The question
Most wearables count steps by tripping a counter when acceleration crosses a fixed threshold. Stride dynamics scale with body geometry, so a single threshold could miss more steps for smaller bodies, which would land unevenly on women. I wear an Oura ring, which is what made the question concrete enough to want to answer with a model rather than a hunch. This was my final project for ISyE 6420, Bayesian Statistics, at Georgia Tech (Spring 2026).
What I built
The Kuopio Gait Dataset records 47 usable subjects (17 women, 30 men) walking at three speeds, captured at the same time by body-worn IMUs, optical motion capture, and floor force plates, with caliper-measured anthropometrics. I wrote my own heel-strike detectors for the IMU and force-plate signals, converted both to cadence, and modeled the difference.
Four nested models in PyMC, from a single pooled Student-t up to one that decomposes the sex effect into leg length, hip width and mass, compared with leave-one-out cross-validation at both the trial and subject level. Escalation rules for when to add model complexity were written down before fitting.
What came back
- Female-minus-male cadence error: -0.32 steps per minute, 95% interval [-2.03, +1.17]
- Every anthropometric coefficient crosses zero; the demographic models add no predictive value over the structural baseline
- Effects larger than about 1.5 steps per minute per standard deviation of body geometry are ruled out
A 47-person barefoot lab study in Finland is not the last word on a real sidewalk in real shoes. The pipeline (23 GB of raw archives in, four diagnosed models out) is fully reproducible, and the public repo has the code, notebooks and full report. Phase two is an interactive tool that takes your measurements and returns the posterior bias for your profile.