Compare restricted and full model that are pooled and p-weighted

[Kerstin Schmidt]

Dear All!

The issue I want to solve is: I have two data points in time, whereas in 8% not the same person (household head) answered all questions but rather the child was the respondent in the second wave. Assuming that this is not a problem as characteristics are inherited from one generation to another, I believe to simply match these two waves. However, I need to also statistically test for this.
Therefore, I want to statistically compare a restricted model (only household heads responded both) with the full model. Now I am puzzled because they are pooled in nature (thus use clustering) and are sample weighted, which does not allow me to use "suest" and the Hausman test.

Would like to do:

mlogit PortfolioDecision  $x1 if SamePerson==1 [pweight=pweight], b(1) nolog
margins, dydx(*)
estimates store restricted

mlogit PortfolioDecision  $x1 [pweight=pweight], b(1) nolog
margins, dydx(*)
estimates store full

suest full restricted, vce(cluster hhid)  noomitted

test [full_0=restricted_0], cons

But the suest command does not allow for both clustering and pweights:

. suest full restricted, vce(cluster hhid)  noomitted
model full was estimated with pweights, you should re-estimate using iweights
r(322);

end of do-file 

Does anyone know what I can do instead? Would it also be possible to run the commands above with (1) unweighted but clustered and (2) weighted but not-clustered data and still get at least a good approximate for a statistical difference between the two?
Any help is highly appreciated!

Thanks

[Andrew Musau]

I would just expand the observations and drop the probability weights. Example:

webuse sysdsn1
set seed 01212021
gen pw= runiformint(1, 5)
*WITH PROBABILITY WEIGHTS
mlogit insure age male nonwhite i.site [pweight= pw], b(1) nolog
preserve
*EXPAND OBS
expand pw
mlogit insure age male nonwhite i.site, b(1) nolog
restore

Res.:

. *WITH PROBABILITY WEIGHTS

. 
. mlogit insure age male nonwhite i.site [pweight= pw], b(1) nolog

Multinomial logistic regression                 Number of obs     =        615
                                                Wald chi2(10)     =      24.98
                                                Prob > chi2       =     0.0054
Log pseudolikelihood = -1629.6254               Pseudo R2         =     0.0298

------------------------------------------------------------------------------
             |               Robust
      insure |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
Indemnity    |  (base outcome)
-------------+----------------------------------------------------------------
Prepaid      |
         age |  -.0095502   .0067665    -1.41   0.158    -.0228122    .0037119
        male |    .295541   .2202639     1.34   0.180    -.1361683    .7272504
    nonwhite |    .928286   .2628762     3.53   0.000     .4130582    1.443514
             |
        site |
          2  |   .1858761   .2324204     0.80   0.424    -.2696595    .6414116
          3  |  -.4475149   .2523063    -1.77   0.076    -.9420261    .0469963
             |
       _cons |   .1372069   .3572138     0.38   0.701    -.5629193     .837333
-------------+----------------------------------------------------------------
Uninsure     |
         age |   .0006181   .0122265     0.05   0.960    -.0233453    .0245815
        male |   .2341683   .3967306     0.59   0.555    -.5434093    1.011746
    nonwhite |   .3329798   .4529042     0.74   0.462    -.5546961    1.220656
             |
        site |
          2  |  -1.126122   .5162875    -2.18   0.029    -2.138027   -.1142168
          3  |  -.2055748   .3991142    -0.52   0.606    -.9878242    .5766747
             |
       _cons |  -1.740114   .6778017    -2.57   0.010    -3.068581   -.4116476
------------------------------------------------------------------------------


. 
. *EXPAND OBS

. 
. expand pw
(1,322 observations created)

. 
. mlogit insure age male nonwhite i.site, b(1) nolog

Multinomial logistic regression                 Number of obs     =      1,877
                                                LR chi2(10)       =     100.08
                                                Prob > chi2       =     0.0000
Log likelihood = -1629.6254                     Pseudo R2         =     0.0298

------------------------------------------------------------------------------
      insure |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
Indemnity    |  (base outcome)
-------------+----------------------------------------------------------------
Prepaid      |
         age |  -.0095502   .0035496    -2.69   0.007    -.0165072   -.0025931
        male |    .295541   .1134714     2.60   0.009     .0731411    .5179409
    nonwhite |    .928286   .1345234     6.90   0.000      .664625    1.191947
             |
        site |
          2  |   .1858761   .1197359     1.55   0.121    -.0488019    .4205541
          3  |  -.4475149   .1290276    -3.47   0.001    -.7004044   -.1946254
             |
       _cons |   .1372069   .1876839     0.73   0.465    -.2306467    .5050605
-------------+----------------------------------------------------------------
Uninsure     |
         age |   .0006181   .0066241     0.09   0.926     -.012365    .0136011
        male |   .2341683   .2142533     1.09   0.274    -.1857605    .6540972
    nonwhite |   .3329798   .2448515     1.36   0.174    -.1469203      .81288
             |
        site |
          2  |  -1.126122   .2727336    -4.13   0.000     -1.66067   -.5915738
          3  |  -.2055748    .214918    -0.96   0.339    -.6268063    .2156568
             |
       _cons |  -1.740114   .3501604    -4.97   0.000    -2.426416   -1.053813
------------------------------------------------------------------------------

[Kerstin Schmidt]

That's a very good idea - thank you very much Andrew!