Figure 3.

Mean percentage of individuals correctly identified to host species according to linear discriminant analyses where models were trained on (A) Bartonella infections of the hosts, (B) flea infestations of the hosts (χ² = 0·02, P = 0·88) and (C) Bartonella infections of the fleas infesting the hosts (χ² = 28·7, P < 0·001), using data from all three woodland sites combined. In each case, models were trained on random selections of 75% of host–parasite associations and used to predict the host identity of the remaining 25% of the data. This was done 1000 times in each case. Grey bars represent models trained on true host–parasite associations, while white bars represent models trained on random host–parasite associations. Differences between the predictive capabilities of each model were assessed using chi‐squared analyses. In Fig. (A), models were trained on host‐Bartonella infections identified either to species level (‘Bartonella species’) or to pITS variant level (‘Bartonella variants’), and ten comparisons were made, represented by the letters a–j. a: χ² = 61·8, P < 0·001, b: χ² = 44·8, P < 0·001, c: χ² = 33·5, P < 0·001, d: χ² = 30·9, P < 0·001, e: χ² = 33·8, P < 0·001, f: χ² = 83·7, P < 0·001, g: χ² = 26·8, P < 0·001, h: χ² = 0·99, P = 0·32, i: χ² = 19·5, P < 0·001, j: χ² = 31·2, P < 0·001. Linear discriminant analyses models could not be computed for Bartonella birtlesii variants alone as the distribution of the one variant shared between host species was highly skewed (birtlesii‐1, found only twice in wood mice but 50 times in bank voles; Table S2).