RCT with missing covariate and outcome data.

[Ronnie Pingel]

Hi,

I am analysing a stratified RCT with four treatment arms, where 150 patients were allocated to each arm.

I am trying figure out if a simple model (which would be more interpretable

for clinicians) rather than a more complicated MI-model will be sufficient.

We have 15-20 baseline covariates (0-10% missing) and one outcome variable with missing data (25%). 

Would the following be a reasonable approach you think?

1. Use imputed mean and missing indicator method to handle missing data for the baseline covariates (as would be the case with no missing outcome).

2. Use imputed mean baseline covariates with the missing indicators in a logistic regression to fit a model for the missingness.

3. Using the weights from 2, re-weight the regression for the outcome, i.e. E(Y|Treatment, Stratum) = b0 + b1*Treatment + b2*Treatment2 + b3*Stratum.

4. Conduct a sensitivity analysis using the pattern mixture model. 

I guess also that Maximum likelihood or MI we would also to a greater extent rely on distributional assumptions compared to this approach.

Best regards,

Ronnie