Approaches To Analyzing Multi-Scale Data For Population Studies: Issues Of Non-Detection And Spatial Dependence
Jeffrey E. Moore, Purdue University, Department of Forestry and Natural Resources, 715 West State St., West Lafayette, IN 47907; (765) 494-9597; jeffmoore@fnr.purdue.edu
Robert K. Swihart, Purdue University, Department of Forestry and Natural Resources, 195 Marsteller Street, West Lafayette, IN 47907
A common problem in ecology is to describe relationships between habitat or landscape characteristics and the prevalence or abundance of a species across a set of locations in space. Traditional approaches to this problem usually involve the measurement of habitat and landscape variables at one or more spatial scales for a set of sample sites, documenting species presence-absence or some index of abundance at those sites, and the use of generalized linear models such as Poisson or logistic regression to statistically elucidate relationships between predictor and dependent variables. However, most traditional analyses fail to account for the influence of non-detection error or dependence inherent to spatial or hierarchical data on occupancy or abundance estimates. Recent statistical developments address these problems. Here we review what we consider some of the most useful and promising methods for analyzing data with inflated zero-counts or correlated error structure. These include zero-inflated binomial (ZIB) and Poisson (ZIP) regression for use with presence-absence and count data, respectively, conditional autoregression (CAR) and autologistic regression for data with spatially-dependent variance structure, and hierarchical linear modeling (HLM) for data with a nested structure (e.g., sites within patches, patches within landscapes, etc). We discuss applications of these methods, software developed to implement them, and provide some examples from vertebrate data collected from several hundred point locations distributed across 35 3x3-mile landscapes in the upper Wabash River basin of northern Indiana.