Bidirectional Dynamics of Cannabis Use and Schizophrenia: A Compound-Causality Research Protocol and Analysis Plan

Background. The question "Are individuals with schizophrenia more likely to use cannabis, or is cannabis more likely to cause schizophrenia?" is usually posed as a choice between two exclusive accounts. Six decades of epidemiology, two decades of genetically informed designs and three legalisation natural experiments indicate that the choice is false: the two directions operate concurrently, through different mechanisms, over different timescales, and in partially different subpopulations. Objective. To specify, pre-register and make falsifiable a compound bidirectional model of the cannabis–schizophrenia relationship, and to define the evidence synthesis, prior elicitation and analytic programme needed to estimate its parameters jointly rather than sequentially. Design. A protocol in four parts: (i) a systematic review and dose-response meta-analysis conducted to PRISMA 2020 and PRISMA-S, with risk of bias assessed by ROBINS-E and certainty by GRADE; (ii) derivation of informative, robust prior distributions from that synthesis; (iii) a prospective multi-arm youth cohort with repeated within-person measurement, polygenic scoring and biologically verified exposure; and (iv) a triangulated analysis combining random-intercept cross-lagged panel models, marginal structural models and the parametric g-formula, bidirectional Mendelian randomization, and within-family comparison. The cohort comprises N = 15,000 young people in three arms — ultra-high-risk youth (Group A, n = 3,000), population-matched adolescents (Group B, n = 10,000) and a first-episode psychosis cohort (Group C, n = 2,000) — each followed for 10 years across 25 assessment occasions (quarterly in years 1–2, semi-annual in years 3–10), with ecological momentary assessment bursts, biochemical verification of cannabis exposure and mandatory registry linkage. All exposure is observational: no cannabis or tetrahydrocannabinol (THC) is administered to any participant at any stage. Hypotheses. H1 (forward psychotomimetic causation): heavy, high-potency, adolescent-onset use raises the hazard of psychotic disorder. H2 (reverse causation): psychosis liability, prodromal symptoms and self-medication raise cannabis use. H3 (common liability): shared genetic and environmental architecture generates the association without either arrow. The compound model nests all three plus their feedback product and liability-dependent effect modification, and the protocol specifies which estimate discriminates between them. Power and feasibility. Monte Carlo simulation (reproducible code in Appendix A) indicates that the co-primary forward endpoint in Group A detects a transition hazard ratio of approximately 1.57 at α = 0.025 under realistic exposure misclassification in its principal supporting analysis (a conditional time-varying hazard model using all observed transitions; power for the confirmatory per-protocol target-trial contrast, which draws on fewer events, has not yet been simulated and is expected to be lower), and the co-primary within-person reverse cross-lag is detectable at a standardised effect of approximately 0.013–0.014. Group B is expected to yield only approximately 32 (European incidence rates) to 106 (Australian treated-incidence bound) incident psychotic disorders, so its primary endpoint is amended to continuous psychotic-experience trajectories. Within-cohort Mendelian randomization is futile (power at most 0.37; instrument F ≈ 2.5–4.5 with the variants entered separately, about 22–170 as a single allele score); two-sample Mendelian randomization against consortium GWAS is adequately powered for odds ratios of about 1.2 per SD of liability only if the cannabis use disorder instruments explain at least about 0.5 per cent of liability variance (an assumption still to be checked), and is underpowered with genome-wide-significant lifetime-use instruments. Recruitment requires a national early-psychosis consortium harmonised with AMP SCZ. Current verdict. The synthesis supports all three processes, unequally. Reverse causation from genetic liability is the most consistently supported causal-genetic direction (Mendelian randomization odds ratios approximately 1.10–1.27). Shared liability removes roughly 40 per cent of the population log-odds in family designs. A modest forward increment, concentrated in heavy, high-potency, adolescent-onset use and in individuals at high polygenic liability, survives these adjustments but is not robust to correlated pleiotropy, and is not visible as rising population schizophrenia incidence. The record contains the full protocol (PDF and Markdown source) and the complete simulation code, parameter register and outputs (MIT licence), which reproduce every power and expected-event result in the protocol.