Report OR with clogit, on imputed data

[Adrian Salinas Fredricson]

Hello,

I am conducting a case control study on individuals matched on age and gender. I want to run a conditional logistic model and control for three variables named; degurbax, utbkat and landkat2 (type of living area, educational level and country of birth).
All these variables that are being controlled for have missing data, on which I want to impute. However, when i run the clogit on the mi estimate, STATA does not report OR, but a coefficient. The code I have been using is the following;


mi set mlong
mi register imputed utbkat landkat2 degurbax
mi impute chained (ologit) utbkat (mlogit) landkat2 (mlogit) degurbax, add(20) rseed(1234) savetrace(trace1,replace)
mi estimate: clogit fall M00 i.degurbax i.utbkat i.landkat2, group(grupp) or

I will attach two printscreens, and as you can see there is no difference when I ask STATA to report OR, from when I am not.

Is there anyone who knows what I am doing wrong? When I use clogit without "mi estimate"-prefix it DOES report OR. Is it valid to use this output?

Kind regards,

Adrian

[daniel klein]

You want

mi estimate , or : clogit ...

to display the final MI estimates in exponentiated form.


You probably also want to account for the nested nature of your data during the imputation process. As an absolute minimum, the group variable should be included in the imputation model. Imputing by(group) or reshaping the data to account for within-group correlations are alternatives. Moreover, you must include the outcome, fall, in the imputation model; failing to do so will bias the associations of the predictors with the outcome towards zero. More generally, all variables that are in the substantive analyses should also appear in the imputation model.

[Adrian Salinas Fredricson]

Thank you so much Daniel, incredibly kind of you. Also thanks for pointing out that I need to account for both my outcome variable and the group variable in my imputation.

I suppose that would look something like this?

mi set mlong
mi register imputed utbkat landkat2 degurbax kon fall
mi impute chained (ologit) utbkat (mlogit) landkat2 (mlogit) degurbax fall kon, add(20) rseed(1234) by(grupp)

When I register kon (gender) and fall (binary outcome variable) as imputed, does it matter that neither of those two variables contain missing data?

Again, thank you for your kind help!

Sincerely,
Adrian

[daniel klein]

When I register kon (gender) and fall (binary outcome variable) as imputed, does it matter that neither of those two variables contain missing data?

If those variables do not have missing values, then you should either not register them or register them as regular. In the imputation model, these variables go to the right-hand side of the equals sign

mi impute chained (ologit) utbkat ... = i.fall i.kon , add(20) by(group)

I am thinking, as the controls are matched on gender, that there is no point in including gender in my imputation model, as I am using "by grupp", and as both controls and cases within each group are the same gender.
Is the coding aswell as the reasoning valid?

If you omit gender, you are essentially assuming that gender is not correlated to any other variable in the model. Even if there is no correlation between gender and group, there might still be a correlation between gender and, say, utbkat, and your imputation model should account for that correlation; if only to improve the imputed values for utbkat.

[Adrian Salinas Fredricson]

mi impute chained (ologit) utbkat ... = i.fall i.kon , add(20) by(group)

Thank you yet again Daniel for taking your time! This has been tremendously helpful and I can now hopefully continue with the project!

Kind regards
Adrian