Presented By: Department of Political Science
Interdisciplinary Seminar in Social Science Methodology (I3SM)
U-M Political Science Doctoral Student Marty Davidson
Marty Davidson will be presenting his paper, "TPRS-Forest Estimation."
Abstract
This paper proposes an estimation strategy – TPRS-Forest, which researchers can use to model spatial data with poorly understood empirical properties. TPRS-Forest estimation uses the projected coordinates of a spatial variable and decomposes it into three components: spatial trend, systematic variation along the spatial trend, and random noise. The first stage uses thin-plate regression splines (TPRS) to approximate a non-stationary spatial trend, or a change in the conditional mean across the region of observation. Because TPRS represents an approximation, this stage does not model all spatial dependencies. To compensate, the second stage performs a random forest regression on the TPRS residuals, which accounts for systematic variation along the spatial trend. The final stage takes the filtered residuals from the second step and tests for any remaining spatial correlative structure, using a global and local Moran test. This proposed method merges together two literatures of spatial modelling – geoadditive models (Kammann & Wand, 2003; Wood, 2017) and spatial random forests (Georganos et al., 2019; Hengl et al., 2004, 2018) – and anchors them in a social scientific conceptual framework. In developing this method, my primary goal is to provide a way for social scientists to descriptively model their data and produce a set of continuous predicted values while minimizing the assumptions on how best to model their data's spatial properties.
The primary function of the Interdisciplinary Seminar in Social Science Methodology (I3SM) is to provide an interdisciplinary forum for students and faculty to present their current projects and to receive feedback on either the methodological component of their project or a methodology under development. Presenters can also present new research questions and ideas and receive ideas about which methodologies would work best to tackle such questions. We define methodology broadly as the approaches to which data is collected and/or organized to give empirical content to social science research. It includes both qualitative and quantitative methodologies.
To join the meeting via Zoom, email skuzushi@umich.edu for the meeting link.
Abstract
This paper proposes an estimation strategy – TPRS-Forest, which researchers can use to model spatial data with poorly understood empirical properties. TPRS-Forest estimation uses the projected coordinates of a spatial variable and decomposes it into three components: spatial trend, systematic variation along the spatial trend, and random noise. The first stage uses thin-plate regression splines (TPRS) to approximate a non-stationary spatial trend, or a change in the conditional mean across the region of observation. Because TPRS represents an approximation, this stage does not model all spatial dependencies. To compensate, the second stage performs a random forest regression on the TPRS residuals, which accounts for systematic variation along the spatial trend. The final stage takes the filtered residuals from the second step and tests for any remaining spatial correlative structure, using a global and local Moran test. This proposed method merges together two literatures of spatial modelling – geoadditive models (Kammann & Wand, 2003; Wood, 2017) and spatial random forests (Georganos et al., 2019; Hengl et al., 2004, 2018) – and anchors them in a social scientific conceptual framework. In developing this method, my primary goal is to provide a way for social scientists to descriptively model their data and produce a set of continuous predicted values while minimizing the assumptions on how best to model their data's spatial properties.
The primary function of the Interdisciplinary Seminar in Social Science Methodology (I3SM) is to provide an interdisciplinary forum for students and faculty to present their current projects and to receive feedback on either the methodological component of their project or a methodology under development. Presenters can also present new research questions and ideas and receive ideas about which methodologies would work best to tackle such questions. We define methodology broadly as the approaches to which data is collected and/or organized to give empirical content to social science research. It includes both qualitative and quantitative methodologies.
To join the meeting via Zoom, email skuzushi@umich.edu for the meeting link.
Explore Similar Events
-
Loading Similar Events...
