[George Ford]

Rejection is evidence against the RE model. So if no rejection, then RE model is fine. See Schunck (2013) at p. 69.

[ajay pasi]

Hello George,
In Schunck and Perales (2017) "In An analogous correlated random-effects model can be estimated using xthybrid by adding the cre option (the results are presented in table2)" (p.102). By RE, do you mean cre option in xthybrid? Or you mean xtlogit Positive_disc01 i.Student_Caste_New i. T_jati_new attendence_percent i.T_nature i.course1 semester i.T_gender i.division , or re vce (robust )?

ref:
Schunck, R., & Perales, F. (2017). Within-and between-cluster effects in generalized linear mixed models: A discussion of approaches and the xthybrid command. The Stata Journal, 17(1), 89-115.

[ajay pasi]

Rejection is evidence against the RE model. So if no rejection, then RE model is fine. See Schunck (2013) at p. 69.

, so it means fixed effect model is not appropriate for the analysis!?

[Carlo Lazzaro]

Ajay:
no, the -fe- estimator is appropriate, as rejection of the null is evidence that -re- is not the way to go.

[ajay pasi]

Thanks Carlo sir, however, when I ran the regression with the fixed effect, it is not reporting the coefficients

. xtlogit Positive_disc01 i.Student_Caste_New i.T_jati_new    attendence_percent i.T_nature i.course1 semester
>    i.T_gender      ,    fe
note: multiple positive outcomes within groups encountered.
note: 8 groups (122 obs) omitted because of all positive or
      all negative outcomes.
note: 2.Student_Caste_New omitted because of no within-group variance.
note: 3.Student_Caste_New omitted because of no within-group variance.
note: 2.course1 omitted because of no within-group variance.
note: 3.course1 omitted because of no within-group variance.
note: 4.course1 omitted because of no within-group variance.
note: 5.course1 omitted because of no within-group variance.
note: 6.course1 omitted because of no within-group variance.
note: 7.course1 omitted because of no within-group variance.

Iteration 0:   log likelihood = -4796.8683  
Iteration 1:   log likelihood = -4781.9886  
Iteration 2:   log likelihood = -4781.9698  
Iteration 3:   log likelihood = -4781.9698  

Conditional fixed-effects logistic regression        Number of obs    =  9,969
Group variable: collegerollno                        Number of groups =    661

                                                     Obs per group:
                                                                  min =      3
                                                                  avg =   15.1
                                                                  max =     16

                                                     LR chi2(6)       =  51.67
Log likelihood = -4781.9698                          Prob > chi2      = 0.0000

------------------------------------------------------------------------------------
   Positive_disc01 | Coefficient  Std. err.      z    P>|z|     [95% conf. interval]
-------------------+----------------------------------------------------------------
 Student_Caste_New |
            SC/ST  |          0  (omitted)
              OBC  |          0  (omitted)
                   |
        T_jati_new |
                2  |  -.0534922   .0658445    -0.81   0.417    -.1825452    .0755607
                3  |   .0108931    .070487     0.15   0.877    -.1272589    .1490451
                   |
attendence_percent |   .0110619   .0017746     6.23   0.000     .0075838      .01454
        2.T_nature |   .0140017   .0547203     0.26   0.798    -.0932482    .1212516
                   |
           course1 |
              eco  |          0  (omitted)
              eng  |          0  (omitted)
            hindi  |          0  (omitted)
          history  |          0  (omitted)
            maths  |          0  (omitted)
              pol  |          0  (omitted)
                   |
          semester |   .0865461   .0209006     4.14   0.000     .0455818    .1275105
        2.T_gender |   .2040773   .0723231     2.82   0.005     .0623267    .3458278
------------------------------------------------------------------------------------

What might be the issue?, I am unable to understand.!

[Carlo Lazzaro]

Ajay:
as Stata reports, the issue is the lack of within panel variation (hence, no variance).
Basically, you're asking the conditional -fe- estimator to return coefficients of time-invariant predictors (something that the it is not able to do).

[ajay pasi]

Thnaks carlo, would you please suggest a way out of this? However, when I am running the regression without fe (with i.var), I am getting this.

. logit Positive_disc01 i.Student_Caste_New i.T_jati_new    attendence_percent i.T_nature i.course1 i.semester
>    i.T_gender, vce (robust)

Iteration 0:   log pseudolikelihood = -6916.3047  
Iteration 1:   log pseudolikelihood = -6201.3585  
Iteration 2:   log pseudolikelihood = -6197.3977  
Iteration 3:   log pseudolikelihood = -6197.3964  
Iteration 4:   log pseudolikelihood = -6197.3964  

Logistic regression                                    Number of obs =  10,091
                                                       Wald chi2(17) = 1242.22
                                                       Prob > chi2   =  0.0000
Log pseudolikelihood = -6197.3964                      Pseudo R2     =  0.1039

------------------------------------------------------------------------------------
                   |               Robust
   Positive_disc01 | Coefficient  std. err.      z    P>|z|     [95% conf. interval]
-------------------+----------------------------------------------------------------
 Student_Caste_New |
            SC/ST  |  -.1396649   .0564622    -2.47   0.013    -.2503289   -.0290009
              OBC  |  -.1004951   .0566178    -1.77   0.076     -.211464    .0104737
                   |
        T_jati_new |
                2  |  -.0330889   .0666478    -0.50   0.620    -.1637162    .0975385
                3  |   .0512036   .0685315     0.75   0.455    -.0831156    .1855229
                   |
attendence_percent |   .0109082   .0012411     8.79   0.000     .0084757    .0133408
        2.T_nature |   .0455933   .0525003     0.87   0.385    -.0573055    .1484921
                   |
           course1 |
              eco  |   .8815111   .0747911    11.79   0.000     .7349233    1.028099
              eng  |  -.3306421   .0822744    -4.02   0.000    -.4918969   -.1693873
            hindi  |   1.269306    .084291    15.06   0.000     1.104099    1.434513
          history  |    .323667   .0984176     3.29   0.001      .130772     .516562
            maths  |   .6725393   .0771584     8.72   0.000     .5213116     .823767
              pol  |   1.994379   .0807284    24.70   0.000     1.836154    2.152603
                   |
          semester |
                2  |   .0811282   .0868599     0.93   0.350     -.089114    .2513704
                3  |   .1893041   .0759795     2.49   0.013     .0403871    .3382211
                4  |   .5681824    .080683     7.04   0.000     .4100465    .7263182
                5  |   .2149754   .0795153     2.70   0.007     .0591284    .3708225
                   |
        2.T_gender |     .20029   .0712817     2.81   0.005     .0605805    .3399996
             _cons |  -1.452294   .1297851   -11.19   0.000    -1.706668    -1.19792
------------------------------------------------------------------------------------

[Carlo Lazzaro]

Ajay:
as expected, the -re- estimator (that you implicitly invoked in your last code) gives yiou back coefficients for time-invariant predictors, too.
The issue is that, as pre your previous tests, the conditional -fe- estimator outperforms its -re- cousin.
I am afraid the only way out is to stick with the conditional -fe- estimator and its drawbacks.

[ajay pasi]

Thanks Carlo, I appreciate all your guidance.