# Interaction Suppresses Main Effects in Repeated Measures GLM? [duplicate]

Possible Duplicate:
What if interaction wipes out my direct effects in regression?

I have three hypotheses within my experiment, and interpreting the interaction in relation to the rest of the results is what I am having difficulty with.

1. The first is that the IV in the model (with two levels) will be related to one DV but not the other (time 1, but not time 2).

A repeated measures GLM indicated that the IV was related to the both DVs.

2. The second hypothesis was that a covariate will be related to one DV and not the other.

Using a repeated measures GLM this hypothesis was supported.

3. Predicted an interaction between the IV, DV and a covariate. (A repeated measures GLM with IV (fixed factor), DVs (within factor), covariate).

What I found was no effect of the covariate, no effect of the interaction, however my main effect for IV and DVs was gone (p=0.075) (hypothesis one).

Now I know p=0.05 isn't everything, so I have two questions.

How would I go about interpreting the fact that my main effect for hypothesis one disappeared IF:

1. you don't consider p=0.05 a black and white cut off point?
2. you do consider p=0.05 a black and white cut off point?

Does it change how I interpret the results from hypothesis one, or their meaning? Does it make it redundant?

• I notice you used the generalized-linear-model tag, but you don't say much about the nature of your DV. Is is binary? Or did you mean the general linear model? Sep 27, 2012 at 22:20
• This seems to be a very common question. You may also want to look at these related Q's on CV: what-if-interaction-wipes-out-my-direct-effects-in-regression?; interaction-suppresses-the-main-effect-how-to-interpret-it; & this is also related (albeit from the other direction): how-to-interpret-inconsistent-beta-values-in-different-steps-of-hierarchical-reg. Sep 27, 2012 at 22:40
• yes i already looked at two of those answers. No sorry thats a mistake, i meant the general linear model... Sep 28, 2012 at 0:06
• No problem, @Rebecca-Anderson, that's one of the issues w/ those; the names are practically identical & easy to confuse. Re the other answers, it might help if you could state what you learned towards answering your questions there, & what wasn't covered / you still don't understand. Sep 28, 2012 at 0:13

Second, certainly adding an interaction can change a main effect - it can make it larger or smaller. When you have an interaction in the model, say $X_1*X_2$, then the main effect for $X_1$ is the effect of change in $X_1$ when $X_2 = 0$ (similarly, the main effect for $X_2$ is the effect of change in $X_2$ when $X_1 = 0$).
• If the interaction is not significant it simply means that, if, in the population from which the sample was drawn the effect there was no difference in the effect for men and women (e.g) then there is a $> 0.05$ chance of getting an interaction term as large as the one you got or larger. Depending on N, large effects can be nonsignificant and small effects can be significant. Sep 27, 2012 at 11:14