Logistic regression, Firth correction, Main Effects and Interaction Are Significant

[marcel]

Codes are attached in a RTF file.

Logistic regression with Firth Correction

I run a stepwise regression analysis in which I included several independent variables and their interactions. The stepwise regression selected a model with two qualitative factors and their interaction, in which ONLY the one factor (Donor) was NOT significant; the output also showed a warning about "quasi-complete separation".

I rerun the analysis with only the selected variables, by including the Firth correction in the new syntax. The output of this run shows that ALL variables are significant (each factor and their interaction). I was confused, because the stepwise indicated that only the interaction is significant.

I rerun the analysis with NO Firth correction and NO selection, and again, the output shows that ONLY one factor (Donor) was NOT significant.

My questions are:

1. I am concerned that something may be wrong because the significance of the variables is different when running a stepwise/no selection logistic regression versus Firth. Any of your kind clarifications on this subject would be highly appreciated.

2. Also, I would like to get some advice on how to report the Firth corrected results (if those results are correct). I am referring to the fact that the main factors are significant as well as the interaction. I learned that when you have a significant interaction, you focus on the interaction and leave the main factors alone. Is this a good approach?

regards,

Marcel

[Ksharp]

FIRTH option is suited for small size sample.

And selection=stepwise is not a good way to select variables . Check Cross-Validate or LASSO method of PROC HPGENSELECT .

For your second question, I don't agree with you. if interaction effect is significant , you can make a slice to see the change in each levels of main effect by LSMEANS  statement. 

[marcel]

Mr. Ksharp,

Thank you for your answer. Size-wise, my sample is fine (compared to the examples in this note). I will take a look at LASSO.

Isn't the SLICE option used to explore interactions?

Regards,

Marcel