Primary Submission Category: Randomized Studies
A tie-breaker design for pragmatic clinical trials
Authors: Minh Nguyen, Tim Morrison, Art Owen, Michael Baiocchi,
Presenting Author: Minh Nguyen*
We study the problem of estimating the causal effect of being admitted to an intensive care unit (ICU) on health outcomes. We employ a tie-breaker design, introduced in Owen and Varian (2020), which limits randomization to a small window of covariate space. Similar to discontinuity designs, tie-breakers sort patients using a running variable. Those above an upper cutoff are given treatment and those below a lower cutoff are not. Those patients between the cutoffs are randomized at some fixed probability of receiving treatment.
We propose a novel, pragmatic study design which modifies the existing tie-breaker design in order to accommodate practical constraints faced by care providers. Using Electronic Health Record data, Nguyen et al. (2021) developed models to predict emergency room patients’ risk for ICU admission – which we use as this study’s running variable. The first modification is forced by the constraint of bed availability, which requires (a) forecasting bed availability and (b) a variable probability of receiving treatment. The second modification allows for physicians to override the study design assignment when necessary. This reframes the tie-breaker design into an encouragement trial. Both modifications afford instrumental variable-type analyses. Finally, we document how a careful analysis of the physicians’ reasons for overriding the assignment to treatment level is valuable for improving clinical decision-making and understanding heterogeneous ICU benefits.