Search
   >> Home >> Products >> Stata 13 >> Endogenous treatment effects

Endogenous treatment effects


Highlights

  • Relaxes conditional independence
    • Correlated unobservables that affect treatment and outcome
    • Selection on unobservables
  • Estimators
    • Linear regression
    • Poisson regression
  • Robust SEs to relax distributional assumptions
  • Cluster–robust SEs to allow for correlated data
  • Predictions
    • Potential-outcome means
    • Observed-outcome means
    • Conditional treatment effects
    • Probability of count and count-ranges for Poisson

Show me

Treatment effects measure the causal effect of a treatment on an outcome. Sometimes, we do not have conditional independence; that is, unobserved variables affecting both treatment and outcome. Endogenous treatment estimators address such cases.

etregress handles continuous outcome variables.

etpoisson handles count outcomes.

Suppose we are studying acne vulgaris in adults. In the data, members of the treated group have a specific genetic marker. That marker is itself determined by covariates, say, x1, x2, and x4. The outcome variable we observe is the number of pimples, and it is determined by x1 and x3. We are also concerned that there are unobserved variables that could cause the presence of the marker and affect the number of pimples.

We fit an endogenous treatment-effects model by typing

. etpoisson pustules x1 x3, treat(marker = x1 x2 x4)

Poisson regression with endogenous treatment      Number of obs   =       1000
(24 quadrature points)                            Wald chi2(3)    =      81.50
Log likelihood = -2297.3045                       Prob > chi2     =     0.0000

Coef. Std. Err. z P>|z| [95% Conf. Interval]
pustules
x1 .0509081 .0442439 1.15 0.250 -.0358084 .1376246
x3 .1281427 .0313078 4.09 0.000 .0667806 .1895049
1.marker 1.329508 .1994485 6.67 0.000 .9385965 1.72042
_cons -.6982827 .1139443 -6.13 0.000 -.9216094 -.4749559
marker
x1 .2638588 .0469415 5.62 0.000 .1718552 .3558624
x2 -.1921751 .048748 -3.94 0.000 -.2877194 -.0966308
x4 .4720434 .0465768 10.13 0.000 .3807546 .5633322
_cons -.4946407 .1121999 -4.41 0.000 -.7145484 -.2747329
/athrho .1350499 .1706788 0.79 0.429 -.1994743 .4695742
/lnsigma -.285167 .0516515 -5.52 0.000 -.386402 -.183932
rho .1342348 .1676033 -.1968701 .4378552
sigma .7518886 .0388361 .6794973 .8319923
Wald test of indep. eqns. (rho = 0): chi2(1) = 0.63 Prob > chi2 = 0.4288

The output reveals no evidence of the correlation that would violate conditional independence; the point estimate of the correlation is 0.1342, and the test against 0 is insignificant.

We can obtain the average treatment effect (ATE) based on these estimates by typing

. margins r.marker

Contrasts of predictive margins
Model VCE    : OIM

Expression   : Potential-outcome mean, predict()

df chi2 P>chi2
marker 1 28.30 0.0000
Delta-method
Contrast Std. Err. [95% Conf. Interval]
marker
(1 vs 0) 2.584824 .485849 1.632577 3.53707

The presence of the marker adds on average 2.58 pustules.

Show me more

Read much more about endogenous treatment in the manuals at etregress and etpoisson.

Back to highlights

See New in Stata 13 for more about what was added in Stata 13.

The Stata Blog: Not Elsewhere Classified Find us on Facebook Follow us on Twitter LinkedIn Google+ Watch us on YouTube