Lecture 20 - Spatial Random Effects Models + Point Reference...
Transcript of Lecture 20 - Spatial Random Effects Models + Point Reference...
Lecture 20Spatial Random Effects Models + Point Reference Spatial Data
Colin Rundel04/03/2017
1
Spatial Random Effects Models
2
Scottish Lip Cancer Data
Observed Expected
8°W 6°W 4°W 2°W 8°W 6°W 4°W 2°W
55°N
56°N
57°N
58°N
59°N
60°N
61°N
0
20
40
60
80
value
3
55°N
56°N
57°N
58°N
59°N
60°N
61°N
8°W 6°W 4°W 2°W
0
2
4
6
Obs/Exp
55°N
56°N
57°N
58°N
59°N
60°N
61°N
8°W 6°W 4°W 2°W
0
5
10
15
20
% Agg Fish Forest
4
Neighborhood / weight matrix
5
Moran’s I
morans_I = function(y, w){n = length(y)y_bar = mean(y)num = sum(w * (y-y_bar) %*% t(y-y_bar))denom = sum( (y-y_bar)^2 )(n/sum(w)) * (num/denom)
}
morans_I(y = lip_cancer$Observed, w = diag(rowSums(W)) %*% W)## [1] 0.2258758
morans_I(y = lip_cancer$Observed / lip_cancer$Expected,w = diag(rowSums(W)) %*% W)
## [1] 0.5323161
ape::Moran.I(lip_cancer$Observed / lip_cancer$Expected,weight = diag(rowSums(W)) %*% W) %>% str()
## List of 4## $ observed: num 0.666## $ expected: num -0.0182## $ sd : num 0.0784## $ p.value : num 0
6
A hierachical model for lip cancer
We have observed counts of lip cancer for 56 districts in Scotland. Let yirepresent the number of lip cancer for district i.
yi ∼ Poisson(λi)
log(λi) = log(Ei) + xiβ + ωi
ω ∼ N (0, σ2(D − ϕW)−1)
where Ei is the expected counts for each region (and serves as an offet).
7
Data prep & JAGS model
D = diag(rowSums(W))X = model.matrix(~scale(lip_cancer$pcaff))log_offset = log(lip_cancer$Expected)y = lip_cancer$Observed
## model{## for(i in 1:length(y)) {## y[i] ~ dpois(lambda[i])## y_pred[i] ~ dpois(lambda[i])## log(lambda[i]) <- log_offset[i] + X[i,] %*% beta + omega[i]## }#### for(i in 1:2) {## beta[i] ~ dnorm(0,1)## }#### omega ~ dmnorm(rep(0,length(y)), tau * (D - phi*W))## sigma2 <- 1/tau## tau ~ dgamma(2, 2)## phi ~ dunif(0,0.99)## }
8
Model Results
35000 40000 45000 50000 55000 60000
−0.
40.
2
Iterations
Trace of beta[1]
−0.5 0.0 0.5
0.0
1.0
2.0
Density of beta[1]
N = 1000 Bandwidth = 0.04924
35000 40000 45000 50000 55000 60000
0.0
0.2
0.4
Iterations
Trace of beta[2]
−0.1 0.0 0.1 0.2 0.3 0.4 0.5
01
23
Density of beta[2]
N = 1000 Bandwidth = 0.02561
9
35000 40000 45000 50000 55000 60000
0.4
0.7
1.0
Iterations
Trace of phi
0.4 0.5 0.6 0.7 0.8 0.9 1.0
04
812
Density of phi
N = 1000 Bandwidth = 0.01076
35000 40000 45000 50000 55000 60000
0.5
1.5
Iterations
Trace of sigma2
0.5 1.0 1.5 2.00.
01.
0
Density of sigma2
N = 1000 Bandwidth = 0.06232
10
Predictions & Residuals
55°N
56°N
57°N
58°N
59°N
60°N
61°N
8°W 6°W 4°W 2°W
10
20
30
Predicted Cases
55°N
56°N
57°N
58°N
59°N
60°N
61°N
8°W 6°W 4°W 2°W
−2
0
2
Residuals
11
Residuals + RMSE + Moran’s I
#RMSElip_cancer_pred$resid %>% .^2 %>% mean() %>% sqrt()## [1] 1.498675
#Moran’s Imorans_I(y = lip_cancer_pred$resid, w = diag(rowSums(W)) %*% W)## [1] 0.05661104
ape::Moran.I(lip_cancer_pred$resid,weight = diag(rowSums(W)) %*% W) %>% str()
## List of 4## $ observed: num 0.0642## $ expected: num -0.0182## $ sd : num 0.0803## $ p.value : num 0.305
12
Point Referenced Data
13
Example - PM2.5 from CSN
The Chemical Speciation Network are a series of air quality monitors run byEPA (221 locations in 2007). We’ll look at a subset of the data from Nov 11th,2007 (n=191) for just PM2.5.
25
30
35
40
45
50
−120 −100 −80
long
lat
pm25
10
20
30
40
10
20
30
40
pm25
14
csn## # A tibble: 191 × 5## site longitude latitude date pm25## <int> <dbl> <dbl> <dttm> <dbl>## 1 10730023 -86.81500 33.55306 2007-11-14 19.43555## 2 10732003 -86.92417 33.49972 2007-11-14 26.40000## 3 10890014 -86.58637 34.68767 2007-11-14 13.40000## 4 11011002 -86.25637 32.40712 2007-11-14 19.70000## 5 11130001 -84.99917 32.47639 2007-11-14 22.60000## 6 40139997 -112.09577 33.50383 2007-11-14 12.30000## 7 40191028 -110.98230 32.29515 2007-11-14 7.20000## 8 51190007 -92.28130 34.75619 2007-11-14 12.70000## 9 60070002 -121.84222 39.75750 2007-11-14 10.00000## 10 60190008 -119.77222 36.78139 2007-11-14 32.26205## # ... with 181 more rows
15
Aside - Splines
16
Splines in 1d - Smoothing Splines
These are a mathematical analogue to the drafting splines representedusing a penalized regression model.
We want to find a function f(x) that best fits our observed datay = y1, . . . , yn while being as smooth as possible.
arg minf(x)
n∑i=1
(yi − f(xi))2 + λ
∫f′′(x)2 dx
Interestingly, this minimization problem has an exact solution which is givenby a mixture of weighted natural cubic splines (cubic splines that are linearin the tails) with knots at the observed data locations (xs).
17
Splines in 1d - Smoothing Splines
These are a mathematical analogue to the drafting splines representedusing a penalized regression model.
We want to find a function f(x) that best fits our observed datay = y1, . . . , yn while being as smooth as possible.
arg minf(x)
n∑i=1
(yi − f(xi))2 + λ
∫f′′(x)2 dx
Interestingly, this minimization problem has an exact solution which is givenby a mixture of weighted natural cubic splines (cubic splines that are linearin the tails) with knots at the observed data locations (xs).
17
Splines in 2d - Thin Plate Splines
Now imagine we have observed data of the form (xi, yi, zi) where we wish topredict zi given xi and yi for all i. We can naturally extend the smoothingspline model in two dimensions,
arg minf(x,y)
n∑i=1
(zi − f(xi, yi))2 + λ
∫ ∫ (∂2f∂x2
+ 2∂2f
∂x ∂y+
∂2f∂y2
)dx dy
The solution to this equation has a natural representation using a weightedsum of radial basis functions with knots at the observed data locations
f(x, y) =n∑i=1
wi d(xi, yi)2 log d(xi, yi).
18
Fitting a TPS
library(fields)coords = select(csn, longitude, latitude) %>% as.matrix()tps = Tps(x = coords, Y=csn$pm25)
pm25_pred = rpm25_pred[cells] = predict(tps, pred_coords)
plot(pm25_pred)points(coords, pch=16, cex=0.5)
−120 −110 −100 −90 −80 −70
2530
3540
4550
5101520253035
19
Gaussin Process Models / Kriging
20
Variogram
library(geoR)coords = csn %>% select(latitude, longitude) %>% as.matrix()d = dist(coords) %>% as.matrix()
variog(coords = coords, data = csn$pm25, messages = FALSE,uvec = seq(0, max(d)/2, length.out=50)) %>% plot()
0 5 10 15 20 25
020
6010
0
distance
sem
ivar
ianc
e
21
variog(coords = coords, data = csn$pm25, messages = FALSE,uvec = seq(0, max(d)/4, length.out=50)) %>% plot()
0 2 4 6 8 10 12 14
020
4060
80
distance
sem
ivar
ianc
e
22
Isotropy / Anisotropy
v4 = variog4(coords = coords, data = csn$pm25, messages = FALSE,uvec = seq(0, max(d)/4, length.out = 50))
plot(v4)
0 2 4 6 8 10 12 14
020
4060
8010
012
0
distance
sem
ivar
ianc
e
0°45°90°135°
23
GP Spatial Model
If we assume that our data is stationary and isotropic then we can use a GaussianProcess model to fit the data. We will assume an exponential covariance structure.
y ∼ N (µ1, Σ)
{Σ}ij = σ2 exp(−r ∥si − sj∥) + σ2n 1i=j
we can also view this as a spatial random effects model where
y(s) = µ(s) + w(s) + ϵ(s)
w(s) ∼ N (0,Σ′)
ϵ(si) ∼ N (0, σ2n)
{Σ′}ij = σ2 exp(−r ∥si − sj∥)
24
GP Spatial Model
If we assume that our data is stationary and isotropic then we can use a GaussianProcess model to fit the data. We will assume an exponential covariance structure.
y ∼ N (µ1, Σ)
{Σ}ij = σ2 exp(−r ∥si − sj∥) + σ2n 1i=j
we can also view this as a spatial random effects model where
y(s) = µ(s) + w(s) + ϵ(s)
w(s) ∼ N (0,Σ′)
ϵ(si) ∼ N (0, σ2n)
{Σ′}ij = σ2 exp(−r ∥si − sj∥)
24
Fitting with spBayes
library(spBayes)
n = nrow(csn)n_samp = 20000coords = select(csn, longitude, latitude) %>% as.matrix()max_range = max(dist(coords)) / 4
starting = list(phi = 3/3, sigma.sq = 33, tau.sq = 17)tuning = list(”phi”=0.1, ”sigma.sq”=0.1, ”tau.sq”=0.1)priors = list(
beta.Norm = list(0, 1000),phi.Unif = c(3/max_range, 6),sigma.sq.IG = c(2, 2),tau.sq.IG = c(2, 2)
)
25
m = spLM(pm25 ~ 1, data = csn, coords = coords, starting = starting, priors = priors,cov.model = ”exponential”, n.samples = n_samp, tuning = tuning,n.report = n_samp/2)
## ----------------------------------------## General model description## ----------------------------------------## Model fit with 191 observations.#### Number of covariates 1 (including intercept if specified).#### Using the exponential spatial correlation model.#### Number of MCMC samples 20000.#### Priors and hyperpriors:## beta normal:## mu: 0.000## cov:## 1000.000#### sigma.sq IG hyperpriors shape=2.00000 and scale=2.00000## tau.sq IG hyperpriors shape=2.00000 and scale=2.00000## phi Unif hyperpriors a=0.21888 and b=6.00000## -------------------------------------------------## Sampling## -------------------------------------------------## Sampled: 10000 of 20000, 50.00%## Report interval Metrop. Acceptance rate: 33.25%## Overall Metrop. Acceptance rate: 33.25%## -------------------------------------------------## Sampled: 20000 of 20000, 100.00%## Report interval Metrop. Acceptance rate: 33.53%## Overall Metrop. Acceptance rate: 33.39%## -------------------------------------------------
26
m = spRecover(m, start=n_samp/2+1, thin = (n_samp/2)/1000)## -------------------------------------------------## Recovering beta and w## -------------------------------------------------## Sampled: 99 of 1000, 9.90%## Sampled: 199 of 1000, 19.90%## Sampled: 299 of 1000, 29.90%## Sampled: 399 of 1000, 39.90%## Sampled: 499 of 1000, 49.90%## Sampled: 599 of 1000, 59.90%## Sampled: 699 of 1000, 69.90%## Sampled: 799 of 1000, 79.90%## Sampled: 899 of 1000, 89.90%## Sampled: 999 of 1000, 99.90%
27
Parameter values
m$p.theta.recover.samples %>% mcmc() %>% plot()
0 200 400 600 800 1000
3070
Iterations
Trace of sigma.sq
20 40 60 80 100
0.00
0.03
Density of sigma.sq
N = 1000 Bandwidth = 2.766
0 200 400 600 800 1000
1020
Iterations
Trace of tau.sq
5 10 15 20 25
0.00
0.10
Density of tau.sq
N = 1000 Bandwidth = 0.802
0 200 400 600 800 1000
0.2
0.6
Iterations
Trace of phi
0.2 0.3 0.4 0.5 0.6 0.7 0.8
02
4Density of phi
N = 1000 Bandwidth = 0.02228
28
m$p.beta.recover.samples %>% mcmc() %>% plot()
0 200 400 600 800 1000
1015
20
Iterations
Trace of (Intercept)
5 10 15 20 250.
000.
050.
100.
150.
200.
25
Density of (Intercept)
N = 1000 Bandwidth = 0.4045
29
Predictions
m_pred = spPredict(m, pred_coords, pred.covars = matrix(1, nrow=nrow(pred_coords)),start=n_samp/2+1, thin=(n_samp/2)/1000)
## ----------------------------------------## General model description## ----------------------------------------## Model fit with 191 observations.#### Prediction at 900 locations.#### Number of covariates 1 (including intercept if specified).#### Using the exponential spatial correlation model.#### -------------------------------------------------## Sampling## -------------------------------------------------## Sampled: 100 of 1000, 9.90%## Sampled: 200 of 1000, 19.90%## Sampled: 300 of 1000, 29.90%## Sampled: 400 of 1000, 39.90%## Sampled: 500 of 1000, 49.90%## Sampled: 600 of 1000, 59.90%## Sampled: 700 of 1000, 69.90%## Sampled: 800 of 1000, 79.90%## Sampled: 900 of 1000, 89.90%## Sampled: 1000 of 1000, 99.90%m_pred_summary = post_summary(t(m_pred$p.y.predictive.samples))
30
splm_pm25_pred = rsplm_pm25_pred[cells] = m_pred_summary$post_mean
plot(splm_pm25_pred)points(coords, pch=16, cex=0.5)
−120 −110 −100 −90 −80 −70
2530
3540
4550
5101520253035
31
JAGs Model
## model{## for(i in 1:length(y)){## y[i] ~ dnorm(beta + w[i], tau)## mu_w[i] <- 0## }#### for(i in 1:length(y)){## for(j in 1:length(y)){## Sigma_w[i,j] <- sigma2_w * exp(-phi * d[i,j])## }## }## Sigma_w_inv <- inverse(Sigma_w)## w ~ dmnorm(mu_w, Sigma_w_inv)#### beta ~ dnorm(0, 1/1000)## sigma2_w ~ dgamma(2, 2)## sigma2 ~ dgamma(2, 2)## tau <- 1/sigma2## phi ~ dunif(3/14, 6)## }
32
10000 11000 12000 13000 14000 15000
1012
14
Iterations
Trace of beta
10 12 14 16
0.0
0.2
0.4
Density of beta
N = 1000 Bandwidth = 0.2477
10000 11000 12000 13000 14000 15000
0.3
0.5
0.7
Iterations
Trace of phi
0.2 0.3 0.4 0.5 0.6 0.7 0.80
24
Density of phi
N = 1000 Bandwidth = 0.02054
33
10000 11000 12000 13000 14000 15000
1015
20
Iterations
Trace of sigma2
5 10 15 20
0.00
0.10
Density of sigma2
N = 1000 Bandwidth = 0.5785
10000 11000 12000 13000 14000 15000
1525
Iterations
Trace of sigma2_w
10 15 20 25 300.
000.
08
Density of sigma2_w
N = 1000 Bandwidth = 0.7447
34