Seeking Stata command suggestion for my outcome variable

[Tariku Getaneh]

Dear Statalists,
Would you please suggest the right Stata command for my typical type of dependent variable? or if any transformation mechanism you may know would be helpful.
I have a panel household survey and I have a categorized dependent variable taking a value of1 if take immediate action, 2 if the next day, 3 if after two days, 4 if between three days and a week, 5 if after a week, and 6 after more than a week.
For now, I used a linear fixed effect regression and found the coefficient of regressor variable of interest X, which is a dummy variable, is 0.19. I interpreted it as 0.19 level delay in those with a value of X=1.
The interpretation as a level doesn’t appear to be reader friendly, do you know of any useful transformation so I can say x percentage or percent delay or any other suggestion.
I am available to clarify if anything is unclear.
BW,

[George Ford]

as a first thought, your DV is ordered, so ologit for estimation. (maybe poisson if you think of the DV as days, but kinda iffy in your case.)

I suspect you'd want to estimate the most likely outcome for X = 1. The likely outcome is the one with the highest probability, which Stata will estimate.

[Tariku Getaneh]

Hi George,
Thank you so much, following your suggestion I used feologit user written command given the panel structure of my data.
How would I interpret the coefficient on my regressor variable of interest X (Which is a dummy=1 YES, 0=NO) . Suppose the ODDS RATIO value is 1.280.

Kind regards

[George Ford]

what does this get you?

logitmarg, dydx(X1)

[Tariku Getaneh]

Here is what I get

 
Marginal    effects    at the average        N. of    observations=    2730
                N. of    panel units =    72
                        
        Margin   Std. err.    z    P>z    [95% conf.    interval]
                        
    1    -.066775   .0362361    -1.84    0.065    -.1377965    .0042464
    2    .0007488   .0004063    -1.84    0.065    -.0015452    .0000476
    3    .0247592   .0134358    1.84    0.065    -.0015745    .0510929
    4    .0228187   .0123828    1.84    0.065    -.0014511    .0470884
    5    .0136723   .0074194    1.84    0.065    -.0008695    .0282141
    6    .0062736   .0034044    1.84    0.065    -.000399    .0129462

[George Ford]

read

http://kevinstaub.com/ewExternalFiles/2020_sj.pdf

add threshold as an option to the feologit command.
then
margins, at(X1 = (0 1))

this will give you the probabilities of the outcomes at X1 = 0 or 1.

if the sign of X1 is negative, it will push the outcomes to the smaller values (and vice versa).

margins, dydx(X1)

will you give the marginal effect for each outcome.

they will vary by outcome.

webuse nhanes2f , clear
sample 20

feologit health age age2 sex race weight iron diabetes, group(location) threshold


Fixed-effects ordered logistic regression

N. of obs. (inc. copies) =    28386
N. of observations       =    2066
N. of panel units        =    62
Wald chi2(10)            =    1381.02
Prob > chi2              =    0.0000
Log conditional likelihood = -11225.712  Pseudo R2                =    0.2888
(Std. err. adjusted for 62 clusters in    location)
    
Robust
health  Coefficient  std. err.      z    P>z     [95% conf.    interval]
    
age   -.0323837   .0199623    -1.62   0.105    -.0715091    .0067417
age2   -.0000407   .0002122    -0.19   0.848    -.0004566    .0003751
sex   -.0393662   .1097884    -0.36   0.720    -.2545475    .175815
race   -.4247452   .1274421    -3.33   0.001    -.6745271    -.1749634
weight   -.0020998   .0032249    -0.65   0.515    -.0084205    .004221
iron    .0033996    .001351     2.52   0.012     .0007516    .0060476
diabetes   -1.229833    .186509    -6.59   0.000    -1.595384    -.8642822
    
/cut1           0  (constrained)
/cut2    1.636488   .1077295                      1.425342    1.847634
/cut3     3.14255   .1198344                      2.907678    3.377421
/cut4    4.374634   .1338231                      4.112346    4.636923

margins, at(diabetes = (0 1))


1._at: diabetes    = 0
2._at: diabetes    = 1

                        
        Delta-method
    Margin    std. err.    z    P>z    [95% conf.    interval]
                        
_predict#_at 
1 1    .0598358    .0309922    1.93    0.054    -.0009078    .1205794
1 2    .172249    .0848464    2.03    0.042    .0059532    .3385448
2 1    .1737577    .0678027    2.56    0.010    .0408669    .3066485
2 2    .3117595    .0645393    4.83    0.000    .1852648    .4382542
3 1    .3125592    .0338245    9.24    0.000    .2462645    .378854
3 2    .3031065    .0494674    6.13    0.000    .2061522    .4000608
4 1    .2413129    .0400506    6.03    0.000    .1628152    .3198106
4 2    .1355122    .0544092    2.49    0.013    .0288722    .2421522
5 1    .2125343    .0914881    2.32    0.020    .033221    .3918476
5 2    .0773728    .0449419    1.72    0.085    -.0107117    .1654572

margins, dydx(diabetes)


                        
        Delta-method
    dy/dx    std. err.    z    P>z    [95% conf.    interval]
                        
diabetes    
_predict    
1    .0731261    .0371754    1.97    0.049    .0002637    .1459885
2    .13284    .0303017    4.38    0.000    .0734498    .1922303
3    .0634835    .0643247    0.99    0.324    -.0625907    .1895577
4    -.0818602    .0462191    -1.77    0.077    -.172448    .0087276
5    -.1875895    .0566985    -3.31    0.001    -.2987166    -.0764624

You see that having diabetes increases the probability of the 2nd outcome by 0.1328 (a change to 0.31176 from 0.1737577).

[Tariku Getaneh]

Thanks a lot George.As I said I used a linear fixed effect initially and the results from the ordered regression appears to be consistent. So, my intention is to use ordered logit result as a robustness check.
If I may ask, in general what is the typical way of presenting results from ordered logit regression? the "Marginal effects" only and/or the odds ratio. Alternatively, marginal effects in the main text and odds ratio in the appendix?

Thanks very much for your time.

Best regards,

[George Ford]

The linear model is modeling the mean, so it will produce the same sort of thing, though less detailed. I think it'd be a stretch to do hypothesis testing in the linear model given the ordered nature of the DV. I'd avoid it and stick with ologit.

As for ologit, look up some papers and see how researchers are describing the results. Use that as a guide, but I encourage you to present what you think is interesting in a way you think is interesting.

[George Ford]

https://liberalarts.tamu.edu/sociology/wp-content/uploads/sites/9/2021/07/notes_logit_ordered_20210713.pdf
https://www3.nd.edu/~rwilliam/stats3/Ologit01.pdf
https://users.stat.ufl.edu/~aa/articles/agresti_tarantola.pdf
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9569299/

go to ssrn.com and search for "ordered logit"

[Tariku Getaneh]

Thank you so much. I will have a look at each document.

Best regards,
TG