Statistics and Data Analysis

Paper 279-25

Qualitative and Limited Dependent Variable Models Using the

New QLIM Procedure
Minbo Kim, SAS Institute Inc., Cary, NC

Abstract

Introductory Example

There are many applications of limited dependent

variable and discrete choice models in a wide variety of areas, including economics, finance, marketing,
political science, and sociology, to name a few. This
paper introduces multinomial discrete choice and simultaneous equations models and presents how to
estimate these models using PROC QLIM. Three data
sets are used for illustration.

For multinomial data analysis, it is important to organize data series in an appropriate way. Here is a simple binary data set that illustrates how you can estimate the multinomial logit model using PROC QLIM.
The first five observations of a simulated data set
(Ben-Akiva and Lerman 1985, p. 88) are shown as
follows:
Example of Binary Choice Data

Introduction
It is not uncommon to encounter econometric models where dependent variables are limited or qualitative. These discrete choice and limited dependent
variable models need to be analyzed using more complicated methods than usual continuous models. The
SAS/ETS QLIM procedure is developed to analyze
mainly cross-sectional data, though you can use the
QLIM procedure for panel or time-series data. PROC
QLIM can analyze the following models:













binary probit and logit

id

auto

1
2
3
4
5

52.9
4.1
4.1
56.2
51.8

transit
4.4
28.5
86.9
31.6
20.2

ttdif

cchoice

alt

48.5
-24.4
-82.8
24.6
31.6

Transit
Transit
Auto
Transit
Transit

0
0
1
0
0

The two travel alternatives in the data set are auto

and transit. The attribute of travel time is recorded as
AUTO for automobile travel and TRANSIT for transit
travel. The binary choice data can be estimated using
the conditional logit model. For conditional logit estimation, we created the new data set and the first ten
observations are shown in the following table:

ordinal probit
Rearranged Binary Data

binary probit and logit with heteroskedasticity

multinomial logit
nested logit
mixed multinomial logit
univariate and bivariate poisson regression

id
1
1
2
2
3
3
4
4
5
5

autodum
1
0
1
0
1
0
1
0
1
0

ttime

cchoice

52.9
4.4
4.1
28.5
4.1
86.9
56.2
31.6
51.8
20.2

Transit
Transit
Transit
Transit
Auto
Auto
Transit
Transit
Transit
Transit

mode
1
2
1
2
1
2
1
2
1
2

choice
0
1
0
1
1
0
0
1
0
1

negative binomial regression

tobit (censored and truncated regression)
simultaneous equations model
endogenous switching regression

The new variable (TTIME) takes the value of automobile travel time (AUTO) for the first record of each individual while it contains transit travel time (TRANSIT)
for the second record. We use an alternative-specific
constant (AUTODUM) for conditional logit estimation.

Statistics and Data Analysis

The probability of choosing the auto mode (MODE=1)

is

P (yi = 1) =

exp(1 AUTODUM + 2 TTIME)

1 + exp(1 AUTODUM + 2 TTIME)

PROC QLIM estimates the conditional logit model

with the TYPE=CLOGIT option. Maximum likelihood estimation uses the Newton-Raphson optimization technique (OPTMETHOD=NR), and the standard errors are calculated from the Hessian metrix
(COVEST=HESS).
proc qlim data=travel2;
model mode = autodum ttime / noint type=clogit id=id
choice=choice covest=hess optmethod=nr;
endogenous discrete=(mode 1 2);
run;

Binary Choice Modeling Using Multinomial Logit

Variable
Parameter Standard
t statistic
Name
Estimate
Error
AUTODUM
-0.2376
0.7505
-0.32
TTIME
-0.0531
0.0206
-2.57
Summary Table
Log L
# Observations
# Records
AIC

-6.166
21
42
16.3321

As explained in the next section, the binary choice

logit can estimate the binomial choice model using the
following specification:

log 1

P (yi = 1)
P (yi = 1)

= (xiA

(probit)

where (
) is the distribution function of standard normal variables and ^1 and ^2 are maximum likelihood
estimates.
proc qlim data=travel1;
model alt = ttdif /
type=blogit covest=hess optmethod=nr;
endogenous discrete=(alt 0 1);
output out=blg p=p_lo;
model alt = ttdif /
type=bprobit covest=hess optmethod=nr;
endogenous discrete=(alt 0 1);
output out=bpr p=p_pr;
run;

Binary Choice Modeling Using Logit & Probit

Variable
Parameter Standard
t statistic
Name
Estimate
Error
Logit Estimates
INTERCEPT
-0.2376
0.7505
-0.32
TTDIF
-0.0531
0.0206
-2.57
Summary Table
Log L
# Observations
# Records
AIC
INTERCEPT
TTDIF

-6.166
21
21
16.3321
Probit Estimates
-0.0644
0.3992
-0.0300
0.0103

Summary Table
Log L
# Observations
# Records
AIC

-6.1652
21
21
16.3303

-0.16
-2.92

xiT ) = xiD 

where xiA = (1, AUTO), xiT = (0, TRANSIT), and

xiD = (1, TTDIF). The explanatory variables, xiA
and xiT , are stored in the separate line with the variable TTIME when multinomial modeling is used. For
binomial logit estimation, one observation per person is needed since the probability of transit choice,
P (yi = 0), can be calculated from the probability of
automobile choice, P (yi = 1).
The binary logit and probit models are estimated using the maximum likelihood method. Parameter estimates are summarized in the following table. The
predicted probabilities are generated by the OUTPUT
statement using the formula

P (yi = 1) =

P (yi = 1) = (^1 + ^2 TTDIF)

exp(^1 + ^2 TTDIF) (logit)

1 + exp(^1 + ^2 TTDIF)

The binomial logit estimates are exactly the same

as the preceding multinomial logit estimates, as expected. The probit estimates are different, since the
error variance of probit is normalized as 1. The prediction probabilities are displayed. The plot indicates
that auto mode would be chosen if the travel by automobile takes less time than the travel by transit
(TTDIF < 0).

Statistics and Data Analysis

be used to calculate a probability of choice j among

1.0

J + 1 alternatives.

0.8

P (yi = j ) =

0.6

The conditional logit model can also be derived from

utility maximization. Let the random utility function be
defined as

0.4
0.2

Uij

0.0
-100

exp(x0 )
PJ ij 0
1 + k=1 exp(xik )

-50

50

100

ttdif
p_lo
p_pr

Discrete Choice Modeling

Binary choice modeling is used to analyze binary response variable models. The regression model with
binary responses can be written as

yi = x0i  + i
where only the sign of the dependent variable is observed as follows:

= 1 if yi > 0
= 0 otherwise
i  standardnormalwithCDF :
Zx 1
p exp( t2 =2)dt
(x) =
1 2
or
logisticwithCDF :
x)
(x) = 1 +exp(
exp(x) (logit)

= Vij + ij

where Vij is a non-stochastic utility function. If we assume linear utility function, then Vij = x0ij  . The error
disturbances are assumed to have iid Gumbel (log
Weibull or type I extreme value) distribution with distribution function, exp( exp( ij )). Then the event
fyi = j g can be expressed using a random utility function as follows:

Uij

 maxk2Ci ;k6=j Uik

Using properties of the Gumbel distribution, the probability of choosing an alternative j from ni choices of
individual i can be derived from utility maximization:

Pi (j )

yi

(probit)

For the binary logit model, the probability of one

choice is denoted

P (yi = 1) =

exp(x0i )
1 + exp(x0i )

The probability of an alternative choice can be derived

easily as P (yi = 0) = 1 P (yi = 1).
When the dependent variable takes multiple discrete
values, the multinomial logit model can be used to
analyze unordered categorical response variables.
Standard multinomial (or conditional) logit models can

= P [x0ij  + ij  maxk2Ci (x0ik  + ik )]

exp(x0ij )
= P exp(
x0ik  )
k2Ci

When explanatory variables contain only individual

characteristics, the multinomial logit model is sometimes defined by

exp(x0  )
PJ i j 0 ;
1 + k=1 exp(xi k )
1
P (yi = 1) =
PJ
1 + k=1 exp(x0i k )
P (yi = j )

j2

This type of multinomial choice modeling has a couple

of weaknesses: it has too many parameters and it is
difficult to interpret.
The odds of success or failure is an important concept
in binary and multinomial modeling. The log-odds ratio is defined

log

P (yi = j )
P (yi = k )

= (xij

xik )0 

This expression also shows that the multinomial logit

has the property of the independence of irrelevant al-

Statistics and Data Analysis

ternatives (IIA). In other words, the log-odds ratio is

only affected by choice j and k .

Multidimensional Discrete Choice Modeling

Suppose a decision maker has a multidimensional
choice Ci , where Ci = C1  C2 


CL Ci and
Ci is defined as the set of infeasible elements in
C1  C2 


CL for the decision maker i. When
elements in Ci share unobserved and observed attributes, utilities of these elements are not independent. In this case, nested logit modeling is appropriate for the multinomial choice data analysis. Multinomial probit modeling can be another approach. It
is convenient to introduce several notations to explain
nested logit modeling. Assume that there are L levels,
with 1 representing the lowest branch of the tree and
L denoting the stem of the tree. The index of a node
at level h is represented by (jh ;


; jL ). Let h denote a node at level h + 1, where h = (jh+1 ;


; jL ).
The choice set Ch contains choices that belong to
branches below the node h . The notation Ch can
also be used to express a set of indices below h .
Note that C0 is a set with a single element, while CL
represents a choice set containing all possible alternatives. The probability of choice at level h is written

Pi (jh jh ) =

where xhi;h 1 is the vector of variables for observation

i related with the node h 1 and

=k;jh h = Ik;jh h k;jh h

The inclusive value of the level h + 1 is written

= ln

j 2Ch

k2Cjh h

exp[(xhi;jh )0 h +
Ik;jh k;jh ]

0  k;1 


 k;L 1  1
When the decision level is at 1, there are no inclusive values. Therefore, the conditional probability is
defined as

Pi (j1 j1 ) =

Uij

= x0ij  + ij + ij

where the error component, ij , is correlated among

alternatives and heteroskedastic and another error
component, ij , independently and identically distributed. When we assume that ij = z0ij , random
coefficients, , have the mixing distribution with the
probability density function f ( j). The choice probability of an alternative j is

Pi (j ) =

exp[(xhi;j1 1 )0 1 ]
1 0 1
j 2C1 exp[(xi;j1 )  ]

Qi (j j )f ( j)d

where

Qi (j j ) =

0
0
P exp(xij  0+ zij )0
k2Ci exp(xik  + zik )

Mixed multinomial logit models are estimated by

simulation-based methods (simulated maximum likelihood or method of simulated moments). The simulated probability is written

P~i (j ) =

exp[(xhi;jh h )0 h + k2Cjh h =k;jh h ]

P
h
0 h P
j 2Ch exp[(xi;jh )  + k2Cjh h =k;jh ]

Ih

The utility function of a decision maker can be specified as

S
1X
Q (j j s )

S s=1 i

where S is the number of simulation replications

drawn from density f (
) and s is a simulated value.
The simulated log-likelihood function is calculated as
P
log L~ = Ni=1 log(P~i (j )). The simulated probability
is an unbiased estimate of Pi (j ), and its variance
decreases as replications (S ) increase. Therefore,
it is critical to find out the most efficient method of
simulation. Train (1999) argues that quasi-random
numbers (Halton sequences) can provide much better simulation-based estimation method than pseudo
random numbers. His finding is that 100 Halton draws
have smaller simulation variance than 1000 random
draws. However, the properties of Halton sequences
have not been fully investigated.

Multinomial Logit Example

Hensher and Greene analyze travel mode choice for
travel between Sydney and Melbourne. They used
data on four travel modes: airplane, train, bus, and
car. The data set contains 210 observations. Refer to Greene (2000) for more detailed data descriptions. In this example, we assume that the choice of
transportation mode is a function of terminal waiting
time (TTIME), in-vehicle time (INVT), in-vehicle cost

Statistics and Data Analysis

(INVC), generalized cost measure (GC), and income

(INCOME). The variable INCOME contains zero value
if the choice is not airplane. Original data is rescaled
dividing by 100. The multinomial choice models are
estimated using the conditional logit, nested logit, and
mixed multinomial logit log-likelihood functions. The
log-likelihood function of the mixed multinomial logit is
calculated using simulation with 125 Halton sequence
draws. All coefficients are assumed to be random
with normal distribution when the mixed logit is estimated. In the following table, the mean and standard
deviation of random coefficients are shown while the
conditional and nested logit models estimate fixed parameters. It is interesting to observe that parameter
estimates of conditional logit and mixed logit models
are almost identical when we assume that the random
parameters have the normal distribution. Asymptotic
standard errors are reported in parentheses.
Multinomial Logit Modeling
Conditional
Nested
Mixed
Parameter
Logit
Logit
Logit
TTIME
-3.5606
-2.4540
-3.5606
(0.4725)
(0.4253) (0.4725)
0.0213
TTIME
(0.6408)
INVC
-2.6820
-0.4933
-2.6820
(1.4694)
(3.7622) (1.4695)
0.0093
INVC
(1.0168)
INVT
-0.6202
-1.6618
-0.6202
(0.1857)
(0.5769) (0.1857)
-0.0001
INVT
(0.0506)
GC
3.4269
2.4142
3.4269
(1.3884)
(3.6993) (1.3885)
-0.0046
GC
(0.7071)
INCOME
1.2462
3.7627
1.2464
(0.6637)
(0.9876) (0.6637)
0.0201
INCOME
(0.8404)
INCLUDE 1
2.1817
(0.4304)
INCLUDE 2
0.1273
(0.0946)
log L
-242.394
-193.666 -242.393

The QLIM statement is given as follows:

proc qlim data=travel id=id;
/*-- conditional logit --*/
model mode = ttime invc invt gc income /
noint type=clogit choice=decision;
endogenous discrete=(mode 1 2 3 4);
/*-- nested logit --*/

model upmode mode = ttime invc invt gc income /

type=nestedlogit choice=decision;
utility u(1, 2 3 4 @ 2) = ttime invc invt gc,
u(2, 1 2) = income;
endogenous discrete=(mode 1 2 3 4, upmode 1 2);
/*-- mixed logit --*/
model mode = ttime invc invt gc income /
type=mixedlogit choice=decision
mixed=(normalvar=ttime invc invt gc income,
randnum=halton);
endogenous discrete=(mode 1 2 3 4);
run;

Heckmans Two-Equation Model Example

Heckmans two-equation simultaneous system is defined

= y2i 1 + x01i 1 + 1 di + u1i

= y1i 2 + x02i 2 + 2 di + u2i

1 if y2i > 0
di =
0 otherwise
2 =
2 1

y1i
y2i

where x1i and x2i are K1  1 and K2  1 exogenous

variable vectors, and (u1i ; u2i ) are iid bivariate normal variables. The reduced form of this system of two
equations can be written

y1
y2

= 1 d + X 1 + e1
= X  2 + e2

where X consists of all distinct column

 vectorsof

X1 and X2 and (e1i ; e2i )

2
0; 1 112
12

Amemiya (1978) suggested the two-stage estimation

method using the following equations:

 1 = J1 1 +  2 1
 2 = J2 2 +  1 2
where J1 and J2 are selection matrices for X1 and X2 ,
respectively.
Copley et al. (1994) studied the relation of quality of
audit services (QUALINDX) with audit fees (LNFEE)
using Heckmans two-equation simultaneous equations model with 1 = 0. They find that the audit fee
is positively related with quality of service in the audit
supply equation, while there is a negative relationship
between the demand for audit quality and audit service fees. However, single equation modeling does
not reveal this relationship because of simultaneous
equations bias. Their specification of simultaneous
equations system is the following:
QUALINDX = f1 (LNSIZE; FINOFFCL;

GOVT; LNFEE)

Statistics and Data Analysis

LNFEE = f2 (LNSIZE; LNBIDS; COSTWGHT;

TENURE; QUALINDX)

Refer to Copley et al. for more details on data and

variable definition. We randomly selected 70% of the
data set (sample size = 118) that was used by Copley et al. and reported two-stage OLS and GLS estimates. The estimates in the demand equation show
sizable difference when we compare with those of
Copley et al. The OLS coefficient estimate of LNFEE
is not significant though the sign is the same as the
GLS. Deis and Hill (1998) carried out bootstrapping.
Their finding is that some asymptotic t-statistics of the
Amemiyas GLS estimates are inflated by 55% when
they use Copley et al.s data set with sample size 118.
proc qlim data=account system;
model qualindx = lnsize finoffcl govt lnfee;
model lnfee = lnsize lnbids costwght tenure qualindx;
endogenous lnfee discrete=(qualindx 0 1);
instruments lnsize lnbids finoffcl
govt costwght tenure;
run;

Simultaneous Equations Estimates of Demand

Function for Audit Quality
Variable
Parameter Standard
t statistic
Name
Estimate
Error
GLS Estimates (sample size = 77)
INTERCEPT
1.9326
2.4245
0.80
LNSIZE
0.7934
0.4924
1.61
FINOFFCL
0.5279
0.5911
0.89
GOVT
2.2846
1.0190
2.24
LNFEE
-1.5916
1.0341
-1.54
OLS Estimates (sample size = 77)
INTERCEPT
0.8625
2.0647
0.42
LNSIZE
0.4248
0.5008
0.85
FINOFFCL
0.4851
0.3270
1.48
GOVT
1.7400
0.8290
2.10
LNFEE
-0.8607
1.0715
-0.80

Simultaneous Equations Estimates of Supply

Function for Audit Quality
Variable
Parameter Standard
t statistic
Name
Estimate
Error
GLS Estimates (sample size = 77)
INTERCEPT
2.5884
1.0489
2.47
LNSIZE
0.4208
0.0730
5.76
LNBIDS
-0.1123
0.2235
-0.50
COSTWGHT
-0.4687
0.3584
-1.31
TENURE
0.0887
0.0428
2.07
QUALINDX
0.4944
0.3188
1.55
OLS Estimates (sample size = 77)
INTERCEPT
2.5543
1.0514
2.43
LNSIZE
0.4292
0.0752
5.70
LNBIDS
-0.1209
0.2243
-0.54
COSTWGHT
-0.4626
0.3586
-1.29
TENURE
0.0892
0.0428
2.09
QUALINDX
0.5021
0.3192
1.57

Conclusion
In this paper, three examples are given to introduce
how the QLIM procedure can be used to solve real
problems. However, there are many other interesting
features. For example, you can use PROC QLIM for
count data and limited dependent variable modeling.
Predicted values and marginal effects are calculated
with the OUTPUT statement in the QLIM procedure.
You can also use PROC QLIM to fit switching regression models.

References
Amemiya, T. (1978), The Estimation of a Simultaneous Equation Generalized Probit Model, Econometrica, 46, 11931205.
Amemiya, T. (1985), Advanced Econometrics, Cambridge: Harvard University Press.
Ben-Akiva, M. and Lerman, S.R. (1985), Discrete
Choice Analysis, Cambridge: MIT Press.
Brownstone, D. and Train, K. (1999), Forecasting
New Product Penetration with Flexible Substitution Patterns, Journal of Econometrics, 89, 109
129.
Copley, P.A., Doucet, M.S., and Gaver, K.M. (1994),
A Simultaneous Equations Analysis of Quality
Control Review Outcomes and Engagement Fees
for Audits of Recipients of Federal Financial Assistance, The Accounting Review, 69, 244256.
Deis, D.R. and Hill, R.C. (1998), An Application of the
Bootstrap Method to the Simultaneous Equations

Statistics and Data Analysis

Model of the Demand and Supply of Audit Services, Contemporary Accounting Research, 15,
8399.
Greene, W.H. (2000), Econometric Analysis, 4th ed.,
Upper Saddle River, N.J.: Prentice Hall.
Heckman, J.J. (1978), Dummy Endogenous Variables in a Simultaneous Equation System,
Econometrica, 46, 931959.
Lee, L.-F. (1981), Simultaneous Equations Models with Discrete and Censored Dependent Variables, in Structural Analysis of Discrete Data with
Econometric Applications, ed. C.F. Manski and D.
McFadden, Cambridge: MIT Press.
Train, K. (1999), Halton Sequences for Mixed Logit,
working paper, University of California, Berkeley.

Contact Information
Minbo Kim, SAS Institute Inc., SAS Campus Drive,
Cary, NC 27513. Email [email protected]
SAS and SAS/ETS are registered trademarks of SAS
Institute Inc. in the USA and other countries. indicates USA registration.