0
$\begingroup$

I am trying to test a model using hierarchical regression analysis. I have one predictor, one outcome and five moderators. My main question is should I put the moderators in the model using forward regression or backward regression so basically:

Option 1: Block 1: co-variables (method Enter). Block 2: predictor (method Enter) Block 3: predictor + 5 main effects (method Enter). Block 4: 5 interaction terms (so centralized main effect x predictor) (method Forward)

Option 2: Option 1: block 1: co-variables (method Enter). Block 2: predictor (method Enter) Block 3: predictor + 5 main effects (method Enter). Block 4: 5 interaction terms (so centralized main effect x predictor) (method Backward)

I tried both ways and they lead to different findings. With option 1 I do find some moderating effects. With option 2, there are no moderating effects. I know there is literature on forward and backward regression but with this many moderators in one model I am lost...

$\endgroup$
0
$\begingroup$

Forward and backward model selection do not result in the same model. "Successful" entry of covariates in a stepwise model does not mean that they are moderating effects. Not surprisingly, the forward model tends to pick fewer variables, but this is not always the case as in your example.

Read the documentation for your software. Be sure the significance level for entering the model is the same as the significance level for staying in the model.

The main issue is that a "moderating effect" is a causal designation, and stepwise model selection is a predictive procedure, so I wouldn't trust the findings from either model. Rather, fit one or more confirmatory models according to the direct acyclic graph (DAG) that describes the proposed causal mechanism, and describe the inference accordingly.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.