Effects of sampling conditions on DNA-based estimates of American black bear abundance

DNA-based capture-mark-recapture techniques are commonly used to estimate American black bear (Ursus americanus) population abundance (N). Although the technique is well established, many questions remain regarding study design. In particular, relationships among N, capture probability of heterogeneity mixtures A and B (p_A and p_B, respectively, or p, collectively), the proportion of each mixture (π), number of capture occasions (k), and probability of obtaining reliable estimates of N are not fully understood. We investigated these relationships using 1) an empirical dataset of DNA samples for which true N was unknown and 2) simulated datasets with known properties that represented a broader array of sampling conditions. For the empirical data analysis, we used the full closed population with heterogeneity data type in Program MARK to estimate N for a black bear population in Great Smoky Mountains National Park, Tennessee. We systematically reduced the number of those samples used in the analysis to evaluate the effect that changes in capture probabilities may have on parameter estimates. Model-averaged N for females and males were 161 (95% CI = 114–272) and 100 (95% CI = 74–167), respectively (pooled N = 261, 95% CI = 192–419), and the average weekly p was 0.09 for females and 0.12 for males. When we reduced the number of samples of the empirical data, support for heterogeneity models decreased. For the simulation analysis, we generated capture data with individual heterogeneity covering a range of sampling conditions commonly encountered in DNA-based capture-mark-recapture studies and examined the relationships between those conditions and accuracy (i.e., probability of obtaining an estimated N that is within 20% of true N), coverage (i.e., probability that 95% confidence interval includes true N), and precision (i.e., probability of obtaining a coefficient of variation ≤20%) of estimates using logistic regression. The capture probability for the larger of 2 mixture proportions of the population (i.e., p_A or p_B, depending on the value of π) was most important for predicting accuracy and precision, whereas capture probabilities of both mixture proportions (p_A and p_B) were important to explain variation in coverage. Based on sampling conditions similar to parameter estimates from the empirical dataset (p_A = 0.30, p_B = 0.05, N = 250, π = 0.15, and k = 10), predicted accuracy and precision were low (60% and 53%, respectively), whereas coverage was high (94%). Increasing p_B, the capture probability for the predominate but most difficult to capture proportion of the population, was most effective to improve accuracy under those conditions. However, manipulation of other parameters may be more effective under different conditions. In general, the probabilities of obtaining accurate and precise estimates were best when p≥ 0.2. Our regression models can be used by managers to evaluate specific sampling scenarios and guide development of sampling frameworks or to assess reliability of DNA-based capture-mark-recapture studies.