On the estimation of detection probabilities for sampling stream-dwelling fishes

List of Tables Table Page 1. Means, standard errors and range of mean density and the negative binomial dispersion parameter, by species, and coefficients, standard errors, and 95% confidence limit from linear regression of species-specific mean abundance and negative binomial dispersion parameter .. 2. Average habitat characteristics, fish abundance, and relative sampling efficiency for monitoring sites used for the comparison of 4 approaches to estimating detection probabilities, for bull and westslope cutthroat List of Figures Figure Page 1. Locations of surveyed and monitored stream channels (gray lines) in the interior Columbia River Basin (bold outline). States are labeled and boundaries (broken line) are shown for reference .. 2. Abundance estimates from the quasi-likelihood Poisson and negative binomial regressions for (a) bull trout, (b) brook trout, and (c) westslope cutthroat trout calculated during the leave-one-out cross-validation procedure. Broken 3. The mean differences between baseline detection probabilities and four detection probability estimators for bull trout, brook trout, and westslope Executive Summary To examine the adequacy of fish probability of detection estimates, I examined distributional properties of survey and monitoring data for bull trout (Salvelinus confluentus), brook trout (Salvelinus fontinalis), westslope cutthroat trout (Oncorhynchus clarki lewisi), chinook salmon parr (Oncorhynchus tshawytscha), and steelhead /redband trout (Oncorhynchus mykiss spp.), from 178 streams in the Interior Columbia River Basin. Negative binomial dispersion parameters varied considerably among species and streams, but were significantly (P<0.05) positively related to fish density. Across streams, the variances in fish abundances differed greatly among species and indicated that the data for all species were overdispersed with respect to the Poisson (i.e., the variances exceeded the means). This significantly affected Poisson probability of detection estimates, which were the highest across species and were, on average, 3.82, 2.66, and 3.47 times greater than baseline values. Required sample sizes for species detection at the 95% confidence level were also lowest for the Poisson, which underestimated sample size requirements an average of 72% across species. Negative binomial and Poisson-gamma probability of detection and sample size estimates were more accurate than the Poisson and generally less than 10% from baseline values. My results indicate the Poisson and binomial assumptions often are violated, which results in probability of detection estimates that are biased high and sample size estimates that are biased low. To increase the accuracy of these estimates, I recommend that future studies use predictive distributions than can incorporate multiple sources of uncertainty or excess variance …