Reporting results ordered logit regression: individual predictors or entire model?

[Anthon Balthon]

Hello,

I am running the ordered logit regression to predict eh041, by regression on the variables aa001, aa004, ba016, ca001, ea104, eb001, eb002, ec023, and dummy.

My question is: Is it better to report the coefficients, standard errors, and p-values for each individual predictor, or for the entire model (if so, which statistics to report?)?

Example of the dataset:

input long id int year byte(ca001 aa001) float aa004 byte(eb001 eb002) float(ea104 ec023) byte(eh041 ba016) int(dummy)
11001 2004 1 1 60 0 1   10 3 2 4 0
11001 2006 . .  . . .    . . . . 1
11002 2004 . 2 65 . .    . . . 4 0
11002 2006 . .  . . .    . . . . 1
25601 2004 1 1 50 0 1   36 5 2 6 0
25601 2006 1 1 52 0 1   36 4 1 6 1

Command for ordered logit regression:

ologit eh041 aa001 aa004 ba016 ca001 ea104 eb001 eb002 ec023 dummy

Output:

note: ca001 omitted because of collinearity
Iteration 0:   log likelihood = -5928.1906  
Iteration 1:   log likelihood = -5880.5609  
Iteration 2:   log likelihood = -5880.4552  
Iteration 3:   log likelihood = -5880.4552  

Ordered logistic regression                     Number of obs     =      6,312
                                                LR chi2(8)        =      95.47
                                                Prob > chi2       =     0.0000
Log likelihood = -5880.4552                     Pseudo R2         =     0.0081

------------------------------------------------------------------------------------
             eh041 |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------------+----------------------------------------------------------------
             aa001 |  -.3323744   .0608015    -5.47   0.000    -.4515431   -.2132056
             aa004 |  -.0066495   .0024361    -2.73   0.006    -.0114241   -.0018748
             ba016 |   .0874215      .0311     2.81   0.005     .0264666    .1483763
             ca001 |          0  (omitted)
             ea104 |  -.0065987   .0030108    -2.19   0.028    -.0124998   -.0006976
             eb001 |   .0052226   .0671317     0.08   0.938    -.1263531    .1367982
             eb002 |  -.2247139   .0795934    -2.82   0.005    -.3807141   -.0687136
             ec023 |  -.2138397   .0323412    -6.61   0.000    -.2772272   -.1504521
             dummy |   .0544896   .0502734     1.08   0.278    -.0440444    .1530237
-------------------+----------------------------------------------------------------
             /cut1 |  -2.209866   .2457687                     -2.691564   -1.728168
             /cut2 |   .8122446   .2445497                       .332936    1.291553
             /cut3 |   3.017259   .2680799                      2.491832    3.542686
------------------------------------------------------------------------------------

[Joseph Coveney]

First, your mode doesn't seem to match up well with the structure of your data, which appears to represent two observations, one in 2004 and the second two years later, for each "id". On that basis alone, I wouldn't report anything from the fitted model.

Second, it looks like some kind of specification search (you didn't realize that two of the candidate predictors are collinear?) If so, then is it the ultimate intention to report what you got with this initial model?

Third, if you're going report anything, be sure to mention the remarkable amount of missing data on each of the variables (that is, if your Example of the dataset is in any way representative of the whole). That would allow your audience to temper any interpretation.