Survival analysis with missing values for some variables

Question

I'm performing a survival analysis. I did the univariant Cox regression model for all my variables (about 40). I even performed the univariant analysis in all those variables that had missing values for some patients at random (about 8 variables, which includes for example the BMI, the CRP or the microbial growth in cultures). None of the variables that had missing values showed significance in this analysis.

It would be ok if I didn't include this variables in the multivariant analysis (since if I only include the patients that have complete date, I loss a lot of sample size)?

Answer 1

I think a much better solution is to do multiple imputation. This can be implemented in major software packages. I know R and SAS have extensive capabilities here, and I am pretty sure that other packages do too (but people who know Python or SPSS or whatever, feel free to correct me).

MI imputes multiple variations on the missing data and then lets you combine the analysis on the imputed sets. This gets rid of most of the problems with the more mundane solutions to missing data, such as variable deletion (which you ask about) or casewise deletion (which you want to avoid) and also mean or median imputation, which have problems.

However, the validity of any solution to missing data depends on the reason for missingness. There are three sorts: Missing completely at random (missing for completely random reasons having nothing to do with anything). Missing at random (missing for reasons that can be accounted for using the variables that are complete) and missing not at random (aka non-ignorable nonresponse). This last one is, as you might guess, much trickier to deal with.