Figure S1 from Variably hungry caterpillars: predictive models and foliar chemistry suggest how to eat a rainforest

A schematic diagram of our analytical steps. All predictive analyses included the standardised Madang data set (Input 1). Firstly, we fit a linear model to the association strengths of our bipartite data set, with a phylogenetic correlation between interacting species (Input 1). We then used the community covariance matrices and phylogenetically standardized interactions strengths derived from this model to perform sequential removal and host prediction for each interactant from the matrix. This allowed us to generate a predicted binary host use matrix filled with the most probable interactions (Output 1). These predictions were then validated by correlating matrix fill and using null models (Validation). Finally, the quantitative data set (Input 1) was extended to include a phylogeny in which the position of hosts from the expanded data set was known, but all additional host associations were predicted based on the original covariance matrices and phylogenetically standardized interactions strengths and this new phylogeny.