The predictive processing account of psychosis does not survive pooling

Design

Random-effects multi-level meta-analysis, PROSPERO registered, Newcastle-Ottawa risk of bias

Population

34 results from 27 studies: 904 adults with schizophrenia-spectrum psychosis and 1039 healthy controls

Primary outcome

Behavioural reliance on prior expectations across perceptual tasks

Effect

g = 0.03 (95 per cent CI -0.27 to 0.34, P = 0.818); delusions r = -0.16 (-0.51 to 0.19); hallucinations r = 0.04 (-0.28 to 0.36)

The proposition that psychotic symptoms arise from a systematically altered weighting of prior expectations against sensory evidence has organised a great deal of computational psychiatry. It has also rested on individual perceptual tasks in modest samples pointing in different directions.

A random-effects multi-level meta-analysis pooled 34 results from 27 studies comparing 904 adults with schizophrenia-spectrum psychosis against 1039 healthy controls on behavioural measures of reliance on priors. The pooled effect was 0.03 (95 per cent CI -0.27 to 0.34, P = 0.818) - which is nothing. Nor was reliance on priors associated with delusions (r = -0.16, -0.51 to 0.19, six results) or hallucinations (r = 0.04, -0.28 to 0.36, ten results). A subgroup analysis testing the common rescue - that priors are affected differently at different levels of a cognitive hierarchy - did not reconcile the conflicting results either (F = 0.1, P = 0.758).

This does not disprove predictive processing as a framework, and the authors are careful to say what they have tested: a generalised deficit spanning multiple aspects of perception. What it does is remove the licence to cite a single task result as evidence for a broad theory. For a clinician the value is defensive - when a paper, a talk or a product claims that a perceptual task indexes psychosis risk or delusion severity, this is the pooled answer, and it is null.

• Treat single-task computational findings in psychosis as preliminary until pooled.
• Do not accept a perceptual task as a marker of delusion or hallucination severity; neither correlation held.
• Note the sample sizes: 904 patients and 1039 controls is a substantial base for a null result.
• The framework is not refuted; a generalised, task-spanning deficit is.
• Nothing here changes assessment or treatment - it changes what can be claimed.

The statistics, in plain English

A Hedges g of 0.03 with an interval from -0.27 to 0.34 is a genuinely informative null: the interval excludes anything larger than a small effect in either direction, which is different from simply failing to reach significance. The correlations with symptoms have wide intervals because they rest on far fewer results - six and ten respectively - so those are underpowered rather than conclusively null. The subgroup test looked for a moderator and found none, which weakens the usual explanation offered for inconsistent findings.

Read the rest in the app

You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.

Get Daily Doctor — free

Scan to keep reading on your phone. No account needed to start.