We showed that a cancer diagnosis and an approaching death were both associated with increased use of symptomatic medications and decreased use of preventive medications. We further illustrated the pharmacoepidemiologic challenges in analyzing and interpreting medication changes at specific timepoints. Analyzing prospectively (i.e., anchoring at cancer diagnosis) and stratifying on length of survival illustrated that the changes in relation to a cancer diagnosis were mainly driven by patients who died within 2 years. In addition, analyzing retrospectively (i.e., anchoring at death) showed that medication usage near death was less dependent on the length of survival and presence of a cancer diagnosis.

The study’s main limitation is that medication adherence is not fully accounted for. However, the data stems from redeemed prescriptions, which increases the likelihood of actual consumption compared with using issued prescriptions [20]. Another limitation is that although cancer types have different prognoses, we do not stratify by individual cancer type, thus limiting the clinical inference and interpretability of our findings. However, such stratification was considered out of scope for this study.

The increase in medication usage prior to a cancer diagnosis is expected to, at least in part, be explained by reverse causation, i.e. early symptoms of the cancer diagnosis triggering new medical treatment [21]. As for medication usage after the cancer diagnosis, a shorter length of survival was correlated to lower use of preventive medication and more use of symptomatic medication.

Our results could be interpreted as physicians, to some extent, were able to predict the life expectancy of patients with cancer and revisit their medication accordingly, i.e., discontinue preventive medication while initiating symptomatic medication to patients with shorter life expectancies. However, it is noteworthy that our analyses were aggregated, and individual patient trajectories were not considered.

A clinical key point from our results is that 60–80% of all patients were treated with preventive medication right up until their death. This might seem excessive considering that these medications should be minimized in patients with limited life expectancy. On a positive note, our results show that these numbers were lower in patients treated after 2010. Measures such as increased education on the often limited beneficial effects and potential side effects of preventive medications might help physicians further accommodate the current guidelines and overcome some barriers related to discontinuing medications near end-of-life. However, predicting life expectancy is challenging. Thus, the clinical usefulness of these retrospective analyses remains limited.

For pharmacoepidemiologists, our findings highlight the need for carefully considering the aim of describing medication changes near end-of-life. If the aim is to inform clinical decision making, a prospective approach is appropriate, i.e., anchoring on an event such as cancer diagnosis or other clinical transitions. In contrast, a retrospective approach, i.e., anchoring on death, would be preferred if the aim is to describe care trajectories up to death. However, as we have shown, researchers need to acknowledge the drawbacks of both methods. Whereas the prospective analysis in the aggregate dilutes marked medication changes among some individuals, the retrospective analysis complicates the clinical inference as it leverages information unavailable to clinicians treating the patients.