What’s this about?

Sample selection arises when the sampled data are not representative of the population of interest. A classic example of sample selection is women’s work participation. Suppose that we want to model the wages of women. If we consider only the sample of women who chose to work, we may end up with a sample in which the wages are too high because women who would have low wages may have chosen not to work. Of course, if the decision whether to work is random, there would be no problem with using only the sample of women who work. This is not a realistic assumption in this case. To obtain valid inference in this example, we must model the outcome, the wages, and the decision to work. We will refer to the two models as the outcome model and the participation model.

In Stata, you can use heckman to fit a Heckman selection model to continuous outcomes, heckprobit to fit a probit sample-selection model to binary outcomes, and heckoprobit to fit an ordered probit model with sample selection to ordinal outcomes. You can now simply prefix these commands with bayes: to fit the corresponding Bayesian sample-selection models.

 

Let’s see it work

Continuing with our example of women’s work participation, we first fit the classical Heckman sample-selection model. Below we model both the wages and the decision to work based on the level of education and age. For the decision to work, we additionally include marriage status and number of children.

 

. heckman wage educ age, select(married children educ age)


Heckman selection model                         Number of obs     =      2,000
(regression model with sample selection)              Selected    =      1,343
                                                      Nonselected =        657

                                                Wald chi2(2)      =     508.44
Log likelihood = -5178.304                      Prob > chi2       =     0.0000

wage Coef. Std. Err. z P>|z| [95% Conf. Interval]
wage5
education .9899537 .0532565 18.59 0.000 .8855729 1.094334
age .2131294 .0206031 10.34 0.000 .1727481 .2535108
_cons .4857752 1.077037 0.45 0.652 -1.625179 2.59673
select
married .4451721 .0673954 6.61 0.000 .3130794 .5772647
children .4387068 .0277828 15.79 0.000 .3842534 .4931601
education .0557318 .0107349 5.19 0.000 .0346917 .0767718
age .0365098 .0041533 8.79 0.000 .0283694 .0446502
_cons -2.491015 .1893402 -13.16 0.000 -2.862115 -2.119915
/athrho .8742086 .1014225 8.62 0.000 .6754241 1.072993
/lnsigma 1.792559 .027598 64.95 0.000 1.738468 1.84665
rho .7035061 .0512264 .5885365 .7905862
sigma 6.004797 .1657202 5.68862 6.338548
lambda 4.224412 .3992265 3.441942 5.006881
LR test of indep. eqns. (rho = 0):   chi2(1) =    61.20   Prob > chi2 = 0.0000

To fit its Bayesian analog, we use bayes: heckman.

 

. bayes: heckman wage educ age, select(married children educ age)

Model summary
Likelihood:
wage ~ heckman(xb_wage,xb_select,{athrho} {lnsigma})
Priors:
{wage:education age _cons} ~ normal(0,10000) (1)
{select:married children education age _cons} ~ normal(0,10000) (2)
{athrho lnsigma} ~ normal(0,10000)
(1) Parameters are elements of the linear form xb_wage. (2) Parameters are elements of the linear form xb_select.
Bayesian Heckman selection model                MCMC iterations   =     12,500
Random-walk Metropolis-Hastings sampling        Burn-in           =      2,500
                                                MCMC sample size  =     10,000
                                                Number of obs     =      2,000
                                                      Selected    =      1,343
                                                      Nonselected =        657
                                                Acceptance rate   =      .3484
                                                Efficiency:   min =     .02314
                                                              avg =     .03657
Log marginal likelihood = -5260.2024                          max =     .05013

Equal-tailed
Mean Std. Dev. MCSE Median [95% Cred. Interval]
wage
education .9919131 .051865 .002609 .9931531 .8884407 1.090137
age .2131372 .0209631 .001071 .2132548 .1720535 .2550835
_cons .4696264 1.089225 .0716 .4406188 -1.612032 2.65116
select
married .4461775 .0681721 .003045 .4456493 .3178532 .5785857
children .4401305 .0255465 .001156 .4402145 .3911135 .4903804
education .0559983 .0104231 .000484 .0556755 .0360289 .076662
age .0364752 .0042497 .000248 .0362858 .0280584 .0449843
_cons -2.494424 .18976 .011327 -2.498414 -2.861266 -2.114334
athrho .868392 .099374 .005961 .8699977 .6785641 1.062718
lnsigma 1.793428 .0269513 .001457 1.793226 1.740569 1.846779
Note: Default priors are used for model parameters.

Unlike heckman, bayes: heckman reports the ancillary parameters only in the estimation metric. We can use bayesstats summary to obtain the parameters in the original metric.

 

. bayesstats summary (rho:1-2/(exp(2*{athrho})+1)) (sigma:exp({lnsigma}))


Posterior summary statistics                      MCMC sample size =    10,000

         rho : 1-2/(exp(2*{athrho})+1)
       sigma : exp({lnsigma})

Equal-tailed
Mean Std. Dev. MCSE Median [95% Cred. Interval]
rho .6970522 .0510145 .003071 .701373 .5905851 .7867018
sigma 6.012205 .1621422 .008761 6.008807 5.700587 6.339366

Parameter rho is a correlation coefficient that measures the dependence between the outcome and participation models. If rho is zero, the two models are independent and can be analyzed separately. In other words, there is no sample selection, and we can model the wages using only the sample of women who work without introducing any bias in our results. In our example, rho is estimated to be between 0.59 and 0.79 with a probability of 0.95, so the decision to work is related to the wages in this example.

We can test for sample selection formally by using, for example, Bayes factors. A Bayes factor of two models is simply the ratio of their marginal likelihoods. The larger the value of the marginal likelihood, the better the model fits the data. To test for sample selection, we can compare the marginal likelihoods of the current model and of the model with rho equal to zero.

First, we store the current Bayesian estimation results from the sample-selection model.

 

. bayes, saving(heckman_mcmc)


. estimates store heckman

Next, we fit a model that assumes no sample selection. When rho equals zero, {athrho} also equals zero. So we specify a strong prior saturated at zero for parameter {athrho}.

 

. bayes, prior({athrho}, normal(0,1e-4)) saving(nosel_mcmc):
  heckman wage educ age, select(married children educ age)

Model summary
Likelihood:
wage ~ heckman(xb_wage,xb_select,{athrho} {lnsigma})
Priors:
{wage:education age _cons} ~ normal(0,10000) (1)
{select:married children education age _cons} ~ normal(0,10000) (2)
{athrho} ~ normal(0,1e-4)
{lnsigma} ~ normal(0,10000)
(1) Parameters are elements of the linear form xb_wage. (2) Parameters are elements of the linear form xb_select.
Bayesian Heckman selection model                MCMC iterations   =     12,500
Random-walk Metropolis-Hastings sampling        Burn-in           =      2,500
                                                MCMC sample size  =     10,000
                                                Number of obs     =      2,000
                                                      Selected    =      1,343
                                                      Nonselected =        657
                                                Acceptance rate   =      .3065
                                                Efficiency:   min =     .03943
                                                              avg =     .09498
Log marginal likelihood = -5283.0246                          max =      .2432

Equal-tailed
Mean Std. Dev. MCSE Median [95% Cred. Interval]
wage
education .8981219 .0509913 .001578 .8973616 .8013416 1.000497
age .1477784 .01854 .00066 .1477496 .1115628 .1850257
_cons 5.994764 .890318 .030657 6.014622 4.150738 7.658942
select
married .4351031 .0748102 .003577 .4377313 .2821176 .5752786
children .4501657 .0285028 .001045 .4492015 .3937091 .5048498
education .0584037 .0110582 .000524 .0579573 .0370387 .0814287
age .034779 .0043677 .00022 .0348894 .0259916 .043139
_cons -2.47607 .1962162 .009818 -2.467739 -2.862694 -2.10733
athrho .0062804 .010209 .00023 .0062746 -.014139 .0261746
lnsigma 1.69586 .019056 .000386 1.695649 1.65948 1.734115
Note: Default priors are used for some model parameters.

. estimates store nosel

We now use bayesstats ic to obtain the Bayes factor of the two models.

 

. bayesstats ic heckman nosel

Bayesian information criteria

DIC log(ML) log(BF)
heckman 10376.05 -5260.202 .
nosel 10435.29 -5283.025 -22.82221
Note: Marginal likelihood (ML) is computed
using Laplace-Metropolis approximation.

 

The value of the log-Bayes factor of -23 indicates a very strong preference for the sample-selection model heckman and thus for the presence of sample selection in these data.

Tell me more

Learn more about the general features of the bayes prefix.

Learn more about Stata’s Bayesian analysis features.

Read more about the bayes prefix and Bayesian analysis in the Stata Bayesian Analysis Reference Manual.