# Comment on: Probabilities of conditionals: The relevance effect might be confounded by the existence of boundary cases (2025) by Zhan and Wang

## TL;DR

This paper critiques a statistical model used by Zhan and Wang (2025) and shows that their conclusions about the relevance effect are invalid.

## Contribution

The paper demonstrates that the statistical model used by Zhan and Wang is inappropriate and provides corrected analyses.

## Key findings

- The binomial model used by Zhan and Wang is invalid for the collected data.
- Re-analyses of both Zhan and Wang's and Skovgaard-Olsen et al.'s data reject the conclusions drawn by Zhan and Wang.
- Valid statistical models support the original relevance effect findings.

## Abstract

Zhan and Wang (2025) claim that the relevance effect found in Skovgaard-Olsen et al. (2016) might be an artifact due to “boundary cases” of zero or one probabilities. The relevance effect refers to the finding that people’s expectations about the probabilistic relevance of the antecedent for the consequent influences people’s probability judgments of indicative conditions, “If A, then C”. In Skovgaard-Olsen et al. (2016, 2017), this effect was found not by using probabilistic notions of relevance, like ΔP=P(C|A)-P(C|-A), as direct predictors of P(if A, then C). Instead, stimulus materials were pre-validated to implement different qualitative reason relation assessments of whether A is a reason for C, A reason against C, or neither, which were operationalized via categories of probabilistic relevance (positive relevance ΔP>0, negative relevance ΔP<0, irrelevance ΔP=0). Despite this, this commentary addresses the fact that the binomial statistical model used by Zhan and Wang (2025) is invalid for the type of data collected. The purpose of this paper is to do a re-analysis of the two data sets of Zhan and Wang (2025). First, we reanalyze the new data that Zhan and Wang (2025) present. Then we re-analyze the data from Skovgaard-Olsen et al. (2016), which Zhan and Wang (2025) in turn re-analyze in Table 5 of their paper. In both cases, we find that invalid statistical models were applied and that the conclusions that Zhan and Wang (2025) draw must be rejected once valid statistical models are applied.