Perseus-style MNAR imputation (left-censoring) on log2 data
perseus_impute.RdApplies the common Perseus imputation where, per sample, missing values are drawn from a narrow normal distribution shifted down from the observed distribution. Works on a plain log2 matrix/data.frame or on a DAList:
If
DAList$data_per_contrastexists and is non-empty, imputes each contrast's matrix and stores the result inDAList$data_per_contrast[[contrast]].Otherwise, imputes the global
DAList$datamatrix in-place.
Usage
perseus_impute(
x,
shift = 1.8,
width = 0.3,
robust = TRUE,
min_obs_per_sample = 5,
seed = NULL,
save_before_after = FALSE,
store_mask = TRUE
)Arguments
- x
A numeric log2 matrix/data.frame (rows=features, cols=samples), or a DAList with
$dataand optionally$data_per_contrast.- shift
numeric, how far to shift down the mean in SD units (default 1.8)
- width
numeric, fraction of SD used as imputation SD (default 0.3)
- robust
logical, use median/MAD (TRUE) or mean/SD (FALSE) per sample
- min_obs_per_sample
Integer. Minimum number of observed (non-missing) log2 values required in a sample before estimating its own mean (mu_j) and standard deviation (sd_j) for imputation. If a sample has fewer than this many observed values, the function falls back to using the pooled distribution across all samples. Default = 5.
- seed
integer or NULL, set for reproducible draws
- save_before_after
logical, store full before/after matrices and mask into the DAList (or return attrs for matrix mode). Default FALSE.
- store_mask
logical, store/log an imputed logical mask. Default TRUE.
Value
If x is a matrix/data.frame, returns the imputed matrix (same dims);
when store_mask=TRUE, attaches attr(result, "imputed_mask").
If x is a DAList, returns the updated DAList with imputed matrices
written to the appropriate slots; when save_before_after=TRUE,
persists before_log2, after_log2, and imputed_mask.
Details
Optionally stores before/after matrices and an imputed mask for diagnostics.
For each sample j, let μ_j and σ_j be location and scale of its observed (log2) values. Missing entries in sample j are drawn from Normal( μ_j - shiftσ_j, (widthσ_j)^2 ). Features missing in all samples remain NA (uninformative).
This function assumes that the input data are on the log2 scale. As a safety check, if the maximum absolute value in a matrix exceeds 1e3, a warning is issued suggesting that the data may not be log2-transformed.
Examples
if (FALSE) { # \dontrun{
# Matrix use
X_imp <- perseus_impute(X, seed=123)
mask <- attr(X_imp, "imputed_mask")
# DAList use (per-contrast if present)
results <- perseus_impute(results, save_before_after = TRUE, seed=1)
} # }