Hi All,

I run the following command

fracglm pass2 location##year location##subject, link(logit) vce(cluster subject)

and get the following estimate for location

Code:
----------------------------------------------------------------------------------
| Robust
pass      |   Coef.     Std. Err.    z      P>|z|     [95% Conf. Interval]
-----------------+----------------------------------------------------------------
location |
         SI | .9647009  .3088148   3.12    0.002       .3594351 1.569967


When I use the heteroskedastic option

fracglm pass location##year location##subject, link(logit) vce(cluster subject) het(i.location)

I get the following estimate for location


Code:
                 |               Robust
            pass |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-----------------+----------------------------------------------------------------
pass             |
        location |
             SI  |  -.7428275   .1674596    -4.44   0.000    -1.071042   -.4146127
                 |
- the coefficient has changed in direction.

However when I run the margins command for the two regressions

margins, dydx(location year)

the estimates for location are very similar to each other.
The diverge at about the 4th or 5th decimal place. Similarly for the estimates for year.

Is anyone able to explain the change in sign of coefficients for location yet almost identical margin effects?

Thank you,

Don