KurisutinaIdentity and access. Daphna Shohamy, Catherine E. Myers, Susan Onlaor, and Mark A. Gluck, Role of the Basal Ganglia in Category Learning: How Do Patients With Parkinson’s Disease Learn? Behavioral Neuroscience 118(4), 676–686. Peer-reviewed primary human experiment. DOI; author-institution PDF; author publication listing. Accessed 2026-09-22. Local PDF, text, provenance.
Factual core
Twelve nondemented patients with mild/moderate Parkinson’s disease, all tested on L-dopa, and fourteen controls completed 600 feedback trials over three days. They predicted sun/rain from fourteen recurring card combinations. Scoring rewarded a combination’s modal outcome, which could differ from the weather realized on that trial. Both groups improved; patients performed worse overall, but the group-by-day interaction was nonsignificant (p=.20). Each participant’s daily responses were classified by their closest ideal multicue, one-cue, or singleton profile. Control profiles increasingly resembled multicue responding; most patient profiles remained singleton-like. Seventy-five of seventy-eight participant-days met the fit threshold. The study reported neither held-out strategy forecasts nor a new awareness questionnaire for these participants. Its authors explicitly allowed either implicit or explicit learning to produce the named strategies and acknowledged alternative accounts involving integration, working memory, and shifting. Primary paper, Methods–Discussion.
Model audit and original interpretation
The cited modeling source defines, for pattern P, its presentation count N_P, observed sun-response count s_P, and a model’s sun-response probability p_PM:
Score(M) = ΣP (N_P p_PM − s_P)² / ΣP N_P².
The closest ideal profile wins; score below .1 counts as an adequate fit. Multicue predicts each pattern’s modal outcome, one-cue responds to one card’s presence/absence, and singleton predicts single-card outcomes but guesses on combinations. The score aggregates responses within patterns. These are targeted method details from Gluck, Shohamy & Myers (2002), Appendix p.418, not a claim to have fully read that additional paper. Local reference.
My deductions: the score weights squared response-proportion errors by N_P² and discards within-pattern temporal order. It is not a trial-level likelihood or a test of future choices. An integrated cue rule and a memorized lookup table can produce exactly the same responses on all trained patterns. Excellent profile fit therefore need not identify an internal strategy, its awareness, or its transfer behavior. A separate fit each day also does not establish that an earlier inferred state predicts a later one.
A further statistical caution is the reported strategy contingency analysis’s N=75: these are repeated days from 26 people, after excluding one control Day1, one patient Day1, and one patient Day2 fit. The described log-linear analysis does not specify within-person dependence handling. Its nominal significance should not be treated as independently replicated transitions. Likewise, comparing accuracy after assigning strategy labels from those same choices cannot independently validate a causal strategy explanation. The nonsignificant accuracy interaction prevents inferring different learning trajectories solely from significant later-day comparisons and a nonsignificant first-day comparison; it does not establish equal trajectories.
Clinical differences support a task-specific association, not a selective basal-ganglia intervention. Medication was not experimentally separated from disease. Matching on screening measures does not rule out differences in executive demands, response speed, or omitted responses. No devaluation or contingency-degradation test was reported, so calling the behavior habitual in the sense of outcome-insensitive control would add an untested claim. Absence of a contemporary awareness measure also blocks using this cohort as evidence that learned dispositions were unavailable to report.
Original implication for functional extraction
A recipient could reproduce trained category choices by receiving a response table, yet fail when a source person generalizes or changes strategy. Extraction should therefore be tested on future decisions selected to distinguish plausible mechanisms, rather than rewarded for recovering a strategy label assigned from its own evaluation choices. A proposed assay would fit on an initial sequence, freeze competing models, and test their different predictions under new combinations or changed task conditions. It must distinguish predicting the person’s answer, making a correct environment prediction, and reproducing the person’s subsequent learning. Use the same elicitation history and allowed information for paired recipient comparisons; add explicit awareness or outcome-sensitivity measurements only if those are the claims being tested. This is a research proposal, not a result of this paper.
Reading scope and limits
Later reading update: the complete Gluck 2002 audit now follows the targeted reading below. It adds an independently checked no-learning control for the original profile-fit tolerance; that calculation does not reanalyze the Shohamy patient data.
Read the complete main introduction, Methods, Results, Discussion, and Conclusions (journal pp.676–684); visually inspected all six figures and both tables. References through p.686 were retained but not independently followed. The supporting Gluck2002 PDF was read only for Table4, relevant methods, and the complete strategy-analysis Appendix; its equation was visually verified. It is not an additional full-paper read. No supplementary document was referenced in the main paper or identified on the checked primary pages; none was read. No raw data, model code, or graph digitization was obtained. Both PDFs parse as eleven pages; title/DOI and primary metadata were checked. A PubMed cookie-challenge download was rejected and removed. No sessions or experiments were run.