# repeated measures ANCOVA; covariate main effect and post-hoc tests

let's say I run a repeated measures ANCOVA, with valence (negative and neutral) as repeated measure and age as a covariate. what I want to know is how age affects my data.

1)I get a significant interaction between the covariate age and valence. As post-hoc I would do a linear regression, where I regress age on the difference between negative and neutral (negative minus neutral), since Anovas work on the difference between levels of the dependent variables.

2) I get a significant covariate effect (between subject table) and as a post-hoc I am doing a linear regression, regressing age on the mean of valence ((negative+neutral)/2).

My questions, first of all, is the approach to perform a repeated measures ANCOVA appropriate since I am interested in the effect of the covariate and do not want to partial it out.

Further, are linear regressions, as described above appropriate as testing the type of the relationship between the covariate and the variables that were significant?

And last, is the way I would create the variable to be regressed on (mean for BS and difference between levels of the independent variable) how it should be? If not, on what should I regress the mean?

Thank you!!!!

For sure is too late, but you should work with linear mixed models instead of ancova or Lm. Your proposals include obtaining a new variable in both cases to consider it as a DV. With linear mixed models you dont need to do this ( and you can also include trial by trial data)

As a general example:

ID Condition value Age 1 neutral 6 35 1 negative 3.5 35 2 neutral 7 21 2 negative 2.3 21 .. . ...

Here, "value" is the average value of whatever you are using in each condition. In short, if you have this data structure you can easily analise it this way:

lmer(value ~ condition + age + (1|ID))

With this model you can test all that you said!

• Sorry, I tried to represent a dataframe but I failed. To be concise, the data must be in long format! Mar 19 at 0:39
• Isn't there a minimum cluster size of 3 observations per ID required for random effects models? Aug 27 at 11:31