Direction of bias from missing data in logistic regression

[Ant Kay]

I've performed a multivariable logistic regression analysis and am now trying to assess the effect of missing data on my effect estimate. I would like to be able to report the direction that the OR would be biased in by the missing data.

Specifically, data is missing for ~15% of my primary exposure. I have create a dummy variable of missing/not missing for the exposure, and have looked at associations between missingness and my other covariates (individually).

I've been told I could do a multivariable analysis of variables associated with missingness to help determine the direction of the bias, but I'm not exactly sure how to go about this, or interpret the results.

Does anyone have any experience with this, and know how you can use information from the above analysis to work out the direction of any bias?

Thanks for your help.

[Marcos Almeida]

Welcome to the Stata Forum/ Statalist.

You may start with the suite of - misstable - commands.

[Andrew Lover]

I'd also try -mvpatterns- (SSC), then do a crude sensitivity analysis where you a) code all missing exposures as '0', rerun your model then b) recode all missing exposures as '1' and rerun, and then compare all the ORs.

Finally, you could use multiple imputation to assess full-case data, but you'll need to think carefully about the nature of your missing exposure (MAR vs MCAR).

https://www.ssc.wisc.edu/sscc/pubs/stata_mi_decide.htm

(Just to note the recoding scheme in the link above is somewhat similar to the sensitivity analysis but has a different goal).

[Carlo Lazzaro]

Ant:
as an aside to previous helpful comments:
- you can also consider the user-written programme -mcartest- (type -search mcartest- from within Stata to access) to check MCAR vs MAR;
- conversely, it your missingness is informative (MNAR, which is empirically untestable), things get trickier;
- you should also investigate the pattern of your missingness (see -misstable patterns-).

[Maarten Buis]

Another thing to look at is the association between your missingness variable and your dependent variable. If they are not associated, then the missingness will not bias your result.