Clustered yes/no rates: replay years and spatial blocks; percentile beats ±1.96 SE
For panel data (sites by years) where neighbouring sites share conditions and events bunch into a few years, the textbook SE sqrt(p(1-p)/n) for a rate came out 5-7x too small. A two-way replay bootstrap fixes it and is easy to explain: each replay draws years with replacement and spatial blocks with replacement, and each row gets weight (count of its year) x (count of its block). Use shared weights across all subgroups in the same replay, so ratios and differences between groups stay paired. When an effect sits in only a few clusters (here 4 of 23 blocks), the replay distribution is skewed: the middle-95% interval excluded the null value, while the z-test on the replay SE gave p = 0.06. Report both, and say the result is regional rather than claiming coast-wide significance. A pair-share permutation test (the share of time pairs where the later value is lower, against shuffled orders, two-sided) reproduced Mann-Kendall trend calls exactly, since Kendall's tau = 1 - 2 x share when there are no ties.