Skip to content

Migrating from 0.32 to 0.33

0.33 corrects the bootstrap intervals of weighted tables and adds weight_scheme= to choose how a resample treats the weights. Unweighted (unit-weight) intervals are bit-identical to 0.32, and no point estimate changes. This page lists what to check coming from 0.32; the changelog has the full list.

Requirements

  • Python >=3.11 (was >=3.10). Python 3.10 has reached end of life; on it, pip now resolves polars-cv 0.32.x instead.

Results that change

Weighted intervals are wider, and now the right width. 0.32 stratified the image-level resample on gt_label crossed with the weight cells, so every replicate kept each cell's positive count. When the weights are estimated over all images (p / q̂ reweighting a vendor or prevalence mix to a target) that count is random, and the weighted statistics depend on it: the bootstrap SE came out 7-17% low and 95% intervals covered 0.88-0.93. The default draw (weight_scheme="reestimate") no longer crosses the label with the cells. In a vendor-reweighting simulation through lroc_auc_ci_lazy the SE went from 0.93x to 1.02x the estimate's actual spread, and coverage from 0.920 to 0.940.

Small weighted groups can report null bounds. A draw that no longer fixes the gt_label counts can, in a small group, draw no positive image (or, for Mann-Whitney, no negative one). That replicate's statistic still had a value (NaN, 0.5 or 0.0) and was scored; it now nulls the group's bounds, as an undefined replicate does. Each replicate draws none with probability about (1 − prevalence)^n, so this needs a group of a few dozen images or fewer. sample_col draws could already make such replicates in 0.32 and scored them the same way; they now null too.

A weight-0 image no longer nulls its group's bounds. It formed a singleton weight cell; cells whose weights are all zero are now exempt from the singleton rule.

Choosing a weight_scheme

Your weights are Use
estimated from the sample: p / q̂ over target distributions (conditional ones included), post-stratification, raking "reestimate" (default)
as above, and each cell's positive count was fixed by the study design "stratified" (0.32's draw)
known in advance (design weights), or continuous (e.g. from a propensity model) "fixed" 
# 0.32's intervals, unchanged:
froc_auc_ci_lazy(table, method="mann_whitney", weight_scheme="stratified")

# A continuous weight: 0.32 nulled every bound (all singleton cells).
lroc_auc_ci_lazy(table, group_by="group_id", weight_scheme="fixed")

# Report.ci takes the same arguments:
report.ci("map", weight_scheme="fixed")
  • weight_rtol=1e300 workarounds (one cell for every weight, to opt out of weight cells) become weight_scheme="fixed". "fixed" forms no cells, so it raises if given strata or weight_rtol.
  • weight_rtol's default is now None, meaning 1e-6 for the schemes that form cells. Passing 1e-6 explicitly still works the same.
  • An unknown weight_scheme raises ValueError.