A review of the effectiveness of blanket curtailment strategies in reducing bat fatalities at terrestrial wind farms in North America

Blanket curtailment of turbine operations during low wind conditions has become an accepted operational minimization tactic to reduce bat mortality at terrestrial wind facilities. Site-specific studies have demonstrated that operational curtailment effectively reduces impacts, but the exact nature of the relationship between increased cut-in speed and fatality reduction in bats remains unclear. To evaluate the efficacy of differing blanket curtailment regimes in reducing bat fatality, we examined data from turbine curtailment experiments in the United States and Canada in a meta-analysis framework. We tested multiple statistical models to explore possible linear and non-linear relationships between turbine cut-in speed and bat fatality reduction while controlling for control cut-in speed. Because the overall sample size for this meta-analysis was small (n = 36 control-treatment studies from 16 field sites from the American Wind Wildlife Information Center and a recent review), we conducted a power analysis to assess the number of control-impact curtailment studies that would be needed to understand the relationship between fatality rate and change in cut-in speed under different fatality reduction scenarios. We also identified the characteristics of individual field studies that may influence their power to detect fatality reduction due to curtailment. Using a response ratio approach, we found any curtailment strategy reduced fatality rates by 56% for studies included in this analysis (p < 0.001). However, we did not find strong evidence for linear (p =0 0.07) or non-linear (p > 0.11) associations between increasing cut-in speeds and fatality reduction. The power analyses showed that the power to detect effects in the meta-analysis was low if fatality reductions were less than 50%. Synthesizing across all analyses, we need more well-designed curtailment studies to determine the effect of increasing curtailment speed and the effect size is likely of a magnitude that we had limited power to detect.


31
Wind energy development is increasing rapidly worldwide and hundreds of thousands of bat 32 fatalities are estimated to occur per year due to collisions with terrestrial wind energy facilities in North

39
Curtailment of turbine operations during low wind conditions, particularly in late summer and fall 40 when fatality rates are highest, has become an accepted operational minimization tactic to reduce bat 41 fatality at terrestrial wind facilities [11]. By increasing the cut-in speed, or the wind speed at which a 42 turbine generator begins to produce electricity, curtailment reduces turbine blade spinning rates. Below 43 the cut-in speed, turbine blades still spin with the wind but do so much more slowly, especially if blades 44 are "feathered" or pitched to catch as little wind as possible. Because bats tend to be more active at lower 45 wind speeds, increasing turbine cut-in speed can significantly reduce bat fatality [1,12]. However, a great 46 deal of variability has been reported in the level of fatality reduction achieved by curtailment, likely due

51
For this study, "blanket" curtailment, in which wind speed and time of day/year are used to 52 determine when to curtail, has both operational and financial implications for wind facility operators [13].

53
At present, the exact nature of the trade-off between turbine energy production and bat fatality 54 minimization is poorly understood. Larger increases in cut-in speeds will further reduce power generation. Still, the implications for fatality reduction are less clear, in part because this type of 56 assessment requires intensive monitoring and is subject to errors introduced by imperfect detection and 57 small sample sizes. Despite limited evidence that raising blanket cut-in speeds above 4.5 m/s will further 58 reduce bat fatalities [14], regulators now have required operational minimization for some new wind 59 projects in the United States and Canada at wind speeds up to 6.9 m/s [15]. A synthesis of the available 60 data from designed curtailment studies will allow us to quantify better the relative benefits of increasing 61 turbine cut-in speed for reducing bat collision fatality.

62
A meta-analysis framework is used to synthesize data across studies to determine the effect of 63 curtailment on bat fatality reduction. Meta-analysis provides a method to account for multiple types of 64 uncertainty and use predictor variables to explain patterns between studies [16]. Random effects meta-65 analyses are needed to account for the uncertainty in effects from each study and the uncertainty in the 66 true effect size to which all studies contribute. Using such an approach, we aim to evaluate the current 67 knowledge of the effectiveness of blanket curtailment regimes in reducing bat fatalities at terrestrial wind 68 projects in North America. We identified three objectives: 1) evaluate existing control-treatment 69 curtailment study data for bats in a meta-analysis framework to examine the relative benefit of increased 70 curtailment cut-in speeds and examine the importance of geography and turbine dimensions on fatality 71 reduction; 2) assess the power of the meta-analysis approach to quantify fatality reduction using a data 72 simulation approach; and 3) understand how different site or survey characteristics (e.g., fatality rates, 73 study length, and carcass persistence) influence the power of individual curtailment field studies to detect 74 a difference in bat fatality rates between control and treatment groups. These analyses are combined to 75 identify the most likely effect of blanket curtailment on bat fatality reduction, how much additional 76 information is needed to be certain of these effects, and how to design curtailment experiments to 77 maximize the value of their results.

79
The study's overall goal was to understand the relationship between blanket curtailment cut-in 80 speed and bat fatality reduction at wind facilities in the United States and Canada. To achieve this goal,

81
we used a response ratio approach that focused on the differences in fatality rates between control and 82 curtailment treatments in available studies. We used a meta-analysis approach (hereafter referred to as the 83 "meta-analysis") to control for variability among studies. As we did not have a predetermined assumption 84 about the nature of the relationship between fatality rate and the change in cut-in speed between control 85 and treatment, we tested multiple statistical models that allowed for both linear and non-linear 86 relationships between cut-in speed and the response to determine which best described the observed 87 pattern. Both the absolute cut-in speed and change in cut-in speed were allowed to influence the predicted 88 fatality rate. Once the best models were selected, we used them to understand how covariates like study 89 location and turbine dimensions could influence the relationship between fatality rate and change in cut-in 90 speed.

91
Because the sample size for this analysis was small (n=36 control-treatment pairs), we also 92 assessed the likelihood that the above meta-analysis would provide statistically significant results and 93 determined the number of control-treatment pairs needed in this meta-analytical framework to be 94 confident in our understanding of the relationship between fatality rate and change in cut-in speed. Thus,

95
we conducted two types of power analyses. The first power analysis (the "meta-analysis power analysis") 96 was designed to quantify the power of the meta-analysis under different hypothetical scenarios about the 97 relationship between fatality rate and change in cut-in speed. The first of these scenarios was an a 98 posteriori scenario based on the results of the best meta-analysis model using existing data, and four 99 additional a priori scenarios with different relationships between fatality reduction and cut-in speed were 100 also examined. The second power analysis (the "fatality estimation power analysis") was designed to 101 inform future curtailment studies and fatality monitoring efforts at operating wind energy facilities. This

102
analysis assessed the relative quality of different fatality studies at the project scale and identified site and 103 survey characteristics (e.g., fatality rate, study length, and carcass persistence) that influenced the power of individual curtailment field studies to detect a difference in bat fatality rates between control and 105 treatment groups. All analyses were conducted in R [17], and all analysis scripts were documented in

125
Individual studies within those projects were determined to be suitable for analysis if they used a blanket curtailment 126 treatment and there were multiple cut-in speeds that could be compared.

128
Fatality estimates in the AWWIC database, which were reported from the original studies, had 129 already been adjusted for detection probability (observer ability to detect carcasses that are present) and 130 carcass persistence (rate of removal of carcasses by scavengers) using searcher efficiency trials and carcass persistence trials, respectively [18]. There are multiple approaches for correcting fatality estimates 132 that differ in their assumptions regarding how to account for detection error resulting from carcass 133 removal and searcher efficiency [18,19].
134 Table 1. Bat fatality curtailment study data from the AWWIC and CanWEA databases, including project name,

135
year, geographic region, and rotor diameter in meters (RD); control cut-in speed (Cont.), experimental cut-in speed 136 (Exp.), and change in cut-in speed (Δ), all in m/s; and treatment effect information, including the mean fatality ratio 137 ± SE (Fatal. Ratio) and percent decrease in fatality between treatments (%). Studies from the same project and year 138 were tested simultaneously and share a control. Some studies lacked information on fatality uncertainty; for these, 139 the global average standard error was applied to the fatality ratio. focusing on different windows of time during fall migration; our approach controls for study-specific 154 variability by pairing control-treatment groups for analysis, but does assume that the relationship between 155 turbine-hours and fatalities is robust to potential variation in the effect of curtailment through time. Few 156 studies reported species-specific fatality rates, so fatality estimates were for all bat species combined.

Cut-in Speed Effect
157

Meta-analysis
158 The effect size of each study was calculated as a log-transformed ratio between the estimated 159 fatality of the treatment and the control, both in the unit of bat fatalities per turbine hour (i.e., the log-160 transformed response ratio, hereafter 'RR'). In instances where only a percent decrease was reported, this 161 was used to calculate the RR (log(1-(% decrease/100)) = RR). This effect size approach controls for 162 differences in study design ranging from site-specific effects to the choice of fatality estimator.

163
Manufacturer cut-in speed can vary among turbine makes and models. In most studies, the control 164 group's cut-in speed was 3.5 m/s (a common cut-in speed set by turbine manufacturers), though values 165 ranged from 3.0 to 5.0 m/s. Experimental cut-in speeds varied from 4.0 to 7.0 m/s. Due to this variation 166 and the small sample size of available studies, the change in cut-in speed between treatment and control of relative rather than absolute change in curtailment cut-in speed.

169
We used a random effects meta-analysis that accounts for heterogeneity in the true effect

170
(between-study variance) and sampling error (within-study variance; [24]). The inclusion of between-171 study variance (i.e., the random effect τ) allows for the incorporation of additional uncertainty in the 172 analysis by assuming the true effect is a random variable that is realized at different magnitudes in 173 different studies. Confidence intervals (90% or 95% depending on the estimator used) for control and 174 treatment fatality estimates were converted into standard error (SE) estimates assuming an approximately 175 normal distribution. While confidence intervals were slightly asymmetrical, a normal approximation was 176 the best available strategy for conversion given the variation in fatality estimators used across studies. For 177 independent studies (i.e., those with no shared control), standard error estimates of the RR were 178 calculated using the delta method [25,26]. In instances where multiple studies shared a common control 179 (i.e., were conducted simultaneously at the same project site; n=23), the correlation among the studies 180 was calculated by decoupling the associated dependence into a single estimate of uncertainty for each 181 study [26,27]. In instances where no estimate of uncertainty was provided in the original study, the mean 182 SE of all studies after decoupling was applied to the estimate.

183
To conduct the meta-analysis and explore the possible relationships between Δ cut-in speed and 184 bat fatality rates, we ran two types of models with the RR as the dependent variable and Δ cut-in as the 210

211
To determine the number of studies required in a random effects meta-analysis to detect relative 212 changes in RR with changing cut-in speed reliably, we implemented a power analysis at the meta-analysis 213 scale using a simulation approach [31]. We conducted meta-analysis power analyses for the non-linear 214 categorical and linear continuous descriptions of the relationship between Δ cut-in speed and RR. For the 215 categorical relationship power analysis, simulations were designed using the Δ cut-in speed categories 216 defined above to replicate the meta-analysis under multiple scenarios. The number of studies per Δ cut-in category, fatality reduction for the first Δ cut-in speed category (β 0 ), and the subsequent reduction in the 218 second and third categories (β 1 , β 2 ), were varied across simulations. The following linear regression 219 equation was used for the categorical model: where X 1 and X 2 are dummy covariates that represent Δ cut-in Categories 2 and 3, respectively. The 222 uncertainty from the Gaussian error term (ε) and inter-study differences (τ) were added by using a normal 223 distribution with a mean of 0 and a standard deviation equal to that observed in the fatality ratio of 224 control-treatment data (Table 1; SD = 0.24). We used a uniform distribution to randomly assign a SE to 225 each simulated study, which ranged between the minimum and maximum of the observed study standard 226 errors (Table 1; range: 0.05-0.34). Once the model was simulated, we used the methods described above 227 to estimate parameters. Each scenario was simulated 10,000 times with 5, 10, 20, and 30 studies per Δ 228 cut-in category to achieve precise estimates of power. The statistical power of each parameter (β 0 , β 1 , and 229 β 2 ) and sign error (the probability that the estimate was the same sign as the given parameter; [32]) were 230 calculated to determine the effectiveness of the model in estimating the scenario parameters. Power was 231 determined by examining whether the results were significantly different from the value of no effect (1 232 for β 0 , and 0 for β 1 and β 2 ; α = 0.05), and the sign error was computed by comparing the signs of the true 233 parameter value and the estimated value.

234
For the linear continuous models, Δ cut-in speed was randomly assigned to each study. To do 235 this, we used the same category framework (where 5, 10, 20, or 30 studies were assigned to each Δ cut-in 236 category), and studies in this category were randomly assigned a Δ cut-in speed from that category that 237 was observed in the studies included in the meta-analysis. These values were then scaled (centered on 238 zero) and used to build a linear model: where 1 is the scaled continuous Δ cut-in speed value for each study. Five scenarios were simulated for the power analysis for the two model types (Table 2). Each 242 scenario was replicated at four different sample sizes (5, 10, 20, and 30 studies per Δ cut-in category).

243
Four of these scenarios were selected a priori to explore our power to detect different types of 244 relationships between fatality ratio and Δ cut-in. We included three scenarios thought to represent 245 plausible hypotheses based on observed results to date: 1) a 25% linear decrease in fatality per 1 m/s 246 increase in cut-in speed; 2) a 50% initial decrease in fatality with Category 1 Δ cut-in speed and 247 subsequently stable fatality rates; and 3) an initial 50% decrease in fatality with Category 1 Δ cut-in speed 248 and then 10% subsequent declines in fatality for Categories 2-3. The fourth scenario, a more extreme 50% 249 exponential decrease per 1 m/s increase in cut-in speed, was intended to provide context for interpreting 250 the results of other scenarios. We also included an a posteriori 'current knowledge' scenario that used 251 parameter estimates obtained from the top model in our meta-analysis (above).

267
New fatalities were generated each night for each turbine using a fatality rate per turbine-night as a

268
Poisson mean. Second, carcass persistence rate was estimated using a carcass persistence trial format.

269
Here, we used the exponential distribution to simulate the survival rates of 50 carcasses at the site based

291
Detection probability was fixed at 50% for all studies, the approximate median of the described studies. in the 25% and 50% reduction treatment groups (along with 95% credible intervals) were estimated using 303 a parametric bootstrapping approach (n = 1000). The 95% credible interval of the difference of the 304 GenEst-derived fatality estimates between these two groups was calculated to determine overlap with 305 zero and used to estimate statistical power for each scenario, and was determined by subtracting the 306 bootstrapped simulations for each treatment group. If a simulation study group did not detect any 307 carcasses, we did not include it in the power analysis calculation.

309
Results Fatality ratios in the database representing fatality reduction due to curtailment ranged from 0.13 311 (87% decrease in fatalities) to 1.00 (0% decrease in fatalities) with an arithmetic mean of 0.46 (53% 312 decrease; n=35 studies). Lower fatality ratio values represented a greater reduction in bat fatality per 313 turbine-hour, while a value of one indicates no difference between curtailment treatment and control (0% 314 decrease). When examining fatality ratios by Δ cut-in category, the mean fatality ratio for Category 1 was 315 0.60 (n = 12), Category 2 was 0.41 (n = 18), and Category 3 was 0.37 (n = 6), suggesting a possible non-316 linear relationship with Δ cut-in speed (Fig. 1). 317

320
Error bars represent the standard error of the fatality ratio. Talbot Wind was excluded from the meta-analysis as an 321 outlier.

Meta-analysis 324
Thirty-five individual studies (from 16 projects) were included in the meta-analysis modeling.

325
The estimated fatality ratio across all studies (i.e., the estimate before controlling for Δ cut-in speed) was

334
Based on the linear model, the RR tended to decrease with increasing Δ cut-in (slope parameter β = -0.17,

361
Error bars are 95% confidence intervals of estimates.

363
The model estimates for fatality ratios for Categories 2 and 3 were 0.42 and 0.34 respectively, but the 364 marginal change of increasing Δ cut-in from Category 1 to Category 2 (β 1 = -0.19, CI: -0.59-0.20, z = -365 0.96, p = 0.33) and from Category 1 to Category 3 (β 2 = -0.46, CI: -1.02-0.11, z = -1.58, p = 0.11) were 366 small, with high amounts of uncertainty in the estimates (Fig. 3) Fig. 4). The exception was the 25% linear decrease 377 scenario, which required over 30 studies to achieve adequate power due to smaller changes at lower Δ 378 cut-in speeds. The statistical power of β 1 and β 2 (Δ cut-in Categories 2-3) were more variable across 379 scenarios (Fig. 4). For β 1 (1.5-2.3 m/s Δ cut-in speed), the 50% exponential decrease scenario had 380 sufficient power at 20 or more studies, and the 25% linear decrease scenario had sufficient power at 30 381 studies per group, but no other scenario met the criteria for sufficient power. For β 2 (3-3.5 m/s Δ cut-in 382 speed), two scenarios achieved sufficient power with less than 10 studies per group (50 % exponential 383 decrease, 25% linear decrease), while another two achieved sufficient power with 20-30 studies per group 384 (50% decrease followed by 10% decreases, and current knowledge scenario). Sign error decreased with 385 increasing sample size for all parameters except those that were set at zero (β 1 and β2 in the 50% decrease 386 then stable scenario) and decreased below 10% at 10 studies per category for most other parameter 387 estimates.
388 Figure 5. The relationship of statistical power and sign error with sample size in the categorical meta-analysis-scale 389 power analysis of curtailment studies to reduce bat fatality rates. We examined the relationship between the number 390 of studies per category of Δ cut-in speed (Category 1 = β 0 = 0.5-1.3 m/s Δ cut-in speed, Category 2 = β 1 = 1.5-2.3 Δ 391 cut-in speed, Category 3 = β 2 = 3-3.5 m/s Δ cut-in speed) and 1) the statistical power to detect change between 392 categories (at top), and 2) the rate at which models would be expected to incorrectly predict the sign of parameter 393 estimates (at bottom). Colors represent different curtailment regime scenarios. The horizontal dashed lines represent 394 the 0.8 power threshold and 50% sign error threshold, respectively. In comparison, the continuous model often had higher power, particularly for constant or 397 increasing relationships between RR and Δ cut-in (Fig. 5). Power to detect linear trends (β 1 ), particularly 398 for the scenarios with decreases at Δ cut-in speeds greater than 1.3 m/s, was greater than 0.8 even with 399 only 5 studies per group. Only the 50% then stable and current knowledge scenarios showed poor power,  (Fig. 6). However, the importance of turbine-nights varied with several variables outside of 418 researcher control, such as effect size and carcass persistence. With a 25% fatality reduction between 419 experimental and control treatments, no tested scenario achieved statistical power of 0.8 when the control 420 fatality rate was low (0.1 mortalities/turbine-night). For scenarios with a 25% reduction in fatality,

421
statistical power was high only when fatality rate, carcass persistence, and turbine-nights were also high 422 (Fig. 6A). The statistical power of studies in the 50% fatality reduction scenarios was more resilient to 423 changes in sampling period and carcasses persistence than the lower-reduction scenarios. Statistical 424 power was above the 0.8 threshold across almost all scenarios with high fatality rates (0.3 fatalities per turbine-night), and a large number of turbine-nights yielded strong statistical power even when the fatality 426 rate was lower (Fig. 6B). Sign error followed a similar pattern, errors occurred more often when fatality 427 rates and fatality reduction from curtailment were low (Fig. 6C). When fatality reduction was 50%, sign 428 error was almost always less than 10% (Fig. 6D). In summary, these simulation results suggest that many 429 curtailment study designs could be effective at detecting differences between treatments in situations with

442
Like past studies, we found evidence that turbine blanket curtailment reduces fatality rates of bats 443 at wind farms at sites that have implemented the technique (as reviewed by Arnett et al. [36]). However, 444 the marginal effect of increasing turbine cut-in speed on fatality rates is more difficult to assess. Using a 445 meta-analysis approach, we estimated that the effect of a 0.5-1.3 m/s increase in cut-in speed resulted in a 446 fatality ratio of 0.52, or a 48% reduction in bat fatalities. Estimated reductions in bat fatalities at higher ∆ 447 cut-in speeds were not found to be significantly different than this value and had high modeled 448 uncertainty. The sample size was small, particularly at higher ∆ cut-in speeds. Within the context of the 449 meta-analysis power analysis, we only had the statistical power to consistently detect reductions of ~50% 450 per 1 m/s ∆ cut-in speed. Combined with our meta-analysis results, it appears unlikely that larger 451 increases in ∆ cut-in speeds beyond Category 1 result in >50% additional fatality reduction (e.g., the 50% 452 exponential decrease scenario). Given that we lacked statistical power to detect changes in fatality ratio less than 50%, it is possible the true effect could still be large enough to be ecologically relevant to bat with lower overall fatality rates could also be instructive and curtailment studies are suggested for sites 502 that typically have high enough fatality rates to elicit conservation concern. Differentiation of fatality rates by species or species group could also help reduce our uncertainty.

504
Species-level traits such as migratory strategy, dispersal distance, and habitat association likely play an 505 important role in fatality risk [44]. For instance, long-distance migrants such as hoary bats, silver-haired 506 bats, and eastern red bats comprise a majority of fatalities at terrestrial wind energy facilities in North

507
America [3,8]. The project-specific risk is then correlated to species distributions, migratory routes, and 508 flight heights, among other characteristics [45]. Incorporating species-level information could improve 509 our understanding of bat fatality reduction, but this would require that species-level fatality estimates, or 510 at least species-group fatality estimates (i.e., migratory tree bats vs. Myotis spp.), be reported from 511 curtailment studies to allow for comparisons. Such estimates were not consistently reported by the studies 512 included in our analysis, often due to insufficient sample size.

513
The precision of meta-analysis parameters is likely to be overestimated in this study. While the 514 random effects meta-analysis framework adds uncertainty to model estimates based on among-study 515 variance [24], we did not account for site dependence as modeling approaches yielded unstable results.

516
Additionally, turbine operation, mortality estimator selection, and blanket curtailment implementation 517 varies substantially between sites (including time of year, time of night, species composition affected, 518 choice of cut-in speed, and turbine feathering), and these differences could affect the results in ways that 519 are difficult to incorporate into meta-analyses due to incomplete documentation of these protocols. While 520 we controlled for some of these potential biases by including variables like control cut-in speed and multi-521 treatment controls, the remaining uncertainties will likely be reduced best with increased sample size or 522 protocol documentation.

523
Recommendations for future studies

524
If blanket curtailment greater than 1.5 m/s above manufacturer specifications continues to be 525 implemented at wind facilities, additional experiments should be conducted to understand the relative 526 benefit of these increased cut-in speeds for reducing bat fatalities. The number of studies that tested ∆ cut-527 in speeds greater than 1.5 m/s were relatively few, and more studies that target these larger changes are needed. Estimates from the meta-analysis power analysis suggest that as many as 25 additional studies at 529 ∆ 2 m/s cut-in speed would be needed to effectively exclude the possibility of a 20% reduction in fatality 530 (and even more are needed to detect an additional 10% reduction). Conducting studies that compare 531 multiple treatment groups against a control during the same time period at the same location would 532 provide greater inferential power to answer such questions; though the costs of each individual study 533 would increase compared to single treatment studies, more could be gained in terms of understanding the 534 benefits of higher ∆ cut-in speeds. At the individual study level, statistical power is dependent on many 535 factors outside of the control of study designers (e.g., fatality rates and carcass persistence). Prior 536 knowledge of these parameters is valuable for designing effective studies, particularly if carcass 537 persistence rates are expected to be lower than average (e.g., due to high scavenging activity at the site).

538
To facilitate inclusion of studies in future meta-analyses, curtailment experiments should report 539 fatality estimates for both control and treatment groups, carcass persistence rates, searcher efficiency, 540 search frequency, search area coverage, number of turbine-nights of study, curtailment regime (including 541 whether feathering occurred), and turbine makes/models, with associated uncertainty values when 542 relevant. When sample size allows, fatality estimates should be reported by species or species group (e.g.,

543
Myotis) rather than for all bat species combined to facilitate taxon-specific assessments of curtailment 544 efficacy.

545
Newer operational minimization strategies have been developed to achieve similar fatality 546 reductions as blanket curtailment but with lower energy loss at higher cut-in speeds [46,47]. "Smart" 547 curtailment strategies, for example, which use additional environmental data besides wind speed to 548 inform the assessment of mortality risk and vary curtailment implementation, show promise to reduce the 549 economic impact of curtailment on wind energy projects [48][49][50]. Several deterrent systems that 550 discourage bats from approaching turbines are also in development and show some promise for reducing 551 fatalities while minimizing power loss [11,[50][51][52], and could be particularly beneficial if used in 552 combination with curtailment at lower wind speeds. While such approaches are still being evaluated, they may eventually represent a more cost-effective alternative to blanket curtailment, particularly blanket 554 curtailment at higher wind speeds.

556
The results of our meta-analysis suggest that blanket curtailment is effective at reducing bat 557 fatalities at terrestrial wind energy facilities, with the meta-analysis describing a mean fatality ratio of 558 0.44, or a 68% reduction in bat fatalities. Given our small sample size, particularly at higher Δ cut-in 559 speeds, our statistical power was limited to test the benefit of increasing cut-in speeds more than 1-1.3 560 m/s above the control cut-in speed. The power analysis suggests that differences in fatality ratio of 50% or 561 greater were often detectable even with small sample sizes (> 80% chance of significance), so it is likely 562 that the true value of incremental increases in Δ cut-in speed is below this 50% threshold. Whitby et al.

563
[37] suggest this is the case and that result combined with our marginally important effect in this study 564 provides more evidence that higher cut-in speeds can yield fewer mortalities. Though the small sample 565 sizes, low power in the present study, and variation in our respective results should engender caution 566 when interpreting these findings. Given the scope of bat fatalities at terrestrial wind farms in North 567 America [3,53], we must learn more about the management effectiveness of curtailment, particularly at 568 larger Δ cut-in speeds. Further development of "smart" curtailment strategies may also reduce fatalities 569 while moderating impacts to project finances [49].

570
The number of available studies in the current analysis limited our analytical options and findings 571 in several ways. If blanket curtailment continues to be a common strategy at wind speeds at ~5 m/s or 572 above (i.e., Δ cut-in speed of >1.5 with a standard factory cut-in speed of 3.5 m/s), we would recommend 573 conducting additional experimental curtailment studies with blanket curtailment treatments at these higher 574 cut-in speeds to strengthen our understanding of the relationship between increasing cut-in speeds and bat 575 fatality rates. Such studies must be carefully designed, ideally using an adaptive management framework

576
[54], to consider such variables as the expected fatality rate and carcass persistence rate when selecting a 577 search interval and defining the number of turbine-nights to monitor. Studies at sites with expected low