Using Surrogates in Covariate-adjusted Response-adaptive Randomization Experiments with Delayed Outcomes

Authors: Lei Shi, Waverly Wei, Jingshen Wang

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Through theoretical investigations and a synthetic HIV study, we show that our design is more efficient than the design without incorporating any surrogate information. We present the case study results in Figure 1. In terms of bias, Figure 1(A) suggests that the point estimators from all three designs have a vanishing bias as the sample size grows, validating the asymptotic unbiasedness of all strategies. Nevertheless, in terms of variance, Figure 1(B) demonstrates that our proposed design yields smaller standard deviations and thus higher estimation efficiency for estimating the ATE.
Researcher Affiliation Academia Lei Shi Division of Biostatistics University of California, Berkeley 2121 Berkeley Way, Berkeley, CA 94704 leishi@berkeley.edu Waverly Wei Marshall School of Business University of Southern California 3670 Trousdale Pkwy, Los Angeles, CA 90089 waverly@marshall.usc.edu Jingshen Wang Division of Biostatistics 2121 Berkeley Way, Berkeley, CA 94704 jingshenwang@berkeley.edu
Pseudocode Yes Algorithm 1 Surrogate-enhanced adaptive experiment with delayed outcomes
Open Source Code Yes We provide open access to the code in the Supplementary Materials.
Open Datasets Yes In this synthetic case study, we calibrate our data-generating process using data collected from an HIV randomized controlled trial (RCT) conducted in Tanzania [15]. [15] Carolyn A Fahey, Prosper F Njau, Emmanuel Katabaro, Rashid S Mfaume, Nzovu Ulenga, Natalino Mwenda, Patrick T Bradshaw, William H Dow, Nancy S Padian, Nicholas P Jewell, et al. Financial incentives to promote retention in care and viral suppression in adults with hiv initiating antiretroviral therapy in tanzania: a three-arm randomised controlled trial. The Lancet HIV, 7(11):e762 e771, 2020.
Dataset Splits No The paper describes a simulation study using parameters from an HIV trial to compare different design strategies ('Our proposed design strategy', 'complete randomization design', 'CARA design that only utilizes primary outcomes'), but it does not specify train/validation/test splits of the synthetic data or a model within these designs.
Hardware Specification Yes The simulation studies require approximately 160 minutes to complete when running on a Mac Book Pro equipped with an 8-core CPU.
Software Dependencies No The paper does not list specific software dependencies with version numbers.
Experiment Setup Yes In our synthetic data generation, we generate the surrogate outcome from a multinomial distribution. The delay mechanism is also generated following multinomial distribution using the parameters in Table B. We generate X as Xit Bernoulli(0.64), and the primary outcome variable Y as Yit|Ait = a, Xit = x, Sit = s N(τ(a, x, s)). The true average treatment effect of the primary outcome is τ = 0.04. Following our proposed design strategy, we set the total number of experimental stages as T = 2(D +1) = 8.