Revisiting Traditional Problems in Spatial Analysis through the Lens of Local Modeling

Project: Research project

Project Details

Description

The analysis of spatial data and spatial processes has long been recognized as having challenges not seen in other forms of analytics. Scale sensitivity, the possible existence of local contextual effects, and the need to account for the special properties of spatial data in inferential testing are issues that challenge the credibility of spatial analytical research. We take a fresh look at each of these challenges through the lens of local modeling, specifically utilizing the recent advances in multiscale geographically weighted regression (MGWR) developed from prior NSF support. In terms of scale sensitivity, we reexamine the modifiable areal unit problem, and its extreme case known as Simpsons paradox in the context of spatially varying processes. These issues have plagued spatial analysis for decades but important insights can be gained by viewing these as problems connected with processes rather than data. In terms of spatial context, we explore how MGWR, through its localized intercept, can identify the existence of contextual effects and measure its importance. As well as forming a bridge between nomothetic and idiographic views on geographical research, this has important ramifications for the current debates in reproducibility and replicability. In terms of inference, we will develop new methods of adjusting for multiple hypothesis testing in local models and a new suite of diagnostics for local models based on the concept of part-data. Both of these inferential developments are critical given the surge in the applications of local models.
Intellectual Merit
Completing this research will advance the analysis of spatial data in three ways:
(i) by providing a new understanding of scale issues through focusing on the processes that generated the data rather than on properties of the data themselves, this will bring a fresh understanding of the modifiable areal unit problem and Simpsons paradox;
(ii) by being able to quantify the role of context in key areas such as the effect of Covid-19 on excess death rates, spatial variations in the incidence of teen pregnancy rates, and the increasingly partisan and acrimonious divisions in voting behavior across the country, we aim to show the relevance of place-based geographical studies and the need to reevaluate the roles of reproducibility and replicability in research where the underlying processes might be spatially varying; and
(iii) by deriving correct inferential tests for local models which account for dependent multiple hypothesis testing and the use of part-data, we will provide the diagnostic tools to better inform the multitude of applications of local models that are now appearing in the literature. These will be incorporated into freely available local modeling software that has already had over 4,400 downloads.

Broader Impacts
Local statistical models are employed extensively across the spatial sciences. For instance, our current bibliography of (M)GWR publications runs to over 2,500 references and there have been over 4,400 downloads of our local modeling software. Consequently, any advances in both the interpretation of scale and context in local modeling and the development of better diagnostic tools will have major impacts across many disciplines and in a multitude of applied contexts. In addition, the insights that will be gained into the MAUP and Simpsons Paradox will have impacts on the understanding of scale effects on spatial analytical techniques outside the realm of local modeling. Similarly, revelations on the importance of context will provide a solid raison detre for geographical studies. Overall, by reducing barriers to credibility across several important areas, the work undertaken through this project will substantially strengthen spatial analytics in general and maintain geographys strong position as a leader in the development of spatial analytics. The methods developed here will be incorporated into freely available software and in workshops and training courses.

StatusActive
Effective start/end date10/1/219/30/24

Funding

  • National Science Foundation (NSF): $399,920.00

Fingerprint

Explore the research topics touched on by this project. These labels are generated based on the underlying awards/grants. Together they form a unique fingerprint.