How does the quality and quantity of time series data affect extinction risk and population decline estimates in population viability analysis?

Background/Question/Methods

Assessing population trends is essential to ascertaining a taxon’s threat status and to the development of conservation management strategies. A widely used tool for this task is population viability analysis (PVA), where population models are used to estimate risk of decline or extinction within a specified time horizon. These risk estimates depend critically on both the quality and quantity of data available. In particular, the length of the time series, natural variability (process error), measurement error, and type of population model can all influence the accuracy of projections generated by PVA.  We evaluate how the above-mentioned sources of variability and uncertainty affect the reliability of predictions of population decline, both in isolation and in concert. Additionally, we assess the accuracy of population projections generated with scalar and matrix population models. Our analyses were performed under a virtual ecologist framework in which hypothetical sets of “true” age-structured time series data were simulated. These “true” time series were parameterized to exhibit different levels of variability and sampled assuming a range of levels of measurement error and data collection time spans. These pseudo-samples of data were then used to parameterize scalar and matrix models. The predicted population decline that arose from different models structures, measurement error, and process error were then compared to actual declines exhibited by the “true” time series.

Results/Conclusions

As expected, scenarios of high variability increased the uncertainty of the predictions generated by both scalar and matrix models. Surprisingly, after a certain point (15 years) the length of the time series sampled did not significantly improve the predictions. In general the uncertainty of the predictions was higher in the simulated models that were generated with larger lambdas. This implies that populations that are declining can be projected with greater confidence than populations that are increasing. Matrix models frequently revealed biases in their predictions whereas scalar models did not. Matrix models slightly underestimated percent decline under scenarios with low measurement error and high and low process error. Additionally, they overestimated percent decline under scenarios with high process and measurement error. In contrast, scalar models exhibited more reliable assessments of the percent decline.