Assessing the transferability of hydraulic habitat models for juvenile Atlantic salmon

Hydraulic habitat models are logistically and technically challenging and expensive to produce. They are therefore frequently transferred between rivers, often with different environmental conditions, without validation. Although studies have recognised problems with model transfer, few have assessed the consequences for model predictions. This study investigated the local (within sub-catchment) transfer of hydraulic habitat models developed for Atlantic salmon (Salmo salar) fry. Two adjacent reaches were chosen for the study, each containing pool, riffle, glide and run habitats. Detailed topographic surveys were used to develop hydraulic models for the study reaches. Substrate and cover were characterised using transects. Seasonal fish habitat use was characterised by electrofishing. Uniform saturation stocking was employed to ensure standard starting densities across all habitats. Generalised additive models were fitted to fry abundance data. Depth, velocity, dominant substrate and cover were used as predictor variables. The distribution of depth, velocity, dominant substrate and cover differed between the two reaches, but was consistent within reaches across seasons. Substrate exhibited the greatest inter-reach difference. Velocity and depth were the most important individual predictors of fry abundance, with the highest densities observed at moderate velocities (∼0.4 m s⁻¹) and low depths (∼0.1 m) across all seasons. When models were transferred locally between reaches, those that were adjusted for inter-reach differences in mean fry abundance (i.e. those predicting changes in relative abundance) performed better than those transferred without adjustment. Complex models that included substrate and cover typically explained some of the variation in abundance in the transferred reach but performed less well than models containing hydraulic parameters alone. This potentially reflected inter-reach differences in the distribution of substrate and cover. This study suggests that (1) uniform stocking is useful for examining habitat-abundance relationships free from the influence of patchy egg deposition; (2) habitat models should be developed at sites offering maximum environmental complexity at a local level; (3) scientists and managers should avoid transferring models between locations with different environmental characteristics, especially in the absence of model validation; (4) complex models should be avoided, and those containing hydraulic variables alone should be considered, if predictions of habitat quality are to be made at new sites.     

A note on the bias due to omitted confounders

In observational studies some confounders may be unknown andtherefore omitted from the analysis while others are adjustedfor. Approximations to the functions defining the relationshipbetween the parameters in the full and reduced models are proposedleading to asymptotic bias estimates. Numerical calculationsfor logistic and Poisson regression models show good agreementbetween asymptotic and simulation bias. A data set containingthe relationship between low birth weight and smoking (Hosmer& Lemeshow, 1989) is used as an illustration.