Ordered Panel models

[NSimmons]

Hi all,

Is there a way of choosing between an Ordered FE Panel model and an Ordered RE Panel model, similar to a Hausman Test? Is this what the Hausman specification test is choosing between when run in Stata 13? I am assuming the stumbling block would be that I am unable to estimate the FE Ordered Model in Stata...

Also, if anybody knows a way of estimating an Ordered FE model in Stata that can allow for 32,000 individuals, it'd be appreciated! A matsize limit of 11,000 seems to be my stumbling block.

Thank you!

Nick

[daniel klein]

Nick,

I suppose you are talking about fixed effects models in the sense of an within estimator, not 'fixed effects' in the sense of mixed models (which do not control for unobserved heterogeneity).

As far as I know there is indeed no such model for an ordered response (i.e. no equivalent conditional likelihood estimator as in the binary case) - but correct me if I assume wrong. However, Allison (2009) discusses so-called 'hybrid' models. He suggests to (i) calculate the within level 1 mean of predictors that vary within units, then (ii) create corresponding predictors as deviations from the means and (iii) finally include both in an ordered logistic model (ologit) with robust standard errors. The predictors in their demeaned form represent within unit effects while the means of the predictors themselves control for between unit differences. As an alternative to the Hausman test, Allison suggests to test the difference between the coefficients for the demeaned predictor and its mean counterpart.

Given Stata 13 now can estimate a mixed-ordered model, you could use meologit instead of ologit to correct for auto-correlation of the standard errors.

Best
Daniel


Allison, Paul, D. (2009). Fixed Effects Regression Models. Thousand Oaks, CA:Sage Publications.


Edit:
Readers might be interested in other approaches to ordered FE models in Stata http://repec.org/usug2011/UK11_Hole.pdf

[NSimmons]

Hi Daniel,

Thanks for the response - the link at the bottom in estimating ordered FE models is actually what I've been looking at (I have estimated all of those models outined by Hole). However, I haven't looked into the Allison book (trying to find access to it!) so will look into this, and of course find code to help!

Regards,
Nick

[daniel klein]

Allison gives complete Stata code to replicate the examples he uses in the book in the appendix. Also you might might be interested in Schnuck (2013) who shows the implementation in the linear case and discusses problems with interactions using factor variable notation in such models and much more.

Best
Daniel


Schnuck, Reinhard (2013). Within and between estimates in random-effects models: Advantages and drawbacks of correlated random effects and hybrid models. The Stata Journal volume 13, number 1, pp. 65-76.
(http://www.stata-journal.com/article...article=st0283)

[Richard Williams]

The Allison book is cheap:

http://www.amazon.com/Effects-Regres.../dp/0761924973

[Richard Williams]

Also, Allison's data sets can be found at

http://www.statisticalhorizons.com/resources/data-sets

I don't think he keeps the code online though, which is unfortunate since the do files can get pretty long.