Within and Between Design in Panel Data

[Simone Nuzzo]

Dear Members,

I have run a market experiment which involves both a within and between subject design. I have 6 separated sessions. In each experimental session, each subject goes through 9 tasks (treatments), playing over 7 periods within each task (treatment). So I have 63 periods in each session. Having 6 sessions poses a problem of serial correlation, since observation within each session might exhibit more correlation than observations between sessions (indeed the same group of subject plays within a session). I want to run a panel model (e.g. Pooled OLS, or fixed effects or random effects), where period is the time unit. Is it correct to set the subject as cross sectional unit and then clustering the standard errors at session level (to account for serial correlation)? My concern comes from the fact that I have a within design in each session but, since I have 6 separated sessions (involving six different groups of subjects), it also seems to me that I have a between subject design in place.

Many Thanks!!!

[Dave Airey]

You are certainly allowed to have both within and between factors in a design. These designs are sometimes called split plot designs. You can have more than one within subjects factor, and more than one between subjects factor. What's not clear to me is whether the sessions have different people within them? So are you running a treatmentXperiod two factor within subjects experiment 6 times on the same people? Or are you doing this on 6 independent groups of people? If every person sees every treatment X period X session, then it seems all factors are within subject as subject is crossed with each factor. If you have independent sessions (independent people in each session), then people are nested in session and you have two within one between factor in your design...

[Simone Nuzzo]

Dear Dave,

thanks for you answer. As said above, I have 6 independent groups (i.e. markets or sessions) of people; within each session I have a within subjects design as described above. In a panel model where each period is the time unit, would you set the subject or the group (i.e. market or session) as cross-sectional unit? Why?

Thanks

[Marcos Almeida]

I'm not sure I understood it right, and the best way to share information is, as recommended in the FAQ, to present the data set - full, abrided or a toy example - by using CODE delimiters or by installing the SSC dataex.

That said, I believe your model has a hierarchical design. Therefore, a two-level mixed effects would be an excellent option.

The time variable could be set "within" IDs, whereas IDs could be set within groups.

Please type in command window

. help mixed

and see whether this model applies to your needs;

[Simone Nuzzo]

Thanks Marcos!

As an aside question about panel models: Assume that in a given independent session each subject goes through tasks from A to I, playing 7 periods (repetitions) within each task. So the total number of periods in that session is equal to (9*7=63), with periods 1 to 7 which belong to task A, periods 8 to 14 which belong to task B and so on. Now, assuming to play in a reverse order (i.e. from task I to A), do I need to re-arrange the periods numeration with periods 1 to 7 which belong to task I, periods 8 to 14 which belong to task H and so on...or can I keep everything unchanged (so in the latter case it just matters that I preserve an increasing period numeration within each task)?

Thanks again!