What happens when researchers stop a clinical trial early—not because something went wrong, but because the evidence became convincing faster than expected?

In this episode of Research Notes, I talk with statistician David Macleod about a fascinating Bayesian adaptive trial embedded directly inside a mobile eye-screening program in Kenya.

The project began with a practical global health problem: people were being screened for eye disease using a smartphone app called Peek Acuity, but many patients who were referred for additional care never showed up at the clinic. Researchers worked with communities to understand why and developed a simple intervention: brief motivational counseling and reminder messages delivered at the time of screening.

Rather than running a traditional fixed-sample randomized controlled trial, the team embedded a Bayesian adaptive trial directly into the app itself. Every week, the researchers updated their estimates using incoming data and evaluated whether the intervention appeared effective enough to stop early.

We discuss:

• How Bayesian adaptive trials differ from traditional randomized trials
• What “priors” are and why they matter
• Why the study stopped after only four weeks
• The tradeoff between speed and certainty in real-world evidence generation
• The risk of false positives in adaptive designs
• How low-cost interventions may justify different statistical thresholds
• Why rigorous methods still matter in implementation science and global health

We also talk about David’s unconventional career path from electrical engineering and telecommunications into epidemiology and biostatistics at the London School of Hygiene and Tropical Medicine.

Read the companion research note on Global Health Research in Practice: https://ghrbook.com/notes/adaptive-trial.html

Research Notes is a podcast and video series exploring how health research actually works: the methods, reasoning, tradeoffs, and decisions behind published studies.

In this episode of Research Notes, I talk with Dr. Gabe Loewinger of the National Institute of Mental Health about a core challenge in psychedelic clinical trials: participants often know whether they received the treatment. This “functional unblinding” means that people in different trial arms can develop very different expectations about what the treatment will do, and those expectations may themselves influence outcomes.

We explore why common approaches—like adjusting for what participants believe or expect—can actually make things worse from a causal inference perspective. Expectancy is not just a nuisance variable to control for; it sits on the pathway between treatment and outcome. Instead, we discuss an alternative framework based on estimating controlled direct effects, which aims to put treatments on a more equal footing by explicitly accounting for how expectations operate. Along the way, we use causal diagrams and thought experiments to clarify why this is such a difficult problem.

In this episode of Research Notes, I talk with Dr. Rohan Khazanchi about his recent paper in JAMA Internal Medicine examining what happened when race was removed from equations used to estimate kidney function.For years, these equations overestimated kidney function for Black patients, delaying access to kidney transplant waitlists and reducing the likelihood of receiving a transplant. We discuss how that happened, what changed, and how a national policy attempted to repair the harm by giving patients time back on the waitlist.We also focus on the methods behind the study—specifically why the team used an interrupted time series design to evaluate the policy and why a difference-in-differences approach ultimately didn’t fit the data.Topics covered:-How race became embedded in kidney function equations-Why removing race changed transplant eligibility-The policy that allowed patients to receive time back on the waitlist-Why the team chose interrupted time series over difference-in-differences-What the data showed after the policy was implemented-Who benefited—and who may have been left outKey finding: Patients received a median of about 1.7 years of additional waitlist time—substantial given typical wait times of 3–5 years.

What does “clinically meaningful” actually mean in psychiatry?

Compass Pathways recently reported Phase 3 results for COMP360, a synthetic psilocybin treatment for treatment-resistant depression. The company said 39% of treated patients achieved a “clinically meaningful” reduction in symptoms.

But who decides what counts as meaningful? And how should we interpret a 3–4 point difference on a scale like MADRS?

In this episode of Research Notes, I talk with Dr. Jerrold “Jerry” Rosenbaum, Stanley Cobb Professor of Psychiatry at Harvard Medical School and director of the Massachusetts General Hospital Center for the Neuroscience of Psychedelics.

Dr. Rosenbaum was not involved in the Compass study, but he has been closely watching the field and was quoted in STAT News saying the results “probably meet the bar for approval” but do not “shout out to you that this is miraculous.”

We discuss:

• What makes a treatment effect clinically meaningful in psychiatry
• How clinicians think about response, remission, and symptom scales like MADRS
• Why Compass introduced a new category of “clinically meaningful” improvement
• How restrictive trial criteria can make psychiatric studies hard to interpret
• Why average effects may hide meaningful benefit in subgroups
• Whether a 3–4 point difference on MADRS matters clinically
• Why durability, cost, and functional unblinding matter for psychedelic treatments

A key point from Dr. Rosenbaum: psychiatric trial outcomes are not just numbers on a page. They are consensus-based tools meant to approximate something much messier and more human — whether a person is suffering less, functioning better, and able to live their life again.

For more:
https://ghrbook.com/notes/clinically-meaningful.html

In this interview, I speak with Dr. Judith Lieber (London School of Hygiene & Tropical Medicine) about her recent paper in The Lancet Global Health examining episiotomy and postpartum haemorrhage in women with moderate or severe anaemia.I originally came across this paper while searching for a real-world example to teach directed acyclic graphs (DAGs). It turned out to be a perfect case: clinically important, analytically rigorous, and explicit about how a DAG guided the study design and adjustment strategy.The study draws on data from the WOMAN-2 trial — a large, international trial of tranexamic acid conducted in Pakistan, Nigeria, Tanzania, and Zambia, focused on postpartum bleeding in women with moderate or severe anaemia. Judy joined the trial team toward the end to conduct exploratory analyses using this rich dataset of over 15,000 women.In this conversation, we focus primarily on methods:-How drawing the DAG clarified the causal question-How it determined what to adjust for — and what to avoid adjusting for-The challenge of distinguishing confounders from mediators-Using proxies when key confounders (like shoulder dystocia) are unmeasured-Conducting a quantitative bias analysis to bound potential unmeasured confounding-Balancing complexity and readability when building a DAGWe also discuss Judy’s pathway into epidemiology, her work at LSHTM’s Clinical Trials Unit, and her current project tackling time-varying treatment decisions with another (even more complicated) DAG.This is a practical, applied conversation about how causal diagrams are actually used in real research — not as theoretical exercises, but as tools for clarifying assumptions, structuring models, and understanding limitations.If you teach causal inference, work with observational data, or are trying to move beyond “control for everything” regression thinking, this is a great example of DAGs in action.