Imputation algorithm selection

imputation-algorithm-selection

Summary

Select and dispatch missing-value matrices to appropriate MNAR imputation methods (GSimp, QRILC, kNN-TN) based on data characteristics and missingness mechanism. This skill routes preprocessed metabolomics data through method-specific wrapper functions to produce algorithm-specific imputed matrices for downstream comparison and evaluation.

When to use

You have a metabolomics dataset with left-censored missing values (e.g., below limit of quantification in LC/MS or GC/MS) and need to evaluate multiple imputation approaches. Select this skill when you must compare GSimp (Gibbs sampler with elastic-net prediction), QRILC (quantile regression with truncated sampling), or kNN-TN (truncation k-nearest neighbors) on the same input matrix and retain separate imputed outputs for each method to assess performance differences.

When NOT to use

• Input data contains only MCAR (missing completely at random) or MAR (missing at random) patterns with no left-censoring; GSimp's truncated bounds (lo=-Inf, hi='min') are designed for MNAR left-censored data and will not provide additional benefit over simpler methods.
• Imputed output is already produced by a single method and no algorithm comparison is required; this skill is for dispatch and multi-method evaluation, not single-pass imputation.
• Data has not been log-transformed or pre-initialized; the wrapper functions assume log-scale preprocessing and will produce incorrect scale-recovery steps if applied to raw-scale input.

Inputs

• Log-transformed metabolomics matrix (samples × variables, numeric)
• QRILC pre-initialized matrix with missing values re-introduced at original NA positions
• Centralized and scaled data frame with scale/centering parameters recorded
• NA position matrix (row-column indices of missing elements)

Outputs

• GSimp-imputed matrix (post-scale-recovery, post-exponential-recovery)
• QRILC-imputed matrix (post-exponential-recovery)
• kNN-TN-imputed matrix (post-exponential-recovery)
• Method-indexed list or separate CSV files of imputed matrices for comparison

How to apply

Begin with log-transformed, QRILC-initialized, and centrally-scaled data. For each imputation method, invoke its corresponding wrapper function: (1) GSimp via pre_processing_GS_wrapper with parameters iters_each=50, iters_all=10, imp_model='glmnet_pred', lo=-Inf, hi='min' for left-censored bounds; (2) QRILC via sim_QRILC_wrapper for direct quantile-regression imputation; (3) kNN-TN via sim_trKNN_wrapper with k=3 and distance='truncation'. Record the NA position matrix before imputation, apply each method independently to the same initialized data, and recover scale/log transformation post-imputation. The choice between methods depends on computational budget (GSimp requires more iterations for MCMC convergence but provides uncertainty estimates) and whether truncated mean estimation (kNN-TN) or regression quantiles (QRILC) better match your data distribution assumptions.

Related tools

• GSimp.R (Core function library for Gibbs sampler imputation; contains GS_impute() for MCMC-based left-censored value sampling) — https://github.com/WandeRum/GSimp
• Impute_wrapper.R (Dispatch and wrapper functions for method routing; contains pre_processing_GS_wrapper, sim_QRILC_wrapper, and sim_trKNN_wrapper) — https://github.com/WandeRum/GSimp
• imputeLCMD R package (Provides impute.QRILC() function for quantile-regression-based left-censored imputation)
• Trunc_KNN/Imput_funcs.r (kNN-TN algorithm implementation with imputeKNN() for truncation-based k-nearest-neighbor imputation) — https://github.com/WandeRum/GSimp

Examples

source('Impute_wrapper.R'); source('GSimp.R'); source('Trunc_KNN/Imput_funcs.r'); after_GS_imp <- pre_processing_GS_wrapper(untargeted_data); after_QRILC_imp <- sim_QRILC_wrapper(log(untargeted_data)) %>% exp(); after_trKNN_imp <- sim_trKNN_wrapper(log(untargeted_data)) %>% exp()

Evaluation signals

• Each method produces a complete imputed matrix with no remaining NA values at original missing positions; verify by intersecting output matrix NA positions with input NA positions (should be disjoint).
• Scale recovery post-imputation is consistent: verify that mean and variance of non-missing features match pre-scaling values (using recorded param_df) within numerical tolerance.
• Imputed values for left-censored positions fall within expected bounds: GSimp and QRILC imputed values should be ≤ minimum of non-missing data in each variable (hi='min' constraint); kNN-TN values should reflect truncated distribution estimates.
• Gibbs sampler convergence (GSimp only): check that repeated imputations with different random seeds produce stable imputed values at traced positions (gibbs_res output); standard deviation of traced elements should stabilize across iterations.
• Method-specific output schemas are met: GSimp returns list with data_imp and gibbs_res; QRILC and kNN-TN return imputed data frame only.

Limitations

• GSimp convergence requires tuning iters_each and iters_all; default iters_all=10, iters_each=50 is a compromise between speed and convergence and may not be sufficient for high-dimensional or extremely sparse data. Larger iterations (iters_all=20, iters_each=100) are recommended for critical analyses but increase runtime quadratically.
• kNN-TN truncated mean estimation assumes Pearson correlation structure and may perform poorly on metabolites with non-linear or multimodal distributions after log transformation.
• QRILC imputation assumes that quantile regression surfaces are smooth; performance degrades when variables have extreme outliers or bimodal distributions that violate underlying quantile regression assumptions.
• All methods assume that log transformation linearizes the relationship between features and makes residuals approximately normal; log-transformation may not be appropriate for datasets with zero or negative values, or for very sparse omics data where most samples are below detection limit.
• The wrapper functions are specific to left-censored MNAR data (lo=-Inf, hi='min'); extending to right-censored or MCAR/MAR data requires parameter reconfiguration (lo/hi bounds, initial method) and may invalidate the preprocessing pipeline.

Evidence

• [intro] wrapper functions for different MNAR imputation methods (GSimp, QRILC, and kNN-TN) and evaluations of these methods: "GSimp provides data pre-processing, simulated data generation, MNAR generation, wrapper functions for different imputation methods (GSimp, QRILC, and kNN-TN) and evaluations of these"
• [readme] Impute_wrapper.R dispatches to GSimp, QRILC, kNN-TN with pre_processing_GS_wrapper: "Impute_wrapper.R contains wrapper functions for different imputation methods (contains pre_processing_GS_wrapper)."
• [readme] pre_processing_GS_wrapper orchestrates log transformation, QRILC initialization, scaling, GSimp imputation, and scale/log recovery: "All aboved steps has been wrapped into the pre_processing_GS_wrapper function for a one-step processing and imputation. The function will give the final imputed dataset."
• [readme] GSimp parameters: iters_each controls per-variable iterations, iters_all controls full-matrix iterations, lo/hi set truncation bounds for left-censored data: "iters_each is the number of iterations for imputing each missing variable (default=100). iters_all is the number of iterations for imputing the whole data matrix (default=20). lo is the"
• [readme] QRILC and kNN-TN are compared as alternative left-censored methods using wrapper functions on log-transformed data with exponential recovery: "data_raw_log <- untargeted_data %>% log() after_trKNN_imp <- sim_trKNN_wrapper(data_raw_log) %>% data.frame() %>% exp() after_QRILC_imp <- sim_QRILC_wrapper(data_raw_log) %>% exp()"