Imputing missing data

missing-data

Introduction to methods for imputing missing values

Author

Jaron Arbet

Published

January 4, 2025

Intro

Missing data analysis: Making it work in the real world

Multiple imputation is the gold standard method for handling missing data but is computationally intensive.
EM algorithm works well, but lack of general software.
If doing likelihood-based regression modeling, then as long as you adjust for variables that influence missingness, results will not be biased due to missing data (although patients with missing values are dropped from the model, thus decreasing power).
- Imputation likely still better due to improved power (i.e. all patients can be included).

Single imputation

There are various scenarios where one wants to impute missing values a single time:

When the goal is not statistical inference, but prediction
If you want to cluster your data before using BoutrosLab.plotting.general::create.heatmap
If the goal is statistical inference, often people will still do single imputation despite multiple imputation being the gold standard. Single imputation is so much easier/faster although more biased.

R notes:

Set the seed beforehand
All of these methods will throw errors if some features have 0 variance (i.e. they only take on 1 value). Remove such features beforehand.
For missForest/missMDA: make sure rownames are set to patient.id beforehand
For impute::impute.knn: make sure rownames are feature names and colnames are patient.id

suppressPackageStartupMessages(library(missForest));
suppressPackageStartupMessages(library(missMDA));
suppressPackageStartupMessages(library(impute));

Continuous

set.seed(123);

data('geno', package = 'missMDA'); 

# impute.knn (features in rows, samples in cols)
# this is my preferred method for high dimensional continuous data since it is relatively fast
knn.imp <- impute::impute.knn(t(geno))$data;
knn.imp <- data.frame(t(knn.imp), check.names = FALSE);

# PCA (samples in rows, features in cols)
ncomp <- missMDA::estim_ncpPCA(geno, ncp.min = 0, ncp.max = 6);
pca.imp <- missMDA::imputePCA(geno, ncp = ncomp$ncp, scale = TRUE)$completeObs;

# missForest
mf.imp <- missForest::missForest(geno)$ximp;

Categorical

data('vnf', package = 'missMDA');

# missForest
mf.imp <- missForest::missForest(vnf)$ximp;

# MCA
#ncomp <- missMDA::estim_ncpMCA(vnf); # slow method to estimate number of components
mca.imp <- missMDA::imputeMCA(vnf, ncp = 3)$complete.obs;

Mixed continuous/categorical

data('snorena', package = 'missMDA'); 

# missForest
mf.imp <- missForest::missForest(snorena)$ximp;

# FAMD
#missMDA::estim_ncpFAMD(snorena); # slow method to estimate number of components
famd.imp <- missMDA::imputeFAMD(snorena, ncp = 3)$completeObs

Multiple imputation

As mentioned in the intro, multiple imputation is the gold standard for handling missing data because it accounts for uncertainty in the imputed values.
Impute dataset multiple times to create multiple imputed datasets. Analyze each dataset separately then pool the results for statistical inference.
Multiple imputation by chained equations: what is it and how does it work?
mice R package
- mice JSS article

Expand for Session Info

R version 4.4.1 (2024-06-14 ucrt)
Platform: x86_64-w64-mingw32/x64
Running under: Windows 10 x64 (build 19045)

Matrix products: default


locale:
[1] LC_COLLATE=English_United States.utf8 
[2] LC_CTYPE=English_United States.utf8   
[3] LC_MONETARY=English_United States.utf8
[4] LC_NUMERIC=C                          
[5] LC_TIME=English_United States.utf8    

time zone: America/Los_Angeles
tzcode source: internal

attached base packages:
[1] stats     graphics  grDevices utils     datasets  methods   base     

other attached packages:
[1] impute_1.80.0  missMDA_1.19   missForest_1.5

loaded via a namespace (and not attached):
 [1] shape_1.4.6.1        gtable_0.3.5         xfun_0.49           
 [4] ggplot2_3.5.1        htmlwidgets_1.6.4    ggrepel_0.9.6       
 [7] lattice_0.22-6       vctrs_0.6.5          tools_4.4.1         
[10] generics_0.1.3       parallel_4.4.1       sandwich_3.1-1      
[13] tibble_3.2.1         fansi_1.0.6          pan_1.9             
[16] cluster_2.1.6        jomo_2.7-6           pkgconfig_2.0.3     
[19] Matrix_1.7-0         rngtools_1.5.2       scatterplot3d_0.3-44
[22] lifecycle_1.0.4      compiler_4.4.1       munsell_0.5.1       
[25] leaps_3.2            codetools_0.2-20     htmltools_0.5.8.1   
[28] glmnet_4.1-8         yaml_2.3.10          mice_3.17.0         
[31] nloptr_2.1.1         pillar_1.9.0         FactoMineR_2.11     
[34] tidyr_1.3.1          MASS_7.3-60.2        flashClust_1.01-2   
[37] DT_0.33              doRNG_1.8.6          iterators_1.0.14    
[40] rpart_4.1.23         boot_1.3-30          mitml_0.4-5         
[43] multcomp_1.4-26      foreach_1.5.2        nlme_3.1-164        
[46] tidyselect_1.2.1     digest_0.6.37        mvtnorm_1.3-1       
[49] dplyr_1.1.4          purrr_1.0.2          splines_4.4.1       
[52] fastmap_1.2.0        grid_4.4.1           colorspace_2.1-1    
[55] cli_3.6.3            magrittr_2.0.3       randomForest_4.7-1.2
[58] survival_3.6-4       utf8_1.2.4           broom_1.0.7         
[61] TH.data_1.1-2        scales_1.3.0         backports_1.5.0     
[64] estimability_1.5.1   rmarkdown_2.28       emmeans_1.10.6      
[67] nnet_7.3-19          lme4_1.1-35.5        zoo_1.8-12          
[70] coda_0.19-4.1        evaluate_1.0.0       knitr_1.48          
[73] doParallel_1.0.17    rlang_1.1.4          itertools_0.1-3     
[76] Rcpp_1.0.13          xtable_1.8-4         glue_1.8.0          
[79] minqa_1.2.8          rstudioapi_0.16.0    jsonlite_1.8.9      
[82] R6_2.5.1             multcompView_0.1-10