Intro to Spatial Analysis & GIS for Spatial Epidemiology in R

SER Workshop

June 13, 2023��go.illinois.edu/SER2023

Marynia Kolak, PhD, MFA, MS

Qinyun Lin, PhD

Today’s workshop

Learning goals for today:

• Understanding the basics of spatial analysis & mapping in R
• Standardizing data at the neighborhood level
• Integrating data from another source
• Calculating new variables
• Generating maps for analysis

Today’s workshop

Follow along with live coding: https://makosak.github.io/Intro2RSpatialMed/

*We will take two 10-minute breaks

Environment Set-Up

• Working version of RStudio
• Install these packages: sf, tmap, tidygeocoder, & rgeoda
• If you don’t have a working R-spatial environment on your computer, create a free account in Posit & utilize today

Introduction to R Spatial

What is Spatial Data?

Spatial Data = Information + Location 🎯�

Spatial data is data that contains information about specific locations.

Note: the information content of the data may change with location.�

On some occasions, spatial data may only include “location.”

But without “location,” the data is no longer spatial.

Spatial Data Types

/ Polygon

Non-Spatial & Spatial Data Formats

• CSVs. Comma Separated Values (or other text files) can store spatial data if they have columns with latitude & longitude to represent coordinate locations

Non-Spatial & Spatial Data Formats

• Shapefile. A proprietary data format by ESRI made of at least 4 extension files (.shp, .shx, .dbf, and .prj)

��

• GeoJSON. An open standard format for encoding a variety of geographic data structures.

Non-Spatial & Spatial Data Formats

In R, we will work with the sf library. ��sf uses the following data structures:

• data.frame - table or a two-dimensional array-like structure.
• tibble - a “modern reimagining” of the data.frame (tidyverse)
• sf - Simple features refers to a formal standard (ISO 19125-1:2004) that describes how objects in the real world can be represented in computers, with emphasis on the spatial geometry of these objects. (Types: point, line, polygon; multipoint, multipolygon)

Coordinate Reference Systems

A Coordinate Reference System (CRS) communicates what method should be used to flatten or project the Earth’s surface onto a 2-dimensional map.

Different CRS imply different ways of projections and can generate substantially different visualizations. For example, the following are some world maps using different projections →

​

Defining your CRS

CRSs can be referred to using a Spatial Reference System Identifier (SRID) including EPSG codes. The EPSG Database is one of the most comprehensive databases that store thousands of different projections.

EPSG 4326 is one of the most common projections used today.

​

In {sf}, you can use the function st_crs to check the CRS used in a data, & st_transform to re-project the data to a particular CRS.

R-spatial Ecosystem

An active and growing online community:

• Geocomputation in R�https://geocompr.robinlovelace.net�
• Spatial Data Science with R�https://rspatial.org/�
• R Spatial Workshops from CSDS�https://spatialanalysis.github.io/tutorials/ �

The popularity of spatial packages in R. The y-axis shows average number of downloads per day, within a 30-day rolling window. Source: Geocomputation in R

Live demo

Mapping Neighborhoods

Thematic Mapping

Thematic or chloropleth maps represent quantitative data through colors or patterns in different geographic areas.

​

There are many different ways to classify & thematically map data:�

• Data classification methods
• Number of classes or bins

Image: Axis Maps

Equal interval classification with 5 bins

Data Classification Methods

Different methods arrange your data using different boundaries to separate classes.

In a quantile map, data is grouped into classes or bins that each have the same number of observations.

3 classes = tertile �4 classes = quartile�5 classes = quintile

Data Classification Methods

Different methods arrange your data using different boundaries to separate classes.

In a natural breaks map (also called Jenks map), data is grouped so that the within-group homogeneity is maximized.

Note the number of observations in each class can be highly unequal.

Data Classification Methods

Different methods arrange your data using different boundaries to separate classes.

A standard deviation map can be useful for helping identify outliers.

The variable is transformed to standard deviational units, covering the range from lowest to highest.

Data Classification Methods

Different methods arrange your data using different boundaries to separate classes.

Additional methods:

• Equal intervals
• Manual intervals
• Box map

Image: Axis Maps

Which method is best?

This depends on your data -- and the findings you are trying to convey

Quantile

Natural Breaks

Standard Deviation

Live demo

Spatial Cluster Detection

Formalizing Tobler’s Law: Spatial Autocorrelation

Using Test Statistics

• A statistic is any value that summarizes characteristics of a distribution
• Calculated from the data and compared to reference distribution
• How likely is the value if it had occurred under the null hypothesis?
• When unlikely (low p-value), the null hypothesis is rejected

When considering spatial autocorrelation, we test against a null hypothesis of spatial randomness.

​

Operationalizing Spatial Randomness

• under spatial randomness, the location of values may be altered without affecting the information content of the data
• Test it: random permutation or reshuffling of values

​

From Anselin, Fall 2017, SOC 20253

Spatial Autocorrelation Statistic

​

• Captures both attribute and locational similarity
• Need to construct index from data that captures both

​

Attribute similarity

Location similarity

Defining Spatial Weights

Before we can test for spatial autocorrelation, we need to define neighbors.

We do this by creating a spatial weight, which is a matrix where neighbor relationships for each unit are recorded.

When creating a spatial weight, think about how far the impact can reach...

Europe, queen contiguity weights

Spatial Weight - 2nd Order Queen Contiguity

Same area, different spatial weight (w)

2nd order queen contiguity

1st order rook contiguity

Contiguity-Based Weights

Rook Contiguity Weight

• Neighbors touch boundary, but not corners
• Can be 1st order and above
• Won’t connect islands or areas on their own

Queen Contiguity Weight

• Neighbors touch boundary and corners
• Can be 1st order and above
• Won’t connect islands or areas on their own

​

Distance-Based Weights

KNN Weight

• K-Nearest Neighbor Weight
• K is a number you choose, like 4
• If K=4, will find 4 nearest neighbors (based on distance of centroids)
• Will connect all areas, even

​

Europe, KNN-4 contiguity weights

Matrix Representation

If 2 are neighbors, the value = 1. Otherwise = 0.

Global Spatial Autocorrelation

Moran’s I Test Statistic:

​

zi = yi - mx : deviations from the mean

Cross product statistic ~ correlation stat.

z statistic: (observed - mean) / SD�Comparable across variables & spatial weights

​

Need to assess whether computed value of Moran’s I is significantly different from a value for a spatially random distribution:

​

1. Compute analytically using normal distribution, etc.
2. Compute computationally, comparing values to a reference distribution obtained from a series of random permutations

Moran’s I & the Permutation Approach

• Moran’s I changes with different spatial weights; use z-scores to compare
• We create a reference distribution and randomly reshuffle observations (using Monte Carlo permutations), computing Moran’s I each time
• Compare observed value to reference dist.
• High/Low Moran’s I; positive vs negative autocorrelation with 0 as the null hypothesis

Tip: More permutations (ex. n=999 vs 499), more stability in findings.

Spatial Autocorrelation Measures

Global Clustering Measure: Is there spatial clustering present?

• Moran’s I

Local Clustering Measures: Where and what type of spatial clustering is present?

• Geary’s C
• Spatial clusters (hot spot, cold spot)�
• LISA ( Local Indicator of Spatial Autocorrelation)
• Spatial clusters (hot spot, cold spot)
• Spatial outliers

Spatial Clusters

Spatial Outliers

Spatial Clusters

https://geodacenter.github.io/workbook/6a_local_auto/lab6a.html

Adding Resources

What is geocoding?

Geocoding is the process of converting addresses into geographic coordinates using a known coordinate reference system (CRS).

We can then use these coordinates (latitude, longitude) to spatially enable data.

​

In R, we use the tidygeocoder package

Read in csv or text file with addresses → data.frame

In R, we use the tidygeocoder package

Geocode addresses to generate latitude, longitude coordinates

Prepare data

In R, we use the tidygeocoder package

→ convert to spatial data sf

Mapping points

Now that addresses have been cleaned, geocoded, and spatially-enabled, they can be added as a point layer on a map

Live demo

Calculating Spatial Metrics

Accessibility Metrics: �How accessible is a resource?

Accessibility is a multidimensional concept, and spatial distance is only one component of accessibility. �

Classical model: (Penchansky & Thomas, 1981)

• Access: a degree of “fit” between clients and the system
• Availability, Accessibility, Affordability, Acceptability, Accommodation

Choosing the “right” measure is complicated

Location optimization considers:

• Supply and demand values at destinations and origins
• Optimal placement for facilities (coverage problem)
• Allocation of demand to given facilities (location-allocation problem)

Different access measures could lead to different conclusions about the existence of spatial mismatch and inequity (Talen & Anselin, 1998)

It is important to consider what characterization of access is most suitable for a specific problem

Choosing the “right” measure is complicated

How should distance between the user/patient and the facility be characterized?

​

What assumptions about travel behaviors are most appropriate?

Access Measurements

Density:

• Container method - buffers, point-in-polygon

Proximity:

• Minimum distance
• Travel cost
• Travel time
• Gravity potential

Container Method

A count of facilities (or measure of services) provided by any geographic unit.

For example:

• How many methadone providers are in each zip code?
• How many methadone providers are in a 1-mile buffer zone?
• What percentage of area is within a threshold of the provider’s service?

Proximity Methods

Minimum distance: the distance from an origin location to the nearest destination (e.g. methadone providers)�

Travel cost: measure of the total or average distance (or travel time) between each origin and destination�

Gravity potential: resources are weighted by size and adjusted for “friction of distance”

Minimum distance to the nearest methadone provider for all contiguous US zip codes

Live demo

Wrap Up & Discussion

Additional R-Spatial Resources

Opioid Environment Toolkit (HEROP resource): https://geodacenter.github.io/opioid-environment-toolkit/index.html

R Spatial Workshops (CSDS/HEROP resource): https://spatialanalysis.github.io/tutorials/ ��Geocomputation in R (Lovelace et al, 2019): https://geocompr.robinlovelace.net��Spatial Data Science with R: https://rspatial.org/

Other topics & useful tools

• Exploratory spatial data analysis (ESDA)
• Kernel density estimates (KDE)
• Spatial regression models
• Spatial lag model (SAR - spatial autoregressive lag model)
• Spatial error model (SEM)
• Some recent applications in public health:
• Principal component analysis (PCA) followed by clustering analysis and spatial regression using principal components as predictors

Kernel density estimates (KDE)

Approximates the distribution of point data along a continuous surface, so that one could understand how likely an event is to occur across spaces, or the “intensity” of activities.

Resources

• adehabitatHR: Home Range Estimation
• Point pattern analysis tutorial by Manuel Gimond

​

Spatial regression models (basics)

​

• Null hypothesis

y = X𝛽 + u, where E[uu’] = σ²I

• Alternative hypotheses
• Spatial lag model (SAR)

y = ρWy + X𝛽 + u

• Spatial error model (SEM)

E[uu’] = 𝚺 ≠ σ²I

Spatial regression models

• Textbook: Anselin, L. and S. Rey. (2014) Modern Spatial Econometrics in Practice: A Guide to GeoDa, GeoDaSpace and PySAL. Chicago: GeoDa Press LLC, Sections 5.1 and 6.1.
• Youtube: Spatial Regression Estimation
• Implementation:
• GeodaSpace
• Pysal

PCA followed by cluster analysis & spatial regression

• Gather multidimensional SDoH variables based on theoretical framework
• Principal component analysis to reduce all SDoH variables to a few PCs (each PC is orthogonal to each other)
• Cluster analysis to identify areas with similar SDoH characteristics
• Spatial regression analysis examining associations between SDoH (using PCs) and health outcomes
• SAR or SEM for spatial dependence (e.g., virus is highly transmissible)
• Spatial regime model to account for spatial heterogeneity (e.g., virus has also spread unevenly across place and populations)

Kolak, M., Bhatt, J., Park, Y. H., Padrón, N. A., & Molefe, A. (2020). Quantification of neighborhood-level social determinants of health in the continental United States. JAMA network open, 3(1), e1919928-e1919928.

Seven SDOH neighborhood typologies were optimized across the continental United States, as presented in this screenshot from the SDOH web application generated to visualize results.

SDOH Neighborhood Typology.

The four indexes/PCs were also used later in spatial regression models to examine how SDoH associates with COVID-19 outcomes differently in rural, urban, suburban areas, accounting for spatial dependence.

Lin, Q., Paykin, S., Halpern, D., Martinez-Cardoso, A., & Kolak, M. (2022). Assessment of structural barriers and racial group disparities of COVID-19 mortality with spatial analysis. JAMA network open, 5(3), e220984-e220984.

Thank you

Marynia Kolak

mkolak@illinois.edu��makosak

@MaryniaKolak

Qinyun Lin

qinyun.lin@gu.se �

qinyun-lin�@lin_qinyun