Screening for post-stroke cognitive impairment: The effect of diagnostic criteria on prevalence and false positive...

Accurately determining prevalence of post-stroke cognitive impairment (PSCI) is essential for research, yet methodological choices vary extensively across studies. This study theoretically and statistically compared four statistical approaches to PSCI classification using a cognitive screen in 1,951 stroke patients and 407 healthy controls. We contrasted (1) the common "at least one positive test" method using 5th centile cut-offs per subtest (uncorrected), (2) a Bonferroni-adjusted variant controlling for multiple comparisons, (3) a total-score summing subtest scores, and (4) a non-parametric Multivariate Normative Comparison (MNC) approach. False-positive rates (i.e., classifying healthy controls as impaired) (FPRs), PSCI prevalence and discrimination of stroke patients and healthy controls were contrasted between methods. PSCI identification using uncorrected 5th centile cut-offs per test resulted in impairment identification in 22% of healthy controls (99% CI [18-27]), while Bonferroni, total-score, and MNC methods maintained a rate of 5% identification of impairment. Corresponding PSCI prevalence estimates in stroke patients were 82%, 63%, 38%, and 55%, respectively. Contrary to predictions, the Bonferroni-adjusted method produced the greatest discrimination between stroke and control groups while ensuring family-wise error control. These findings demonstrate that uncorrected screening criteria can substantially inflate PSCI classifications in healthy controls even for ceiling-bounded screening tests, and that simple correction methods can achieve robust statistical control without loss of interpretability. Bonferroni-adjusted centile cut-offs are recommended for research studies in which PSCI is identified across multiple tests, as they maximize diagnostic specificity and maintain conceptual transparency compared to more complex multivariate approaches.