st: Inconsistent number and maximum size of groups using -xtlogit, fe-

["David W. Harless" <dwharles@vcu.edu>]

I am estimating a model using conditional logit/fixed-effects logit, but using the command 
-xtlogit, fe- because it reports (and saves in e()) information on the number and size of 
groups.  But in a sample with some very large group sizes, the number of groups and the 
maximum size of group from -xtlogit, fe- is different than what I calculate directly. 
For example:

********************************************************************************
.         xtlogit adverse x , fe i(g1)
note: multiple positive outcomes within groups encountered.
note: 19918 groups (525174 obs) dropped due to all positive or
      all negative outcomes.

Iteration 0:   log likelihood = -11648.872
Iteration 1:   log likelihood = -11647.283
Iteration 2:   log likelihood = -11647.282
Iteration 3:   log likelihood = -11647.282

Conditional fixed-effects logistic regression   Number of obs      =   1522444
Group variable (i): g1                          Number of groups   =       597

                                                Obs per group: min =         2
                                                               avg =    2121.8
                                                               max =     31712

                                                LR chi2(1)         =      3.76
Log likelihood  = -11647.282                    Prob > chi2        =    0.0524

------------------------------------------------------------------------------
     adverse |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
           x |  -.1458831   .0743596    -1.96   0.050    -.2916254   -.0001409
------------------------------------------------------------------------------

.         keep if e(sample)
(525174 observations deleted)

.         collapse (count) adverse, by(g1)

.         summ adverse, d

                       (count) adverse
-------------------------------------------------------------
      Percentiles      Smallest
 1%            2              2
 5%           11              2
10%           36              2       Obs                 603
25%          145              2       Sum of Wgt.         603

50%          505                      Mean           2524.783
                        Largest       Std. Dev.      6166.928
75%         1734          36928
90%         5423          47192       Variance       3.80e+07
95%        14684          48473       Skewness        4.51094
99%        31712          51907       Kurtosis       26.67479

.         disp r(sum)
1522444

.         count if adverse<=31712
  597
*******************************************************************************

Hence, when I use -collapse- to calculate the number and sizes of groups, I obtain 603 
groups (compared to 597 in -xtlogit, fe-) and find six groups having more members than the 
maximum group size reported by -xtlogit, fe- (31,712).


I note that if I go back to my original data set and use only the groups with more than 
31712 observations, the number of groups is indicated as 0 and there are missing values 
for the number of observations per group:

**********************************************************************************
.         quietly bysort g1: gen grp_cnt = _N

.         xtlogit adverse x if grp_cnt>31712, fe i(g1)
note: multiple positive outcomes within groups encountered.

Iteration 0:   log likelihood =  -1205.835
Iteration 1:   log likelihood = -1205.8343
Iteration 2:   log likelihood = -1205.8343

Conditional fixed-effects logistic regression   Number of obs      =    255748
Group variable (i): g1                          Number of groups   =         0

                                                Obs per group: min =         .
                                                               avg =         .
                                                               max =         .

                                                LR chi2(1)         =      1.31
Log likelihood  = -1205.8343                    Prob > chi2        =    0.2533

------------------------------------------------------------------------------
     adverse |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
           x |   .3934376   .3650677     1.08   0.281    -.3220821    1.108957
------------------------------------------------------------------------------

**********************************************************************************

I am using an updated Stata/MP for 64-bit Windows:
. about

Stata/MP 9.2 for Windows 64-bit x86-64
Born 06 Feb 2007

Dave Harless





*
*   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/