JAV with MICE & Poisson regression?

[melissaja...@gmail.com]

Hi Jonathan,

I am looking to perform MI  using MICE prior to analysing a modified Poisson model with robust variance. The data is MAR. Both restricted cubic spline terms and interactions are to be included in the analysis model. Specifically, the interactions are between the spline term & a multi-level categorical variable. I am unsure whether to use just another variable or passive imputation to deal with the non-linear terms and interactions. As far as I know the JAV approach is only advocated in the context of a linear regression analysis model (or possibly logistic), multivariate normal distribution imputation model and MCAR data. A categorical variable within the interaction adds an additional layer of complexity. As multiple interactions are to be assessed, imputing by subgroup is not realistic. I realise using passive imputation will likely bias the interactions towards the null but this may be a more appropriate approach given my context. Do you have any advice on this?

Thank you