topic Re: how to evaluate the model fit for logistic regression in Statistical Procedures

how to evaluate the model fit for logistic regression

dart926 — Fri, 12 Sep 2014 18:07:00 GMT

Hi all,

I am trying to build a logistic regression model: using 7 variables (see below) to predict college enrollment (Enroll vs. Not-Enroll). When I put all seven variables in the model, the Hosmer and Lemeshow Goodness-of-Fit Test is significant, I think it suggests that the model does not fit the data well. I then tried variable selection and allowed two-way interactions to enter the selection, some of the interaction terms were selected for the final model, and the Hosmer and Lemeshow Goodness-of-Fit Test is still significant.

I also tried variable selection without including interaction terms. Four out of the seven variables were selected for the final model (race, sat score, legacy and year) and the Hosmer and Lemeshow Goodness-of-Fit p-value is significant as well. Does any one have suggestions on what I should do next?

A general question: Is the Hosmer and Lemeshow Goodness-of-Fit Test a good way to evaluate model fit? What do you usually use to evaluate model fit for logistic regression?

Any suggestion is appreciated.

Yanmin

Gender (Female vs. Male)
Race (Non-US vs. Minority vs. White)
SAT scores (numerical variable)
Academic Interest (Sciences vs. Interdisciplinary vs. Humanities vs. Undecided vs. Social Sciences)
Legacy (Yes vs. No)
First generation (Yes vs. No)
Year(2012 vs. 2013 vs. 2014)

Re: how to evaluate the model fit for logistic regression

stat_sas — Fri, 12 Sep 2014 19:06:27 GMT

Hi,

Did you compare default model fit statistics before considering Hosmer and Lemeshow?

Re: how to evaluate the model fit for logistic regression

dart926 — Fri, 12 Sep 2014 19:14:34 GMT

Hi there,

Thanks so much for responding. My SAS code is below.

proc logistic data=asq;

class gender (ref='Male') race (ref='White') Aca_Ins (ref='Social Sciences') legacy (ref='No') first_gen(ref='No') year(ref='2014') /param=ref;

model enroll (event='Enroll')= gender_r race SAT_sum Aca_Ins legacy first_gen year/lackfit ;

run;

for the default default model fit statistics, do you mean the following table? what should I compare to? Thank you!!

AIC	4018.729	3781.145
SC	4024.708	3858.872
-2 Log L	4016.729	3755.145

Re: how to evaluate the model fit for logistic regression

stat_sas — Fri, 12 Sep 2014 19:21:24 GMT

Yes, what about Testing Global Null Hypothesis: BETA=0?

Re: how to evaluate the model fit for logistic regression

dart926 — Fri, 12 Sep 2014 20:25:37 GMT

Thank you!!!

I will read a bit more about logistic regression before continuing the analysis. Will come back to you later.

Re: how to evaluate the model fit for logistic regression

Reeza — Fri, 12 Sep 2014 21:41:02 GMT

SAS Annotated Output: Proc Logistic

SAS Data Analysis Examples: Logit Regression

Re: how to evaluate the model fit for logistic regression

dart926 — Fri, 12 Sep 2014 21:48:49 GMT

Thank you so much Reeza! This is very helpful!