pca_imp: PCA Imputation for Numeric Matrices

In slideimp: Numeric Matrices K-NN and PCA Imputation

PCA Imputation for Numeric Matrices

Description

Impute missing values in a numeric matrix using regularized or expectation-maximization (EM) PCA imputation. Supports warm-start LOBPCG with both the previous eigenblock and search direction.

Usage

pca_imp(
  obj,
  ncp = 2,
  scale = TRUE,
  method = c("regularized", "EM"),
  coeff.ridge = 1,
  row.w = NULL,
  threshold = 1e-06,
  seed = NULL,
  nb.init = 1,
  maxiter = 1000,
  miniter = 5,
  solver = c("auto", "exact", "lobpcg"),
  lobpcg_control = NULL,
  colmax = 0.9,
  post_imp = TRUE,
  na_check = TRUE,
  clamp = NULL,
  .progress = FALSE
)

Arguments

obj	A numeric matrix with samples in rows and features in columns.
ncp	Integer. Number of principal components used to predict missing entries.
scale	Logical. If TRUE, columns are scaled to unit variance.
method	Character. PCA imputation method: either "regularized" or "EM".
coeff.ridge	Numeric. Ridge regularization, used only when method = "regularized". Values ⁠< 1⁠ move toward EM PCA; values ⁠> 1⁠ move toward mean imputation.
row.w	Row weights, normalized to sum to 1. NULL (equal weights), a positive numeric vector of length nrow(obj), or "n_miss" (down-weight rows with more missing values).
threshold	Numeric. Convergence threshold.
seed	Integer, numeric, or NULL. Random seed for reproducibility.
nb.init	Integer. Number of random initializations. The first initialization is always mean imputation.
maxiter	Integer. Maximum number of iterations.
miniter	Integer. Minimum number of iterations.
solver	Character. Eigensolver: "auto" (default), "exact", or "lobpcg". "auto" runs a short timed probe and picks "lobpcg" only when clearly faster. Consecutive EM calls warm-start LOBPCG with both the previous eigenblock and search direction. When nb.init > 1, the auto choice from the first init is reused. See Performance tips.
lobpcg_control	A list of LOBPCG eigensolver control options, usually created by lobpcg_control(). A plain named list is also accepted. Ignored when solver = "exact".
colmax	Numeric scalar between 0 and 1. Columns with a missing-data proportion greater than colmax are excluded from the main imputation method. Excluded columns are left unchanged unless post_imp = TRUE, in which case remaining missing values are replaced by column means when possible.
post_imp	Logical. If TRUE, replace missing values remaining after the main imputation method with column means when possible.
na_check	Logical. If TRUE, check whether the returned matrix still contains missing values.
clamp	Optional numeric vector c(lower, upper) bounding PCA-imputed values (use -Inf/Inf for one-sided, NULL for none). E.g., c(0, 1) for DNAm beta values. Observed values are not clamped.
.progress	Logical. If TRUE, show imputation progress.

Details

This function reimplements the PCA imputation method from the missMDA package by Francois Husson and Julie Josse, based on Josse and Husson (2016).

Value

A numeric matrix of the same dimensions as obj, with missing values imputed. The returned object has class slideimp_results.

PCA Performance tips

Speed comes from three levers: solver (through LOBPCG with warm-start), threshold, and scale. Tune these first, then accuracy parameters (ncp, coeff.ridge) on a representative subset.

Exact vs. LOBPCG with warm-start. Whether "lobpcg" beats "exact" depends on size and low-rankness: "lobpcg" is preferred for large, approximately low-rank matrices with small ncp, and "exact" for small matrices (including slide_imp() windows), where it is faster and more robust. Separately, the warm-start makes each successive solve cheap: pca_imp() warm-starts LOBPCG with the previous eigenblock and search direction, so once imputed values stabilize, later solves converge in a few iterations. The payoff therefore grows with the number of EM iterations, independent of low-rankness. solver = "auto" (default) probes both and is a safe start.

Threshold. The default 1e-6 is conservative; 1e-5 is often faster with very similar values.

Scale. For columns on a common scale (e.g., DNAm beta values in ⁠[0, 1]⁠), scale = FALSE can be faster and more accurate.

Parallel and BLAS. In parallel via tune_imp() or group_imp() with a multithreaded BLAS, set pin_blas = TRUE to avoid thread oversubscription. On Windows, the stock BLAS can be slow. Advanced users can swap in OpenBLAS.

See Speeding up PCA imputation for the full workflow.

References

Josse J, Husson F (2013). Handling missing values in exploratory multivariate data analysis methods. Journal de la SFdS, 153(2), 79-99.

Josse J, Husson F (2016). missMDA: A Package for Handling Missing Values in Multivariate Data Analysis. Journal of Statistical Software, 70(1), 1-31. \Sexpr[results=rd]{tools:::Rd_expr_doi("10.18637/jss.v070.i01")}

Examples

set.seed(123)
obj <- sim_mat(10, 10)$input
sum(is.na(obj))
obj[1:4, 1:4]

# Randomly initialize missing values 5 times. The first initialization is
# mean imputation. Select `ncp` with `tune_imp()`.
pca_imp(obj, ncp = 2, nb.init = 5, seed = 123)

slideimp documentation built on June 17, 2026, 1:08 a.m.