ECN 421 – Labor Economics

Practice Statistics and Methods

Andreas Kostol
([email protected])

Spring 2021

1/34
Summary

The identification problem:

we only observe one of two (or more) potential outcomes
−→ hard to estimate the counterfactual

Data contamination:
Observational data and OLS provides results that are (biased
and) hard to interpret

Potential solutions:
We can run real experiments when feasible or look for natural
experiments in the data

2/34
Causation 6= Correlation

Since so much of modern labor economics is about causation

and because business and policy is about what works,

3/34
Causation 6= Correlation

Since so much of modern labor economics is about causation

and because business and policy is about what works,
Correlation is not the same as causation.

3/34
Causation 6= Correlation

Since so much of modern labor economics is about causation

and because business and policy is about what works,
Correlation is not the same as causation.
Random assignment solves the identification problem.

3/34
Causation 6= Correlation

Since so much of modern labor economics is about causation

and because business and policy is about what works,
Correlation is not the same as causation.
Random assignment solves the identification problem.
Need to understand when and why we apply the tools OLS,
DID, IV, RDD.

3/34
MULTIPLE CHOICE

4/34
Question

Econometrics is (one true)

1. another word for rocket science
2. a combination of economic theory and statistical methods
3. is the quantitative theory of labor supply

5/34
Question

Econometrics is (one true)

1. another word for rocket science
2. a combination of economic theory and statistical methods
3. is the quantitative theory of labor supply

Answer: a combination of economic theory and statistical

methods

5/34
Question

Following is a histogram of ages of 10 candidates applying for

a particular high-school teaching position.

I. The median age is between 24 and 25.

II. The mean age is between 22 and 23.
III. The mean age is greater than the median age.

6/34
Question

Following is a histogram of ages of 10 candidates applying for

a particular high-school teaching position.

I. The median age is between 24 and 25.

II. The mean age is between 22 and 23.
III. The mean age is greater than the median age.

Answer: The mean age is greater than the median age.

6/34
Question

Which of the following options are examples of natural

experiments?
I. Reform that changes mandatory years of schooling.
II. Sudden change in minimum wage at the state level.
III. Union status of workers.
IV. Changes in the duration of paid maternity leave based on
the date of birth of the child.

7/34
Question

Which of the following options are examples of natural

experiments?
I. Reform that changes mandatory years of schooling.
II. Sudden change in minimum wage at the state level.
III. Union status of workers.
IV. Changes in the duration of paid maternity leave based on
the date of birth of the child.

Answer: I, II and IV

7/34
Question

What is the main problem with identification of causes and

effects in economics?
1. We only observe samples, and not the population
2. We only observe one potential outcome per individual
3. Our computers are not strong enough to handle the
amount of data that we need

8/34
Question

What is the main problem with identification of causes and

effects in economics?
1. We only observe samples, and not the population
2. We only observe one potential outcome per individual
3. Our computers are not strong enough to handle the
amount of data that we need

Answer: We only observe one potential outcome per individual

8/34
Question
Consider a sample of AZ workers and their hourly wages
{ 10; 10; 12; 13; 15; 15; 25; 29; 30; 31} at time period 1, and
{10; 10; 10; 12; 15; 16; 18; 32; 35; 37} at time period 2.
a. Calculate the mean and 50th percentile in period 1
b. Calculate the mean and 50th percentile in period 2
c. Calculate the variance in both periods.
d. Did inequality in wages increase? Why?

9/34
Question
Consider a sample of AZ workers and their hourly wages
{ 10; 10; 12; 13; 15; 15; 25; 29; 30; 31} at time period 1, and
{10; 10; 10; 12; 15; 16; 18; 32; 35; 37} at time period 2.
a. Calculate the mean and 50th percentile in period 1
b. Calculate the mean and 50th percentile in period 2
c. Calculate the variance in both periods.
d. Did inequality in wages increase? Why?

Answer: w̄1 = 19, 50pct,1 = 15; w̄2 = 19.5 , 50pct,2 = 15.5;

1
(wi,t − w̄t )2 .
P
var (wt ) = 10−1

Yes, wage inequality went up. The variance clearly went up, so
did the dispersion in terms of 90th to 10th percentile
differences. 9/34
Question

What is the main challenge identifying the causal effect of a

policy?
1. Recovering the counterfactual.
2. Estimating the slope of a line
3. Estimating the correlation between two variables

10/34
Question

What is the main challenge identifying the causal effect of a

policy?
1. Recovering the counterfactual.
2. Estimating the slope of a line
3. Estimating the correlation between two variables

Answer: Recovering the counterfactual.

10/34
Question

The graph below shows cumulative proportions plotted against

land values (in dollars per acre) for farms on sale in a rural
community.

What is the median land value? 1000, 2000, 2750, or 3300

11/34
Question

The graph below shows cumulative proportions plotted against

land values (in dollars per acre) for farms on sale in a rural
community.

What is the median land value? 1000, 2000, 2750, or 3300

Answer: 2000

11/34
Question

What assumptions do we need to hold for OLS to give an

unbiased estimate of a training program on wages?
1. Residuals are normally distributed.
2. Random assignment
3. The median wage in the population is equal to the mean.

12/34
Question

What assumptions do we need to hold for OLS to give an

unbiased estimate of a training program on wages?
1. Residuals are normally distributed.
2. Random assignment
3. The median wage in the population is equal to the mean.

Answer: 2

12/34
Question

Which of the following are true statements?

I. If there is sufficient evidence to reject a null hypothesis at
the 10% level, then there is sufficient evidence to reject it at
the 5% level.
II. When evaluating whether we can reject a null hypothesis we
tend to us the critical value 42.
III. If a hypothesis test is conducted at the 1% level, there is a
1% chance of rejecting a true null hypothesis.

13/34
Question

Which of the following are true statements?

I. If there is sufficient evidence to reject a null hypothesis at
the 10% level, then there is sufficient evidence to reject it at
the 5% level.
II. When evaluating whether we can reject a null hypothesis we
tend to us the critical value 42.
III. If a hypothesis test is conducted at the 1% level, there is a
1% chance of rejecting a true null hypothesis.

Answer: If a hypothesis test is conducted at the 1%

level, there is a 1% chance of rejecting a true null
hypothesis.

13/34
Question

Which of the following statements are correct?

1. The correlation between Y and X equals the covariance
divided by the product of the standard deviation of Y and
X
2. The correlation measures the causal effect of an
explanatory variable X on the dependent variable Y .

14/34
Question

Which of the following statements are correct?

1. The correlation between Y and X equals the covariance
divided by the product of the standard deviation of Y and
X
2. The correlation measures the causal effect of an
explanatory variable X on the dependent variable Y .

Answer: 1.

14/34
Question

In the presence of selection into treatment which of the

following sentences is true?
1. The OLS estimator is unbiased.
2. The OLS estimator equals the average effect among the
treated and a selection bias.
3. The slope of a line equals the optimal labor supply

15/34
Question

In the presence of selection into treatment which of the

following sentences is true?
1. The OLS estimator is unbiased.
2. The OLS estimator equals the average effect among the
treated and a selection bias.
3. The slope of a line equals the optimal labor supply

Answer: The OLS estimator equals theaverage effect among

the treated and a selection bias.

15/34
Question

Random assignment solves the potential outcome problem by

design
1. True
2. False

16/34
Question

Random assignment solves the potential outcome problem by

design
1. True
2. False

Answer: True

16/34
Question

Which of following conditions are required for a

Differences-in-Differences estimator being implemented:
I - Existence of a control group and a treatment group
II - Data available at least in two periods of both groups
III - Differential time trends between groups before treatment

17/34
Question

Which of following conditions are required for a

Differences-in-Differences estimator being implemented:
I - Existence of a control group and a treatment group
II - Data available at least in two periods of both groups
III - Differential time trends between groups before treatment

Answer: I and II.

17/34
Question

Which of following conditions are required for a

Differences-in-Differences estimator being implemented:
I - Existence of a control group and a treatment group
II - Data available at least in two periods of both groups
III - Differential time trends between groups before treatment

Answer: I and II.

We need parallel trends before the treatment date. If they are
differential, that indicates the identifying assumption is
violated.

17/34
Question: OLS, IV, DiD, or RDD

A researcher wants to compute the impact of new on drug the

heart disease occurrence on a randomly selected sample of
patients. The patients are randomly assigned to control and
treatment groups. The effect of the drug can be correlated
with non-observable individual characteristics. The researcher
has information about heart disease 3 months after the
treatment.

18/34
Question: OLS, IV, DiD, or RDD

A researcher wants to compute the impact of new on drug the

heart disease occurrence on a randomly selected sample of
patients. The patients are randomly assigned to control and
treatment groups. The effect of the drug can be correlated
with non-observable individual characteristics. The researcher
has information about heart disease 3 months after the
treatment.

Answer: OLS. Why?

18/34
Question: OLS, IV, DiD, or RDD

A researcher wants to compute the impact of new on drug the

heart disease occurrence on a randomly selected sample of
patients. The patients are randomly assigned to control and
treatment groups. The effect of the drug can be correlated
with non-observable individual characteristics. The researcher
has information about heart disease 3 months after the
treatment.

Answer: OLS. Why?

Random assignment!

18/34
Question: OLS, IV, DiD, or RDD

A researcher wants to compute the impact of an additional

year of school on wages. It is known that the years of
schooling decision are correlated with non-observable
individual characteristics those are correlated with wages.
However, the researcher has information on a schooling reform
that partially affected the years of schooling in the population.

19/34
Question: OLS, IV, DiD, or RDD

A researcher wants to compute the impact of an additional

year of school on wages. It is known that the years of
schooling decision are correlated with non-observable
individual characteristics those are correlated with wages.
However, the researcher has information on a schooling reform
that partially affected the years of schooling in the population.

Answer: IV

19/34
Question: OLS, IV, DiD, or RDD

A researcher wants to compute the impact of an additional

year of school on wages. It is known that the years of
schooling decision are correlated with non-observable
individual characteristics those are correlated with wages.
However, the researcher has information on a schooling reform
that partially affected the years of schooling in the population.

Answer: IV
Quasi-randomization: Potential outcomes are independent of
reform. Reform affects years of schooling, but not wages other
than via schooling.

19/34
Question: OLS, IV, DiD, or RDD

A researcher wants to understand the impact of payroll taxes

on individual wages. Payroll taxes increases with the level of
earnings, but firms have to pay higher taxes for everyone born
before 15th of May 1964. The data available to the researcher
include date of birth, payroll tax, hours worked and wages.

20/34
Question: OLS, IV, DiD, or RDD

A researcher wants to understand the impact of payroll taxes

on individual wages. Payroll taxes increases with the level of
earnings, but firms have to pay higher taxes for everyone born
before 15th of May 1964. The data available to the researcher
include date of birth, payroll tax, hours worked and wages.

Answer: RDD.

20/34
Question: OLS, IV, DiD, or RDD

A researcher wants to understand the impact of payroll taxes

on individual wages. Payroll taxes increases with the level of
earnings, but firms have to pay higher taxes for everyone born
before 15th of May 1964. The data available to the researcher
include date of birth, payroll tax, hours worked and wages.

Answer: RDD.
There exists an assignment variable, cutoff is “random”, and
the cutoff determines treatment

20/34
Question

What are we minimizing when an OLS regression is

performed?
1. We are maximizing the utility of the data
2. We are minimizing the sum of the squared residuals.
3. We are calculating ordinary labor supply models

21/34
Question

What are we minimizing when an OLS regression is

performed?
1. We are maximizing the utility of the data
2. We are minimizing the sum of the squared residuals.
3. We are calculating ordinary labor supply models

Answer: We are minimizing the sum of squared residuals. In

other words, we are minimizing the distance between the
model and the data.

21/34
Question

Which of the following is an example of selection bias in

attempting to assess the wage-returns to schooling
1. High ability people select into schooling, and high ability
persons tend to have higher wages
2. Schooling is randomly assigned in the population

22/34
Question

Which of the following is an example of selection bias in

attempting to assess the wage-returns to schooling
1. High ability people select into schooling, and high ability
persons tend to have higher wages
2. Schooling is randomly assigned in the population

Answer: High ability people select into schooling, and high

ability persons tend to have higher earnings potential.

22/34
PROBLEM 1

23/34
Problem 1

a) (10 points) Do you think that the coefficients in the regression are likely to suffer
from selection bias? Explain.

24/34
Problem 1

a) (10 points) Do you think that the coefficients in the regression are likely to suffer
from selection bias? Explain.
A: Nope. This sounds good. It satisfies the randomization assumption required to
estimate ordinary least squares models.

24/34
Problem 1

a) (10 points) Do you think that the coefficients in the regression are likely to suffer
from selection bias? Explain.
A: Nope. This sounds good. It satisfies the randomization assumption required to
estimate ordinary least squares models.
b) (10 points) Does the labor supply curve ever bend backwards (i.e., start falling for
higher wages)?

24/34
Problem 1

a) (10 points) Do you think that the coefficients in the regression are likely to suffer
from selection bias? Explain.
A: Nope. This sounds good. It satisfies the randomization assumption required to
estimate ordinary least squares models.
b) (10 points) Does the labor supply curve ever bend backwards (i.e., start falling for
higher wages)?
A: Differentiate with respect to wage, set equal to zero: 0.5 = 2 × .05W . Hence,
labor supply starts falling when wage is above $5.

24/34
Problem 1
c) (10 points) The backward bending part of this model is given by parameter
estimate on c. When thinking about inference on this parameter, how would you
describe the null hypothesis?

25/34
Problem 1
c) (10 points) The backward bending part of this model is given by parameter
estimate on c. When thinking about inference on this parameter, how would you
describe the null hypothesis?
A: We want to test whether c=0. The alternative hypothesis is that it is not equal
to 0.

25/34
Problem 1
c) (10 points) The backward bending part of this model is given by parameter
estimate on c. When thinking about inference on this parameter, how would you
describe the null hypothesis?
A: We want to test whether c=0. The alternative hypothesis is that it is not equal
to 0.
d) (10 points) Does the inference on the parameter estimates of c alter your
conclusion from question (b)?

25/34
Problem 1
c) (10 points) The backward bending part of this model is given by parameter
estimate on c. When thinking about inference on this parameter, how would you
describe the null hypothesis?
A: We want to test whether c=0. The alternative hypothesis is that it is not equal
to 0.
d) (10 points) Does the inference on the parameter estimates of c alter your
conclusion from question (b)?
A: Yes. Look at the ratio of the estimate -0.05 over standard error. We observe
that the (absolute value of the) ratio is much smaller than the critical value of 2.
Hence, we definitely do not reject the null, so we do not believe that the labor
supply is backward bending anymore. Or put more precisely, it could be, but we do
not have the evidence yet to support the idea.

25/34
Problem 1
c) (10 points) The backward bending part of this model is given by parameter
estimate on c. When thinking about inference on this parameter, how would you
describe the null hypothesis?
A: We want to test whether c=0. The alternative hypothesis is that it is not equal
to 0.
d) (10 points) Does the inference on the parameter estimates of c alter your
conclusion from question (b)?
A: Yes. Look at the ratio of the estimate -0.05 over standard error. We observe
that the (absolute value of the) ratio is much smaller than the critical value of 2.
Hence, we definitely do not reject the null, so we do not believe that the labor
supply is backward bending anymore. Or put more precisely, it could be, but we do
not have the evidence yet to support the idea.
e) (10 points) Do you think that the coefficients in the regression are likely to suffer
from selection bias? [YES or NO]

25/34
Problem 1
c) (10 points) The backward bending part of this model is given by parameter
estimate on c. When thinking about inference on this parameter, how would you
describe the null hypothesis?
A: We want to test whether c=0. The alternative hypothesis is that it is not equal
to 0.
d) (10 points) Does the inference on the parameter estimates of c alter your
conclusion from question (b)?
A: Yes. Look at the ratio of the estimate -0.05 over standard error. We observe
that the (absolute value of the) ratio is much smaller than the critical value of 2.
Hence, we definitely do not reject the null, so we do not believe that the labor
supply is backward bending anymore. Or put more precisely, it could be, but we do
not have the evidence yet to support the idea.
e) (10 points) Do you think that the coefficients in the regression are likely to suffer
from selection bias? [YES or NO]
A: Yes, the wage is correlated with preferences for leisure. So when wages are high,
drivers do no want to drive. So there is a unobserved determinant of hours that is
correlated with positive changes in the wage. This means that there should be a
negative bias in the estimates, where taxi drivers’ preferences for soccer look like a
backward bending labor supply curve.

25/34
PROBLEM 2

26/34
Problem 2

a) (20 points) Calculate the difference-in-difference estimate of the effect of the

mandated maternity benefits. Show how do you obtain this estimate. Does the simple
comparison understate or overstate the effect? By how much?

27/34
Problem 2

a) (20 points) Calculate the difference-in-difference estimate of the effect of the

mandated maternity benefits. Show how do you obtain this estimate. Does the simple
comparison understate or overstate the effect? By how much?
A: -0.30. Understate the effect by 0.14

27/34
Problem 2

a) (20 points) Calculate the difference-in-difference estimate of the effect of the

mandated maternity benefits. Show how do you obtain this estimate. Does the simple
comparison understate or overstate the effect? By how much?
A: -0.30. Understate the effect by 0.14
b) (20 points) What are the two main assumptions that the difference-in-difference
estimator needs to satisfy in order to provide causal estimates of the treatment effects?

27/34
Problem 2

a) (20 points) Calculate the difference-in-difference estimate of the effect of the

mandated maternity benefits. Show how do you obtain this estimate. Does the simple
comparison understate or overstate the effect? By how much?
A: -0.30. Understate the effect by 0.14
b) (20 points) What are the two main assumptions that the difference-in-difference
estimator needs to satisfy in order to provide causal estimates of the treatment effects?
A: Diff-in Diff uses a control group to remove the effects of other factors from your
esimates. Assumptions: Common shocks and parallel trends

27/34
Problem 2

a) (20 points) Calculate the difference-in-difference estimate of the effect of the

mandated maternity benefits. Show how do you obtain this estimate. Does the simple
comparison understate or overstate the effect? By how much?
A: -0.30. Understate the effect by 0.14
b) (20 points) What are the two main assumptions that the difference-in-difference
estimator needs to satisfy in order to provide causal estimates of the treatment effects?
A: Diff-in Diff uses a control group to remove the effects of other factors from your
esimates. Assumptions: Common shocks and parallel trends
c) (10 points) Imagine that now you may want to obtain your difference-in-difference
estimator by performing a regression. What is the econometric specification that you
would run? Explain.

27/34
Problem 2

a) (20 points) Calculate the difference-in-difference estimate of the effect of the

mandated maternity benefits. Show how do you obtain this estimate. Does the simple
comparison understate or overstate the effect? By how much?
A: -0.30. Understate the effect by 0.14
b) (20 points) What are the two main assumptions that the difference-in-difference
estimator needs to satisfy in order to provide causal estimates of the treatment effects?
A: Diff-in Diff uses a control group to remove the effects of other factors from your
esimates. Assumptions: Common shocks and parallel trends
c) (10 points) Imagine that now you may want to obtain your difference-in-difference
estimator by performing a regression. What is the econometric specification that you
would run? Explain.
A: estimating equation is
wage = b0 + b1 ∗ treat · state + b2 ∗ post + b3 ∗ (treat · state ∗ post) +  where b3 is
the diff-in-diff estimate.

27/34
PROBLEM 3

28/34
Problem 3

Define the treatment effect of a training program in terms of

potential outcomes.

29/34
Problem 3

Define the treatment effect of a training program in terms of

potential outcomes.

Answer: βi = Yi1 − Yi0 , or β = ∆ = E [Y 1 − Y 0 ],

or β = E [Y 1 | d = 1] − E [Y 0 | d = 0],
or with words

29/34
Problem 3

Describe the average treatment effect and how that may differ
from the treatment effect among the treated.

30/34
Problem 3

Describe the average treatment effect and how that may differ
from the treatment effect among the treated.

Answer: β ATE = E [Y 1 ] − E [Y 0 ], β ATT = E [Y 1 − Y 0 | d = 1]

They differ if effects are heterogeneous in the population. It is
reasonable to believe that effects are systematically different
among those who chose to be treated than those who do not.
ATE captures the average, including the untreated.

30/34
Extra Questions

31/34
R-Squared

When do we know that the R-squared is good enough?

32/34
R-Squared

When do we know that the R-squared is good enough?

In labor economics, we normally don’t care about how much of

the variance our model explains. We care about a precise and
robust answer to a well-defined question.

In macro economics, people sometimes think a high R-squared

means that the model is good. However, a model is judged by
it’s ability to predict unseen events. High predictive power for
the past is not necessarily a good measure of its predictive
power of the future.

32/34
R-Squared

In Statistics or Supervised Machine Learning, the goal is often

to find a model that can best predict unseen events. That is
how you make money, and where large amounts of data
become valuable.
With this goal, there is a bias–variance trade-off. This is
normally addressed by training a model (in-sample-bias), and
then testing it on a different data set (out-of-sample-variance).

33/34
R-Squared

The bias–variance dilemma or bias–variance problem is the

conflict in trying to simultaneously minimize these two sources
of error that prevent supervised learning algorithms from
generalizing beyond their training set.
High bias can cause an algorithm to miss the relevant relations
between features and target outputs (underfitting).
The variance is an error from sensitivity to small fluctuations
in the training set.
High variance can cause an algorithm to model the random
noise in the training data, rather than the intended outputs
(overfitting).

34/34