A Hybrid Approach to Identifying Local Spatial Variation

• Activity Number: 410
• Type: Contributed
• Date/Time: Tuesday, July 31, 2012 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics and the Environment
• Author(s): Marc Scott*+ and Sean Jack Buckley
• Companies: NYU and National Center for Education Statistics
• Keywords: multilevel model ; spatial autocorrelation ; neighborhood effects ; school test scores

Abstract

We combine approaches of spatial data analysis with multi-level modeling to formally assess global and local spatially driven variation. Cluster-based areal units are used to capture variation residing between given areal units and individual sites. These empirically constructed neighborhoods are then incorporated into a multilevel model to decompose variation into crossed and nested components that systematically measure different levels of spatial variation. An intermediate level is introduced, which offers an alternative to spatial error models. When attempting to control for unobserved neighborhood effects, this approach integrates well with the multilevel framework. We compare the resulting decompositions to equivalent constructs from spatial error models. We apply this approach to six years of New York City public elementary school test scores. NYC school districts are natural areal units, representing administrative practices and coarse neighborhood effects, but local neighborhood effects using data of this nature have heretofore been difficult to identify. Our modeling approach addresses this. We also discuss the robustness of the approach.