A clinical trial should be sized with full rigor, the freedom to weigh a result as part of a larger picture comes after the data.

A meditation on cognitive difference, prompted by a friend who sees colors in numbers, that ends up being a case for pairing every visualization with the statistics that back it up.

Find out why.

Using the history of the mile record as an analogy, this piece argues that drug development's stubborn 10–15% success rate isn't stagnation — it is a rising bar. Even with the advent of AI, industry will once again find itself settling there due to the rising bar: then what?

Why "highly improbable" and "impossible" deserve very different treatment — in cosmology, theology, and in how you build a statistical model.

More data always helps — but this post works through why that answer is trite, and what a rigorous, decision-focused measure of "confidence" actually looks like.

Using math derived from standard trial design, the piece argues that a 'failed' trial implies lack of clinical meaningfulness whose certainty is governed by the designed power. This only holds if your variance assumptions are correct, thus checking a confidence interval matched to the study's power checks against being misled.

The problem was never having early knowledge of a trial's outcome — it's stopping or acting on that knowledge outside a pre-planned framework — and a simple mixture model shows the knowledge was partially recoverable all along.

True excellence in statistics can never be fully attained, only endlessly approached through a widening circle of knowledge that statistics alone can never provide.

AI will increasingly absorb operational work, forcing statistics organizations to rethink how they develop scientific judgment.

Like the parable of Falamusha, how people begin their careers shapes their future value — and smart organizations in the AI age must invest in that journey rather than abandon it for short-term efficiency.

The same moral calculus that separates the bystander from the busybody applies surprisingly well to the obligations of a scientist who finds errors in published work.

Numbers are a language, and like any language, partial fluency is enough to get by but not enough to avoid being misled.

By inverting the question asked in a widely cited 2000 paper, a new metric reveals that dose proportionality is better understood as a continuous property of a compound than as a binary regulatory verdict.

Through the lens of bioequivalence, the study everyone dismisses as cookie cutter, a statistician discovers that the simple and boring can be anything but.

A statistician's argument that biology driven statistical sophistication is worth much more than statistical sophistication alone.

Good regulation solves the problem it was designed for and quietly creates the conditions for a deeper problem. Science and scientists have not been spared.

Precision is a curse until it becomes a virtue. Born from a snarky LinkedIn comment and precision, a useful metric emerges.

MCP‑Mod literally changed the dose response world for inference. Using this as inspiration, I propose a means by which inference can also be applied to multiple concentration response models.

If you think ‘not significant’ means ‘nothing learned,’ this post politely disagrees, and draws the line between uncertainty and inadequacy.