Inspect multiple imputation model

Check multiple imputation is valid under the proposed imputation model(s) and directed acyclic graph (DAG). Validity means that the proposed approach will allow unbiased estimation of the estimand(s) of interest, including regression parameters, associations, and causal effects.

Usage

checkMI(dep, preds = NULL, r_cra, mdag)

Arguments

dep: The partially observed variable(s) to be imputed, specified as a string (space delimited) or a list
preds: Optional fully observed imputation model predictor(s), specified as a string (space delimited) or a list
r_cra: The complete record indicator, specified as a string
mdag: The DAG, specified as a string using dagitty syntax, or as a dagitty graph object

Value

A message indicating whether multiple imputation is valid under the proposed DAG and imputation model predictor(s)

Details

Imputation model(s) should include all other analysis model variables as predictors, as well as any auxiliary variables. The DAG should include all observed and unobserved variables related to the analysis model variables and their missingness, as well as the complete record ("missingness") indicator.

In principle, multiple imputation is valid if all partially observed variables are unrelated to missingness, given the (fully observed) imputation model predictors. This is determined using the proposed DAG by checking whether all the partially observed variables are 'd-separated' from the complete record indicator, conditional on the imputation model predictors. It is assumed that all the specified imputation model predictors are fully observed and will be used to impute all the specified partially observed variables.

References

Curnow E, Tilling K, Heron JE, Cornish RP, Carpenter JR. 2023. Multiple imputation of missing data under missing at random: including a collider as an auxiliary variable in the imputation model can induce bias. Frontiers in Epidemiology. doi:10.3389/fepid.2023.1237447

Examples

# Example DAG for which multiple imputation is valid, because collider
## variable 'bwt' is not included as a predictor
checkMI(dep="bmi7", preds="matage mated pregsize", r_cra="r",
        mdag="matage -> bmi7 mated -> matage mated -> bmi7
              sep_unmeas -> mated sep_unmeas -> r pregsize -> bmi7
              pregsize -> bwt sep_unmeas -> bwt")
#> Based on the proposed directed acyclic graph (DAG), the partially
#> observed variable(s) and complete record indicator are independent
#> given the fully observed imputation model predictor(s). Hence, multiple
#> imputation methods which assume data are missing at random are valid in
#> principle.

# Example DAG for which multiple imputation is not valid
checkMI(dep="bmi7", preds="matage", r_cra="r",
        mdag="matage -> bmi7 mated -> matage mated -> bmi7
               sep_unmeas -> mated sep_unmeas -> r")
#> Based on the proposed directed acyclic graph (DAG), the partially
#> observed variable(s) and complete record indicator are not independent
#> given the fully observed imputation model predictor(s). Hence, multiple
#> imputation methods which assume data are missing at random may not be
#> not valid.
#> 
#> Consider using a different imputation model and/or strategy (e.g.
#> not-at-random fully conditional specification).  For example, the
#> partially observed variable(s) and complete record indicator are
#> independent if each of the following sets of variables are used as
#> predictors in the imputation model(s):
#> 
#> mated
#> 
#> c("matage", "mated")
#> 
#> sep_unmeas
#> 
#> c("matage", "sep_unmeas")
#> 
#> c("mated", "sep_unmeas")
#> 
#> c("matage", "mated", "sep_unmeas")