Analysis of Roadkill Rates from a Trial of a Virtual Fence for Reducing Wombat Road Mortalities. Are Severely Under-Powered Studies Worth the Effort?

• Published in: Current Journal of Applied Science and Technology pp. 1 - 13
• Main Authors: Candy, Steven G., Englefield, Bruce
• Format: Journal Article
• Language: English
• Published: 14.09.2022
• ISSN: 2457-1024; 2457-1024
• DOI: 10.9734/cjast/2022/v41i333946

Aims: Evaluation of the effectiveness of methods of roadkill mitigation requires field experiments and thus experimental designs. We evaluated such a published experiment that used a virtual road-fence (VF) system to attempt to reduce roadkill of bare-nosed wombats (Vombatus ursinus) and draw general conclusions on the ability of the implemented design, given overall mean rates of roadkill, to detect a substantial reduction of 50%. Study Design: The study site consisted of contiguous fenced and unfenced road sections that were each monitored for road kill for 995 and 322 days pre- and post-installation, respectively. The study design was an unreplicated BACI (Before-After-Control-Impact) design. Methodology: No statistical analysis of the roadkill data were carried out in the original study. We used a Poisson/log-link Generalized Linear Model fit to their BACI data as a single 2x2 table of counts and define a single interaction parameter as a function of control-adjusted rate reduction due to the VF. This parameter can also be defined as the logarithm of the ratio of odds of a random kill in the fenced section occurring in the post-installation period to the corresponding odds for the unfenced section. A null hypothesis of no effect of the VF corresponds to a log odds-ratio of zero. We use simulation to show that estimates of this parameter have close to a Gaussian distribution and from this derive an estimate of the statistical power of the design to detect a hypothetical effect of a 50% reduction in rate. We also used simulation to estimate the corresponding power to detect this reduction if the BACI had been physically replicated three times. Results: The estimated log odds-ratio and the corresponding percent reduction inferred for the VF were -0.02 (SE=0.53) and 1.8%, respectively, and the statistical power of the design to detect a hypothetical effect of a 50% reduction in rate was only .35 but gave an improved power of .7 under the hypothetical replication. Conclusion: Our results showed that there was very low confidence in this design combined with similar overall roadkill rates being able to detect a 50% reduction. This increased to medium-to-high confidence if the experiment had been replicated in three locations. The very small point estimate of reduction could have been so low compared to a true but modest effect purely by chance, thus our conclusion is that there is very low confidence in any estimate of the VF’s effect obtained using this study’s design in the context of common roadkill rates.