Bookmark and Share

Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.


[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: testing fixed effects versus random effects for clustered data using overiden


From   Gordon Hughes <[email protected]>
To   [email protected]
Subject   Re: st: testing fixed effects versus random effects for clustered data using overiden
Date   Tue, 01 Mar 2011 11:42:50 +0000

A somewhat belated response, but you need to think more carefully about what you are trying to do. Some points to bear in mind:

A. Why are you trying to test FE against RE for your data? I know it is standard in this literature, but does it make sense in this case?

B. Look at your models. The only worthwhile explanatory variables are soil_type_1 & soil_type_2. The classic justification for the FE specification is correlation between the individual effect and some of the explanatory variables, perhaps due to omitted variable bias. Do you have reason to believe that, for example, better or worse farmers tend to have plots with better or worse soil quality? It is possible, but this is the type of question to think about.

C. What are your options if you conclude that FE is the better specification? You have 86 households with an average of 2.6 plots per household, so you will lose 40% of your data. You need to think about alternative specifications that are less vulnerable to the problems that lead the RE estimates to be inconsistent. You might look at Chapter 21 in the Cameron & Trivedi text on Microeconometrics plus the references cited there.

D. Why exactly are you combining robust panel estimation with cluster-adjusted standard errors using the panel variable as the cluster variable? I am surprised that this has even worked, but the results may be meaningless. You might want to cluster households that live in the same village, but there is no reason to duplicate the adjustment for the panel variable.

Ultimately, the suspicion has to be that the model which you are using tells you little more than that soil types influence crop yields. This may or may not be correct, but the solution lies in a better specification of your model, not in trying an ever larger range of econometric techniques to a limited and noisy set of data.

Gordon Hughes
[email protected]


------------------------------

Date: Mon, 28 Feb 2011 14:46:15 +0000
From: Ridhima Gupta <[email protected]>
Subject: st: testing fixed effects versus random effects for clustered data using overiden

Hello,

I don't have a panel data in the strict sense of the term i.e. I have
data on farmers who have several plots/fields. I first perform a
standard hausman test and I do not reject the
null hypothesis of random effects. The result is as follows:

hausman fixed ., sigmamore

                 ---- Coefficients ----
             |      (b)          (B)            (b-B)     sqrt(diag(V_b-V_B))
             |     fixed          .          Difference          S.E.
- -------------+----------------------------------------------------------------
Happy_Seeder |    .1271123     .5428827       -.4157704         .429518
   Rotavator |   -.3621306    -.3157851       -.0463455         .840699
  Seed_Drill |    -.083707      .860967        -.944674        .5539517
quantity_s~c |    .0056369    -.0031733        .0088101        .0188915
plotsize_hec |    .0252054     .0310519       -.0058465        .0525993
   ferti_hec |    .0027516      .003003       -.0002514        .0045816
exp_weedi_~s |    .1077859     .0282141        .0795719        .0585852
 soil_type_1 |   -1.034988    -1.900228        .8652395        .6801256
 soil_type_2 |   -.4202215    -2.165068        1.744846        .8051986
- ------------------------------------------------------------------------------ b = consistent under Ho and Ha; obtained from xtreg B = inconsistent under Ha, efficient under Ho; obtained from xtreg

    Test:  Ho:  difference in coefficients not systematic

                  chi2(9) = (b-B)'[(V_b-V_B)^(-1)](b-B)
                          =       12.70
                Prob>chi2 =      0.1766


But when I perform the robust version of this test, I reject the null
hypothesis of random effects.

Random-effects GLS regression                   Number of obs      =       227
Group variable: hh_id                           Number of groups   =        86

R-sq:  within  = 0.0106                         Obs per group: min =         1
between = 0.0922 avg = 2.6 overall = 0.0674 max = 6

Random effects u_i ~ Gaussian                   Wald chi2(9)       =     15.76
corr(u_i, X)       = 0 (assumed)                Prob > chi2        =    0.0720

(Std. Err. adjusted for 86 clusters in hh_id) - ------------------------------------------------------------------------------
             |               Robust
yield_per_~c |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
- -------------+----------------------------------------------------------------
Happy_Seeder |   .5428827   1.024596     0.53   0.596    -1.465289    2.551054
Rotavator | -.3157851 1.425605 -0.22 0.825 -3.10992 2.47835 Seed_Drill | .860967 1.036252 0.83 0.406 -1.170049 2.891983
quantity_s~c |  -.0031733    .041733    -0.08   0.939    -.0849685     .078622
plotsize_hec |   .0310519   .0467653     0.66   0.507    -.0606065    .1227103
ferti_hec | .003003 .0053299 0.56 0.573 -.0074435 .0134495
exp_weedi_~s |   .0282141   .0792311     0.36   0.722    -.1270759    .1835041
soil_type_1 | -1.900228 .7639786 -2.49 0.013 -3.397599 -.4028574 soil_type_2 | -2.165068 .8668351 -2.50 0.013 -3.864033 -.4661022 _cons | 41.85728 4.370738 9.58 0.000 33.29079 50.42376 - -------------+----------------------------------------------------------------
     sigma_u |  4.4041886
     sigma_e |  4.0351113
         rho |  .54364968   (fraction of variance due to u_i)
- ------------------------------------------------------------------------------

. xtoverid

Test of overidentifying restrictions: fixed vs random effects
Cross-section time-series model: xtreg re  robust cluster(hh_id)
Sargan-Hansen statistic  26.327  Chi-sq(9)    P-value = 0.0018


Is there any inconsistency here? Are there any tests in the literature
that allow me to test the assumption of homoskedasticity and no
auto-correlation in the random effects model?

Thanks a lot,
Ridhima

*
*   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/


© Copyright 1996–2018 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   Site index