Modeling and validation

Modeling estimates the underlying allele-frequency distribution between sparse measurements while making uncertainty and unsupported regions first-class results.

More than one model class

The working group builds Bayesian spatial models and welcomes geometric deep learning and other statistically defensible approaches. Models compete on scientific acceptance evidence rather than on visual smoothness or novelty.

One shared result format

Every model runs ahead of publication and produces the same kind of versioned result file. That file records the estimated frequency, its uncertainty, whether the available data support the estimate, the effective sample information, and the exact model and data versions. Published files are never overwritten. The data service can read them without needing to know how a particular model was fitted.

Validation before publication

The blocking checks are published geography and disease burden for haemoglobin S (HbS), published frequency and X-linked inheritance results for glucose-6-phosphate dehydrogenase (G6PD) deficiency, and measured carrier-screening rates. Calibration, prediction in locations withheld from model training, statistical diagnostics, correction for how participants were recruited, and stability all matter. A model that cannot beat an appropriate baseline or reproduce known science is refused.

Models are fitted with PyMC rather than INLA. Issue #34 chose a Hilbert-space Gaussian process approximation. The published surfaces and the HbS validation run instead use an inducing-point Gaussian process on an H3 grid over the sphere (#105), with a beta-binomial likelihood. Read Issue #34 before proposing a different runtime, and use the scientific objectives to keep alternative models on the same acceptance contract.