Table 2.

Table of recommendations for the use of multiple imputation, regression imputation and missing indicators in the development and deployment of clinical prediction models.

Recommendations
Assuming no missing data will be present at deployment, multiple imputation (including the outcome) or regression imputation (omitting the outcome) are recommended as the best strategies and estimates of predictive performance was comparable between the two.
Where missingness is allowed at deployment, and multiple imputation is impossible at deployment (e.g. where the original development data, or sufficient computational power is not available), regression imputation can be used as an alternative.
Always omit the outcome from the imputation model under regression imputation.
Where data are assumed to be MNAR-X or MNAR-Y and missingness is allowed at deployment, the inclusion of a missing indicator can offer marginal improvements in model performance, and does not harm performance under MCAR or MAR mechanisms
The use of missing indicators under MNAR-Y can harm model performance when missingness is not allowed at deployment, and is not recommended