Dealing with missing data for linear mixed effects model (APIM) in SPSS

Question

I plan to use SPSS for an actor-partner interdependence model (APIM). My sample will have a significant amount (possibly >20%) missing data on one of the predictor variables (edit: I'm emphasizing predictor because I just saw something about ML only being useful when missing data are on the outcome variable; not sure if this is true). I'm somewhat familiar with issues related to missing data mechanisms (I believe my data are MAR, though I haven't systematically examined it yet) and the choice of different strategies, e.g., it seems there is a consensus that both multiple imputation (MI) and maximum likelihood (ML) are best practices. My questions are about implementation.

1. I've been told that SPSS does ML, but can anyone confirm this is possible when using APIM, and if so, is it the default?

2. I've also been told that SPSS can do MI with a particular add-on. But how does pooling work? Can an APIM be estimated following MI? It seems that ML is the more convenient option when using SPSS, but I'm curious if there is an automated process for MI like there is in the R mice package (though I'm also not sure with mice whether it's possible to do multilevel models, etc.)

Answer 1

A (RE)ML approach to missing data only applies to outcome variables. I.e. to apply it to missing predictors, you would not just have a model for your outcome of interest, but instead a joint model for the outcome of interest and the covariates (at least the ones that have some missingness). One can certainly "trick" software (like SAS or R into doing that without having to find a dedicated package for it, no idea about SPSS) into doing this kind of modeling, but it is a lot more challenging / needs a lot more care that just modeling one outcome. I'd guess that SPSS like SAS or R will probably not do anything like that by default, but rather (inappropriately) fit with only the complete records or refuse to fit the model.

A simpler to apply approach is multiple imputation (MI) creating $M$ imputed datasets, followed by fitting your model of interest on each of the $M$ datasets and combining the quantity of interest across the $M$ datasets into a single overall results e.g. using Rubin's rule. The good thing about MI is that there are good functions/packages (e.g. the Amelia package in R, again, not sure what SPSS offers) in any modern statistical software for imputing all types of data in quite a flexible way, e.g. it would not really matter whether your covariates of interest are continuous, categorical, ordinal or something else. Additionally, it is usually a little less "dangerous" (i.e. you are less likely to badly violate some model assumptions) than joint modeling.

Answer 2

SPSS does not have a procedure specifically designed for an APIM model, but if you're intending to use a linear mixed models approach, then ML estimation using the MIXED procedure is available (the default is restricted maximum likelihood or REML).

There is a multiple imputation procedure to create imputed data, and MIXED does produce pooled estimates for model parameters. Pooling is done based on Rubin's rules, as is standard in analyses of multiply-imputed data.