Aligning stratum classes to the reference

Machinery for the stratified refit drift read. That read re-estimates the four-class general finite mixture model within each stratum, so it is the refit path the archived refit pilot used; the primary invariance read is now the kernel effect size (computing the invariance effect size), which freezes the reference and needs no alignment.

A stratum fit numbers its classes in whatever order it converges on, so before any movement can be measured the stratum’s classes have to be matched to the reference’s. Every proband in a stratum carries two labels, the class its stratum fit gave it and the class the pooled reference gave it, and the alignment compares those two labellings directly.

The alignment is a swappable method, chosen at the command line and recorded in the run hash. Two are defined:

• Membership alignment (the default) pairs classes by who is in them, maximising the total Jaccard overlap across the four classes by Hungarian assignment. It reports each pair’s Jaccard as a confidence and the adjusted Rand index for the partition as a whole.

• Centroid alignment pairs classes by profile, matching on the standardised distance between centroids. It is a cross-check, not the default, because it cannot tell a class that moved from one that reorganised its members.

The gap between the two is the point: a class has genuinely drifted only when its members are largely unchanged and its profile has moved. A class whose matched Jaccard falls below one half is flagged as reorganised and set aside from the drift count, which is also why a distance on the order of the whole between-class separation is read as a warning rather than a finding.

uv run analysis drift --axis age_at_diagnosis --alignment membership

--alignment takes membership or centroid. Because the alignment reads the stored fit summaries, switching methods re-measures without re-fitting. Measuring how far an aligned class then moved is the sibling step.