This study represents one of the largest observational investigations to date—and the first to incorporate MR—examining the relationship between cannabis use, cognitive function and dementia risk in older adults. Cannabis users demonstrated significantly better cognitive performance in cross-sectional analyses; however, these findings were not supported by longitudinal or genetic data, suggesting that residual confounding rather than a causal relationship is likely. Importantly, cannabis use was not associated with increased risk of dementia in either observational or MR analyses, indicating no supporting evidence of a causal link with cognitive decline in later life within the limits of the present data.

Cannabis users demonstrated better performance on numeric memory and fluid intelligence tests, consistent with some previous research,²¹ but contrasting with evidence that acute cannabis intoxication leads to small-to-moderate impairments in working memory and executive functioning, and that chronic or heavy use is associated with deficits particularly in verbal learning and memory, with more mixed evidence for executive domains.⁷ These inconsistencies may be attributed to variations in age, timing of exposure, dosage, frequency of use or sample characteristics and may also reflect differences between the biology of CanLU, associated with higher educational attainment, and CanUD, associated with lower educational attainment.

The observed association between moderate cannabis use and higher cognitive performance in cross-sectional analyses is unlikely to reflect a direct cognitive benefit of cannabis. Rather, it is more plausibly explained by residual confounding and selection effects. In UKB, cannabis use is socially patterned and associated with higher educational attainment and socioeconomic status, factors that are themselves strongly linked to cognitive performance and cognitive reserve. Individuals engaging in recreational, non-dependent cannabis use may therefore differ systematically from non-users in ways that are incompletely captured by available covariates. Consistent with this interpretation, the absence of supporting evidence from longitudinal or MR analyses, together with the lack of a dose–response relationship, further argues against a causal effect of cannabis use on cognition.

In addition to residual confounding, selection bias may also influence the observed associations. Participants included in the analytic sample were more likely to be socioeconomically advantaged and to have higher educational attainment than those with incomplete cannabis exposure or cognitive outcome data. Higher socioeconomic position and educational attainment are strongly associated with better cognitive performance and, in many settings, lower prevalence of heavy or dependent substance use. As a result, this non-random selection into the analytic sample could plausibly attenuate associations between cannabis use and adverse cognitive outcomes, biasing estimates towards the null in the context of true harmful effect of cannabis on cognition.

In the sex-stratified analysis, a significant interaction emerged only for fluid intelligence showing a positive association in males than females. Although previous studies have reported that sex moderates the cognitive effects of cannabis, with greater vulnerability to impaired visual memory among males and more pronounced impairments in attention and executive functioning among females,¹⁴ these patterns were not replicated in the present analysis. Instead, our findings suggest that sex-specific effects may extend to reasoning ability, highlighting the need for further research to elucidate the underlying mechanisms, including potential hormonal, neurobiological and behavioural factors. Given the absence of significant associations in the overall longitudinal analyses, we did not pursue sex-stratified longitudinal models, as further stratification would substantially reduce statistical power and was unlikely to yield robust estimates. Similarly, sex-stratified MR analyses were not feasible due to the lack of sufficiently powered sex-specific GWAS for cannabis-related traits.

Cannabis dose did not materially modify the association between cannabis use and cognitive performance, with no clear evidence of a dose–response relationship observed, raising important questions whether confounding factors, such as baseline cognitive abilities, personality traits or environmental influences, play a more substantial role. Cannabis users across all frequency categories performed better than non-users on numeric memory and fluid intelligence tests. However, it remains unclear whether cannabis use directly enhances cognitive function or if this association is confounded. Individuals with higher education and socioeconomic status are more inclined to use cannabis recreationally and less likely to develop dependence and may have greater cognitive reserve, which provides resilience against cognitive decline.²⁶

Consistent with these cross-sectional associations reflecting something other than a direct effect of cannabis on cognition, neither longitudinal analyses nor genetic analyses supported a causal link between cannabis use and cognitive performance. In the longitudinal analysis, counterintuitively, performance on some tests appeared to improve over time. We hypothesise this is due to the learning effect.²⁷ MR analysis revealed no significant associations between genetically predicted cannabis use and cognition. No significant associations were observed with dementia in either observational or MR analyses, contrasting with a recent study reporting higher dementia risk in individuals with incident acute care due to cannabis use (a more severe exposure phenotype and likely to reflect CanUD). The study did not, however, explore the duration, frequency or type of cannabis use, which may be highly relevant in understanding relationships.²⁸ The null MR findings should be interpreted in the context of modest genetic effect sizes for cannabis-related phenotypes and limited statistical power, particularly for cognitive outcomes. Rather than providing precise genetic prediction, the MR analyses offer complementary evidence that helps reduce confounding and reverse causation, strengthening confidence that the observed cross-sectional associations are unlikely to reflect a causal effect of cannabis use.

Dementia outcomes in the MVP were ascertained from linked EHRs. Although these records provide longitudinal clinical information, the exact onset of dementia is generally unknown and may precede the first documented diagnosis. Consequently, the null associations observed with dementia should be interpreted in the context that event times reflect the timing of first recorded diagnosis rather than true disease onset, introducing potential interval censoring. In addition, death was treated as a censoring event rather than a competing risk; if cannabis use or related covariates are associated with mortality, this could contribute to underestimation of dementia incidence and attenuation of effect estimates. Cannabis use has been most consistently associated with increased risk of cardiovascular disease, which may partly explain the absence of a clear association with dementia through competing risks, whereby individuals at higher cardiovascular risk may die before reaching the age at which dementia typically manifests. This is relevant despite vascular dementia accounting for approximately 10% of dementia cases, as competing mortality may attenuate observed associations with all-cause dementia in epidemiological analyses. While these factors are unlikely to substantially alter relative hazard estimates in this large cohort, future studies applying methods that explicitly account for interval-censored outcomes and competing risks may further refine estimates of dementia risk.

Our findings are broadly consistent with prior population-based longitudinal studies that have not observed accelerated age-related cognitive decline associated with cannabis use.^{9 10} Earlier work from the Epidemiologic Catchment Area study likewise demonstrated that cognitive decline over a 12-year follow-up period was closely related to ageing and educational attainment, but not to cannabis use.²⁹ The present study reinforces and extends these findings by leveraging substantially larger and more diverse population-based cohorts, examining a broader range of cognitive domains and incorporating rigorous adjustment for sociodemographic and health-related factors alongside complementary genetic analyses. In contrast, studies reporting cognitive decline associated with cannabis use, including findings from the New Zealand Dunedin cohort, have primarily focused on persistent, adolescent-onset and heavy cannabis use.³⁰ Differences in exposure severity, timing and study design likely explain these discrepancies, as the present study primarily captures lifetime or later-life cannabis use in population-based cohorts, and midlife cognitive differences observed in such cohorts may reflect early-life exposure or reduced cognitive reserve rather than progressive neurodegeneration attributable to cannabis use.

Discrepancies between observational and MR associations could result from residual confounding, statistical power differences or differences in exposures (later life compared with lifetime use). Higher educational performance of cannabis users in UKB raises the possibility that positive cross-sectional associations observed with cannabis may be confounded by educational background. Although we adjusted for a range of sociodemographic and health-related factors, including education and household income, these variables act as proxies and may not fully capture heterogeneity in socioeconomic circumstances, educational quality or early-life cognitive ability. In addition, broader lifestyle and environmental influences, such as diet, physical activity and health-related behaviours across the life course, are incompletely measured and may further contribute to residual confounding. This emphasises the need for cautious interpretation of observational studies.