Secondary exposures and outcomes

[regina...@gmail.com]

Dear Jonathan,

I'm currently working on an analysis in which I am interested in the effect of an exposure, measured in four different ways (all binary: self-reported depression , antidepressant use, hospital admission, and the fourth one is a combination of the three: any indication of depression vs. none). All of these variables are fully observed. I'm also looking at three different survival outcomes (stroke, myocardial infarction, and a combination of those) which are also fully observed. In the final Cox PH model of the any.vs. none exposure variable, I am also interested in interactions between my exposure and some covariates. I have missing covariates with missingness up to 15%. Since I include a number of covariates in my fully adjusted model, overall missingness is high (~40%). My sample size is ~450,000.

I've been trying to find the best way of approaching the multiple exposure and outcome variables in a multiple imputation model. Is it sensible to include multiple outcomes in the same imputation model? Would you recommend running separate imputations for each of the exposure variable? I've struggled to find literature on what to do when there are multiple exposure and outcome measures. I would very much appreciate your advice.

Many thanks in advance.

Kind regards,

Regina

[Jonathan Bartlett]

Hi Regina

Good question. My personal view would be to tackle each of the exposure/outcome combinations as a separate exercise, and do the imputation separately for each combination of these. This way you can ensure that each set of imputations is appropriate for that particular substantive/outcome model. Given that you say you have interactions in the model, you may want to consider doing the imputation using my smcfcs package, which can impute missing covariates for Cox models (and others) in a way which is compatible with your specified Cox model (which includes interactions). R version is here (https://cran.r-project.org/web/packages/smcfcs/index.html) and Stata version is in SSC (https://econpapers.repec.org/software/bocbocode/s457968.htm).

Best wishes

Jonathan

--
You received this message because you are subscribed to the Google Groups "Missing Data" group.
To unsubscribe from this group and stop receiving emails from it, send an email to missing-data+unsubscribe@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Regina Prigge]

Dear Jonathan,

Thank you very much for your prompt response and advice. I had hoped that I could safe some time by combining different exposure/outcome versions in one imputation model but I will follow your advice. Your help is much appreciated! And yes - I aim to use the smcfcs R package to run the imputations. 

Best wishes,

Regina

You received this message because you are subscribed to a topic in the Google Groups "Missing Data" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/missing-data/XNWUXjK0B-g/unsubscribe.
To unsubscribe from this group and all its topics, send an email to missing-data+unsubscribe@googlegroups.com.