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

Observational studies in behavioural health often produce conflicting evidence because exposures are entangled with familial, clinical and social determinants. Rather than treating non-convergence as an obstacle to synthesis, we propose bias-domain triangulation, a causal-structure-aware framework that treats it as a diagnostic target. The framework distinguishes actual covariate adjustments from candidate background causal structures, adjudicates the causal roles of adjustment variables and maps core back-door pathways into distinct bias domains before quantitative synthesis, thereby reframing the question from whether estimates are heterogeneous to which bias structure changes the estimate. We apply this framework to a systematically derived literature on prenatal paracetamol exposure and offspring autism spectrum disorder or attention-deficit/hyperactivity disorder, comprising 22 studies and 33 adjusted estimates. Stronger overall control was associated with attenuation of the pooled association from 2.08 under weak control to 0.98 under strong control. Domain-specific analyses showed that pooled estimates approached the null only under strong familial/genetic control, whereas strong control of clinical-indication or social-behavioural domains left residual associations. This pattern suggests that shared familial liability is the bias structure most consistently associated with attenuation in the current evidence. Beyond this case, bias-domain triangulation offers a reusable strategy for diagnosing why observational evidence may persistently diverge, moving evidence synthesis beyond heterogeneity toward structural explanation. This study was registered on PROSPERO (CRD420261365276).