Title: Estimating the effect of insulin use on outcomes in people with cystic fibrosis-related diabetes using causal prediction
Authors: Emily Granger - London School of Hygiene and Tropical Medicine (United Kingdom) [presenting]
Freddy Frost - University of Liverpool (United Kingdom)
Ruth Keogh - London School of Hygiene and Tropical Medicine (United Kingdom)
Abstract: Cystic fibrosis-related diabetes (CFRD) is associated with poor clinical outcomes for people with cystic fibrosis. Insulin is recommended as the only treatment for CFRD, but little is known about the consequences of long-term insulin use. Given the relatively low numbers of people with CFRD, running sufficiently powered randomised control trials is extremely challenging. We, therefore, aimed to estimate the effect of long-term insulin use on health outcomes in people with CFRD using longitudinal observational data. In a randomised trial, we would compare average outcomes in those randomised to treatment to those randomised to non-treatment. However, with observational data treatment status depends on individual characteristics that also affect the outcome and the treated and untreated are not directly comparable. We use counterfactual prediction models to predict the counterfactual outcome for each individual, i.e., the expected outcome we would have observed if the treated group were untreated and vice-versa. We used data from the UK cystic fibrosis registry. The longitudinal nature of these data gave rise to several challenges, such as time-dependent confounding, time-dependent eligibility and uncertainty in the direction of causal pathways. We will discuss how we tackled these challenges using two different approaches to counterfactual prediction: inverse-probability-of-treatment weighting of marginal structural models and the g-formula.