Fig 4. Type I error of different resampling methods in multivariate count simulations from a lognormal-Poisson model, varying n (left) and p (right).

Mean abundances were manipulated by multiplying the matrix of the means μ by a factor (a) δ = 1, (b) δ = 2.5, (c) δ = 5. The shaded grey region represents a 95% confidence band around the nominal Type I error rate of 0.05. The PIT-trap performed reasonably well in simulations varying n, but became inflated at large p because of failure of model assumptions, for low-medium abundances (a-b). The parametric bootstrap, assuming an unstructured correlation matrix, performed similarly.