We examined the effects of different regionalization schemes on the performance of River InVertebrate Prediction and Classification System (RIVPACS)-type predictive models in assessing the biological conditions of streams of the US for the National Wadeable Streams Assessment (WSA). Three regionalization schemes were considered: a single national predictive model (MOD1), separate predictive models for each of the 9 WSA aggregated Omernik level III ecoregions (MOD9), and 3 predictive models roughly corresponding to the western US, the Appalachian Mountains, and the Central and Coastal Plains (MOD3). The goal of the WSA was to assess stream condition at the national scale and at the scale of WSA aggregated ecoregions, so we compared the performance of the ratio of the observed number of taxa to the expected number of taxa (O/E) index estimated using different regionalization schemes at both of these spatial scales. We assessed model performance with a randomized resampling procedure, in which we set aside 10% of the reference sites, calibrated the model with the remaining sites, and applied the model to the set-aside sites. Performance statistics for the set-aside reference sites were accumulated over 10 iterations. When summarized at the national scale, mean model predictions of O/E for set-aside reference sites from the 3 different regionalization schemes were all reasonably close to 1. When summarized by the 9 aggregated ecoregions, MOD1 and MOD3 predictions of O/E differed systematically from 1 in certain aggregated ecoregions. Over all 9 ecoregions, the magnitude of these differences was significantly greater than observed with MOD9 predictions. Results from our analysis suggest that O/E values at test sites should be interpreted with respect to mean and SD of O/E of reference sites from the same region to minimize the effects of systematic biases in the predictions. RIVPACS-type predictive models also should be calibrated at a spatial scale similar to the scale at which summary statistics are reported.