Expand description
Discrete contingency-table machinery operating on factorized Dataset codes.
This is the discrete back-end for the power-divergence family of
conditional-independence tests (see crate::ci_tests): it builds (X, Y)
count tables from integer codes, applies the power-divergence statistic for a
given parameter λ (with optional Yates’ continuity correction on 2×2
tables), and aggregates over the strata defined by the conditioning set Z.
The statistic matches scipy.stats.chi2_contingency(table, lambda_=λ, correction=yates), which is how the golden fixture was generated. For a table
with observed counts O, marginals R_i, C_j, total N and expected
E_ij = R_i * C_j / N:
λ = 0(log-likelihood / G-test):2 * Σ O·ln(O/E).λ = −1(modified log-likelihood):2 * Σ E·ln(E/O).- otherwise (incl.
λ = 1,2/3,−1/2):(2 / (λ·(λ+1))) · Σ ( O^(λ+1)/E^λ − O ).
The O^(λ+1)/E^λ − O form is used (rather than O·((O/E)^λ − 1)) because it
is O = 0-safe for every λ > −1: at O = 0 the term is 0 instead of the
0·∞ = NaN the naive form would produce for λ < 0.
Conditional tests consume a cached row partition for the conditioning set, tabulating each stratum in turn.