Re: Re: st: Problems with FE and Hausman test

[Christopher Baum <kit.baum@bc.edu>]

<>
On Feb 20, 2014, at 2:33 AM, John wrote:

> So then the solution would be start adding dummy variables: 22 for states (23-1) and 2 for time (3-1)?

> xtreg PERCENTGRAD IV1 IV2 IV3 IV4 IV5 IV6 IV7 IV8 State1, State2, State3, State4, State5, State6, State7, State8, State9, State10,
> State11, State12, State13, State14, State15, State16, State17, State18, State19, State20, State21, State22, Timeis2007, Timeis2008, re
> estimates store fixed_group
> xtreg PERCENTGRAD IV1 IV2 IV3 IV4 IV5 IV6 IV7 IV8 State1, State2, State3, State4, State5, State6, State7, State8, State9, State10, State11, State12, State13, State14, State15, State16, State17, State18, State19, State20, State21, State22, Timeis2007, Timeis2008, fe
> hausman fixed_group
> 
> As for the "quality of education factors", my research and the literature presupposes that in fact yes there are such factors, the average teacher salary for example you mentioned. But the education debate and literature in some states has also suggested taking power away from local school boards and giving it to the state board of education may (or may not) be helpful in this regard. One IV, for example, is on the purchase of school texts: can the local board go out and buy whatever it wishes, pick only from a list of books "certified" by the state board of ed, etc. Another is on the power of the state board of education to put a school into administrative receivership/takeover the local school board.

No, not at all. The time dummies are fine, and should be added. But the state dummies do exactly the same thing as the fixed effects estimator. You cannot run FE because the factors IV1-IV8 vary by state but not over time, as I understand it. Thus running an OLS regression on IV1-IV8, state dummies and time dummies will cause all IV factors to be dropped, as you found when you tried the FE estimator. It also makes no sense to introduce fixed effects (in whatever form) and random effects, as they are alternative ways of modeling the state-specific unobserved heterogeneity.

If you write down a model in which the outcome is a function of state-specific factors IV1-IV8 and macro variation over time (as captured by two time dummies), you're saying that variations in the outcome that are not explained by state or time end up in the error term, as you have nothing else to explain them. Thus one approach would be to apply the between estimator (xtreg, be) and just run the regression on the cross-sectional averages. 

My concern with your basic model and the argument that these "powers given to the locality" are important is that your model, as written, says that they are the ONLY factors (other than geography and time) that affect the outcome. I do not question that these "powers" measures might have explanatory power, over and above other measures such as teacher salaries, expenditures per pupil, etc. But the model as you state it argues that ONLY the "powers" variables matter, and that everything else can be subsumed in a state fixed effect. If during the period some states applied policies to improve graduation rates, e.g. by spending more on tutoring weak students, those would not be captured in your model, and would end up in what is left unexplained. But if you ask a teacher or principal in that state why the graduation rate increased, they could tell you.

Kit

Kit Baum
Professor of Economics and Social Work, Boston College, Chestnut Hill MA, USA
DIW Research Professor, Department of Macroeconomics, DIW Berlin, Berlin, Germany
baum@bc.edu  |  http://ideas.repec.org/e/pba1.html












*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/