Habitat affinity and density-dependent movement as indicators of fish habitat restoration efficacy

In mark-recapture assays from four different study years, the affinity of young-of-the-year Chinook Salmon (Oncorhynchus tschawytscha) and steelhead (O. mykiss) for stream pools restored with or created by engineered log structures was greater than that for pools without restoration, though with high interannual variability. From corresponding distribution and abundance data, it was clear that behavioral data are not always concordant with single observations of abundance. The same was true of the correlation between either behavior or abundance and physical characteristics of pools, although depth and current velocity had some explanatory power for both responses in Chinook. Density-dependent immigration into pools by Chinook indicated that restored pools have greater capacity for this species than unrestored pools; however no such pattern emerged for steelhead. Variation among individuals in body condition has implications for population-wide fitness and low variation was correlated with stronger affinity for pools. This suggests that pools mediate habitat-related trade-offs and that restoring them might have positive effects on fitness. Thus, behavioral data appear to provide stronger indications of restoration effectiveness than observational data alone.

INTRODUCTION the mainstem Columbia River if suitable overwintering habitat is not available in the river (Hillman et al. 1987), with smolt outmigration after one year of freshwater residency. Steelhead can rear in 89 the streams for 1-3 years before outmigration. Predation risk primarily comes from birds (belted 90 kingfisher, Ceryle alcyon; great blue heron, Ardea herodias) and semi-aquatic mammals (e.g., river 91 otter, Lontra canadensis. Larger, predatory fish such as resident and fluvial bull trout (Salvelinus 92 confluentus have been observed in deeper pools created by larger in-stream structures such as 93 channel-spanning weirs, but not in the smaller pools created by the engineered log jams (ELJs) 94 that comprised restoration projects in the river.  Captured fish were placed in insulated, aerated buckets and mildly anaesthetized with MS-125 222 (< 0.1 g · l -1 ) for 2-3 minutes. Sub-yearling fish in this study system range from 50-75 126 mm (Chinook) and 35-70 mm (steelhead), depending on growth rates. Following identification 127 and recording of size data (standard length, SL, in mm and mass in g), fish were marked with 128 a subcutaneous injection of visual implant elastomer (VIE; Northwest Marine Inc.). Following 129 marking, fish were transferred to another insulated, aerated bucket where they were allowed to 130 fully recover from anaesthetization. The recovery period was at least 10 min, or after a full righting 131 response with fish appearing alert and responsive, before they were released to the capture pool.

132
After 24 hours, the pool was re-sampled and the number of both recaptured individuals and newly 133 captured unmarked fish were noted.  Differences between the slopes (e.g., β restored > β unrestored ), evaluated by using a N marked · habi-142 tat type term in a linear model would indicate differences in affinity for restored and unrestored 143 habitat.

144
I specified the models for each species as generalized linear mixed-effects models (GLMMs), 145 assuming a Poisson error distribution. I assumed a priori that the relationship between N recaptured on Day 2 and N marked on Day 1 passed through the origin because pools with no fish captured (or 147 marked) on Day 1 were excluded. Habitat type and number of fish marked were designated as 148 fixed effects and year as the random effect. Because pool area is a strong positive correlate of 149 fish abundance (Polivka et al. 2015), and thus affected the number of fish marked to begin with, 150 it was entered into each model as an offset parameter to prevent fitting a negative value (Zuur 151 et al. 2009). This offset also enables some indication that observed habitat selection and affinity 152 for restored pools is not simply an artifact of restoration creating larger pools. I compared four 153 candidate GLMMs for each species to determine the importance of habitat and annual effects: 1) 154 equal slopes of the regression lines for the two habitat types (i.e., no N marked · habitat interaction 155 term) with the random effect (year) included, 2) equal slopes and no random effect, 3) unique 156 slopes (including interaction term) plus the random effect, 4) unique slopes and no random effect.

157
I selected the best model using the Akaike Information Criterion (AIC; Burnham and Anderson 158 2002). If the best model was one of the models that included a N marked · habitat type interaction 159 term, I concluded that fish differed in their affinity for restored vs. unrestored habitat and calculated 160 habitat-specific βs. Given the multi-year nature of the study, I expected the best model to include 161 year as a random factor, justifying within-year analyses. To ensure that the offset parameter did 162 not cause some systematic lack of model fit, I re-ran the analyses with pool area designated as a 163 simple predictor.

164
To address annual differences and to compare behavior with observations of distribution and 165 abundance, I used two sets of GLMs for each year: one for behavior (including the interaction term 166 as above) and one to describe fish abundance on Day 1. I also compared whether physical habitat Model output for GLM in R (R Core Team 2018) provides AIC scores and I used these to confirm that the model with only significant predictors also had the lowest AIC score. These models identi-177 fied 1) the years in which β restored and β unrestored indicated different levels of habitat affinity if the 178 best model had a significant interaction term, 2) whether any differences in affinity corresponded 179 to differences among habitats in abundance, and 3) whether affinity and abundance were associated where I = number of immigrants, R = density of recaptured fish and a and b describe the shape of 199 the response curve. The peak level of immigration is I = a be at recapture density R = 1 b and λ is  Examination of density dependent emigration is a simpler process as emigration is expected to 226 increase linearly with density. However it may also be an artifact of the total number of fish marked 227 in a pool, total pool area or there may be differences in total emigration by habitat type. Therefore,

228
I constructed another linear model using each of these parameters. To determine whether habitat 229 type affected density dependence, I included a habitat × density interaction term. I performed 230 analysis of variance on these models to identify significant predictors of total emigrants.

231
To examine whether condition variability among individuals was correlated with habitat affin-232 ity as predicted by habitat selection trade-offs, I tested the correlation between the coefficient of variation in the Fulton Condition Index for fish (K; Anderson and Neumann 1996) and habitat 234 affinity β. The Fulton Index relates length (L) and mass (m) as: Although the scaling exponent for L can vary among species, I used a log(mass) vs. log(length) 236 regression to determine that the exact value was 2.997 ± 0.013 and thus not meaningfully different  (Table 1). With all years combined, the affinity 253 of each species was greater for restored habitat; however, for Chinook, the difference was very  (Table 2). Mean depth in each habitat type 264 was 56.5 ± 20.1 cm (restored) and 44.8 ± 9.6 cm (unrestored). Mean current velocity was 18.0 265 ± 10.0 cm · s -1 (restored) and 32.8 ± 15.1 cm · s -1 (unrestored). Temperature was indicated as a 266 significant correlate in some models, but the correlation often was opposite in direction for affinity 267 vs. abundance (Table 2). suggests that the habitats in this system might already be fairly well saturated, and linear models 328 may still be appropriate to estimate basic site fidelity.

329
The linear approach to habitat selection behavior here is analogous to the use of isodars (Morris    (2009,2012,2013,2016). Models considered were: 1) fit lines with equal slopes for both habitats (no N marked × habitat interaction term included) and including study year as a random factor, 2) equal slopes and no random effect, 3) unique slopes (with interaction term) and the random effect, 4) unique slopes and no random effect. Model selection by AIC (best fit model in boldface). year by year affinity differences for restored or unrestored habitat identified by N marked × habitat (significant differences among habitats in slope (β) shown in bold). From separate GLMs, significant abundance differences among restored and unrestored habitats. For both affinity and abundance, models were selected by stepwise removal of non-significant terms. Significant positive (+) and negative (-) correlations of physical habitat parameters shown for each group of models. *Non-significant slopes; NA: Zero recaptures in N=3 pools, only 1-2 fish marked per pool Year β restored β unrestored Affinity correlates Abundance Abundance correlates a) Chinook  FIG. 1: Map of the segment of the Entiat River where study reaches containing restored and unrestored pools were located (RK = river kilometer, measured upstream from confluence with Columbia River).  Table 1; slopes of lines and significant within-year differences given in Table 2. describing density-dependent immigration into restored or unrestored pools (see Fig. 3). See text for description of the derivation of the HDI. The 95% bounds (dashed lines) show a restored −a unrestored > 0, indicating restored pools have a higher immigration capacity for a given density of individuals remaining in the pool (i.e., a restored > a unrestored .  Table 2) and the coefficient of variation in condition index (K; CVCI) of recaptured individuals in each habitat type and year. When considering all years and habitats, stronger affinity for pools (regardless of restoration) was correlated with a lower CVCI (p = 0.010). Data from restored habitat in 2016 were omitted because β was not significantly different from zero.