We computed the codispersion of each pair of species at spatial lags ranging from 20 to 120 m. The maximum spatial lag equaled just under one-fourth of the length of the shortest side of the plot and was used to ensure adequate sample sizes for the largest spatial lag. To assess the significance of the observed codispersion patterns, we compared the observed codispersion values for each species pair calculated for each spatial lag and direction to values generated using two null models. The first was a CSR model, where one species distribution was fixed and the point locations of the other species were distributed completely spatially randomly across the plot. The second was a toroidal shift model, where the positions of trees were fixed, thus maintaining their autocorrelation structure, but the entire plot was shifted in a random direction and distance around a torus (Wiegand and Moloney 2014).
For each comparison, the null models were used to generate 199 new data sets for one of the species of each pair; 199 null simulations was a large enough number to confidently determine significant differences between observed and expected, and small enough to generate expected values on a desktop computer within a few days. Only one of the species pair needed to be randomized because this was enough to break their spatial association, allowing us to test the significance of their co-variation. The observed codispersion values at each spatial lag were then compared to the vector of codispersion values at the same spatial lags and directions under each null model to estimate tail probabilities; if the observed value was ≥195th value or ≤5th value, we deemed it to be significantly different from expected (i.e., a two-tailed test; P < 0.05). Finally, we calculated the Type I error rate of the CSR and toroidal shift null models by comparing the observed codispersion between two CSR simulated patterns (Appendix S2) to values generated under the CSR and toroidal shift models.