Tempered Thompson Algorithm
RCTClinical Trial
Standard randomized controlled trials allocate subjects equally across treatment arms to maximize statistical precision, but this can assign many vulnerable individuals to inferior interventions [1, 2]. Pure bandit algorithms assign subjects to high-performing arms to maximize participant welfare, but vanishing allocation probabilities for suboptimal arms cause severe inferential bias and high variance [3].
Picture this
Imagine a doctor testing four different medical treatments on arriving patients [3, 4]. Instead of flipping a fair four-sided coin for every single patient throughout the entire study, the doctor tracks which treatment is working best [2, 5]. For each new patient, the doctor flips a biased coin: 80% of the time, the patient receives whichever treatment currently looks best based on historical data, while 20% of the time, the patient is completely randomized among all four options [4]. This guarantees that every treatment maintains a minimum chance of being tested (e.g., 5%), allowing researchers to reliably gather data while still giving most patients the top treatment [4, 6].
What the evidence says
At gamma = 0.2, the minimum allocation probability for any of the 4 treatment arms is bounded below at 5% [4, 6, 7]. In counterfactual simulations using 2-month employment outcomes, the algorithm doubled the employment gains compared to a standard randomized controlled trial [11, 12].
- Who was studied
- N = 3,770 individuals (1,663 Syrian refugees and 2,107 Jordanian jobseekers) in urban Jordan [7-9].
- How
- Adaptive targeted field experiment using a Tempered Thompson Algorithm with mixing parameter gamma = 0.2 and 16 demographic strata, evaluating rapid 6-week employment as well as 2-month and 4-month follow-up outcomes [4, 7, 10].
What to do
Set the mixing parameter gamma in a multi-armed bandit algorithm to enforce a lower bound gamma/k on assignment probabilities for all k treatment arms to balance participant welfare and frequentist inference bounds [4, 13].
From the source
"Our Tempered Thompson Algorithm combines the estimation objective of conventional RCTs with the welfare-maximizing objective of bandit algorithms." [5]
An Adaptive Targeted Field Experiment: Job Search Assistance for Refugees in Jordan