Samples of size n come from a bivariate normal with unit variances, correlation ρ and mean δ = c(cos θ, sin θ); we test H0: δ = 0. Left: each dot is one sample's pair of t-statistics (t1, t2); the Bonferroni test accepts inside the square, the unadjusted t-tests inside the dashed square, and red dots are the samples that Hotelling's T² rejects (its boundary uses the sample correlation, so it is an ellipse only approximately). Top right: power against the length c of the shift, exact for T², simulated for the t-tests. Bottom right: the shift against the population's ellipse. A shift that goes against the correlation is far easier to detect.
T² = n X̄TS−1X̄ rejects when (n − 2)T²/(2(n − 1)) exceeds the upper-α point of F2, n−2; its exact power uses the noncentral F law with noncentrality nδTΣ−1δ. The t-tests use tj = √n X̄j/sj and reject when either |tj| exceeds tn−1 at α/2 (unadjusted) or α/4 (Bonferroni). Power curves reuse the same simulated samples at every shift (only the mean moves; S does not change).
code/ch14_multivariate_models.py (hotelling_demo, hotelling_t2, hotelling_power, f2_crit, t_tests).