piiR provides a clean, standalone implementation of the Predictive Information Index (PII), a post-measurement metric designed to quantify how well a score predicts an outcome compared to a full model.
Version 0.3.0 includes only the canonical scoring functions and visualization tool. All non-canonical or experimental features have been removed.
install.packages("piiR") # From CRAN
What is PII? The Predictive Information Index (PII) quantifies the utility of a score relative to the full predictive model. It answers the question:
How much information does this score preserve from the full model when predicting an outcome?
PII can be computed using three methods:
"r2": squared correlation ratio (R²-based)
"rm": root-mean-squared error ratio
"v": predicted value variance ratio
pii(y, score_pred, full_pred, type = "r2")
y: observed outcome
score_pred: predictions from your score-based model
full_pred: predictions from the full model
type: one of "r2", "rm", or "v"
Returns a single PII value between 0 and 1.
Visualizes score vs. outcome prediction alignment.
pii_plot(y, score_pred, full_pred)
set.seed(123)
y <- rnorm(100)
full <- y + rnorm(100, sd = 0.2)
score <- y + rnorm(100, sd = 0.5)
pii(y, score, full, type = "r2")
pii_plot(y, score, full)
Wells (2025). Predictive Information Index (PII): A model-agnostic utility metric for post-measurement evaluation. https://github.com/TheotherDrWells/piiR