ETH Zurich Sociology Working Paper No. 5
A Stata implementation of the Blinder-Oaxaca decomposition
Ben Jann
May 2008
ETH Zurich, Chair of Sociology
SEW E 21, Scheuchzerstrasse 70 8092 Zurich, Switzerland
Tel. +41 44 632 55 56 Fax+41446321054 info@soz.gess.ethz.ch www.socio.ethz.ch
A Stata implementation of the Blinder-Oaxaca decomposition
Ben Jann
ETH Zurich jannb@ethz.ch
May 2008
Abstract.
Keywords: st0001, Blinder-Oaxaca decomposition, outcome differential, wage gap
1 Introduction
An often used methodology to study labor market outcomes by groups (sex, race, and so on) is to decompose mean differences in log wages based on regression models in a counterfactual manner. The procedure is known in the literature as the Blinder- Oaxaca decomposition (Blinder 1973; Oaxaca 1973) and divides the wage differential between two groups into a part that is “explained” by group differences in productivity characteristics such as education or work experience and a residual part that cannot be accounted for by such differences in wage determinants. This “unexplained” part is often used as a measure for discrimination, but it also subsumes the effects of group differences in unobserved predictors. Most applications of the technique can be found in the labor market and discrimination literature (for meta studies see, e.g., Stanley and Jarrell 1998 or Weichselbaumer and Winter-Ebmer 2005). However, the method may also be useful in other fields. In general, the technique can be employed to study group differences in any (continuous and unbounded) outcome variable. For example, O’Donnell et al. (2008) use it to analyze health inequalities by poverty status.
The purpose of this article is to give an overview of the Blinder-Oaxaca decomposi- tion and introduce a new command implementing the technique in Stata called oaxaca. In the next section, the most common variants of the procedure are summarized and a number of issues such as the identification of the contribution of categorical predictors or the estimation of standard errors are addressed. The third section then describes the syntax and options of the new oaxaca command and the fourth section provides examples for its application using labor market data.
The counterfactual decomposition technique popularized by Blinder (1973) and Oaxaca (1973) is widely used to study mean outcome differences be- tween groups. For example, the technique is often used to analyze wage gaps by sex or race. The present paper summarizes the technique and addresses a num- ber of complications such as the identification of effects of categorical predictors in the detailed decomposition or the estimation of standard errors. A new Stata command called oaxaca is introduced and examples illustrating its usage are given.
2 A Stata implementation of the Blinder-Oaxaca decomposition 2 Methods and formulas
Given are two groups A and B, an outcome variable Y , and a set of predictors. For example, think of males and females, (log) wages as the outcome variable, and human capital indicators such as education and work experience as predictors. The question now is how much of the mean outcome difference
R = E(YA) − E(YB) (1) where E(Y ) denotes the expected value of the outcome variable, is accounted for by
group differences in the predictors. Based on the linear model
Yl=Xl′βl+εl, E(εl)=0, l∈{A,B} (2)
where X is a vector containing the predictors and a constant, β contains the slope parameters and the intercept, and ε is the error, the mean outcome difference can be expressed as the difference in the linear prediction at the group-specific means of the regressors. That is
R = E(YA) − E(YB) = E(XA)′βA − E(XB)′βB (3)
since
E(Yl) = E(Xl′βl + εl) = E(Xl′βl) + E(εl) = E(Xl)′βl with E(βl) = βl and E(εl) = 0 by assumption.
To identify the contribution of group differences in predictors to the overall outcome difference, equation (3) can be rearranged, for example, as follows (see Winsborough and Dickinson 1971; Jones and Kelley 1984; Daymont and Andrisani 1984):
R = [E(XA) − E(XB)]′ βB + E(XB)′(βA − βB) + [E(XA) − E(XB)]′ (βA − βB) (4) This is a “three-fold” decomposition, that is, the outcome difference is divided into three
parts:
The first summand
R=E+C+I
E = [E(XA) − E(XB)]′ βB
amounts to the part of the differential that is due to group differences in the predictors (the “endowments effect”). The second component
C = E(XB)′(βA − βB)
measures the contribution of differences in the coefficients (including differences in the
intercept). The third summand
I = [E(XA) − E(XB)]′ (βA − βB)
Ben Jann 3 is an interaction term accounting for the fact that differences in endowments and coef-
ficients exist simultaneously between the two groups.
Decomposition (4) is formulated from the viewpoint of Group B. That is, the group differences in the predictors are weighted by the coefficients of Group B to determine the endowments effect (E). In other words, the E component measures the expected change in Group B’s mean outcome, if Group B had Group A’s predictor levels. Similarly, for the second component (C), the differences in coefficients are weighted by Group B’s predictor levels. That is, the second component measures the expected change in Group B’s mean outcome, if Group B had Group A’s coefficients. Naturally, the differential can analogously be expressed from the viewpoint of Group A, yielding the reverse three-fold decomposition
R = [E(XA) − E(XB)]′ βA + E(XA)′(βA − βB) − [E(XA) − E(XB)]′ (βA − βB) (5)
Now the “endowments effect” amounts to the expected change of Group A’s mean outcome, if Group A had Group B’s predictor levels. The “coefficients effect” quantifies the expected change in Group A’s mean outcome, if Group A had Group B’s coefficients.
An alternative decomposition that is prominent in the discrimination literature re- sults from the concept that there is some nondiscriminatory coefficients vector that should be used to determine the contribution of the differences in the predictors. Let β∗ be such a nondiscriminatory coefficients vectors. The outcome difference can then be written as
R = [E(XA) − E(XB)]′ β∗ + [E(XA)′(βA − β∗) + E(XB)′(β∗ − βB)] (6) We now have a “two-fold” decomposition
R=Q+U
where the first component
Q = [E(XA) − E(XB)]′ β∗
is the part of the outcome differential that is “explained” by group differences in the
predictors (the “quantity effect”) and the second summand
U = E(XA)′(βA − β∗) + E(XB)′(β∗ − βB)
is the “unexplained” part. The latter is usually attributed to discrimination, but it is important recognize that it to also captures all potential effects of differences in unobserved variables.
The “unexplained” part in (6) is sometimes further decomposed. Let βA = β∗ + δA and βB = β∗ + δB with δA and δA as group-specific discrimination parameter vectors (positive or negative discrimination, depending on sign). U can then be expressed as
U = E(XA)′δA − E(XB)′δB (7)
4 A Stata implementation of the Blinder-Oaxaca decomposition
that is, the unexplained component of the differential can be subdivided into a part UA = E(XA)′δA
that measures discrimination in favor of Group A and a part UB = −E(XB)′δB
that quantifies discrimination against Group B.1 Again, however, this interpretation hinges on the assumption that there are no relevant unobserved predictors.
The estimation of the components of the three-fold decompositions (4) and (5) is straightforward. Let βˆA and βˆB be the least squares estimates for βA and βB, obtained separately from the two group-specific samples. Furthermore, use the group means X ̄A and X ̄B as estimates for E(XA) and E(XB). Based on these estimates the decomposi- tions (4) and (5) are computed as
R = Y ̄A − X ̄B = (X ̄A − X ̄B)′βˆB + X ̄B′ (βˆA − βˆB) + (X ̄A − X ̄B)′(βˆA − βˆB) (8) and
R=Y ̄A −X ̄B =(X ̄A −X ̄B)′βˆA +X ̄A′ (βˆA −βˆB)−(X ̄A −X ̄B)′(βˆA −βˆB) (9)
The determination of the components of the two-fold decomposition (6) is more involved because an estimate for the unknown nondiscriminatory coefficients vector β∗ is needed. Several suggestions have been made in the literature. For example, there may be reason to assume that discrimination is directed towards one of the groups only, so that β∗ = βA or β∗ = βB (see Oaxaca 1973, who speaks of an “index number problem”). Again assume that Group A are males and Group B are females. If, for instance, wage discrimination is only directed against women and there is no (positive) discrimination of men, then we may use βˆA as an estimate for β∗ and compute decomposition (6) as
R=(X ̄A−X ̄B)′βˆA+X ̄B′ (βˆA−βˆB) (10) Similarly, if there is no discrimination of women, but only (positive) discrimination of
men, the decomposition is
R = (X ̄A − X ̄B)′βˆB + X ̄A′ (βˆA − βˆB) (11)
Often, however, there is no specific reason to assume that the coefficients of one or the other group are nondiscriminating. Moreover, economists have argued that the undervaluation of one group comes along with an overvaluation of the other (e. g. Cotton 1988). Reimers (1983) therefore proposes to use the average coefficients over both groups as an estimate for the nondiscriminatory parameter vector, that is
βˆ∗ = 0.5βˆA + 0.5βˆB (12)
1. UA and UB have opposite interpretations. A positive value for UA reflects positive discrimination of Group A; a positive value for UB indicates negative discrimination of Group B.
Ben Jann 5 Similarly, Cotton (1988) suggests to weight the coefficients by the group sizes nA and
nB, that is
Furthermore, based on theoretical derivations, Neumark (1988) advocates the usage of
βˆ∗ = nA βˆA + nB βˆB (13) nA + nB nA + nB
the coefficients from a pooled regression over both groups as an estimate for β∗.
Note that, as has been pointed out by Oaxaca and Ransom (1994) and others,
decomposition (6) can also be expressed as
R = [E(XA) − E(XB)]′ [WβA + (I − W)βB]+[(I − W)′E(XA) + W′E(XB)]′ (βA−βB) (14) where W is a matrix of relative weights given to the coefficients of Group A and I is the identity matrix. For example, choosing W = I is equivalent to setting β∗ = βA. Similarly, W = 0.5I is equivalent to β∗ = 0.5βA + 0.5βB. Furthermore, Oaxaca and
Ransom (1994) show that
W = Ω = (X′AXA + X′BXB)−1X′AXA (15)
with X as the observed data matrix, is equivalent to using the coefficients from a pooled model over both groups as the reference coefficients.2
An issue with the approach by Neumark (1988) or Oaxaca and Ransom (1994) is that it may inappropriately transfer some of the unexplained part of the differential into the explained component, although this does not seem to have received much attention in the literature.3 Assume a simple model of log wages (lnW) on education (Z) with sex-specific intercepts αM and αF due to discrimination. The model is
αM +γZ+ε if“male”
lnW = (16)
αF + γZ + ε if “female”
Let αM = α and αF = α + δ, where δ is the discrimination parameter. Then the model
may also be expressed as
ln W = α + γZ + δF + ε (17)
with F as an indicator for “female”. Assume that γ > 0 (positive relation between education and wages) and δ < 0 (discrimination against women). If we now use γ∗ from a pooled model
lnW =α∗ +γ∗Z+ε∗ (18) in decomposition (6), then, following from the theory on omitted variables (see, e.g.,
Gujarati 2003, 510f.), the explained part of the differential is
∗ Cov(Z,G)
Q=[E(ZM)−E(ZF)]γ =[E(ZM)−E(ZF)] γ+δ Var(Z) (19)
2. Another solution is to set W = diag(β − βB ) · diag(βA − βB )−1 , where β without subscript denotes the coefficients from the pooled model. Note that, although the decomposition results are the same, this approach yields a weighting matrix that is quite different from Oaxaca and Ransom’s Ω. For example, whereas W computed as described in this footnote is a diagonal matrix, Ω has off-diagonal elements unequal zero and is not even symmetric.
3. An exception its Fortin (2006)
6 A Stata implementation of the Blinder-Oaxaca decomposition
with Var(Z) as the variance of Z and Cov(Z,G) as the covariance between Z and G. If men on average are better educated than women, then the covariance between Z and G is negative and the explained part of the decomposition gets overstated (given γ > 0 and δ < 0). In essence, the difference in wages between men and women is explained by sex.
To avoid such a distortion of the decomposition results due to the residual group dif- ference spilling over into the slope parameters of the pooled model, my recommendation is to always include a group indicator in the pooled model as an additional covariate.
Estimation of sampling variances
Given the popularity of the Blinder-Oaxaca procedure it is astonishing how little at- tention has been paid to the issue of statistical inference. Most studies in which the procedure is applied only report point estimates for the decomposition results, but do notmakeanyindicationaboutsamplingvariancesorstandarderrors.4 However,foran adequate interpretation of the results approximate measures of statistical precision are indispensable.
Approximate variance estimators for certain variants of the decomposition were first proposed by Oaxaca and Ransom (1998) with Greene (2003, 53–54) making similar suggestions. The estimators by Oaxaca and Ransom (1998) and Greene (2003) are a good starting point, but they neglect an important source of variation. Most social- science studies on discrimination are based on survey data where all (or most) variables are random variables. That is, not only the outcome variable, but also the predictors are subject to sampling variation (an exception would be experimental factors set by the researcher). Whereas an important result for regression analysis is that it does not matter for the variance estimates whether regressors are stochastic or fixed, this is not true for the Blinder-Oaxaca decomposition. The decomposition is based on multiplying regression coefficients by means of regressors. If the regressors are stochastic, then the means have sampling variances. These variances are of the same asymptotic order than the variances of the coefficients (think of the means as the intercepts from regression models without covariates). To get consistent standard errors for the decomposition results it seems therefore important to take the variability induced by the randomness of the predictors into account.
Consider the expression
Y ̄ = X ̄′βˆ (20)
where X ̄ is the vector of mean estimates for the predictors and βˆ contains the least- squares estimates of the regression coefficients. If the predictors are stochastic, then X ̄ and βˆ are both subject to sampling variation. Assuming that X ̄ and βˆ are uncorrelated (which follows from the standard regression assumption that the conditional expectation of the error is zero for all covariate values; of course this is only true if the model is
4. Exceptions are for example Oaxaca and Ransom (1994, 1998), Silber and Weber (1999), Horrace and Oaxaca (2001), Fortin (2006), Heinrichs and Kennedy (2007) and Lin (2007). Furthermore, Jackson and Lindley (1989) and Shrestha and Sakellariou (1996) propose statistical tests for discrimination.
Ben Jann 7 correctly specified), the variance of (20) can be written as
V (X ̄ ′ βˆ) = E (X ̄ )′ V (βˆ)E (X ̄ ) + E (βˆ)′ V (X ̄ )E (βˆ) + trace V (X ̄ )V (βˆ) (21)
where V (X ̄ ) and V (βˆ) are the variance-covariance matrices for X ̄ and βˆ (see the proof in Jann 2005b; for the variance of the product of two independent random variables also see Mood et al. 1974, 180). By inserting estimates for the expectations and variance matrices, we get the variance estimator
V(X ̄′βˆ)=X ̄′V(βˆ)X ̄+βˆ′V(X ̄)βˆ+trace V(X ̄)V(βˆ) (22)
V(βˆ) is simply the variance-covariance matrix obtained from the regression procedure. AnaturalestimatorforV(X ̄)isV(X ̄)=X′X/[n(n−1)],whereX isthecentereddata matrix,i.e. X =X−1X ̄′.
The variances for the components of the Blinder-Oaxaca decomposition can be de- rived analogously. For example, ignoring the asymptotically vanishing5 last term in (22) and assuming that the two groups are independent, the approximate variance estimators for the two terms of decomposition (10) are
V([X ̄A−X ̄B]′βˆA)≈(X ̄A−X ̄B)′V(βˆA)(X ̄A−X ̄B)+βˆA′ V(X ̄A)+V(X ̄B) βˆA (23)
and
V(X ̄B′ [βˆA −βˆB])≈X ̄B′ V(βˆA)+V(βˆB) X ̄B +(βˆB −βˆB)′V(X ̄B)(βˆB −βˆB) (24)
where use of the fact is made that the variance of the sum of two uncorrelated random variables is equal to the sum of the individual variances. An interesting point about formulas (23) and (24) is that ignoring the stochastic nature of the predictors will primarily affect the variance of the first term of the decomposition (the “explained” part). This is because in most applications group differences in coefficients and means are much smaller then the levels of coefficients and means.
It is possible to develop similar formulas for all the decomposition variants outlined above, but derivations can get complicated once a pooled model is used and covariances between the pooled model and the group models have to be taken into account or if the assumption of independence between the two groups is loosened (e. g. if dealing with a cluster sample). An alternative approach that is simple and general and produces equivalent results is to estimate the joint variance-covariance matrix of all used statistics (see Weesie 1999, [R] suest) and then apply the “delta method” (see [R] nlcom and the references therein). In fact, in the case of independence between the two groups the results of the delta method for decomposition (4) are formally equal to (23) and (24). Furthermore, a general result for the delta method is that if the input variance matrix is asymptotically normal, then also the variance matrix of the transformed statistics is
5. Whereas the first and second terms are of order O(n−1), the last term is O(n−2).
8 A Stata implementation of the Blinder-Oaxaca decomposition
asymptotically normal (see, e. g., Greene 2003, 70). That is, since asymptotic normality holds for regression coefficients and mean estimates under very general conditions, the variances obtained by the delta method can be used to construct approximate confidence intervals for the decomposition results in the usual manner.
Detailed decomposition
Often not only is the total decomposition of the outcome differential into an explained and an unexplained part of interest, but also the detailed contributions of the single predictors or sets of predictors are subject to investigation. For example, one might want to evaluate how much of the gender wage gap is due to sex differences in education and how much is due to differences in work experience. Similarly, it might be informative to determine how much of the unexplained gap is related to differing returns to education and how much to differing returns to work experience.
Identifying the contributions of the individual predictors to the explained part of the differential is easy because the total component is a simple sum over the individual contributions. For example, for decomposition (10),
Q=(X ̄A−X ̄B)′βˆA =(X ̄1A−X ̄1B)βˆ1A+(X ̄2A−X ̄2B)βˆ2A+... (25)
where X ̄1, X ̄2, . . . are the means of the single regressors and βˆ1, βˆ2, . . . are the associated coefficients. The first summand reflects the contribution of the group differences in X ̄1, the second of differences in X ̄2, and so on. Also the estimation of standard errors for the individual contributions is straightforward.
Similarly, using decomposition (10) as an example, the individual contributions to the unexplained part are the summands in
U=X ̄′ (βˆ −βˆ )=X ̄′ (βˆ −βˆ )+X ̄′ (βˆ −βˆ )+... (26) B A B 1B 1A 1B 2B 2A 2B
Note, however, that other than for the explained part of the decomposition, the con- tributions to the unexplained part may depend on arbitrary scaling decisions if the predictors do not have natural zero points (e. g. Jones and Kelley 1984, 334). Without loss of generality, assume a simple model with just one explanatory variable:
Yl=β0l+β1lZl+εl, l∈{A,B} The unexplained part of the decomposition based on (10) then is
U = ( βˆ 0 A − βˆ 0 B ) + ( βˆ 1 A − βˆ 1 B ) Z ̄ B
The first summand is the part of the unexplained gap that is due to “group membership” (Jones and Kelley 1984); the second summand reflects the contribution of differing returns to Z. Now assume that the zero point of Z is shifted by adding a constant a. The effect of such a shift on the decomposition results is as follows:
U = (βˆ0A − aβˆ1A) − (βˆ0B − aβˆ1B) + (βˆ1A − βˆ1B)(Z ̄B + a)
Ben Jann 9
Evidently, the scale shift changes the results; a portion amounting to a(βˆ1A − βˆ1B ) is transferred from the group membership component to the part that is due to different slope coefficients. The conclusion is that the detailed decomposition results for the unexplained part only have a meaningful interpretation for variables for which scale shifts are not allowed, that is, for variables that have a natural zero point.6
A related issue that has received much attention in the literature is that the decompo- sition results for categorical predictors depend on the choice of the omitted base category (Jones 1983; Jones and Kelley 1984; Oaxaca and Ransom 1999; Nielsen 2000; Horrace and Oaxaca 2001; Gardeazabal and Ugidos 2004; Polavieja 2005; Yun 2005b). The ef- fect of a categorical variable is usually modeled by including 0/1-variables (“dummy” variables) for the different categories in the regression equation, where one of the cat- egories (the “base” category) is omitted to avoid collinearity. It is easy to see that the decomposition results for the single 0/1-variables depend on the choice of the base category, since the associated coefficients quantify differences with respect to the base category. If the base category changes, the decomposition results change.
For the explained part of the decomposition this may not be critical because the sum of the contributions of the single indicator variables, that is, the total contribution of the categorical variable is unaffected by the choice of the base category. For the unexplained part of the decomposition, however, there is again a tradeoff between the group membership component (the difference in intercepts) and the part attributed to differences in slope coefficients. For the unexplained part changing the base category not only alters the results for the singly dummy variables but also changes the contribution of the categorical variable as a whole.
An intuitively appealing solution to the problem has been proposed by Gardeazabal and Ugidos (2004) and Yun (2005b). The idea is to restrict the coefficients for the single categories to sum to zero, that is, to express effects as deviations from the grand mean. This can be implemented by restricted least squares estimation or by transforming the dummy variables before model estimation as proposed by Gardeazabal and Ugidos (2004).7 AmoreconvenientmethodinthecontextoftheBlinder-Oaxacadecomposition is to estimate the group models using the standard dummy coding and then transform the coefficients vectors so that deviations from the grand mean are expressed and the (redundant) coefficient for the base category is added (Suits 1984; Yun 2005b). If applied to such transformed estimates, the results of the Blinder-Oaxaca decomposition are independent of the choice of the omitted category. Furthermore, the results are equal to the simple averages of the results one would get from a series of decompositions in which the categories are used one after another as the base category (Yun 2005b).
6. The problem does not occur for the explained part of the decomposition or the interaction compo- nent in the three-fold decomposition because a cancels out in these cases. Furthermore, stretching or compressing the scales of the X variables (multiplication by a constant) does not alter any of the de- composition results because such multiplicative transformations are counterbalanced by the coefficient estimates.
7. In fact, the approach by Gardeazabal and Ugidos (2004) is simply what is known as the “effects coding” (Hardy 1993, 64–71) or the “deviation contrast coding” (Hendrickx 1999).
10 A Stata implementation of the Blinder-Oaxaca decomposition
The deviation contrast transform works as follows. Given is the model Y =β0 +β1D1 +...+βk−1Dk−1 +ε
where β0 is the intercept and Dj , j = 1, . . . , k − 1, are the dummy variables representing a categorical variable with k categories. Category k is the base category. Alternatively, the model may be formulated as
Y =β0 +β1D1 +...+βk−1Dk−1 +βkDk +ε where βk is constrained to zero. Now let
and define
The transformed model then is
c=(β1 +...+βk)/k
β ̃0=β0+c and β ̃j=βj−c, j=1,...,k
k
Y =β ̃0 +β ̃1D1 +...+β ̃kDk +ε, β ̃j =0
j=1
Note that the transformed model is mathematically equivalent to the untransformed model. For example, the two models produce identical predictions. The variance- covariance matrix for the transformed model can be obtained by applying the general formula for weighted sums of random variables given in, e. g., Mood et al. (1974, 179). Models with several sets of dummy variables can be transformed by applying the for- mulas to each set separately. Furthermore, the transformation can be applied to the interaction terms between a categorical and a continuous variable in an analogous man- ner except that now c is added to the main effect of the continuous variable instead of the intercept. Also note that the application of the transform is not restricted to linear regression. It can be used with any model as long as the effects of the dummies are expressed as additive effects.
Other restrictions to identify the contribution of a categorical variable to the unex- plained part of the decomposition are imaginable. For example, the restriction could be
k
w j β ̃ j = 0 j=1
where wj are weights proportional to the relative frequencies of the categories, so that the coefficients reflect deviations from the overall sample mean (Kennedy 1986; Haisken- DeNew and Schmidt 1997). Hence, there is still some arbitrariness in the method by Gardeazabal and Ugidos (2004) and Yun (2005b).
3 The oaxaca command
The methods presented above are implemented in a new command called oaxaca. The command first estimates the group models and possibly a pooled model over both
Ben Jann 11
groups using regress ([R] regress) or any user specified estimation command. suest is then applied, if necessary, to determine the combined variance-covariance matrix of the models ([R] suest) and the group means of the predictors are estimated using mean ([R] mean). Finally, the various decomposition results and their standard errors (and covariances) are computed based on the combined parameter vector and variance- covariance matrix of the models’ coefficients and the mean estimates.8 The standard errors are obtained by the delta method.9
3.1 Syntax
The syntax of the oaxaca command is
oaxaca depvar indepvars if in weight , by(groupvar ) swap detail(dlist) adjust(varlist) threefold(reverse) weight(# # ...) pooled(model opts) omega(model opts) reference(name) split x1(names and values) x2(names and values) categorical(clist) svy(svyspec) vce(vcetype) cluster(varname) fixed(varlist) nosuest nose model1(model opts) model2(model opts) xb noisily level(#) eform
where depvar is the outcome variable of interest (e.g. log wages) and indepvars are predictors (e. g. education, work experience, etc.). groupvar identifies the groups to be compared. oaxaca typed without arguments replays the last results.
fweights, aweights, pweights, and iweight are allowed; see [U] 11.1.6 weight. Furthermore, bootstrap, by, jackknife, statsby, and xi are allowed; see [U] 11.1.10 Prefix commands. Weights are not allowed with the bootstrap prefix and aweights are not allowed with the jackknife prefix. vce(), cluster(), and weights are not allowed with the svy option.
3.2 Options
Main
by(groupvar) specifies the groupvar that defines the two groups that will be compared. by() is not optional.
swap reverses the order of the groups.
detail(dlist) requests that the detailed results for the individual predictors be re-
8. The covariances between the models’ coefficients and the mean estimates are assumed zero in any case. This assumption may be violated in misspecified models.
9. nlcom ([R] nlcom) could be used to compute the variance-covariance matrix of the decomposition results. However, nlcom employs general methods based on numerical derivatives and is slow if the models contain many covariates. oaxaca therefore has its own specific implementation of the delta method based on analytic derivatives.
12
A Stata implementation of the Blinder-Oaxaca decomposition
ported. Use dlist to subsume the results for sets of regressors (variables not appear- ing in dlist are listed individually). The syntax for dlist is
name:varlist , name:varlist ...
The usual shorthand conventions apply to the varlists specified in dlist (see help varlist; additionally, cons is allowed). For example, specify detail(exp:exp*) to subsume exp (experience) and exp2 (experience squared). name is any valid Stata name and labels the set.
adjust(varlist) causes the differential to be adjusted by the contribution of the specified variables before performing the decomposition. This is useful, for example, if the specified variables are selection terms. Note that adjust() is not needed for heckman models.
Decomposition type
threefold(reverse) computes the three-fold decomposition. This is the default unless weight(), pooled, omega, or reference() is specified. The decomposition is expressed from the viewpoint of Group 2 (B). Specify threefold(reverse) to express the decomposition from the viewpoint of Group 1 (A).
weight(# # ...) computes the two-fold decomposition where # # ... are the weights given to Group 1 (A) relative to Group 2 (B) in determining the refer- ence coefficients (weights are recycled if there are more coefficients than weights). For example, weight(1) uses the Group 1 coefficients as the reference coefficients, weight(0) uses the Group 2 coefficients.
pooled(model opts) computes the two-fold decomposition using the coefficients from a pooled model over both groups as the reference coefficients. groupvar is included in the pooled model as an additional control variable. Estimation details may be specified in parentheses; see the model1() option below.
omega(model opts) computes the two-fold decomposition using the coefficients from a pooled model over both groups as the reference coefficients (without including groupvar as a control variable in the pooled model). Estimation details may be specified in parentheses; see the model1() option below.
reference(name) computes the two-fold decomposition using the coefficients from a stored model. name is the name under which the model was stored; see [R] esti- mates store. Do not combine the reference() option with bootstrap or jackknife methods.
split causes the ”unexplained” component in the two-fold decomposition to be split into a part related to Group 1 (A) and a part related to Group 2 (B). split is effective only if specified with weight(), pooled, omega, or reference().
Only one of threefold, weight(), pooled, omega, and reference() is allowed.
Ben Jann 13 X-Values
x1(names and values) and x2(names and values) provide custom values for specific predictors to be used for Group 1 (A) and Group 2 (B) in the decomposition. The default is to use the group means of the predictors. The syntax for names and values is
varname = value , varname = value . . . Example: x1(educ 12 exp 30)
categorical(clist) identifies sets of dummy variables representing categorical variables and transforms the coefficients so that the results of the decomposition are invariant to the choice of the (omitted) base category (deviation contrast transform). The syntax for clist is
varlist , varlist . . .
Each variable list must contain a variable for the base category (that is, the base category indicator must exist in the data). The transform can also be applied to interactions between a categorical and a continuous variable. Specify the continuous variable in parentheses at the end of the list in this case, i.e.
varlist (varname) , ...
and also include a list for the main effects. Example:
detail(d1 d2 d3, xd1 xd2 xd3 (x))
where x is the continuous variable, and d1 etc. and xd1 etc. are the main effects and interaction effects. The code for implementing the categorical() option has been taken from the devcon user command (Jann 2005a).
SE/SVY
svy(vcetype , svy options) executes oaxaca while accounting for the survey set- tings identified by svyset (this is essentially equivalent to applying the svy prefix command, although the svy prefix is not allowed with oaxaca due to some technical issues). vcetype and svy options are as described in [SVY] svy.
vce(vcetype) specifies the type of standard errors reported. vcetype may be may be analytic (the default), robust, cluster clustvar, bootstrap, or jackknife; see [R] vce option.
cluster(varname) adjusts standard errors for intragroup correlation; this is Stata 9 syntax for vce(cluster clustvar).
fixed(varlist) identifies fixed regressors (all if specified without argument; an example for fixed regressors are experimental factors). The default is to treat regressors as stochastic. Stochastic regressors inflate the standard errors of the decomposition components.
14 A Stata implementation of the Blinder-Oaxaca decomposition
nosuest prevents or enforces using suest to obtain the covariances between the mod- els/groups. suest is implied by pooled, omega, reference(), svy, vce(cluster), and cluster(). Specifying nosuest may cause biased standard errors and is strongly discouraged.
nose suppresses the computation of standard errors. Model estimation
model1(model opts) and model2(model opts) specify the estimation details for the two group-specific models. The syntax for model opts is
estcom , addrhs(spec) estcom options
where estcom is the estimation command to be used and estcom options are options allowed by estcom. The default estimation command is regress. addrhs(spec) adds spec to the “right-hand side” of the model. For example, use addrhs() to add extra variables to the model. Examples:
model1(heckman, select(varlist s) twostep)
model1(ivregress 2sls, addrhs((varlist2=varlist iv)))
Note that oaxaca uses the first equation for the decomposition if a model contains
multiple equations.
Furthermore, coefficients that occur in one of the model only are assumed zero for the other group. It is important, however, that the associated variables contain non-missing values for all observations in both groups.
noisily displays the models’ estimation output. Reporting
xb displays a table containing the regression coefficients and predictor values on which the decomposition is based.
level(#) specifies the confidence level, as a percentage, for confidence intervals. The default is level(95) or as set by set level.
eform specifies that the results be displayed in exponentiated form.
Ben Jann
3.3 Returned results
15
Scalars
e(N) e(N1)
Macros
e(cmd)
e(depvar)
e(group1)
e(wtype)
e(clustvar)
e(vcetype)
Matrices
e(b) e(b0)
Functions
e(sample)
number of observations number of obs. in Group 1
oaxaca
name of dependent variable value defining Group 1 weight type
name of cluster variable title used to label Std. Err.
decomposition results coefficients and X-values
marks estimation sample
e(N clust) e(N1)
number of clusters number of obs. in Group 2
name group variable value defining Group 2 weight expression
vcetype specified in vce()
4 Examples
Three-fold decomposition
e(by)
e(group2)
e(wexp)
e(vce) e(properties) b V
e(V) variance matrix of e(b) e(V0) variance matrix of e(b0)
The standard application of the Blinder-Oaxaca technique is to divide the wage gap between, say, men and women into a part that is explained by differences in determinants of wages such as education or work experience and a part that cannot be explained by such group differences. An example using data from the Swiss Labor Market Survey 1998 (Jann 2003) is as follows:
. use oaxaca, clear
(Excerpt from the Swiss Labor Market Survey 1998)
. oaxaca lnwage educ exper tenure, by(female) noisily
Model for group 1
Source SS df MS
Model 49.613308 3 16.5377693 Residual 122.143834 747 .163512495
Total 171.757142 750 .229009522
Number of obs = 751 F( 3, 747) = 101.14 Prob>F = 0.0000 R-squared = 0.2889 Adj R-squared = 0.2860 Root MSE = .40437
[95% Conf. Interval]
lnwage
educ
exper
tenure
_cons
Model for group 2
Source
Coef.
.0820549
.0098347
.0100314
2.24205
Std. Err. t
.0060851 13.48
.0016665 5.90
.0020397 4.92
.0778703 28.79
P>|t|
0.000 .070109
0.000 .0065632
0.000 .0060272
0.000 2.08918
.0940008
.0131062
.0140356
2.394921
Model 33.5197344 3 Residual 188.08041 679
SS df
MS
11.1732448
.276996185
Number of obs = 683 F( 3, 679) = 40.34 Prob>F = 0.0000 R-squared = 0.1513 Adj R-squared = 0.1475 Root MSE = .5263
Total 221.600144 682 .324926897
16
A Stata implementation of the Blinder-Oaxaca decomposition
lnwage
Coef.
Std. Err. t
.0087108 10.07
.0028971 4.52
.0035374 1.03
.1091691 19.22
Std. Err. z
.0174874 196.73
.0218522 149.49
.027988 6.20
.015693 5.43
.0255804 3.23
.010966 0.51
P>|t| [95% Conf. Interval]
educ .0877579 exper .0131074 tenure .0036577 _cons 2.097806
Blinder-Oaxaca decomposition
1: female = 0
2: female = 1
Coef.
Differential
Prediction_1 3.440222 Prediction_2 3.266761
Difference .1734607
Decomposit~n
Endowments .0852798
Coefficients .082563 Interaction .005618
0.000 .0706546
0.000 .0074191
0.301 -.0032878
0.000 1.883457
Number of obs =
.1048611
.0187958
.0106032
2.312156
1434
P>|z| [95% Conf. Interval]
0.000 3.405947
0.000 3.223932
0.000 .1186052
0.000 .0545222
0.001 .0324263
0.608 -.0158749
3.474497
3.309591
.2283163
.1160375
.1326996
.0271109
As is evident from the example, oaxaca first estimates two group-specific regression models and then performs the decomposition (the noisily option causes the group models’ results to be displayed and is specified in the example for purpose of illustra- tion). The default decomposition performed by oaxaca is the three-fold decomposition according to equation (4). To compute the reverse three-fold decomposition given in (5) specify threefold(reverse).
The decomposition output reports the mean predictions by groups and their differ- ence in the first panel. In our sample, the mean of the log wages is 3.44 for men and 3.27 for women, yielding a wage gap of 0.17. In the second panel of the decomposition output the wage gap is divided into three parts. The first part reflects the mean increase in women’s wages if they had the same characteristics as men. The increase of 0.085 in the example indicates that differences in endowments account for about half of the wage gap. The second term quantifies the change in women’s wages when applying the men’s coefficients to the women’s characteristics. The third part is the interaction term that measures the simultaneous effect of differences in endowments and coefficients.
Two-fold decomposition
Alternatively, a two-fold decomposition according to equation (6) can be requested, where weight(), pooled, or omega determines the choice of the reference coefficients. For example weight(1) corresponds to decomposition (10), weight(0) to decomposi- tion (11). omega causes the coefficients from a pooled model over both samples to be used as the reference coefficients, which is equivalent to Oaxaca and Ransom’s approach based on (15). The pooled option also causes the coefficients from a pooled model to
Ben Jann 17
be used, but now the pooled models also contains a group membership indicator. Based on the argumentation outlined in Section 2 my suggestion is to use pooled instead of omega.
For our example data, the results with the pooled option are as follows: . oaxaca lnwage educ exper tenure, by(female) pooled
Blinder-Oaxaca decomposition
1: female = 0
2: female = 1
Coef.
Differential
Prediction_1 3.440222 Prediction_2 3.266761
Difference .1734607
Decomposit~n
Explained .089347
Unexplained .0841137
Number of obs = 1434
Robust
Std. Err. z
.0174586 197.05
.0218042 149.82
.0279325 6.21
.0137531 6.50
.025333 3.32
P>|z|
0.000
0.000
0.000
0.000 0.001
[95% Conf. Interval]
3.406004
3.224026
.118714
.0623915
.034462
3.47444
3.309497
.2282075
.1163026
.1337654
Again the conclusion is that differences in endowments account for about half of the wage gap.10
A further possibility is to provide a stored reference model using the reference() option. For example, for the decomposition of the wage gap between blacks and whites the reference model is sometimes estimated based on all races, not just blacks and whites. In such a case, the reference model would have to be estimated first using all observations and then be provided to oaxaca via the reference() option.
Exponentiated results
The results in the example above are expressed on the logarithmic scale (remember that log wages are used as the dependent variable) and it might be sensible to re-transform the results to the original scale (Swiss francs in this case) using the eform option:
. oaxaca, eform
Blinder-Oaxaca decomposition
1: female = 0
2: female = 1
Robust
exp(b) Std. Err.
Number of obs = 1434
z P>|z| [95% Conf. Interval]
10. Note that unlike in the first example robust standard errors are reported (oaxaca uses suest to estimate the joint variance matrix for all coefficients if pooled is specified; suest implies robust standard errors). To compute robust standard errors in the first example you would have to add vce(robust) to the command.
18
A Stata implementation of the Blinder-Oaxaca decomposition
Differential
Prediction_1 31.19388 Prediction_2 26.22626
Difference 1.189414
Decomposit~n
Explained 1.09346
Unexplained 1.087753
.5446007
.5718438
.0332234
.0150385
.027556
197.05 0.000
149.82 0.000
6.21 0.000
6.50 0.000
3.32 0.001
30.14454
25.12908
1.126048
1.064379
1.035063
32.27975
27.37135
1.256346
1.123336
1.143125
The (geometric) means of wages are 31.2 Swiss francs for men and 26.2 francs for women, which amounts to a difference of 18.9 percent. Adjusting women’s endowments levels to the levels of men would increase women’s wages by 9.3 percent. A gap of 8.8 percent remains unexplained.
Survey estimation
oaxaca supports complex survey estimation, but svy has to be specified as an option and is not allowed as a prefix command (which does not restrict functionality). For example, variable wt provides sampling weights for the Swiss Labor Market Survey 1998. The weights (and strata or PSU’s if there were any) can be taken into account as follows:
. svyset [pw=wt]
pweight: wt
VCE: linearized
Single unit: missing
Strata 1:
SU 1:
FPC 1:
. oaxaca lnwage educ exper tenure, by(female) pooled svy
Blinder-Oaxaca decomposition
Number of strata = 1
Number of PSUs = 1434
1: female = 0
2: female = 1
Linearized
Coef. Std. Err. t
Differential
Prediction_1 3.405696 .0226315 150.48 Prediction_2 3.193847 .0276466 115.52
Difference .2118488 .0357284 5.93
Decomposit~n
Explained .1107614 .0189974 5.83
Unexplained .1010875 .0315911 3.20
Number of obs
Population size
Design df
= 1647
= 1424.3797
= 1433
P>|t|
0.000
0.000
0.000
0.000 0.001
[95% Conf. Interval]
3.361302
3.139615
.1417633
.0734956
.0391178
3.450091
3.248079
.2819344
.1480271
.1630572
Ben Jann 19 Detailed decomposition
Use the detail option to compute the individual contributions of the predictors to the components of the decomposition. detail specified without argument reports the contribution of each predictor individually. Alternatively, one can define groups of pre- dictors for which the results be subsumed in parentheses. Furthermore, one might apply the deviation contrast transform to dummy variable sets so that the contribution of a categorical predictor to the unexplained part of the decomposition does not depended on the choice of the base category. Example:
. tabulate isco, nofreq generate(isco)
. oaxaca lnwage educ exper tenure isco2-isco9, by(female) pooled ///
> detail(exp_ten: exper tenure, isco: isco?) categorical(isco?)
Blinder-Oaxaca decomposition
1: female = 0
2: female = 1
Coef.
Differential
Prediction_1 3.440222 Prediction_2 3.266761
Difference .1734607
Explained
educ .0395615
exp_ten .0399316 isco -.0056093 Total .0738838
Unexplained
educ -.1324971
exp_ten .0129955 isco -.0159367 _cons .2350152 Total .0995769
Selectivity bias adjustment
Number of obs = 1434
Robust
Std. Err. z
.0174589 197.05
.0218047 149.82
.0279331 6.21
.0097334 4.06
.0089081 4.48
.012445 -0.45
.017772 4.16
.1788045 -0.74
.0400811 0.32
.0296549 -0.54
.195018 1.21
.0266887 3.73
P>|z|
0.000
0.000
0.000
0.000
0.000
0.652
0.000
0.459
0.746
0.591
0.228
0.000
[95% Conf. Interval]
3.406003
3.224025
.118713
.0204843
.022472
-.0300009
.0390513
-.4829475
-.0655619
-.0740592
-.1472132
.047268
3.474441
3.309498
.2282085
.0586387
.0573911
.0187824
.1087163
.2179533
.0915529
.0421858
.6172435
.1518859
In labor market research it is common to include a correction for sample selection bias in the wage equations based on the procedure by Heckman (1976, 1979). Wages are only observed for people who are participating in the labor force and this might be a selective group. The most straightforward approach to account for selection bias in the decomposition is to deduct the selection effects from the overall differential and then apply the standard decomposition formulas to this adjusted differential (Reimers 1983; an alternative approach is followed by Dolton and Makepeace 1986; see Neuman and Oaxaca 2004 for an in-depth treatment of this issue).
If oaxaca is used with heckman, the decomposition is automatically adjusted for
20 A Stata implementation of the Blinder-Oaxaca decomposition selection. For example, the following command includes a selection correction in the
wage equation for women and decomposes the adjusted wage gap:
. oaxaca lnwage educ exper tenure, by(female) model2(heckman, twostep ///
> select(lfp = age agesq married single divorced kids6 kids714))
Blinder-Oaxaca decomposition
1: female = 0
2: female = 1
Coef.
Differential
Prediction_1 3.440222 Prediction_2 3.275643
Difference .164579
Decomposit~n
Endowments .0858436
Coefficients .0736812 Interaction .0050542
Std. Err. z
.0174874 196.73
.0281554 116.34
.0331442 4.97
.0157566 5.45
.031129 2.37
.0109895 0.46
Number of obs = 1434
P>|z| [95% Conf. Interval]
0.000 3.405947
0.000 3.220459
0.000 .0996176
0.000 .0549613
0.018 .0126695
0.646 -.0164849
3.474497
3.330827
.2295404
.116726
.134693
.0265932
Comparing the results to the output in the first example reveals that the uncorrected wages of women are slightly biased downwards (3.267 versus the selectivity corrected 3.276) and the wage gap is somewhat overestimated (0.173 versus the corrected 0.165).
Alternatively, it is sometimes sensible to compute the selection variables outside of oaxaca and then use the adjust() option to correct the differential. Example:
. probit lfp age agesq married single divorced kids6 kids714 if female==1
(output omitted )
. predict xb if e(sample), xb
(759 missing values generated)
. generate mills = normalden(-xb) / (1 – normal(-xb))
(759 missing values generated)
. replace mills = 0 if female==0
(759 real changes made)
. oaxaca lnwage educ exper tenure mills, by(female) adjust(mills)
Blinder-Oaxaca decomposition
1: female = 0
2: female = 1
Coef.
Differential
Prediction_1 3.440222 Prediction_2 3.266761
Difference .1734607 Adjusted .164579
Decomposit~n
Endowments .0858436
Coefficients .0736812 Interaction .0050542
Number of obs
= 1434
Std. Err. z
.0174874 196.73
.0218659 149.40
.0279987 6.20
.033215 4.95
.0157766 5.44
.0312044 2.36
.0110181 0.46
P>|z| [95% Conf. Interval]
0.000 3.405947
0.000 3.223905
0.000 .1185843
0.000 .0994788
0.000 .0549221
0.018 .0125217
0.646 -.0165409
3.474497
3.309618
.2283372
.2296792
.1167651
.1348407
.0266493
Ben Jann 21
Using oaxaca with nonstandard models
You can also use oaxaca, for example, with binary outcome variables and employ a command such as logit to estimate the models. You have to understand, however, that oaxaca will always apply the decomposition to the linear predictions from the models (based on the first equation if a model contains multiple equations). With logit models, for example, the decomposition computed by oaxaca is expressed in terms of log odds and not in terms of probabilities or proportions.11
5 References
Blinder, A. S. 1973. Wage Discrimination: Reduced Form and Structural Estimates. The Journal of Human Resources 8: 436–455.
Cotton, J. 1988. On the Decomposition of Wage Differentials. The Review of Economics and Statistics 70: 236–243.
Daymont, T. N., and P. J. Andrisani. 1984. Job Preferences, College Major, and the Gender Gap in Earnings. The Journal of Human Resources 19: 408–428.
Dolton, P. J., and G. H. Makepeace. 1986. Sample Selection and Male-Female Earnings Differentials in the Graduate Labour Market. Oxford Economic Papers 38: 317–341.
Fairlie, R. W. 2005. An extension of the Blinder-Oaxaca decomposition technique to logit and probit models. Journal of Economic and Social Measurement 30: 305–316.
Fortin, N. M. 2006. Greed, Altruism, and the Gender Wage Gap. University of British Columbia. Available from http://www.econ.ubc.ca/nfortin/Fortinat8.pdf.
Gardeazabal, J., and A. Ugidos. 2004. More on identification in detailed wage decom- positions. The Review of Economics and Statistics 86: 1034–1036.
Gomulka, J., and N. Stern. 1990. The Employment of Married Women in the United Kingdom 1970-83. Economica 57: 171–199.
Greene, W. H. 2003. Econometric Analysis. 5th ed. Upper Saddle River, NJ: Pearson Education.
Gujarati, D. N. 2003. Basic Econometrics. 4th ed. New York: McGraw-Hill.
Haisken-DeNew, J. P., and C. M. Schmidt. 1997. Inter-Industry and Inter-Regional Differentials: Mechanics and Interpretation. The Review of Economics and Statistics 79: 516–521.
11. Approaches to decompose differences in proportions are provided by, e. g., Gomulka and Stern (1990), Fairlie (2005), or Yun (2005a).
22 A Stata implementation of the Blinder-Oaxaca decomposition Hardy, M. A. 1993. Regression With Dummy Variables. Newbury Park, CA: Sage.
Heckman, J. J. 1976. The Common Structure of Statistical Models of Truncation, Sample Selection and Limited Dependent Variables and a Simple Estimator for Such Models. Annals of Econometrics and Social Measurement 5: 475–492.
———. 1979. Sample Selection Bias as a Specification Error. Econometrica 47: 153–161. Heinrichs, J., and P. Kennedy. 2007. A computational trick for calculating the Blinder-
Oaxaca decomposition and its standard error. Economics Bulletin 3(66): 1–7. Hendrickx, J. 1999. dm73: Using categorical variables in Stata. Stata Technical Bulletin
52: 2–8.
Horrace, W. C., and R. L. Oaxaca. 2001. Inter-Industry Wage Differentials and the Gender Wage Gap: An Identification Problem. Industial and Labor Relations Review 54: 611–618.
Jackson, J. D., and J. T. Lindley. 1989. Measuring the extent of wage discrimination: a statistical test and a caveat. Applied Economics 21: 515–540.
Jann, B. 2003. The Swiss Labor Market Survey 1998 (SLMS 98). Journal of Applied Social Science Studies 123: 329–335.
———. 2005a. devcon: Stata module to apply the deviation contrast transform to estimation results. Statistical Software Components S450603. Available from http://ideas.repec.org/c/boc/bocode/s450603.html.
———. 2005b. Standard Errors for the Blinder-Oaxaca Decomposi- tion. German Stata Users’ Group Meetings 2005. Available from http://repec.org/dsug2005/oaxaca se handout.pdf.
Jones, F. L. 1983. On Decomposing the Wage Gap: A Critical Comment on Blinder’s Method. The Journal of Human Resources 18: 126–130.
Jones, F. L., and J. Kelley. 1984. Decomposing Differences Between Groups. A Cau- tionary Note on Measuring Discrimination. Sociological Methods and Research 12: 323–343.
Kennedy, P. 1986. Interpreting Dummy Variables. The Review of Economics and Statistics 68: 174–175.
Lin, E. S. 2007. On the standard errors of Oaxaca-type decompositions for inter-industry gender wage differentials. Economics Bulletin 10(6): 1–11.
Mood, A. M., F. A. Graybill, and D. C. Boes. 1974. Introduction to the Theory of Statistics. New York: McGraw-Hill.
Neuman, S., and R. L. Oaxaca. 2004. Wage decompositions with selectivity-corrected wage equations: A methodological note. Journal of Economic Inequality 2: 3–10.
Ben Jann 23 Neumark, D. 1988. Employers’ Discriminatory Behavior and the Estimation of Wage
Discrimination. The Journal of Human Resources 23: 279–295.
Nielsen, H. S. 2000. Wage discrimination in Zambia: an extension of the Oaxaca-Blinder
decomposition. Applied Economics Letters 7: 405–408.
Oaxaca, R. 1973. Male-Female Wage Differentials in Urban Labor Markets. Interna-
tional Economic Review 14: 693–709.
Oaxaca, R. L., and M. Ransom. 1998. Calculation of approximate variances for wage
decomposition differentials. Journal of Economic and Social Measurement 24: 55–61. Oaxaca, R. L., and M. R. Ransom. 1994. On discrimination and the decomposition of
wage differentials. Journal of Econometrics 61: 5–21.
———. 1999. Identification in Detailed Wage Decompositions. The Review of Eco-
nomics and Statistics 81: 154–157.
O’Donnell, O., E. van Doorslaer, A. Wagstaff, and M. Lindelow. 2008. Analyzing Health Equity Using Household Survey Data. A Guide to Techniques and Their Implemen- tation. Washington, DC: The World Bank.
Polavieja, J. G. 2005. Task Specificity and the Gender Wage Gap. Theoretical Consid- erations and Empirical Analysis of the Spanish Survey on Wage Structure. European Sociological Review 21: 165–181.
Reimers, C. W. 1983. Labor Market Discrimination Against Hispanic and Black Men. The Review of Economics and Statistics 65: 570–579.
Shrestha, K., and C. Sakellariou. 1996. Wage discrimination: a statistical test. Applied Economics Letters 3: 649–651.
Silber, J., and M. Weber. 1999. Labor market discrimination: are there dignificant differences between the various decomposition procedures? Applied Economics 31: 359–365.
Stanley, T. D., and S. B. Jarrell. 1998. Gender Wage Discrimination Bias? A Meta- Regression Analysis. The Journal of Human Resources 33: 947–973.
Suits, D. B. 1984. Dummy Variables: Mechanics V. Interpretation. The Review of Economics and Statistics 66: 177–180.
Weesie, J. 1999. sg121: Seemingly unrelated estimation and the cluster-adjusted sand- wich estimator. Stata Technical Bulletin 52: 34–47.
Weichselbaumer, D., and R. Winter-Ebmer. 2005. A Meta-Analysis of the International Gender Wage Gap. Journal of Economic Surveys 19: 479–511.
Winsborough, H. H., and P. Dickinson. 1971. Components of Negro-White Income Differences. In Proceedings of the Social Statistics Section, 6–8. Washington, DC: American Statistical Association.
24 A Stata implementation of the Blinder-Oaxaca decomposition Yun, M.-S. 2005a. Hypothesis tests when decomposing differences in the first moment.
Journal of Economic and Social Measurement 30: 295–304.
———. 2005b. A Simple Solution to the Indentification Problem in Detailed Wage Decompositions. Economic Inquiry 43: 766–772.