Despite a large body of research examining the associations between medication use and dementia, the issue remains unresolved. This thesis uses administrative healthcare data to examine two drug classes, proton pump inhibitors (PPIs) and benzodiazepines, and dementia risk using a methodological sensitivity analysis approach to better understand why the evidence remains equivocal, and whether these drugs do increase the risk of dementia. Cox regression models were used to model risk using a population level cohort. Subgroup analysis was used when needed to reduce indication bias. Finally, high dimensional propensity score matched cohorts were used to reduce unmeasured confounding.

We found that while PPI users had modestly higher risk of dementia, this increased risk was due to higher rates of baseline comorbid conditions that are also risk factors for dementia. An analysis of PPI initiators also found that these conditions were predictors of treatment initiation and increased duration of use. After adjusting for the comorbid conditions, the association between PPIs and dementia was null. Similarly, benzodiazepine users had higher risk of dementia but also higher rates of dementia risk factors at baseline. Adjusting for these conditions reduced the estimated increased risk, although it remained significant. However, this class of drugs is most used in those with depression and anxiety, risk factors for dementia, resulting in potential indication bias. When this bias was reduced by examining cohorts of depressed persons, or of those with anxiety no increase in dementia risk was found.

The pharmacoepidemiological research into dementia risk associated with prescription drugs is messy. Insufficiently controlling for the noise present in non-randomized and observational data can lead to detecting signals of uncertain validity. It is hoped that this systematic approach will raise the bar for future research in this area, and that future researchers would, before publishing alarming findings, assess more closely whether they have truly controlled for confounding, reduce the risk of bias, and that their study design can answer the question they are trying to ask.