Age-Period-Cohort Mixed Models

[cecilmeeusen]

I am performing an Age-Period-Cohort Mixed Model analysis on the European Social Survey Data (ESS 2002-2010).
To estimate such a model, period and cohort effects must be specified as random effects and age as a fixed effects.
The problem is that I have only 5 survey years. Is it allowed to specify a period variable with only 5 time points as random?
In other articles on APCMM they perform random effects of period variables with only 8 datapoints. What is the limit?

Is it allowed to fix the period effects as well, and only randomise cohort effects?

In Mlwin: I perform a cross-classified model with the SETX command. That is correct?

Thank you,

Cecil

[joneskel]

Two answers and a warning

1) to perform cross-classified model I would use the MCMC procedures ( see MCMC Manual)

2) this simulation study reports on small n for random effects: daniel-stegmueller.com/files/Multilevel_samplesize.pdf

A Warning - this is a personal view: the 'trick' of using age in the fixed part and (grouped) cohort and period in the random effects in a cross classified model is, from my perspective, problematic, as it relies on getting the specification of the data generating process correct and there is no way of knowing that for certain. More on this when some simulation studies have been completed.

[andrewjdbell]

To follow up on this, see this paper that Kelvyn and I have written and is currently in review, which uses simulations to assess the HAPC model via simulations.

http://www.mendeley.com/download/public ... 41c/dl.pdf