MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway...
Transcript of MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway...
![Page 1: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/1.jpg)
MODULE 16: Spatial Statistics in Epidemiologyand Public Health
Lecture 5: Spatial regression
Jon Wakefield and Lance Waller
1 / 1
![Page 2: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/2.jpg)
References
I Waller and Gotway (2004, Chapter 9) Applied SpatialStatistics for Public Health Data. New York: Wiley.
I Elliott, P., et al. (2000) Spatial Epidemiology: Methods andApplications, Oxford: Oxford University Press.
I Haining, R. (2003). Spatial Data Analysis: Theory andPractice. Cambridge: Cambridge University Press.
I Bailey, T.C. and Gatrell, A.C. (1995) Interactive Spatial DataAnalysis. Essex: Addison Wesley Longman Limited.
2 / 1
![Page 3: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/3.jpg)
What do we have so far?
I Point process ideas (intensities, K -functions).I Data: (x , y) event locations.I Where are the clusters? Use intensities.I How are events clusters? Use K -functions.
I Disease clustering with point data.
I Disease clustering with regional counts.
3 / 1
![Page 4: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/4.jpg)
What’s left?
I So we know how to describe and evaluate spatial patterns inhealth outcome data.
I What about linking patterns in health outcomes to patterns inexposures?
I With independent observations we know how to use linear andgeneralized linear models such as linear, Poisson, logisticregression.
I What happens with dependent observations?
4 / 1
![Page 5: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/5.jpg)
Caveat
“...all models are wrong. The practical question is howwrong do they have to be to not be useful.”Box and Draper (1987, p. 74)
5 / 1
![Page 6: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/6.jpg)
What changes with dependence?
I In statistical modeling, we are often trying to describe themean of the outcome as a function of covariates, assumingerror terms are mutually independent.
I That means we usually model any trend in the data as a trendin expectations.
I Allows estimation of covariate effects.
I With dependent error terms, observed trends may be due tocovariates, correlation, or both.
I May impact the identifiability of covariate effects.
I Could have different effects equally likely under differentcorrelation models.
6 / 1
![Page 7: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/7.jpg)
Residual correlation
I Where do correlated errors come from?
I Perhaps outcomes truly correlated (infectious disease).
I Perhaps we omitted an important variable that has spatialstructure itself.
I If temperature is important and we left it out of a modelapplied to the continental U.S., what would the residuals looklike?
7 / 1
![Page 8: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/8.jpg)
Residual maps important
I If high temperatures associated with high outcomes, we wouldunderfit in southern states (observations > model predictions⇒ positive residuals), and overfit in northern states(observations < model prediction ⇒ negative residuals).
I The “missing covariate” idea suggests that maps of residualsare important spatial diagnostics.
I Also, we may want to apply tests of clustering or to detectclusters to residuals.
I Moran’s I , LISAs.
8 / 1
![Page 9: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/9.jpg)
Our plan
I We will take the NY leukemia data and add some covariates.
I We will fit linear and Poisson regression models with variousspatial correlation structures and compare inferences.
I Remember, all of these models are wrong, but some may beuseful.
9 / 1
![Page 10: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/10.jpg)
Illustrating regression models
I New York leukemia data from Waller et al. (1994)
I 281 census tracts (1980 Census).
I 8 counties in central New York.
I 592 cases for 1978-1982.
I 1,057,673 people at risk.
10 / 1
![Page 11: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/11.jpg)
Crude Rates (per 100,000)
Central New York Census Tracts, 1980Leukemia rates 1978-1982
0.000000 - 0.000021
0.000022 - 0.000324
0.000325 - 0.000588
0.000589 - 0.001096
0.001097 - 0.006993
11 / 1
![Page 12: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/12.jpg)
Building the model
I Let Yi = count for region i .
I Let Ei = expected count for region i .
I xi ,TCE = inverse distance to TCE site.
I xi ,65 = percent over age 65 (census).
I xi ,home = percent who own own home (census).
I The model:
Yi = β0 + xi ,TCEβTCE + xi ,65β65 + xi ,homeβhome + εi .
12 / 1
![Page 13: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/13.jpg)
Assumptions for regression
I The error terms, εiind∼ N(0, σ2);
I The data have a constant variance, σ2;
I The data are uncorrelated (OLS) or have a specifiedparametric covariance structure (GLS);
13 / 1
![Page 14: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/14.jpg)
Y normally distributed?
Histogram
Incidence Proportions
Freq
uenc
y
0.000 0.004
050
100
150
●●
●●
●
●●
●
●
●
●
●
●
●
●
●●
●
●●
●
●●●
●●
●
●●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●●●●●
●●
●
●
●●
●
●●
●●●
●
● ●●
●●
●
●●
●●●
●
●●
●●
●
●
●
●●
●
●
●●●●
●●
●●
●●
●
●●●●
●
●
●●●
●
●
●
●
●
● ●
●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●●●
●●
●
●
●●●
●
●
●
●
●
●
●
●●
●
●●● ●
●
●
●
●
●
●
●
●
●●●
●
●
● ●●
●
●
●●● ●
●
●
●
●●
●
●
●●
●●●
●
●●
●
●●●
●●●
●●
●
●●●
●●
●●
●
●●
●●
●
●●●
●
●
●
●
●
●
●●
●●
●
●
●
●
●●
●●
●●●
● ●
●
●
●●●●
●
● ●●
●●●
●●
●
●
●●
●●
●●
●●
●
● ●
●
●●
●
−3 −1 0 1 2 30.00
00.
003
0.00
6
Normal Q−Q Plot
Theoretical Quantiles
Sam
ple
Qua
ntile
s
14 / 1
![Page 15: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/15.jpg)
Transformation?
Zi = log
(1000(Yi + 1)
ni
).
Transformed Incidence Proportions
Freq
uenc
y
−2 0 1 2 3 4 5
020
4060
●●
●●
●
●●●
●
●
●
●
●
●●
●●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●●
●
●●
●
●
●
●
●
●
●
●
●●
●●●
●
●●
●
●
●
●●
●
●●
●
●●
●
●
●●
●●
●
●
●
●●
●●
●●
●●
●
●
●
●●
●
●
●●
●
●●●
●●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●●
●
●
●
●
●●
●
●●●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●●
●
●●
●
●●
●●
●●
●
●
●●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●●
●
●
●
●
●
●
●●
●●
●
●●●
●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●●
●
●
●●
●
●
●
●●
●●
●
●
●
●
−3 −1 0 1 2 3
−20
12
34
Normal Q−Q Plot
Theoretical Quantiles
Sam
ple
Qua
ntile
s
15 / 1
![Page 16: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/16.jpg)
Outliers, where are the top 3?
16 / 1
![Page 17: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/17.jpg)
Scatterplots
●●
●●
●
●●●●●●
●
●
●●●●●
●●
●●
●●●●
●
●
●
●
●
●
●
●
●●
● ●
●
●●
●
● ●
●
●●
●
●
●
●
●●●
●●●
●●●
●
●
●
● ●
●
●●
●●●
●
●●●
●●
●
●
●
●●●●●
●
●●
●
●
●
●●
●
●
●●●
●●●
●●
●
●
●
●●●
●
●
●●●●●
●
●
●
●●
●
●●●
●
●
●●
●
●●
●
●
●
●
●●
●
●●●●●
●
●
●●●
●
●
●
●
●
●●●
●
●
●
●●●
●
●
●●
●
●●
●
●●●
●
●
●●
●●●●●
●
●●
●
●
●
●
●
●●●●
●●
●●
●●
●●●
●●●●●●●
●●●
●
●●
●
●●
●
●
●●●●●●
●
●
●
●
●●
●
●
●●
●
●
●●●●
●●●●
●
●●●●
●●
●
●●
●
●
●
●
●●
●●
●
●
● ●
●
●
●
●●
●●
●
●
●●
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5
−20
24
Inverse Distance
Tran
sfor
med
Cou
ntInverse Distance vs. Outcome
●●
●●
●
●●●●
●●
●
●
●●
●● ●
● ●
●●
●●●
●
●
●
●
●
●
●
●
●
●●
● ●
●
●●
●
● ●
●
●●
●
●
●
●
●●●
●●●
●● ●
●
●
●
● ●
●
●●
●● ●
●
●●
●
●●
●
●
●
● ●●●●
●
●●
●
●
●
●●
●
●
●●●
● ●●
●●
●
●
●
●●●
●
●
●●●●●
●
●
●
●●
●
●●●
●
●
●●
●
●●
●
●
●
●
●●
●
●●●●●
●
●
●●●
●
●
●
●
●
●●●
●
●
●
●●●
●
●
●●
●
●●
●
●●●
●
●
●●
●●●●●
●
●●
●
●
●
●
●
●●●●
●●
●●
●●
●●●
●●●●●●
●
●●●
●
●●
●
●●
●
●
●●●●●●
●
●
●
●
●●
●
●
●●
●
●
●●●
●
●●●●
●
●●●●
●●
●
●●
●
●
●
●
●●
●●
●
●
● ●
●
●
●
● ●
●●
●
●
●●
1 2 3 4 5 6
−20
24
Log(100*Inverse Distance)
Tran
sfor
med
Cou
nt
Log(100*Inverse Distance) vs. Outcome
●●
●●
●
● ●●
●●
●
●
●
●●●●●
●●
●●
●●
●●
●
●
●
●
●
●
●
●
●●
● ●
●
●●
●
●●
●
●●
●
●
●
●
●●●
●● ●
●●●
●
●
●
● ●
●
●●
●●●
●
●●
●
●●
●
●
●
●●●●
●●
●●
●
●
●
● ●
●
●
●● ●
●● ●
●●
●
●
●
● ●●
●
●
●●
●●
●
●
●
●
●●
●
●●
●
●
●
●●
●
●●
●
●
●
●
●●
●
●●●●
●
●
●
●●●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●●
●
●●●
●
●
●●
● ●●
●●
●
●●
●
●
●
●
●
●●● ●
●●
●●
●●
● ●●
●●
●●
●●
●
●●●
●
●●
●
●●
●
●
●● ●●
●●
●
●
●
●
●●
●
●
●●
●
●
●●
●●
●●●●
●
●●●●●
●
●
●●
●
●
●
●
●●
●●●
●
● ●
●
●
●
● ●
●●
●
●
●●
0.0 0.1 0.2 0.3 0.4 0.5
−20
24
% > 65 years
Tran
sfor
med
Cou
nt
Percent > 65 vs. Outcome
●●
●●
●
● ●●
●●
●
●
●
●●●●●
● ●
●●
●●●
●
●
●
●
●
●
●
●
●
●●
●●
●
●●
●
●●
●
●●
●
●
●
●
●●●
●● ●
●● ●
●
●
●
●●
●
●●
●● ●
●
●●
●
●●
●
●
●
●●● ●
●●
●●
●
●
●
● ●
●
●
●●●
● ●●
● ●
●
●
●
● ●●
●
●
●●
●●
●
●
●
●
●●
●
●●
●
●
●
●●
●
●●
●
●
●
●
●●
●
●● ●●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●●
●
●●
●
●
●
●●
●●●
● ●
●
●●
●
●
●
●
●
●●●●
●●
●●
● ●
●●●
●●
●●
●●
●
● ●●
●
●●
●
●●
●
●
●● ● ●
●●
●
●
●
●
●●
●
●
●●
●
●
●●
●●
●● ●●
●
●●
●●
●●
●
●●
●
●
●
●
●●
●●
●
●
●●
●
●
●
●●
●●
●
●
●●
0.0 0.2 0.4 0.6 0.8 1.0
−20
24
% Own Home
Tran
sfor
med
Cou
nt
Percent Own Home vs. Outcome
17 / 1
![Page 18: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/18.jpg)
Linear Regression (OLS)
Parameter Estimate Std. Error p-value
β0 (Intercept) -0.5173 0.1586 0.0012
β1 (TCE) 0.0488 0.0351 0.1648
β2 (% Age > 65) 3.9509 0.6055 <0.0001
β3 (% Own home) -0.5600 0.1703 0.0011
σ2 0.4318 277 df
R2=0.1932 AIC=567.5
18 / 1
![Page 19: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/19.jpg)
Is OLS appropriate?
I Z s roughly Gaussian (symmetric).
I Do Z s have constant variance?
I No, since population sizes vary.
I Var(Zi ) = Var(
log(1000(Yi+1)
ni
))I Try weighted least squares with weights 1/ni .
19 / 1
![Page 20: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/20.jpg)
Linear Regression (WLS)
Parameter Estimate Std. Error p-value
β0 (Intercept) -0.7784 0.1412 <0.0001
β1 (TCE) 0.0763 0.0273 0.0056
β2 (% Age > 65) 3.8566 0.5713 <0.0001
β3 (% Own home) -0.3987 0.1531 0.0097
σ2 1121.94 277 df
R2=0.1977 AIC=513.5
20 / 1
![Page 21: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/21.jpg)
What changed?
I The three outliers are all in regions with small ni .
I Weighting reduced their impact on estimates.
I Most profound effect is with respect to TCE.
21 / 1
![Page 22: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/22.jpg)
WLS fitted values
22 / 1
![Page 23: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/23.jpg)
Residual plot
-1.50 -0.75 0.00 0.75 1.50fit
-4
-2
0
2
4st
uden
tized
resid
uals
23 / 1
![Page 24: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/24.jpg)
Residual map
24 / 1
![Page 25: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/25.jpg)
What are we looking for?
I Patterns in locations of residuals.
I Model underfit (predictions too low) near cities?
I Correlations in residuals?
I Let’s try semivariograms for the residuals.
I Let’s try local Moran’s I for residuals.
25 / 1
![Page 26: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/26.jpg)
Residual correlation? (Tip your head to the right.)
020
4060
80
distance
0.0 0.2 0.4 0.6 0.8 1.0 1.2gamma
26 / 1
![Page 27: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/27.jpg)
Comments
I Residual semivariogram not too impressive.
I We can try maximum likelihood fit incorporating residualcorrelation via the semivariogram (which defines covariancematrix).
27 / 1
![Page 28: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/28.jpg)
Linear Regression, Correlated Errors (ML)
Parameter Estimate Std. Error p-value
β0 (Intercept) -0.7222 0.1972 <0.0001
β1 (TCE) 0.0826 0.0434 0.0576
β2 (% Age > 65) 3.7093 0.6188 <0.0001
β3 (% Own home) -0.3245 0.2044 0.1136
c0=0.3740 cs=0.0558 a=6.93
AIC=565.6 277 df
28 / 1
![Page 29: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/29.jpg)
Weighting?
I We also need to include weights to account forheteroskedasticity.
I Again we use weights equal to 1/ni .
I What changes?
29 / 1
![Page 30: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/30.jpg)
Linear regression, Correlated, Weighted
Parameter Estimate Std. Error p-value
β0 (Intercept) -0.9161 0.1648 <0.0001
β1 (TCE) 0.0956 0.0322 0.0032
β2 (% Age > 65) 3.5763 0.5920 <0.0001
β3 (% Own home) -0.2285 0.1761 0.1956
c0=997.65 cs=127.12 a=6.86
AIC=514.7 277 df
30 / 1
![Page 31: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/31.jpg)
Fitted values (correlated, weighted)
31 / 1
![Page 32: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/32.jpg)
Modelling counts directly
I Using linear regression required a fair amount of datatransformation, just to meet modelling assumptions.
I Can we model the counts directly?
I In epidemiology, common to use logistic or Poisson regression.
I For rare disease, little difference between logistic and Poisson.
I Both are examples of generalized linear models (McCullaghand Nelder, 1989).
32 / 1
![Page 33: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/33.jpg)
Building the model
I Let Yi = count for region i .
I Let Ei = expected count for region i .
I Let (xi ,TCE , xi ,65, xi ,home) be the associated covariate values.
I Poisson regression:
Yi ∼ Poisson(Eiζi )
where
log(ζi ) = β0 + xi ,TCEβTCE + xi ,65β65 + xi ,homeβhome .
33 / 1
![Page 34: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/34.jpg)
What’s different?
I Poisson distribution for counts, rather than transformingproportions for normality.
I Link function: Natural log of mean of Yi is a linear functionof covariates.
I So βs represent multiplicative increases in expected counts, eβ
a measure of relative risk associated with one unit increase incovariate.
I Ei an offset, what we expect if the covariates have no impact.
I Age, race, sex adjustments in either Ei (standardization) orcovariates.
34 / 1
![Page 35: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/35.jpg)
How do we add spatial correlation?
I Trickier than in regression, since mean and variance arerelated for Poisson observations.
I Two general approaches:I Marginal specification defining correlation among means.I Conditional specification defining correlation through the use
of random effects.
35 / 1
![Page 36: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/36.jpg)
Marginal and conditional models
I We often think of a model representing the marginal mean,E (Y) as a function of fixed, unknown parameters.
I That is, the parameters define the population average effectof the covariates (“On average, how does a given level of airpollution impact a person?”)
I Another approach is to consider a model of the conditionalmean for each subject.
I In this setting we think of fixed effects of parameters andrandom effects specific to the subjects.
36 / 1
![Page 37: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/37.jpg)
Marginal versus conditional interpretation
I For us: fixed effects apply equally to all subjects, randomeffects apply to a particular subject.
I Interpret fixed effects conditional on levels of the randomeffects.
I “What is the effect of aspirin on a headache averaged over allindividuals in the study?” (Marginal effect).
I “What is the effect of aspirin on a headache in thisindividual?” (Conditional effect).
I Random effects allow different parameter values forindividuals, following some distribution.
37 / 1
![Page 38: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/38.jpg)
Random intercepts
I A model with fixed and random effects is a mixed model.
I A very common formulation is to have fixed parameter valuesand a random intercept. This says everyone has the sameresponse to the treatment, but that individuals have differentstarting points.
I In Poisson regression setting, if we add random effects wegenerate a generalized linear mixed model (GLMM).
38 / 1
![Page 39: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/39.jpg)
Random effects and the conditional specification
I We add a random effect (intercept).
I Represents an impact of region i , not accounted for in Ei orthe covariates.
I We define this random effect to have a spatial distribution.
39 / 1
![Page 40: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/40.jpg)
Building the model
I Let Yi denote the observed number of cases in region i .
I Let Ei denote the expected number of cases, ignoringcovariate effects.
I Assume Ei known, perhaps age-standardized, or based onglobal (external or internal) rates.
I First stage:
Yi |ζiind∼ Poisson(Eiζi )
I ζi represent a relative risk associated with region i notaccounted for by the Ei .
40 / 1
![Page 41: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/41.jpg)
Building the model
I Note Yi/Ei = SMRi , the MLE of ζi .
I Also note, E [Yi |ζi ] 6= Ei , since Ei does not include the impactof the random effect.
I Create a GLMM with log link by
log(E [Yi |ζi ]) = log(Ei ) + log(ζi )
I If we add covariates and rename log(ζi ) = ψi , then
log(ζi ) = x′iβ + ψi
41 / 1
![Page 42: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/42.jpg)
New York data
I So our model is
Yi |β, ψiind∼ Poisson(Ei exp(x′iβ + ψi )),
log(ζi ) = β0 + xi ,TCEβTCE + xi ,65β65 + xi ,homeβhome + ψi .
I The ψi represent the random interecpts.
I Add overdispersion via ψiind∼ N(0, vψ).
I Add spatial correlation via
ψ ∼ MVN(0,Σ).
42 / 1
![Page 43: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/43.jpg)
Priors and “shrinkage”
I Overdispersion model (i.i.d. ψi ) results in each estimate beinga compromise between the local SMR and the global averageSMR.
I “Borrows information (strength)” from other observations toimprove precision of local estimate.
I “Shrinks” estimate toward global mean. (Note: “shrink” doesnot mean “reduce”, rather means “moves toward”).
43 / 1
![Page 44: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/44.jpg)
Local shrinkage
I Spatial model (correlated ψi ) results in each estimate begin acompromise between the lcoal SMR and the local averageSMR.
I Shrinks each ψi toward the average of its neighbors.
I Can also include both global and local shrinkage (Besag, York,and Mollie 1991).
I How do we fit these models?
44 / 1
![Page 45: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/45.jpg)
Bayesian inference
Bayesian inference regarding model parameters based on posteriordistribution
Pr [β,ψ|Y]
proportional to the product of the likelihood times the prior
Pr [Y|β,ψ]Pr [ψ]Pr [β].
Defers spatial correlation to the prior rather than the likelihood.
45 / 1
![Page 46: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/46.jpg)
Spatial priors
I Could model joint distribution
ψ ∼ MVN(0,Σ).
I Could also model conditional distribution
ψi |ψj 6=i ∼ N
(∑j 6=i cijψj∑
j 6=i cij,
1
vCAR∑
j 6=i cij
), i = 1, . . . ,N.
where cij are weights defining the neighbors of region i .
I Adjacency weights: cij = 1 if j is a neighbor of i .
46 / 1
![Page 47: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/47.jpg)
CAR priors
I The conditional specification defines the conditionalautoregressive (CAR) prior (Besag 1974, Besag et al. 1991).
I Under certain conditions on the cij , the CAR prior defines avalid multivariate joint Gaussian distribution.
I Variance covariance matrix a function of the inverse of thematrix of neighbor weights.
47 / 1
![Page 48: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/48.jpg)
Perspective: Generalized linear mixed model
I Given the values of the random effects (ψi s), observations(Yi s) are independent.
I Taking into account correlation in the ψi s, the Yi s arecorrelated.
I Conditionally independent Yi |ψi give likelihood function.
I (Spatially correlated) distribution of the ψi s a priordistribution.
48 / 1
![Page 49: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/49.jpg)
Fitting Bayesian models: Markov chain Monte Carlo
I Posterior often difficult to calculate mathematically.
I Iterative simulation approach to model fitting.
I Given full conditional distributions, simulate a new value foreach parameter, holding the other parameter values fixed.
I The set of simulated values converges to a sample from theposterior distribution.
I WinBUGS software.www.mrc-bsu.cam.ac.uk/bugs/welcome.shtml
49 / 1
![Page 50: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/50.jpg)
Conceptual MCMC example
I Suppose we have a model with data Y and three parametersθ1, θ2, and θ3.
I “Gibbs sampler” simulates values from the full conditionaldistributions
f (θ1|θ2, θ3,Y),
f (θ2|θ1, θ3,Y),
f (θ3|θ1, θ2,Y).
50 / 1
![Page 51: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/51.jpg)
Conceptual MCMC
I Start with values θ(1)1 , θ
(1)2 , and θ
(1)3 .
sample θ(2)1 from f (θ1|θ(1)2 , θ
(1)3 ,Y),
sample θ(2)2 from f (θ2|θ(2)1 , θ
(1)3 ,Y),
sample θ(2)3 from f (θ3|θ(2)1 , θ
(2)2 ,Y).
I As we continue to update θ, sampled values becomeindistinguishable from a sample from the joint posteriordistribution f (θ1, θ2, θ3|Y).
51 / 1
![Page 52: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/52.jpg)
MCMC example
I Gelman et al. (2004). Theoretical and MCMC results.[Y1
Y2
]∼ MVN
([θ1θ2
],
[1 ρρ 1
]).
I Uniform priors on θ1, θ2, yield posterior[θ1θ2
]∼ MVN
([Y1
Y2
],
[1 ρρ 1
]).
52 / 1
![Page 53: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/53.jpg)
Full conditionals
I Multivariate results give full conditionals
θ1|θ2,Y ∼ N(Y1 + ρ(θ2 − Y2), 1− ρ2),
θ2|θ1,Y ∼ N(Y2 + ρ(θ1 − Y1), 1− ρ2).
I Let’s try a Gibbs sampler and compare to the theoreticalresults.
53 / 1
![Page 54: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/54.jpg)
MCMC example
●
−4 −2 0 2 4
−4−2
02
4
theta1
theta2
First 10 iterations
●●●
●●
● ●
●●
●●●●●●
● ●● ●
●●
●
●
●●●
●●
−4 −2 0 2 4
−4−2
02
4
theta1
theta2
500 iterations
●
● ●
●●●●●●●●
●●
●●
●●
● ●●●
●●
●●● ●
● ●
●●
● ●
●●●●●●
●●
● ●●●
● ●
●●● ●●●
● ●●●
● ●
●●●●
● ●
●●
●●●●
● ●●●
● ●
●●● ●
●●
●●
●●● ●
●●
●●● ●
●●●●
● ●●●
●●
● ●
●●●●
●●●●●●●●●●
● ●●●
●●
● ●
●●●●●●●●
● ●
●●●●
● ●
●●●●● ●
●●●●
● ●
●●●●
●●
● ●●●
●●
● ●● ●
●●
●●●●● ●
●●
●●
● ●●●
● ●●●
●●●●
● ●
●●
● ●
● ●
●●
●●● ●● ●●●
● ●
●●●●
● ●
●●
●●
●●● ●
●●●●
● ●
●●
●●●●
●●
● ●●●
●●
●●
●●
● ●
●●
●●
●●
●●● ●●●
● ●●●●●
● ●
●●
● ●
●●
● ●
●●
●●● ●
●●●●●●
●●
● ●●●
●●
●●
●●
●●
● ●●●●●
● ●
●●● ●●●
● ●
●●
●●●●●●
●●●●
●●
● ●
●●●●
● ●● ●
●●
●●● ●
● ●
●●
●●
● ●
●●●●●●● ●
● ●
●●● ●
●●
● ●●●●●
●●●●●●●●
● ●●●
●●
● ●
●●
● ●●●
●●
● ●
●●
● ●
●●●●
● ●
●●
● ●●●
●●● ●
● ●
●●
●●●●
●●
● ●
●●
●●
● ●●●●●
●●●●● ●
●●● ●
●●
●●●●
● ●
●●
●●●●● ●●●
●●
●●
● ●
● ●
●●
● ●●●
● ●
●●● ●●●
● ●
●●● ●
●●
●●●●
●●
●●
● ●
● ●
●●● ●
●●●●●●
●●
● ●●●●●
● ●●●
● ●
●●
●●
● ●
● ●●●
●●
●●
● ●
●●
●●
● ●
●●
● ●
●●
● ●
●●●●
●●● ●
● ●
● ●
●●
●●● ●
●●
●●
●●
●●
●●●●
● ●●●
● ●
●●●●
●●
●●●●
● ●
●●●●
●●
● ●●●
● ●
●●● ●●●● ●
●●
●●
● ●
●●●●
● ●
●●
●●
● ●●●
●●
●●
● ●●●●●●●
●●
● ●
●●
●●
● ●●●●●
● ●
●●●●
● ●●●● ●
●●
●●●●
●●
●●●●
● ●●●
●●
●●
● ●●●● ●
● ●●●
●●
● ●●●
● ●●●
●●● ●
●●●●
●●●●
● ●
● ●
●●●●●●
●●● ●
● ●●●
● ●●●
●●● ●
●●
● ●
●●●●●●
●●
●●
●●
● ●
● ●●●
●●
●●● ●
●●
●●
● ●
●●
●●
● ●●●
● ●● ●
●●
●●●●
● ●●●●●
● ●●●●●
● ●●●
● ●●●
● ●
●●
●●
●●
● ●
●●●●
● ●
●●
●●● ●
● ●●●
● ●●●
● ●●●
● ●
●●●●
●●●●
● ●
●●
●●
● ●● ●
●●
●●
● ●●●
● ●
●●
● ●●●
● ●●●
●●
● ●
●●●●● ●
● ●
●●
●●
●●● ●●●● ●●●
●●
● ●
●●
● ●
●●● ●●●
●●● ●
● ●
●●
●●
●
Theta1
Density
−4 −2 0 2 4
0.0
0.2
0.4
Theta2
Density
−4 −2 0 2 40.0
0.2
0.4
54 / 1
![Page 55: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/55.jpg)
Back to CAR prior
I Almost custom-made for MCMC.
I Defined for ψi , given ψj for j 6= i .
I We define neighborhood weights cij .
55 / 1
![Page 56: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/56.jpg)
Complete model specification
Yi |β, ψiind∼ Poisson(Ei exp(x′iβ + ψi )),
log(ζi ) = β0 + xi ,TCEβTCE + xi ,65β65 + xi ,homeβhome + ψi .
βk ∼ Uniform.
ψi |ψj 6=i ∼ N
(∑j 6=i cijψj∑
j 6=i cij,
1
vCAR∑
j 6=i cij
), i = 1, . . . ,N.
1
vCAR∼ Gamma(0.5, 0.0005).
56 / 1
![Page 57: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/57.jpg)
MCMC trace plots
0 2000 4000 6000 8000
−6−4
−20
Iteration
Inte
rcep
t
0 2000 4000 6000 8000
26
1014
Iteration
Age
Effe
ct
0 2000 4000 6000 8000
−1.5
−0.5
Iteration
Expo
sure
Effe
ct
0 2000 4000 6000 8000
−11
23
Iteration
Hom
e O
wner
ship
Effe
ct
57 / 1
![Page 58: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/58.jpg)
Posterior densities
−0.5 0.0 0.5
0.0
1.0
2.0
Intercept
Dens
ity
1 2 3 4 5 6
0.0
0.2
0.4
0.6
Age Effect
Dens
ity
0.00 0.10 0.20 0.30
02
46
8
Exposure Effect
Dens
ity
−1.0 −0.5 0.0 0.5
0.0
1.0
2.0
Home Ownership Effect
Dens
ity
58 / 1
![Page 59: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/59.jpg)
MCMC posterior estimates
Covariate Posterior 95% CredibleMedian Set
β0 0.048 (-0.355, 0.408)β65 3.984 (2.736, 5.330)βTCE 0.152 (0.066, 0.226)βhome -0.367 (-0.758, 0.049)
59 / 1
![Page 60: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/60.jpg)
But there’s more!
I A nifty thing about MCMC estimates:
We get posterior samples from any function of modelparameters by taking that function of the sampled posteriorparameter values.
I Gives us posterior inference for SMRi = Yi ,fit/Ei .
I Also can get Pr [SMRi > 200|Y] and map these exceedenceprobabilities.
60 / 1
![Page 61: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/61.jpg)
Posterior median SMRs
61 / 1
![Page 62: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/62.jpg)
Posterior exceedence probabilities
62 / 1
![Page 63: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/63.jpg)
Example 2
I Cryptozoology Example: Waller and Carlin (2010) DiseaseMapping. In Handbook of Spatial Statistics, Gelfand et al.(eds.). Boca Raton: CRC/Chapman and Hall.
63 / 1
![Page 64: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/64.jpg)
Cryptozoology example
I County-specific reports of encounters with Sasquatch(Bigfoot).
I “...which brings us to the appropriateness of the Bigfootexample.”
I Data downloaded from www.bfro.net
I Sightings from counties in Oregon and Washington (PacificNorthwest).
I Probability of report related to population density?
I (Hopefully) rare events in small areas.
I Perhaps spatial smoothing will stabilize local rate estimates.
I Fit models with no random effects, exchangeable randomeffects, CAR random effects, convolution random effects.
64 / 1
![Page 65: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/65.jpg)
Sasquatch Data
Wasco
Number of Reports Reports per 2000 Population
2000 Population per Square Mile
0 200 400 600 800100
Kilometers
Legend
Population/Sq. Mi.Reports
0.7 - 12.9
13.0 - 32.1
32.2 - 69.9
70.0 - 180.1
180.2 - 414.0
414.1 - 793.3
793.4 - 1419.3
Reports/Person
0.00000 - 0.00003
0.00003 - 0.00008
0.00008 - 0.00016
0.00016 - 0.00026
0.00026 - 0.00046
0.00046 - 0.00076
0.00076 - 0.00517
Skamania
King
0
1-5
6-10
11-15
16-20
21-25
25-51
65 / 1
![Page 66: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/66.jpg)
Reports vs. Population Density
02
46
0.000 0.001 0.002 0.003 0.004 0.005
log(Census 2000 population density)
Reports per Census 2000 population size
Log of Population Density
02
46
0 e+00 2 e−04 4 e−04 6 e−04 8 e−04 1 e−03
log(Census 2000 population density)
Reports per Census 2000 population size
Ignoring outlier
66 / 1
![Page 67: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/67.jpg)
Observed vs. Expected
0 20 40 60 80 100 120
010
2030
4050
Expected (constant per person risk of reporting)
Obs
erve
dObserved versus Expected Number of Reports
Skamania
King
67 / 1
![Page 68: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/68.jpg)
Predicted relative risks and credible sets
Filled circle = Skamania, Filled square = Wasco
0 5 10 15 20 25 30 35
−2−1
01
23
4
log(
rela
tive
risk)
No Random Effects Exchangeable CAR Convolution
68 / 1
![Page 69: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/69.jpg)
Mapped relative risksNo random effect RRs Exchangeable RRs
CAR RRs Convolution RRs
Legend 0.00 - 1.00
1.01 - 2.00
2.01 - 3.00
3.01 - 4.00
4.01 - 15.00
approxim ately 70.00
69 / 1
![Page 70: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/70.jpg)
Skamania Sasquatch Ordinances
I http://www.skamaniacounty.org/commissioners/
homepage/ordinances-2/
I Big Foot Ordinance 69-1: “THEREFORE BE IT RESOLVEDthat any premeditated, willful and wonton slaying of any suchcreature shall be deemed a felony punishable by a fine not toexceed Ten Thousand Dollars ($10,000.00) and/orimprisonment in the county jail for a period not to exceedFive (5) years. ADOPTED this 1st day of April, 1969.”
I Big Foot Ordinance 1984-2:I Repealed felony and jail sentence.I Established a Sasquatch Refuge (Skamania County).I Clarified penalty (gross misdemeanor vs. misdemeanor) and
penalty (fine and jail time), disallowed insanity defense, andclarified distinction between coroner designation of victim ashumanoid (murder) or anthropoid (this ordinance).
70 / 1
![Page 71: MODULE 16: Spatial Statistics in Epidemiology and Public ...€¦ · References I Waller and Gotway (2004, Chapter 9) Applied Spatial Statistics for Public Health Data. New York:](https://reader036.fdocuments.net/reader036/viewer/2022070716/5eda75fdb3745412b5716132/html5/thumbnails/71.jpg)
Conclusions
I What method to use depends on what you want data youhave and what question you want to answer.
I All methods try to balance trend (fixed effects) withcorrelation (here, with random effects).
I All models wrong, some models useful.
I Trying more than one approach often sensible.
I Few methods (including Monte Carlo simulation) in currentGIS packages.
71 / 1