Improving cut-off determination of Neuropsychological tests: a Bayesian approach

Traditional validation approaches in neuropsychology typically rely on normative data from healthy individuals to establish cut-off scores. While this method provides estimates of specificity, it does not inform sensitivity, and the resulting cut-off is generic rather than pathology-specific. This limitation becomes critical when assessing clinical populations, as neuropsychological tools are often validated on very small patient samples. Such constraints make it difficult to apply standard Receiver Operating Characteristic (ROC) analyses, which require sufficiently large and balanced datasets to produce reliable estimates. To address this issue, we propose a Bayesian adaptation of the ROC framework. The rationale behind this approach is that Bayesian methods naturally incorporate uncertainty arising from sample size, allowing for more robust inference even when patient data are scarce. Unlike simple simulation techniques, Bayesian modeling provides posterior distributions for key parameters, enabling credible intervals for cut-off scores, sensitivity, and specificity. In our conceptual example, we consider a dataset with 200 healthy controls and only 10 patients. Using Bayesian regression, we estimate model parameters and generate simulated datasets from posterior distributions. For each simulated dataset, we compute ROC curves and derive credible intervals for diagnostic metrics. This approach leverages prior assumptions about the relationship between patients and controls and models data distributions using Student’s t-distribution, which accounts for heavier tails and greater variability in small samples. Consequently, the Bayesian framework offers a principled solution for neuropsychological validation in contexts where traditional ROC analysis is impractical.