Skip to contents

Summarise per-feature MNAR scores into a single dataset-level index

mnar_index.Rd

Aggregates the per-feature Tjur R² scores produced by mnar_score() into a single value in 0-1 representing how strongly experimental condition predicts missingness across the dataset.

Usage

mnar_index(
  summary_df,
  weight_by = c("miss_frac", "equal"),
  coverage_penalty = TRUE
)

Arguments

summary_df

A data.frame as returned in the $summary element of mnar_score(). Must contain columns: mnar_score, miss_frac, and mnar_class.

weight_by

Character. How to weight features when computing the mean score. One of:

"miss_frac"

(Default) Weight each feature by its missingness fraction. Features with more missing values contribute more — they represent a larger share of the imputation/modelling challenge.

"equal"

Unweighted mean across all informative features.

coverage_penalty

Logical. If TRUE (default), scales the weighted mean score by the fraction of features that are informative (i.e. have at least min_observed observed and min_missing missing values as set in mnar_score()). When most features are fully observed and therefore uninformative, the dataset-level index is attenuated accordingly. If FALSE, the index reflects only the informative features without penalising for their proportion.

Value

A list with the following elements:

index

Numeric in 0-1. The dataset-level MNAR index. Values near 0 indicate MAR-like missingness; values near 1 indicate missingness that is strongly structured by experimental condition.

weighted_mean_score

Numeric. The weighted mean Tjur R² across informative features, before any coverage penalty is applied.

coverage

Numeric. Fraction of features that were informative (had scoreable missingness).

n_informative

Integer. Number of informative features.

n_total

Integer. Total number of features.

weight_by

Character. The weighting scheme used.

coverage_penalty

Logical. Whether the coverage penalty was applied.

Details

Unlike mnar_global_score(), which fits a pooled logistic regression and can miss condition effects that cancel across features (e.g. peptide A missing in condition X and peptide B missing in condition Y), this function aggregates absolute per-feature scores and correctly identifies such opposing patterns as condition-structured missingness.

The index is computed as follows:

  1. Uninformative features (NA score) are excluded.

  2. The weighted mean Tjur R² is computed across informative features, using miss_frac or equal weights depending on weight_by.

  3. If coverage_penalty = TRUE, the result is multiplied by the fraction of all features that were informative.

The coverage penalty ensures that a dataset where only 5\ have any missingness does not score the same as one where 80\ are missing in a condition-structured way, even if the informative features have identical per-feature scores.

When coverage_penalty = FALSE the index reflects the strength of condition-missingness association among features that do have missing values, which may be preferable when missingness prevalence is not itself of interest.

See also

mnar_score, mnar_global_score

Examples

if (FALSE) { # \dontrun{
result <- mnar_score(obj, i = "peptides", group_cols = "condition")

# Default: miss_frac weighted, with coverage penalty
idx <- mnar_index(result$summary)
cat("MNAR index:", round(idx$index, 3), "\n")

# Strength among missing features only, no coverage penalty
idx2 <- mnar_index(result$summary,
                   weight_by        = "equal",
                   coverage_penalty = FALSE)
} # }