Help needed...Conflicts on logit model with binary regressors!

[Ngozi ADELEYE]

Dear all,

Please I will appreciate any clarification regarding the SPECIFICATION and INTERPRETATION of logit model with BINARY regressors. I have a logit model with 2 regressors (REPUTATION and RD) that are binary coded 0 and 1.

These are the two model specifications":

(1) logit innovation employees exports reputation RD reputation#RD, r

(2) logit innovation employees exports i.reputation i.RD i.reputation#i.RD, r

...and these are the results:

**(1)
logit innovation employees exports reputation RD reputation#RD, r

Logistic regression                             Number of obs     =      2,476
                                                Wald chi2(5)      =     173.43
                                                Prob > chi2       =     0.0000
Log pseudolikelihood = -1403.7525               Pseudo R2         =     0.0613

-------------------------------------------------------------------------------
              |               Robust
   innovation |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
--------------+----------------------------------------------------------------
    employees |  -.0039655   .0013319    -2.98   0.003    -.0065761   -.0013549
      exports |   .0055913   .0031888     1.75   0.080    -.0006586    .0118413
   reputation |   .5641152   .1202647     4.69   0.000     .3284008    .7998296
           RD |   .9932076   .1121152     8.86   0.000     .7734658    1.212949
              |
reputation#RD |
         0 1  |  -.1586253   .2306959    -0.69   0.492     -.610781    .2935304
         1 0  |          0  (omitted)
         1 1  |          0  (omitted)
              |
        _cons |  -1.509243     .09378   -16.09   0.000    -1.693049   -1.325438
-------------------------------------------------------------------------------

**(2)
logit innovation employees exports i.reputation i.RD i.reputation#i.RD, r

Logistic regression                             Number of obs     =      2,476
                                                Wald chi2(5)      =     173.43
                                                Prob > chi2       =     0.0000
Log pseudolikelihood = -1403.7525               Pseudo R2         =     0.0613

-------------------------------------------------------------------------------
              |               Robust
   innovation |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
--------------+----------------------------------------------------------------
    employees |  -.0039655   .0013319    -2.98   0.003    -.0065761   -.0013549
      exports |   .0055913   .0031888     1.75   0.080    -.0006586    .0118413
 1.reputation |   .5641152   .1202647     4.69   0.000     .3284008    .7998296
         1.RD |   .8345823   .2018024     4.14   0.000     .4390569    1.230108
              |
reputation#RD |
         1 1  |   .1586253   .2306959     0.69   0.492    -.2935304     .610781
              |
        _cons |  -1.509243     .09378   -16.09   0.000    -1.693049   -1.325438
-------------------------------------------------------------------------------

Queries:
(1) Both produced almost identical results except for RD. Please why is this so since RD is a dummy variable?

(2) The coefficient of the interaction term (reputation#RD) is the same in both results but the appearance differs. While 1 1 is dropped in the 1st Table, it reflects in the 2nd Table. Please why is this so?

I know it's got to do with Stata programming but any constructive contribution will be greatly appreciated.

Thanks in advance,
Ngozi

[Carlo Lazzaro]

Ngozi:
the difference seems to rest on the fact that, in your first code, Stata reads -RD- as a continuous variable.
You can delve into this issue comparing the -predict p- resulting from the two codes.

[Ngozi ADELEYE]

Ngozi:
the difference seems to rest on the fact that, in your first code, Stata reads -RD- as a continuous variable.
You can delve into this issue comparing the -predict p- resulting from the two codes.

Thanks so much, Carlo.
You are correct. Both yielded the same predicted values.
Not surprising since both "reputation" and "RD" are 0, 1 dummy variables.
Once again, thanks for the constructive feedback, appreciated!
Ngozi