Assessing Additive interaction for binary response using linear probability

[Kaeslock223_]

I am trying to figure out what distribution would be the most appropriate. My professor said to model the log regression as a linear probability to assess for additive scale. I didn't find statically significant interaction when using link=identity distr=bin, but when I stratified by the suspected effect modifier, I could see a difference. When I went back and changed the distr=bin to distr=normal, it displayed significant interaction. She said I could do it, but this doesn't make sense to me because the response variable is binomial. Could someone explain?

[Rick_SAS]

Please post both versions of the code.