Stata Technical Bulletin
19
Example: Robust cluster seemingly unrelated probit regression
In this example, we use data collected as part of the NLSY study (Center for Human Resource Research 1989). We would
like to model whether a person is part of a union by their age, race, the log of wages earned, and whether they live in the
south. We would like to simultaneously model whether a person is a college graduate by the log of wages earned, age, race,
and whether they are part of the SMSA.
Since we have data collected according to occupation codes, we would like to obtain standard errors that are robust to
heteroscedasticity taking into account our clusters on occupation.
. eq union age race ln-wage south
. eq coIlgrad ln-wage age race not-smsa
. suprob union coilgrad, robust cluster(occ-code)
Fitting constant only model
(output omitted )
Iteration 320: Log Likelihood = -3361.815
Fitting full model
(output omitted )
Iteration 3: Log Likelihood = -2204.3038
Seemingly unrelated probit regression Number of obs = 1819
Model chi2(8) = 2315.02
Prob > chi2 = 0.0000
Log Likelihood = -2204.3037683 Pseudo R2 = 0.3443
(standard errors adjusted for clustering on occ-code)
— |
I |
Coef. |
Robust Std. Err. |
z |
P>∣z∣ |
[957. Conf. |
— Interval] |
— union age „cons |
I I I |
-.0188275 .2544691 .0377699 -.3085786 -.2690615 |
.0077825 |
-2.419 3.721 0.235 -2.536 -1.319 |
0.016 0.000 0.814 0.011 0.187 |
-.0340808 .1204186 -.2775495 -.5470361 -.6687357 |
— -.0035742 |
— coIlgrad not-smsa |
I I I |
.7299636 -.0601856 -.7237232 .3649792 .3998985 |
.2924354 .129846 |
2.496 -4.610 -5.574 2.479 0.888 |
0.013 0.000 0.000 0.013 0.374 |
.1568008 -.0857711 -.9782167 .0764424 -.4825412 |
— 1.303126 -.0346 -.4692296 .653516 1.282338 |
— rho „cons |
I |
.0632956 |
.0682244 |
0.928 |
0.354 |
-.0704218 |
— .197013 |
Note in the above that the constant-only model took 320 iterations! The likelihood for the constant only model is very flat
so that this is not unusual. If you are willing to see only the output of the full model without the likelihood ration and pseudo
Tf2 of the model, you can use the nochi option to bypass fitting the constant only model.
Example: Nested probit regression
In this example, we will simulate nested data by altering the data from Pindyck and Rubinfeld (1981). We altered the data
using
. tab priv vote
I vote
priv I |
0 |
1 I |
Total |
— | |||
0 I |
24 |
46 I |
70 |
1 I |
5 |
5 I |
10 |
— | |||
Total I |
29 |
51 I |
80 |
. replace priv = . if vote==0
(29 real changes made, 29 to missing)