fractional response model for longitudinal data?

[Dave Airey]

I had experimental percentage outcome data with a good deal of 0s and 100s (histogram for percentage outcome was U shaped). I used "fracreg logit" on these [0,1] data as well as some non-parametric methods. I've also explored beta regression after bumping the 0s and 100s larger and small, respectively. Now longitudinal data on the same outcome to look at some time effects are being considered. Are there mixed models for this kind of outcome in Stata?

The data are generated without a denominator for the percentage. There is a fixed resource that can be used during the experiment. Some use all of it, none of it, or some of it. How much gets used is the outcome variable.

[Stephen Jenkins]

Dave: is the following paper of relevance? "Panel data methods for fractional response variables with an application to test pass rates", by Leslie E. Papke, Jeffrey M. Wooldridge, Journal of Econometrics 145 (2008) 121–133

[Dave Airey]

Thank you. I should have looked in the econometric literature. I was in the med lit. Do you know if it is available as a SAS/R/Stata implementation? I'll take the paper up to some others to look at.

Dave

[Dave Airey]

I found some implementation here. I was blocked from another location.

http://econ.msu.edu/faculty/papke/

Dave

[John Mullahy]

Dave: See section 3 in the Papke-Wooldridge paper that Stephen mentioned (which I also believe is the key reference on this topic). Their method should be pretty straightforward to implement in Stata with -glm- (using -cluster-) or maybe -xtgee- I'd think. I'd guess possible in R or SAS too, but don't know the particulars.

The key in the P-W approach is that it involves a probit, not logit, link. That said, it's a pretty flexible modeling approach for panels with fractional outcomes.

[Nick Cox]

See also earlier references in http://www.statalist.org/forums/foru...-robust-option esp; #10 from Maarten Buis.