Spatial Data Analysis
Download
Report
Transcript Spatial Data Analysis
Advanced Spatial Analysis
Spatial Regression Modeling
Paul R. Voss
and
Katherine J. Curtis
Day 5
GISPopSci
Review of yesterday
• Dealing with heterogeneity in relationships
across space
• Discrete & continuous spatial heterogeneity
in relationships
• GWR
- concept & motivation
- how to do it
- what it all means
GISPopSci
Questions?
GISPopSci
Plan for today
• Shift away from spatial econometrics
– away from the classical, frequentist perspective
– away from the MLE world
• Spatial data analysis from a Bayesian
perspective
– example using count data & Poisson likelihood
– meaning of Bayes’ rule
– MCMC
• Afternoon lab
– brief demonstration of spatial modeling in WinBUGS & R
– your analyses; your data; your issues
GISPopSci
Bayesian orientation…
GISPopSci
Situate this in a statistical framework (1)
Back in my day…
OLS linear regression models
continuous dependent variables
independent & normally distributed errors
(Legendre 1805; Gauss 1809; Galton 1875)
Adiren-Marie Legendre
(1752-1853)
Carl Friedrich Gauss
(1777-1855)
GISPopSci
Francis Galton
(1822-1911)
Situate this in a statistical framework (2)
Back in my day…
OLS linear regression models
continuous dependent variables
independent & normally distributed errors
(Legendre 1805; Gauss 1809; Galton 1875)
ascendency of the “frequentists”; NHST; Maximum Likelihood Estimation (MLE)
decision theory (Neyman & E. Pearson 1933)
evidential interpretation of statistical data; p-values (Fisher 1922 – 1960s)
Statistical
significance tests:
What should I do?
Jerzy Neyman
(1894-1981)
Egon Sharp Pearson
(1895-1980)
GISPopSci
Statistical
significance tests:
What’s the
Ronald Aylmer Fisher
evidence?
(1890-1962)
Situate this in a statistical framework (3)
Back in my day…
OLS linear regression models
continuous dependent variables
independent & normally distributed errors
ascendency of the “frequentists”;
Generalized Linear Models…
fixed effects regression
accommodate many kinds of outcomes:
Gaussian, Poisson, binomial
maximum likelihood estimation
independent observations
(Nelder & Wedderburn 1972; McCullagh & Nelder 1989)
John Ashworth Nelder
(1924 – 8/7/10)
R. W. M. Wedderburn
(1947 – 1975)
GISPopSci
Peter McCulloch
University of Chicago
Situate this in a statistical framework (4)
Back in my day…
OLS linear regression models
continuous dependent variables
independent & normally distributed errors
ascendency of the “frequentists”
Generalized Linear Models…
fixed effects regression
accommodate many kinds of outcomes:
Gaussian, Poisson, binomial
maximum likelihood estimation
(Nelder & Wedderburn 1972; McCullagh & Nelder 1989)
Generalized Linear Mixed Models…
linear predictor contains random effects
MLE (difficult integrals over the random effects)
Iterative approximations of the likelihood:
Laplace, PQL, MQL, numerical quadrature
(Breslow & Clayton 1993)
Norman E. Breslow
Univ. of Washington
GISPopSci
David G. Clayton
Cambridge Univ.
Situate this in a statistical framework (5)
Back in my day…
OLS linear regression models
continuous dependent variables
independent & normally distributed errors
(Legendre 1805; Gauss 1809; Galton 1875)
Generalized Linear Models…
fixed effects regression
accommodate many kinds of outcomes:
Gaussian, Poisson, binomial
maximum likelihood estimation
(Nelder & Wedderburn 1972; McCullagh & Nelder 1989)
Generalized Linear Mixed Models…
linear predictor contains random effects
MLE (difficult integrals over the random effects)
Iterative approximations of the likelihood:
Laplace, PQL, MQL, numerical quadrature
(Breslow & Clayton 1993)
MCMC…
1950s onward
simulations on Bayesian posterior distributions
samplers:
Gibbs, M-H, others
(enormous literature)
GISPopSci
While this was going on, indeed,
before all this, there was a
somewhat parallel development of
what originally was referred to as
“inverse probability”
19th Century:
Inverse Probability &
Linear regression
20th Century:
Frequentist paradigm
21st Century:
Bayesian paradigm?
And, at times, a rather fierce debate between what
came to be known as “classical (likelihood)
statisticians” and “Bayesians”
GISPopSci
The Reverend Thomas Bayes
• First discovered the theorem that
now bears his name
• Written up in paper, “An Essay
Towards a Problem on the Doctrine
of Chances”
• Published posthumously in the
Philosophical Transactions of the
Royal Society in 1763
• Independently(?) discovered and
formalized by Laplace (1774;
1781); “theory of inverse
probability” became the dominant
paradigm
• Interest (theoretical) renewed in
mid-20th century; interest (applied)
exploded around 1985-90
GISPopSci
Thomas Bayes
(1702 – 1761)
Pierre Simon,
Marquis de Laplace
(1749 – 1827)
GISPopSci
Bayesian methods are often turned to when
likelihood estimation becomes difficult or
unrealistic (especially in the world of spatial
data analysis)
• Advantages of a Bayesian perspective
-
more natural interpretation of parameter intervals & probability
intervals
clear & overt assumptions regarding prior assumptions
ease with which the true parameter density is obtained
ability to update as new data become available
likelihood maximization not required (for probit & logit models,
maximization of the simulated likelihood function can be difficult)
desirable estimation properties (consistency & efficiency)
• Disadvantages of a Bayesian perspective
-
steep learning curve; for those of us trained in a classical
perspective, things don’t come easy
determining a reasonable prior
GISPopSci
From early understanding of
the laws of probability…
P(AB) = P(A|B)P(B) = P(B|A)P(A)
from which we obtain “Bayes’ formula”:
P(B|A) = P(A|B)P(B)
P(A)
This statement is true whether you take a
Bayesian perspective or not. It’s only in a
certain understanding of this formula that
you become a Bayesian
GISPopSci
So… Bayes’ formula
P(B|A) = P(A|B)P(B)
P(A)
Now, from the “Law of Total Probability”,
we have for mass probabilities:
c
P( A) P( A | B) P( B) P( A | B ) P( B )
c
or, for density probabilities:
P( A) P ( A | B ) P ( B )dB
B
P(B|A) P(A|B)P(B)
“Posterior” probability
“Likelihood”
GISPopSci
“Prior” probability
Usually the focus is on using the Bayes
formulation to estimate model parameters
(rather than expressing relationships
among event probabilities
P(B|A) P(A|B)P(B)
P(parameters|data) P(data|parameters)P(parameters)
“Posterior” probability
“Data Likelihood”
“Prior” probability
The “Bayesian mantra”
The goal generally is to achieve “optimal”
parameter estimates where attention is focused
on precision (std. errors of the parameters)
GISPopSci
From this we have the
“Bayesian mantra”
POSTERIOR LIKELIHOOD x PRIOR
P[θ|D] P[D|θ] P[θ]
In full Bayesian modeling the prior
distribution is formulated before
consulting the data
In so-called empirical Bayesian
modeling the prior distribution is
derived from the data
GISPopSci
Approaches to Model Building
i.e., approaches to making good “guesses” about
the values of the parameters in the model and
inferences about them
• Classical statistics
• Bayesian statistics
• Regardless of approach, the process of
statistics involves:
– formulating a research question
– collecting data
– developing a probability model for the data
– estimating the model
– assessing the quality of the “fit” & modifying the model as necessary
– summarizing the results in order to answer the research question
(“statistical inference”)
GISPopSci
Classical approach (1)
• Three basic steps:
– specify the model
– estimate & evaluate the model (MLE)
– inference
• Fundamental idea:
– a good choice for the estimate of a parameter (viewed as fixed)
is that value of the parameter that makes the observed data
most likely to have occurred
GISPopSci
Classical approach (2)
• The process of MLE involves:
1. construction of a likelihood function of the parameter(s) of interest
L( | D) f ( D | ) P ( D | )
2. for example, normal likelihood:
L( | D) L( , | X ) f ( X | , )
n
i 1
( xi ) 2
exp
2
2
2
2
1
– Note: observations {xi, x2,…,xn) are assumed (conditionally)
independent
GISPopSci
Classical approach (3)
• The process of MLE involves:
(3) Simplify the likelihood function:
– for example, for the normal likelihood:
L( , | X ) (2 2)
n
2
1
exp 2
2
2
( xi )
i 1
n
(4) Take logarithm of likelihood function:
( , | X ) n ln( )
GISPopSci
1
2 2
n
i 1
( xi )2
Classical approach (4)
• The process of MLE involves:
(5) Take partial derivative of the log-likelihood function with respect to
each parameter (for example, for the normal log likelihood):
n( x )
2
and
n 1
3
n
( xi )2
i 1
(6) Set the partial derivatives to zero and solve for the parameters:
n
ˆ x and ˆ 2
These should look familiar
GISPopSci
( xi ) 2
i 1
n
Classical approach (5)
• The process of MLE involves:
(7) Take partial second derivatives to obtain estimates of the variability
in the estimates for the mean & standard deviation (Hessian matrix):
2
2
2
T
2
2
2
2
2
4
n
2
n( x )
4
n
2
GISPopSci
4
n( x )
4
n
( xi ) 2
i 1
6
Bayesian approach
• Three basic steps (same as classical approach),
with a rather different notion about the parameters
• Fundamental idea:
– a good choice for the estimate of the parameters (not fixed) are those
values of the parameter estimated from a probability distribution
involving both the observed data and our subjective notion of the
value of the parameter
– Bayesian approach tells you how to alter your beliefs
• The modern Bayesian approach involves:
– specification of a “likelihood function” (“sampling density”) for the data,
conditional on the parameters
– specification of a “prior distribution” for the model parameters
– derivation of the “posterior distribution” for the model parameters,
given the likelihood function and the prior distribution
– obtain a sample (Monte Carlo simulation) from the posterior
distribution of the parameters
– Summarize the parameter samples and draw inferences
GISPopSci
I’m going to use as a running
example an application in spatial
epidemiology
South Carolina congenital
abnormality deaths 1990
Data set developed & much used by
Andrew Lawson,
Spatial epidemiologist,
Biostatistics, Bioinformatics, and
Epidemiology program
Medical University of South Carolina
GISPopSci
South Carolina congenital
abnormality deaths 1990
4
11
13
6
5
0
0
2
0
2
1
3
5
7
3
0
4
0
14
7
0
1
1
3
3
0
7
8
6
11
0
8
1
1
2
8
16
Some considerations:
1
1
3
• Data are a complete realization, not a sample
17population
1
• Need to know something about
the underlying
1
0
to understand elevated counts;
understand the
relative risk
0 just
5 a list of numbers
• Spatial structure is relevant; it’s not
• Areal units are arbitrary; have nothing to do with health
outcomes
• Areal units are irregular in shape and
size
GISPopSci
Mapping issues
• Relative risk estimation
– key concept
– we want to capture & understand the data structure to get good
estimates of relative risk
– generally measured by comparing observed counts to expected
counts; standardization.
– helps to guide health resource allocations
• Disease clustering
– main issue is to understand apparent clustering of risk
– clustering of both data and risk is common (Tobler’s 1st Law)
– extent of clustering of risk and location of clusters of high risk
• Ecological analysis
– more research oriented
– identifying covariates to find causes of elevated risk or clusters
– especially useful when hypothesis involves a putative source of
pollution
GISPopSci
Relative risk estimation
• SMRs (Standardized Mortality/Morbidity
Ratios); or SIRs (Standardized Incidence
Ratios)
– yi :
Observed counts
– ei :
Expected counts (derived through standardization)
– yi / ei : Estimated SMR; estimate of i (a “saturated est.”)
• Some issues:
– the SMR is notoriously unstable; it can blow up when ei is
small; SMR is unbiased estimate of i but is inefficient;
need to compute and study the std. errors
– zero counts aren’t differentiated; estimated SMR is zero
when yi is zero, even when ei may vary considerably
– confusing notation: ei has different meanings
GISPopSci
If we assume the underlying process
is a Poisson process…
• Individuals within the study population behave
independently wrt disease propensity, after allowance is
made for observed or unobserved confounding variables;
“conditional independence”
• The underlying “at risk” background intensity (RR) has a
continuous & constant (i.e., i = ) spatial distribution within
the study area
• Events are unique (occur as spatially separate events)
• When these (Poisson process) assumptions are met, the
data can be modeled via a likelihood approach with an
intensity function representing the “at risk” background
GISPopSci
MLE of Relative Risk parameters…
• Likelihood: We usually assume that the counts of disease
result from a Poisson process such that yi has a Poisson
distribution with expected value eii
• We write: yi ~ Pois(eii)
• The counts have a Poisson likelihood:
L(eii | yi )
(eii | yi )
m
i 1
m
(eii ) yi exp(eii )
yi!
m
m
y (ln(e )) e ln(y !)
i
i 1
i i
i i
i 1
i
i 1
• Differentiate wrt eii and set to zero yields:
MLE
ˆ
i yi ei SMR
GISPopSci
So, with MLE…
• The counts have a joint probability of arising
based on the likelihood, L(eii |yi)
• L(eii |yi) is the product of Poisson probabilities for
each of the areas (again, why can we say this?)
• It tells us how likely the parameter estimates of i
are given the data, yi , under a set of Poisson
process assumptions
• This is a common approach for obtaining
estimates of parameters
GISPopSci
Cautions:
• Visual message can change with change
in quantile
declaration
Mean
= 0.9703
• Large literature on cognitive aspects of
Std. dev. = 0.8024
disease mapping
• Our eye is drawn to the extreme values;
so remember: unstable variances
GISPopSci
So… what have we got here?
• Choropleth map of SMRs of
infant mortality from congenital
abnormalities, SC counties, 1990
• SMRi = yi/ei, where yi is the
observed count & ei is the
expected count (note: we permit
ei to be fractional)
• In this example, the ei are simple expectations generated
using “indirect standardization”; i.e., by applying the crude
rate (per 1,000 births) for SC, aggregated over the years
1990-98, to county births in 1990:
46
1990
ei
1990
Births i
*
1990 98
(congenital abnormality deaths)i
i 1
46
1990 98
(births)i
i 1
GISPopSci
The SMR are the ML estimates of
the true unknown relative risks
• yi is the count of disease in the ith region (these are fixed
data assumed to follow a Poisson(ei i ))
ind
yi | i ~ Poisson(eii )
• ei is the expected count in the ith region (also fixed)
• We will say that i is the “true,” fixed, relative risk in the ith
region (a parameter); this is what we wish to estimate
• E(yi) = ei E(i) = ei i under assumptions of Poisson process
• yi / ei : SMR; an unbiased estimate of i
• So… we have some estimates of the relative risks in each
county; what about the variance of these estimates?
• It turns out that there are different opinions about this
GISPopSci
We now move beyond MLE…
In particular, we drop the classical notion that
i is a vector of fixed parameters; we assume
that the i are random variables (having a
distribution) and we will make some parametric
assumptions concerning these parameters
Which as Bayesians we will call “priors”
GISPopSci
We will establish a simple hierarchy…
• yi ~ Pois(eii)
• i ~ Gamma(, )… (don’t panic; next slide)
• This is a very simple example which allows the
risk to vary according to a distribution
• and are unknown here and we can either try
to estimate them from the data, or…
• Perhaps also give them a distribution:
e.g., ~ Exp(), ~ Exp()
• This is how Bayesian hierarchies are established:
– yi is the lowest level in hierarchy
– i is the next level
– & affect i ; i affects yi ; & affect &
GISPopSci
Why assume a gamma prior?
• That is, why i ~ Gamma(, ) ?
• It makes conceptual sense; as estimates of
relative risk, we want a distribution for i that
requires the i are non-zero, and have a mean
probably not too far from unity while allowing for
some observations well above unity
– Gamma is one such distribution, although in its
simplicity has some shortcomings
– Log-normal distribution shares these distributional
attributes (and opens up opportunities not permitted by
the Gamma distribution
• Important: Gamma prior is the “conjugate” prior
for the Poisson likelihood
• And what’s & ?
GISPopSci
We look at the notion of “conjugacy” in this
context by first examining the distributions
• We started out with a Poisson likelihood:
L ( | Y )
m
i 1
y
ie
yi!
im1 yi m
e
m
y!
i 1 i
• and then defined Gamma “prior”:
1
p ( )
e
( )
• Excluding the normalizing constants, the product
of Poisson likelihood & Gamma prior is:
p ( | Y )
im1 yi 1 ( m )
e
GISPopSci
which is
Gamma(yi+, m+)
Now…define “conjugacy”:
• When the prior & likelihood are of such a form that
the posterior distribution follows the same form as
the prior, the prior and likelihood are said to be
“conjugate”
• By way of illustration, the previous slide reveals
that a Poisson likelihood and a Gamma prior yield
a Gamma posterior
• Very important matter prior to modern MCMC
estimation
GISPopSci
A basic hierarchy
Hyperparameter
Data
Parameter
Hyperparameter
Data
1st level
Likelihood
Distribution
2nd level
Prior
Distribution
GISPopSci
Our model hierarchy
Defined here by a DAG (Directed Acyclic Graph)
= k is a shape parameter;
=1/ is an inverse scale
parameter
y
g ( | , )
GISPopSci
1
e
( 1)!
for 0
What do we do when
conjugacy analysis fails us?
• In general, a posterior distribution will be so
complex that we must use simulation to
obtain samples
• So… (cue the music!) Modern Posterior
Inference (MCMC) to the rescue!
GISPopSci
Modern Posterior Inference
• Unlike the ML parameters (estimates of risk), Bayesian
model parameters are described by a distribution;
consequently a range of values of risk will arise (some more
likely than others); not just the single most likely value
• Posterior distributions are sampled to give a range of these
values (posterior sample). This sample will contain a large
amount of information about the parameter of interest
• Again, a Bayesian model consists of a likelihood and prior
distribution(s)
• The product of the likelihood and the prior distribution(s)
yields the posterior distribution
• In Bayesian modeling, all the inference about parameters is
made from the posterior distribution
GISPopSci
Posterior Sampling
• There are several methods used for this.
The two most common are:
– Gibbs Sampling
– Metropolis-Hastings Sampling
• Both are examples of Markov Chain Monte
Carlo (MCMC) methods
Nicholas Constantine
Metropolis
(1915-1999)
W. Keith Hastings
(1930 - )
GISPopSci
Josiah Willard Gibbs
(1839-1903)
Gibbs Sampler
• Fast algorithm; it yields a new sample value at
each iteration
• Requires knowledge of the conditional
distributions of the parameters given the other
parameters
• For parameter set [{i }, , ]:
– Fix the i ‘s and to estimate
– Then, fix the new and , say, to estimate 1
– etc. etc. etc.
• BUGS (later WinBUGS) was developed for this
method (Bayesian inference Using Gibbs
Sampling)
GISPopSci
Gibbs sampler…
Posterior distribution is a joint pdf of the unknown parameters:
{ , , 1, 2, ..., n}
ˆ { | , 1, 2, ..., n}
we can do this because we have assigned
probability distributions to the parameters
and have initialized them
ˆ { | ˆ , 1, 2, ..., n}
ˆ1 { 1 | ˆ , ˆ , 2, ..., n}
ˆn {n | ˆ , ˆ , ˆ2, ..., ˆn 1}
one update cycle is complete
return to the top and begin next update cycle; do
this many times until convergence is achieved
GISPopSci
Metropolis-Hastings (MH) sampler
• This is a simple algorithm (more simple than
Gibbs sampling) for updating parameters
and sampling posterior distributions
• It does not require knowledge of the
conditional distributions, but it does not
guarantee a useful new sample at each
iteration
• Simple to implement
• WinBUGS now includes MH updating
GISPopSci
Metropolis-Hastings (MH) sampler
Works off a very simple rule: accept an updated
parameter value if it moves the simulation toward the
Posterior limiting distribution:
Let (, ’) be the acceptance probability of moving from
current parameter estimate to proposed update ’:
p{ ' k | 1, 2 , ..., k 1, k 1, ..., n}
(k , 'k ) min 1,
p{k | 1, 2 , ..., k 1, k 1, ..., n}
Update is accepted if the proposal function is > 1
GISPopSci
Brief history
• Thomas Bayes (1702 – 1761); Oh… him
• Gibbs sampler: Geman & Geman (1984)
• MCMC: Gelfand & Smith (JASA, 1990)
• BUGS: Biostatistics Unit, Medical Research
Council, Cambridge University; 1989
• WinBUGS: Imperial College School of Medicine
at St Mary's, London; 2000
• Continued development:
–
–
–
–
–
OpenBUGS
DoodleBUGS
GeoBUGS
BRUGS
R2WinBUGS
GISPopSci
Using WinBUGS
• WinBUGS is a windows version of the
BUGS package. BUGS stands for
Bayesian inference using Gibbs
Sampling
• The software must be programmed to
sample form Bayesian models
• For simple models there is an
interactive Doodle editor; more complex
models must be written out fully.
GISPopSci
WinBUGS Introduction
GISPopSci
Doodle Editor
• The doodle editor allows you to visually
set up the ingredients of a model
• It then automatically writes the BUGS
code for the model
GISPopSci
BUGS code and Doodle stages
for simple Poisson-Gamma model
GISPopSci
BUGS code and Doodle stages
for simple Poisson-Gamma model
GISPopSci
Doodle model code:
Original model code:
GISPopSci
Sometimes we may choose to just
set the hyperparameters
Or let the data determine the values
(so-called “Empirical Bayes”)
GISPopSci
For example, we might set the Gamma
hyperparameters to, say, 0.1 & 0.1
These are somewhat “uninformative” or “diffuse” priors
GISPopSci
Even a very non-informative prior has the
effect of slightly pulling the likelihood estimates
(the SMRs) toward the mean of the prior
“Bayesian smoothing”
GISPopSci
Returning to the South Carolina
congenital abnormality deaths 1990
4
11
13
6
5
0
2
0
2
0
1
3
3
0
4
5
7
0
14
7
0
1
1
3
3
8
6
11
and recalling how
the expected deaths
were calculated…
(indirect
standardization)
1
1
0
0
8
2
8
7
16
1
1
3
1
1
17
0
0
46
1990
ei
1990
Births i
*
5
1990 98
(congenital abnormality deaths)i
i 1
46
1990 98
(births)i
i 1
GISPopSci
The data (including 2 covariates)
GISPopSci
GISPopSci
Berkeley County, SC
1990 Deaths: 16; Expected: 8.54
SMR: 1.87 (1.07 – 3.04)
2000 pop: 143,000
68% white
Poverty rate: 11.8%
Density: 129 pers/mi2
Part of the Charleston-North Charleston-SommervilleMetropolitan Statistical Area
Principal employer: U.S. Naval Weapons Station (18,450)
Saluda County, SC
1990 Deaths: 3; Expected: 0.77
SMR: 3.92 (0.81 – 11.46)
2000 pop: 18,000
66% white
Poverty rate: 15.6%
Density: 41 pers/mi2
GISPopSci
Confidence intervals for
the SMRs
It turns out that there
are many options
We’ll examine some of
this in lab this afternoon
GISPopSci
GISPopSci
Saluda Co.
Deaths = 3
McCormick Co.
Deaths = 0
Allendale Co.
Deaths = 1
Calhoun Co.
Deaths = 0
Lee & Bamberg Cos.
Deaths = 1 each
Barnwell Co.
Deaths = 1
Berkeley Co.
Deaths = 16
GISPopSci
GISPopSci
GISPopSci
GISPopSci
GISPopSci
GISPopSci
This afternoon we will see what happens
when various Bayesian models are fit to
these data to estimate relative risk
•
•
•
•
•
•
•
Simple Poisson-Gamma
Poisson-Log-Normal with pov. covariate
Poisson-Log-Normal with inc. covariate
Poisson-Log-Normal with both pov. & inc.
Poisson-Log-Normal with pov. & UH
Poisson-Log-Normal with pov., UH & CH
Preview: Bayesian smoothing tells us there’s
probably nothing interesting in these data
regarding the incidence of infant congenital deaths
in 1990
GISPopSci
Mean
= 0.9703
Std. dev. = 0.8024
GISPopSci
Mean
= 1.0373
Std. dev. = 0.0814
GISPopSci
Original SMR MLE
saturated estimates of
relative risk
Mean
= 0.9703
Std. dev. = 0.8024
Full Bayes mean
posterior estimates of
relative risk (model:
PLNPovUHCH)
GISPopSci
Mean
= 1.0373
Std. dev. = 0.0814
GISPopSci
GISPopSci
GISPopSci
GISPopSci
GISPopSci
GISPopSci
GISPopSci
GISPopSci
Things I’ve learned
• This stuff is not easy
– the price of entry is high, despite the free software
• Learning to efficiently code in R takes practice
– my R script is not particularly elegant
• Look at carefully at the WinBUGS log file
– the model doesn’t always converge (see next slide)
– may have to try it 3-4 times
Good
• Because full Bayes modeling
is based on a
simulation, the results vary from one run to
another
– This takes some time getting used to
Not good
GISPopSci
Readings for today
• Besag, Julian, Jeremy York, & Annie Mollié.
1991. “Bayesian Image Restoration with Two
Applications in Spatial Statistics.” Annals of the
Institute of Statistical Mathematics 43(1):1-20. [In
the beginning…]
• Package ‘R2WinBUGS: Running WinBUGS and
OpenBUGS from R / S-PLUS. Version 2.1-18.
March 22, 2011. [Useful acces to R functionality
in a Bayesian framework]
GISPopSci
Afternoon Lab
Bayesian Modeling in
WinBUGS & R
Presentations & discussion
GISPopSci
Areas of needed research:
• Spatial panel models; space-time
interactions
• Latent continuous variables; binary
dependent variable; counts; etc.
• Flow models
• Endogenous weighting matrices
• More users of the Bayesian perspective in
the social sciences
GISPopSci
Thanks for your participation!!
Special thanks Stephen Matthews, Don
Janelle, & the terrific PSU support!
See you this afternoon
GISPopSci