Interaction terms in fixed effect model

[Pranshu Tripathi]

I am working on panel data. My regression model includes a three-way interaction term, so I am using a fixed effects model. Three-way interaction consists of two dummies: treatment (pre and post-treatment) and health of a firm (distress and non-distress). The three-way interaction has a continuous variable with these two dummies.

1. We cannot use the VIF command to check multicollinearity after running panel data regression in the Stata. So I am using the OLS model with all the interactions (including the lower-order interactions) to check for multicollinearity. I have found a higher VIF for two-way interaction. So, I dropped this interaction based on a high correlation with the three-way interaction.
2. BP LM test suggests a random effect model, and the Hausman test suggests a fixed effects model (I have included all the interactions except the dropped one).
2. For heteroskedasticity and autocorrelation, I used the modified Wald and Woolridge tests, respectively.

Are these procedures right? Can I use interaction terms in the fixed effect model?

[Maarten Buis]

You can use interactions in a fixed effects model, but you have to include all lower level interactions. This is not optional. High multicollinearity is expected and not a problem.

[Pranshu Tripathi]

Thanks, Maarten

[Pranshu Tripathi]

Hey, Maarten, when I include the less relevant lower-order interaction, the result is insignificant, whereas after excluding it, the result is significant.