Keywords: Instrumental variables, perrson-centred treatment effects
A major challenge in comparative effectiveness research is that some effect modifiers are unobserved. We consider local instrumental variable methods for estimating person-level treatment effects. We assess the comparative effectiveness of ICU transfer versus continued general ward care on 7-day hospital mortality, for deteriorating ward patients in 49 UK hospitals. To address unobserved confounding, we use ICU bed availability as an instrumental variable (IV) for whether or not a patient is ICU-transferred. Our local-IV methods estimate the patient-level effects of ICU transfer on 7-day mortality. We report effects overall, and for pre-specified subgroups including the National Early Warning Score (NEWS). A total of 9,192 patients were included in the IV analysis. The predicted effect of ICU-transfer on 7-day mortality was a risk difference of -10.5% (95% CI -24.6% to 3.5%). The effect of ICU transfer differed by subgroup; for low risk patients, the estimated risk difference was -0.1% (CI: -9.8% to 8.6%), versus -34.2% (CI: -64.7% to -3.8%) for high-risk patients. We conclude by considering the wider role for local-IV methods in reporting person-level treatment effects.