Glen Johnson, PhD Lehman College / CUNY School of Public Health [email protected]

25
University at Albany School of Public Health EPI 621, Geographic Information Systems and Public Health Glen Johnson, PhD Lehman College / CUNY School of Public Health [email protected] Introduction to Smoothing and Spatial Regression

description

University at Albany School of Public Health EPI 621, Geographic Information Systems and Public Health. Introduction to Smoothing and Spatial Regression. Glen Johnson, PhD Lehman College / CUNY School of Public Health [email protected]. Consider points distributed in space. - PowerPoint PPT Presentation

Transcript of Glen Johnson, PhD Lehman College / CUNY School of Public Health [email protected]

Page 1: Glen Johnson,  PhD Lehman College / CUNY School of Public Health glen.johnson@lehman.cuny

University at Albany School of Public HealthEPI 621, Geographic Information Systems and Public Health

Glen Johnson, PhDLehman College / CUNY School of Public Health

[email protected]

Introduction to Smoothing and Spatial

Regression

Page 2: Glen Johnson,  PhD Lehman College / CUNY School of Public Health glen.johnson@lehman.cuny

Consider points distributed in space

“Pure” Point process:Only coordinates locating some “events”.

Set of points, S ={s1, s2, … , sn}

Points represent locations of something that is measured. Values of a random variable, Z, are observed for a set S of locations, such that the set of measurements areZ(s) ={Z(s1), Z(s2), … , Z(sn)}

_____________________Examples include• location of burglaries• location of disease cases• location of trees, etc.

___________________________Examples include• cases and controls (binary outcome)

identified by location of residence• Population-based count

(integer outcome) tied to geographic centroids

• PCBs measured in mg/kg (continuous outcome) in soil cores taken at specific point locations

Page 3: Glen Johnson,  PhD Lehman College / CUNY School of Public Health glen.johnson@lehman.cuny

Example of a Pure Point Process: Baltimore Crime Events

Question: How to interpolate a smoothed surface that shows varying “intensity” of the points?

(source: http://www.people.fas.harvard.edu/~zhukov/spatial.html)

Page 4: Glen Johnson,  PhD Lehman College / CUNY School of Public Health glen.johnson@lehman.cuny

From: Cromely and McLafferty. 2002. GIS and Public Health.

Kernel Density Estimation

Page 5: Glen Johnson,  PhD Lehman College / CUNY School of Public Health glen.johnson@lehman.cuny

Kernel Density EstimationEstimate “intensity” of events at regular grid points as a function of nearby observed events. General formula for any point x is:

where xi are “observed” points for i = 1, …, n locations in the study area, k(.) is a kernel function that assigns decreasing weight to observed points as they approach the bandwidth h. Points that lie beyond the bandwidth, h, are given zero weighting.

Page 6: Glen Johnson,  PhD Lehman College / CUNY School of Public Health glen.johnson@lehman.cuny

Baltimore Crime Locations (Kernel Density)

Bandwidth = 0.007 Bandwidth = 0.05

Bandwidth = 0.1 Bandwidth = 0.15

0

20000

40000

60000

80000

100000

120000

140000

160000

Results from Kernel Density Smoothing in R

Page 7: Glen Johnson,  PhD Lehman College / CUNY School of Public Health glen.johnson@lehman.cuny

Source: http://spatialityblog.com/2011/09/29/spatial-analysis-of-nyc-bikeshare-maps/

Kernel Density Surface of Bike Share Locations in NYC

Page 8: Glen Johnson,  PhD Lehman College / CUNY School of Public Health glen.johnson@lehman.cuny

Examples of Values Observed at Point Locations, Z(s) :

Question: How to interpolate a smoothed surface that captures variation in Z(s)?

Page 9: Glen Johnson,  PhD Lehman College / CUNY School of Public Health glen.johnson@lehman.cuny

First, consider “deterministic” approaches to spatial interpolation:

• Deterministic models do not acknowledge uncertainty.

• Only real advantage is simplicity; good for exploratory analysis

• Several options, all with limitations. We will consider Inverse Distance Weighted (IDW) because of its common usage.

Page 10: Glen Johnson,  PhD Lehman College / CUNY School of Public Health glen.johnson@lehman.cuny

Inverse Distance Weighted Surface Interpolation

Define search parameters

Define power of distance-decay function

0,

0,1

0

01

Interpolate value at point as

( ) ( )

for neighboring observed values ( ),

where the weight

for distance .

pi

npi

i

n

i ii

i

di

d

s

Z s Z s

n Z s

d

Page 11: Glen Johnson,  PhD Lehman College / CUNY School of Public Health glen.johnson@lehman.cuny

Illustration: Tampa Bay sediment total organic carbon

Page 12: Glen Johnson,  PhD Lehman College / CUNY School of Public Health glen.johnson@lehman.cuny

True “geostatistical” models assume the data, Z(S) = {Z(s1), Z(s2), … , Z(sn)}, are a partial realization of a random field.

Note that the set of locations S are a subset of some 2-dimensional spatial domain D, that is a subset of the real plane.

Page 13: Glen Johnson,  PhD Lehman College / CUNY School of Public Health glen.johnson@lehman.cuny

General Protocol:

1. Characterize properties of spatial autocorrelation through variogram modeling;

2. Predict values for spatial locations where no data exist, through Kriging.

Page 14: Glen Johnson,  PhD Lehman College / CUNY School of Public Health glen.johnson@lehman.cuny

A semivariogram is defined as

for distance h between the two locations, and is estimated as

for nh pairs separated by distance hj (called a “lag”).

After repeating for different lags, say j =1, … 10, the semivariance can be plotted as a function of distance.

21(h) E( ( ) ( ))2

Z s Z s h

2

1

1ˆ( ) ( ( ) ( ))2

hn

j i iih

h Z s Z s hn

Page 15: Glen Johnson,  PhD Lehman College / CUNY School of Public Health glen.johnson@lehman.cuny

Given any location si, all other locations are treated as within distance h if they fall within a search window defined by the direction, lag h, angular tolerance and bandwidth.

Adapted from Waller and Gotway. Applied Spatial Statistics for Public Health. Wiley, 2004.

bandwidth

Page 16: Glen Johnson,  PhD Lehman College / CUNY School of Public Health glen.johnson@lehman.cuny

Example semivariogram cloud for pairwise differences (red dots) , with the average semivariance for each lag (blue +), and a fitted semivariogram model (solid blue line)

Page 17: Glen Johnson,  PhD Lehman College / CUNY School of Public Health glen.johnson@lehman.cuny

Characteristics of a semivariogram

Range = the distance within which positive spatial autocorrelation exists

Nugget = spatial discontinuity + observation errorSill = maximum semivariance

Page 18: Glen Johnson,  PhD Lehman College / CUNY School of Public Health glen.johnson@lehman.cuny

If the variogram form does not depend on direction, the spatial process is isotropic. If it does depend on direction, it is anisotropic.

Multiple semi-variograms for different directions. Note changing parameter is the range.

Surface map of semivariance shows values more similar in NW-SE direction and more different in SW-NE direction.

Page 19: Glen Johnson,  PhD Lehman College / CUNY School of Public Health glen.johnson@lehman.cuny

Kriging then uses semivariogram model results to define weights used for interpolating values where no data exists.The result is called the “Best Linear Unbiased Predictor”. The basic form is

01

( ) ( )p

i ii

Z s Z s

Where the λi assign weights to neighboring values according to semivariogram modeling that defines a distance-decay relation within the range, beyond which the weight goes to zero.

Page 20: Glen Johnson,  PhD Lehman College / CUNY School of Public Health glen.johnson@lehman.cuny

Several variations of Kriging:• Simple (assumes known mean)• Ordinary (assumes constant mean, though

unknown) [our focus this week]• Universal (non-stationary mean)• Cokriging (prediction based on more than one

inter-related spatial processes)• Indicator (probability mapping based on binary

variable) [you will see in the lab work]• Block (areal prediction from point data)• And other variations …

Page 21: Glen Johnson,  PhD Lehman College / CUNY School of Public Health glen.johnson@lehman.cuny

Example of two types of Kriging for California O3:

1. Ordinary Kriging (Detrended, Anisotropic)

-continuous surface

2. Indicator Kriging

- probability isolines

Page 22: Glen Johnson,  PhD Lehman College / CUNY School of Public Health glen.johnson@lehman.cuny

What if point locations are centroids of polygons and the value Z(si) represents aggregation within polygon i ?

Page 23: Glen Johnson,  PhD Lehman College / CUNY School of Public Health glen.johnson@lehman.cuny

With polygon data, we can still define neighbors as some function of Euclidean distance between polygon centroids, as we do for point-level data,

but now we have other ways to define neighbors and their weights …

Page 24: Glen Johnson,  PhD Lehman College / CUNY School of Public Health glen.johnson@lehman.cuny

i

Defining spatial “Neighborhoods”

Raster or Lattice:

Rook

Queen- 1st orderQueen- 2nd order

iii

Page 25: Glen Johnson,  PhD Lehman College / CUNY School of Public Health glen.johnson@lehman.cuny

Spatial Regression Modeling as a method for both • assessing the effects of covariates

and…• smoothing a response variable