Hypothetical Estimands in Randomised Controlled Trials: Unifying Causal Inference and Semiparametric Theory
Authors
Jespersen, Louise Østerby ; Strøm, Silje Post
Term
4. term
Education
Publication year
2025
Submitted on
2025-05-24
Pages
95
Abstract
In randomized clinical trials, events that occur after treatment starts—such as taking rescue medication or discontinuing treatment—can blur the true treatment effect because they affect the link between treatment and outcome. A common approach is to target the hypothetical estimand, meaning the outcome we would expect if these intercurrent events did not occur. This thesis outlines a practical causal inference workflow for such trials and introduces key ideas from semiparametric models and the targeted learning framework as background. In current practice, the hypothetical estimand is often analyzed with a Mixed Model for Repeated Measures (MMRM). We propose using Longitudinal Targeted Maximum Likelihood Estimation (LTMLE) as an alternative for estimating the hypothetical estimand. Through simulations and an empirical analysis, we assess how MMRM and LTMLE handle varying amounts of intercurrent events. Our findings suggest that while MMRM is straightforward and easy to interpret, LTMLE is more robust because it more faithfully represents causal relationships when rescue medication and treatment discontinuation occur.
I randomiserede kliniske forsøg kan hændelser, der opstår efter behandlingsstart – som brug af redningsmedicin eller behandlingsophør – sløre den egentlige behandlingseffekt, fordi de påvirker sammenhængen mellem behandling og udfald. En udbredt tilgang er at fokusere på det hypotetiske estimand, dvs. det udfald man ville forvente, hvis disse interkurrente hændelser ikke fandt sted. Dette speciale præsenterer en praktisk arbejdsgang for kausal inferens i sådanne forsøg og introducerer centrale ideer fra semiparametriske modeller og den målrettede læringsramme som baggrund. I praksis analyseres det hypotetiske estimand ofte med et mixed model for gentagne målinger (MMRM). Vi foreslår at bruge Longitudinal Targeted Maximum Likelihood Estimation (LTMLE) som alternativ til at estimere det hypotetiske estimand. Med simuleringer og en empirisk analyse undersøger vi, hvordan MMRM og LTMLE håndterer forskellige mængder af interkurrente hændelser. Resultaterne tyder på, at MMRM er enkel og let at fortolke, mens LTMLE er mere robust, fordi den bedre afspejler kausale sammenhænge, når der forekommer redningsmedicin og behandlingsophør.
[This apstract has been rewritten with the help of AI based on the project's original abstract]
Keywords