We used the Multivariate Imputation by Chained Equations method [ 47] with m = 5 multiple imputations, and with logistic regression and predictive mean matching imputation models respectively for binary and continuous variables.
Polytomous logistic regression and also predictive mean matching regression as imputation models for T-, N- and M-stage yield good estimations on the individual stage value, the stage-specific numbers of cases and the stage-specific survival curves, as long as the amount of missing values is not too high.
For statistical analysis of tumour stage information in cancer registry data, both on the individual and the aggregated level, multiple imputation with chained equations using polytomous regression or predictive mean matching as an imputation model was in this limited simulation study found to be an appropriate method for dealing with missing data in tumour stage.
Incomplete dichotomous variables were imputed using a logistic regression model, while predictive mean matching was used to impute incomplete continuous variables.
The approach we call 'linear imputation model with predictive mean matching' (or just 'PMM') is the same as passive imputation up to the generation of σ * 2 and γ *.
The missing tumour stage values were then treated with multiple imputation with chained equations, using polytomous regression, predictive mean matching, random forests and proportional sampling as imputation models.
For continuous variables, we use predictive mean matching, and for binary variables, we use logistic regression models.
These were complete case analysis (CC), single imputation (SI) using predictive mean matching [ 5], MI fitting separate flexible additive imputation models to each incomplete covariate with predictive mean matching [ 21] (MI-aregImpute), MI using regression switching (MI-MICE) and the addition of predictive mean matching (MI-MICE-PMM) [ 5].
Predictive mean matching was used for the imputation of thickness, logit models for lymph nodes (none vs one or more) and ulceration (yes, no), and multinomial logit models for Clark's level and subsite.
Researchers should spend efforts on specifying the imputation model correctly, rather than expecting predictive mean matching or local residual draws to do the work.
