Transcript random effects modeling

Spatially
structured
random
effects
by
Daniel A. Griffith
Ashbel Smith Professor
of Geospatial Information
Sciences
ABSTRACT
Researchers increasingly are employing random effects
modeling to analyze data. When data are
georeferenced, a random effect term needs to be
spatially structured in order to account for spatial
autocorrelation. Spatial structuring can be achieved in
various ways, including the use of semivariogram,
spatial autoregressive, and spatial filter models. SAS
implements the semivariogram option for linear mixed
models. GeoBUGS implements the spatial
autoregressive option for either linear or generalized
linear mixed modeling. Recently developed spatial
filtering methodology can be used in either case, as well
as with the SAS generalized linear mix model procedure,
and furnishes one means of estimating space-time mixed
models. This presentation summarizes comparisons of
these three forms of spatial structuring, illustrating
implementations with selected ecological data for the
municipalities of Puerto Rico.
From Legendre
Spatial structures in communities indicate that some process
hasbeen at work to create them. Two families of mechanisms
cangenerate spatial structures in communities:
• Autocorrelation model: the spatial structures are generated
by thespecies assemblage themselves (response variables).
• Induced spatial dependence model: forcing (explanatory)
variablesare responsible for the spatial structures found in
the speciesassemblage. They represent environmental or
biotic control of thespecies assemblages, or historical
dynamics
To understand the mechanisms that generate these
structures, we needto explicitly incorporate the spatial
community structures, at all scales, into the statistical model.
Spatial autocorrelation (SA) is technically defined as the dependence, due to geographic proximity,
present in the residuals of a [regression-type] model of a response variable y whicht akes into account all
deterministic effects due to forcing variables.
Spatial autocorrelation can be
interpreted in different ways
As a spatial process mechanism – the cartoon
As a diagnostic tool – the Cliff-Ord Eire example (the model specification should be nonlinear)
As a nuisance parameter – eliminating spatial dependency to avoid statistical complications
As a spatial spillover effect – georeferencing of pediatric lead poisoning cases in Syracuse, NY
As an outcome of areal unit demarcation – the modifiable areal unit problem (MAUP)
As redundant information – spatial sampling; map interpolation
As map pattern – spatial filtering (to be discussed in this course)
As a missing variables indicator/surrogate – a possible implication of spatial filtering
As self-correlation – what is discussed next
The magic box is a physical model
of spatial autocorrelation
The permutation perspective
The SASIM
game
http://www.nku.edu/~longa/cgi-bin/cgi-tcl-examples/generic/SA/SA.cgi
Measures of spatial autocorrelation
MC: Moran Coefficient; GR: Geary Ratio; semivariogram
n
MC 
 (y
n
n
 c (y
ij
j
 y)
j1

c ij
(y i  y) 2
i 1
n

2
i 1 j1
c ij (y i  y j ) 2
i 1 j1
n
n
c ij

i 1
γ(d k ) 
nk

(y i  y j )
i 1
n

n -1
n
 y)
n
i 1 j1
GR 
i
i 1
n

n
(y i  y) 2
2*n k
2

f( d k ), k  1,2,..., K
Spherical
Exponential
Bessel function (1st order, 2nd kind)
Georeferenced data scatterplots
• The horizontal axis is the measurement
scale for some attribute variable
• The vertical axis is the measurement scale
for neighboring values (topological
distance-based) of the same attribute
variable
OR
• The horizontal axis is (usually) Euclidean
distance between geocoded locations
• The vertical axis is the measurement scale
for geographic variability
Describing a scatterplot trend
positive relationship:
High Y with High X
& Medium Y with Medium X
& Low Y with Low X
negative relationship:
High Y with Low X
& Medium Y with Medium X
& Low Y with High X
Description of the Moran scatterplot
Positive spatial autocorrelation
2002 population
- high values tend to be
density
surrounded by nearby high values
- intermediate values tend to be surrounded
by nearby intermediate values
- low values tend to be surrounded by
nearby low values
MC = 0.49
GR = 0.58
Description of the Moran scatterplot
Negative spatial autocorrelation competition for space
- high values tend to be
surrounded by nearby low values
- intermediate values tend to be surrounded
by nearby intermediate values
- low values tend to be surrounded by
nearby high values
MC = -0.16
sMC = 0.075
GR = 1.04
Graphical
portrayals of spatial
autocorrelation
latent in
transformed As
data
Constructing eigenfunctions for
filtering spatial autocorrelation
out of georeferenced variables:
Moran Coefficient = (n/1T C1)x
YT(I – 11T/n)C (I – 11T/n)Y/ YT(I – 11T/n)Y
the eigenfunctions come from
(I – 11T/n)C (I – 11T/n)
Random effects model
Y  f( Xβ  ξ, ε)
ξ is a random observation effect (differences
among individual observational units)
ε is a time-varying residual error (links to
change over time)
The composite error term is the sum of the
two.
Random effects model: normally
distributed intercept term
• ξ ~ N(0, σ ) and uncorrelated with covariates
• supports inference beyond the nonrandom
sample analyzed
• simplest is where intercept is allowed to vary
across areal units (repeated observations are
individual time series)
• The random effect variable is integrated out
(with numerical methods) of the likelihood fcn
• accounts for missing variables & within unit
correlation (commonality across time periods)
2
Spatial structuring of random effects
• CAR: conditional autoregressive model
• ICAR: improper conditional autoregressive
model (spatial autocorrelation set to 1,and
a spatially structured and a spatially
unstructured variance component is
estimated)─should be specified as a
convolution prior (spatially structured &
unstructured random effects)
• SF: spatial filter identified with a
frequentist GLM
Frequentist
Bayesian
Definition
of
probability
Long-run expected
Relative degree of
frequency in repeated
belief in the state of the
(actual or hypothetical) world
experiments (Law of LN)
Point
estimate
Maximum likelihood
estimate
Uncertainty
intervals
for
parameters
“confidence intervals”
“credible intervals”
based on the Likelihood based on the posterior
Ratio Test (LRT)
i.e., probability distribution
the expected probability
distribution of the
maximum likelihood
estimate over many
experiments
Mean, mode or median
of the posterior
probability distribution
Uncertainty
intervals of
nonparameters
Based on likelihood
profile/LRT, or by
resampling from the
sampling distribution of
the parameter
Calculated directly from
the distribution of
parameters
Model
selection
Discard terms that are not
significantly different from
a nested (null) model at a
previously set confidence
level
Retain terms in models,
on the argument that
processes are not absent
simply because they are
not statistically significant
Difficulties
Confidence intervals are
Subjectivity; need to
confusing (range that will specify priors
contain the true value in a
proportion α of repeated
experiments); rejection of
model terms for “nonsignificance”
Impact of sample size
•prior
•distribution
As the sample size
increases, a prior distribution has less and less
impact on results; BUT
•likelihood
•distribution
•effective
•sample size
•for spatially
•autocorrelated
•data
What is BUGS?
Bayesian inference Using Gibbs Sampling
• is a piece of computer software for the
Bayesian analysis of complex statistical
models using Markov chain Monte Carlo
(MCMC) methods.
• It grew from a statistical research project at
the MRC BIOSTATISTICAL UNIT in
Cambridge, but now is developed jointly
with the Imperial College School of
Medicine at St Mary’s, London.
•Classic BUGS
•BUGS
•WinBUGS (Windows Version)
•
GeoBUGS (spatial models)
•
PKBUGS (pharmokinetic modeling)
• The Classic BUGS program uses textbased model description and a commandline interface, and versions are available
for major computer platforms (e.g., Sparc,
Dos). However, it is not being further
developed.
What is WinBUGS?
• WinBUGS, a windows program with an option of
a graphical user interface, the standard ‘pointand-click’ windows interface, and on-line
monitoring and convergence diagnostics. It also
supports Batch-mode running (version 1.4).
• GeoBUGS, an add-on to WinBUGS that fits
spatial models and produces a range of maps
as output.
• PKBUGS, an efficient and user-friendly interface
for specifying complex population
pharmacokinetic and pharmacodynamic
(PK/PD) models within the WinBUGS software.
What is GeoBUGS?
• Available via
http://www.mrc-bsu.cam.ac.uk/
bugs/winbugs/geobugs.shtml
• Bayesian inference is used to spatially
smooth the standardized incidence ratios
using Markov chain Monte Carlo (MCMC)
methods. GeoBUGS implements models for
data that are collected within discrete regions
(not at the individual level), and smoothing is
done based on Markov random field models
for the neighborhood structure of the regions
relative to each other.
What is MCMC?
MCMC is used to simulate from some
distribution p known only up to a constant
factor, C:
pi = Cqi
where qi is known but C is unknown and too
horrible to calculate.
MCMC begins with conditional (marginal)
distributions, and MCMC sampling outputs a
sample of parameters drawn from their
posterior (joint) distribution.
The geographic distribution of
elevation across the island of
Puerto Rico
From a USGS DEM containing 87,358,136
points. Darkness of gray scale is directly
proportional to elevation.
SAS PROC MIXED summary results for a
quadratic gradient LMM: LN( elev + 17.5)
Semivariogram model
none
spherical
exponential
Gaussian
power
Bessel
---
0.0331
0.2151
0.2210
0.2514
0.2450
Variance
(nugget)
Spatial
correlation
Residual
b0
---
< 0.0001
0.7643
0.5730
0.2702
0.6089
0.217
6.080***
0.169
6.080***
< 0.0001
6.157***
0.0253
6.201***
< 0.0001
6.157***
0.0057
6.202***
bu2
-0.349*** -0.349***
-0.463*** -0.486*** -0.463*** -0.486***
buv
-0.263*** -0.263***
-0.254**
-0.255**
-0.254**
-0.257**
bv
-0.270*** -0.270***
-0.114
-0.168
-0.114
-0.168
bv2
-0.527*** -0.527***
-0.529*** -0.561*** -0.529*** -0.569***
The average random effects term example
MCMC chain from a WinBUGS run
ICAR
spatial filter (SF)
WinBUGS: geographic distributions of
unstructured (left) and spatially
structured (right) random effects
spatial filter (SF)
WinBUGS: ICAR
Comparative parameter estimates for a LMM
quadratic gradient description of LN( elev + 17.5)
Param- SAS semivariogram
eter
(Bessel) model
estimate
b0
bu2
buv
bv2
var
varure
varssre
6.1906
-0.5048
-0.2229
-0.5314
0.0055
0.2856
0.7205
se
SAS SF
estimate
se
0.287 6.1101 0.055
0.122 -0.3881 0.031
0.123 -0.2939 0.030
0.125 -0.5193 0.032
0.019
0
--0.091 0.0001
--0.221 0.0282
---
GeoBUGS-ICAR WinBUGS-SF
(100 weeded replications)
estimate
se
estimate
se
6.5175
-0.7507
-0.2031
-0.5683
0.0049
0.0047
0.4854
0.168 6.1101 0.061
0.153 -0.3878 0.035
0.101 -0.2920 0.030
0.061 -0.5190 0.037
0.007 0.0305 0.024
0.007 0.0318 0.025
0.093 0.0301
---
varure denotes the variance of the unstructured random effects
varssre denotes the variance of the spatially structured random effects
binomial GLMM random effects
SAS SF
WinBUGS SF
WinBUGS ICAR
SF
GLMM
statistic
b0
b elevelev
b E1
b E4
μ̂ ξ
σ̂ ξ2
σ̂ ξ2SS
P(S-W)
MCss
GRss
MC ξ̂
GR ξ̂
MC ξ̂ SS
GR ξ̂ SS
r ξ̂,elev
r ξ̂,E
1
r ξ̂,E
4
SAS NLMIXED
(SF)
estimate
se
WinBUGS (100 weeded replications)
SF
ICAR
estimate
se
estimate
se
-1.2867
-0.0111
3.0646
3.2182
0.0015
0.7045
0.9787
<0.0001
0.967
0.158
0.119
1.045
0.356
0.739
0.001
-0.001
0.001
-1.3114
-0.0110
3.0600
3.0116
0.0054
0.7144
0.9783
< 0.0001
0.975
0.154
0.132
1.000
0.357
0.739
0.001
-0.009
0.022
0.2624
0.0013
1.1559
1.3433
0.2852
0.0014
1.3632
1.4256
-1.5340
-0.0100
***
***
0.0066
0.3727
0.9583
< 0.0001
0.787
0.177
0.036
1.129
0.388
0.696
0.011
***
***
0.2419
0.0013
Graphical diagnostics of residuals
for the GLMM estimated with SAS
Scatterplot of the
SAS and mean
WinBUGS
estimated spatially
structured random
effects terms
Individual GLMM estimation results
for each Puerto Rican sugar cane
crop year
Crop
year
Individ- Raw perual SF
centages
eigenvector MC GR
#s
Point-in-time estimation
#
0s
-a
-belevelev
μ̂ ξ
σ̂ ξ2
P(SW)
1965/
0.484 0.458 3 1.1959 0.0064 0.0020 1.0135 0.0055
66
1966/
1,4,6,24 0.490 0.454 4 1.3786 0.0060 0.0021 1.1138 0.0027
67
1967/
0.474 0.434 6 1.7354 0.0050 0.0018 1.0856 0.0017
68
Space-time data:
preliminaries
random effect (re)
re + ss: 1966/67
re + ss: 1965/66
re + ss: 1967/68
Random effects term is constant across time; spatial
structuring changes over time
Space-time GLMM: Puerto Rican sugar
cane crop years 1965/66-1967/68 when all
fixed effects are year-specific
crop year 1965/66
statistic
b0
b elevelev
b E1
b E4
b E6
b E24
pseudo-R2
MC ξ̂ SS
GR ξ̂ SS
MCresiduals
GRresiduals
estimate
se
-1.2291
0.2336
-0.0065
0.0009
4.5040
1.1684
4.9713
1.2432
-4.5091
1.1620
-4.0290
1.0994
0.9950
0.449
0.580
0.019
0.916
crop year 1966/7
estimate
se
-1.3122
0.2336
-0.0064
0.0009
4.5785
1.1684
5.3372
1.2432
-4.8209
1.1620
-3.9657
1.0994
0.9976
0.467
0.562
0.042
0.839
crop year 1967/68
estimate
se
-1.4520
0.2336
-0.0065
0.0009
4.9226
1.1684
5.8053
1.2432
-5.0203
1.1621
-4.0285
1.0995
0.9929
0.493
0.537
0.009
0.808
Discussion & Implications
1. All three common specifications of spatial
structuring—semivariogram, spatial
autoregressive and SF models—for a random
effect term in mixed statistical models perform in
an equivalent fashion.
2. Matching Bayesian model priors with their
implicit frequentist counterparts yields estimation
results from both approaches that are essentially
the same.
3. making use of spatially structured random
effects tends to furnish an alternative to quasilikelihood estimation techniques for GLMMs
4. Semivariogram models offer a geostatistical theoretical
basis and have been implemented in SAS for LMMs.
•
A spatial statistics practitioner with the necessary computer
programming skills can employ WinBUGS in order to utilize
them with GLMMs.
5. Spatial autoregressive modeling offers a theoretical
basis for spatial structuring, and is available in
GeoBUGS.
•
This would be very difficult to trick SAS into doing.
6. Spatial filtering, which can be derived from spatial
autoregressive model specifications,
•
•
•
tends to be more exploratory in nature (being akin to principal
components analysis)
can be implemented in either SAS or WinBUGS for either
LMMs or GLMMs, and
can be easily extended to space-time datasets with either of
these software packages.
7. Illustrative Puerto Rico sugar cane
examples tend to have a random effect
term that virtually equates to the
corresponding LMM/GLMM residual
variate.
•
This is not always the case, as is highlighted
by the extension of a GLMM specification to
a space-time sugar cane dataset.
8. All of the estimated random effects terms
for the various Puerto Rico examples
tend to be non-normal.
9. once a random effect term has been
estimated with a frequentist approach,
using it when calculating a deviance
statistic allows its number of degrees of
freedom to be approximated for GLMMs.
• Although n values are estimated, because
they are correlated, the resulting number of
degrees of freedom is less than n.
• This particular finding should help spatial
statistics practitioners better understand the
cost of employing a statistical mixed model.
A df aside: future research
• Spiegelhalter et al. (2002) address the df problem
for complex hierarchical models in which the
number of parameters is not clearly defined
because, for instance, of the presence of random
effects.
• An information-theoretic argument is used to
approximate the effective number of parameters in
a model, equivalent to the trace of the product of
the Fisher information and the posterior covariance
matrices.
– this particular approximation is equivalent to the trace of
the ‘hat’ matrix for linear models with a normally
distributed error term.
k dfs for random effects
deviance binomial
k  n  (p  1) 
1
deviance Poisson
k  n  (p  1) 
deviance neativebinomial