LIMITATIONS AND SUGGESTIONS FOR FUTURE WORK In general, while the estimates of probability of presence, index scores, CNQI, and CASI generated through this assessment represent a useful and objective means for assessing aquatic habitat and prioritizing habitats for restoration or protection, there are some limitations that are important to consider. Results generated through the modeling process are ultimately limited by the quality of data used to generate them. In the future, the model can be improved by improving the resolution and precision of the data. For example, some county‐level data were used as predictor variables although the data likely generalize conditions at the catchment scale. In some cases, this resulted in generalizations in CASI or in the individual CASI metrics,which is evidenced by the visibly unnatural hard break lines at some county boundaries. Although these variables—such as network cattle density and network surface water use—were limited in spatial resolution,they still had high relative influence in some BRT models and were important to retain for predictive performance. In the future, refinement of these county‐level variables or inclusion of higher resolution surrogates could improve both the precision of the BRT model predictions and post‐modeling indices.A second limitation is that the data and maps represent only a snapshot in time. Therefore, the models maynot represent conditions before or after the data were collected or created. For example, any habitat lost orgained due to increased impervious surface cover since the 2006 National Land Cover Database (NLCD) wasnot considered in this assessment. Similarly, a portion of the uncertainty can be attributable to the temporal mismatches between the fish collection data and landscape data. As such, improving the temporal matchbetween those datasets for future work would be beneficial.Although the BRT statistical modeling algorithm automatically accounts for interactions between predictor variables, the post modeling process used to generate CNQI and CASI does not fully account for interactions.The post modeling outputs were derived from the relative influence and function plot outputs from BRT.Those outputs themselves are objective approximations useful for model interpretation and interrogation;however, they are not used in estimating the predicted values in BRT. Therefore, the individual function plotsfrom which CASI and CNQI and their metrics were derived did not account for variable interactions because they are not accounted for in the BRT relative influence values or function plots. For example, while appropriately accounting for an interaction between ecoregion and stream size would require separate function plots for the effect of drainage area in each separate ecoregion, the function plot generated by BRT represents an average effect of drainage area across all ecoregions. In that example, the effect of drainagearea is overestimated in some ecoregions and underestimated in others. This is not a limitation specific to the methodology used in this assessment, but common in other popular predictive modeling approaches. The advantage of the approach using BRT, however, is in the improved ability to predict current conditions (i.e.,probability of presence) relative to other methods.While continuous response variables can be modeled, binomial response variables can generally be modeled with greater precision in cases where the response data vary in collection method or date. Throughout this assessment, we have generally found that binomial (i.e., presence‐absence) response variable models have performed better than continuous (i.e., abundance‐based) variable models. In the future, basing diversity metrics on the presence‐absence of targeted species, rather than relative abundance, may improve their precision.Future models could be well served by including additional data (should it become available) that may be important in structuring fish populations in this region. Data on dams is currently of incomplete and at a scale that limits its usefulness at this scale. Another particularly useful predictor variable for future assessments would be data that quantifies confined animal feeding operations (CAFO). The cattle density variable used in this model did not include any information from these types of feedlots, which could have significant impacton water quality and fish communities.There were also a few important issues that were beyond the scope of this project. Acid precipitation,biological interactions, and local habitat variation are all important in structuring fish communities. These variables were not directly used as predictor variables, although, when possible, surrogates were used to approximate variation in the model resulting from these processes.Local habitat measures such as water quality (pH, alkalinity, instream temperature), physical habitat complexity, and substrate size are examples of local measures important to structuring fish communities.These measures could not be directly quantified in this analysis given the scope and scale of the project. However, since each catchment’s land cover and geology was included in the analysis, some aspects of waterquality were indirectly modeled. Likewise, habitat complexity and substrate size could be partially captured by the combination of stream slope and bedrock and surficial geology. Nonetheless, exclusion of detailed local measures likely accounts for some uncertainty in the model results. Thus, the results from this analysis should be combined with local expert knowledge and additional field data to arrive at the most accurate representation of habitat conditions.In addition, inclusion of biological interactions in future models could improve the precision of the model andthe ability to quantify its influence on the response variables. Specifically, important biological interactions in this system could include the negative interactions resulting from the introduction of non‐native or other stocked fishes, such as brown trout or Asian carp. Finally, another important consideration for managing aquatic habitat at this scale, which was not considered directly in this analysis, is climate change. Potential impacts from climate change include altered thermal regimes, stream flow regimes, and physical habitat. Particularly for coldwater fishes such as trout and sculpin,future warming could result in increased population isolation due to confinement to headwater habitats ormore localized thermal refugia. Specifically, identifying catchments vulnerable to climate change andimportant to species of interest in this system—for example, those on the fringe of meeting the upper thermal criteria for a species—could represent an important and supplementary next step in the identification of restoration and protection priorities for targeted aquatic populations.
