Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.
From | Maarten buis <maartenbuis@yahoo.co.uk> |
To | statalist@hsphsun2.harvard.edu |
Subject | Re: st: Estimating firm level data on regional level data using a within estimator. |
Date | Sun, 6 Jun 2010 14:34:16 +0000 (GMT) |
--- On Sun, 6/6/10, John Antonakis wrote: > The assumptions of the estimators must be met. If the > assumptions of the random effects estimator are not met, and > if the Hausman test shows that the estimator is not > consistent then the researcher has to bite the bullet! > It is the same thing as estimating a regression model where > you know that x correlates with the disturbance and yet you > go ahead and estimate the model in any case. The coefficient > of x could be higher, lower, or of a different sign. What is > the use to society to report estimates that one knows to be > inconsistent? Inconsistent and bias critically depends on what you want to know: linear regression and random effects will give you a consistent estimate on how the averages differ between groups. So a statement that estimate XYZ is inconsistent or biased is meaningless unless you first specify (explicitly or implicitly) what it is that you exactly want to know. Fixed effects estimators are controlling for all unobserved variables that are constant on the higher level unit. However, you often do _not_ want to controll for all variables, e.g. intervening variables. Since fixed effects indiscrimately controlls for all higher level variables, the fixed effects regression will be a inconsistent estimate for a large (probably the largest) subset of parameters of interest. On the other hand, fixed effects do not control for all unobserved variables you might want to control for, in particular those that aren't constant on the higher level. So again there are a large set of parameters of interest for which fixed effects are inconsistent. In essence the only way to reliably controll for unobserved variables is to observed them. Even a randomized experiment will only work if the paremeter of interest is a linear combination of means (and I am ignoring the problem of external validity). All this is not to deny that randomized experiments and fixed effects regression are useful tools in ones statistical toolbox, but they are just that, a tool with advantages and disadvantages. I am being a bit hard, I guess this has more to do with my frustration with some of the recent converts in my discipline. As is often the case, the recent converts are the worst fundamentalists. My impression is that in many cases where people in my discipline use fixed effects regression those people have no idea what they want to controll for (other than that they want to control for "everything"), which to me means that those estimates are exactly meaningless. Anyhow, as I said, my frustration is with people who abuse a method they don't understand, and obviously none of this applies to you. So don't take any of this personally. -- Maarten -------------------------- Maarten L. Buis Institut fuer Soziologie Universitaet Tuebingen Wilhelmstrasse 36 72074 Tuebingen Germany http://www.maartenbuis.nl -------------------------- * * For searches and help try: * http://www.stata.com/help.cgi?search * http://www.stata.com/support/statalist/faq * http://www.ats.ucla.edu/stat/stata/