Skip to contents

Quantify the degree of MNAR vs MAR missingness in a QFeatures assay

mnar_score.Rd

For each feature, fits a logistic regression of binary missingness against one or more experimental group variables from colData. Tjur's R² (discrimination coefficient) of this model is used as a per-feature MNAR score: high values indicate that missingness is structured by experimental condition (MNAR-like), while values near zero indicate missingness is unrelated to condition (MAR-like).

Usage

mnar_score(
  obj,
  i,
  group_cols,
  min_observed = 2,
  min_missing = 1,
  store_results = TRUE
)

Arguments

obj

A QFeatures object.

i

Integer or character. Index or name of the assay to analyse.

group_cols

Character vector. One or more column names from colData(obj) that define experimental groups (e.g. c("condition", "treatment")). Multiple columns are combined into an interaction term.

min_observed

Integer. Minimum number of observed (non-missing) values required for a feature to be included. Features below this threshold are returned as NA. Default: 2.

min_missing

Integer. Minimum number of missing values required for a feature to be included. Features with no missingness are uninformative and returned as NA. Default: 1.

store_results

Logical. If TRUE, MNAR scores and classifications are stored in rowData of the assay within the returned QFeatures object. Default: TRUE.

Value

A list with the following elements:

scores

Named numeric vector of per-feature MNAR scores (Tjur's R²).

summary

A data.frame with per-feature details: n_observed, n_missing, miss_frac, mnar_score, mnar_class.

global

A named numeric vector with dataset-level summaries: mean_mnar_score, median_mnar_score, prop_mnar, prop_mar, prop_uninformative.

obj

The input QFeatures object, optionally with results written into rowData if store_results = TRUE.

Details

The MNAR score for each feature is Tjur's R² (discrimination coefficient) from a logistic regression:

missing_indicator ~ group

where missing_indicator is 1 if the value is missing and 0 if observed, and group is a factor combining the specified colData columns. Tjur's R² is computed as the difference in mean predicted probabilities between missing and observed values:

Tjur's R² = mean(P(missing | truly missing)) - mean(P(missing | truly observed))

It ranges from 0 (no discrimination — MAR) to 1 (perfect discrimination — MNAR). Logistic regression is used in preference to a linear model as it is the correct model for a binary outcome.

If multiple group_cols are provided they are combined into a single interaction factor (e.g. condition:timepoint), so each unique combination of levels is treated as a distinct group.

MNAR classification thresholds (mnar_class):

  • "MNAR" : Tjur's R² >= 0.5

  • "mixed" : Tjur's R² in [0.2, 0.5)

  • "MAR" : Tjur's R² < 0.2

  • "uninformative" : excluded due to min_observed / min_missing filters

These thresholds are heuristic — inspect the score distribution before applying fixed cutoffs to your dataset.

See also

mnar_global_score

Examples

if (FALSE) { # \dontrun{
library(QFeatures)

# Single grouping variable
result <- mnar_score(obj, i = "peptides", group_cols = "condition")

# Multiple variables combined as an interaction
result <- mnar_score(obj, i = 1,
                     group_cols = c("condition", "timepoint"))

# Inspect per-feature results
hist(result$scores, breaks = 40,
     xlab = "MNAR score (Tjur's R²)", main = "Missingness structure")
print(result$global)
head(result$summary)
} # }