2l.pmm and too many random effects

[james-gwinnutt]

Hi Stef,

I am trying to impute longitudinal missing data using the 2 level procedures in mice. One of the variables only takes certain values, but there are many of these values and so doesn't make sense to treat it as a discrete outcome. I wanted to use 2l.pmm to impute this variable. When I try this I get the error:

Error: number of observations (=4153) <= number of random effects (=17094) for term (1 + v1+ v2 + v3 + v4 + v5 + v6 + v7 + v8 + v9 + v10| id); the random-effects parameters and the residual variance (or scale parameter) are probably unidentifiable

I guess this is because it is trying to calculate random intercepts and slopes for each individual and there are not enough observations (max obs = 3, min = 2 per person). If I go to the prediction matrix and switch the values from 2 to 1 (i.e. removing the random slope) the imputation works. I was wondering whether there is any point to doing this, or whether I should give up on trying to impute in a multi-level manner given that there are not many observations per person? It seems to me that the other variables that can be imputed with 2l.norm will benefit from the 2 level structure, and only the variable that I am trying to use 2l.pmm will loose this advantage (but the imputed values will be possible values).

I hope this makes sense. Please let me know if I can clarify anything. Thanks a lot for your time reviewing this query.
James

[stefvanbuuren]

I suspect the 2l.pmm() model is too general for your data. My general advice is to carefully build up the multilevel model, e.g., in the way I described it in http://stefvanbuuren.name/fimd/sec-mlguidelines.html. Start simple, inspect the solution, and only complicate the model when essential features are not present in your imputations.