COS 314-6 - An empirical test of the number of visits required to estimate occupancy in a forest game bird when detection probabilities vary by site

Abstract: Occupancy models are a widely used method for evaluating the associations between species occurrence and environmental variables. Fitting these models requires species detection data from multiple visits at each survey location. As repeated surveys can be both costly and logistically difficult, investigators are frequently faced with the challenge of minimizing visits while obtaining robust estimates of occupancy and its predictors. Previous simulation and theoretical analyses have suggested a range of 3-5 visits as the minimum necessary to estimate parameters in an unbiased manner. However, there have been few empirical tests of this assertion, and none that have considered more than a maximum of 6 visits. Here we apply two emerging technologies, automated recording units (ARUs) and automated classification methods, to conduct an empirical occupancy study for American Woodcock (Scolopax minor) across 99 survey points in Pennsylvania, USA. These technologies allowed us to efficiently construct an unusually long 19 day detection history at each point. We subsample this full detection history to create surveys of 2 through 19 visits, allowing us to evaluate the influence of the number of visits on occupancy estimates and the influence of eight habitat covariates.

We find strong evidence of heterogeneity in detection probabilities across sites in our empirical data, allowing us to examine the issue of heterogeneity and appropriate means to address it. First, we show that heterogeneity such as that found in our data can arise from two potentially common spatial processes. Second, we demonstrate that a little used approach for occupancy modeling that allows for arbitrary detection heterogeneity, the non-parametric Chao2 estimator, is able to produce stable and unbiased estimates of occupancy after 6 visits, whereas a traditional occupancy model ignoring heterogeneity produces monotonically increasing estimates of occupancy with visit length. Third, we note that estimates of the effect of habitat covariates on occupancy should be unbiased even if some occupied sites remain undetected, so long as detection probability is not correlated with occupancy covariates. We show that in this data set nearly unbiased occupancy covariate estimates can be obtained from any number of visits using simple logistic regression. Together, this work suggests a straightforward multi-part approach for modeling occupancy in a frequentist framework when detection heterogeneity is present and data from sufficient visits are available.