Table 3.

Strategies for exploring reasons for failed imputation procedures

Strategy	Problem identified
Remove variables from the imputation model in turn	If the model runs successfully after omitting a particular variable, this might provide some insight into which variable(s) is causing the problem
Create cross-tabulations of categorical variables in the imputation model (such as that shown in Table 1)	Look for sparse or empty cells as these may be causing perfect prediction. It may be necessary to explore patterns across > 2 variables, as perfect prediction can occur for strata produced by combinations of multiple variables
Explore correlations between variables	This can help identify possible sources of collinearity
Examine any output the software produces prior to breakdown of the MI procedure e.g. interim estimates of model parameters	Look for signs of collinearity such as large standard errors and unstable coefficients across iterations. Omission of variables from a model might also signal perfect prediction or collinearity. If the imputation procedure iterates for a substantial amount of time, it might be advisable to run a small number of iterations in order to obtain some output
For problems with MICE, the univariate imputation models can be tested outside the MICE framework by fitting models to observed data (i.e. complete cases)	Check whether the software removes any variables or issues warnings when fitting the univariate models (as these error messages might provide information that is not provided after imputation model failure). When fitting the univariate models, it is also possible to use additional diagnostics such as the variance inflation factor, which provides an indication of whether standard errors are inflated due to collinearity [22]