[Paper Review] On Probability Leakage

William M. Briggs|arXiv (Cornell University)|Jan 17, 2012

Statistical Mechanics and Entropy8 references3 citations

TL;DR

This paper introduces 'probability leakage' as a critical model error where a statistical model assigns positive probability to events deemed impossible by the evidence (E). It demonstrates that common regression models often exhibit severe leakage—especially in bounded data like GPA or counts—rendering them uncalibrated and empirically invalid, even when standard diagnostics suggest otherwise.

ABSTRACT

The probability leakage of model M with respect to evidence E is defined. Probability leakage is a kind of model error. It occurs when M implies that events $y$, which are impossible given E, have positive probability. Leakage does not imply model falsification. Models with probability leakage cannot be calibrated empirically. Regression models, which are ubiquitous in statistical practice, often evince probability leakage.

Motivation & Objective

• To define and formalize 'probability leakage' as a model error when a model assigns positive probability to events ruled out by evidence.
• To challenge the assumption that standard model diagnostics (e.g., p-values, residual plots) are sufficient for model validation, especially in bounded data.
• To argue that models with probability leakage cannot be empirically calibrated and thus fail as predictive tools.
• To demonstrate through real-world examples (e.g., GPA, call center data) that even well-fitting regression models exhibit massive leakage.
• To advocate for a predictive, objectivist Bayesian framework where model validity is judged by adherence to evidence, not posterior parameter estimates.

Proposed method

• Defines probability leakage as the positive probability assigned by model M to events y that are impossible under evidence E (i.e., Pr(y < ya or y > yb | E) = 0, but Pr(y < ya or y > yb | x, z, M) > 0).
• Uses the predictive distribution p(y|x, z, M) = ∑θ p(y|x, z, θi, M) p(θi|z, M) as the logical consequence of model M and evidence, which must be checked for consistency with E.
• Applies the framework to regression models by computing P(y < 0 | x, z, M) for bounded outcomes (e.g., abandonment rates, GPA), showing leakage even when diagnostics appear satisfactory.
• Quantifies leakage as the probability mass assigned to impossible values, with values near 1 indicating severe model failure.
• Proposes that leakage should be a primary criterion in model selection, especially when M implies continuous distributions over inherently discrete or bounded observables.
• Suggests future work on modeling leakage as a function of x, using new models to predict leakage across different input values.

Experimental results

Research questions

• RQ1 How can a model be both statistically well-fitting and yet fundamentally flawed due to probability leakage?
• RQ2 To what extent do standard regression models fail in their predictive calibration when applied to bounded or discrete data?
• RQ3 Can probability leakage serve as a stronger criterion for model falsification than traditional diagnostic tests?
• RQ4 How does the presence of evidence E that restricts the range of observable y affect the validity of continuous probability models like the normal distribution?
• RQ5 What is the relationship between model calibration, probability leakage, and the posterior distribution p(θ|z, M) in Bayesian inference?

Key findings

• Regression models applied to bounded data—such as GPA (0–4) or call center abandonment rates—can exhibit severe probability leakage, with up to 92% of predicted probability mass assigned to impossible values (e.g., negative abandonment rates).
• Even when p-values and residual plots indicate good fit, models can assign 38% and 2% probability to negative outcomes at Location A and B, respectively, under median predictor values.
• For a null model (no predictors), the probability leakage for negative outcomes in the same data was over 10%, indicating substantial model error.
• Probability leakage is not detectable via standard diagnostics but is a direct consequence of model misalignment with evidence E.
• Models with probability leakage cannot be empirically calibrated because they assign non-zero probability to events that are known to be impossible.
• The paper concludes that probability leakage is a strong, often overlooked indicator of model inadequacy, even when all other diagnostics appear satisfactory.