*2.4. Statistical Analysis*

We reduced our "combined dataset" to 1741 samples from 74 species and 20 localities because there was insu fficient information to accurately classify nine records of the species *Diasporus tigrillo*, *D. vocator*, *Hyloscirtus palmeri, Triprion spinosus,* and *Cruziohyla calcarifer* in the FRHI. For our analyses, we pooled all species together instead of using species as a predictor because the samples sizes per species were highly variable (from 1–177), which could produce significant models that may be an artifact of opportunistic sampling instead of a real pattern. Instead, we used the FRHI, which is highly correlated with taxonomic group. We were unable to include time of sampling as a predictor in our analyses because these data were missing in several of the amphibian assemblages sampled. All our analyses were performed with the R package "stats" [71].

We analyzed *Bd* prevalence with fixed-e ffects generalized linear models (GLMs) using infected status as a binomial response variable (uninfected or infected) and herpetological province, altitudinal belt, and the FRHI as predictors. Ranking of the candidate GLMs followed the Akaike's information criterion (AIC) where the model with the lowest AIC was considered the most robust [72]. To analyze *Bd* infection intensity (estimated as the number of genomic equivalents), we analyzed the 351 *Bd*-positive swabs where *Bd* infection intensity was quantified through qPCR (see Section 2.3). We used linear models (LMs) to compare *Bd* infection intensity (response variable) across herpetological provinces, altitudinal belts, and FRHI (predictor variables). We log-transformed the *Bd* infection intensity to reduce skewness. Statistical significance of models was tested with ANOVA. For both, GLMs and LMs, we conducted post hoc, pairwise comparisons (Tukey's honestly significant di fference; HSD-test) to confirm where the di fferences occurred between significant predictors. We were unable to run mix-e ffects models or fixed-e ffects interaction models because some combinations of predictors presented missing or low values, causing convergence di fficulties.
