How many predictors are in your model? My guess is LOTS. Correlations are a terrible way to look at collinearity because they only consider 2 variable relationships. My guess is that you have so many IVs in your model that you have Multicollinearity anyway. If this is the case, try to subset your research question into simpler questions where less complicated Models can be studied first and go from there. Since you seem to interested in prediction I would suggest trying an adaptive lasso approach with your data because adaptive lasso is resistant to multicollinearity in general. I am attaching a paper of ours which is under review that does this with what I expect are simpler models with a sample size of approx. 5000. See if this doesn't help. I would be happy to send you our R programs if you think that might help. Good luck, David Booth

Hi

Just some thoughts from me. I am really not sure if I understand this properly, as I aggree that this is somewhat difficult to understand. It is intuitive that the two models should produce the same conclusions. However, I would trust the multiple logistic regression results, as you analyze the net contribution of two main predictors together with the interaction term (I assume you only have these two predictor variables included in the model, as is shown in the output word file ). These statistical effects are adjusted for each other, and in additon one predictor is shown to vary over different levels of the the other predictor. The other bivariate results represent unadjusted effects. Besides, I have seen several articles warning against dichotomizing continous variables (f.ex in low and high level groups), as this gives decreased variance and loss of information. The differences could of course be due to several other factors which I am not aware of.

Best,

Rolf

Rolf Gjestad I was assuming all possible interactions rather than just 2-way. interactions. Intuition will not work here unless your intuition takes all possible model terms into account, which is what I was really trying to say. Best wishes, David Booth

David Eugene Booth Rolf Gjestad Thank you for your responses. I just wanted to clarify that the only variales I have entered into the model are "stressors", "hopelessness" & the interaction term "stressors X hopelessness". No other variales have been entered.

Well you really need to do more than that or your model could be really misspecified. All this comes from full factorial experimental designs. If you have only 1 interaction term in your model and have 42 other significant terms that you haven't included you are forcing all of that into the wrong place. Always start with a full model and then if you want to, reduce variables. That's where the terms full and reduced model come from. NEVER start from a reduced model and work up.

Best wishes, David Booth

@James a preprint just came across my screen that seems to apply to your question the link is ::

https://www.researchgate.net/publication/341817242_Sufficient_Dimension_Reduction_for_Interactions

Hopefully this will help. Best wishes, David Booth

David Eugene Booth Thank you for your responses on this. I really appreciate it.

James Knowles A few comments:

The main issue I see is that you have so many levels for each factor (10 x 5 = 50 cells). This makes for a very messy process when you want to compare individual cells. Furthermore, the n-size could vary considerably across cells. Perhaps condensing the levels a bit would help simplify things. For instance, stressors could be recoded into three levels: 1-3 (low), 4-6 (medium), 7-9 (high) instead of ten (0-9).

Is subdividing the sample in this way appropriate and necessary? Did you assess whether the subgroups are comparable? How do the residuals looks? For argument's sake, data for the 'high' hopelessness subgroup could be highly variable whereas data for the 'low' hopelessness subgroup could be more tightly clustered. (What's true of the entire sample is not necessarily true of the subgroups). This is why I'd exercise caution in subdividing my sample post-hoc. In some cases, the results will be counterintuitive and misleading because you're using different models: the same way an F-test can be significant and yet follow up t-tests for the same data can be non-significant.

Allowing for all that, recall that the coefficients in the logistic regression show the probability that a participant has suicidal thoughts as hopelessness increases. Looking at your output, I don't see where the coefficients for the subgroups came from: .7, .54, and 1.53? However, assuming your analyses are correct, all coefficients are positive and the 'high' hopelessness subgroup shows the largest coefficient (1.53). Is this not what you would expect?

Blaine Tomkins Exactly. I believe you are saying in ANOVA language what I said in regression language. Am I correct? David