0
$\begingroup$

I am having a few problems with my analysis and hope that somebody can help me out. About my design: I have 120 respondents, who I showed each 7 facebook posts out of a pool of 16 (randomly assigned). Each facebook posts was "ranked" by them on 1 DV (shares) and 3 IV (entertainment, information and emotion).

Due to the repeated measures design I decided to use ANCOVA and include "participant" as a factor, controlling for general differences among participants (so 'groups'), and my three IV's as covariates. I am not interested in group differences though. Up to now makes sense, doesnt it? I could also perform a linear regression but it would be kind of crazy to create a dummy for 120 participants?

Now I have the problem that I dont meet the assumptions. If I check for interactions between the factor (participants) and IV's, they are all significant. And in a model with participant as factor and the three covariates, the Levene's test is significant as well...Nevertheless, overall there is a linear relationship between DV and IV's. But each group is only 7 responses, this may be too small?

Anybody knows why and can help? Since I use ANCOVA for another purpose than most of the people anyway, I am not sure if I can ignore the violations or how I can help it.

Highly appreciate any help!! :) Wiebke

Improve this question
$\endgroup$

1 Answer 1

Reset to default
0
$\begingroup$

First, ANCOVA and regression are the same model. In matrix terms both are Y + XB + e.

Second, when you have repeated measures then the assumption of independent errors is necessarily violated. Neither regression nor ANCOVA deals with this.

Third, two general methods of dealing with dependence like this are generalized estimating equations and multi-level models. Both have been discussed here a lot.

Improve this answer
$\endgroup$
7
  • $\begingroup$ Thank you for your response Peter! I will have a look into both models, although I assume that they are too complex for my paper. $\endgroup$
    – Wiebke
    Commented May 17, 2014 at 16:55
  • $\begingroup$ If they are too complex for your paper, then you need to change your paper. As is, your analysis violates the assumptions of the model. $\endgroup$
    – Peter Flom
    Commented May 17, 2014 at 17:18
  • $\begingroup$ I dont really understand the mixed model. Changing my paper is a bit too late, as I have performed questionnaire etc. already. But if I use ANCOVA as linear regression (since thats what it is) and include participant as a random factor (to control for that) and I am not interested in group differences anyway, do I need to look the ANOVA assumptions anyway or rather at the multiple regressoin assumptions? Sorry for the confusion!! :/ $\endgroup$
    – Wiebke
    Commented May 18, 2014 at 13:22
  • 1
    $\begingroup$ Thanks Peter! I appreciate your help. I guess then I have to go into the mixed models. I just looked into GEE, and was wondering it its correct that my participant is my subject variable and the post number is my within-subject variable? Or the way around? Thanks!! $\endgroup$
    – Wiebke
    Commented May 18, 2014 at 14:04
  • 1
    $\begingroup$ I have performed a GEE and my results seam reasonable :-) It was easier than expected. Thank you for your help again! $\endgroup$
    – Wiebke
    Commented May 19, 2014 at 12:23

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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