Getting contrasts after regression with multiple imputation and complex survey design / mimrgns

[Kate Nguyen]

Hi,

I am using a large-scale survey dataset with a complex survey design. I have also used multiple imputation to account for missing data.

Now, I am trying to run a regression model on the imputed data, taking into account its complex survey design. I would then like to conduct a hypothesis test to determine whether people at, say, the 25th and 75th percentiles of BMI had the same value of the outcome variable. However, I am not sure how to conduct this test, given that my data are both multiply imputed and have a complex survey design.

To conduct this analysis in a non-survey, non-mi context, I think I would use something like:

reg smoker BMI
margins, at((p25) BMI) at((p75) BMI) contrast(atcontrast(r._at) marginswithin)

To conduct the regression in the mi/svyset context, I think I would use this for the regression model:
mi estimate: svy: reg smoker BMI

However, I am not sure how to conduct the postestimation test of whether people at the 25th and 75 percentiles are the same on the outcome. Is there a way to do this using ssc's -mimrgns-, or some other method? (I can see in the current -mimrgns- help file a note that it doesn't support contrast, although I also found this thread here noting that mimrgns now has support for contrast operators:

http://www.statalist.org/forums/foru...updated-on-ssc

Many thanks in advance for any help. I very much appreciate it.

[daniel klein]

Currently, mimrgns does not support the atcontrast() suboption. From a quick glance, I think it should, and I will implement this in an updated version when I find the time. Until then, you might want to implement this yourself.

capture program drop mimargins
program mimargins , eclass properties(mi)
    version 12.1
    svy : regress attack bmi
    margins , at((p25) bmi) at((p75) bmi) ///
        contrast(atcontrast(r._at) marginswithin) ///
        vce(unconditional) post
end

mi estimate : mimargins whatever

Note that I added the vce(unconditional) option that you probably want to use with survey data.

Best
Daniel

[Kate Nguyen]

Dear Daniel,

Thank you so much for this very helpful response. I was able to run this code and generate the contrasts. I greatly appreciate it.

Many thanks,
Kate