Environmental management actions in the Upper Mississippi River System (UMRS) typically require pre-project assessments of predicted benefits under a range of project scenarios. The U.S. Army Corps of Engineers (USACE) now requires certified and peer-reviewed models to conduct these assessments. Previously, habitat benefits were estimated for fish communities in the UMRS using the Aquatic Habitat Appraisal Guide (AHAG v.1.0; AHAG from hereon). This spreadsheet-based model used a habitat suitability index (HSI) approach that drew heavily upon Habitat Evaluation Procedures (HEP; U.S. Fish and Wildlife Service, 1980) by the U.S. Fish and Wildlife Service (USFWS). The HSI approach requires developing species response curves for different environmental variables that seek to broadly represent habitat. The AHAG model uses species-specific response curves assembled from literature values, data from other ecosystems, or best professional judgment.

A recent scientific review of the AHAG indicated that the model’s effectiveness is reduced by its dated approach to large river ecosystems, uncertainty regarding its data inputs and rationale for habitat-species response relationships, and lack of field validation (Abt Associates Inc., 2011). The reviewers made two major recommendations: (1) incorporate empirical data from the UMRS into defining the empirical response curves, and (2) conduct post-project biological evaluations to test pre-project benefits estimated by AHAG.

Our objective was to address the first recommendation and generate updated response curves for AHAG using data from the Upper Mississippi River Restoration-Environmental Management Program (UMRR-EMP) Long Term Resource Monitoring Program (LTRMP) element. Fish community data have been collected by LTRMP (Gutreuter and others, 1995; Ratcliff and others, in press) for 20 years from 6 study reaches representing 1,930 kilometers of river and >140 species of fish. We modeled a subset of these data (28 different species; occurrences at sampling sites as observed in day electrofishing samples) using multiple logistic regression with presence/absence responses. Each species’ probability of occurrence, at each sample site, was modeled as a function of 17 environmental variables observed at each sample site by LTRMP standardized protocols. The modeling methods used (1) a forward-selection process to identify the most important predictors and their relative contributions to predictions; (2) partial methods on the predictor set to control variance inflation; and (3) diagnostics for LTRMP design elements that may influence model fits.

Models were fit for 28 species, representing 3 habitat guilds (Lentic, Lotic, and Generalist). We intended to develop “systemic models” using data from all six LTRMP study reaches simultaneously; however, this proved impossible. Thus, we “regionalized” the models, creating two models for each species: “Upper Reach” models, using data from Pools 4, 8, and 13; and “Lower Reach” models, using data from Pool 26, the Open River Reach of the Mississippi River, and the La Grange reach of the Illinois River. A total of 56 models were attempted. For any given site-scale prediction, each model used data from the three LTRMP study reaches comprising the regional model to make predictions. For example, a site-scale prediction in Pool 8 was made using data from Pools 4, 8, and 13. This is the fundamental nature and trade-off of regionalizing these models for broad management application.

Model fits were deemed “certifiably good” using the Hosmer and Lemeshow Goodness-of-Fit statistic (Hosmer and Lemeshow, 2000). This test post-partitions model predictions into 10 groups and conducts inferential tests on correspondences between observed and expected probability of occurrence across all partitions, under Chi-square distributional assumptions. This permits an inferential test of how well the models fit and a tool for reporting when they did not (and perhaps why). Our goal was to develop regionalized models, and to assess and describe circumstances when a good fit was not possible.

Seven fish species composed the Lentic guild. Good fits were achieved for six Upper Reach models. In the Lower Reach, no model produced good fits for the Lentic guild. This was due to (1) lentic species being much less prominent in the Lower Reach study areas, and (2) those that do express greater prominence principally do so only in the La Grange reach of the Illinois River. Thus, developing Lower Reach models for Lentic species will require parsing La Grange from the other two Lower Reach study areas and fitting separate models. We did not do that as part of this study, but it could be done at a later time.

Nine species comprised the Lotic guild. Good fits were achieved for seven Upper Reach models and six Lower Reach models. Four species had good fits for both regions (flathead catfish, blue sucker, sauger, and shorthead redhorse). Three species showed zoogeographic zonation, with a good model fit in one of the regions, but not in the region in which they were absent or rarely occurred (blue catfish, rock bass, and skipjack herring).

Twelve species comprised the Generalist guild. Good fits were achieved for five Upper Reach models and eight Lower Reach models. Six species had good fits for both regions (brook silverside, emerald shiner, freshwater drum, logperch, longnose gar, and white bass). Two species showed zoogeographic zonation, with a good model fit in one of the regions, but not in the region in which they were absent or rarely occurred (red shiner and blackstripe topminnow).

Poorly fit models were almost always due to the diagnostic variable “field station,” a surrogate for river mile. In these circumstances, the residuals for “field station” were non-randomly distributed and often strongly ordered. This indicates either fitting “pool scale” models for these species and regions, or explicitly model covariances between “field station” and the other predictors within the existing modeling framework. Further efforts on these models should seek to resolve these issues using one of these two approaches.

In total, nine species, representing two of the three guilds (Lotic and Generalist), produced well-fit models for both regions. These nine species should comprise the basis for AHAG 2.0. Additional work, likely requiring downscaling of the regional models to pool-scale models, will be needed to incorporate additional species. Alternately, a regionalized AHAG could be comprised of those species, per region, that achieved well-fit models. The number of species and the composition of the regional species pools will differ among regions as a consequence. Each of these alternatives has both pros and cons, and managers are encouraged to consider them fully before further advancing this approach to modeling multi-species habitat suitability.