2
$\begingroup$

We have carried out an experiment that measures how many references it takes within a group to find the most expert person in that group on a specific topic (DV). We have 7 topics (4 of which are considered "visible" and 3 of which are considered "invisible"). The visible/invisible distinction is our treatment. We have 42 groups (school classes). We conducted the experiment with all 7 topics in each group so that groups and topics are crossed. We believe that the number of references required to find the most expert person in a class depends on the visibility/invisibility of the search topic. I am struggling to come up with the proper design. I believe that search topics are nested in the treatment (visible/invisible). I would also like to consider the fact that I have 7 measures for each class. I am not interested in assessing a direct effect of class but just want to control for the fact that the 7 measures within one class may not be independent from each other. Can this still be done with an ANOVA (and if so what type?) or should I rather use a GLM?

$\endgroup$

1 Answer 1

1
$\begingroup$

First, thanks for providing context, that always helps

Second, your dependent variable is a count (number of references); you should account for that in your choice of analysis (if the counts were all quite large, this might be ignorable, but it looks like your counts are low).

Third, since you did the experiment multiple times in each class, therefore your data are not independent and you should deal with that in your analysis as well.

One choice would be a nonlinear mixed model, these are quite complex models; if you have little experience with statistics they may be kind of daunting. That said, if you are using R you can look at the lme4 and nlme packages. If you are using SAS then GLIMMIX is the PROC to at least start with.

$\endgroup$
3
  • $\begingroup$ thanks for your help! You are correct, the DV is comparatively low. It is an approximation of the mean number of references that is obtained by subjecting the data from the experiment to a monte carlo simulation. So while these are usually not integer counts they cannot be negative. The DV is also very close to normal distribution, i.e. it does not exhibit typical problems of counts such as excess zeros. I have access to Stata and HLM. Could you point me towards the better alternative of these two and maybe even a command? $\endgroup$
    – Christoph
    Oct 1, 2013 at 12:52
  • $\begingroup$ Also, I was wondering about whether I also need to account for the topics being nested in visible/invisible? Or is this the wrong way of looking at things? $\endgroup$
    – Christoph
    Oct 1, 2013 at 12:53
  • $\begingroup$ Nestedness is OK, as far as I can tell. $\endgroup$
    – Peter Flom
    Oct 1, 2013 at 17:29

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.