Incorrect Probit estimates

[Ferdinand Berr]

Hello,

I'm currently trying to replicate a Probit model from a research paper. I tried to rebuild the data set used and then ran the profit command for the baseline specification without controls. My dependent variable "Effort" is a dummy, the explanatory variable "Dresden" is a dummy too, and "Bezirk" is a categorical variable with 8 different values used for clustering the error term.

Here is an excerpt of my data:

* Example generated by -dataex-. To install: ssc install dataex
clear
input float(Effort Dresden) byte Bezirk
1 1 6
1 1 6
1 1 6
0 1 6
1 1 6
1 1 6
0 1 6
0 1 6
1 1 6
0 1 6
0 1 6
1 1 6
0 1 6
0 1 6
0 1 6
1 1 6
1 1 6
1 1 6
0 1 6
0 1 6
end
label values Bezirk u88_0600
label def u88_0600 6 "Dresden", modify

And here the profit command incl. results:

probit Effort Dresden, vce(cluster Bezirk)

Iteration 0:   log pseudolikelihood = -2297.4395  
Iteration 1:   log pseudolikelihood = -2292.1496  
Iteration 2:   log pseudolikelihood = -2292.1495  

Probit regression                               Number of obs     =      3,381
                                                Wald chi2(0)      =          .
                                                Prob > chi2       =          .
Log pseudolikelihood = -2292.1495               Pseudo R2         =     0.0023

                                 (Std. Err. adjusted for 8 clusters in Bezirk)
------------------------------------------------------------------------------
             |               Robust
      Effort |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
     Dresden |  -.1705294   .0607965    -2.80   0.005    -.2896884   -.0513704
       _cons |    .245526   .0607965     4.04   0.000      .126367    .3646851
------------------------------------------------------------------------------

The coefficient in the paper I try to replicate is around - 0.07. When I estimate a model with many control variables, all of them have the same signs and similar significance as in my reference paper, but the magnitude of coefficients is "scaled" by a factor of around 2.5 for all as seen in my baseline result here.

I guess I must have made some basic mistake with organizing my data or running the probit in Stata, I'd be very grateful for any help!

[Shuhong Peng]

I can't run your excerpted data. Stata displayed "note: Dresden omitted because of collinearity". I don't your full sample. Maybe your could check your data such as unit of measure.

[Ferdinand Berr]

You are right, in the data excerpt all individuals are from the city of Dresden, which is one of the 8 districts ("Bezirk") in my data, therefore the collinearity.

One thing which is weird to me: When I use the predict function after estimating the model, and then display the fitted values for "Dresden" = 1 and "Dresden" = 0 separately, the means of fitted values actually look good; that is, they are different by around -0.07, which would be the expected estimate. But the coefficient I got is around -0.17 as shown above.

[Maarten Buis]

That suggest that what is reported in the paper you are trying to replicate are not the probit coefficients but the marginal effects. This should be clearly mentioned in the text, can you check that?

[Ferdinand Berr]

That suggest that what is reported in the paper you are trying to replicate are not the probit coefficients but the marginal effects. This should be clearly mentioned in the text, can you check that?

That's it, thank you a lot! It is mentioned in the table notes.