My experiment had a 2×3 design with one covariate. If I analyse the results using an ANOVA, I get a strongly significant interaction between the two main factors (p<.001). If I add the covariate into the analysis and do an ANCOVA, the interaction between the two main factors is not significant (p=.252).
The covariate itself was not significant, nor did it have any significant interactions.
There are similar questions on this website, but I was hoping to understand what does such a result pattern say about the covariate? How should I interpret the results?
Adding the covariate was the novel feature of my study and therefore it is important to me to understand this interaction properly.
Is the covariate influencing the results?
I’ve been thinking about this all day and I guess the possible interpretation would be:
When controlling for the covariate, the interaction of the main effects is no longer significant.
In which case I don’t understand a couple of things:
a) Does that mean the covariate is responsible for the interaction?
b) Why is the interaction between FactorA*FactorB*Covariate not significant then?
c) Why isn’t the covariate significantly interacting with anything at all, when its presence seems to be influencing the results?
Maybe I am doing something wrong, because it is not making much sense to me.
Here are my results:
The last two pictures show when I run an ANOVA without the covariate (repeated measures, Factor 1 (2 levels). Factor 2 (3 levels).
The first three when I run an ANCOVA on the same factors, except I put in OptimismScore as a covariate.
Ordinarily I would say that each term you add diminishes the residual degrees of freedom, and that in turn means your F-statistic has to be larger in order to attain a given level of significance. That does not appear to be the case here because your F-values themselves are very different when you include the covariate.
Some statistical pakcages get higher order F-value calculations wrong in the presence of covariate interactions (specifically, what goes into the denominator). One way to sanity test this is to obtain the coefficient estimates for the model with and without the covariate and see if any of the estimates change a lot (several-fold) between the two. That would be an indicator that the covariate really does make a difference.
But I also am mystified why centering the covariate would fix this. That’s weird enough that if I were you I wouldn’t rely on these calculations until speaking to a statistician.