Bias-domain triangulation of non-convergent observational evidence in mental health research

Observational estimates in mental health can diverge because exposure is structured by familial, clinical and social factors. We developed bias-domain triangulation to test whether this divergence follows specific sources of confounding. The framework separates reported adjustment sets from an independently generated causal structure, assigns variables to causal roles and maps back-door pathways into bias domains before pooling. We applied it to prenatal paracetamol exposure and offspring autism spectrum disorder or attention-deficit/hyperactivity disorder. The review included 24 articles and 39 adjusted estimates. The pooled association declined from 2.08 under weak overall control to 0.98 under strong overall control. Only strong familial/genetic control brought the pooled estimate to the null. Strong control of clinical indication and social-behavioural factors left residual associations. These results link attenuation most consistently to shared familial liability in the available evidence. Bias-domain triangulation offers a reusable, pre-pooling test of which unresolved bias structure accompanies non-convergent observational estimates. The study was registered on PROSPERO (CRD420261365276).