ln(t ) = 0 + p ln(t p) +

i ln(t wi )

(4.8)

ln t + ln = 0 + p ln t + p ln p +

i ln t +

i ln wi

(4.9)

ln t + ln = 0 + p ln p +

i ln wi + ln t p +

(4.10)

ln = ln + ln t

p +

i 1

(4.11)

0 = ln t

p +

i 1

(4.12)

0 = p +

i 1

(4.13)

(4.14)

1 = p +

X
i

Hence, the homogeneity condition is only fulfilled if the coefficient of the (logarithmic) output
price and the coefficients of the (logarithmic) input prices sum up to one. As they sum up to
2.71 + (0.73) + (0.19) + 0.16 = 1.95, the homogeneity condition is not fulfilled in our estimated
model.

4.3.4 Estimation with linear homogeneity in all prices imposed

In order to derive a Cobb-Douglas profit function with linear homogeneity in input prices imposed,
we re-arrange the homogeneity condition (4.14) to get
p = 1

N
X

(4.15)

i=1

and replace p in the profit function (4.6) by the right-hand side of the above equation:
!

ln = 0 + 1

X
i

175

i ln p +

X
i

i ln wi

(4.16)

4 Dual Approach: Profit Function

ln = 0 + ln p

i ln p +

ln ln p = 0 +

i ln wi

(4.17)

i (ln wi ln p)

(4.18)

ln

wi
= 0 +
i ln
p
p
i

(4.19)

This Cobb-Douglas profit function with linear homogeneity imposed can be estimated by following
command:
> profitCDHom <- lm( log( profit / pOut ) ~ log( pCap / pOut ) +
+

log( pLab / pOut ) + log( pMat / pOut ), data = dat )

> summary( profitCDHom )

Call:
lm(formula = log(profit/pOut) ~ log(pCap/pOut) + log(pLab/pOut) +
log(pMat/pOut), data = dat)
Residuals:
Min

1Q

Median

3Q

Max

-3.6045 -0.2724

0.0972

0.6013

2.0385

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept)

14.27961

0.45962

31.068

< 2e-16 ***

log(pCap/pOut) -0.82114

0.16953

-4.844 3.78e-06 ***

log(pLab/pOut) -0.90068

0.25591

-3.519 0.000609 ***

log(pMat/pOut) -0.02469

0.23530

-0.105 0.916610

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Residual standard error: 0.9909 on 122 degrees of freedom

(14 observations deleted due to missingness)
Multiple R-squared:

0.3568,

F-statistic: 22.56 on 3 and 122 DF,

Adjusted R-squared:

0.341

p-value: 1.091e-11

The coefficient of the (logarithmic) output price can be obtained by the homogeneity restriction (4.15). Hence, it is 1 (0.82) (0.9) (0.02) = 2.75. Now, all monotonicity conditions
are fulfilled: profit is increasing in the output price and decreasing in all input prices. We can
use a Wald test or a likelihood-ratio test to test whether the model and the data contradict the
homogeneity assumption:

176

4 Dual Approach: Profit Function

> library( "car" )
> linearHypothesis( profitCD, "log(pOut) + log(pCap) + log(pLab) + log(pMat) = 1" )
Linear hypothesis test
Hypothesis:
log(pOut)

+ log(pCap)

+ log(pLab)

+ log(pMat) = 1

Model 1: restricted model

Model 2: log(profit) ~ log(pOut) + log(pCap) + log(pLab) + log(pMat)
Res.Df

RSS Df Sum of Sq

122 119.78

121 116.57

Pr(>F)

3.2183 3.3407 0.07005 .

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

> lrtest( profitCD, profitCDHom )

Likelihood ratio test
Model 1: log(profit) ~ log(pOut) + log(pCap) + log(pLab) + log(pMat)
Model 2: log(profit/pOut) ~ log(pCap/pOut) + log(pLab/pOut) + log(pMat/pOut)
#Df

LogLik Df

Chisq Pr(>Chisq)

6 -173.88

5 -175.60 -1 3.4316

0.06396 .

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Both tests reject the null hypothesis, linear homogeneity in all prices, at the 10% significance
level but not at the 5% level. Given the importance of microeconomic consistency and that 5%
is the standard significance level, we continue our analysis with the Cobb-Douglas profit function
with linear homogeneity imposed.

4.3.5 Checking Convexity in all prices

The last property that we have to check is the convexity in all prices. A continuous and twice
continuously differentiable function is convex, if its Hessian matrix is positive semidefinite. A
necessary condition for positive semidefiniteness is that all diagonal elements are non-negative,
while a sufficient condition is that all principal minors are non-negative (e.g. Chiang, 1984). The

177

4 Dual Approach: Profit Function

first derivatives of the Cobb-Douglas profit function with respect to the input prices are:

ln
=
= i
wi
ln wi wi
wi

(4.20)

and the first derivative with respect to the output price is:

ln

=
= p
p
ln p p
p

(4.21)

Now, we can calculate the second derivatives as derivatives of the first derivatives (4.20)
and (4.21):
2
wi wj

2
wi p

w
i

i wi

=
wj
wj
i

=
ij i 2
wi wj
wi

i
=
j
ij i 2
wi wj
wi

= i (j ij )
wi wj
=

w
i

i wi

=
p
i
=
wi p
i
=
p
wi p

= i p
wi p

(4.22)
(4.23)
(4.24)
(4.25)

(4.26)
(4.27)
(4.28)
(4.29)

p p

2
p
=
=
p2
p
p

p
=
p 2
p p
p
p

=
p p 2
p
p
p

= p (p 1) 2 ,
p

(4.30)
(4.31)
(4.32)
(4.33)

where ij (again) denotes Kroneckers delta (2.66).

As all elements of the Hessian matrix include as a multiplicative term, we can ignore this
variable in the calculation of the Hessian matrix, because the value neither changes the signs of
the (diagonal) elements of the matrix, nor the signs of the principal minors and the determinant
(as long as is positive, i.e. the non-negativity condition is fulfilled) given the general rule that
| M | = |M |, where M denotes a quadratic matrix, denotes a scalar, and the two vertical
bars denote the determinant function.

178

4 Dual Approach: Profit Function

We start with checking convexity in all prices of the Cobb-Douglas profit function without
homogeneity imposed.
To simplify the calculations, we define short-cuts for the coefficients:
> gCap <- coef( profitCD )[ "log(pCap)" ]
> gLab <- coef( profitCD )[ "log(pLab)" ]
> gMat <- coef( profitCD )[ "log(pMat)" ]
> gOut <- coef( profitCD )[ "log(pOut)" ]
Using these coefficients, we compute the second derivatives of our estimated Cobb-Douglas profit
function:
> hpCapCap <- gCap * ( gCap - 1 ) / dat$pCap^2
> hpLabLab <- gLab * ( gLab - 1 ) / dat$pLab^2
> hpMatMat <- gMat * ( gMat - 1 ) / dat$pMat^2
> hpCapLab <- gCap * gLab / ( dat$pCap * dat$pLab )
> hpCapMat <- gCap * gMat / ( dat$pCap * dat$pMat )
> hpLabMat <- gLab * gMat / ( dat$pLab * dat$pMat )
> hpCapOut <- gCap * gOut / ( dat$pCap * dat$pOut )
> hpLabOut <- gLab * gOut / ( dat$pLab * dat$pOut )
> hpMatOut <- gMat * gOut / ( dat$pMat * dat$pOut )
> hpOutOut <- gOut * ( gOut - 1 ) / dat$pOut^2
Now, we prepare the Hessian matrix for the first observation:
> hessian <- matrix( NA, nrow = 4, ncol = 4 )
> hessian[ 1, 1 ] <- hpCapCap[1]
> hessian[ 2, 2 ] <- hpLabLab[1]
> hessian[ 3, 3 ] <- hpMatMat[1]
> hessian[ 1, 2 ] <- hessian[ 2, 1 ] <- hpCapLab[1]
> hessian[ 1, 3 ] <- hessian[ 3, 1 ] <- hpCapMat[1]
> hessian[ 2, 3 ] <- hessian[ 3, 2 ] <- hpLabMat[1]
> hessian[ 1, 4 ] <- hessian[ 4, 1 ] <- hpCapOut[1]
> hessian[ 2, 4 ] <- hessian[ 4, 2 ] <- hpLabOut[1]
> hessian[ 3, 4 ] <- hessian[ 4, 3 ] <- hpMatOut[1]
> hessian[ 4, 4 ] <- hpOutOut[1]
> print( hessian )
[,1]

[,2]

[,3]

[,4]

[1,]

0.185633270

0.060331020 -0.005072286 -1.14920901

[2,]

0.060331020

0.286060178 -0.003907371 -0.88527867

[3,] -0.005072286 -0.003907371 -0.001709437

[4,] -1.149209014 -0.885278673

0.07442915

0.074429148 10.64451706

179

4 Dual Approach: Profit Function

As the third element on the diagonal of this Hessian matrix is negative, the necessary condition
for positive semidefiniteness is not fulfilled. Hence, we do not need to calculate the principal
minors of the Hessian matrix, as we already can conclude that the Hessian matrix is not positive
semidefinite and hence, the estimated profit function is not convex at the first observation.2
We can check whether the third element on the diagonal of the Hessian matrix is non-negative
at other observations:
> sum( hpMatMat >= 0, na.rm = TRUE )
[1] 0
As it is non-negative not at a single observation, we must conclude that the estimated CobbDouglas profit function without homogeneity imposed violates the convexity property at all observations.
Now, we will check, whether our Cobb-Douglas profit function with linear homogeneity imposed
is convex in all prices. Again, we create short-cuts for the estimated coefficients:
> ghCap <- coef( profitCDHom )["log(pCap/pOut)"]
> ghLab <- coef( profitCDHom )["log(pLab/pOut)"]
> ghMat <- coef( profitCDHom )["log(pMat/pOut)"]
> ghOut <- 1- ghCap - ghLab - ghMat
We compute the second derivatives:
> hphCapCap <- ghCap * ( ghCap - 1 ) / dat$pCap^2
> hphLabLab <- ghLab * ( ghLab - 1 ) / dat$pLab^2
> hphMatMat <- ghMat * ( ghMat - 1 ) / dat$pMat^2
> hphCapLab <- ghCap * ghLab / ( dat$pCap * dat$pLab )
> hphCapMat <- ghCap * ghMat / ( dat$pCap * dat$pMat )
> hphLabMat <- ghLab * ghMat / ( dat$pLab * dat$pMat )
> hphCapOut <- ghCap * ghOut / ( dat$pCap * dat$pOut )
> hphLabOut <- ghLab * ghOut / ( dat$pLab * dat$pOut )
> hphMatOut <- ghMat * ghOut / ( dat$pMat * dat$pOut )
> hphOutOut <- ghOut * ( ghOut - 1 ) / dat$pOut^2
And we prepare the Hessian matrix for the first observation:
> hessianHom <- matrix( NA, nrow = 4, ncol = 4 )
> hessianHom[ 1, 1 ] <- hphCapCap[1]
> hessianHom[ 2, 2 ] <- hphLabLab[1]
> hessianHom[ 3, 3 ] <- hphMatMat[1]
2

Please note that this Hessian matrix is not negative semidefinite either, because the other three principal minors
are positive. Hence, the Cobb-Douglas profit function is neither concave nor convex at the first observation.

180

4 Dual Approach: Profit Function

> hessianHom[ 1, 2 ] <- hessianHom[ 2, 1 ] <- hphCapLab[1]
> hessianHom[ 1, 3 ] <- hessianHom[ 3, 1 ] <- hphCapMat[1]
> hessianHom[ 2, 3 ] <- hessianHom[ 3, 2 ] <- hphLabMat[1]
> hessianHom[ 1, 4 ] <- hessianHom[ 4, 1 ] <- hphCapOut[1]
> hessianHom[ 2, 4 ] <- hessianHom[ 4, 2 ] <- hphLabOut[1]
> hessianHom[ 3, 4 ] <- hessianHom[ 4, 3 ] <- hphMatOut[1]
> hessianHom[ 4, 4 ] <- hphOutOut[1]
> print( hessianHom )
[,1]

[,2]

[,3]

[,4]

[1,]

0.2198994186

0.315188197

0.0008740851 -1.30964735

[2,]

0.3151881974

2.114366248

0.0027786041 -4.16320062

[3,]

0.0008740851

0.002778604

0.0003198275 -0.01154546

[4,] -1.3096473550 -4.163200623 -0.0115454561 11.00022565

As all diagonal elements of this Hessian matrix are positive, the necessary conditions for positive semidefiniteness are fulfilled. Now, we calculate the principal minors in order to check the
sufficient conditions for positive semidefiniteness:
> hessianHom[1,1]
[1] 0.2198994
> det( hessianHom[1:2,1:2] )
[1] 0.3656043
> det( hessianHom[1:3,1:3] )
[1] 0.0001151481
> det( hessianHom )
[1] 3.268657e-19
The conditions for the first three principal minors are fulfilled and the fourth principal minor is
close to zero, where it is positive on some computers but negative on other computers. As Hessian
matrices of linear homogeneous functions are always singular, it is expected that the determinant
of the Hessian matrix (the N th principal minor) is zero. However, the computed determinant
of our Hessian matrix is not exactly zero due to rounding errors, which are unavoidable on
digital computers. Given that the determinant of the Hessian matrix of our Cobb-Douglas cost
function with linear homogeneity imposed should always be zero, the N th sufficient condition for
positive semidefiniteness (sign of the determinant of the Hessian matrix) should always be fulfilled.

181

4 Dual Approach: Profit Function

Consequently, we can conclude that our Cobb-Douglas profit function with linear homogeneity
imposed is convex in all prices at the first observation. In order to avoid problems due to rounding
errors, we can just check the positive semidefiniteness of the first N 1 rows and columns of the
Hessian matrix:
> semidefiniteness( hessianHom[1:3,1:3], positive = TRUE )
[1] TRUE
In the following, we will check whether convexity in all prices is fulfilled at each observation in
the sample:
> dat$convexCDHom <- NA
> for( obs in 1:nrow( dat ) ) {
+

hessianPart <- matrix( NA, nrow = 3, ncol = 3 )

hessianPart[ 1, 1 ] <- hphCapCap[obs]

hessianPart[ 2, 2 ] <- hphLabLab[obs]

hessianPart[ 3, 3 ] <- hphMatMat[obs]

hessianPart[ 1, 2 ] <- hessianPart[ 2, 1 ] <- hphCapLab[obs]

hessianPart[ 1, 3 ] <- hessianPart[ 3, 1 ] <- hphCapMat[obs]

hessianPart[ 2, 3 ] <- hessianPart[ 3, 2 ] <- hphLabMat[obs]

dat$convexCDHom[obs] <-

semidefiniteness( hessianPart, positive = TRUE )

+ }
> sum( !dat$convexCDHom, na.rm = TRUE )
[1] 0
This result indicates that the convexity condition is violated not at a single observation. Consequently, our Cobb-Douglas profit function with linear homogeneity imposed is convex in all prices
at all observations.

4.3.6 Predicted profit

As the dependent variable of the Cobb-Douglas profit function without homogeneity imposed is
ln(), we have to apply the exponential function to the fitted dependent variable, in order obtain
the fitted profit . Furthermore, we have to be aware of that the fitted method only returns
the predicted values for the observations that were included in the estimation. Hence, we have
to make sure that the predicted profits are only assigned to the observations that have a positive
profit and hence, were included in the estimation:
> dat$profitCD[ dat$profit > 0 ] <- exp( fitted( profitCD ) )

182

4 Dual Approach: Profit Function

We obtain the predicted profit from the Cobb-Douglas profit function with homogeneity imposed by:
> dat$profitCDHom[ dat$profit > 0 ] <- exp( fitted( profitCDHom ) ) * dat$pOut

4.3.7 Optimal Profit Shares

Given Hotellings Lemma, the coefficients of the (logarithmic) output price and the (logarithmic)
input prices are equal to the optimal profit shares derived from a Cobb-Douglas profit function:
ln (p, w)
p
(p, w)
p
p y(p, w)
=
= y(p, w)
=
r(p, w) 1
(4.34)
ln p
p
(p, w)
(p, w)
(w, y)
ln (p, w)
(p, w) wi
wi
wi xi (p, w)
i =
=
= xi (p, w)
=
ri (p, w) 0 (4.35)
ln wi
wi (p, w)
(p, w)
(w, y)

p =

In contrast to real shares, these profit shares are never between zero and one but they sum
up to one, as do real shares:
r+

py X
wi x i
+

i


ri =

py

i wi

xi

=1

(4.36)

For instance, an optimal profit share of the output of p = 2.75 means that profit maximization
would result in a total revenue that is 2.75 times as large as the profit, which corresponds to
a return on sales of 1/2.75 = 36%. Similarly, an optimal profit share of the capital input of
cap = 0.82 means that profit maximization would result in total capital costs that are 0.82
times as large as the profit.
The following commands draw histograms of the observed profit shares and compare them to
the optimal profit shares, which are predicted by our Cobb-Douglas profit function with linear
homogeneity imposed:
> hist( ( dat$pOut * dat$qOut / dat$profit )[
+

dat$profit > 0 ], 30 )

> lines( rep( ghOut, 2), c( 0, 100 ), lwd = 3

> hist( ( - dat$pCap * dat$qCap / dat$profit )[

+

dat$profit > 0 ], 30 )

> lines( rep( ghCap, 2), c( 0, 100 ), lwd = 3

> hist( ( - dat$pLab * dat$qLab / dat$profit )[

+

dat$profit > 0 ], 30 )

> lines( rep( ghLab, 2), c( 0, 100 ), lwd = 3

> hist( ( - dat$pMat * dat$qMat / dat$profit )[

+

dat$profit > 0 ], 30 )

> lines( rep( ghMat, 2), c( 0, 100 ), lwd = 3

The resulting graphs are shown in figure 4.3. These results somewhat contradict previous results.

183

60
0

20

40

Frequency

40
20
0

Frequency

60

80

80

4 Dual Approach: Profit Function

10

15

20

profit share cap

60
40
0

20

Frequency

40
20
0

Frequency

60

80

profit share out

profit share lab

profit share mat

Figure 4.3: Cobb-Douglas profit function: observed and optimal profit shares

184

4 Dual Approach: Profit Function

While the results based on production functions and cost functions indicate that the apple producers on average use too much capital and too few materials, the results of the Cobb-Douglas
profit function indicate that almost all apple producers use too much materials and most apple
producers use too less capital and labor. However, the results of the Cobb-Douglas profit function
are consistent with previous results regarding the output quantity: all results suggest that most
apple producers should produce more output.

4.3.8 Derived Output Supply Input Demand Functions

Hotellings Lemma says that the partial derivative of a profit function with respect to the output
price is the output supply function and that the partial derivatives of a profit function with
respect to the input prices are the negative (unconditional) input demand functions:
(p, w)
(p, w)
= p
p
p
(p, w)
(p, w)
= i
xi (p, w) =
wi
wi
y(p, w) =

(4.37)
(4.38)

These output supply and input demand functions should be homogeneous of degree zero in all
prices:
y(t p, t w) = y(p, w)

(4.39)

xi (t p, t w) = xi (p, w)

(4.40)

This condition is fulfilled for the output supply and input demand functions derived from a
linearly homogeneous Cobb-Douglas profit function:
(t p, t w)
t (p, w)
(p, w)
= p
= p
= y(p, w)
tp
tp
p
(t p, t w)
t (p, w)
(p, w)
xi (t p, t w) = i
= i
= i
= xi (p, w)
t wi
t wi
wi
y(t p, t w) = p

(4.41)
(4.42)

4.3.9 Derived Output Supply and Input Demand Elasticities

Based on the derived output supply function (4.37) and the derived input demand functions (4.38),
we can derive the output supply elasticities and the (unconditional) input demand elasticities:
y(p, w) p
p y(p, w)
p (p, w)
p
(p, w)
p
=
p
p
p
y(p, w)
p2
y(p, w)
p
p
(p, w)
=
y(p, w)
p
p
y(p, w)
p y(p, w)
p
= p
r(w, y)

yp (p, w) =

185

(4.43)
(4.44)
(4.45)
(4.46)

4 Dual Approach: Profit Function

= p 1

(4.47)

y(p, w) wj
wj y(p, w)
p (p, w) wj
=
p
wj
y(p, w)
wj
p
xj (p, w)
=
p
y(p, w)
(p, w) wj xj (p, w)
= p
p y(p, w) (p, w)
p rj (w, y)
=
r(w, y)

(4.50)

= j

(4.53)

yj (p, w) =

p
xi (p, w)
p
xi (p, w)
i (p, w)
p
=
wi
p
xi (p, w)
i
p
=
y(p, w)
wi
xi (p, w)
(p, w) p y(p, w)
= i
wi xi (p, w) (p, w)
i p
=
ri (w, y)

ip (p, w) =

= p

(4.48)
(4.49)

(4.51)
(4.52)

(4.54)
(4.55)
(4.56)
(4.57)
(4.58)
(4.59)

xi (p, w) wj
wj xi (p, w)
i (p, w)
wj
(p, w)
wj
=
+ ij i
2
wi
wj
xi (p, w)
xi (p, w)
wj

ij (p, w) =

i
wj
(p, w)
xj (p, w)
+ ij i
wi
xi (p, w)
wi xi (p, w)
(p, w) wj xj (p, w)
i
= i
ij
wi xi (p, w) (p, w)
ri (w, y)
i
i rj (w, y)
=
ij
ri (w, y)
ri (w, y)
=

= j ij

(4.60)
(4.61)
(4.62)
(4.63)
(4.64)
(4.65)

All derived input demand elasticities based on our Cobb-Douglas profit function with linear

186

4 Dual Approach: Profit Function

homogeneity imposed are presented in table 4.1. If the output price increases by one percent, the
profit-maximizing firm will increase the use of capital, labor, and materials by 2.75% each, which
increases the production by 1.75%. The proportional increase of the input quantities (+2.75%)
results in a less than proportional increase in the output quantity (+1.75%). This indicates
that the model exhibits decreasing returns to scale, which is not surprising, because a profit
maximum cannot be in an area of increasing returns to scale (if all inputs are variable and all
markets function perfectly). If the price of capital increases by one percent, the profit-maximizing
firm will decrease the use of capital by 1.82% and decrease the use of labor and materials by 0.82%
each, which decreases the production by 0.82%. If the price of labor increases by one percent,
the profit-maximizing firm will decrease the use of labor by 1.9% and decrease the use of capital
and materials by 0.9% each, which decreases the production by 0.9%. If the price of materials
increases by one percent, the profit-maximizing firm will decrease the use of materials by 1.02%
and decrease the use of capital and labor by 0.02% each, which will decrease the production by
0.02%.
Table 4.1: Output supply and input demand elasticities derived from Cobb-Douglas profit function (with linear homogeneity imposed)
p
wcap wlab wmat
y
1.75 -0.82 -0.9 -0.02
xcap 2.75 -1.82 -0.9 -0.02
xlab 2.75 -0.82 -1.9 -0.02
xmat 2.75 -0.82 -0.9 -1.02

4.4 Cobb-Douglas Short-Run Profit Function

4.4.1 Specification
The specification of a Cobb-Douglas short-run profit function is

v = A pp

wii

Y
iN 1

xj j ,

(4.66)

jN 2

This Cobb-Douglas short-run profit function can be linearized to

ln v = 0 + p ln p +

X
iN 1

with 0 = ln A.

187

i ln wi +

X
jN 2

j ln xj

(4.67)

4 Dual Approach: Profit Function

4.4.2 Estimation
We can estimate a Cobb-Douglas short-run profit function with capital as a quasi-fixed input
using the following commands. Again, function lm automatically removes the observations (apple
producers) with negative gross margin:
> profitCDSR <- lm( log( vProfit ) ~ log( pOut ) + log( pLab ) + log( pMat ) +
+

log( qCap ), data = dat )

> summary( profitCDSR )

Call:
lm(formula = log(vProfit) ~ log(pOut) + log(pLab) + log(pMat) +
log(qCap), data = dat)
Residuals:
Min

1Q

Median

3Q

Max

-4.7422 -0.0646

0.2578

0.4931

0.8989

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept)

3.2739

1.2261

2.670 0.008571 **

log(pOut)

3.1745

0.2263

14.025

log(pLab)

-1.6188

0.4434

-3.651 0.000381 ***

log(pMat)

-0.7637

0.2687

-2.842 0.005226 **

log(qCap)

1.0960

0.1245

< 2e-16 ***

8.802 8.31e-15 ***

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Residual standard error: 0.9659 on 127 degrees of freedom

(8 observations deleted due to missingness)
Multiple R-squared:
F-statistic:

0.6591,

Adjusted R-squared:

61.4 on 4 and 127 DF,

p-value: < 2.2e-16

0.6484

4.4.3 Properties
This short-run profit function fulfills all microeconomic monotonicity conditions: it is increasing
in the output price, it is decreasing in the prices of all variable inputs, and it is increasing in the
quasi-fixed input quantity. However, the homogeneity condition is not fulfilled, as the coefficient
of the output price and the coefficients of the prices of the variable inputs do not sum up to one
but to 3.17 + (1.62) + (0.76) = 0.79.

188

4 Dual Approach: Profit Function

4.4.4 Estimation with linear homogeneity in all prices imposed

We can impose the homogeneity condition on the Cobb-Douglas short-run profit function using
the same method as for the Cobb-Douglas (long-run) profit function:
> profitCDSRHom <- lm( log( vProfit / pOut ) ~ log( pLab / pOut ) +
+

log( pMat / pOut ) + log( qCap ), data = dat )

> summary( profitCDSRHom )

Call:
lm(formula = log(vProfit/pOut) ~ log(pLab/pOut) + log(pMat/pOut) +
log(qCap), data = dat)
Residuals:
Min

1Q

Median

3Q

Max

-4.7302 -0.0677

0.2598

0.5160

0.8916

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept)

3.3145

1.2184

log(pLab/pOut)

-1.4574

0.2252

-6.471 1.88e-09 ***

log(pMat/pOut)

-0.7156

0.2427

-2.949

1.0847

0.1212

log(qCap)

2.720

0.00743 **
0.00380 **

8.949 3.50e-15 ***

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Residual standard error: 0.9628 on 128 degrees of freedom

(8 observations deleted due to missingness)
Multiple R-squared:

0.5227,

F-statistic: 46.73 on 3 and 128 DF,

Adjusted R-squared:

0.5115

p-value: < 2.2e-16

We can obtain the coefficient of the output price from the homogeneity condition (4.15): 1
(1.457) (0.716) = 3.173. All microeconomic monotonicity conditions are still fulfilled: the
Cobb-Douglas short-run profit function with homogeneity imposed is increasing in the output
price, decreasing in the prices of all variable inputs, and increasing in the quasi-fixed input
quantity.
We can test the homogeneity restriction by a likelihood ratio test:
> lrtest( profitCDSRHom, profitCDSR )
Likelihood ratio test

189

4 Dual Approach: Profit Function

Model 1: log(vProfit/pOut) ~ log(pLab/pOut) + log(pMat/pOut) + log(qCap)
Model 2: log(vProfit) ~ log(pOut) + log(pLab) + log(pMat) + log(qCap)
#Df

LogLik Df

5 -180.27

6 -180.17

Chisq Pr(>Chisq)

1 0.1859

0.6664

Given the large P -value, we can conclude that the data do not contradict the linear homogeneity
in the output price and the prices of the variable inputs.

4.4.5 Returns to scale

The sum over the coefficients of all quasi-fixed inputs indicates the percentage change of the gross
margin if the quantities of all quasi-fixed inputs are increased by one percent: If this sum is larger
than one, the increase in gross margin is more than proportional to the increase in the quasi-fixed
inputs. Hence, the technology has increasing returns to scale. If this sum over the coefficients of
all quasi-fixed inputs is smaller than one, the increase in gross margin is less than proportional
to the increase in the quasi-fixed inputs and the technology has decreasing returns to scale. As
the coefficient of our (single) quasi-fixed input is larger than one (1.085), we can conclude that
the technology has increasing returns to scale.

4.4.6 Shadow prices of quasi-fixed inputs

The partial derivatives of the short-run profit function with respect to the quantities of the
quasi-fixed inputs denote the additional gross margins that can be earned by an additional unit
of these quasi-fixed inputs. These internal marginal values of the quasi-fixed inputs are usually
called shadow prices. In case of the Cobb-Douglas short-run profit function, the shadow prices
can be computed by

v
ln v v
v
=
= j
xj
ln xj xj
xj

(4.68)

Before we can calculate the shadow price of the capital input, we need to calculate the predicted
gross margin v . As the dependent variable of the Cobb-Douglas short-run profit function with
homogeneity imposed is ln( v / ln p), we have to apply the exponential function to the fitted
dependent variable and then we have to multiply the result with p, in order obtain the fitted
gross margins v . Furthermore, we have to be aware of that the fitted method only returns
the predicted values for the observations that were included in the estimation. Hence, we have
to make sure that the predicted gross margins are only assigned to the observations that have a
positive gross margin and hence, were included in the estimation:
> dat$vProfitCDHom[ dat$vProfit > 0 ] <+

exp( fitted( profitCDSRHom ) ) * dat$pOut[ dat$vProfit > 0 ]

Now, we can calculate the shadow price of the capital input for each apple producer who has a
positive gross margin and hence, was included in the estimation:

190

4 Dual Approach: Profit Function

> dat$pCapShadow <- with( dat, coef(profitCDSRHom)["log(qCap)"] *
+

vProfitCDHom / qCap )

The following commands show the variation of the shadow prices of capital and compare them
to the observed capital prices:
> hist( dat$pCapShadow, 30 )
> hist( dat$pCapShadow[ dat$pCapShadow < 30 ], 30 )

100

200

300

400

shadow price of capital

5.0

50.0

0.2

1.0

shadow prices

6
0

Frequency

40
30
20
10

Frequency

50

10

60

500.0

> compPlot( dat$pCap, dat$pCapShadow, log = "xy" )

10

15

20

25

30

shadow price of capital

0.2

1.0

5.0 20.0

200.0

observed prices

Figure 4.4: Shadow prices of capital

The resulting graphs are shown in figure 4.4. The two histograms show that most shadow prices
are below 30 and many shadow prices are between 3 and 11 but there are also some apple
producers who would gain much more from increasing their capital input. Indeed, all apple
producers have a higher shadow price of capital than the observed price of capital, where the
difference is small for some producers and large for other producers. These differences can be
explained by risk aversion and market failures on the credit market or land market (e.g. marginal
prices are not equal to average prices).

191

5 Stochastic Frontier Analysis

5.1 Theory
5.1.1 Different Efficiency Measures
5.1.1.1 Output-Oriented Technical Efficiency with One Output
The output-oriented technical efficiency according to Shepard is defined as
TE =

y
y

y = T E y

0 T E 1,

(5.1)

where y is the observed output quantity and y is the maximum output quantity that can be
produced with the observed input quantities x.
The output-oriented technical efficiency according to Farrell is defined as
y
y

TE =

y = T E y

T E 1.

(5.2)

These efficiency measures are graphically illustrated in Bogetoft and Otto (2011, p. 26, figure 2.2).
5.1.1.2 Input-Oriented Technical Efficiency with One Input
The input-oriented technical efficiency according to Shepard is defined as
TE =

x
x

x = T E x

T E 1,

(5.3)

where x is the observed input quantity and x is the minimum input quantity at which the
observed output quantities y can be produced.
The input-oriented technical efficiency according to Farrell is defined as
TE =

x
x

x = T E x

0 T E 1.

(5.4)

These efficiency measures are graphically illustrated in Bogetoft and Otto (2011, p. 26, figure 2.2).

192

5 Stochastic Frontier Analysis

5.1.1.3 Output-Oriented Technical Efficiency with Two or More Outputs
The output-oriented technical efficiencies according to Shepard and Farrell assume a proportional
increase of all output quantities, while all input quantities are held constant.
Hence, the output-oriented technical efficiency according to Shepard is defined as
TE =

y1
y2
yM
= = ... =

y1
y2
yM

yi = T E yi i

0 T E 1,

(5.5)

are the maximum output

where y1 , y2 , . . . , yM are the observed output quantities, y1 , y2 , . . . , yM

quantities (given a proportional increase of all output quantities) that can be produced with the
observed input quantities x, and M is the number of outputs.
The output-oriented technical efficiency according to Farrell is defined as
TE =

y1
y
y
= 2 = ... = M
y1
y2
yM

yi = T E yi i

T E 1.

(5.6)

These efficiency measures are graphically illustrated in Bogetoft and Otto (2011, p. 27, figure 2.3, right panel).
5.1.1.4 Input-Oriented Technical Efficiency with Two or More Inputs
The input-oriented technical efficiencies according to Shepard and Farrell assume a proportional
reduction of all inputs, while all outputs are held constant.
Hence, the input-oriented technical efficiency according to Shepard is defined as
TE =

x1
x2
xN
= = ... =

x1
x2
xN

xi = T E xi i

TE 1

(5.7)

where x1 , x2 , . . . , xN are the observed input quantities, x1 , x2 , . . . , xN are the minimum input
quantities (given a proportional decrease of all input quantities) at which the observed output
quantities y can be produced, and N is the number of inputs.
The input-oriented technical efficiency according to Farrell is defined as
TE =

x1
x
x
= 2 = ... = N
x1
x2
xN

xi = T E xi i

0 T E 1.

(5.8)

These efficiency measures are graphically illustrated in Bogetoft and Otto (2011, p. 27, figure 2.3, left panel).
5.1.1.5 Output-Oriented Allocative Efficiency and Revenue Efficiency
According to equation (5.6), the output-oriented technical efficiency according to Farrell is
TE =

y2
yM
p y
y1
=
= ... =
=
,
y1
y2
yM
py

193

(5.9)

5 Stochastic Frontier Analysis

where y is the vector of technically efficient output quantities and p is the vector of output prices.
The output-oriented allocative efficiency according to Farrell is defined as
AE =

p y
p y
=
,
p y
p y

(5.10)

where y is the vector of technically efficient and allocatively efficient output quantities and y is
the vector of output quantities so that p y = p y and yi /
yi = AE i.
Finally, the revenue efficiency according to Farrell is
RE =

p y
p y p y
=
= AE T E
py
p y p y

(5.11)

All these efficiency measures can also be specified according to Shepard by just taking the inverse of the Farrell specifications. These efficiency measures are graphically illustrated in Bogetoft
and Otto (2011, p. 40, figure 2.11).
5.1.1.6 Input-Oriented Allocative Efficiency and Cost Efficiency
According to equation (5.8), the input-oriented technical efficiency according to Farrell is
TE =

x
1
wx

x
2
x
N
=
,
=
= ... =
x1
x2
xN
wx

(5.12)

where x
is the vector of technically efficient input quantities and w is the vector of output prices.
The input-oriented allocative efficiency according to Farrell is defined as
AE =

w x
wx

=
,
wx

wx

(5.13)

where x is the vector of technically efficient and allocatively efficient input quantities and x
is
the vector of output quantities so that w x
= w x and x
i /
xi = AE i.
Finally, the cost efficiency according to Farrell is
CE =

w x
w x w x

=
= AE T E
wx
wx
wx

(5.14)

All these efficiency measures can also be specified according to Shepard by just taking the inverse of the Farrell specifications. These efficiency measures are graphically illustrated in Bogetoft
and Otto (2011, p. 36, figure 2.9).
5.1.1.7 Profit Efficiency
The profit efficiency according to Farrell is defined as
PE =

p y w x
,
pyw x

194

(5.15)

5 Stochastic Frontier Analysis

where y and x denote the profit maximizing output quantities and input quantities, respectively
(assuming full technical efficiency). The profit efficiency according to Shepard is just the inverse
of the Farrell specifications.
5.1.1.8 Scale efficiency
In case of one input x and one output y = f (x), the scale efficiency according to Farrell is defined
as
SE =

AP
,
AP

(5.16)

where AP = f (x)/x is the observed average product AP = f (x )/x is the maximum average
product, and x is the input quantity that results in the maximum average product.
The first-order condition for a maximum of the average product is
AP
f (x) 1 f (x)
=
2 =0
x
x x
x

(5.17)

This condition can be re-arranged to

f (x) x
=1
x f (x)

(5.18)

Hence, a necessary (but not sufficient) condition for a maximum of the average product is an
elasticity of scale equal to one.

5.2 Stochastic Production Frontiers

5.2.1 Specification
In section 2, we have estimated average production functions, where about half of the observations
were below the estimated production function and about half of the observations were above the
estimated production function (see left panel of figure 5.1). However, in microeconomic theory,
the production function indicates the maximum output quantity for each given set of input
quantities. Hence, theoretically, no observation could be above the production function and an
observations below the production function would indicate technical inefficiency.
This means that all residuals must be negative or zero. A production function with only
non-positive residuals could look like:
ln y = ln f (x) u with u 0,

(5.19)

where u 0 are the non-positive residuals. One solution to achieve this could be to estimate
an average production function by ordinary least squares and then simply shift the production
function up until all residuals are negative or zero (see right panel of figure 5.1). However, this

195

5 Stochastic Frontier Analysis

y

o
o
o

o
o o

o o

o
o
o

o
o
o

Source: Bogetoft and Otto (2011)

Figure 5.1: Production function estimation: ordinary regression and with intercept correction
procedure does not account for statistical noise and is very sensitive to positive outliers.1 As
virtually all data sets and models are flawed with statistical noise, e.g. due to measurement
errors, omitted variables, and approximation errors, Meeusen and van den Broeck (1977) and
Aigner, Lovell, and Schmidt (1977) independently proposed the stochastic frontier model that
simultaneously accounts for statistical noise and technical inefficiency:
ln y = ln f (x) u + v

with u 0,

(5.20)

where u 0 accounts for technical inefficiency and v accounts for statistical noise. This model
can be re-written (see, e.g. Coelli et al., 2005, p. 243):
y = f (x) eu ev

(5.21)

Output-oriented technical efficiencies are usually defined as the ratio between the observed
output and the (individual) stochastic frontier output (see, e.g. Coelli et al., 2005, p. 244):
TE =

y
f (x) eu ev
=
= eu
f (x) ev
f (x) ev

(5.22)

The stochastic frontier model is usually estimated by an econometric maximum likelihood

estimation, which requires distributional assumptions of the error terms. Most often, it is assumed
that the noise term v follows a normal distribution with zero mean and constant variance v2 ,
the inefficiency term u follows a positive half-normal distribution or a positive truncated normal

This is also true for the frequently-used Data Envelopment Analysis (DEA).

196

5 Stochastic Frontier Analysis

distribution with constant scale parameter u2 , and all vs and all us are independent:
v N (0, v2 )

(5.23)

u N + (, u2 ),

(5.24)

where = 0 for a positive half-normal distribution and 6= 0 for a positive truncated normal
distribution. These assumptions result in a left-skewed distribution of the total error terms =
u + v, i.e. the density function is flat on the left and steep on the right. Hence, it is very rare
that a firm has a large positive residual (much higher output than the production function) but
it is not so rare that a firm has a large negative residual (much lower output than the production
function).
5.2.1.1 Marginal products and output elasticities in SFA models
Given the multiplicative specification of stochastic production frontier models (5.21) and assuming
that the random error v is zero, we can see that the marginal products are downscaled by the
level of the technical efficiency:
y
f (x) u
f (x)
f (x)
=
e = TE
= T E i
xi
xi
xi
xi

(5.25)

However, the partial production elasticities are unaffected by the efficiency level:
y xi
f (x) u xi
f (x) xi
ln f (x)
=
e
=
=
= i
xi y
xi
f (x)eu
xi f (x)
ln xi

(5.26)

As the output elasticities do not depend on the firms technical efficiency, also the elasticity of
scale does not depend on the firms technical efficiency.

5.2.2 Skewness of residuals from OLS estimations

The following commands plot histograms of the residuals taken from the Cobb-Douglas and the
Translog production function:
> hist( residuals( prodCD ), 15 )
> hist( residuals( prodTL ), 15 )
The resulting graphs are shown in figure 5.2. The residuals of both production functions are
left-skewed. This visual assessment of the skewness can be confirmed by calculating the skewness
using the function skewness that is available in the package moments:
> library( "moments" )
> skewness( residuals( prodCD ) )
[1] -0.4191323

197

10 15 20
0

Frequency

10 15 20
5
0

Frequency

5 Stochastic Frontier Analysis

1.5

0.5

0.5 1.0 1.5

1.5

residuals prodCD

0.5

0.5 1.0 1.5

residuals prodTL

Figure 5.2: Residuals of Cobb-Douglas and Translog production functions

> skewness( residuals( prodTL ) )
[1] -0.3194211
As a negative skewness means that the residuals are left-skewed, it is likely that not all apple
producers are fully technically efficient.
However, the distribution of the residuals does not always have the expected skewness. Possible
reasons for an unexpected skewness of OLS residuals are explained in section 5.3.2.

5.2.3 Cobb-Douglas Stochastic Production Frontier

We can use the command sfa (package frontier) to estimate stochastic production frontiers. The
basic syntax of the command sfa is similar to the syntax of the command lm. The following
command estimates a Cobb-Douglas stochastic production frontier assuming that the inefficiency
term u follows a positive halfnormal distribution:
> library( "frontier" )
> prodCDSfa <- sfa( log( qOut ) ~ log( qCap ) + log( qLab ) + log( qMat ),
+

data = dat )

> summary( prodCDSfa )

Error Components Frontier (see Battese & Coelli 1992)
Inefficiency decreases the endogenous variable (as in a production function)
The dependent variable is logged
Iterative ML estimation terminated after 12 iterations:
log likelihood values and parameters of two successive iterations
are within the tolerance limit
final maximum likelihood estimates

198

5 Stochastic Frontier Analysis

Estimate Std. Error z value

Pr(>|z|)

(Intercept) 0.228813

1.247739

0.1834 0.8544981

log(qCap)

0.160934

0.081883

1.9654 0.0493668 *

log(qLab)

0.684777

0.146797

4.6648 3.089e-06 ***

log(qMat)

0.465871

0.131588

3.5404 0.0003996 ***

sigmaSq

1.000040

0.202456

4.9395 7.830e-07 ***

gamma

0.896664

0.070952 12.6375 < 2.2e-16 ***

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

log likelihood value: -133.8893

cross-sectional data
total number of observations = 140
mean efficiency: 0.5379937
The parameters of the Cobb-Douglas production frontier can be interpreted as before. The
estimated production function is monotonically increasing in all inputs. The output elasticity of
capital is 0.161, the output elasticity of labor is 0.685, The output elasticity of materials is 0.466,
and the elasticity of scale is 1.312.
The estimation algorithm re-parameterizes the variance parameter of the noise term (v2 ) and
the scale parameter of the inefficiency term (u2 ) and instead estimates the parameters 2 = v2 +u2
and = u2 / 2 . The parameter lies between zero and one and indicates the importance of the
inefficiency term. If is zero, the inefficiency term u is irrelevant and the results should be equal
to OLS results. In contrast, if is one, the noise term v is irrelevant and all deviations from the
production frontier are explained by technical inefficiency. As the estimate of is 0.897, we can
conclude that both statistical noise and inefficiency are important for explaining deviations from
the production function but that inefficiency is more important than noise. As u2 is not equal to
the variance of the inefficiency term u, the estimated parameter cannot be interpreted as the
proportion of the total variance that is due to inefficiency. In fact, the variance of the inefficiency
term u is

V ar(u) = u2 1

 2

  ,

(5.27)

where (.) indicates the cumulative distribution function and (.) the probability density function
of the standard normal distribution. If the inefficiency term u follows a positive halfnormal
distribution (i.e. = 0), the above equation reduces to
h

V ar(u) = u2 1 (2 (0))2 ,

199

(5.28)

5 Stochastic Frontier Analysis

We can calculate the estimated variances of the inefficiency term u and the noise term v by
following commands:
> gamma <- unname( coef(prodCDSfa)["gamma"] )
[1] 0.8966641
> sigmaSq <- unname( coef(prodCDSfa)["sigmaSq"] )
[1] 1.00004
> sigmaSqU <- gamma * sigmaSq
[1] 0.8966997
> varU <- sigmaSqU * ( 1 - ( 2 * dnorm(0) )^2 )
[1] 0.3258429
> varV <- sigmaSqV <- ( 1 - gamma ) * sigmaSq
[1] 0.10334
Hence, the proportion of the total variance (V ar(u + v) = V ar(u) + V ar(v))2 that is due to
inefficiency is estimated to be:
> varU / ( varU + varV )
[1] 0.7592169
This indicates that around 75.9% of the total variance is due to inefficiency.
The frontier package calculates these additonal variance parameters (and some further variance
parameters) automatically, if argument extraPar of the summary() method is set to TRUE:
> summary( prodCDSfa, extraPar = TRUE )
Error Components Frontier (see Battese & Coelli 1992)
Inefficiency decreases the endogenous variable (as in a production function)
The dependent variable is logged
Iterative ML estimation terminated after 12 iterations:
log likelihood values and parameters of two successive iterations
are within the tolerance limit
2

This equation relies on the assumption that the inefficiency term u and the noise term v are independent, i.e.
their covariance is zero.

200

5 Stochastic Frontier Analysis

final maximum likelihood estimates
Estimate Std. Error z value

Pr(>|z|)

(Intercept) 0.228813

1.247739

0.1834 0.8544981

log(qCap)

0.160934

0.081883

1.9654 0.0493668 *

log(qLab)

0.684777

0.146797

4.6648 3.089e-06 ***

log(qMat)

0.465871

0.131588

3.5404 0.0003996 ***

sigmaSq

1.000040

0.202456

4.9395 7.830e-07 ***

gamma

0.896664

0.070952 12.6375 < 2.2e-16 ***

sigmaSqU

0.896700

0.241715

3.7097 0.0002075 ***

sigmaSqV

0.103340

0.055831

1.8509 0.0641777 .

sigma

1.000020

0.101226

9.8791 < 2.2e-16 ***

sigmaU

0.946942

0.127629

7.4195 1.176e-13 ***

sigmaV

0.321465

0.086838

3.7019 0.0002140 ***

lambdaSq

8.677179

6.644542

1.3059 0.1915829

lambda

2.945705

1.127836

2.6118 0.0090061 **

varU

0.325843

NA

NA

NA

sdU

0.570827

NA

NA

NA

gammaVar

0.759217

NA

NA

NA

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

log likelihood value: -133.8893

cross-sectional data
total number of observations = 140
mean efficiency: 0.5379937
The additionally returned parameter are defined as follows: sigmaSqU = u2 = 2 , sigmaSqV

p
p
= v2 = 2 (1 ) = Var (v), sigma = = 2 , sigmaU = u = u2 , sigmaV = v = v2 ,
lambdaSq = 2 = u2 /v2 , lambda = = u /v , varU = Var (u), sdU =

Var (u), and gammaVar

= Var (u)/(Var (u) + Var (v)).

The t-test for the coefficient (e.g. reported in the output of the summary method) is not valid,
because is bound to the interval [0, 1] and hence, cannot follow a t-distribution. However, we can
use a likelihood ratio test to check whether adding the inefficiency term u significantly improves
the fit of the model. If the lrtest method is called just with a single stochastic frontier model,
it compares the stochastic frontier model with the corresponding OLS model (i.e. a model with
equal to zero):
> lrtest( prodCDSfa )

201

5 Stochastic Frontier Analysis

Likelihood ratio test
Model 1: OLS (no inefficiency)
Model 2: Error Components Frontier (ECF)
#Df

LogLik Df

5 -137.61

6 -133.89

Chisq Pr(>Chisq)

1 7.4387

0.003192 **

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Under the null hypothesis (no inefficiency, only noise), the test statistic asymptotically follows a
mixed 2 -distribution (Coelli, 1995).3 The rather small P-value indicates that the data clearly
reject the OLS model in favor of the stochastic frontier model, i.e. there is significant technical
inefficiency.
As neither the noise term v nor the inefficiency term u but only the total error term = u + v
is known, the technical efficiencies T E = eu are generally unknown. However, given that the
parameter estimates (including the parameters 2 and or v2 and u2 ) and the total error term
are known, it is possible to determine the expected value of the technical efficiency (see, e.g.
Coelli et al., 2005, p. 255):
Td
E = E eu


(5.29)

These efficiency estimates can be obtained by the efficiencies method:

> dat$effCD <- efficiencies( prodCDSfa )
Now, we visualize the variation of the efficiency estimates using a histogram and we explore
the correlation between the efficiency estimates and the output as well as the firm size (measured
as aggregate input use by a Fisher quantity index of all inputs):
> hist( dat$effCD, 15 )
> plot( dat$qOut, dat$effCD, log = "x" )
> plot( dat$X, dat$effCD, log = "x" )
The resulting graphs are shown in figure 5.3. The efficiency estimates are rather low: the firms
only produce between 10% and 90% of the maximum possible output quantities. As the efficiency
directly influences the output quantity, it is not surprising that the efficiency estimates are highly
correlated with the output quantity. On the other hand, the efficiency estimates are only slightly
correlated with firm size. However, the largest firms all have an above-average efficiency estimate,
while only a very few of the smallest firms have an above-average efficiency estimate.
3

As a standard likelihood ratio test assumes that the test statistic follows a (standard) 2 -distribution under the
null hypothesis, a test that is conducted by the command lrtest( prodCD, prodCDSfa ) returns an incorrect
P-value.

202

0.2

0.4

0.6

0.8

1e+05

5e+05

effCD

0.8

0.4

0.6

0.2

0.2
0

effCD

0.6
0.4

effCD

10
5

Frequency

15

0.8

5 Stochastic Frontier Analysis

5e+06

0.5

1.0

qOut

2.0

5.0

Figure 5.3: Efficiency estimates of Cobb-Douglas production frontier

5.2.4 Translog Production Frontier

As the Cobb-Douglas functional form is very restrictive, we additionally estimate a Translog
stochastic production frontier:
> prodTLSfa <- sfa( log( qOut ) ~ log( qCap ) + log( qLab ) + log( qMat )
+

+ I( 0.5 * log( qCap )^2 ) + I( 0.5 * log( qLab )^2 )

+ I( 0.5 * log( qMat )^2 ) + I( log( qCap ) * log( qLab ) )

+ I( log( qCap ) * log( qMat ) ) + I( log( qLab ) * log( qMat ) ),

data = dat )

> summary( prodTLSfa, extraPar = TRUE )

Error Components Frontier (see Battese & Coelli 1992)
Inefficiency decreases the endogenous variable (as in a production function)
The dependent variable is logged
Iterative ML estimation terminated after 23 iterations:
log likelihood values and parameters of two successive iterations
are within the tolerance limit
final maximum likelihood estimates
Estimate Std. Error z value

Pr(>|z|)

(Intercept)

-8.8110424 19.9181626 -0.4424 0.6582271

log(qCap)

-0.6332521

log(qLab)

4.4511064

log(qMat)

-1.3976309

2.0855273 -0.3036 0.7614012

4.4552358

0.9991 0.3177593

3.8097807 -0.3669 0.7137284

I(0.5 * log(qCap)^2)

0.0053258

I(0.5 * log(qLab)^2)

-1.5030433

0.6812813 -2.2062 0.0273700 *

I(0.5 * log(qMat)^2)

-0.5113559

0.3733348 -1.3697 0.1707812

I(log(qCap) * log(qLab))

0.4187529

0.1866174

0.2747251

203

0.0285 0.9772324

1.5243 0.1274434

5 Stochastic Frontier Analysis

I(log(qCap) * log(qMat)) -0.4371561

0.1902856 -2.2974 0.0215978 *

I(log(qLab) * log(qMat))

0.9800294

0.4216637

2.3242 0.0201150 *

sigmaSq

0.9587307

0.1968009

4.8716 1.107e-06 ***

gamma

0.9153387

0.0647478 14.1370 < 2.2e-16 ***

sigmaSqU

0.8775633

0.2328364

3.7690 0.0001639 ***

sigmaSqV

0.0811674

0.0497448

1.6317 0.1027476

sigma

0.9791480

0.1004960

9.7432 < 2.2e-16 ***

sigmaU

0.9367835

0.1242744

7.5380 4.771e-14 ***

sigmaV

0.2848989

0.0873025

3.2634 0.0011010 **

10.8117751

9.0334816

1.1969 0.2313628

lambda

3.2881264

1.3736518

2.3937 0.0166789 *

varU

0.3188892

NA

NA

NA

sdU

0.5647027

NA

NA

NA

gammaVar

0.7971103

NA

NA

NA

lambdaSq

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

log likelihood value: -128.0684

cross-sectional data
total number of observations = 140
mean efficiency: 0.5379939
A likelihood ratio test confirms that the stochastic frontier model fits the data much better than
an average production function estimated by OLS:
> lrtest( prodTLSfa )
Likelihood ratio test
Model 1: OLS (no inefficiency)
Model 2: Error Components Frontier (ECF)
#Df

LogLik Df Chisq Pr(>Chisq)

11 -131.25

12 -128.07

1 6.353

0.005859 **

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

A further likelihood ratio test indicates that it is not really clear whether the Translog stochastic
frontier model fits the data significantly better than the Cobb-Douglas stochastic frontier model:
> lrtest( prodCDSfa, prodTLSfa )

204

5 Stochastic Frontier Analysis

Likelihood ratio test
Model 1: prodCDSfa
Model 2: prodTLSfa
#Df

LogLik Df

6 -133.89

12 -128.07

Chisq Pr(>Chisq)

6 11.642

0.07045 .

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

While the Cobb-Douglas functional form is accepted at the 5% significance level, it is rejected in
favor of the Translog functional form at the 10% significance level.
The efficiency estimates based on the Translog stochastic production frontier can be obtained
(again) by the efficiencies method:
> dat$effTL <- efficiencies( prodTLSfa )
The following commands illustrate their variation, their correlation with the output level, and
their correlation with the firm size (measured as input use):
> hist( dat$effTL, 15 )
> plot( dat$qOut, dat$effTL, log = "x" )

0.2

0.4

0.6

0.8

1e+05

5e+05

effTL

5e+06

0.8

0.4

0.6

0.2

0.2

effTL

0.6
0.4

effTL

8
6
4

Frequency

10

0.8

12

> plot( dat$X, dat$effTL, log = "x" )

0.5

qOut

1.0

2.0

5.0

Figure 5.4: Efficiency estimates of Translog production frontier

The resulting graphs are shown in figure 5.4. These efficiency estimates are rather similar to the
efficiency estimates based on the Cobb-Douglas stochastic production frontier. This is confirmed
by a direct comparison of these efficiency estimates:
> compPlot( dat$effCD, dat$effTL )

205

0.6

0.8

5 Stochastic Frontier Analysis

effTL

0.2

0.4

0.2

0.4

0.6

0.8

effCD

Figure 5.5: Efficiency estimates of Cobb-Douglas and Translog production frontier

The resulting graph is shown in figure 5.5. Most efficiency estimates only slightly differ between
the two functional forms but a few efficiency estimates are considerably higher for the Translog
functional form. The inflexibility of the Cobb-Douglas functional form probably resulted in an
insufficient adaptation of the frontier to some observations, which lead to larger negative residuals
and hence, lower efficiency estimates in the Cobb-Douglas model.

5.2.5 Translog Production Frontier with Mean-Scaled Variables

As argued in section 2.6.15, it is sometimes convenient to estimate a Translog production (frontier)
function with mean-scaled variables. The following command estimates a Translog production
function with mean-scaled output and input quantities:
> prodTLmSfa <- sfa( log( qmOut ) ~ log( qmCap ) + log( qmLab ) + log( qmMat )
+

+ I( 0.5 * log( qmCap )^2 ) + I( 0.5 * log( qmLab )^2 )

+ I( 0.5 * log( qmMat )^2 ) + I( log( qmCap ) * log( qmLab ) )

+ I( log( qmCap ) * log( qmMat ) ) + I( log( qmLab ) * log( qmMat ) ),

data = dat )

> summary( prodTLmSfa )

Error Components Frontier (see Battese & Coelli 1992)
Inefficiency decreases the endogenous variable (as in a production function)
The dependent variable is logged
Iterative ML estimation terminated after 17 iterations:
log likelihood values and parameters of two successive iterations

206

5 Stochastic Frontier Analysis

are within the tolerance limit
final maximum likelihood estimates
Estimate Std. Error z value

Pr(>|z|)

(Intercept)

0.6388793

0.1311531

4.8712 1.109e-06 ***

log(qmCap)

0.1308903

0.1003318

1.3046

log(qmLab)

0.7065404

0.1555606

4.5419 5.575e-06 ***

log(qmMat)

0.4657266

0.1516483

3.0711

0.002133 **

I(0.5 * log(qmCap)^2)

0.0053227

0.1848995

0.0288

0.977034

I(0.5 * log(qmLab)^2)

-1.5030266

0.6761522 -2.2229

0.026222 *

I(0.5 * log(qmMat)^2)

-0.5113617

0.3749803 -1.3637

0.172661

0.2686428

0.119047

I(log(qmCap) * log(qmLab))

0.4187571

1.5588

0.192038

I(log(qmCap) * log(qmMat)) -0.4371473

0.1886950 -2.3167

0.020521 *

I(log(qmLab) * log(qmMat))

0.9800162

0.4201674

2.3324

0.019677 *

sigmaSq

0.9587158

0.1967744

4.8722 1.104e-06 ***

gamma

0.9153349

0.0659588 13.8774 < 2.2e-16 ***

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

log likelihood value: -128.0684

cross-sectional data
total number of observations = 140
mean efficiency: 0.5379969
> all.equal( coef( prodTLmSfa )[-c(1:4)], coef( prodTLmSfa )[-c(1:4)] )
[1] TRUE
> all.equal( efficiencies( prodTLmSfa ), efficiencies( prodTLSfa ) )
[1] "Mean relative difference: 7.059684e-06"
While the intercept and the first-order parameters have adjusted to the new units of measurement, the second-order parameters, the variance parameters, and the efficiency estimates remain (nearly) unchanged. From the estimated coefficients of the Translog production frontier
with mean-scaled input quantities, we can immediately see that the monotonicity condition is
fulfilled at the sample mean, that the output elasticities of capital, labor, and materials are
0.131, 0.707, and 0.466, respectively, at the sample mean, and that the elasticity of scale is
0.131 + 0.707 + 0.466 = 1.303 at the sample mean.

207

5 Stochastic Frontier Analysis

5.3 Stochastic Cost Frontiers

5.3.1 Specification
The general specification of a stochastic cost frontier is
ln c = ln c(w, y) + u + v

with u 0,

(5.30)

where u 0 accounts for cost inefficiency and v accounts for statistical noise. This model can be
re-written as:
c = c(w, y) eu ev

(5.31)

The cost efficiency according to Shepard is

CE =

c
f (x) eu ev
=
= eu ,
v
c(w, y) e
c(w, y) ev

(5.32)

while the cost efficiency according to Farrell is

CE =

c(w, y) ev
c(w, y) ev
=
= eu .
c
f (x) eu ev

(5.33)

Assuming a normal distribution of the noise term v and a positive half-normal distribution of
the inefficiency term u, the distribution of the residuals from a cost function is expected to be
right-skewed in the case of cost inefficiencies.

5.3.2 Skewness of residuals from OLS estimations

The following commands visualize the distribution of the residuals of the OLS estimations of the
Cobb-Douglas and Translog cost functions with linear homogeneity in input prices imposed:
> hist( residuals( costCDHom ) )
> hist( residuals( costTLHom ) )
The resulting graphs are shown in figure 5.6. The distributions of the residuals look approximately
symmetric and rather a little left-skewed than right-skewed (although we expected the latter).
This visual assessment of the skewness can be confirmed by calculating the skewness using the
function skewness that is available in the package moments:
> library( "moments" )
> skewness( residuals( costCDHom ) )
[1] -0.05788105
> skewness( residuals( costTLHom ) )
[1] -0.03709506

208

10 20 30 40

Frequency

15
0 5

Frequency

25

5 Stochastic Frontier Analysis

0.5

0.0

0.5

0.5

residuals costCDHom

0.0

0.5

1.0

residuals costTLHom

Figure 5.6: Residuals of Cobb-Douglas and Translog cost functions

The residuals of the two cost functions have both a small (in absolute terms) but negative
skewness, which means that the residuals are slightly left-skewed, although we expected rightskewed residuals. It could be that the distribution of the unknown true total error term (u + v)
in the sample is indeed symmetric or slightly left-skewed, e.g. because
there is no cost inefficiency (but only noise) (the distribution of residuals is correct),
the distribution of the noise term is left-skewed, which neutralizes the right-skewed distri-

bution of the inefficiency term (misspecification of the distribution of the noise term in the
SFA model),
the distribution of the inefficiency term is symmetric or left-skewed (misspecification of the
distribution of the inefficiency term in the SFA model),
the sampling of the observations by coincidence resulted in a symmetric or left-skewed distribution of the true total error term (u+v) in this specific sample, although the distribution
of the true total error term (u + v) in the population is right-skewed, and/or
the farm managers do not aim at maximizing profit (which implies minimizing costs) but
have other objectives.
It could also be that the distribution of the unknown true residuals in the sample is right-skewed,
but the OLS estimates are left-skewed, e.g. because
the parameter estimates are imprecise (but unbiased),
the estimated functional forms (Cobb-Douglas and Translog) are poor approximations of

the unknown true functional form (functional-form misspecification),

there are further relevant explanatory variables that are not included in the model specification (omitted-variables bias),
there are measurement errors in the variables, particularly in the explanatory variables
(errors-in-variables problem), and/or
the output quantity or the input prices are not exogenously given (endogeneity bias).

209

5 Stochastic Frontier Analysis

Hence, a left-skewed distribution of the residuals does not necessarily mean that there is no
cost inefficiency, but it could also mean that the model is misspecified or that this is just by
coincidence.

5.3.3 Estimation of a Cobb-Douglas stochastic cost frontier

The following command estimates a Cobb-Douglas stochastic cost frontier with linear homogeneity in input prices imposed:
> costCDHomSfa <- sfa( log( cost / pMat ) ~ log( pCap / pMat ) +
+

log( pLab / pMat ) + log( qOut ), data = dat,

ineffDecrease = FALSE )

> summary( costCDHomSfa )

Error Components Frontier (see Battese & Coelli 1992)
Inefficiency increases the endogenous variable (as in a cost function)
The dependent variable is logged
Iterative ML estimation terminated after 53 iterations:
cannot find a parameter vector that results in a log-likelihood value
larger than the log-likelihood value obtained in the previous step
final maximum likelihood estimates
Estimate Std. Error z value
(Intercept)

6.74916371 0.74012850

Pr(>|z|)

9.1189 < 2.2e-16 ***

log(pCap/pMat) 0.07241581 0.04506614

1.6069

0.1081

log(pLab/pMat) 0.44641744 0.07888311

5.6592 1.521e-08 ***

log(qOut)

0.37415457 0.02962967 12.6277 < 2.2e-16 ***

sigmaSq

0.11117637 0.01424404

7.8051 5.945e-15 ***

gamma

0.00019319 0.06608496

0.0029

0.9977

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

log likelihood value: -44.87812

cross-sectional data
total number of observations = 140
mean efficiency: 0.9963129
The parameter , which indicates the proportion of the total residual variance that is caused by
inefficiency is close to zero and a t-test suggests that it is statistically not significantly different
from zero. As the t-test for the parameter is not always reliable, we use a likelihood ratio test
to verify this result:

210

5 Stochastic Frontier Analysis

> lrtest( costCDHomSfa )
Likelihood ratio test
Model 1: OLS (no inefficiency)
Model 2: Error Components Frontier (ECF)
#Df

LogLik Df Chisq Pr(>Chisq)

5 -44.878

6 -44.878

0.4991

This test confirms that the fit of the OLS model (which assumes that is zero and hence, that
there is no inefficiency) is not significantly worse than the fit of the stochastic frontier model.
In fact, the cost efficiency estimates are all very close to one. By default, the efficiencies()
method calculates the efficiency estimates as E [eu ], which means that we obtain estimates
of Farrell-type cost efficiencies (5.33). Given that E [eu ] is not equal to 1/E [eu ] (as the expectation operator is an additive operator), we cannot obtain estimates of Shepard-type cost
efficiencies (5.32) by taking the inverse of the estimates of the Farrell-type cost efficiencies (5.33).
However, we can obtain estimates of Shepard-type cost efficiencies (5.32) by setting argument
minusU of the efficiencies() method equal to FALSE, which tells the efficiencies() method
to calculate the efficiency estimates as E [eu ].
> dat$costEffCDHomFarrell <- efficiencies( costCDHomSfa )
> dat$costEffCDHomShepard <- efficiencies( costCDHomSfa, minusU = FALSE )
> hist( dat$costEffCDHomFarrell, 15 )

25
0.99630

0.99634

15

Frequency
0.99626

0 5

15
0 5

Frequency

25

> hist( dat$costEffCDHomShepard, 15 )

1.00366

costEffCDHomFarrell

1.00370

1.00374

costEffCDHomShepard

Figure 5.7: Efficiency estimates of Cobb-Douglas cost frontier

The resulting graphs are shown in figure 5.7. While the Farrell-type cost efficiencies are all slightly
below one, the Shepard-type cost efficiencies are all slightly above one. Both graphs show that
we do not find any relevant cost inefficiencies, although we have found considerable technical
inefficiencies.

211

5 Stochastic Frontier Analysis

5.4 Analyzing the Effects of z Variables

In many empirical cases, the output quantity does not only depend on the input quantities but
also on some other variables, e.g. the managers education and experience and in agricultural
production also the soil quality and rainfall. If these factors influence the production process,
they must be included in applied production analyses in order to avoid an omitted-variables bias.
Our data set on French apple producers includes the variable adv, which is a dummy variable
and indicates whether the apple producer uses an advisory service. In the following, we will
apply different methods to figure out whether the production process differs between users and
non-users of an advisory service.

5.4.1 Production Functions with z Variables

Additional factors that influence the production process (z) can be included as additional explanatory variables in the production function:
y = f (x, z).

(5.34)

This function can be used to analyze how the additional explanatory variables (z) affect the
output quantity for given input quantities, i.e. how they affect the productivity.
In case of a Cobb-Douglas functional form, we get following extended production function:
ln y = 0 +

i ln xi + z z

(5.35)

Based on this Cobb-Douglas production function and our data set on French apple producers,
we can check whether the apple producers who use an advisory service produce a different output
quantity than non-users with the same input quantities, i.e. whether the productivity differs
between users and non-users. This extended production function can be estimated by following
command:
> prodCDAdv <- lm( log( qOut ) ~ log( qCap ) + log( qLab ) + log( qMat ) + adv,
+

data = dat )

> summary( prodCDAdv )

Call:
lm(formula = log(qOut) ~ log(qCap) + log(qLab) + log(qMat) +
adv, data = dat)
Residuals:
Min

1Q

Median

3Q

Max

-1.7807 -0.3821

0.0022

0.4709

1.3323

212

5 Stochastic Frontier Analysis

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) -2.33371

1.29590

-1.801

0.0740 .

log(qCap)

0.15673

0.08581

1.826

0.0700 .

log(qLab)

0.69225

0.15190

4.557 1.15e-05 ***

log(qMat)

0.62814

0.12379

5.074 1.26e-06 ***

adv

0.25896

0.10932

2.369

0.0193 *

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Residual standard error: 0.6452 on 135 degrees of freedom

Multiple R-squared:
F-statistic:

0.6105,

Adjusted R-squared:

52.9 on 4 and 135 DF,

p-value: < 2.2e-16

0.599

The estimation result shows that users of an advisory service produce significantly more than
non-users with the same input quantities. Given the Cobb-Douglas production function (5.35),
the coefficient of an additional explanatory variable can be interpreted as the marginal effect on
the relative change of the output quantity:
z =

ln y
ln y y
y 1
=
=
z
y z
z y

(5.36)

Hence, our estimation result indicates that users of an advisory service produce approximately
25.9% more output than non-users with the same input quantity but the large standard error
of this coefficient indicates that this estimate is rather imprecise. Given that the change of a
dummy variable from zero to one is not marginal and that the coefficient of the variable adv is
not close to zero, the above interpretation of this coefficient is a rather poor approximation. In
fact, our estimation results suggest that the output quantity of apple producers with advisory
service is on average exp(z ) = 1.296 times as large as (29.6% larger than) the output quantity of
apple producers without advisory service given the same input quantities. As users and non-users
of an advisory service probably differ in some unobserved variables that affect the productivity
(e.g. motivation and effort to increase productivity), the coefficient az is not necessarily the
causal effect of the advisory service but describes the difference in productivity between users
and non-users of the advisory service.

5.4.2 Production Frontiers with z Variables

A production function that includes additional factors that influence the production process (5.34)
can also be estimated as a stochastic production frontier. In this specification, it is assumed that
the additional explanatory variables influence the production frontier.
The following command estimates the extended Cobb-Douglas production function (5.35) using
the stochastic frontier method:

213

5 Stochastic Frontier Analysis

> prodCDAdvSfa <- sfa( log( qOut ) ~ log( qCap ) + log( qLab ) + log( qMat ) + adv,
+

data = dat )

> summary( prodCDAdvSfa )

Error Components Frontier (see Battese & Coelli 1992)
Inefficiency decreases the endogenous variable (as in a production function)
The dependent variable is logged
Iterative ML estimation terminated after 14 iterations:
log likelihood values and parameters of two successive iterations
are within the tolerance limit
final maximum likelihood estimates
Estimate Std. Error z value

Pr(>|z|)

(Intercept) -0.247751

1.357917 -0.1824 0.8552300

log(qCap)

0.156906

0.081337

1.9291 0.0537222 .

log(qLab)

0.695977

0.148793

4.6775 2.904e-06 ***

log(qMat)

0.491840

0.139348

3.5296 0.0004162 ***

adv

0.150742

0.111233

1.3552 0.1753583

sigmaSq

0.916031

0.231604

3.9552 7.648e-05 ***

gamma

0.861029

0.114087

7.5471 4.450e-14 ***

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

log likelihood value: -132.8679

cross-sectional data
total number of observations = 140
mean efficiency: 0.5545099
The estimation result still indicates that users of an advisory service have a higher productivity
than non users, but the coefficient is smaller and no longer statistically significant. The result of
the t-test is confirmed by a likelihood-ratio test:
> lrtest( prodCDSfa, prodCDAdvSfa )
Likelihood ratio test
Model 1: prodCDSfa
Model 2: prodCDAdvSfa
#Df

LogLik Df

6 -133.89

7 -132.87

Chisq Pr(>Chisq)

1 2.0428

0.1529

214

5 Stochastic Frontier Analysis

The model with advisory service as additional explanatory variable indicates that there are
significant inefficiencies (at 5% significance level):
> lrtest( prodCDAdvSfa )
Likelihood ratio test
Model 1: OLS (no inefficiency)
Model 2: Error Components Frontier (ECF)
#Df

LogLik Df Chisq Pr(>Chisq)

6 -134.76

7 -132.87

3.78

0.02593 *

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

The following commands compute the technical efficiency estimates and compare them to the
efficiency estimates obtained from the Cobb-Douglas production frontier without advisory service
as an explanatory variable:
> dat$effCDAdv <- efficiencies( prodCDAdvSfa )
> compPlot( dat$effCD[ dat$adv == 0 ],
+

dat$effCDAdv[ dat$adv == 0 ] )

> points( dat$effCD[ dat$adv == 1 ],

+

dat$effCDAdv[ dat$adv == 1 ], pch = 20 )

The resulting graph is shown in figure 5.8. It appears as if the non-users of an advisory service
became somewhat more efficient. This is because the stochastic frontier model that includes
the advisory service as an explanatory variable has in fact two production frontiers: a lower
frontier for the non-users of an advisory service and a higher frontier for the users of an advisory
service. The coefficient of the dummy variable adv, i.e. adv , can be interpreted as a quick
estimate of the difference between the two frontier functions. In our empirical case, the difference
is approximately 15.1%. However, a precise calculation indicates that the frontier of the users of
the advisory service is exp (adv ) = 1.163 times (16.3% higher than) the frontier of the non-users
of advisory service. And the frontier of the non-users of the advisory service is exp (adv ) =
0.86 times (14% lower than) the frontier of the users of advisory service. As the non-users of
an advisory service are compared to a lower frontier now, they appear to be more efficient now.
While it is reasonable to have different frontier functions for different soil types, it does not seem
to be too reasonable to have different frontier functions for users and non-users of an advisory
service, because there is no physical reasons, why users of an advisory service should have a
maximum output quantity that is different from the maximum output quantity of non-users.

215

0.8

0.6

0.4

0.2

Production frontier with advisory service

5 Stochastic Frontier Analysis

0.2

0.4

0.6

0.8

Production frontier without advisory service

Figure 5.8: Technical efficiency estimates of Cobb-Douglas production frontier with and without
advisory service as additional explanatory variable (circles = producers who do not
use an advisory service, solid dots = producers who use an advisory service

5.4.3 Efficiency Effects Production Frontiers

As explained above, it does not seem to be too reasonable to have different frontier functions for
users and non-users of an advisory service. However, it seems to be reasonable to assume that
users of an advisory service have on average different efficiencies than non-users. A model that
can account for this has been proposed by Battese and Coelli (1995). In this stochastic frontier
model, the efficiency level might be affected by additional explanatory variables: The inefficiency
term u follows a positive truncated normal distribution with constant scale parameter u2 and a
location parameter that depends on additional explanatory variables:
u N + (, u2 )

with = z,

(5.37)

where is an additional parameter (vector) to be estimated. Function sfa can also estimate
these efficiency effects frontiers. The additional variables that should explain the efficiency
level must be specified at the end of the model formula, where a vertical bar separates them from
the (regular) input variables:
> prodCDSfaAdvInt <- sfa( log( qOut ) ~ log( qCap ) + log( qLab ) + log( qMat ) |
+

adv, data = dat )

> summary( prodCDSfaAdvInt )

Efficiency Effects Frontier (see Battese & Coelli 1995)
Inefficiency decreases the endogenous variable (as in a production function)

216

5 Stochastic Frontier Analysis

The dependent variable is logged
Iterative ML estimation terminated after 18 iterations:
log likelihood values and parameters of two successive iterations
are within the tolerance limit
final maximum likelihood estimates
Estimate Std. Error z value
(Intercept)

-0.090747

Pr(>|z|)

1.215594 -0.0747 0.9404914

log(qCap)

0.168625

0.082138

2.0529 0.0400775 *

log(qLab)

0.653868

0.142849

4.5773

log(qMat)

0.513527

0.132644

3.8715 0.0001082 ***

4.71e-06 ***

Z_(Intercept) -0.016714

1.220842 -0.0137 0.9890769

Z_adv

1.057873 -1.0186 0.3083926

-1.077550

sigmaSq

1.096441

0.747990

1.4659 0.1426891

gamma

0.863087

0.094468

9.1363 < 2.2e-16 ***

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

log likelihood value: -130.516

cross-sectional data
total number of observations = 140
mean efficiency: 0.600435
One can use the lrtest() method to test the statistical significance of the entire inefficiency
model, i.e. the null hypothesis is H0 : = 0 and j = 0 j:
> lrtest( prodCDSfaAdvInt )
Likelihood ratio test
Model 1: OLS (no inefficiency)
Model 2: Efficiency Effects Frontier (EEF)
#Df

LogLik Df

5 -137.61

8 -130.52

Chisq Pr(>Chisq)

3 14.185

0.001123 **

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

The test indicates that the fit of this model is significantly better than the fit of the OLS model
(without advisory service as explanatory variable).

217

5 Stochastic Frontier Analysis

The coefficient of the advisory service in the inefficiency model is negative but statistically
insignificant. By default, an intercept is added to the inefficiency model but it is completely
statistically insignificant. In many econometric estimations of the efficiency effects frontier model,
the intercept of the inefficiency model (0 ) is only weakly identified, because the values of 0 can
often be changed with only marginally reducing the log-likelihood value, if the slope parameters of
the inefficiency model (i , i 6= 0) and the variance parameters ( 2 and ) are adjusted accordingly.
This can be checked by taking a look at the correlation matrix of the estimated parameters:
> round( cov2cor( vcov( prodCDSfaAdvInt ) ), 2 )
(Intercept) log(qCap) log(qLab) log(qMat) Z_(Intercept) Z_adv
(Intercept)

1.00

-0.06

-0.48

-0.21

0.07

0.08

log(qCap)

-0.06

1.00

-0.37

-0.17

log(qLab)

-0.48

-0.37

1.00

-0.57

log(qMat)

-0.21

-0.17

-0.57

1.00

Z_(Intercept)

0.07

-0.13

0.20

-0.11

1.00

0.91

Z_adv

0.08

-0.13

0.26

-0.20

0.91

1.00

sigmaSq

0.03

0.10

-0.14

0.00

-0.95 -0.89

gamma

0.29

-0.03

0.07

-0.34

-0.55 -0.46

-0.13 -0.13
0.20

0.26

-0.11 -0.20

sigmaSq gamma
(Intercept)

0.03

log(qCap)

0.10 -0.03

log(qLab)
log(qMat)

-0.14

0.29
0.07

0.00 -0.34

Z_(Intercept)

-0.95 -0.55

Z_adv

-0.89 -0.46

sigmaSq

1.00

0.73

gamma

0.73

1.00

The estimate of the intercept of the inefficiency model (0 ) is very highly correlated with the
estimate of the (slope) coefficient of the advisory service in the inefficiency model (1 ) and the
estimate of the parameter 2 and it is considerably correlated with the estimate of the parameter .
The intercept can be suppressed by adding a -1 to the specification of the inefficiency model:
> prodCDSfaAdv <- sfa( log( qOut ) ~ log( qCap ) + log( qLab ) + log( qMat ) |
+

adv - 1, data = dat )

> summary( prodCDSfaAdv )

Efficiency Effects Frontier (see Battese & Coelli 1995)
Inefficiency decreases the endogenous variable (as in a production function)
The dependent variable is logged

218

5 Stochastic Frontier Analysis

Iterative ML estimation terminated after 14 iterations:
log likelihood values and parameters of two successive iterations
are within the tolerance limit
final maximum likelihood estimates
Estimate Std. Error z value

Pr(>|z|)

(Intercept) -0.090455

1.247496 -0.0725

0.94220

log(qCap)

0.168471

0.077008

2.1877

0.02869 *

log(qLab)

0.654341

0.139669

4.6849 2.800e-06 ***

log(qMat)

0.513291

0.130854

3.9226 8.759e-05 ***

Z_adv

-1.064859

0.545950 -1.9505

0.05112 .

sigmaSq

1.086417

0.255371

4.2543 2.097e-05 ***

gamma

0.862306

0.081468 10.5845 < 2.2e-16 ***

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

log likelihood value: -130.5161

cross-sectional data
total number of observations = 140
mean efficiency: 0.599406
A likelihood ratio test against the corresponding OLS model indicates that the fit of this SFA
model is significantly better than the fit of the corresponding OLS model (without advisory
service as explanatory variable):
> lrtest( prodCDSfaAdv )
Likelihood ratio test
Model 1: OLS (no inefficiency)
Model 2: Efficiency Effects Frontier (EEF)
#Df

LogLik Df

5 -137.61

7 -130.52

Chisq Pr(>Chisq)

2 14.185

0.0002907 ***

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

A likelihood ratio test confirms the t-test that the intercept in the inefficiency model is statistically
insignificant:
> lrtest( prodCDSfaAdv, prodCDSfaAdvInt )

219

5 Stochastic Frontier Analysis

Likelihood ratio test
Model 1: prodCDSfaAdv
Model 2: prodCDSfaAdvInt
#Df

LogLik Df Chisq Pr(>Chisq)

7 -130.52

8 -130.52

1 2e-04

0.9892

The coefficient of the advisory service in the inefficiency model is now significantly negative
(at 10% significance level), which means that users of an advisory service have a significantly
smaller inefficiency term u, i.e. are significantly more efficient. The size of the coefficients of the
inefficiency model () cannot be reasonably interpreted. However, if argument margEff of the
efficiencies method is set to TRUE, this method does not only return the efficiency estimates but
also the marginal effects of the variables that should explain the efficiency level on the efficiency
estimates (see Olsen and Henningsen, 2011):
> dat$effCDAdv2 <- efficiencies( prodCDSfaAdv, margEff = TRUE )
The marginal effects differ between observations and are available in the attribute margEff. The
following command extracts and visualizes the marginal effects of the variable that indicates the
use of an advisory service on the efficiency estimates:

15
5
0

Frequency

> hist( attr( dat$effCDAdv2, "margEff" ), 20 )

0.02

0.03

0.04

0.05

0.06

marginal effect

Figure 5.9: Marginal effects of the variable that indicates the use of an advisory service on the
efficiency estimates
The resulting graph is shown in figure 5.9. It indicates that apple producers who use an advisory
service are between 6.3 and 6.4 percentage points more efficient than apple producers who do not
use an advisory service.

220

6 Panel Data and Technological Change

Until now, we have only analyzed cross-sectional data, i.e. all observations refer to the same period
of time. Hence, it was reasonable to assume that the same technology is available to all firms
(observations). However, when analyzing time series data or panel data, i.e. when observations
can originate from different time periods, different technologies might be available in the different
time periods due to technological change. Hence, the state of the available technologies must be
included as an explanatory variable in order to conduct a reasonable production analysis. Often,
a time trend is used as a proxy for a gradually changing state of the available technologies.
We will demonstrate how to analyze production technologies with data from different time
periods by using a balanced panel data set of annual data collected from 43 smallholder rice
producers in the Tarlac region of the Philippines between 1990 and 1997. We loaded this data set
(riceProdPhil) in section 1.3.2. As it does not contain information about the panel structure,
we created a copy of the data set (pdat) that includes information on the panel structure.

6.1 Average Production Functions with Technological Change

In case of an applied production analysis with time-series data or panel data, usually the time (t)
is included as additional explanatory variable in the production function:
y = f (x, t).

(6.1)

This function can be used to analyze how the time (t) affects the (available) production technology.
The average production technology (potentially depending on the time period) can be estimated
from panel data sets by the OLS method (i.e. pooled) or by any of the usual panel data methods
(e.g. fixed effects, random effects).

6.1.1 Cobb-Douglas Production Function with Technological Change

In case of a Cobb-Douglas production function, usually a linear time trend is added to account
for technological change:
ln y = 0 +

i ln xi + t t

221

(6.2)

6 Panel Data and Technological Change

Given this specification, the coefficient of the (linear) time trend can be interpreted as the rate
of technological change per unit of the time variable t:
y
ln y
ln y y
y
t =
=

t
y t
x

(6.3)

6.1.1.1 Pooled estimation of the Cobb-Douglas Production Function with Technological

Change
The pooled estimation can be done by:
> riceCdTime <- lm( log( PROD ) ~ log( AREA ) + log( LABOR ) + log( NPK ) +
+

mYear, data = riceProdPhil )

> summary( riceCdTime )

Call:
lm(formula = log(PROD) ~ log(AREA) + log(LABOR) + log(NPK) +
mYear, data = riceProdPhil)
Residuals:
Min

1Q

Median

3Q

Max

-1.83351 -0.16006

0.05329

0.22110

0.86745

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) -1.665096

0.248509

-6.700 8.68e-11 ***

log(AREA)

0.333214

0.062403

5.340 1.71e-07 ***

log(LABOR)

0.395573

0.066421

5.956 6.48e-09 ***

log(NPK)

0.270847

0.041027

6.602 1.57e-10 ***

mYear

0.010090

0.008007

1.260

0.208

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Residual standard error: 0.3299 on 339 degrees of freedom

Multiple R-squared:

0.86,

Adjusted R-squared:

F-statistic: 520.6 on 4 and 339 DF,

0.8583

p-value: < 2.2e-16

The estimation result indicates an annual rate of technical change of 1%, but this is not statistically different from 0%, which means no technological change.
The command above can be simplified by using the pre-calculated logarithmic (and meanscaled) quantities:

222

6 Panel Data and Technological Change

> riceCdTimeS <- lm( lProd ~ lArea + lLabor + lNpk + mYear, data = riceProdPhil )
> summary( riceCdTimeS )
Call:
lm(formula = lProd ~ lArea + lLabor + lNpk + mYear, data = riceProdPhil)
Residuals:
Min

1Q

Median

3Q

Max

-1.83351 -0.16006

0.05329

0.22110

0.86745

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) -0.015590

0.019325

-0.807

0.420

lArea

0.333214

0.062403

5.340 1.71e-07 ***

lLabor

0.395573

0.066421

5.956 6.48e-09 ***

lNpk

0.270847

0.041027

6.602 1.57e-10 ***

mYear

0.010090

0.008007

1.260

0.208

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Residual standard error: 0.3299 on 339 degrees of freedom

Multiple R-squared:

0.86,

Adjusted R-squared:

F-statistic: 520.6 on 4 and 339 DF,

0.8583

p-value: < 2.2e-16

The intercept has changed because of the mean-scaling of the input and output quantities but
all slope parameters are unaffected by using the pre-calculated logarithmic (and mean-scaled)
quantities:
> all.equal( coef( riceCdTime )[-1], coef( riceCdTimeS )[-1],
+

check.attributes = FALSE )

[1] TRUE
6.1.1.2 Panel data estimations of the Cobb-Douglas Production Function with
Technological Change
The panel data estimation with fixed individual effects can be done by:
> riceCdTimeFe <- plm( lProd ~ lArea + lLabor + lNpk + mYear, data = pdat )
> summary( riceCdTimeFe )
Oneway (individual) effect Within Model

223

6 Panel Data and Technological Change

Call:
plm(formula = lProd ~ lArea + lLabor + lNpk + mYear, data = pdat)
Balanced Panel: n=43, T=8, N=344
Residuals :
Min. 1st Qu.

Median 3rd Qu.

-1.5900 -0.1570

0.0456

0.1780

Max.
0.8180

Coefficients :
Estimate Std. Error t-value
lArea

Pr(>|t|)

0.5607756

0.0785370

7.1403 7.195e-12 ***

lLabor 0.2549108

0.0690631

3.6910 0.0002657 ***

lNpk

0.1748528

0.0484684

3.6076 0.0003625 ***

mYear

0.0130908

0.0071824

1.8226 0.0693667 .

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Total Sum of Squares:

43.632

Residual Sum of Squares: 24.872

R-Squared

0.42995

Adj. R-Squared :

0.3712

F-statistic: 56.0008 on 4 and 297 DF, p-value: < 2.22e-16

And the panel data estimation with random individual effects can be done by:
> riceCdTimeRan <- plm( lProd ~ lArea + lLabor + lNpk + mYear, data = pdat,
+

model = "random" )

> summary( riceCdTimeRan )

Oneway (individual) effect Random Effect Model
(Swamy-Arora's transformation)
Call:
plm(formula = lProd ~ lArea + lLabor + lNpk + mYear, data = pdat,
model = "random")
Balanced Panel: n=43, T=8, N=344
Effects:
var std.dev share

224

6 Panel Data and Technological Change

idiosyncratic 0.08375 0.28939

0.8

individual

0.2

theta:

0.02088 0.14451

0.4222

Residuals :
Min. 1st Qu.

Median 3rd Qu.

-1.7500 -0.1430

0.0485

0.1910

Max.
0.8520

Coefficients :
Estimate Std. Error t-value

Pr(>|t|)

(Intercept) -0.0213044

0.0292268 -0.7289

0.4665

lArea

0.4563002

0.0662979

6.8826 2.854e-11 ***

lLabor

0.3190041

0.0647524

4.9265 1.311e-06 ***

lNpk

0.2268399

0.0426651

5.3168 1.921e-07 ***

mYear

0.0115453

0.0071921

1.6053

0.1094

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Total Sum of Squares:

117.05

Residual Sum of Squares: 29.195

R-Squared

0.75058

Adj. R-Squared :

0.73968

F-statistic: 255.045 on 4 and 339 DF, p-value: < 2.22e-16

A variable-coefficient model for panel model with individual-specific coefficients can be estimated by:
> riceCdTimeVc <- pvcm( lProd ~ lArea + lLabor + lNpk + mYear, data = pdat )
> summary( riceCdTimeVc )
Oneway (individual) effect No-pooling model
Call:
pvcm(formula = lProd ~ lArea + lLabor + lNpk + mYear, data = pdat)
Balanced Panel: n=43, T=8, N=344
Residuals:
Min.

1st Qu.

Median

Mean

3rd Qu.

Max.

-0.817500 -0.081970

0.006677

0.000000

0.093980

0.554100

225

6 Panel Data and Technological Change

Coefficients:
(Intercept)
Min.

:-3.8110

lArea
Min.

:-5.2850

lLabor
Min.

:-2.72761

lNpk
Min.

:-1.3094

1st Qu.:-0.3006

1st Qu.:-0.4200

1st Qu.:-0.30989

1st Qu.:-0.1867

Median : 0.1145

Median : 0.6978

Median : 0.08778

Median : 0.1050

Mean

Mean

Mean

Mean

: 0.1839

: 0.5896

: 0.06079

: 0.1265

3rd Qu.: 0.5617

3rd Qu.: 1.8914

3rd Qu.: 0.61479

3rd Qu.: 0.3808

Max.

Max.

: 4.7633

Max.

Max.

NA's

:18

: 3.7270

: 1.75595

: 1.7180

mYear
Min.

:-0.471049

1st Qu.:-0.044359
Median :-0.008111
Mean

:-0.012327

3rd Qu.: 0.054743

Max.

: 0.275875

Total Sum of Squares: 2861.8

Residual Sum of Squares: 8.9734
Multiple R-Squared: 0.99686
A pooled estimation can also be done by
> riceCdTimePool <- plm( lProd ~ lArea + lLabor + lNpk + mYear, data = pdat,
+

model = "pooling" )
This gives the same estimated coefficients as the model estimated by lm:

> all.equal( coef( riceCdTimeS ), coef( riceCdTimePool ) )

[1] TRUE
A Hausman test can be used to check the consistency of the random-effects estimator:
> phtest( riceCdTimeRan, riceCdTimeFe )
Hausman Test
data:

lProd ~ lArea + lLabor + lNpk + mYear

chisq = 14.6204, df = 4, p-value = 0.005557

alternative hypothesis: one model is inconsistent

226

6 Panel Data and Technological Change

The Hausman test clearly shows that the random-effects estimator is inconsistent (due to correlation between the individual effects and the explanatory variables).
Now, we test the poolability of the model:
> pooltest( riceCdTimePool, riceCdTimeFe )
F statistic
data:

lProd ~ lArea + lLabor + lNpk + mYear

F = 3.4175, df1 = 42, df2 = 297, p-value = 4.038e-10

alternative hypothesis: unstability
> pooltest( riceCdTimePool, riceCdTimeVc )
F statistic
data:

lProd ~ lArea + lLabor + lNpk + mYear

F = 1.9113, df1 = 210, df2 = 129, p-value = 4.022e-05

alternative hypothesis: unstability
> pooltest( riceCdTimeFe, riceCdTimeVc )
F statistic
data:

lProd ~ lArea + lLabor + lNpk + mYear

F = 1.3605, df1 = 168, df2 = 129, p-value = 0.03339

alternative hypothesis: unstability
The pooled model (riceCdTimePool) is clearly rejected in favour of the model with fixed individual effects (riceCdTimeFe) and the variable-coefficient model (riceCdTimeVc). The model
with fixed individual effects (riceCdTimeFe) is rejected in favor of the variable-coefficient model
(riceCdTimeVc) at 5% significance level but not at 1% significance level.

6.1.2 Translog Production Function with Constant and Neutral Technological

Change
A Translog production function that accounts for constant and neutral (unbiased) technological
change has following specification:
ln y = 0 +

X
i

i ln xi +

1 XX
ij ln xi ln xj + t t
2 i j

(6.4)

In this specification, the rate of technological change is

ln y
= t
t

227

(6.5)

6 Panel Data and Technological Change

and the output elasticities are the same as in the time-invariant Translog production function (2.105):
i =

X
ln y
= i +
ij ln xj
ln xi
j

(6.6)

In order to be able to interpret the first-order coefficients of the (logarithmic) input quantities
(i ) as output elasticities (i ) at the sample mean, we use the mean-scaled input quantities. We
also use the mean-scaled output quantity in order to use the same variables as Coelli et al. (2005,
p. 250).
6.1.2.1 Pooled estimation of the Translog Production Function with Constant and Neutral
Technological Change
The following command estimates a Translog production function that can account for constant
and neutral technical change:
> riceTlTime <- lm( lProd ~ lArea + lLabor + lNpk +
+

I( 0.5 * lArea^2 ) + I( 0.5 * lLabor^2 ) + I( 0.5 * lNpk^2 ) +

I( lArea * lLabor ) + I( lArea * lNpk ) + I( lLabor * lNpk ) + mYear,

data = riceProdPhil )

> summary( riceTlTime )

Call:
lm(formula = lProd ~ lArea + lLabor + lNpk + I(0.5 * lArea^2) +
I(0.5 * lLabor^2) + I(0.5 * lNpk^2) + I(lArea * lLabor) +
I(lArea * lNpk) + I(lLabor * lNpk) + mYear, data = riceProdPhil)
Residuals:
Min

1Q

Median

3Q

Max

-1.52184 -0.18121

0.04356

0.22298

0.87019

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept)

0.013756

0.024645

0.558

lArea

0.588097

0.085162

6.906 2.54e-11 ***

lLabor

0.191764

0.080876

2.371

0.01831 *

lNpk

0.197875

0.051605

3.834

0.00015 ***

-0.435547

0.247491

-1.760

0.07935 .

I(0.5 * lLabor^2) -0.742242

0.303236

-2.448

0.01489 *

I(0.5 * lNpk^2)

0.020367

0.097907

0.208

0.83534

I(lArea * lLabor)

0.678647

0.216594

3.133

0.00188 **

I(0.5 * lArea^2)

228

0.57712

6 Panel Data and Technological Change

I(lArea * lNpk)

0.063920

0.145613

0.439

0.66097

I(lLabor * lNpk)

-0.178286

0.138611

-1.286

0.19926

0.012682

0.007795

1.627

0.10468

mYear
--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Residual standard error: 0.3184 on 333 degrees of freedom

Multiple R-squared:

0.8719,

Adjusted R-squared:

F-statistic: 226.6 on 10 and 333 DF,

0.868

p-value: < 2.2e-16

In the Translog production function that accounts for constant and neutral technological change,
the monotonicity conditions are fulfilled at the sample mean and the estimated output elasticities
of land, labor and fertilizer are 0.588, 0.192, and 0.198, respectively, at the sample mean. The
estimated (constant) annual rate of technological progress is around 1.3%.
Conduct a Wald test to test whether the Translog production function outperforms the CobbDouglas production function:
> library( "lmtest" )
> waldtest( riceCdTimeS, riceTlTime )
Wald test
Model 1: lProd ~ lArea + lLabor + lNpk + mYear
Model 2: lProd ~ lArea + lLabor + lNpk + I(0.5 * lArea^2) + I(0.5 * lLabor^2) +
I(0.5 * lNpk^2) + I(lArea * lLabor) + I(lArea * lNpk) + I(lLabor *
lNpk) + mYear
Res.Df Df
1

339

333

Pr(>F)

6 5.1483 4.451e-05 ***

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

The Cobb-Douglas specification is clearly rejected in favour of the Translog specification for the
pooled estimation.
6.1.2.2 Panel-data estimations of the Translog Production Function with Constant and
Neutral Technological Change
The following command estimates a Translog production function that can account for constant
and neutral technical change with fixed individual effects:
> riceTlTimeFe <- plm( lProd ~ lArea + lLabor + lNpk +
+

I( 0.5 * lArea^2 ) + I( 0.5 * lLabor^2 ) + I( 0.5 * lNpk^2 ) +

229

6 Panel Data and Technological Change

+

I( lArea * lLabor ) + I( lArea * lNpk ) + I( lLabor * lNpk ) + mYear,

data = pdat, model = "within" )

> summary( riceTlTimeFe )

Oneway (individual) effect Within Model
Call:
plm(formula = lProd ~ lArea + lLabor + lNpk + I(0.5 * lArea^2) +
I(0.5 * lLabor^2) + I(0.5 * lNpk^2) + I(lArea * lLabor) +
I(lArea * lNpk) + I(lLabor * lNpk) + mYear, data = pdat,
model = "within")
Balanced Panel: n=43, T=8, N=344
Residuals :
Min. 1st Qu.

Median 3rd Qu.

-1.0100 -0.1450

0.0191

0.1680

Max.
0.7460

Coefficients :
Estimate Std. Error t-value Pr(>|t|)
lArea

0.5828102

0.1173298

4.9673 1.16e-06 ***

lLabor

0.0473355

0.0848594

0.5578 0.577402

lNpk

0.1211928

0.0610114

1.9864 0.047927 *

I(0.5 * lArea^2)

-0.8543901

0.2861292 -2.9860 0.003067 **

I(0.5 * lLabor^2) -0.6217163

0.2935429 -2.1180 0.035025 *

I(0.5 * lNpk^2)

0.0429446

0.0987119

0.4350 0.663849

I(lArea * lLabor)

0.5867063

0.2125686

2.7601 0.006145 **

I(lArea * lNpk)

0.1167509

0.1461380

0.7989 0.424995

I(lLabor * lNpk)

-0.2371219

mYear

0.0165309

0.1268671 -1.8691 0.062619 .

0.0069206

2.3887 0.017547 *

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Total Sum of Squares:

43.632

Residual Sum of Squares: 21.912

R-Squared

0.49781

Adj. R-Squared :

0.42111

F-statistic: 28.8456 on 10 and 291 DF, p-value: < 2.22e-16

And the panel data estimation with random individual effects can be done by:

230

6 Panel Data and Technological Change

> riceTlTimeRan <- plm( lProd ~ lArea + lLabor + lNpk +
+

I( 0.5 * lArea^2 ) + I( 0.5 * lLabor^2 ) + I( 0.5 * lNpk^2 ) +

I( lArea * lLabor ) + I( lArea * lNpk ) + I( lLabor * lNpk ) + mYear,

data = pdat, model = "random" )

> summary( riceTlTimeRan )

Oneway (individual) effect Random Effect Model
(Swamy-Arora's transformation)
Call:
plm(formula = lProd ~ lArea + lLabor + lNpk + I(0.5 * lArea^2) +
I(0.5 * lLabor^2) + I(0.5 * lNpk^2) + I(lArea * lLabor) +
I(lArea * lNpk) + I(lLabor * lNpk) + mYear, data = pdat,
model = "random")
Balanced Panel: n=43, T=8, N=344
Effects:
var std.dev share
idiosyncratic 0.07530 0.27440

0.79

individual

0.21

theta:

0.01997 0.14130

0.434

Residuals :
Min. 1st Qu.
-1.3900 -0.1620

Median 3rd Qu.

0.0456

0.1840

Max.
0.7980

Coefficients :
Estimate Std. Error t-value

Pr(>|t|)

(Intercept)

0.0213211

0.0347371

0.6138

lArea

0.6831045

0.0922069

7.4084 1.061e-12 ***

lLabor

0.0974523

0.0804060

1.2120

0.226370

lNpk

0.1708366

0.0546853

3.1240

0.001941 **

I(0.5 * lArea^2)

0.539776

-0.4275328

0.2468086 -1.7322

0.084156 .

I(0.5 * lLabor^2) -0.6367899

0.2872825 -2.2166

0.027326 *

I(0.5 * lNpk^2)

0.0307547

0.0957745

0.3211

0.748324

I(lArea * lLabor)

0.5666863

0.2059076

2.7521

0.006245 **

I(lArea * lNpk)

0.1037657

0.1421739

0.7299

0.465995

I(lLabor * lNpk)

-0.2055786

0.1277476 -1.6093

0.108508

0.0070184

0.043549 *

mYear

0.0142202

2.0261

231

6 Panel Data and Technological Change

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Total Sum of Squares:

114.08

Residual Sum of Squares: 26.624

R-Squared

0.76662

Adj. R-Squared :

0.74211

F-statistic: 109.386 on 10 and 333 DF, p-value: < 2.22e-16

The Translog production function cannot be estimated by a variable-coefficient model for panel
model with our data set, because the number of time periods in the data set is smaller than the
number of the coefficients.
A pooled estimation can be done by
> riceTlTimePool <- plm( lProd ~ lArea + lLabor + lNpk +
+

I( 0.5 * lArea^2 ) + I( 0.5 * lLabor^2 ) + I( 0.5 * lNpk^2 ) +

I( lArea * lLabor ) + I( lArea * lNpk ) + I( lLabor * lNpk ) + mYear,

data = pdat, model = "pooling" )

> summary(riceTlTimePool)
Oneway (individual) effect Pooling Model
Call:
plm(formula = lProd ~ lArea + lLabor + lNpk + I(0.5 * lArea^2) +
I(0.5 * lLabor^2) + I(0.5 * lNpk^2) + I(lArea * lLabor) +
I(lArea * lNpk) + I(lLabor * lNpk) + mYear, data = pdat,
model = "pooling")
Balanced Panel: n=43, T=8, N=344
Residuals :
Min. 1st Qu.
-1.5200 -0.1810

Median 3rd Qu.

0.0436

0.2230

Max.
0.8700

Coefficients :
Estimate Std. Error t-value

Pr(>|t|)

(Intercept)

0.0137557

0.0246454

0.5581 0.5771201

lArea

0.5880972

0.0851622

6.9056 2.542e-11 ***

lLabor

0.1917638

0.0808764

2.3711 0.0183052 *

lNpk

0.1978747

0.0516045

3.8344 0.0001505 ***

I(0.5 * lArea^2)

-0.4355466

0.2474913 -1.7598 0.0793520 .

232

6 Panel Data and Technological Change

I(0.5 * lLabor^2) -0.7422415

0.3032362 -2.4477 0.0148916 *

I(0.5 * lNpk^2)

0.0203673

0.0979072

0.2080 0.8353358

I(lArea * lLabor)

0.6786472

0.2165937

3.1333 0.0018822 **

I(lArea * lNpk)

0.0639200

0.1456135

0.4390 0.6609677

I(lLabor * lNpk)

-0.1782859

mYear

0.0126820

0.1386111 -1.2862 0.1992559

0.0077947

1.6270 0.1046801

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Total Sum of Squares:

263.52

Residual Sum of Squares: 33.761

R-Squared

0.87189

Adj. R-Squared :

0.84401

F-statistic: 226.623 on 10 and 333 DF, p-value: < 2.22e-16

This gives the same estimated coefficients as the model estimated by lm:
> all.equal( coef( riceTlTime ), coef( riceTlTimePool ) )
[1] TRUE
A Hausman test can be used to check the consistency of the random-effects estimator:
> phtest( riceTlTimeRan, riceTlTimeFe )
Hausman Test
data:

lProd ~ lArea + lLabor + lNpk + I(0.5 * lArea^2) + I(0.5 * lLabor^2) +

I(0.5 * lNp

chisq = 66.071, df = 10, p-value = 2.528e-10

alternative hypothesis: one model is inconsistent
The Hausman test clearly rejects the consistency of the random-effects estimator.
The following command tests the poolability of the model:
> pooltest( riceTlTimePool, riceTlTimeFe )
F statistic
data:

lProd ~ lArea + lLabor + lNpk + I(0.5 * lArea^2) + I(0.5 * lLabor^2) +

F = 3.7469, df1 = 42, df2 = 291, p-value = 1.525e-11

alternative hypothesis: unstability

233

I(0.5 * lNp

6 Panel Data and Technological Change

The pooled model (riceCdTimePool) is clearly rejected in favour of the model with fixed individual effects (riceCdTimeFe), i.e. the individual effects are statistically significant.
The following commands test if the fit of Translog specification is significantly better than the
fit of the Cobb-Douglas specification:
> waldtest( riceCdTimeFe, riceTlTimeFe )
Wald test
Model 1: lProd ~ lArea + lLabor + lNpk + mYear
Model 2: lProd ~ lArea + lLabor + lNpk + I(0.5 * lArea^2) + I(0.5 * lLabor^2) +
I(0.5 * lNpk^2) + I(lArea * lLabor) + I(lArea * lNpk) + I(lLabor *
lNpk) + mYear
Res.Df Df
1

297

291

Chisq Pr(>Chisq)

6 39.321

6.191e-07 ***

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

> waldtest( riceCdTimeRan, riceTlTimeRan )

Wald test
Model 1: lProd ~ lArea + lLabor + lNpk + mYear
Model 2: lProd ~ lArea + lLabor + lNpk + I(0.5 * lArea^2) + I(0.5 * lLabor^2) +
I(0.5 * lNpk^2) + I(lArea * lLabor) + I(lArea * lNpk) + I(lLabor *
lNpk) + mYear
Res.Df Df
1

339

333

Chisq Pr(>Chisq)

6 30.077

3.8e-05 ***

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

> waldtest( riceCdTimePool, riceTlTimePool )

Wald test
Model 1: lProd ~ lArea + lLabor + lNpk + mYear
Model 2: lProd ~ lArea + lLabor + lNpk + I(0.5 * lArea^2) + I(0.5 * lLabor^2) +
I(0.5 * lNpk^2) + I(lArea * lLabor) + I(lArea * lNpk) + I(lLabor *
lNpk) + mYear
Res.Df Df Chisq Pr(>Chisq)

234

6 Panel Data and Technological Change

1

339

333

6 30.89

2.66e-05 ***

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

The Cobb-Douglas functional form is rejected in favour of the Translog functional for for all three
panel-specifications that we estimated above. The Wald test for the pooled model differs from
the Wald test that we did in section 6.1.2.1, because waldtest by default uses a finite sample
F statistic for models estimated by lm but uses a large sample Chi-squared statistic for models
estimated by plm. The test statistic used by waldtest can be specified by argument test.

6.1.3 Translog Production Function with Non-Constant and Non-Neutral

Technological Change
Technological change is not always constant and is not always neutral (unbiased). Therefore,
it might be more suitable to estimate a production function that can account for increasing or
decreasing rates of technological change as well as biased (e.g. labor saving) technological change.
This can be done by including a quadratic time trend and interaction terms between time and
input quantities:
ln y = 0 +

i ln xi +

X
1 XX
1
ij ln xi ln xj + t t +
ti ln xi + tt t2
2 i j
2
i

(6.7)

In this specification, the rate of technological change depends on the input quantities and the
time period:
X
ln y
= t +
ti ln xi + tt t
t
i

(6.8)

and the output elasticities might change over time:

i =

X
ln y
ij ln xj + ti t.
= i +
ln xi
j

(6.9)

6.1.3.1 Pooled Estimation of a Translog Production Function with Non-Constant and

Non-Neutral Technological Change
The following command estimates a Translog production function that can account for nonconstant rates of technological change as well as biased technological change:
> riceTlTimeNn <- lm( lProd ~ lArea + lLabor + lNpk +
+

I( 0.5 * lArea^2 ) + I( 0.5 * lLabor^2 ) + I( 0.5 * lNpk^2 ) +

I( lArea * lLabor ) + I( lArea * lNpk ) + I( lLabor * lNpk ) +

mYear + I( mYear * lArea ) + I( mYear * lLabor ) + I( mYear * lNpk ) +

235

6 Panel Data and Technological Change

+

I( 0.5 * mYear^2 ), data = riceProdPhil )

> summary( riceTlTimeNn )

Call:
lm(formula = lProd ~ lArea + lLabor + lNpk + I(0.5 * lArea^2) +
I(0.5 * lLabor^2) + I(0.5 * lNpk^2) + I(lArea * lLabor) +
I(lArea * lNpk) + I(lLabor * lNpk) + mYear + I(mYear * lArea) +
I(mYear * lLabor) + I(mYear * lNpk) + I(0.5 * mYear^2), data = riceProdPhil)
Residuals:
Min

1Q

Median

3Q

Max

-1.54976 -0.17245

0.04623

0.21624

0.87075

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept)

0.001255

0.031934

0.039

lArea

0.579682

0.085892

6.749 6.73e-11 ***

lLabor

0.187505

0.081359

2.305

lNpk

0.207193

0.052130

3.975 8.67e-05 ***

-0.468372

0.265363

-1.765

0.07849 .

I(0.5 * lLabor^2) -0.688940

0.308046

-2.236

0.02599 *

I(0.5 * lNpk^2)

0.055993

0.099848

0.561

0.57533

I(lArea * lLabor)

0.676833

0.223271

3.031

0.00263 **

I(lArea * lNpk)

0.082374

0.151312

0.544

0.58654

I(lLabor * lNpk)

-0.226885

0.145568

-1.559

0.12005

mYear

0.008746

0.008513

1.027

0.30497

I(mYear * lArea)

0.003482

0.028075

0.124

0.90136

I(mYear * lLabor)

0.034661

0.029480

1.176

0.24054

I(mYear * lNpk)

-0.037964

0.020355

-1.865

I(0.5 * mYear^2)

0.007611

0.007954

0.957

I(0.5 * lArea^2)

0.96867
0.02181 *

0.06305 .
0.33933

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Residual standard error: 0.3184 on 329 degrees of freedom

Multiple R-squared:

0.8734,

Adjusted R-squared:

F-statistic: 162.2 on 14 and 329 DF,

0.868

p-value: < 2.2e-16

We conduct a Wald test to test whether the Translog production function with non-constant
and non-neutral technological change outperforms the Cobb-Douglas production function and
the Translog production function with constant and neutral technological change:

236

6 Panel Data and Technological Change

> waldtest( riceCdTimeS, riceTlTimeNn )
Wald test
Model 1: lProd ~ lArea + lLabor + lNpk + mYear
Model 2: lProd ~ lArea + lLabor + lNpk + I(0.5 * lArea^2) + I(0.5 * lLabor^2) +
I(0.5 * lNpk^2) + I(lArea * lLabor) + I(lArea * lNpk) + I(lLabor *
lNpk) + mYear + I(mYear * lArea) + I(mYear * lLabor) + I(mYear *
lNpk) + I(0.5 * mYear^2)
Res.Df Df

Pr(>F)

339

329 10 3.488 0.00022 ***

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

> waldtest( riceTlTime, riceTlTimeNn )

Wald test
Model 1: lProd ~ lArea + lLabor + lNpk + I(0.5 * lArea^2) + I(0.5 * lLabor^2) +
I(0.5 * lNpk^2) + I(lArea * lLabor) + I(lArea * lNpk) + I(lLabor *
lNpk) + mYear
Model 2: lProd ~ lArea + lLabor + lNpk + I(0.5 * lArea^2) + I(0.5 * lLabor^2) +
I(0.5 * lNpk^2) + I(lArea * lLabor) + I(lArea * lNpk) + I(lLabor *
lNpk) + mYear + I(mYear * lArea) + I(mYear * lLabor) + I(mYear *
lNpk) + I(0.5 * mYear^2)
Res.Df Df
1

333

329

F Pr(>F)

4 0.9976 0.4089

The fit of the Translog specification with non-constant and non-neutral technological change is
significantly better than the fit of the Cobb-Douglas specification but it is not significantly better
than the fit of the Translog specification with constant and neutral technological change.
In order to simplify the calculation of the output elasticities (with equation 6.9) and the
annual rates of technological change (with equation 6.8), we create shortcuts for the estimated
coefficients:
> a1 <- coef( riceTlTimeNn )[ "lArea" ]
> a2 <- coef( riceTlTimeNn )[ "lLabor" ]
> a3 <- coef( riceTlTimeNn )[ "lNpk" ]
> at <- coef( riceTlTimeNn )[ "mYear" ]

237

6 Panel Data and Technological Change

> a11 <- coef( riceTlTimeNn )[ "I(0.5 * lArea^2)" ]
> a22 <- coef( riceTlTimeNn )[ "I(0.5 * lLabor^2)" ]
> a33 <- coef( riceTlTimeNn )[ "I(0.5 * lNpk^2)" ]
> att <- coef( riceTlTimeNn )[ "I(0.5 * mYear^2)" ]
> a12 <- a21 <- coef( riceTlTimeNn )[ "I(lArea * lLabor)" ]
> a13 <- a31 <- coef( riceTlTimeNn )[ "I(lArea * lNpk)" ]
> a23 <- a32 <- coef( riceTlTimeNn )[ "I(lLabor * lNpk)" ]
> a1t <- at1 <- coef( riceTlTimeNn )[ "I(mYear * lArea)" ]
> a2t <- at2 <- coef( riceTlTimeNn )[ "I(mYear * lLabor)" ]
> a3t <- at3 <- coef( riceTlTimeNn )[ "I(mYear * lNpk)" ]
Now, we can use the following commands to calculate the partial output elasticities:
> riceProdPhil$eArea <- with( riceProdPhil,
+

a1 + a11 * lArea + a12 * lLabor + a13 * lNpk + a1t * mYear )

> riceProdPhil$eLabor <- with( riceProdPhil,

+

a2 + a21 * lArea + a22 * lLabor + a23 * lNpk + a2t * mYear )

> riceProdPhil$eNpk <- with( riceProdPhil,

+

a3 + a31 * lArea + a32 * lLabor + a33 * lNpk + a3t * mYear )

We can calculate the elasticity of scale by taken the sum over all partial output elasticities:
> riceProdPhil$eScale <- with( riceProdPhil, eArea + eLabor + eNpk )
We can visualize (the variation of) the output elasticities and the elasticity of scale with
histograms:
> hist( riceProdPhil$eArea, 15 )
> hist( riceProdPhil$eLabor, 15 )
> hist( riceProdPhil$eNpk, 15 )
> hist( riceProdPhil$eScale, 15 )
The resulting graphs are shown in figure 6.1. If the firms increase the land area by one percent,
the output of most firms will increase by around 0.6 percent. If the firms increase labor input by
one percent, the output of most firms will increase by around 0.2 percent. If the firms increase
fertilizer input by one percent, the output of most firms will increase by around 0.25 percent. If
the firms increase all input quantities by one percent, the output of most firms will also increase
by around 1 percent. These graphs also show that the monotonicity condition is not fulfilled for
some observations:
> sum( riceProdPhil$eArea < 0 )
[1] 20

238

40
0

20

Frequency

40
20
0

Frequency

60

6 Panel Data and Technological Change

0.5

0.0

0.5

1.0

0.5

0.0

0.5

1.5

60
20
0

20

40

Frequency

eLabor

Frequency

eArea

1.0

0.1

0.1

0.3

0.5

0.8

1.0

eNpk

eScale

Figure 6.1: Output elasticities and elasticities of scale

239

1.2

1.4

6 Panel Data and Technological Change

> sum( riceProdPhil$eLabor < 0 )
[1] 63
> sum( riceProdPhil$eNpk < 0 )
[1] 7
> riceProdPhil$monoTl <- with( riceProdPhil, eArea >0 & eLabor > 0 & eNpk > 0 )
> sum( !riceProdPhil$monoTl )
[1] 85
20 firms have a negative output elasticity of the land area, 63 firms have a negative output elasticity of labor, and 7 firms have a negative output elasticity of fertilizers. In total the monotonicity
condition is not fulfilled at 85 out of 344 observations. Although the monotonicity conditions
are fulfilled for a large part of firms in our data set, these frequent violations indicate a possible
model misspecification.
We can use the following command to calculate the annual rates of technological change:
> riceProdPhil$tc <- with( riceProdPhil,
+

at + at1 * lArea + at2 * lLabor + at3 * lNpk + att * mYear )

We can visualize (the variation of) the annual rates of technological change with a histogram:

40
20
0

Frequency

> hist( riceProdPhil$tc, 15 )

0.05

0.00

0.05

0.10

tc

Figure 6.2: Annual rates of technological change

The resulting graph is shown in figure 6.2. For most observations, the annual rate of technological
change was between 0% and 3%.

240

6 Panel Data and Technological Change

6.1.3.2 Panel-data estimations of a Translog Production Function with Non-Constant and
Non-Neutral Technological Change
The panel data estimation with fixed individual effects can be done by:
> riceTlTimeNnFe <- plm( lProd ~ lArea + lLabor + lNpk +
+

I( 0.5 * lArea^2 ) + I( 0.5 * lLabor^2 ) + I( 0.5 * lNpk^2 ) +

I( lArea * lLabor ) + I( lArea * lNpk ) + I( lLabor * lNpk ) +

mYear + I( mYear * lArea ) + I( mYear * lLabor ) + I( mYear * lNpk ) +

I( 0.5 * mYear^2 ), data = pdat )

> summary( riceTlTimeNnFe )

Oneway (individual) effect Within Model
Call:
plm(formula = lProd ~ lArea + lLabor + lNpk + I(0.5 * lArea^2) +
I(0.5 * lLabor^2) + I(0.5 * lNpk^2) + I(lArea * lLabor) +
I(lArea * lNpk) + I(lLabor * lNpk) + mYear + I(mYear * lArea) +
I(mYear * lLabor) + I(mYear * lNpk) + I(0.5 * mYear^2), data = pdat)
Balanced Panel: n=43, T=8, N=344
Residuals :
Min. 1st Qu.
-1.0100 -0.1430

Median 3rd Qu.

0.0175

0.1670

Max.
0.7490

Coefficients :
Estimate Std. Error t-value

Pr(>|t|)

lArea

0.5857359

0.1191164

4.9173 1.479e-06 ***

lLabor

0.0336966

0.0869044

0.3877

0.698494

lNpk

0.1276970

0.0623919

2.0467

0.041599 *

I(0.5 * lArea^2)

-0.8588620

0.2952677 -2.9088

0.003912 **

I(0.5 * lLabor^2) -0.6154568

0.2979094 -2.0659

0.039733 *

I(0.5 * lNpk^2)

0.0673038

0.1014542

0.6634

0.507613

I(lArea * lLabor)

0.6016538

0.2164953

2.7791

0.005811 **

I(lArea * lNpk)

0.1205064

0.1549834

0.7775

0.437479

I(lLabor * lNpk)

-0.2660519

0.1353699 -1.9654

0.050336 .

mYear

0.0148796

0.0076143

1.9542

0.051654 .

I(mYear * lArea)

0.0105012

0.0270130

0.3887

0.697752

I(mYear * lLabor)

0.0230156

0.0286066

0.8046

0.421743

0.0199045 -1.4044

0.161277

I(mYear * lNpk)

-0.0279542

241

6 Panel Data and Technological Change

I(0.5 * mYear^2)

0.0058526

0.0069948

0.8367

0.403458

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Total Sum of Squares:

43.632

Residual Sum of Squares: 21.733

R-Squared

0.50189

Adj. R-Squared :

0.41872

F-statistic: 20.6552 on 14 and 287 DF, p-value: < 2.22e-16

And the panel data estimation with random individual effects can be done by:
> riceTlTimeNnRan <- plm( lProd ~ lArea + lLabor + lNpk +
+

I( 0.5 * lArea^2 ) + I( 0.5 * lLabor^2 ) + I( 0.5 * lNpk^2 ) +

I( lArea * lLabor ) + I( lArea * lNpk ) + I( lLabor * lNpk ) +

mYear + I( mYear * lArea ) + I( mYear * lLabor ) + I( mYear * lNpk ) +

I( 0.5 * mYear^2 ), data = pdat, model = "random" )

> summary( riceTlTimeNnRan )

Oneway (individual) effect Random Effect Model
(Swamy-Arora's transformation)
Call:
plm(formula = lProd ~ lArea + lLabor + lNpk + I(0.5 * lArea^2) +
I(0.5 * lLabor^2) + I(0.5 * lNpk^2) + I(lArea * lLabor) +
I(lArea * lNpk) + I(lLabor * lNpk) + mYear + I(mYear * lArea) +
I(mYear * lLabor) + I(mYear * lNpk) + I(0.5 * mYear^2), data = pdat,
model = "random")
Balanced Panel: n=43, T=8, N=344
Effects:
var std.dev share
idiosyncratic 0.07573 0.27518 0.796
individual
theta:

0.01941 0.13933 0.204

0.4275

Residuals :
Min. 1st Qu.
-1.3900 -0.1620

Median 3rd Qu.

0.0456

0.1800

Max.
0.7900

242

6 Panel Data and Technological Change

Coefficients :
Estimate Std. Error t-value

Pr(>|t|)

(Intercept)

0.0101183

0.0389961

0.2595

lArea

0.6809764

0.0930789

7.3161 1.965e-12 ***

lLabor

0.0865327

0.0813309

1.0640

0.288128

lNpk

0.1800677

0.0554226

3.2490

0.001278 **

I(0.5 * lArea^2)

0.795434

-0.4749163

0.2627102 -1.8078

0.071557 .

I(0.5 * lLabor^2) -0.6146891

0.2907148 -2.1144

0.035232 *

I(0.5 * lNpk^2)

0.0614961

0.0980315

0.6273

0.530891

I(lArea * lLabor)

0.5916989

0.2113078

2.8002

0.005409 **

I(lArea * lNpk)

0.1224789

0.1488815

0.8227

0.411297

I(lLabor * lNpk)

-0.2531048

0.1350400 -1.8743

0.061776 .

mYear

0.0116511

0.0077140

1.5104

0.131907

I(mYear * lArea)

0.0028675

0.0265731

0.1079

0.914134

I(mYear * lLabor)

0.0355897

0.0279156

1.2749

0.203242

I(mYear * lNpk)

-0.0344049

I(0.5 * mYear^2)

0.0069525

0.0195392 -1.7608

0.079198 .

0.0071510

0.331650

0.9722

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Total Sum of Squares:

115.71

Residual Sum of Squares: 26.417

R-Squared

0.77169

Adj. R-Squared :

0.73804

F-statistic: 79.4317 on 14 and 329 DF, p-value: < 2.22e-16

The Translog production function cannot be estimated by a variable-coefficient model for panel
model with our data set, because the number of time periods in the data set is smaller than the
number of the coefficients.
A pooled estimation can be done by
> riceTlTimeNnPool <- plm( lProd ~ lArea + lLabor + lNpk +
+

I( 0.5 * lArea^2 ) + I( 0.5 * lLabor^2 ) + I( 0.5 * lNpk^2 ) +

I( lArea * lLabor ) + I( lArea * lNpk ) + I( lLabor * lNpk ) +

mYear + I( mYear * lArea ) + I( mYear * lLabor ) + I( mYear * lNpk ) +

I( 0.5 * mYear^2 ), data = pdat, model = "pooling" )

This gives the same estimated coefficients as the model estimated by lm:

> all.equal( coef( riceTlTimeNn ), coef( riceTlTimeNnPool ) )

[1] TRUE

243

6 Panel Data and Technological Change

A Hausman test can be used to check the consistency of the random-effects estimator:
> phtest( riceTlTimeNnRan, riceTlTimeNnFe )
Hausman Test
data:

lProd ~ lArea + lLabor + lNpk + I(0.5 * lArea^2) + I(0.5 * lLabor^2) +

I(0.5 * lNp

chisq = 21.7306, df = 14, p-value = 0.08432

alternative hypothesis: one model is inconsistent
The Hausman test rejects the consistency of the random-effects estimator at the 10% significance
level but it cannot reject the consistency of the random-effects estimator at the 5% significance
level.
The following command tests the poolability of the model:
> pooltest( riceTlTimeNnPool, riceTlTimeNnFe )
F statistic
data:

lProd ~ lArea + lLabor + lNpk + I(0.5 * lArea^2) + I(0.5 * lLabor^2) +

F = 3.6544, df1 = 42, df2 = 287, p-value = 4.266e-11

alternative hypothesis: unstability
The pooled model (riceCdTimePool) is clearly rejected in favor of the model with fixed individual
effects (riceCdTimeFe), i.e. the individual effects are statistically significant.
The following commands test if the fit of Translog specification is significantly better than the
fit of the Cobb-Douglas specification:
> waldtest( riceTlTimeNnFe, riceCdTimeFe )
Wald test
Model 1: lProd ~ lArea + lLabor + lNpk + I(0.5 * lArea^2) + I(0.5 * lLabor^2) +
I(0.5 * lNpk^2) + I(lArea * lLabor) + I(lArea * lNpk) + I(lLabor *
lNpk) + mYear + I(mYear * lArea) + I(mYear * lLabor) + I(mYear *
lNpk) + I(0.5 * mYear^2)
Model 2: lProd ~ lArea + lLabor + lNpk + mYear
Res.Df

Df Chisq Pr(>Chisq)

287

297 -10 41.45

9.392e-06 ***

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

244

I(0.5 * lNp

6 Panel Data and Technological Change

> waldtest( riceTlTimeNnRan, riceCdTimeRan )
Wald test
Model 1: lProd ~ lArea + lLabor + lNpk + I(0.5 * lArea^2) + I(0.5 * lLabor^2) +
I(0.5 * lNpk^2) + I(lArea * lLabor) + I(lArea * lNpk) + I(lLabor *
lNpk) + mYear + I(mYear * lArea) + I(mYear * lLabor) + I(mYear *
lNpk) + I(0.5 * mYear^2)
Model 2: lProd ~ lArea + lLabor + lNpk + mYear
Res.Df

Df

Chisq Pr(>Chisq)

329

339 -10 33.666

0.0002103 ***

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

> waldtest( riceTlTimeNnPool, riceCdTimePool )

Wald test
Model 1: lProd ~ lArea + lLabor + lNpk + I(0.5 * lArea^2) + I(0.5 * lLabor^2) +
I(0.5 * lNpk^2) + I(lArea * lLabor) + I(lArea * lNpk) + I(lLabor *
lNpk) + mYear + I(mYear * lArea) + I(mYear * lLabor) + I(mYear *
lNpk) + I(0.5 * mYear^2)
Model 2: lProd ~ lArea + lLabor + lNpk + mYear
Res.Df

Df Chisq Pr(>Chisq)

329

339 -10 34.88

0.0001309 ***

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Finally, we test whether the fit of Translog specification with non-constant and non-neutral
technological change is significantly better than the fit of Translog specification with constant
and neutral technological change:
> waldtest( riceTlTimeNnFe, riceTlTimeFe )
Wald test
Model 1: lProd ~ lArea + lLabor + lNpk + I(0.5 * lArea^2) + I(0.5 * lLabor^2) +
I(0.5 * lNpk^2) + I(lArea * lLabor) + I(lArea * lNpk) + I(lLabor *
lNpk) + mYear + I(mYear * lArea) + I(mYear * lLabor) + I(mYear *
lNpk) + I(0.5 * mYear^2)

245

6 Panel Data and Technological Change

Model 2: lProd ~ lArea + lLabor + lNpk + I(0.5 * lArea^2) + I(0.5 * lLabor^2) +
I(0.5 * lNpk^2) + I(lArea * lLabor) + I(lArea * lNpk) + I(lLabor *
lNpk) + mYear
Res.Df Df

Chisq Pr(>Chisq)

287

291 -4 2.3512

0.6715

> waldtest( riceTlTimeNnRan, riceTlTimeRan )

Wald test
Model 1: lProd ~ lArea + lLabor + lNpk + I(0.5 * lArea^2) + I(0.5 * lLabor^2) +
I(0.5 * lNpk^2) + I(lArea * lLabor) + I(lArea * lNpk) + I(lLabor *
lNpk) + mYear + I(mYear * lArea) + I(mYear * lLabor) + I(mYear *
lNpk) + I(0.5 * mYear^2)
Model 2: lProd ~ lArea + lLabor + lNpk + I(0.5 * lArea^2) + I(0.5 * lLabor^2) +
I(0.5 * lNpk^2) + I(lArea * lLabor) + I(lArea * lNpk) + I(lLabor *
lNpk) + mYear
Res.Df Df

Chisq Pr(>Chisq)

329

333 -4 3.6633

0.4535

> waldtest( riceTlTimeNnPool, riceTlTimePool )

Wald test
Model 1: lProd ~ lArea + lLabor + lNpk + I(0.5 * lArea^2) + I(0.5 * lLabor^2) +
I(0.5 * lNpk^2) + I(lArea * lLabor) + I(lArea * lNpk) + I(lLabor *
lNpk) + mYear + I(mYear * lArea) + I(mYear * lLabor) + I(mYear *
lNpk) + I(0.5 * mYear^2)
Model 2: lProd ~ lArea + lLabor + lNpk + I(0.5 * lArea^2) + I(0.5 * lLabor^2) +
I(0.5 * lNpk^2) + I(lArea * lLabor) + I(lArea * lNpk) + I(lLabor *
lNpk) + mYear
Res.Df Df

Chisq Pr(>Chisq)

329

333 -4 3.9905

0.4073

The tests indicate that the fit of Translog specification with constant and neutral technological
change is not significantly worse than the fit of Translog specification with non-constant and
non-neutral technological change.
The difference between the Wald tests for the pooled model and the Wald test that we did in
section 6.1.3.1 is explained at the end of section 6.1.2.2.

246

6 Panel Data and Technological Change

6.2 Frontier Production Functions with Technological Change

The frontier production technology can be estimated by many different specifications of the
stochastic frontier model. We will focus on three specifications that are all nested in the general
specification:
ln ykt = ln f (xkt , t) ukt + vkt ,

(6.10)

where the subscript k = 1, . . . , K indicates the firm, t = 1, . . . , T indicates the time period, and
all other variables are defined as before. We will apply the following three model specifications:
1. time-invariant individual efficiencies, i.e. ukt = uk , which means that each firm has an
individual fixed efficiency that does not vary over time;
2. time-variant individual efficiencies, i.e. ukt = uk exp( (t T )), which means that each
firm has an individual efficiency and the efficiency terms of all firms can vary over time
with the same rate (and in the same direction); and
3. observation-specific efficiencies, i.e. no restrictions on ukt , which means that the efficiency
term of each observation is estimated independently from the other efficiencies of the firm
so that basically the panel structure of the data is ignored.

6.2.1 Cobb-Douglas Production Frontier with Technological Change

We will use the specification in equation (6.2).
6.2.1.1 Time-invariant Individual Efficiencies
We start with estimating a Cobb-Douglas production frontier with time-invariant individual
efficiencies. The following commands estimate two Cobb-Douglas production frontiers with timeinvariant individual efficiencies, the first does not account for technological change, while the
second does:
> riceCdSfaInv <- sfa( lProd ~ lArea + lLabor + lNpk, data = pdat )
> summary( riceCdSfaInv )
Error Components Frontier (see Battese & Coelli 1992)
Inefficiency decreases the endogenous variable (as in a production function)
The dependent variable is logged
Iterative ML estimation terminated after 9 iterations:
log likelihood values and parameters of two successive iterations
are within the tolerance limit
final maximum likelihood estimates
Estimate Std. Error z value

Pr(>|z|)

247

6 Panel Data and Technological Change

(Intercept) 0.182636

0.035340

5.1680 2.366e-07 ***

lArea

0.453900

0.064382

7.0501 1.788e-12 ***

lLabor

0.288922

0.063853

4.5248 6.045e-06 ***

lNpk

0.227542

0.040644

5.5983 2.164e-08 ***

sigmaSq

0.155377

0.024144

6.4354 1.232e-10 ***

gamma

0.464317

0.088270

5.2602 1.439e-07 ***

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

log likelihood value: -86.43042

panel data
number of cross-sections = 43
number of time periods = 8
total number of observations = 344
thus there are 0 observations not in the panel
mean efficiency: 0.8187935
> riceCdTimeSfaInv <- sfa( lProd ~ lArea + lLabor + lNpk + mYear, data = pdat )
> summary( riceCdTimeSfaInv )
Error Components Frontier (see Battese & Coelli 1992)
Inefficiency decreases the endogenous variable (as in a production function)
The dependent variable is logged
Iterative ML estimation terminated after 11 iterations:
log likelihood values and parameters of two successive iterations
are within the tolerance limit
final maximum likelihood estimates
Estimate Std. Error z value

Pr(>|z|)

(Intercept) 0.1832751

0.0345895

5.2986 1.167e-07 ***

lArea

0.4625174

0.0644245

7.1792 7.011e-13 ***

lLabor

0.3029415

0.0641323

4.7237 2.316e-06 ***

lNpk

0.2098907

0.0418709

5.0128 5.364e-07 ***

mYear

0.0116003

0.0071758

1.6166

sigmaSq

0.1556806

0.0242951

6.4079 1.475e-10 ***

gamma

0.4706143

0.0869549

5.4122 6.227e-08 ***

0.106

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

log likelihood value: -85.0743

248

6 Panel Data and Technological Change

panel data
number of cross-sections = 43
number of time periods = 8
total number of observations = 344
thus there are 0 observations not in the panel
mean efficiency: 0.8176333
In the Cobb-Douglas production frontier that accounts for technological change, the monotonicity
conditions are globally fulfilled and the (constant) output elasticities of land, labor and fertilizer
are 0.463, 0.303, and 0.21, respectively. The estimated (constant) annual rate of technological
progress is around 1.2%. However, both the t-test for the coefficient of the time trend and a
likelihood ratio test give rise to doubts whether the production technology indeed changes over
time (P-values around 10%):
> lrtest( riceCdTimeSfaInv, riceCdSfaInv )
Likelihood ratio test
Model 1: riceCdTimeSfaInv
Model 2: riceCdSfaInv
#Df

LogLik Df

Chisq Pr(>Chisq)

7 -85.074

6 -86.430 -1 2.7122

0.09958 .

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Further likelihood ratio tests show that OLS models are clearly rejected in favor of the corresponding stochastic frontier models (no matter whether the production frontier accounts for
technological change or not):
> lrtest( riceCdSfaInv )
Likelihood ratio test
Model 1: OLS (no inefficiency)
Model 2: Error Components Frontier (ECF)
#Df

LogLik Df

5 -104.91

-86.43

Chisq Pr(>Chisq)

1 36.953

6.051e-10 ***

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

249

6 Panel Data and Technological Change

> lrtest( riceCdTimeSfaInv )
Likelihood ratio test
Model 1: OLS (no inefficiency)
Model 2: Error Components Frontier (ECF)
#Df

LogLik Df

6 -104.103

-85.074

Chisq Pr(>Chisq)

1 38.057

3.434e-10 ***

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

This model estimates only a single efficiency estimate for each of the 43 firms. Hence, the vector
returned by the efficiencies method only has 43 elements by default:
> length( efficiencies( riceCdSfaInv ) )
[1] 43
One can obtain the efficiency estimates for each observation by setting argument asInData equal
to TRUE:
> pdat$effCdInv <- efficiencies( riceCdSfaInv, asInData = TRUE )
Please note that the efficiency estimates for each firm still do not vary between time periods.
6.2.1.2 Time-variant Individual Efficiencies
Now we estimate a Cobb-Douglas production frontier with time-variant individual efficiencies.
Again, we estimate two Cobb-Douglas production frontiers, the first does not account for technological change, while the second does:
> riceCdSfaVar <- sfa( lProd ~ lArea + lLabor + lNpk,
+

timeEffect = TRUE, data = pdat )

> summary( riceCdSfaVar )

Error Components Frontier (see Battese & Coelli 1992)
Inefficiency decreases the endogenous variable (as in a production function)
The dependent variable is logged
Iterative ML estimation terminated after 11 iterations:
log likelihood values and parameters of two successive iterations
are within the tolerance limit
final maximum likelihood estimates

250

6 Panel Data and Technological Change

Estimate Std. Error z value

Pr(>|z|)

(Intercept) 0.182016

0.035251

5.1635 2.424e-07 ***

lArea

0.474919

0.066213

7.1726 7.360e-13 ***

lLabor

0.300094

0.063872

4.6983 2.623e-06 ***

lNpk

0.199461

0.042740

4.6669 3.058e-06 ***

sigmaSq

0.129957

0.021098

6.1598 7.285e-10 ***

gamma

0.369639

0.104045

3.5527 0.0003813 ***

time

0.058909

0.030863

1.9087 0.0563017 .

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

log likelihood value: -84.55036

panel data
number of cross-sections = 43
number of time periods = 8
total number of observations = 344
thus there are 0 observations not in the panel
mean efficiency of each year
1

0.7848433 0.7950303 0.8048362 0.8142652 0.8233226 0.8320146 0.8403483 0.8483313

mean efficiency: 0.817874
> riceCdTimeSfaVar <- sfa( lProd ~ lArea + lLabor + lNpk + mYear,
+

timeEffect = TRUE, data = pdat )

> summary( riceCdTimeSfaVar )

Error Components Frontier (see Battese & Coelli 1992)
Inefficiency decreases the endogenous variable (as in a production function)
The dependent variable is logged
Iterative ML estimation terminated after 14 iterations:
log likelihood values and parameters of two successive iterations
are within the tolerance limit
final maximum likelihood estimates
Estimate Std. Error z value

Pr(>|z|)

(Intercept)

0.1817461

0.0358969

5.0630 4.127e-07 ***

lArea

0.4761159

0.0654855

7.2706 3.580e-13 ***

lLabor

0.2987926

0.0646926

4.6187 3.862e-06 ***

251

6 Panel Data and Technological Change

lNpk

0.1991403

mYear

-0.0031916

0.0428435

4.6481 3.350e-06 ***

0.0150900 -0.2115

0.83249

sigmaSq

0.1255571

0.0287422

4.3684 1.252e-05 ***

gamma

0.3478592

0.1488417

2.3371

0.01943 *

time

0.0711209

0.0657939

1.0810

0.27971

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

log likelihood value: -84.52871

panel data
number of cross-sections = 43
number of time periods = 8
total number of observations = 344
thus there are 0 observations not in the panel
mean efficiency of each year
1

0.7780268 0.7905799 0.8025751 0.8140192 0.8249213 0.8352924 0.8451453 0.8544942

mean efficiency: 0.8181318
In the Cobb-Douglas production frontier that accounts for technological change, the monotonicity
conditions are globally fulfilled and the (constant) output elasticities of land, labor and fertilizer
are 0.476, 0.299, and 0.199, respectively. The estimated (constant) annual rate of technological change is around -0.3%, which indicates technological regress. However, the t-test for the
coefficient of the time trend and a likelihood ratio test indicate that the production technology
(frontier) does not change over time, i.e. there is neither technological regress nor technological
progress:
> lrtest( riceCdTimeSfaVar, riceCdSfaVar )
Likelihood ratio test
Model 1: riceCdTimeSfaVar
Model 2: riceCdSfaVar
#Df

LogLik Df

Chisq Pr(>Chisq)

8 -84.529

7 -84.550 -1 0.0433

0.8352

A positive sign of the coefficient (named time) indicates that efficiency is increasing over
time. However, in the model without technological change, the t-test for the coefficient and

252

6 Panel Data and Technological Change

the corresponding likelihood ratio test indicate that the effect of time on the efficiencies only is
significant at the 10% level:
> lrtest( riceCdSfaInv, riceCdSfaVar )
Likelihood ratio test
Model 1: riceCdSfaInv
Model 2: riceCdSfaVar
#Df LogLik Df
1

6 -86.43

7 -84.55

Chisq Pr(>Chisq)

1 3.7601

0.05249 .

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

In the model that accounts for technological change, the t-test for the coefficient and the
corresponding likelihood ratio test indicate that the efficiencies do not change over time:
> lrtest( riceCdTimeSfaInv, riceCdTimeSfaVar )
Likelihood ratio test
Model 1: riceCdTimeSfaInv
Model 2: riceCdTimeSfaVar
#Df

LogLik Df

7 -85.074

8 -84.529

Chisq Pr(>Chisq)

1 1.0912

0.2962

Finally, we can use a likelihood ratio test to simultaneously test whether the technology and the
technical efficiencies change over time:
> lrtest( riceCdSfaInv, riceCdTimeSfaVar )
Likelihood ratio test
Model 1: riceCdSfaInv
Model 2: riceCdTimeSfaVar
#Df

LogLik Df

6 -86.430

8 -84.529

Chisq Pr(>Chisq)

2 3.8034

0.1493

All together, these tests indicate that there is no significant technological change, while it remains
unclear whether the technical efficiencies significantly change over time.

253

6 Panel Data and Technological Change

In econometric estimations of frontier models, where one variable (e.g. time) can affect both the
frontier and the efficiency, the two effects of this variable can often be hardly separated, because
the corresponding parameters can be simultaneous adjusted with only marginally reducing the
log-likelihood value. This can be checked by taking a look at the correlation matrix of the
estimated parameters:
> round( cov2cor( vcov( riceCdTimeSfaVar ) ), 2 )
(Intercept) lArea lLabor

lNpk mYear sigmaSq gamma

(Intercept)

1.00

0.17

-0.12

0.04

0.43

0.47 -0.17

lArea

0.17

1.00

-0.68 -0.38 -0.08

0.02

0.05

lLabor

-0.12 -0.68

0.02

time

1.00 -0.28

0.10

0.09

-0.06 -0.08 -0.01

lNpk

0.02 -0.38

-0.28

1.00

0.02

0.03

0.01 -0.12

mYear

0.04 -0.08

0.10

0.02

1.00

0.70

0.69 -0.87

sigmaSq

0.43

0.02

-0.06

0.03

0.70

1.00

0.94 -0.85

gamma

0.47

0.05

-0.08

0.01

0.69

0.94

1.00 -0.84

time

-0.17

0.09

-0.01 -0.12 -0.87

-0.85 -0.84

1.00

The estimate of the parameter for technological change (mYear) is highly correlated with the
estimate of the parameter that indicates the change of the efficiencies (time).
Again, further likelihood ratio tests show that OLS models are clearly rejected in favor of the
corresponding stochastic frontier models:
> lrtest( riceCdSfaVar )
Likelihood ratio test
Model 1: OLS (no inefficiency)
Model 2: Error Components Frontier (ECF)
#Df

LogLik Df

5 -104.91

-84.55

Chisq Pr(>Chisq)

2 40.713

4.489e-10 ***

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

> lrtest( riceCdTimeSfaVar )

Likelihood ratio test
Model 1: OLS (no inefficiency)
Model 2: Error Components Frontier (ECF)
#Df

LogLik Df

Chisq Pr(>Chisq)

254

6 Panel Data and Technological Change

1

6 -104.103

-84.529

2 39.149

9.85e-10 ***

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

In case of time-variant efficiencies, the efficiencies method returns a matrix, where each row
corresponds to one of the 43 firms and each column corresponds to one of the 0 time periods:
> dim( efficiencies( riceCdSfaVar ) )
[1] 43

One can obtain a vector of efficiency estimates for each observation by setting argument asInData
equal to TRUE:
> pdat$effCdVar <- efficiencies( riceCdSfaVar, asInData = TRUE )
6.2.1.3 Observation-specific efficiencies
Finally, we estimate a Cobb-Douglas production frontier with observation-specific efficiencies.
The following commands estimate two Cobb-Douglas production frontiers, the first does not
account for technological change, while the second does:
> riceCdSfa <- sfa( lProd ~ lArea + lLabor + lNpk, data = riceProdPhil )
> summary( riceCdSfa )
Error Components Frontier (see Battese & Coelli 1992)
Inefficiency decreases the endogenous variable (as in a production function)
The dependent variable is logged
Iterative ML estimation terminated after 9 iterations:
log likelihood values and parameters of two successive iterations
are within the tolerance limit
final maximum likelihood estimates
Estimate Std. Error z value

Pr(>|z|)

(Intercept) 0.333747

0.024468 13.6400 < 2.2e-16 ***

lArea

0.355511

0.060125

5.9128 3.363e-09 ***

lLabor

0.333302

0.063026

5.2883 1.234e-07 ***

lNpk

0.271277

0.035364

7.6709 1.708e-14 ***

sigmaSq

0.238627

0.025941

9.1987 < 2.2e-16 ***

gamma

0.885382

0.033524 26.4103 < 2.2e-16 ***

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

255

6 Panel Data and Technological Change

log likelihood value: -86.20268
cross-sectional data
total number of observations = 344
mean efficiency: 0.7229764
> riceCdTimeSfa <- sfa( lProd ~ lArea + lLabor + lNpk + mYear,
+

data = riceProdPhil )

> summary( riceCdTimeSfa )

Error Components Frontier (see Battese & Coelli 1992)
Inefficiency decreases the endogenous variable (as in a production function)
The dependent variable is logged
Iterative ML estimation terminated after 10 iterations:
log likelihood values and parameters of two successive iterations
are within the tolerance limit
final maximum likelihood estimates
Estimate Std. Error z value

Pr(>|z|)

(Intercept) 0.3375352

0.0240787 14.0180 < 2.2e-16 ***

lArea

0.3557511

0.0596403

5.9649 2.447e-09 ***

lLabor

0.3507357

0.0631077

5.5577 2.733e-08 ***

lNpk

0.2565321

0.0351012

7.3083 2.704e-13 ***

mYear

0.0148902

0.0068853

2.1626

sigmaSq

0.2418364

0.0259495

9.3195 < 2.2e-16 ***

gamma

0.8979766

0.0304374 29.5024 < 2.2e-16 ***

0.03057 *

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

log likelihood value: -83.76704

cross-sectional data
total number of observations = 344
mean efficiency: 0.7201094
Please note that we used the data set riceProdPhil for these estimations, because the panel
structure should be ignored in these specifications and the data set riceProdPhil does not
include information on the panel structure.
In the Cobb-Douglas production frontier that accounts for technological change, the monotonicity conditions are globally fulfilled and the (constant) output elasticities of land, labor and

256

6 Panel Data and Technological Change

fertilizer are 0.356, 0.351, and 0.257, respectively. The estimated (constant) annual rate of technological change is around 1.5%.
A likelihood ratio test confirms the t-test for the coefficient of the time trend, i.e. the production
technology significantly changes over time:
> lrtest( riceCdTimeSfa, riceCdSfa )
Likelihood ratio test
Model 1: riceCdTimeSfa
Model 2: riceCdSfa
#Df

LogLik Df

Chisq Pr(>Chisq)

7 -83.767

6 -86.203 -1 4.8713

0.02731 *

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

6.2.2 Translog Production Frontier with Constant and Neutral Technological

Change
The specification of a Translog production function that accounts for constant and neutral (unbiased) technological change is given in (6.4).1
6.2.2.1 Observation-Specific Efficiencies
The following commands estimate a two Translog production frontiers with observation-specific
efficiencies, the first does not account for technological change, while the second can account for
constant and neutral technical change:
> riceTlSfa <- sfa( log( prod ) ~ log( area ) + log( labor ) + log( npk ) +
+

I( 0.5 * log( area )^2 ) + I( 0.5 * log( labor )^2 ) + I( 0.5 * log( npk )^2 ) +

I( log( area ) * log( labor ) ) + I( log( area ) * log( npk ) ) +

I( log( labor ) * log( npk ) ), data = riceProdPhil )

> summary( riceTlSfa )

Error Components Frontier (see Battese & Coelli 1992)
Inefficiency decreases the endogenous variable (as in a production function)
The dependent variable is logged
Iterative ML estimation terminated after 16 iterations:
1

We use not only mean-scaled input quantities but also the mean-scaled output quantity in order to obtain
the same estimates as Coelli et al. (2005, p. 250). Please note that the order of coefficients/regressors is
different in Coelli et al. (2005, p. 250): intercept, mYear, log(area), log(labor), log(npk), 0.5*log(area)^2,
log(area)*log(labor), log(area)*log(npk), 0.5*log(labor)^2, log(labor)*log(npk), 0.5*log(npk)^2.

257

6 Panel Data and Technological Change

log likelihood values and parameters of two successive iterations
are within the tolerance limit
final maximum likelihood estimates
Estimate

Std. Error z value

Pr(>|z|)

(Intercept)

3.3719e-01

2.8747e-02 11.7298 < 2.2e-16 ***

log(area)

5.3429e-01

7.9139e-02

6.7513 1.466e-11 ***

log(labor)

2.0910e-01

7.4439e-02

2.8090 0.0049699 **

log(npk)

2.2145e-01

4.5141e-02

4.9057 9.309e-07 ***

I(0.5 * log(area)^2)

-5.1502e-01

2.0692e-01 -2.4889 0.0128124 *

I(0.5 * log(labor)^2)

-5.6134e-01

2.7039e-01 -2.0761 0.0378885 *

I(0.5 * log(npk)^2)

-7.1029e-05

9.4128e-02 -0.0008 0.9993979

I(log(area) * log(labor))

6.2604e-01

1.7284e-01

3.6221 0.0002922 ***

I(log(area) * log(npk))

8.1749e-02

1.3867e-01

0.5895 0.5555218

I(log(labor) * log(npk))

-1.5750e-01

1.4027e-01 -1.1228 0.2615321

sigmaSq

2.1856e-01

2.4990e-02

8.7458 < 2.2e-16 ***

gamma

8.6930e-01

3.9456e-02 22.0319 < 2.2e-16 ***

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

log likelihood value: -76.95413

cross-sectional data
total number of observations = 344
mean efficiency: 0.7326115
> riceTlTimeSfa <- sfa( log( prod ) ~ log( area ) + log( labor ) + log( npk ) +
+

I( 0.5 * log( area )^2 ) + I( 0.5 * log( labor )^2 ) + I( 0.5 * log( npk )^2 ) +

I( log( area ) * log( labor ) ) + I( log( area ) * log( npk ) ) +

I( log( labor ) * log( npk ) ) + mYear, data = riceProdPhil )

> summary( riceTlTimeSfa )

Error Components Frontier (see Battese & Coelli 1992)
Inefficiency decreases the endogenous variable (as in a production function)
The dependent variable is logged
Iterative ML estimation terminated after 17 iterations:
log likelihood values and parameters of two successive iterations
are within the tolerance limit
final maximum likelihood estimates

258

6 Panel Data and Technological Change

Estimate Std. Error z value

Pr(>|z|)

(Intercept)

0.3423626

0.0285089 12.0090 < 2.2e-16 ***

log(area)

0.5313816

0.0786313

6.7579 1.400e-11 ***

log(labor)

0.2308950

0.0744167

3.1027 0.0019174 **

log(npk)

0.2032741

0.0448189

4.5355 5.748e-06 ***

I(0.5 * log(area)^2)

-0.4758612

0.2021533 -2.3540 0.0185745 *

I(0.5 * log(labor)^2)

-0.5644708

0.2652593 -2.1280 0.0333374 *

I(0.5 * log(npk)^2)

-0.0072200

0.0923371 -0.0782 0.9376756

I(log(area) * log(labor))

0.6088402

0.1658019

3.6721 0.0002406 ***

I(log(area) * log(npk))

0.0617400

0.1383298

0.4463 0.6553627

I(log(labor) * log(npk))

-0.1370538

0.1407360 -0.9738 0.3301377

mYear

0.0151111

0.0069164

2.1848 0.0289024 *

sigmaSq

0.2217092

0.0251305

8.8223 < 2.2e-16 ***

gamma

0.8835549

0.0367095 24.0688 < 2.2e-16 ***

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

log likelihood value: -74.40992

cross-sectional data
total number of observations = 344
mean efficiency: 0.7294192
In the Translog production frontier that accounts for constant and neutral technological change,
the monotonicity conditions are fulfilled at the sample mean and the estimated output elasticities
of land, labor and fertilizer are 0.531, 0.231, and 0.203, respectively, at the sample mean. The
estimated (constant) annual rate of technological progress is around 1.5%. A likelihood ratio test
confirms the t-test for the coefficient of the time trend, i.e. the production technology (frontier)
significantly changes over time:
> lrtest( riceTlTimeSfa, riceTlSfa )
Likelihood ratio test
Model 1: riceTlTimeSfa
Model 2: riceTlSfa
#Df

LogLik Df

Chisq Pr(>Chisq)

13 -74.410

12 -76.954 -1 5.0884

0.02409 *

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

259

6 Panel Data and Technological Change

Two further likelihood ratio tests indicate that the Translog specification is superior to the CobbDouglas specification, no matter whether the two models allow for technological change or not.
> lrtest( riceTlSfa, riceCdSfa )
Likelihood ratio test
Model 1: riceTlSfa
Model 2: riceCdSfa
#Df
1

LogLik Df

Chisq Pr(>Chisq)

12 -76.954

6 -86.203 -6 18.497

0.005103 **

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

> lrtest( riceTlTimeSfa, riceCdTimeSfa )

Likelihood ratio test
Model 1: riceTlTimeSfa
Model 2: riceCdTimeSfa
#Df
1

LogLik Df

Chisq Pr(>Chisq)

13 -74.410

7 -83.767 -6 18.714

0.004674 **

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

6.2.3 Translog Production Frontier with Non-Constant and Non-Neutral

Technological Change
The specification of a Translog production function with non-Constant and non-Neutral technological change is given in (6.7).
6.2.3.1 Observation-Specific Efficiencies
The following command estimates a Translog production frontier with observation-specific efficiencies that can account for non-constant rates of technological change as well as biased technological change:
> riceTlTimeNnSfa <- sfa( log( prod ) ~ log( area ) + log( labor ) + log( npk ) +
+

I( 0.5 * log( area )^2 ) + I( 0.5 * log( labor )^2 ) + I( 0.5 * log( npk )^2 ) +

I( log( area ) * log( labor ) ) + I( log( area ) * log( npk ) ) +

I( log( labor ) * log( npk ) ) + mYear + I( mYear * log( area ) ) +

260

6 Panel Data and Technological Change

+

I( mYear * log( labor ) ) + I( mYear * log( npk ) ) + I( 0.5 * mYear^2 ),

data = riceProdPhil )

> summary( riceTlTimeNnSfa )

Error Components Frontier (see Battese & Coelli 1992)
Inefficiency decreases the endogenous variable (as in a production function)
The dependent variable is logged
Iterative ML estimation terminated after 21 iterations:
log likelihood values and parameters of two successive iterations
are within the tolerance limit
final maximum likelihood estimates
Estimate Std. Error z value

Pr(>|z|)

(Intercept)

0.3106562

0.0314015

9.8931 < 2.2e-16 ***

log(area)

0.5126722

0.0786138

6.5214 6.965e-11 ***

log(labor)

0.2380479

0.0745020

3.1952 0.0013974 **

log(npk)

0.2151253

0.0443911

4.8461 1.259e-06 ***

I(0.5 * log(area)^2)

-0.5095014

0.2245191 -2.2693 0.0232500 *

I(0.5 * log(labor)^2)

-0.5394564

0.2627781 -2.0529 0.0400826 *

I(0.5 * log(npk)^2)

0.0212605

0.0923038

0.2303 0.8178341

I(log(area) * log(labor))

0.6132444

0.1687103

3.6349 0.0002781 ***

I(log(area) * log(npk))

0.0683940

0.1436778

0.4760 0.6340580

I(log(labor) * log(npk))

-0.1590173

0.1481054 -1.0737 0.2829677

mYear

0.0090026

0.0074634

1.2062 0.2277309

I(mYear * log(area))

0.0050527

0.0235232

0.2148 0.8299271

I(mYear * log(labor))

0.0241186

0.0254828

0.9465 0.3439117

I(mYear * log(npk))

-0.0335256

0.0176150 -1.9032 0.0570093 .

I(0.5 * mYear^2)

0.0149772

0.0069298

2.1613 0.0306744 *

sigmaSq

0.2227259

0.0243844

9.1339 < 2.2e-16 ***

gamma

0.8957679

0.0320612 27.9393 < 2.2e-16 ***

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

log likelihood value: -70.5919

cross-sectional data
total number of observations = 344
mean efficiency: 0.728398
At the mean values of the input quantities and the middle of the observation period, the monotonicity conditions are fulfilled, the estimated output elasticities of land, labor and fertilizer are

261

6 Panel Data and Technological Change

0.513, 0.238, and 0.215, respectively, and the estimated annual rate of technological progress is
around 0.9%.
The following likelihood ratio tests compare the Translog production frontier that can account
for non-constant rates of technological change as well as biased technological change with the
Translog production frontier that does not account for technological change and with the Translog
production frontier that only accounts for constant and neutral technological change:
> lrtest( riceTlTimeNnSfa, riceTlSfa )
Likelihood ratio test
Model 1: riceTlTimeNnSfa
Model 2: riceTlSfa
#Df

LogLik Df

Chisq Pr(>Chisq)

17 -70.592

12 -76.954 -5 12.725

0.0261 *

--Signif. codes:

0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

> lrtest( riceTlTimeNnSfa, riceTlTimeSfa )

Likelihood ratio test
Model 1: riceTlTimeNnSfa
Model 2: riceTlTimeSfa
#Df

LogLik Df Chisq Pr(>Chisq)

17 -70.592

13 -74.410 -4 7.636

0.1059

These tests indicate that the Translog production frontier that can account for non-constant
rates of technological change as well as biased technological change is superior to the Translog
production frontier that does not account for any technological change but it is not significantly
better than the Translog production frontier that accounts for constant and neutral technological
change. Although it seems to be unnecessary to use the Translog production frontier that can
account for non-constant rates of technological change as well as biased technological change, we
use it in our further analysis for demonstrative purposes.
The following commands create short-cuts for some of the estimated coefficients and calculate
the rates of technological change at each observation:
> at <- coef(riceTlTimeNnSfa)["mYear"]
> atArea <- coef(riceTlTimeNnSfa)["I(mYear * log(area))"]
> atLabor <- coef(riceTlTimeNnSfa)["I(mYear * log(labor))"]

262

6 Panel Data and Technological Change

> atNpk <- coef(riceTlTimeNnSfa)["I(mYear * log(npk))"]
> att <- coef(riceTlTimeNnSfa)["I(0.5 * mYear^2)"]
> riceProdPhil$tc <- with( riceProdPhil, at + atArea * log( area ) +
+

atLabor * log( labor ) + atNpk * log( npk ) + att * mYear )

The following command visualizes the variation of the individual rates of technological change:

30
0 10

Frequency

> hist( riceProdPhil$tc, 20 )

0.05

0.00

0.05

0.10

technological change

Figure 6.3: Annual rates of technological change

The resulting graph is shown in figure 6.3. Most individual rates of technological change are
between 4% and +7%, i.e. there is technological regress at some observations, while there
is strong technological progress at other observations. This wide variation of annual rates of
technological change is not unusual in applied agricultural production analysis because of the
stochastic nature of agricultural production.

6.2.4 Decomposition of Productivity Growth

In the beginning of this course, we have discussed and calculated different productivity measures, of which the total factor productivity (T F P ) is a particularly important determinant of a
firms competitiveness. During this course, we haveamongst other thingsanalyzed all three
measures that affect a firms total factor productivity, i.e.
the current state of the technology (T ) in the firms sector, which might change due to

technological change,
the firms technical efficiency (T E), which might change if the firms distance to the current
technology changes, and
the firms scale efficiency (SE), which might change if the firms size relative to the optimal
firm size changes.
Hence, changes of a firms (or a sectors) total factor productivity (T F P ) can be decomposed into technological changes (T ), technical efficiency changes (T E), and scale efficiency

263

6 Panel Data and Technological Change

changes (SE):
T F P T + T E + SE

(6.11)

This decomposition often helps to understand the reasons for improved or reduced total factor
productivity and competitiveness.

6.3 Data Envelopment Analysis (DEA)

> library( "Benchmarking" )
We create a matrix of input quantities and a vector of output quantities:
> xMat <- cbind( riceProdPhil$AREA, riceProdPhil$LABOR, riceProdPhil$NPK )
> yVec <- riceProdPhil$PROD
For now, we ignore the panel structure of the data and conduct an input-oriented DEA with
VRS:
> deaVrsIn <- dea( xMat, yVec )
> hist( eff( deaVrsIn ) )
Display the peers of the first 14 observations:
> peers( deaVrsIn )[ 1:14, ]
peer1 peer2 peer3 peer4
[1,]

189

303

333

NA

[2,]

189

303

333

NA

[3,]

189

303

333

NA

[4,]

90

227

256

333

[5,]

90

303

333

NA

[6,]

227

256

333

NA

[7,]

189

320

333

NA

[8,]

189

303

333

NA

[9,]

256

303

333

NA

[10,]

189

303

333

NA

[11,]

12

227

331

NA

[12,]

12

NA

NA

NA

[13,]

227

256

333

NA

[14,]

256

303

333

NA

Display the s of the first 14 observations:

> lambda( deaVrsIn )[ 1:14, ]

264

6 Panel Data and Technological Change

L12

L90 L116 L141 L163

L189 L206 L217

L227

[1,] 0.0000000 0.0000000

0 0.069365011

0 0.0000000

[2,] 0.0000000 0.0000000

0 0.177065580

0 0.0000000

[3,] 0.0000000 0.0000000

0 0.128631252

0 0.0000000

[4,] 0.0000000 0.2338441

0 0.000000000

0 0.2424113

[5,] 0.0000000 0.6369664

0 0.000000000

0 0.0000000

[6,] 0.0000000 0.0000000

0 0.000000000

0 0.8329082

[7,] 0.0000000 0.0000000

0 0.009313996

0 0.0000000

[8,] 0.0000000 0.0000000

0 0.029415433

0 0.0000000

[9,] 0.0000000 0.0000000

0 0.000000000

0 0.0000000

[10,] 0.0000000 0.0000000

0 0.101147709

0 0.0000000

[11,] 0.7520222 0.0000000

0 0.000000000

0 0.2380402

[12,] 1.0000000 0.0000000

0 0.000000000

0 0.0000000

[13,] 0.0000000 0.0000000

0 0.000000000

0 0.3755759

[14,] 0.0000000 0.0000000

0 0.000000000

0 0.0000000

L256

L303 L313 L319

L320

L331 L332

L333

[1,] 0.00000000 0.09397232

0 0.00000000 0.000000000

0 0.83666267

[2,] 0.00000000 0.21810667

0 0.00000000 0.000000000

0 0.60482775

[3,] 0.00000000 0.25497935

0 0.00000000 0.000000000

0 0.61638940

[4,] 0.02919074 0.00000000

0 0.00000000 0.000000000

0 0.49455386

[5,] 0.00000000 0.34624742

0 0.00000000 0.000000000

0 0.01678618

[6,] 0.10499234 0.00000000

0 0.00000000 0.000000000

0 0.06209945

[7,] 0.00000000 0.00000000

0 0.06897294 0.000000000

0 0.92171307

[8,] 0.00000000 0.01309000

0 0.00000000 0.000000000

0 0.95749456

[9,] 0.03549837 0.38499720

0 0.00000000 0.000000000

0 0.57950444

[10,] 0.00000000 0.21619497

0 0.00000000 0.000000000

0 0.68265733

[11,] 0.00000000 0.00000000

0 0.00000000 0.009937601

0 0.00000000

[12,] 0.00000000 0.00000000

0 0.00000000 0.000000000

0 0.00000000

[13,] 0.24094633 0.00000000

0 0.00000000 0.000000000

0 0.38347778

[14,] 0.39092224 0.23629835

0 0.00000000 0.000000000

0 0.37277941

The following command conducts an input-oriented DEA with CRS:

> deaCrsIn <- dea( xMat, yVec, RTS = "crs" )
> hist( eff( deaCrsIn ) )
We can calculate the scale efficiencies by:
> se <- eff( deaCrsIn ) / eff( deaVrsIn )
> hist( se )
The following command conducts an input-oriented DEA with DRS

265

6 Panel Data and Technological Change

> deaDrsIn <- dea( xMat, yVec, RTS = "drs" )
> hist( eff( deaDrsIn ) )
And we check if firms are too small or too large. This is the number of observations that
produce at the scale below the optimal scale size:
> sum( eff( deaVrsIn ) - eff( deaDrsIn ) > 1e-4 )
[1] 187
The following command conducts an output-oriented DEA with VRS:
> deaVrsOut <- dea( xMat, yVec, ORIENTATION = "out" )
> hist( efficiencies( deaVrsOut ) )
The following command conducts an output-oriented DEA with CRS:
> deaCrsOut <- dea( xMat, yVec, RTS = "crs", ORIENTATION = "out" )
> hist( eff( deaCrsOut ) )
In case of CRS, input-oriented efficiencies are equivalent to output-oriented efficiencies:
> all.equal( eff( deaCrsIn ), 1 / eff( deaCrsOut ) )
[1] TRUE
The following command obtains super efficiencies for an input-oriented DEA with CRS:
> sdeaVrsIn <- sdea( xMat, yVec )
> hist( eff( sdeaVrsIn ) )
The following command conducts a DEA with graph hyperbolic efficiencies and VRS:
> deaVrsGraph <- dea( xMat, yVec, ORIENTATION = "graph" )
> hist( eff( deaVrsGraph ) )
> plot( eff( deaVrsIn ), eff( deaVrsGraph ) )
> abline(0,1)
The following commands display the slack of the first 14 observations in an input-oriented
DEA with VRS:
> deaVrsIn <- dea( xMat, yVec, SLACK = TRUE )
> sum( deaVrsIn$slack )
[1] 279
> deaVrsIn$sx[ 1:14, ]

266

6 Panel Data and Technological Change

sx1

sx2

sx3

[1,] 0.00000000 18.7385598

0.00000

[2,] 0.00000000

0.5002562

0.00000

[3,] 0.00000000

7.4030979

0.00000

[4,] 0.00000000

0.0000000

0.00000

[5,] 0.04397651

0.0000000

0.00000

[6,] 0.00000000

2.3670434

0.00000

[7,] 0.00000000

0.0000000 39.68889

[8,] 0.00000000 13.0362360

0.00000

[9,] 0.00000000 59.0884136

0.00000

[10,] 0.00000000 17.3290527

0.00000

[11,] 0.00000000

0.0000000

0.00000

[12,] 0.00000000

0.0000000

0.00000

[13,] 0.00000000

5.4720202

0.00000

[14,] 0.00000000 11.6460484

0.00000

> deaVrsIn$sy[ 1:14, ]

[1] 0.0000000 0.0000000 0.0000000 0.0000000 0.0000000 0.0000000 0.0000000
[8] 0.0000000 0.0000000 0.0000000 0.1412618 0.0000000 0.0000000 0.0000000
The following commands calculate and decompose productivity changes:
> xMat0 <- xMat[ riceProdPhil$YEARDUM == 1, ]
> xMat1 <- xMat[ riceProdPhil$YEARDUM == 2, ]
> yVec0 <- yVec[ riceProdPhil$YEARDUM == 1 ]
> yVec1 <- yVec[ riceProdPhil$YEARDUM == 2 ]
> c00 <- eff( dea( xMat0, yVec0, RTS = "crs" ) )
> c01 <- eff( dea( xMat0, yVec0, XREF = xMat1, YREF = yVec1, RTS = "crs" ) )
> c11 <- eff( dea( xMat1, yVec1, RTS = "crs" ) )
> c10 <- eff( dea( xMat1, yVec1, XREF = xMat0, YREF = yVec0, RTS = "crs" ) )
Productivity changes (Malmquist):
> dProd0 <- c10 / c00
> hist( dProd0 )
> dProd1 <- c11 / c01
> plot( dProd0, dProd1 )
> dProd <- sqrt( dProd0 * dProd1 )
> hist( dProd )
Technological changes:

267

6 Panel Data and Technological Change

> dTech0 <- c00 / c01
> dTech1 <- c10 / c11
> plot( dTech0, dTech1 )
> dTech <- sqrt( dTech0 * dTech1 )
> hist( dTech )
Efficiency changes:
> dEff <- c11 / c00
> hist( dEff )
Checking Malmquist decomposition:
> all.equal( dProd, dTech * dEff )
[1] TRUE

268

Bibliography
Aigner, D., C.A.K. Lovell, and P. Schmidt. 1977. Formulation and Estimation of Stochastic
Frontier Production Function Models. Journal of Econometrics 6:2137.
Battese, G.E., and T.J. Coelli. 1995. A Model for Technical Inefficiency Effects in a Stochastic
Frontier Production Function for Panel Data. Empirical Economics 20:325332.
Bogetoft, P., and L. Otto. 2011. Benchmarking with DEA, SFA, and R, vol. 157 of International
Series in Operations Research & Management Science. Springer.
Chambers, R.G. 1988. Applied Production Analysis. A Dual Approach. Cambridge University
Press, Cambridge.
Chand, R., and J.L. Kaul. 1986. A Note on the Use of the Cobb-Douglas Profit Function.
American Journal of Agricultural Economics 68:162164.
Chiang, A.C. 1984. Fundamental Methods of Mathematical Economics, 3rd ed. McGraw-Hill.
Coelli, T.J. 1995. Estimators and Hypothesis Tests for a Stochastic: A Monte Carlo Analysis.
Journal of Productivity Analysis 6:247268.
Coelli, T.J., D.S.P. Rao, C.J. ODonnell, and G.E. Battese. 2005. An Introduction to Efficiency
and Productivity Analysis, 2nd ed. New York: Springer.
Croissant, Y., and G. Millo. 2008. Panel Data Econometrics in R: The plm Package. Journal
of Statistical Software 27:143.
Czekaj, T., and A. Henningsen. 2012. Comparing Parametric and Nonparametric Regression
Methods for Panel Data: the Optimal Size of Polish Crop Farms. FOI Working Paper No.
2012/12, Institute of Food and Resource Economics, University of Copenhagen.
Hayfield, T., and J.S. Racine. 2008. Nonparametric Econometrics: The np Package. Journal of
Statistical Software 27:132.
Henning, C.H.C.A., and A. Henningsen. 2007. Modeling Farm Households Price Responses in
the Presence of Transaction Costs and Heterogeneity in Labor Markets. American Journal of
Agricultural Economics 89:665681.

269

Bibliography
Hurvich, C.M., J.S. Simonoff, and C.L. Tsai. 1998. Smooting Parameter Selection in Nonparametric Regression Using an Improved Akaike Information Criterion. Journal of the Royal
Statistical Society Series B 60:271293.
Ivaldi, M., N. Ladoux, H. Ossard, and M. Simioni. 1996. Comparing Fourier and Translog
Specifications of Multiproduct Technology: Evidence from an Incomplete Panel of French
Farmers. Journal of Applied Econometrics 11:649667.
Kleiber, C., and A. Zeileis. 2008. Applied Econometrics with R. New York: Springer.
Li, Q., and J.S. Racine. 2007. Nonparametric Econometrics: Theory and Practice. Princeton:
Princeton University Press.
McClelland, J.W., M.E. Wetzstein, and W.N. Musserwetz. 1986. Returns to Scale and Size in
Agricultural Economics. Western Journal of Agricultural Economics 11:129133.
Meeusen, W., and J. van den Broeck. 1977. Efficiency Estimation from Cobb-Douglas Production
Functions with Composed Error. International Economic Review 18:435444.
Olsen, J.V., and A. Henningsen. 2011. Investment Utilization and Farm Efficiency in Danish
Agriculture. FOI Working Paper No. 2011/13, Institute of Food and Resource Economics,
University of Copenhagen.
Racine, J.S. 2008. Nonparametric Econometrics: A Primer. Foundations and Trends in Econometrics 3:188.
Teetor, P. 2011. R Cookbook . OReilly Media.
Zuur, A., E.N. Ieno, and E. Meesters. 2009. A Beginners Guide to R. Use R!, Springer.

270