2
$\begingroup$

I have data with 6 predictor variables and a response variable (default ). When I do a logistic regression using probit estimation, I get a result different from SAS.

(the link to my data https://drive.google.com/open?id=1TZvEsOpuzLQ_BvGBR26n5M-yzlp2yUWH)

equation <- as.formula(default ~ dlagR3+loglaginf6+dloglagImp6+loglagRR9+dlagFX6+laglogRGDP6 )

model    <- glm(equation , family = binomial('probit'), data=setDF(data)) 

But R is giving the following error:

 glm.fit: fitted probabilities numerically 0 or 1 occurred

Coefficients:
        Estimate Std. Error z value Pr(>|z|)    
  (Intercept) -5.13523    1.55705  -3.298 0.000974 ***
  dlagR3       0.04719    0.03696   1.277 0.201711    
  loglaginf6  -0.39395    0.05637  -6.989 2.76e-12 ***
  dloglagImp6  0.13151    0.02981   4.412 1.02e-05 ***
  loglagRR9    6.78246    0.09228  73.497  < 2e-16 ***
  dlagFX6      0.02099    0.01835   1.144 0.252535    
  laglogRGDP6 -1.52335    0.32653  -4.665 3.08e-06 ***

But when I use SAS or Eviews, this error is not thrown and they both agree on the coefficients. SAS Output

Eviews Output

So I tried implementing a lot of supposed solutions as given in a lot of forums but no matter what I do, R is giving the same coefficients. The results given by Eviews and SAS is the accurate one.

Note: SAS gives the probability of 0 (not 1) so I have created a new response variable default2 which is opposite of default

proc logistic data=data;
model default2 = dlagR3 loglaginf6 dloglagImp6 loglagRR9 dlagFX6 laglogRGDP6 
/link=probit;
run;
$\endgroup$
  • 1
    $\begingroup$ There might be complete or quasi-complete separation in your data, which will effect the maximum likelihood estimates. SAS must have a different way of dealing with that (I don't know much about the other software you use). $\endgroup$ – Demetri Pananos Nov 16 '18 at 6:09
  • $\begingroup$ @DemetriPananos That's fine. I tried using a lot of solutions to variable separation but they don't seem to be working. $\endgroup$ – Dom Jo Nov 16 '18 at 6:13
  • $\begingroup$ The thing is that no matter whichever alternate function i use, the coefficients are all same and they don't change $\endgroup$ – Dom Jo Nov 16 '18 at 6:14
  • $\begingroup$ See this answer. Seems like this person has the same problem. One answer talks about PROC LOGISTIC stats.stackexchange.com/questions/11109/… $\endgroup$ – Demetri Pananos Nov 16 '18 at 6:15
  • $\begingroup$ Out of curiosity, what happens when you remove some predictors from the model? Try regressing on just one or two variables. Do the results look the same between the three? $\endgroup$ – Demetri Pananos Nov 16 '18 at 6:19

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.