Adaptive Management of Flows in a Regulated River: Flow-ecology Relationships Revealed by a 26-year, Five-treatment Flow Experiment

Adaptive management (AM) is often proposed as a means to resolve uncertainty in the management of socio-ecological systems but successful implementation of AM is rare. We report results from a 26 year, five-treatment, AM experiment designed to inform decision makers about the response of juvenile salmonids (Oncorhynchus spp.) to flow releases from a dam on the regulated Bridge River, British Columbia, Canada. Treatments consisted of a baseline (no dam release) and four different dam release regimes that followed a semi-natural hydrograph but varied in the magnitude of spring-summer freshet flows. We found total salmonid biomass was highest at the lowest flow release, and decreased with increasing flow, consistent with a priori predictions made by an expert solicitation process. Species-specific responses were observed that in some cases could be attributed to interactions between the flow regime and life history. The relationship between juvenile biomass and flow resulting from the experiment can inform decisions on water management for this river. The documentation of successful AM experiments is sorely needed to allow for reflection on the circumstances when AM is likely to deliver desirable outcomes, and to improve other decision processes that require fewer resources and less time to implement.


Appendix 1: Evaluation of effects of priors on capture probability hyper distribution parameters
To estimate juvenile salmon and steelhead abundance in this study, we initially applied the same Hierarchical Bayesian Model (HBM) used in Bradford et al. (2011) to the extended dataset .
The model would not converge for more recent years when catches were much lower. Site-specific estimates of capture probability, which drive estimates of the hyper-distribution of capture probability, depend on the magnitude of the reduction in catches across passes.There is no information about capture probabilty at a site if no fish of a given species-age class are captured, and very little information when the catch is very low. If this pattern occurs at many sites, the hyper-distribution of capture probability will be poorly defined and more information on capture probability in the prior distribution is required to obtain reliable estimates of capture probability and abundance.
In the original application of the HBM we used an uninformative prior for the mean capture probability across sites centered at 0.5 (beta distribution with parameters beta(1,1)), and a minimally informative prior for the standard deviation in capture probabilities across sites (half-cauchy distribution with scale parameters 0 and 0.3, see Gelman 2006). To obtain more reliable estimates with sparser data from more recent years, we used a more informative prior on the mean capture probability across sites.
The prior was still centered at 0.5 but we assumed more certainty about this mean level (beta(50,50)). We used a uniform prior on the precison (inverse of variance) of capture probability across sites (unif(10,500)) which constrained the maximum extent of variation in capture probability aross sites. In In cases where capture probability was well defined in all years because the species-age class was abundant and widely distributed across sites (e.g. age-0 rainbow trout), model estimates of the mean of the capture probability hyper distribution were similar (Fig. A1). For other species and age classes, where there was less information in the data on capture probability, the means of the capture probability distributions from the model with the original priors were lower and less certain compared to those estimated from the revised model. This occurred because the more informative priors had a greater effect on the posterior when there was less information about parameter values from the data.
Estimates of the standard deviation of the capture probability hyper distribution, determining the extent of variability acros sites, were considerably higher and less certain based on the less informative priors used in Bradford et al. (2011), compared to the more informative ones used here (Fig. A2). Similar to the analysis of the means of the capture probability hyper distribution, differences between models were substantive for species and age classes with less information about capture probability.
The effects of revised capture probability priors on estimates of total abundance across reaches were modest (Fig. A3). Year-specific estimates of abundance and uncertainty in abundance for age-1 rainbow trout and age-0 chinook salmon based on the original model were higher than estimates based on the revised model with more informative priors. This occurred because capture probability estimates were lower in the original model with less informative priors. Abundance estimates for age-0 rainbow trout and coho salmon based on models with different priors on capture probability were similar. Estimates of flow treatment effects from the mixed effects model will largely be determined by relative differences in abundance estimates across years. Effects of prior assumptions on abundance estimates led to modest differences in the scale of abundance estimates, but did not effect relative differences in estimates over time based on the correlation in abundance estimates between models. The square of the Pearson correlation coefficient (r 2 ) of annual abundance estimates from the two models was 0.98 and 0.88 for age-0 and age-1 rainbow trout, 0.94 for age-0 coho salmon, and 0.92 for age-0 chinook salmon.
To better understand the effects of low catch and occupancy on estimates of abundance from the HBM, we simulated a set of catch depletions across 50 sites based on a zero-inflated log-normal distribution of fish densities. We then applied the HBM to the simulated data and compared estimates of abundance and capture probability to the values used drive the simulation. We found that capture probability was underestimated and abundance was overestimated, and the extent of bias increased with the degree of zero-inflation in simulated fish densities. For example, when we assumed that 30% of the sample sites were unoccupied and mean density was low, abundance was overestimated by 50%. This occurred because the HBM assumes a log-normal distribution in fish density across sites and does not explicitly account for zero-inflation. When the true distribution of densities is zero-inflated, a better fit is obtained by lowering the capture probability because this increases the likelihood for sites with low or zero catch. This in turn results in an overestimate of abundance. Increasing information on capture probability in prior distributions reduces the tendency of the model to underestimate capture probability and therefore reduces the extent of positive bias in abundance. We attempted to revise the structure of the HBM to directly estimate the extent of zero-inflation, but this additional parameter was not estimable because the degree of zero-inflation and the magnitude of capture probability were confounded. That is, the model could not distinguish between cases where capture probabiltiy was high and a large fraction of sites were unoccupied, and the opposte pattern. Although directly accounting for zero-inflation in animal distributions can be accomodated in a mark-recapture framework (Conroy et al. 2008), confounding between capture probability and abundance precludes its use in studies that use depletion-based estimates.