UC Berkeley

Reliance on traditional research methods in situations where innovative methodologies are superior can hinder inference. In global health research, where wasted resources can be equated with lives lost, non-adaptation of promising methods has dire consequences. Given the high cost of randomized controlled trials (RCTs), one major obstacle in these studies is the Type II error, in which an investigator fails to reject the null hypothesis when the null hypothesis is false. Unexpected disease transmission rates may lead studies to be underpowered and yield Type II errors. Furthermore, even if an RCT correctly rejects or fails to reject the null, data may not be fully utilized if investigators only analyze and report the intervention effect in the total study sample. Only focusing on quantifying the overall impact of the intervention of interest may neglect investigation into important observational relationships between covariates of interest or heterogeneity of the intervention effect in population subgroups, although these observational analyses and subgroups of interest should be pre-specified to maintain transparency and replicability. As participants were not randomized to receive exposures represented by covariate values, additional analytic rigor and inferential caution must be exercised in order to enable meaningful interpretation. My dissertation aims to explore three examples of potential methodological solutions to these challenges to inference in RCTs: a systematic review of the ring trial design, a targeted learning analysis of treatment heterogeneity, and an observational analysis of multiple biomarkers using trial data. Chapter 1. Spatiotemporal clustering: ring trials. For cluster RCTs in emergent and elimination disease settings, unpredictable spatiotemporal clustering often leads to imbalanced clusters, or even clusters with zero incident cases, which can reduce study power and limit inference. Ring trials, a trial design in which the units of randomization are responsively-defined clusters of individuals in social or physical proximity to an index case, may improve investigators’ ability to make inferences in these settings. Despite this potential utility, this RCT design remains under-examined and under-utilized. Investigators conducted a systematic review of the ring trial design to examine the existing applications of this study design as well as its benefits and drawbacks. We identified 26 ring trials, 15 cluster-randomized trials that used ring interventions, five trials that used ring recruitment and randomized within rings, and one individually-randomized trial that used a ring intervention. Ring trial designs require strong disease surveillance and contact tracing mechanisms, rapid intervention delivery systems, and a treatment with a strong post-exposure prophylactic effect. In these settings, ring trials can retain power despite unpredictable spatiotemporal clustering of the outcome of interest. Chapter 2. Heterogeneity in treatment effect: targeted learning analysis of treatment heterogeneity. Even if there is an effect of the intervention on the outcome among certain individuals, a study may fail to detect this relationship if the effect is heterogeneous in the study sample. Investigators can gain additional insight through analysis of the conditional average treatment effect, which is the treatment effect based on individual covariate status. Using data from the WASH Benefits study, which enrolled pregnant mothers and young children in rural Bangladesh, analysis of treatment heterogeneity can improve our understanding of child growth in low and middle-income countries. Despite the widespread use of water, sanitation, hygiene (WSH), nutrition (N), and combined (N+WSH) interventions, investigators have found mixed evidence regarding these interventions’ impact on child growth and development. Insufficient reduction of pathogens may explain WSH’s lack of impact, and environmental enteric dysfunction (EED), a condition of impaired intestinal permeability and inflammation, may modify the impact of WSH and N interventions on child growth. This study applied targeted machine learning methods to assess treatment heterogeneity of N+WSH, WSH, and N interventions on child growth by pathogen and EED biomarker status. We found that children with greater levels of myeloperoxidase, a gut inflammation biomarker associated with EED, and Campylobacter, a genus of bacteria that is associated with EED onset, had a greater effect of all treatments on growth. These results contribute to the body of literature characterizing individual predictors of N+WSH, WSH, and N intervention effectiveness as well as our understanding of EED. Chapter 3. Maximizing data utilization: observational analysis of high-dimensional data nested within a randomized controlled trial. Trials devote enormous resources to evaluating the effect of the randomized intervention in the study sample, but limiting analyses to only include this intervention may neglect the wealth of data that these RCTs can provide. In addition to analysis of the effect of the intervention, investigators can conduct observational analyses nested within an RCT, although these analyses will require additional methodological rigor in order to limit confounding and bias to enable meaningful inference. Data from the WASH Benefits study provide an opportunity to assess the relationship between stress neurobiology, an exposure that could not be ethically randomized, and child development. Stress has been implicated as a key pathway by which adverse circumstances can lead to developmental impairment, and prior studies have indicated a possible link between stress and subsequent development. This study evaluated the relationship between stress and development through an observational analysis nested within an RCT. We assessed physiologic measures of stress using measures of the hypothalamic-pituitary-adrenal (HPA) axis, the sympathetic-adrenal-medullary (SAM) system, and oxidative status. We constructed generalized additive models to compare development outcomes of children at the 75th and 25th percentiles of stress biomarker distributions while adjusting for potential confounders. We found that measures of HPA axis activity were associated with poor development outcomes. These observations support the use of HPA axis biomarkers, particularly cortisol and glucocorticoid receptor methylation, to indicate children who are at risk of poor developmental outcomes. This study explores and provides applied examples of RCT design and analysis methods that may improve efficiency in global epidemiologic research. This research serves to 1) improve our understanding of a neglected trial design and explore innovative methods of analyzing trial data; 2) identify which characteristics define amenability to N+WSH, WSH, and N interventions, providing insights that EED may be associated with treatment effectiveness; 3) evaluate the relationship between stress biomarkers and child development through observational analyses nested within an RCT, which supports the use of HPA axis biomarkers to indicate children at risk for poor developmental outcomes.