gologit2 output interpretation

[Yawo Kokuvi]

I am fitting a model predicting a person's perceived odds of having cardio-risk (measured via 3 -categories)
-- No Risk
-- Low Risk
-- High Risk

Brant test indicated need for ordered logit approach, so I am using gologit2 for the analysis. I've consulted the documentation on gologit2 and read Dr. William's materials as well, and attempted an interpretation of the output, but I want to be sure I am making the right contrasts/comparisons appropriately:

Here is my command and output, and my interpretation follows:

set more off
gologit2 cardiorisk age i.sex i.fam_hist yrsusa ///
i.lowcarb i.friends_sameethnic educ i.friends_risk ///
 , auto lrforce


------------------------------------------------------------------------------
Testing parallel lines assumption using the .05 level of significance...

Step  1:  Constraints for parallel lines imposed for age (P Value = 0.6247)
Step  2:  Constraints for parallel lines imposed for i.fam_hist (P Value = 0.4300)
Step  3:  Constraints for parallel lines imposed for i.sex (P Value = 0.3468)
Step  4:  Constraints for parallel lines imposed for yrsusa (P Value = 0.2823)
Step  5:  Constraints for parallel lines are not imposed for
          educ (P Value = 0.00081)
          i.lowcarb (P Value = 0.00543)
          i.friends_sameethnic (P Value = 0.00227)
          i.friends_risk (P Value = 0.01911)

------------------------------------------------------------------------------

Generalized Ordered Logit Estimates             Number of obs     =        496
                                               
---------------------------------------------------------------------------------------
         cardiorisk   |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
----------------------+----------------------------------------------------------------
none  
                      |
                educ  |   .0267677   .0509569     0.53   0.599     -.073106    .1266415
                      |
            lowcarb   |   .9560091   .2417372     3.95   0.000     .4822129    1.429805
                      |
   friends_sameethnic |   .1248707    .392147     0.32   0.750    -.6437233    .8934647
   
                age   |   .0193308   .0113127     1.71   0.087    -.0028417    .0415033
           sex_female |  -.2504234    .195122    -1.28   0.199    -.6328555    .1320088
            fam_hist  |   .6090692   .1892519     3.22   0.001     .2381422    .9799962
            yrsusa    |  -.8336397   .3744464    -2.23   0.026    -1.567541   -.0997383
     
                _cons |  -1.148632   .9356265    -1.23   0.220    -2.982427    .6851618
----------------------+----------------------------------------------------------------
low                   |
 
        friends_risk  |   2.038412   .3153763     6.46   0.000     1.420285    2.656538
        
                age   |   .0193308   .0113127     1.71   0.087    -.0028417    .0415033
           sex_female |  -.2504234    .195122    -1.28   0.199    -.6328555    .1320088
            fam_hist  |   .6090692   .1892519     3.22   0.001     .2381422    .9799962
            yrsusa    |  -.8336397   .3744464    -2.23   0.026    -1.567541   -.0997383
                      |
                _cons |  -.3649326   .9573218    -0.38   0.703    -2.241249    1.511384
                
=======================================================================================

MY ATTEMPTED INTERPRETATION

Constrained Factors - Interpretation

- Higher cardiovascular risk perception is higher for those with a family history of cardiovascular illness
- By contrast, lower risk perception associated with increasing US residence


Unconstrained Factors (No Risk vrs Low/High Risk) - Interpretation

- Eating low carb diet is associated with having NO risk perception, compared to having at least (low + high) risk perception


Unconstrained Factors (No+Low vrs High Risk) - Interpretation

- a persons self-perceived cardio risk is associated that of their friends: having friends who are similarly
perceived as having lower risk perception (none + low) compared to high, tends to positively influence one's risk perception.


I would very much appreciate some thoughts on whether i am interpreting this output correctly, particularly with regard to the comparisons between the various categories of the DV.

Sincerely, cY

[Richard Williams]

Something seems off in that the two panels contain different independent variables, e.g. educ and lowcarb should be in the 2nd panel too. Make sure you are posting the correct output.

[Yawo Kokuvi]

Thanks very much for the quick response.

Yes, I am sorry I made an error and copied one of the preliminary models I have been working on, and messed things up.
Here is the candidate model I am focusing on, and would appreciate some feedback with regard to the comparisons:


cheers, CY

================================================== =============================

. gologit2 cardiorisk i.sex age educ i.lowcarb ///
>  friends_risk i.fam_hist friends_sameethnic yrsusa, auto lrforce

------------------------------------------------------------------------------
Testing parallel lines assumption using the .05 level of significance...

Step  1:  Constraints for parallel lines imposed for age (P Value = 0.6594)
Step  2:  Constraints for parallel lines imposed for 1.sex (P Value = 0.3154)
Step  3:  Constraints for parallel lines imposed for yrsusa (P Value = 0.2723)
Step  4:  Constraints for parallel lines imposed for 1.fam_hist (P Value = 0.1162)
Step  5:  Constraints for parallel lines are not imposed for 
          educ (P Value = 0.00063)
          1.lowcarb (P Value = 0.00621)
          friends_risk (P Value = 0.01461)
          friends_sameethnic (P Value = 0.00316)

Wald test of parallel lines assumption for the final model:

 ( 1)  [none]age - [low]age = 0
 ( 2)  [none]1.sex - [low]1.sex = 0
 ( 3)  [none]1.fam_hist - [low]1.fam_hist = 0
 ( 4)  [none]yrsusa - [low]yrsusa = 0

           chi2(  4) =    4.69
         Prob > chi2 =    0.3205

An insignificant test statistic indicates that the final model
does not violate the proportional odds/ parallel lines assumption

If you re-estimate this exact same model with gologit2, instead
of autofit you can save time by using the parameter

pl(age 0b.sex 1.sex 0b.lowcarb 0b.fam_hist 1.fam_hist yrsusa)

------------------------------------------------------------------------------

Generalized Ordered Logit Estimates             Number of obs     =        504
                                                LR chi2(12)       =     123.15
                                                Prob > chi2       =     0.0000
Log likelihood = -454.64741                     Pseudo R2         =     0.1193

 ( 1)  [none]age - [low]age = 0
 ( 2)  [none]1.sex - [low]1.sex = 0
 ( 3)  [none]1.fam_hist - [low]1.fam_hist = 0
 ( 4)  [none]yrsusa - [low]yrsusa = 0
------------------------------------------------------------------------------------
        cardiorisk |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------------+----------------------------------------------------------------
none               |
               sex |
                1  |  -.2290532   .1920955    -1.19   0.233    -.6055535    .1474471
                   |
               age |   .0189786   .0111708     1.70   0.089    -.0029157    .0408729
              educ |   .0291178   .0504111     0.58   0.564    -.0696861    .1279216
                   |
           lowcarb |
                1  |   .8694071   .5397422     1.61   0.107    -.1884682    1.927282
                   |
      friends_risk |   .2055355   .3585004     0.57   0.566    -.4971124    .9081835
                   |
          fam_hist |
                1  |   .6009144   .1867137     3.22   0.001     .2349623    .9668665
                   |
friends_sameethnic |   1.025543   .2392585     4.29   0.000     .5566055    1.494481
            yrsusa |   .0013285   .0004581     2.90   0.004     .0004306    .0022265
             _cons |   -1.18831    .898648    -1.32   0.186    -2.949628    .5730076
-------------------+----------------------------------------------------------------
low                |
               sex |
                1  |  -.2290532   .1920955    -1.19   0.233    -.6055535    .1474471
                   |
               age |   .0189786   .0111708     1.70   0.089    -.0029157    .0408729
              educ |   -.140776   .0498742    -2.82   0.005    -.2385277   -.0430244
                   |
           lowcarb |
                1  |  -.7214967   .5889009    -1.23   0.221    -1.875721    .4327278
                   |
      friends_risk |  -.7764095   .3344735    -2.32   0.020    -1.431966   -.1208535
                   |
          fam_hist |
                1  |   .6009144   .1867137     3.22   0.001     .2349623    .9668665
                   |
friends_sameethnic |   2.068744   .3139869     6.59   0.000     1.453341    2.684146
            yrsusa |   .0013285   .0004581     2.90   0.004     .0004306    .0022265
             _cons |  -.3629457   .9207698    -0.39   0.693    -2.167621     1.44173
------------------------------------------------------------------------------------

and here is my attempted interpretation:




Constrained Factors - Interpretation

-- Family History: Having a family history of cardio risk associated is positively associated with no risk perception
-- Age: The older one is, the higher the probability of no risk perception

Unconstrained Factors (No Risk vrs Low/High Risk) - Interpretation

- Having friends of the same ethnic group is associated with having NO risk perception, compared to perceiving at least some risk (low + high risk)

Unconstrained Factors (No+Low vrs High Risk) - Interpretation

-- Education: higher levels of education is associated with having lower risk perception ( (none + low), compared to higher
-- Friends Risk: a person’s self-perceived cardio risk is associated that of their friends: respondents are likely to perceive themselves as having lower (none + low) risk (compared to high), if they similarly perceive their friends as having such lowered risk.
-- Being of the same ethnic group as one’s friends is strongly associated with having lower (none+low) risk perception, compared to higher



================================================== ========================================

[Richard Williams]

I just looked quickly but I think your interpretations may be reversed. A gologit model is like a series of binary logits. The first panel is category 1 versus 2, 3. The 2nd is categories 1 and 2 versus 3.

Positive increases say increases in X make it MORE likely that a respondent will be in a HIGHER category. A negative coefficent means that increases in X make it MORE likely that a respondent will be in THIS category or a LOWER one.

So, for example, the positive coefficient for age means that older people are more likely to have higher perceptions of risk. Which, to me, even knowing as little as I do about your research, seems far more reasonable than what you said.

Likewise with family history. If you have a family history of cardio risk, you are more likely to feel you are at risk. Again, far more sensible than what you said.

I think I've tried to make these interpretations clear in my writings but people still screw it up. Because gologit output looks like mlogit output, people think they should interpret gologit like mlogit but they shouldn't. If I ever get wildly ambitious I may try to come up with some form of alternative formatting for the output.

I also think this table formatting can help when presenting results: https://www3.nd.edu/~rwilliam/gologi...2016Table2.pdf

Also see https://www3.nd.edu/~rwilliam/xsoc73994/Margins05.pdf

[Yawo Kokuvi]

Dr. Williams: thanks very much for your explanation - I now get it, and understand the way the contrasts are interpreted. I will brood on it further and get back to you all if I have any further questions. best wishes - cY