Articles, newest first
Staggered Rollouts and Difference-in-Differences: When Two-Way Fixed Effects Get It Wrong
A feature rolls out to regions in three waves. The panel regression with region and month fixed effects says the effect is 0.6. The true average effect on th...
Read article
Reproducible Randomness Is More Than Calling set.seed()
A function can be reproducible and still behave badly. Calling set.seed() internally fixes its own Monte Carlo result but also replaces the caller's random-n...
Read article
Manuel Pinto Coelho: Preventive Medicine, Scientific Overreach and the Weight of Evidence
Several public claims by Manuel Pinto Coelho begin with legitimate preventive medicine concerns and real biological mechanisms. The problem is the recurrent ...
Read article
Propensity Scores: Matching, Weighting and the Estimator That Forgives One Mistake
Four estimators agree on the effect when both models are correct. Misspecify the outcome model and regression adjustment is off by 0.19; misspecify the treat...
Read article
When Unlabelled Data Makes Semi-Supervised Learning Worse
More data is only useful when the assumptions connecting it to the labelled problem are reasonable. In a controlled experiment, self-training begins with a s...
Read article
Sampling Uncertainty Can Dominate Representation Uncertainty
A better representation can recover structure that a linear summary cannot see, but that does not mean representation choice is immediately the main source o...
Read article
Ratio Metrics in A/B Tests: The Session-Level Test Lies and the Delta Method Fixes It
An A/A test on conversion per session, with two thousand users per arm and about three sessions each, comes back significant one time in ten. Nothing changed...
Read article
Inflammation Is Not a Diagnosis
"Inflammation" has become a universal explanation online. In biology it is a family of processes that differ by tissue, trigger and duration, and its most us...
Read articleWhen One Simulation Metric Tells the Wrong Story
A parameter setting can have the smaller typical estimation error and still be much more sensitive to starting values. Another can look poor under a toleranc...
Read article
Sample Ratio Mismatch: The One Diagnostic That Invalidates an Experiment
The experiment assigned a million users evenly and logged 49.75 percent in the treatment arm. That is 2,470 users missing, a one-in-a-million coincidence, an...
Read article