Dementia and Cancer: Unravelling Methodological Biases in a Population-Based Cohort

An inverse association between dementia and cancer has been consistently reported and often attributed to opposing biological mechanisms. However, methodological biases such as collider stratification and surveillance bias may underlie this finding.

We examined the association between dementia and subsequent cancer diagnoses among 4,532 participants of the population-based Rotterdam Study. Individuals with incident dementia were matched to up to two participants without dementia by age and sex; importantly, we also aligned the start of follow-up between these groups to reduce immortal time bias and ensure comparable observation windows. Cancer outcomes included both pathology-confirmed and non-pathology-confirmed (NPC) diagnoses to address potential surveillance bias. We applied cause-specific Cox proportional hazards models and Fine and Gray subdistribution hazard models to estimate cancer risk, accounting for the competing risk of death.

An initial analysis of the full cohort, assessing the association between ever having dementia and ever having cancer, replicated the previously reported inverse association (hazard ratio [HR] 0.58; 95% CI: 0.50-0.68). However, after matching and aligning follow-up periods, Kaplan-Meier curves indicated a lower probability of remaining cancer-free among individuals with dementia compared to those without dementia - particularly when including NPC diagnoses, which were more frequent among participants with dementia. In these analyses, the direction of the association reversed (HR 5.23; 95% CI: 3.65-7.48), suggesting a higher cancer risk among individuals with dementia. This elevated risk persisted in competing risks analysis (subdistribution HR 2.54; 95% CI: 1.80-3.58), suggesting that the elevated cancer risk was not solely attributable to differential mortality.

The frequently reported inverse association between dementia and cancer may reflect methodological artefacts rather than biological antagonism. Accounting for collider stratification and surveillance bias revealed a potential positive association, underscoring the importance of bias-aware epidemiological methods in research on comorbidities of ageing.