Incorporating the efforts of senior academic and consulting biostatisticians, post-doctoral researchers, and advanced graduate students, CAPSL contributes novel clinical and translational science (CTS) solutions for the statistical design and analysis of clinical and translational research (CTR) studies.
Carolina Pragmatic Trials Laboratory (CAPSL)
The Carolina Pragmatic Trials Laboratory (CAPSL) supports the design and statistical analysis of pragmatic trials testing interventions in real-world settings.
Incorporating the efforts of senior academic and consulting biostatisticians, post-doctoral researchers, and advanced graduate students, CAPSL contributes novel clinical and translational science (CTS) solutions for the statistical design and analysis of clinical and translational research (CTR) studies.
CAPSL aims to achieve translational benefits and tackle translational roadblocksall in support of TraCS-wide efforts to translate scientific discoveries into new treatments and approaches to medical care.
The laboratory has two primary focus areas:
Cluster randomized trials are increasingly complex, often having multiple periods, multiple levels, or multi-component interventions. They include cohort and repeated cross-sectional samples in parallel, cluster crossover, and stepped wedge designs, respectively.
They span patient-centered outcomes, surveys, administrative, and electronic health records in a variety of settings including hospitals, jails, nursing homes, and primary medical practicesi.e., clustersoften targeting disadvantaged and vulnerable populations. They often include small numbers of clusters, resulting in challenges for valid design and statistical analysis.
CAPSL will build upon a recently completed PCORI-funded award that advances marginal modeling methods for stepped wedge and other complex cluster randomized trial designs.
TraCS supports cluster randomized trial requests that can be enhanced with novel CTS methods and tools developed by CAPSL. Tools include well-tested and user-friendly software to support the dissemination and implementation of novel methods.
Design and analysis methods for dynamic treatment regime estimation are essential ingredients of evidence-based precision health. These methods center on using data to create algorithms for assigning treatment to the right patient at the right time.
The designs address pragmatic concerns by seeking to reflect the structure of decision points in clinical care and public health policy as well as practical limitations and context. Such study designs include sequential multiple assignment randomized trials (SMARTs), micro-randomized trials for mHealth, observational study designs for precision health, and hybrid designs.
The main analytical tool for such data is off-policy reinforcement learning. CAPSL seeks to build on several recent research projects which arose out of TraCS research requests and collaborations to refine methodology for dynamic treatment regime estimation in a number of settings. These methods are included in the following projects:
- Design and analysis of the BEST (Biomarkers for Evaluating Spine Treatments) clinical triala two-stage SMART design that is part of the funded BACPAC study coordinated by UNC's Collaborative Studies Coordinating Center. BACPAC is a part of the NIH HEAL initiative.
- The HEDRA projectdeveloping artificial intelligence methodology for real-time guidance of exercise in patients with type 1 diabetes. Research here involves off-policy reinforcement learning and real-time mHealth data.
- Use of electronic health record data from patients with Crohn's disease to develop dynamic treatment regime estimation to prolong time to second surgery.
TraCS supports such dynamic treatment regime estimation project-related requests that can be enhanced with novel design and reinforcement learning methods and tools developed by CAPSL. Tools include well-tested and user-friendly software to support the dissemination and implementation of novel methods.
View recent CAPSL Highlights:
CAPSL Director John Preisser supported Dr. Emily Ciccone, a member of UNC's Institute for Global Health and Infectious Diseases, providing data analysis for a novel stepped wedge cluster randomized trial to improve the stewardship of acute respiratory illness (STAR).
Enrolling 1,280 children in 15 rural villages in Uganda, this STAR trial demonstrated that point-of-care C-reactive protein measurement by 65 community health workers effectively reduced antibiotic use among children with respiratory illness, while having little to no impact on clinical outcomes. Preisser, assisted by TraCS GRA Ms. Di Hu, employed novel statistical methods in the design (including power computations) and analysis of the trial.
Preisser believes that this may be the first comprehensive application to a CRT of the statistical methods and software tools generated by Preisser's PCORI-funded methods study completed in 2024.
Dr. Ciccone, Preisser, Hu, and others published results in PLOS Medicine on August 19, 2024:
Ciccone EJ, Hu D, Preisser JS, Cassidy CA, Kabugho L, et al. (2024) Point-of-care C-reactive protein measurement by community health workers safely reduces antimicrobial use among children with respiratory illness in rural Uganda: A stepped wedge cluster randomized trial. PLOS Medicine 21(8): e1004416. doi.org/10.1371/journal.pmed.1004416.
The UNC Institute for Global Health and Infectious Disease described the study and its prospective wider positive implications for antibiotics stewardship, including in low-resource settings, in the news release: Point-of-Care Diagnostics Guide Antibiotic Use in Rural Uganda (August 22, 2024).
Crohn's disease (CD) is an autoimmune disease that has not been well characterized in the pediatric population due to the limited number of prospective cohort-based studies. Consequently, current postoperative pediatric practice guidelines are informed by expert consensus as opposed to evidence-based guidance, which leaves significant variance in postoperative management in the real world.
CAPSL Deputy Director Michael Kosorok and GRA Ms. Jane She teamed with Dr. Matthew Egberg, Assistant Professor, UNC Pediatrics/Gastroenterology, to approach this problem using novel methods. The team was interested in using administrative insurance claims data, which reflect everyday clinical practice, to bridge the gap in knowledge of treatment characterization in the pediatric population without expending the resources to recruit patients for prospective studies, which may take years to complete.
As a key CAPSL research objective, Kosorok and She proposed the development of a novel dynamic treatment regime estimator for time-to-event data characterized by a large number of treatment stages. If successfully developed, the dynamic treatment regime estimator could be used in the pediatric CD population to prolong patient survival time between the first and second surgeries. The estimator could also be applied to any time-to-event disease setting with a large number of treatment stages.
GRA Jane She has taken the lead in conducting novel simulation work critical to ultimate development of the dynamic treatment regime estimator. She explains,
"As a GRA, my job was to convert any clinical discussions and decisions to a data pipeline and code. This began with data preprocessing, to simplify use of administrative claims, where present many challenges in defining stages clearly. We then began the coding of the algorithm, which was an iterative process, translating the theory and big picture from meetings with Dr. Kosorok to functional, generalizable code. After the code started working, we began working on simulation settings to examine finite-sample performance of our method in various scenarios such as greater/fewer stages, more/less censoring, and so forth.
Our early attempts focused on base case scenarios, with a small number of stages, running the code line by line, just to figure out which portions were causing errors, and making sure each step was giving the expected output. After extensive debugging, we began running a large number of simulation replicates, which sometimes might take days to run.
Currently, we are able to demonstrate promising simulation results. We are now focusing our efforts on proving the theoretical properties of our estimator using mathematics, and on applying our method to our Crohn's disease administrative claims data of interest."
View a summary of Ms. She's research progress to date (pdf).
CAPSL post-doctoral researcher Dr. John Sperger is lead author on the following publication describing novel stepped wedge study design:
Sperger J, Kosorok MR, Linnan L, Kneipp SM. Multilevel Intervention Stepped Wedge Designs (MLI-SWDs). Prevention Science. 2024 May 15:1-3. doi-org.libproxy.lib.unc.edu/10.1007/s11121-024-01657-y
This work grew out of a collaboration between the TraCS CAPSL research team and investigators leading the North Carolina Works for Health (NCW4H) study. The NCW4H study inspired the study design generally, but the intricacies of the study required additional adjustments beyond the standard use case.
NCW4H aims to develop and evaluate a multilevel intervention for reducing the negative health consequences stemming from unemployment. Unemployed individuals are enrolled and randomized to either no intervention beyond the standard support provided by the local department of health and human services (DHHS) or, alternatively, to the Chronic Disease Prevention Program, a lifestyle and behavior change intervention, in addition to DHHS support. Once a study participant gains employment, their supervisor is contacted and, if consented, the individual is then randomized to either no intervention or, alternatively, to the Supervising for Success (S4S) program. NCW4H was planned and launched as a 2×2 factorial design, but COVID-related and logistical challenges required adjusting the design after enrollment. The study population and nature of the interventionsone pre-employment, one post-employmentcreates the unique challenge that participants do not begin in a cluster (those managed by the same supervisor at an employer) but join one over the course of the study. Also, once a supervisor has been randomized, any additional hires of study participants will belong to the same cluster-level condition because the supervisors can't be re-randomized.
Sperger discusses his approach in guiding novel study design appropriate to these considerations:
"Behavioral and public health interventions are often packages comprised of multiple component interventions that target different levels of the socio-ecological determinants of health, such as the individual, family, organizational, or state levels. Yet the common approach to developing and testing these packages of interventions involves taking intervention components that were validated in isolation and then testing the combined intervention by randomizing participants to receive all or none of the combined intervention. This precludes estimating the effect of each component of the intervention or their interaction, only the combined intervention effect can be estimated.
Our proposed design combines randomization at both the individual and cluster levels and utilizes a stepped wedge design at the cluster level. Randomizing at both levels enables estimation of the individual effect of each intervention component as well as their interaction, and importantly doesn't reduce the power to test the combined intervention effect in most common settings. This can be a statistical 'free lunch,' but there is an operational cost to randomizing at both the individual and cluster levels instead of only the cluster level as in current standard practice.
There are also important limitations on when the design is appropriate. First, it must be logistically feasible to randomize individuals. Second, investigators should be confident that spillover effects from one individual to another in the same cluster are unlikely. The plausibility of this assumption depends on both the specifics of the individual-level intervention and how individuals are clustered. An intervention that has a high risk of spillover for individuals in the same household may not be risky if the cluster units are cities."