I'm evaluating the effectiveness of an intervention. We have a sample of 129 participants. We analyse the primary outcome one month after study inclusion by a linear mixed-effects model with the outcome at baseline and group (intervention / comparision) as fixed effects. A random intercept is modelled per participant (to account for the non-independence of participants(parents)). I work within R.
Unfortunately, our intervention and comparison groups' characteristics differ substantially. Therefore, we need to account for the unbalanced samples. At the same time, our sample is rather small. If we limit the analysis to complete cases, we loose up to 16% of participants.
Therefore, I need to 1) impute the missing data and 2) perform propensity score weighting 3) pool the imputated and weighted data for the final regression analysis.
- Multiple Imputation: For the multiple imputation, I relied on the "Multiple Imputation using R" document by Sarah R Haile (file:///C:/Users/pybiwu60/Downloads/mi_intro20191001-1.pdf) - so far so good; no issues.
2 + 3) Propensity Score Weighting + Pooling: When comparing different weighting methods, overlap weights outperform the "standard approach" of IPW. So, overlap weights would be the way to optimise our groups' balance.
With the R package PSweight, overlap weights can be calculated and added to the completed imputed dataframes. But then I can't pool the regression analysis.
Has anyone an suggestion on how I could use PSweight with imputed datasets and then pool regression results?
Thank you!
When considering IPW, the MatchThem package offers the function "weightthem", which "creates" weights (IPW) in all imputed datasets, and then the regression analysis can be pooled with the "with" and "pool" functions. I was able to do so.