The Duck, the Data, and the Sceptic: Eight Years of Learning to Ask Better Questions
In my second year of my MSc, I hosted my first journal club. The paper was on Mixed Effect Machine Learning, a framework for handling the correlated observations that show up constantly in clinical data, where the same patient appears in your dataset dozens of times.
I was genuinely excited. The core idea was elegant: decompose the outcome into a population-level signal and a subject-specific noise term, train your ML model on the clean signal, and you get better predictions on clustered data.
My excitement was entirely about the mechanics. I remember being absorbed by the EM algorithm’s convergence logic: the way it iteratively partitioned variance until the fixed and random effects stabilized. I presented it the way I understood it: as a solution to a technical problem. The random effects were a nuisance to be mathematically isolated. The win condition was a better AUROC.
I hadn’t yet thought to ask what information was just left on the table.
Before the MSc, I was a dental house officer in Yangon. Our team spent days in the field measuring periodontal pocket depths — six sites per tooth, twenty-eight teeth per patient — and we were still short of our target sample size. We knew, without being able to articulate it precisely, that a smarter way to prioritize candidates existed. We just didn’t have the tools.
That frustration is what sent me to Mahidol. I arrived wanting tools. For the first two years, tools were all I wanted.
Last month, I hosted another journal club, this time later in my academic journey, on a paper titled Assessing the Replicability of RCTs in RWE Emulations. The paper comes out of the RCT DUPLICATE initiative, which has spent years trying to replicate randomized trial findings using routine healthcare claims data. The central question isn’t “can we model this?” It’s “should anyone believe what the model is telling them?”
The paper’s proposed framework of the Sceptical P-Value approaches replication the way a rational skeptic would. You start by quantifying the maximum disbelief the original RCT result can survive, given its own uncertainty. Then you ask whether the real-world evidence conflicts strongly enough with that disbelief to reject it. It’s not two separate significance tests; it’s a joint credibility assessment. The win condition is epistemic, not predictive.
What struck me preparing this presentation was how different my instincts were compared to 2021. I wasn’t excited about the math first. I was thinking about the TRITON-TIMI 38 emulation in the case study which is a trial of prasugrel versus clopidogrel and about who the evidence was quietly built for. The formal exclusion list was short: intracranial pathology, anemia, thrombocytopenia, pregnancy. But the enrolled population tells the real story. 92.5% Caucasian. A third from North America. Virtually everyone with immediate access to a catheterization lab. In a fragmented healthcare environment like the one I came from, that profile isn’t the majority — it’s the exception.
The variance I once called a nuisance, I now call uncertainty. The patients I once called exclusions, I now call selection bias. These aren’t just terminological upgrades. They reflect a different relationship to the question.
In 2021, I was asking: how do I fit this correctly?
In 2026, I’m asking: is this evidence trustworthy enough to act on — and for whom?
The Yangon fieldwork and the sceptical p-value are, at some level, the same problem at different scales. A measurement you couldn’t take because the patient never came back, and a treatment effect you can’t trust because the trial never enrolled people like them, both are absences that distort the evidence. Learning to see those absences, rather than just optimizing around them, is the work I didn’t know I was signing up for.
In the intersection of clinical logic and data science, there is a specific kind of in-between existence. The duck belongs to both the water and the sky, and the land between them. Eight years in, I think the most important thing that land has taught me is this: the data tells you what it saw. Your job is to remember everything it didn’t.
