Revisiting Simpson's Paradox: a statistical misspecification perspective

TL;DR

This paper reexamines Simpson's paradox through the lens of statistical misspecification, emphasizing that association reversals can be spurious and highlighting the importance of valid probabilistic assumptions for trustworthy inference.

Contribution

It introduces a statistical misspecification perspective to distinguish between trustworthy and untrustworthy association reversals in Simpson's paradox.

Findings

Spurious associations often result from invalid assumptions.

Trustworthy reversals are distinguishable from spurious ones.

Revisits causal explanations of the paradox with new insights.

Abstract

The primary objective of this paper is to revisit Simpson's paradox using a statistical misspecification perspective. It is argued that the reversal of statistical associations is sometimes spurious, stemming from invalid probabilistic assumptions imposed on the data. The concept of statistical misspecification is used to formalize the vague term `spurious results' as `statistically untrustworthy' inference results. This perspective sheds new light on the paradox by distingusing between statistically trustworthy vs. untrustworthy association reversals. It turns out that in both cases there is nothing counterintuitive to explain or account for. This perspective is also used to revisit the causal `resolution' of the paradox in an attempt to delineate the modeling and inference issues raised by the statistical misspecification perspective. The main arguments are illustrated using both actual and hypothetical data from the literature, including Yule's "nonsense-correlations" and the Berkeley admissions study.