AStA (2008) 92: 197206

DOI 10.1007/s10182-008-0056-3
ORI GI NAL PAPER
An extension of the BlinderOaxaca decomposition
to nonlinear models
Thomas K. Bauer Mathias Sinning
Received: 23 April 2007 / Accepted: 11 October 2007 / Published online: 6 February 2008
Springer-Verlag 2008
Abstract This paper develops a general BlinderOaxaca decomposition that allows
the differences in an outcome variable between two groups to be decomposed into
(i) a part that is explained by differences in observed characteristics and (ii) a part
attributable to differences in the estimated coefcients also for nonlinear regression
models. Based on this general model, we show how it can be applied to different
models with discrete and limited dependent variables.
Keywords BlinderOaxaca decomposition Nonlinear regression models General
econometric methods
1 Introduction
The decomposition method developed by Blinder (1973) and Oaxaca (1973) and gen-
eralized by Juhn et al. (1991), Neumark (1988), and Oaxaca and Ransom (1988,
1994) is a popular descriptive tool in empirical economics. The method allows the
decomposition of outcome variables between two groups into (i) a part that is ex-
plained by differences in observable characteristics and (ii) a part attributable to dif-
ferences in the returns to these characteristics. So far, these decomposition methods
have been applied mainly in the context of linear regression models. In many research
questions, however, the specic outcome variable requires the estimation of nonlin-
ear models. In the case of discrete or limited dependent variables, for example, OLS
yields inconsistent parameter estimates and in turn misleading decomposition results.
In particular, since the parameter estimates of nonlinear models typically differ from
T.K. Bauer M. Sinning ()
Rheinisch-Westflisches Institut fr Wirtschaftsforschung (RWI Essen), Hohenzollernstr. 1-3,
45128 Essen, Germany
e-mail: sinning@rwi-essen.de
198 T.K. Bauer, M. Sinning
the marginal effects of the latent outcome variable, they cannot be used to perform a
standard BlinderOaxaca decomposition.
A decomposition method for models with binary dependent variables was devel-
oped by Fairlie (1999, 2003). Yun (2004) provided an extension of this method that
allows a detailed decomposition for all explanatory variables included in the respec-
tive model. Decomposition methods for quantile regression models were developed
by Machado and Mata (2005) and Melly (2005). In the linear context, Montgomery
and Powell (2003) determined the BlinderOaxaca decomposition based on FIML
tobit estimates. Finally, Fitzenberger et al. (2006) adapted the BlinderOaxaca de-
composition to probit models to decompose changes in net union densities over time.
In this paper, we generalize the BlinderOaxaca decomposition method to other non-
linear models. Based on this generalized decomposition, we demonstrate how the
BlinderOaxaca decomposition can be applied to different models with discrete and
limited dependent variables.
2 An extension of the BlinderOaxaca decomposition to nonlinear models
Consider the following linear regression model, which is estimated separately for the
groups g =(A, B),
Y
ig
=X
ig

g
+
ig
,
for i = 1, . . . , N
g
, and

g
N
g
= N. For these models, Blinder (1973) and Oaxaca
(1973) proposed the decomposition
Y
A
Y
B
=
OLS
=
_
X
A
X
B
_

A
+X
B
_

B
_
, (1)
where Y
g
=N
1
g

N
g
i=1
Y
ig
and X
g
=N
1
g

N
g
i=1
X
ig
. The rst termon the right-hand
side of (1) displays the difference in the outcome variable between the two groups A
and B that is due to differences in observable characteristics, whereas the second
term shows the differential that is due to differences in the estimated coefcients.
A decomposition of the outcome variable similar to (1) is not appropriate for non-
linear (NL) models, because the conditional expectations E(Y
ig
| X
ig
) may differ
from X
g

g
. Therefore, we propose to decompose the mean difference of Y
i
using
conditional expectations evaluated at different coefcient estimates:

A,NL
=
_
E

A
(Y
iA
| X
iA
) E

A
(Y
iB
| X
iB
)
_
+
_
E

A
(Y
iB
| X
iB
) E

B
(Y
iB
| X
iB
)
_
, (2)
where E

g
(Y
ig
| X
ig
) refers to the conditional expectation of Y
ig
and E

g
(Y
ih
| X
ih
)
to the conditional expectation of Y
ih
evaluated at the parameter vector
g
, with g, h =
(A, B) and g = h. Changing the reference group, an alternative expression for the
decomposition is

B,NL
=
_
E

B
(Y
iA
| X
iA
) E

B
(Y
iB
| X
iB
)
_
+
_
E

A
(Y
iA
| X
iA
) E

B
(Y
iA
| X
iA
)
_
. (3)
An extension of the BlinderOaxaca decomposition 199
In both equations, the rst term on the right-hand side displays the part of the differ-
ential in the outcome variable between the two groups that is due to differences in the
covariates X
ig
, and the second term the part of the differential in Y
ig
that is due to
differences in the coefcients
g
.
To apply this decomposition to different nonlinear models, one has to derive the
respective sample counterparts S(

g
, X
ig
) and S(

h
, X
ig
) of the conditional expecta-
tions E

g
(Y
ig
| X
ig
) and E

h
(Y
ig
| X
ig
) for g, h =(A, B) and g =h. The following
section illustrates the application of (2) for different models with discrete and lim-
ited dependent variables. An estimation of the corresponding components of (3) is
straightforward.
1
Note that this decomposition shares all problems of the original Blinder
Oaxaca decomposition, e.g., a potential sensitivity of the results with respect to
the choice of the reference group and the specication of the regression model
that results in biased coefcients and hence misleading decomposition results (see
Oaxaca and Ransom 1999).
3 Discrete dependent variable models
3.1 Logit and probit models
Discrete dependent variable models comprise binary and ordered logit and probit
models as well as models for count data.
2
Because binary logit and probit models
may be considered as a special case of ordered logit and probit models, the decom-
position method for binary dependent variables proposed by Fairlie (1999, 2003)
represents a special case of the BlinderOaxaca decomposition for ordered choice
models. Ordered logit and probit models (O) are frequently used, for example, as a
framework for analyzing outcomes of opinion surveys. These models are based on a
latent model of the form
Y

ig
=X
ig

g,O
+
ig,O
,
where Y

ig
is unobserved. Instead of Y

ig
, only the following realizations are observed:
Y
ig
= 0 if Y

ig
0,
= 1 if 0 <Y

ig

1
,
= 2 if
1
<Y

ig

2
,

= J if
J1
Y

ig
,
1
The bootstrap method may be applied to obtain standard errors and condence intervals of the compo-
nents of (2) and (3). The program code for the estimation was written in STATA(R) and is available online
at www.rwi-essen.de/sinning.
2
Note that the proposed method cannot be applied to multinomial models. Our decomposition method
requires the calculation of conditional expectations. These, however, cannot be calculated for unordered
discrete dependent variables.
200 T.K. Bauer, M. Sinning
where the s are unknown parameters to be estimated together with the coefcients

g,O
. The conditional expectation of Y
ig
evaluated at the parameter vector
g,O
can
be written as
E

g,O
(Y
ig
| X
ig
) = F(
1
X
ig

g,O
) F(X
ig

g,O
)
+2
_
F(
2
X
ig

g,O
) F(
1
X
ig

g,O
)
_
+
+J
_
1 F(
J1
X
ig

g,O
)
_
.
Assuming that the error term
ig,O
is normally distributed across observations leads
to the ordered probit model, where F() is dened as the cumulative standard normal
distribution (). The logit model is obtained when the error term
ig,O
is assumed
to follow a logistic distribution, i.e., when F() represents a cumulative logistic dis-
tribution ().
Given the estimates of the parameter vector
g,O
, the sample counterparts of the
single components of the decomposition equation can be calculated as
S
_

g,O
, X
ig
_
=
1
N
N

i=1
__
F
_

1
X
ig

g,O
_
F
_
X
ig

g,O
__
+2
_
F
_

2
X
ig

g,O
_
F
_

1
X
ig

g,O
__
+
+J
_
1 F
_

J1
X
ig

g,O
___
. (4)
The sample counterpart of E

h,O
(Y
ig
| X
ig
), S(

h,O
, X
ig
), is obtained by replacing

g,O
with

h,O
in (4).
3
These sample counterparts can then be used to calculate the
single parts of the decomposition (2) as

O
=
_
S
_

A,O
, X
iA
_
S
_

A,O
, X
iB
__
+
_
S(

A,O
, X
iB
) S
_

B,O
, X
iB
__
.
The BlinderOaxaca decomposition for ordered choice models reduces to the decom-
position method for binary choice models if J =1.
3.2 Count data models
The Poisson regression model (P), which represents the basic model used to study
count data, assumes that the dependent variable Y
ig
conditional on the covariates X
ig
is Poisson distributed with density
f (Y
ig
| X
ig
) =
exp(
ig
)
Y
ig
ig
Y
ig
!
, Y
ig
=0, 1, 2, . . .
3
Because the calculation of S(

h
, X
ig
) is straightforward, we will present just the calculation of
S(

g
, X
ig
) in the remainder of the paper.
An extension of the BlinderOaxaca decomposition 201
and conditional expectation
E(Y
ig
| X
ig
) =
ig
=exp(X
ig

g,P
).
The sample counterpart of E

g,P
(Y
ig
| X
ig
) is given by
S
_

g,P
, X
ig
_
=Y
g,

g,P
=
1
N
g
N
g

i=1
exp
_
X
ig

g,P
_
.
A well-known problem of the Poisson model is the assumption that the dependent
variable has the same mean and variance
ig
= exp(X
ig

g,P
). If this assumption is
violated, an alternative conditional distribution of the dependent variable needs to be
used, one that permits a more exible specication of the variance of the dependent
variable. The negative binomial (negbin) regression model (NB) represents such an
alternative. The negbin regression model relaxes the assumption of equality of the
conditional mean and the variance of the dependent variable, while assuming the
same form of the conditional mean as the Poisson model.
The sample counterpart of the conditional mean of the negbin regression model is
S
_

g,NB
, X
ig
_
=Y
g,

g,NB
=
1
N
g
N
g

i=1
exp
_
X
ig

g,NB
_
.
In contrast to the Poisson model, the negbin model assumes a quadratic relationship
between the variance and the mean, i.e.
V(Y
ig
| X
ig
) =
ig
+
2
ig
,
where is a scalar parameter to be estimated together with
g,NB
.
In addition to the Poisson and negbin regression models, zero-inated models are
frequently used when analyzing count data. These models take into account that real-
life data may contain excess zeros, causing a higher probability of zero values than
is consistent with the Poisson and the negative binomial distribution. In this case it
could be assumed that zeros and positive values do not come from the same data-
generating process (Winkelmann 2000).
In order to investigate the probability of excess zeros, Lambert (1992) proposed
a zero-inated Poisson model that allows for two different data-generating regimes:
the outcome of regime 1 (R
1
) is always zero, whereas the outcome of regime 2 (R
2
)
is generated by a Poisson process. In this model, the conditional expectation of Y
ig
given X
ig
consists of the conditional probability of observing regime 2 given X
ig
and
the conditional expectation of the zero-truncated density:
E(Y
ig
| X
ig
) =
_
1 Pr(R
1
| X
ig
)
_
E(Y
ig
| R
2
, X
ig
). (5)
Lambert (1992) specied the conditional probability of regime 1, that always leads
to a zero outcome, as a logit model:
Pr(R
1
| X
ig
) =
exp(Z
ig

g
)
1 +exp(Z
ig

g
)
,
202 T.K. Bauer, M. Sinning
where Z
ig
contains the covariates of the conditional probability of excess zeros and

g
is the parameter vector to be estimated. The mean of the dependent variable spec-
ied by (5) can then be estimated for the zero-inated Poisson (ZIP) and the zero-
inated negbin (ZINB) model by
S
_

g,j
, X
ig
_
=
1
N
g
N
g

i=1
_
1
_

Pr(R1) | X
ig
__

ig
=
1
N
g
N
g

i=1
exp(X
ig

g,j
)
1 +exp(Z
ig

g,j
)
,
for j =ZIP, ZINB.
The hurdle model represents another modication of count data models. This
model can be interpreted as a two-part model, where the rst part is a binary outcome
model and the second part a truncated count data model. The conditional expectation
of Y
ig
in this model is given by
E(Y
ig
| X
ig
) =Pr(Y
ig
>0 | X
ig
)E(Y
ig
| Y
ig
>0, X
ig
).
According to Cameron and Trivedi (1998) the conditional expected values of Y
ig
given Y
ig
> 0 and X
ig
of the hurdle Poisson (HP) and the hurdle negbin (HNB)
model are given by
E(Y
ig
| Y
ig
>0, X
ig
) =
exp(X
ig

HP
g
)
1 exp(exp(X
ig

HP
g
))
,
and
E(Y
ig
| Y
ig
>0, X
ig
) =
exp(X
ig

HNB
g
)
1 (1 + exp(X
ig

HNB
g
))

,
respectively. Assuming a logistic distribution for the underlying zero-generating
process, these expected values can be estimated by
S
_

HP
g
, X
ig
_
=
1
N
g
N
g

i=1
exp(X
ig

HP
g
)
(1 exp(exp(X
ig

HP
g
)))(1 +exp(Z
ig

HP
g
))
,
and
S
_

HNB
g
, X
ig
_
=
1
N
g
N
g

i=1
exp(X
ig

HNB
g
)
(1 (1 + exp(X
ig

HNB
g
))

)(1 +exp(Z
ig

HNB
g
))
.
4 Limited dependent variable models
4.1 Tobit models
Limited dependent variable models include truncated regression models and models
for censored and corner solution outcome variables. Technically, models with cen-
sored and corner solution outcome variables may be estimated appropriately by a
An extension of the BlinderOaxaca decomposition 203
tobit model (see, e.g., Wooldridge 2002). While censored outcome variables are not
observable for a part of the population (such as top-coded wage information or pre-
ferred labor supply), corner solution outcome variables take on the value zero with
positive probability but represent a continuous random variable over strictly positive
values (such as actual labor supply). In a general version of the tobit model (TB),
the dependent variable takes on the values a
1
and a
2
with positive probability and
represents a continuous random variable over values between a
1
and a
2
, i.e.
Y

ig
= X
ig

g,TB
+
ig,TB
,
Y
ig
= a
1
if Y

ig
a
1
,
Y
ig
= a
2
if Y

ig
a
2
,
Y
ig
= Y

ig
=X
ig

g,TB
+
ig,TB
if a
1
<Y

ig
<a
2
,

ig
N
_
0,
2
g,TB
_
.
If one is interested in the marginal effects of a latent censored outcome variable,
the strategy would be to use the tobit estimator in the standard BlinderOaxaca de-
composition depicted in (1). However, the conventional decomposition method leads
to erroneous predictions of the components of the decomposition equation if one
aims at analyzing the observable corner solution outcome variable Y
ig
. In this case,
an alternative decomposition method must be applied.
Assuming homoscedastic and normal distributed error terms
ig,TB
, the condi-
tional expectation of Y
ig
given X
ig
consists of the respective conditional expectations
and probabilities of observing a
1
, a
2
, or a value between a
1
and a
2
, i.e.
E(Y
ig
| X
ig
) = a
1

1
(
g,TB
, X
g
,
g,TB
) +a
2

2
(
g,TB
, X
g
,
g,TB
)
+(
g,TB
, X
g
,
g,TB
)

_
X
ig

g,TB
+
g,TB
(
g,TB
, X
g
,
g,TB
)
(
g,TB
, X
g
,
g,TB
)
_
, (6)
where
(
g,TB
, X
g
,
g,TB
) =1
_
(a
1
X
ig

g,TB
)

g,TB
_

_
(a
2
X
ig

g,TB
)

g,TB
_
,
and
(
g,TB
, X
g
,
g,TB
) =
_
(a
1
X
ig

g,TB
)

g,TB
_

_
(a
2
X
ig

g,TB
)

g,TB
_
.
() represents the standard normal density function. Using the sample counterpart
of (6),
S
_

g,TB
, X
ig
,
g,TB
_
=
1
N
N

i=1
a
1

1
_

g,TB
, X
g
,
g,TB
_
204 T.K. Bauer, M. Sinning
+a
2

2
_

g,TB
, X
g
,
g,TB
_
+
_

g,TB
, X
ig
,
g,TB
_

_
X
ig

g,TB
+
g,TB
(

g,TB
, X
ig
,
g,TB
)
(

g,TB
, X
ig
,
g,TB
)
_
,
the single parts of the decomposition equation can be estimated by

TB
=
_
S
_

A,TB
, X
iA
,
A,TB
_
S
_

A,TB
, X
iB
,
A,TB
__
(7)
+
_
S
_

A,TB
, X
iB
,
A,TB
_
S
_

B,TB
, X
iB
,
B,TB
__
.
If the dependent variable is not truncated, i.e. if a
1
and a
2
, (7) reduces
to the original BlinderOaxaca decomposition described in (1).
Depending on the choice of the standard error of the residual,
g,TB
, used in the
counterfactual part of the decomposition equation, alternative decompositions can be
derived. These versions of the decomposition equation may differ from each other if
there are large differences in the variance of the error term between the two groups
(see Bauer and Sinning 2005a for a comprehensive description of the BlinderOaxaca
decomposition for tobit models).
4.2 Truncated regression models
The results derived for the tobit model can easily be transferred to a truncated regres-
sion model of the form
Y
ig
=X
ig

g,TR
+
ig,TR
,
where the dependent variable is truncated at a lower limit a
1
and a higher limit a
2
.
The error terms
ig,TR
are assumed to be homoscedastic and distributed normally
with mean zero and variance
2
g,TR
. Consequently,
Y
ig
| X
ig
N
_
X
ig

g,TR
,
2
g,TR
_
.
In this model, the conditional expectation of Y
ig
given X
ig
consists of the conditional
expectation of Y
ig
given X
ig
and a
1
< Y
ig
< a
2
weighted with the probability of
observing a value between a
1
and a
2
:
E

g,TR
,
g,TR
(Y
ig
| X
ig
) = (
g,TR
, X
g
,
g,TR
)

_
X
ig

g,TR
+
g,TR
(
g,TR
, X
g
,
g,TR
)
(
g,TR
, X
g
,
g,TR
)
_
. (8)
Similar to (7), the components of the decomposition equation of the truncated regres-
sion model can be estimated by using the sample counterpart of (8):
S
_

g,TR
, X
ig
,
g,TR
_
=
1
N
N

i=1

g,TR
, X
ig
,
g,TR
_

_
X
ig

g,TR
+
g,TR
(

g,TR
, X
ig
,
g,TR
)
(

g,TR
, X
ig
,
g,TR
)
_
.
An extension of the BlinderOaxaca decomposition 205
5 Summary
This paper extends the decomposition method proposed by Blinder (1973) and Oax-
aca (1973) to nonlinear regression models. This extension permits a decomposition
of differences in an outcome variable between two groups into (i) a part that may be
explained by differences in observed characteristics and (ii) a part that is attributable
to differences in the estimated coefcients.
The paper illustrates how this general version of the BlinderOaxaca decomposi-
tion can be applied to models with discrete and limited dependent variables. In par-
ticular, a BlinderOaxaca decomposition method for ordered logit and probit models
is derived, which represents a generalization of the decomposition method for binary
logit and probit models proposed by Fairlie (1999, 2003). Moreover, the Blinder
Oaxaca decomposition is applied to count data models, including Poisson and neg-
ative binomial models, zero-inated Poisson and negative binomial models as well
as hurdle Poisson and negative binomial models. An empirical application of the de-
composition method for count data models was performed by Bauer et al. (2007).
Finally, the BlinderOaxaca decomposition is extended to truncated regression and
tobit models, where the latter was used by Bauer and Sinning (2005b) to analyze
differences in the savings behavior between natives and immigrants in Germany.
Acknowledgements The authors thank Michael Fertig, Markus Hahn, John P. Haisken-DeNew and
Harald Tauchmann for helpful comments.
References
Bauer, T.K., Sinning, M.: BlinderOaxaca decomposition for tobit models. RWI Discussion Papers No. 32,
110 (2005a)
Bauer, T.K., Sinning, M.: The savings behavior of temporary and permanent migrants in Germany. RWI
Discussion Papers No. 29, 137 (2005b)
Bauer, T.K., Ghlmann, S., Sinning, M.: Gender differences in smoking behavior. Health Econ. 19, 895
909 (2007)
Blinder, A.S.: Wage discrimination: reduced form and structural estimates. J. Hum. Resour. 8, 436455
(1973)
Cameron, A.C., Trivedi, P.K.: Regression Analysis of Count Data. Cambridge University Press, Cambridge
(1998)
Fairlie, R.W.: The absence of the african-american owned business: an analysis of the dynamics of self-
employment. J. Labor Econ. 17, 80108 (1999)
Fairlie, R.W.: An extension of the BlinderOaxaca decomposition technique to logit and probit models.
Yale University Economic Growth Center Discussion Paper No. 873, 111 (2003)
Fitzenberger, B., Kohn, K., Wang, Q.: The erosion of Union membership in Germany: determinants, den-
sities, decompositions. IZA Discussion Papers No. 2193, 136 (2006)
Juhn, C., Murphy, K.M., Pierce, B.: Accounting for the slowdown in blackwhite wage convergence. In:
Kosters, M.H. (ed.) Workers and Their Wages: Changing Patterns in the United States. American
Enterprise Institute, Washington (1991)
Lambert, D.: Zero-inated Poisson regression with an application to defects in manufacturing. Techno-
metrics 34, 114 (1992)
Machado, J.A.F., Mata, J.: Counterfactual decomposition of changes in wage distributions using quantile
regression. J. Appl. Econom. 20, 445465 (2005)
Melly, B.: Decomposition of differences in distribution using quantile regression. Labour Econ. 12, 577
590 (2005)
Montgomery, M., Powell, I.: Does an advanced degree reduce the gender wage gap? Evidence from MBAs.
Ind. Relat. 42, 396418 (2003)
206 T.K. Bauer, M. Sinning
Neumark, D.: Employers discriminatory behavior and the estimation of wage discrimination. J. Hum.
Res. 23, 279295 (1988)
Oaxaca, R.L.: Malefemale wage differentials in urban labor markets. Int. Econ. Rev. 14, 693709 (1973)
Oaxaca, R.L., Ransom, M.: Searching for the effect of unionism on the wages of union and nonunion
workers. J. Labor Res. 9, 139148 (1988)
Oaxaca, R.L., Ransom, M.: On discrimination and the decomposition of wage differentials. J. Econom.
61, 521 (1994)
Oaxaca, R.L., Ransom, M.: Identication in detailed wage decompositions. Rev. Econ. Stat. 81, 154157
(1999)
Winkelmann, R.: Econometric Analysis of Count Data. Springer, Berlin (2000)
Wooldridge, J.M.: Econometric Analysis of Cross Section and Panel Data. MIT Press, Cambridge (2002)
Yun, M.: Decomposing differences in the rst moment. Econ. Letters 82, 275280 (2004)