Smooth Curves, Similar Conclusions? Comparing Linear Regression and GAMLSS Neuropsychological Norms
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Objective
Regression-based normative approaches are widely used in neuropsychology but often rely on score transformations to satisfy model assumptions. We compared previously published linear regression (LR)-based norms with norms derived using Generalized Additive Models for Location, Scale and Shape (GAMLSS) for the brief cognitive battery used in the Norwegian Dementia Disease Initiation (DDI) cohort.
Method
GAMLSS norms were developed using the same normative samples as the original LR norms for the Consortium to Establish a Registry for Alzheimer’s Disease (CERAD) word list test, Trail Making Test (TMT) A and B, FAS phonemic fluency, and Visual Object and Space Perception Battery (VOSP) Silhouettes. Expected low-score frequencies and empirical base rates were assessed in a normative subsample (n = 131). Clinical implications were evaluated in the DDI clinical cohort (n = 643) using Mild Cognitive Impairment (MCI) classification, two-year diagnostic stability and change, and cerebrospinal fluid (CSF) biomarkers.
Results
Compared with LR norms, GAMLSS yielded lower frequencies of low scores, primarily driven by CERAD delayed recall. Nevertheless, concordance between approaches was high (κ = 0.91), with only 4.2% discordant classifications. Two-year diagnostic stability and change were broadly similar across approaches, and CSF biomarker profiles did not clearly favor either normative method.
Conclusions
GAMLSS provided a more faithful representation of neuropsychological score distributions, particularly for bounded and non-normal outcomes. However, downstream clinical differences were modest in this setting, suggesting that well-calibrated LR norms may remain robust for clinical classification.
