Performance of the GenEst Mortality Estimator Compared to The Huso and Shoenfeld Estimators
The impacts of wind power development on bat and bird populations are commonly assessed by estimating the number of fatalities at wind power facilities through post-construction monitoring (PCM) studies. Standard methodology involves periodic carcass searches on plots beneath turbines (Strickland et al. 2011, US Fish and Wildlife Service 2012). The resulting counts are adjusted to compensate for bias due to imperfect carcass detection by searchers, removal of carcasses by scavengers or other processes (Korner-Nievergelt et al. 2011), and carcasses that may have fallen outside of searched areas. To account for the bias in counts due to imperfect detection and carcass removal, investigators typically conduct bias trial experiments to inform models of carcass detection probability. Many different estimators have been proposed that combine information about the bias trial experiments to estimate a detection probability for carcasses (g) and ultimately obtain an estimate of total mortality (M). The two estimators that have seen the most widespread use in North America recently are the Huso (Huso 2011, Huso et al. 2012) and Shoenfeld (Shoenfeld 2004; also called the Erickson estimator) estimators. GenEst (Dalthorp et al. 2018a, 2018b, 2018c) is the newest statistical estimator to become available and was designed to improve upon the Huso and Shoenfeld estimators by generalizing the key assumptions in both, and to improve comparability among new PCM studies. In addition to relaxing some of the assumptions inherent to the Huso and Shoenfeld estimators, GenEst uses a parametric bootstrap applied to a novel approach to variance estimation (Madsen et al. 2019). The current study was undertaken to document the performance of GenEst relative to the Huso and Shoenfeld estimators. We took a simulation approach to the study because simulation data provides the basis to compare mortality estimators under conditions where the “truth” is known. The estimators were compared on three metrics: 1) bias—the tendency of an estimator to over- or under-estimate actual mortality, 2) precision—the ability of an estimator to constrain an estimate to a narrow range (measured here as the width of a 90% confidence interval [CI] around the point estimate divided by the true, known mortality), and 3) CI coverage—the probability a CI with a specified level of confidence actually includes the true level of mortality. Although our simulations were conceived and designed—and are discussed—with respect to wind power facilities, it is important to note that the estimators and results discussed here are relevant to any post-construction fatality monitoring study that may occur (such as at solar facilities) where detection is imperfect. Although our study treats the problem of mortality estimation when detection is imperfect, it is also important to note that all of the estimators considered here are Horvitz-Thompson (Horvitz and Thompson 1952) style estimators, that is, none are designed to estimate the mortality of rare species as might be necessary under an Incidental Take Permit. The Evidence of Absence estimator (Dalthorp et al. 2017) is still the most appropriate statistical tool for rare event estimation. The simulations cover a broad range of conditions that may occur in field studies and complete results are presented without commentary in the appendix. The main body of this report does not provide a comprehensive treatment of our results; rather, we try to identify some of the more important differences among the estimators and some conditions under which reliable mortality estimates are especially challenging.