Bootstrapping

Bootstrapping is a resampling method introduced by Bradley Efron in 1979 for estimating the sampling distribution of a statistic by repeatedly drawing random samples with replacement from the observed data. In phylogenetics, bootstrapping is the standard method for assessing the robustness of a phylogenetic tree: the original sequence alignment is resampled hundreds of times, a tree is built from each resampled dataset, and the proportion of trees that recover each branch is reported as a bootstrap support value.

The bootstrap has become a ritual in phylogenetics — a number that must be reported for a tree to be publishable. But the interpretation of bootstrap values is fraught. A bootstrap proportion is not a probability that a branch is correct; it is the probability that the branch would be recovered if the data were resampled in the same way. This distinction matters because bootstrap values conflate sampling error with model error. If the evolutionary model is misspecified — if the true history involves reticulation or horizontal gene transfer that the tree model cannot represent — bootstrap values can be arbitrarily high while the tree is systematically wrong.

The method also assumes that the sites in a sequence alignment are independent and identically distributed, an assumption violated by linkage, codon structure, and spatial correlation in protein structures. Despite these limitations, bootstrapping remains useful as a rough heuristic for identifying unstable branches, provided it is not treated as a measure of truth.

Bootstrapping in phylogenetics is a confidence ritual masquerading as a confidence interval. It tells you how stable your tree is to perturbation of the data, not how close your tree is to biological reality. The field treats bootstrap values of 70% as acceptable and 95% as strong, but these thresholds are arbitrary conventions with no formal justification — another case of statistical practice outrunning statistical understanding.