Computes random forest-based AUC

The function trains and fits a random forest model to assess the ability to predict missingness for the specified covariate(s). If missing indicator can be predicted as a function of observed covariates, MAR may be a likely scenario and would imply that imputation may be feasible.

Important: don't include variables like ID variables, ZIP codes, dates, etc.

Usage

smdi_rf(
  data = NULL,
  covar = NULL,
  train_test_ratio = c(0.7, 0.3),
  tune = FALSE,
  set_seed = 42,
  ntree = 1000,
  n_cores = 1
)

Arguments

data: dataframe or tibble object with partially observed/missing variables
covar: character covariate or covariate vector with partially observed variable/column name(s) to investigate. If NULL, the function automatically includes all columns with at least one missing observation and all remaining covariates will be used as predictors
train_test_ratio: numeric vector to indicate the test/train split ratio, e.g. c(.7, .3) which is the default
tune: logical,if TRUE, a 5-fold cross validation is performed combined with a random search for the optimal number of optimal number of variables randomly sampled as candidates at each split (mtry). FALSE is the default due to potentially extensive computation times.
set_seed: seed for reproducibility, defaults to 42
ntree: integer, number of trees (defaults to 1000 trees)
n_cores: integer, if >1, computations will be parallelized across amount of cores specified in n_cores (only UNIX systems)

Value

returns an rf object which comes as a list that contains the ROC AUC value and corresponding variable importance in training dataset (latter as ggplot object). That is, for each covar, the following outputs are provided:

rf_table: The area under the receiver operating curve (AUC) as a measure of the ability to predict the missingness of the partially observed covariate
rf_plot: ggplot object illustrating the variable importance for the prediction made expressed by the mean decrease in accuracy per predictor. That is how much would the accuracy of the prediction (# of correct predictions/Total # of predictions made) decrease, had we left out this specific predictor.
OOB: estimated OOB error for each investigated partially observed confounder (indicates the performance of the random forest model for data points that are not used in training a tree.)

Details

The random forest utilizes the randomForest engine.

CAVE: If the missingness indicator variables of other partially observed covariates (indicated by suffix _NA) have an extremely high variable importance (combined with an unusually high AUC), this might be an indicator of a monotone missing data pattern. In this case it is advisable to exclude other partially observed covariates and run missingness diagnostics separately.

References

Sondhi A, Weberpals J, Yerram P, Jiang C, Taylor M, Samant M, Cherng S. A systematic approach towards missing lab data in electronic health records: A case study in non-small cell lung cancer and multiple myeloma. CPT Pharmacometrics Syst Pharmacol. 2023 Jun 15. <doi: 10.1002/psp4.12998.> Epub ahead of print. PMID: 37322818.

Examples

library(smdi)

smdi_rf(data = smdi_data, covar = "ecog_cat")
#> # A tibble: 1 × 2
#>   covariate rf_auc
#> * <chr>     <chr> 
#> 1 ecog_cat  0.510