Parallel lines assumption of ordered logit regressions with multiply imputed data

[Johanna Turgetto]

Dear all,

I am working with an imputed data set (M=50, N=1359) and use an ordered logisitc regression (ologit) on my ordinal dependent variable wellbeing well82 (5 levels).

local SESvar "SEX_G2 ib2.class i.education"            
mi estimate, post or: ologit well82 `SESvar'
esttab using "Pap1/Model15_wb3_1.rtf", replace ///
            b(3) ar2 not star obslast

This runs without a problem.
Subsequently, I would like to check the parallel lines assumption of ordered logit regressions. I am failing at this and would appreciate your help. I already installed the brant command as suggested.
To get around the problem of testing the assumption, I have already tested out the gologit2 command as well. Unfortunately without success as the command runs into an error right away.

Thank you!

[Johanna Turgetto]

In case someone was faced with the same issue. By luck I found the following thread that will help:

https://www.statalist.org/forums/for...dds-assumption

[daniel klein]

Just a thought: if you impute the missing values using an ordered model, you are basically imposing the assumption of parallel odds. I wonder whether it makes sense to then test whether the assumption is/was justified. I would tend to look at this before then imputation then pick the appropriate model to impute missing values. There is probably less to worry about if the respective variable (outcome) does not have missing values that need to be imputed.

[Johanna Turgetto]

Daniel,
thank you for your thoughts! In the meantime I did run the test as suggested in #2 on my original data (pre imputation) and run into the problem of violation of the assumption.

mi xeq: omodel logit devar indepvars

Hence, I wanted to use gologit2 instead- though the problem of negative probabilities occurs now - but wonder whether the violation of assumption implies I should redo my imputation first of all. I used mi impute chained (and was actually just happy to have a working imputation model as I've never done imputations before and had some issues with convergence).

mi impute chained  ///
    (ologit) well02 v3937 alkohol02 alkohol82 alkohol12 rauch82 rauch12 schul_m BMI82_o BMI12_o BMI02_o  v2039 e25a1h v20l1j v16l2  ///
    (logit) v13a7ane_12 e20g2a v2036  ///
    (pmm, knn(5)) well12 well82 family12 family02 schock_youth death abschluss82 v12a2 ///
                Sport12 Sport02 kidNO KidNO_12 beruf_m scheidung02 scheidung12 e164a1e rv12a1 rauch02 e13f5 v17l7  ///
    (regress) Netto_35 Netto_45 ISEI35 ISEI45 weeklyh weeklyh02 alter_m_geburt ///
    (truncreg) v4171 bdi bdi_12 v4144 v4164 v4149  ///
               v4260 sozinteg sozinteg_12 ichstark ichstark_12 ///
               v14a14se selbswiberuf_12 v4190 v4154 ///
               = i.SEX_G2 [additional varlist] ///
               , add(50) replace rseed(1359) augment noisily showcommand
            

What do you think?
Thank you so much!

[Johanna Turgetto]

My apologies,
I did run the omodel test on my imputed data but then used my original data to use the autofit option to find what variables violate the assumption

 gologit2 well82 `SESvar', autofit difficult

.

Nevertheless, as the results after using gologit2 are mostly negative, I believe there must be a bigger issue?!

[daniel klein]

I find it hard to give specific advice as I do not know your research questions, your data, your results, etc. Generally speaking, do not look at the p-values for your tests of parallel odds; look at the coefficients. Also, ask yourself questions such as: How different are the coefficients, substantially? How would the differences in coefficients change the substantive conclusions that you draw regarding your research questions?

[Johanna Turgetto]

I will do that, thanks again!