Predictive models of fish microhabitat selection in multiple sites accounting for abundance overdispersion. (28th April 2020)

• Record Type: Journal Article
• Authors: Plichard, Laura; Forcellini, Maxence; Le Coarer, Yann; Capra, Hervé; Carrel, Georges; Ecochard, René; Lamouroux, Nicolas
• Is Part Of: River research and applications. Volume 36:Number 7(2020)
• Page Start: 1056
• Page End: 1075
• Publication Date: 2020-04-28
• Subjects: abundance overdispersion -- fish microhabitat selection modelling -- fish preference -- hydraulic habitat -- mixed‐effect models -- negative binomial distribution; Rivers -- Regulation -- Periodicals; Rivers -- Periodicals; 551.483
• DOI: 10.1002/rra.3631
• Languages: English
• ISSNs: 1535-1459

Abstract

Abstract: Microhabitat selection models are frequently used in rivers to evaluate anthropogenic effects on aquatic organisms. Fish models are generally developed from few rivers, with debatable statistical treatments for coping with overdispersed abundance distributions. Analyses of data from multiple rivers are needed to test their transferability and increase their relevance for stakeholders. Using 3, 528 microhabitats sampled in nine French rivers during 129 surveys, we developed models for 35 specific size classes of 22 fish species. We used mixed‐effects generalized linear models (accounting for multiple surveys), involving B‐spline transformations (accounting for nonlinear responses) and assuming a negative binomial distribution (accounting for abundance overdispersion). We compared models of increasing complexity: no selection (M1), an "average" selection similar in all surveys (M2), two models with different selection across surveys (M3–M4). Of 132 univariate cases (specific size classes by habitat), 63% indicated selection for depth, 71% for velocity, 45% for substratum size and 13% for substratum heterogeneity. A total of 50 models were retained, involving 26/35 specific size classes. Model fits indicated low explained deviance ( R 2 MF < 0.19) and higher rank correlations (ρ < 0.69) between observed and modelled values. However, Bayesian posterior predictive checks validated these results since excellent fits would generate R 2 MF lower than 0.59 and ρ lower than 0.78. We found high transferability among rivers and dates, because (a) M2 was the most appropriate in 26/50 cases; (b) the R 2 MF and ρ values by M2 was, respectively, 72% and 75% of that explained by the complex M4 and (c) independent river cross‐validations showed good transferability. Bivariate models for selected specific size classes improved univariate model fits (ρ from 0.30 to 0.38). Overall, using a nonlinear mixed‐effect approach, our results confirmed the relevance of "average" models based on several rivers for developing helpful e‐flow tools. Finally, our modelling approach opens opportunities for integrating additional effects as the spatial distribution of competitors.