I am collecting individual participant data to perform meta analyses. Some of the cohorts did not collect data on potential confounders. I am wondering what is the best way to deal with this missingness.
1) I could delete these cohorts from multivariable analyses, however, this would be wasteful of data and would make the univariable analyses non-comparable to the multivariable analyses.
2) I could provide sensitivity analyses depending ong the data variability.
3) I could use multiple imputation to impute the missing variables in one cohort using the data from the other cohorts.
Does anyone have experience in dealing with such issues? Can anyone tell me the validity of multiple imputation when a variable has 100% missingness in one cohort but not in the whole dataset?
As I understand it, MI expects the variable to be missing at random. It might be completely at random, or subgroups of patients might have higher rates of missing data, but it has to be random.
So you have two options, and you might want to try it both ways:
* use all variables and only those cohorts with all variables
* use all cohorts, and only those variables in all cohorts
Distribution attributes of the available data might be used to simulate the data series and produce the missing co-founder(s). Just an idea I thought might help find a fix.
To answer my own question, see https://www.ncbi.nlm.nih.gov/m/pubmed/25663182/ and package micemd.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics