Thematic Session (SDS Symposium 23')

Leveraging Geographic Context at Multiple Scales: The Salience of the Neighborhood in Statistical Learning & Causal Analysis

In cutting-edge spatial learning methods, the 'neighborhood' is often used as a way to pool information, improving predictions and regularising estimates. But, the correspondence between the statistically-useful neighborhoods that our methods identify and the actual neighborhoods that matter to people is generally unknown and under-examined. Instead, new methods proceed apace, coming up with new, better, more efficient ways of learning from context. This session seeks to provide a platform for those interested in neighborhood effects themselves and their correspondence with the neighborhoods that are salient for individual behavior. In addition, this session seeks to provide a home for those interested in developing new local spatial learning methods that learn from geographical context to improve predictions or regularise estimates.

Photo of Taylor M. Oshan (University of Maryland)

Taylor M. Oshan (University of Maryland)

Generalized Additive Spatial Smoothing (GASS)

The geographical context of individuals, households, or places of interest is often conceptualized using ‘neighborhoods’ that encapsulate the characteristics of the surrounding environment. ‘Neighborhood effects’ capture the relationship between an outcome and one or more variables defined within the neighborhood around the outcome. The definition of neighborhoods is therefore critical for accurately measuring neighborhood effects in a wide variety of applications across disciplines, such as sociology, epidemiology, economics, urban planning, and criminology. Though administrative boundaries are sometimes conveniently used as neighborhoods, they may not actually align with neighborhoods as they are experienced and are limited in their ability to represent their inherently ‘fuzzy’ nature. Another strategy, sometimes referred to as ‘egohoods’, employs potentially overlapping buffers and distance-weighted smoothing functions around each observation in order to construct fuzzy, idiosyncratic neighborhoods. However, it is often not clear how to parameterize egohoods, such as selecting the appropriate buffer size or smoothing function, and the same parameters (i.e., spatial scale) are assumed to apply for each relationship in a model. This paper introduces a new methodology called generalized additive spatial smoothing (GASS) that provides a flexible multiscale smoothing framework for simultaneously selecting variable-specific neighborhoods and modeling neighborhood effects.

Photo of Levi John Wolf (University of Bristol)

Levi John Wolf (University of Bristol)

A Quadtree-based algorithm for clustering regression

Recent work on 'local' geographical models has focused on exploring model specifications where covariates are allowed to vary at distinctive spatial scales. In these models, the 'spatial scale' that is learned reflects a kind of spatial smooth parameter that is held constant across the map. An alternative collection of techniques, known as “clustering regression,” seek to find latent 'clusters' in the data, and use these clusters to improve predictions and construct local parameter estimates. These discrete local regressions have not been systematically compared to other kinds of local regression but offer important practical and theoretical differences for their application. This paper will provide an overview of contemporary theory and model frameworks for clustering regression and outline a novel hierarchical technique for clustering regression and inference relying on quadtrees. The model will be compared to existing approaches for both discrete and continuous local regression, and recommendations on usage will be provided.

Photo of Ana Petrović (Delft university of Technology)

Ana Petrović (Delft university of Technology)

Segregation in the Netherlands over time, space and scale

Spatial segregation of socioeconomic and ethnic groups affects economic and social functioning of cities as integral urban systems as well as individual outcomes of people, such as income, education or health. Both causes and consequences of segregation include many different processes, such as those related to housing or labour markets, which occur at different spatial scales, ranging from small neighbourhoods to urban regions. Segregation at all these spatial scales changes over time, which becomes particularly relevant in the conditions of increasing economic inequalities, international and internal migration, and population aging. However, most of the empirical evidence about segregation in the Netherlands is cross-sectional and uses single spatial scales. Therefore, it is not clear at which spatial extents segregation is increasing or decreasing, and, therefore, also not straightforward what drives segregation trends and how to deal with the segregation in different places. Using individual-level register data from 1999 onwards, geocoded at 100m by 100m grid cells, on the OSSC (ODISSEI Secure Supercomputer), we investigate segregation trends in the Netherlands at multiple spatial scales, taking into account various sociodemographic characteristics of people.

Photo of Mehak Sachdeva (New York University)

Mehak Sachdeva (New York University)

Exploring the Influence of Profile Similarity in Places on Human Behavior in Spatial Contextual and Behavioral Models

This research talk explores the influence of similarity in place characteristics on human behavior within spatial contextual and behavioral models. While such models traditionally assume that processes are similar for nearby places and employ spatial weighting, they often overlook the impact of similarity in attributes for spatially distant places. Our study aims to investigate whether places sharing comparable measurable factors affecting behavior also influence the values, beliefs, attitudes, and behavior of their inhabitants similarly. Additionally, we examine the significance of this influence in conjunction with spatial proximity and the novel insights it provides about places. By employing advanced statistical techniques, such as multidimensional scaling, we aim to identify commonalities among different places with similar contextual determinants of behavior beyond their geographical locations. Establishing a connection between the intensity of contextual effects and distinct locational profiles enables us to gain a deeper understanding of the interplay between place characteristics and behavioral outcomes. The findings from these inquiries hold transformative potential across numerous fields in the social sciences, providing a deeper understanding of the complex interactions between places and human behavior.

Photo of Ziqi Li (Florida State university)

Ziqi Li (Florida State university)

Measuring contextual effects through local eXplainable Artificial Intelligence (XAI)

Recent advancements in local eXplainable Artificial Intelligence (XAI) methods open up new avenues to leverage the flexibility and accuracy of machine learning models for studying spatial phenomena. These methods offer model explainability at the individual level, which helps us gain insights into how individuals behave and interact within specific spatial and covariate contexts. This short presentation will introduce methods that can measure contextual effects within machine learning models. The combination of model flexibility, accuracy, and interpretability holds great promise in advancing our understanding of complex spatial processes.