st: xtreg fe with time varying categorical predictors

[Vaidyanathan Ganapathy <vaidyang@usc.edu>]

Hi Statalisters,

I have longitudinal claims data of children with a certain type of
insurance provided for low income families. I am interested in running
a fixed effects regression of healthcare expenditures regressed upon the
   category of insurance eligibility (changes over time with change in family
  income) and a set of medical conditions (coded as 0 = not present, 1 =
  if present, etc.). These medical diagnoses change over time, depending on
   whether or not these diagnoses were observed in each 6-month period, over
  the length of the panel (2 years). I have N > 100,000 and T = 4 periods.

  Now, when I run the panel regressions assuming individual random or fixed
   effects, the random effect estimates appear to be meaningful (I.e., the
  signs of the coefficients are as expected). The fixed effect estimates
 behave quite interestingly as explained below -

  1) for the category of insurance eligibility variable, the coefficient
   signs on some of the categories change (with significance) than seen for
  random effects. This is not meaningful (for e.g., I would expect that a
  change from non-disabled category to disabled category, over time, should
   increase the overall healthcare expenditures whereas I see a decrease in
  costs using the fixed effects estimates, p < 0.05).

  2) for the medical conditions, all of which are supposed to have a positive
  impact on healthcare expenditures, to different magnitudes though, I see
   the effects are negative if I use lagged diagnoses and positive if I use
  level variables. The reason for using lagged diagnoses (one lag) is some of
  the diagnoses could be pre-determined.

  The Hausman test shows that the fixed effects estimates are consistent
   under H0 and Ha (under xtreg).

  Any thoughts on why the fixed effect estimates are behaving this way would
  be highly appreciated. As you would have noted, my regressors are all
  categorical variables, although they vary over time (some with more
   temporal variation as compared to others). Is there something that I need
  to know when running fixed effects with categorical regressors?

 Thanks,

Vicky
*
*   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/