[Leon Edelman]

Thank you http://www.statalist.org/forums/memb...lyde-schechter. If that is the case that it won't be estimated with much precision is there a different method to use? How can I avoid this issue? Or perhaps How can I estimate the age period and cohort effects?

[Clyde Schechter]

Well, when you have exact information, there is the identity: age = period - cohort. So there is inherent colinearity among these constructs, and any model that tries to estimate effects of all three is inherently unidentified. There are two basic ways to approach this difficulty.

1. Assume some constraint. For example, historical information, or scientific theory sometimes enables one to stipulate that one of these effects is zero, or that it follows some specific linear trend, or something like that. If you can do that, then the identification problem goes away. Before you introduced the age variable, that is what you had implicitly done: you had constrained the age effect to be zero by not including it in your model.

2. Use imprecise data. This is what you have, implicitly, done here. Your birth cohort variable is coarse: each category covers decades. Your period variable is also somewhat coarse, with 5 time points that, I cannot tell, may not actually represent 5 equally spaced points in "real" time. (I assume nothing here is moving close to the speed of light, nor operating in the gravitational field of a nearby black hole.) By using imprecise data, the exact colinear relationship is broken and estimates are possible. But near-colinearity remains. From that the inevitable consequence is imprecise parameter estimates.

I am not aware of any alternatives to these approaches.