Stata The Stata listserver
[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

st: Testing constraints on ISCs in -clogit-


From   "Clive Nicholas" <Clive.Nicholas@newcastle.ac.uk>
To   statalist@hsphsun2.harvard.edu
Subject   st: Testing constraints on ISCs in -clogit-
Date   Sat, 5 Feb 2005 06:26:56 -0000 (GMT)

All,

Hiya! After a very helpful and constructive private dialogue with May
Boggess at StataCorp, I was told that, in fitting a -clogit- model, the
'individual'-specific variables (ISVs) (i.e., independent variables which
are interacted with the alternative-specfic constants) _must_ have their
parameters initially constrained. Quite a revelation, given that no
mention is made of this in quite a number of texts that set out the
conditional (mixed) logit model.

After eventually working out how to set about doing it, my final (edited)
output for the 'null' model was this:

. clogit winner edchange edchgsqd spending spendsqd letoutXcon letoutXlab
letoutXldm clmargXcon clmargXlab cdmargXcon cdmargXldm ldmargXlab
dmargXldm classXcon classXlab classXldm, group(id) constraints(1 2 3 4 5 6
7 8 9 10 11 12) or

Iteration 0:   log likelihood = -2604.3979
Iteration 1:   log likelihood = -1531.3854
Iteration 2:   log likelihood = -1457.0306
Iteration 3:   log likelihood = -1455.6393
Iteration 4:   log likelihood = -1455.6368
Iteration 5:   log likelihood = -1455.6368

Conditional (fixed-effects) logit regression    Number of obs   =      10985
                                                Wald chi2(4)    =    1535.80
Log likelihood = -1455.6368                     Prob > chi2     =     0.0000

 ( 1)  [winner]letoutXcon = 0
 ( 2)  [winner]letoutXlab = 0
 ( 3)  [winner]letoutXldm = 0
 ( 4)  [winner]clmargXcon = 0
 ( 5)  [winner]clmargXlab = 0
 ( 6)  [winner]cdmargXcon = 0
 ( 7)  [winner]cdmargXldm = 0
 ( 8)  [winner]ldmargXlab = 0
 ( 9)  [winner]ldmargXldm = 0
 (10)  [winner]classXcon = 0
 (11)  [winner]classXlab = 0
 (12)  [winner]classXldm = 0
----------------------------------------------------------------------------
    winner | Odds Ratio   Std. Err.      z    P>|z|     [95% Conf. Interval]
-----------+----------------------------------------------------------------
  edchange |   .0001503   .0000605   -21.86   0.000     .0000683     .000331
  edchgsqd |    1020589    2276743     6.20   0.000     12881.91    8.09e+07
  spending |   .2908725   .1378165    -2.61   0.009     .1149212    .7362161
  spendsqd |   234.0945   117.4051    10.88   0.000     87.59698    625.5951
letoutXcon |          1          .        .       .            .           .
letoutXlab |          1          .        .       .            .           .
letoutXldm |          1          .        .       .            .           .
clmargXcon |          1          .        .       .            .           .
clmargXlab |          1          .        .       .            .           .
cdmargXcon |          1          .        .       .            .           .
cdmargXldm |          1          .        .       .            .           .
ldmargXlab |          1          .        .       .            .           .
ldmargXldm |          1          .        .       .            .           .
 classXcon |          1          .        .       .            .           .
 classXlab |          1          .        .       .            .           .
 classXldm |          1          .        .       .            .           .
----------------------------------------------------------------------------

. est store nmodel

. lrtest fmodel nmodel
LR test likely invalid for models with robust vce
r(498);

. lrtest fmodel nmodel, force

likelihood-ratio test                                LR chi2(12) =   1813.60
(Assumption: nmodel nested in fmodel)                Prob > chi2 =    0.0000

Although May didn't actually say the parameters were to be constrained to
zero, I assumed this was what she meant and it does appear sensible.
However, although I finally got what I wanted (a test result which appears
to strongly confirm that the ISVs included in the full model should be
retained), -lrtest- had to be -force-d to run, as you can see.

Is it safe to accept this result given the rather ominous error message
that appeared after -lrtest- was first run? I'm not sure, but it does
represent an emormous difference in the model log-likelihoods for only 12
df (the critical chi-squared value at p=.05 is 21.026). There is no other
way do such a test after -clogit- as far as I'm aware.

Thanks in anticipation.

CLIVE NICHOLAS        |t: 0(044)7903 397793
Politics              |e: clive.nicholas@ncl.ac.uk
Newcastle University  |http://www.ncl.ac.uk/geps

*
*   For searches and help try:
*   http://www.stata.com/support/faqs/res/findit.html
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/



© Copyright 1996–2014 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index