[SPSS - BLR] Adding interaction term leads to multicollinearity


I’m re-analysing survival data from another author

Data: 2 Treatments (no controls); 4 Temperatures; Event = Death on day 12 post-inoculation with treatment
Treatment 1:
25C = 100% mortality (60/60 deaths)
30C = 100% mortality (60/60 deaths)
35C = 68% mortality (41/60 deaths)
40C = 22% mortality (13/60 deaths)
Treatment 2:
25C = 100% mortality (60/60 deaths)
30C = 100% mortality (60/60 deaths)
35C = 100% mortality (60/60 deaths)
40C = 42% mortality (25/60 deaths)

Arranged in long format in SPSS, column for Treatment, Temperature, Event (1=dead, 0 = alive) for each individual subject (480 in total)

Output required: I need accurate coefficients to plot the probability of death as a function of temperature (range 25-45C).

I run the BLR in SPSS with the main effects only (Treatment as a categorical covariate and temperature (scale predictor) as a covariate). The output makes sense (both predictors are significant) and the plot looks reasonable. However, my main concern is that an interaction term is important. I run the BLR with the interaction term only and again the output is reasonable (interaction term is significant). Note the plots from the main effects only model and the interaction term only model are the exact same (by and large). Any idea why this is?
(Note: this is the same situation when I run complementary log-log in GLM)

When I include the main effects (treatment and temperature) and an interaction term (treatment*temperature) the model goes haywire (multicollinearity). I ran collinearity diagnostics – Main effects only VIF =1; Interaction term only VIF = 1; Full model- Treatment VIF = 34.8, Temperature VIF = 10.0, Treatment*Temperature = 43.8.

My feeling is to use the main effects only model for parsimony and interpretability of coefficients but my concern is that I have missed something. I’ve tried centering to remove multicollinearity but this doesn’t work. Is there a way to have a full model – main effects and interaction term? Or is my data set just not complex enough for that? Would PLS be more appropriate? Can probability be estimated from PLS coefficients? Any thoughts on what is happening here would be very gratefully appreciated.

Thank you very much,