Sharks are experiencing population declines worldwide, making the collection and analysis of abundance data critically important for understanding shark population ecology and conservation. However, shark abundance time series are typically short, highly variable, and often have a high proportion of zeros, which are properties that make typical analysis difficult or potentially unreliable and can be influenced by shark ecology. Therefore, the statistical methodology and sampling effort employed can affect conclusions about shark population ecology and conservation. We aim to evaluate how analytical methods influence estimates of shark population trends, using simulation models and empirical data. We assessed the effects of the magnitude of changes in population abundance (trend strength), variability in population abundance, and the degree of zero inflation on the ability of generalized linear models (GLMs) with differing distributions for the response to infer accurate conclusions about populations for common, rare, and schooling model shark species as well as for populations with different movement patterns. We also compare the results of different GLMs fit on a long-term time series of shark and ray observations at Cocos Island off Costa Rica to infer the consequences of using a particular statistical modeling framework for our knowledge of real shark populations.
Results/Conclusions
Stronger trends and decreasing variability result in more reliable conclusions about the decline of shark populations. In addition, schooling and mobile behaviors and zero inflation can increase dispersion, leading to overestimates of mild decline or underestimates of severe declines. Overall, the conclusions from linear regressions, Poisson GLMs, and negative binomial GLMs do not differ substantially on simulations, particularly when declines are not severe, but binomial GLMs on presence-absence data differ drastically from other models. Additionally, linear regressions can greatly overestimate severe declines due to an assumed underlying normal distribution. In contrast, zero inflated negative binomial models have the highest accuracy due to their ability to account for multiples sources of dispersion in the data. On real data, linear regressions and binomial GLMs infer less severe declines than Poisson or negative binomial GLMs, indicating that dispersion in real shark populations likely calls for the use of appropriate GLMs and the collection of abundance data rather than presence-absence or compositional data. Monitoring decisions require reliable information on shark population trajectories to be efficient and effective. Careful consideration of the biology of the target species and the property of the time series need to be considered when analyzing abundance data.