This may be the author’s version of a work that was submitted/accepted

for publication in the following source:

Cramb, Susanna, Duncan, Earl, White, Nicole, Baade, Peter, &

Mengersen, Kerrie
(2016)
Spatial Modelling Methods.
Cancer Council Queensland and Queensland University of Technology
(QUT), Brisbane, Qld.

This file was downloaded from: https://eprints.qut.edu.au/204103/

c Consult author(s) regarding copyright matters

This work is covered by copyright. Unless the document is being made available under a
Creative Commons Licence, you must assume that re-use is limited to personal use and
that permission from the copyright owner must be obtained for all other uses. If the docu-
ment is available under a Creative Commons License (or other specified license) then refer
to the Licence for details of permitted re-use. It is a condition of access that users recog-
nise and abide by the legal requirements associated with these rights. If you believe that
this work infringes copyright please provide details by email to [email protected]

Notice: Please note that this document may not be the Version of Record
(i.e. published version) of the work. Author manuscript versions (as Sub-
mitted for peer review or as Accepted for publication after peer review) can
be identified by an absence of publisher branding and/or typeset appear-
ance. If there is any doubt, please refer to the published source.

https:// cancerqld.blob.core.windows.net/ site/ content/ uploads/ 2018/ 12/

Statistical-Methods-Report-2016.pdf
Spatial Modelling Methods

Prepared for the National Health Performance Authority

June 2016
Suggested citation
Cramb SM, Duncan EW, White NM, Baade PD, Mengersen KL, 2016. Spatial Modelling
Methods. Brisbane: Cancer Council Queensland and Queensland University of Technology
(QUT).

Author affiliations
Viertel Cancer Research Centre, Cancer Council Queensland: Susanna Cramb and Peter
Baade.

ARC Centre of Excellence for Mathematical and Statistical Frontiers, Queensland University
of Technology (QUT): Susanna Cramb, Earl Duncan, Nicole White and Kerrie Mengersen.

Acknowledgements
The authors wish to thank the following people for their helpful feedback and advice:
 Professor Joanne Aitken, Head of Research and Director of Registries, Viertel Cancer
Research Centre, Cancer Council Queensland.
 Dr Paramita Dasgupta, Senior Research Officer, Viertel Cancer Research Centre,
Cancer Council Queensland.

ii
Executive Summary

Spatial information and spatial technologies can bring significant value to health agencies
through improved decision support, resource management and allocation, and clinical
outcomes. Disease mapping is used to explain and predict patterns of diseases outcomes
across geographical areas, identify areas of increased risk, and assist in understanding the
causes of diseases. As such, its use in informing policy recommendations is growing, and
diseases of national importance, such as cancer, are being increasingly mapped across small
regions.

Yet there are many potential methodological approaches for examining disease data over
small areas, and understanding the benefits and disadvantages of any single approach when
applied to a given situation is critical.

The aims of this report are threefold. First, to provide an accessible overview of the methods
used in analysing spatial public health data, ranging from raw (unsmoothed) estimates
through to complex Bayesian hierarchical models. Secondly, to outline the practical
computational implementation of these methods. Finally, by comparing the advantages and
disadvantages of these methods, to provide general guidelines and recommendations for their
use.

Examples of the methods used in existing cancer atlases and other small-area analyses are
also provided, as well as Bayesian approaches to incorporating multiple nested regions;
considering the combined influence of related variables, such as remoteness and area-level
socioeconomic disadvantage; small-area estimation from survey data; and extending the
spatial analyses to also consider differences over time (spatio-temporal models).

Key issues to consider when using spatial data include data quality, including the reliability
of location measures, and the degree of similarity between nearby areas (spatial correlation).

Although unsmoothed estimates such as crude or age-standardised rates may be useful for
exploratory analyses, they are rarely appropriate for small-area analyses due to the small
numbers involved, and should not be used when:
1. The addition of one event (disease case/death), or one more person at risk, results in a
large difference (such as 25% or more) in at least one area’s rates.
2. The number of events (rate numerator) is less than three for at least one area.
3. The population at risk per area is small (typically less than 500 people), and these
numbers vary by an order of magnitude across the areas.

Smoothing methods may be either direct (e.g. locally-weighted, kernel smoothing) or model-
based (e.g. Poisson kriging, Empirical Bayes or fully Bayesian). In general, direct smoothing
methods are also more appropriate for exploratory analyses, but less useful when
investigating contributing factors as they have more limited capacity for adjusting for
covariates.

iii
Model-based smoothing approaches have several advantages over the direct smoothing
methods, and their use is recommended when assessing the impact of covariates is important,
or the underlying pattern of risk needs to be understood.

There is no one model that represents the ultimate approach for disease mapping. The aims of
the analysis, data quality, and expected results (such as disparate risks between nearby areas)
can all influence the selection of the final model. Nonetheless, Bayesian hierarchical models
are increasingly used in disease mapping, have been shown to perform well overall, and with
the more recent application of approximation methods are able to generate results quickly.

For a cancer atlas, we generally recommend the use of Bayesian hierarchical models. The
fully Bayesian approach enables the development of more complex, realistic models with
reliable disease rates in low population areas, clearer summaries of spatial and temporal
correlation, more precise and interpretable confidence intervals, and greater ability to account
for and quantify measured sources of uncertainty than other possible approaches. The
Bayesian approach also has excellent flexibility in handling changing inferential goals, such
as obtaining smoothed risk maps as well as identifying motivating predictors of disease such
as ethnicity or socioeconomic status.

iv
Contents
Executive Summary...................................................................................................................................... iii
List of Boxes …. ............................................................................................................................................. vii
List of Tables .............................................................................................................................................. vii
List of Figures .............................................................................................................................................viii
List of Abbreviations.................................................................................................................................... ix
1. Introduction ................................................................................................................................................ 1
1.1 Background ................................................................................................................ 1
1.2 Definition of small area/spatial ................................................................................ 2
1.3 Aims ............................................................................................................................ 3
1.4 Structure of report ..................................................................................................... 3
2. Calculating small-area health estimates ........................................................................................... 5
2.1 Analysing spatial correlation .................................................................................... 5
2.2 Defining the neighbourhood ..................................................................................... 8
2.3 Unsmoothed estimates ............................................................................................ 10
2.4 Direct smoothing ..................................................................................................... 13
2.4.1 Locally-weighted average/median ............................................................... 13
2.4.2 Kernel smoothers ........................................................................................... 14
2.5 Model-based smoothing .......................................................................................... 15
2.5.1 Foundational approaches .............................................................................. 15
2.5.2 Poisson Kriging............................................................................................... 17
2.5.3 Empirical Bayes .............................................................................................. 18
2.5.4 Fully Bayesian................................................................................................. 20
2.5.4.1 Incidence/mortality data .................................................................. 20
2.5.4.2 Survival data....................................................................................... 26
2.6 Computation............................................................................................................. 28
2.6.1 Numerical........................................................................................................ 28
2.6.2 Markov chain Monte Carlo ............................................................................ 30
2.6.3 MCMC approximation methods .................................................................... 31
2.6.4 Creating the neighbourhood matrix ............................................................. 31
2.6.5 Software .......................................................................................................... 32
3. Current approaches for small-area cancer estimates ............................................................... 35
3.1 Small-area cancer screening estimates.................................................................. 35
3.2 Small-area cancer incidence/mortality estimates ................................................ 36
3.3 Small-area cancer survival estimates .................................................................... 37
3.4 Summary and conclusion ........................................................................................ 38

v
4. Further topics in Bayesian models ................................................................................................... 39
4.1 Inclusion of multiple nested geographies.............................................................. 39
4.2 Inclusion of remoteness and area-level socioeconomic status............................ 41
4.3 Use with survey data ............................................................................................... 42
4.4 Spatio-temporal data ............................................................................................... 44
5. Recommendations .................................................................................................................................. 47
5.1 When should smoothing/modelling replace direct estimation? ......................... 47
5.2 What type of smoothing/modelling should be used? ........................................... 47
5.3 What methods should be used for a cancer atlas? ................................................ 48
5.4 Conclusion ................................................................................................................ 50
References .............................................................................................................................................. 51
Appendix A Glossary ............................................................................................................................. 65
Appendix B Bayesian disease mapping tutorial ......................................................................... 69
Appendix C Computational software ............................................................................................. 93
Appendix D Recommended further reading ................................................................................ 97

vi
List of Boxes
Box 2.1 Global measures of spatial correlation ......................................................................... 6
Box 2.2 Local measures of spatial correlation ........................................................................... 7
Box 2.3 Weighting mechanism examples .................................................................................. 9
Box 2.4 Crude rates.................................................................................................................. 11
Box 2.5 Directly age-standardised rates .................................................................................. 11
Box 2.6 The Standardised Morbidity/Mortality Ratio (SMR) ................................................. 12
Box 2.7 Variance and covariance ............................................................................................ 13
Box 2.8 The Nadaraya-Watson kernel estimator ................................................................ 15
Box 2.9 GLMs and GLMMs .................................................................................................... 16
Box 2.10 The Poisson model ................................................................................................... 16
Box 2.11 The Semivariogram .................................................................................................. 17
Box 2.12 Bayesian inference ................................................................................................... 19
Box 2.13 Calculating the effective sample size of the conjugate prior distribution ................ 19
Box 2.14 The BYM model....................................................................................................... 20
Box 2.15 Alternative priors to the BYM ................................................................................. 22
Box 2.16 Lawson and Clark’s model ....................................................................................... 23
Box 2.17 Hidden Potts-Markov random field model............................................................... 25
Box 2.18 Bayesian spatial relative survival model .................................................................. 27
Box 2.19 BLUE, Maximum likelihood and Least Squares ..................................................... 29
Box 2.20 MCMC ..................................................................................................................... 30
Box 2.21 INLA ........................................................................................................................ 31
Box 4.1 Nested geographies as a hierarchical model .............................................................. 40
Box 4.2 The basic area-level model ......................................................................................... 43
Box 4.3 The Bernardinelli spatio-temporal model................................................................... 44
Box B.1 Bayesian model.......................................................................................................... 72
Box B.2 Normal distribution.................................................................................................... 73
Box B.3 Selecting regional scale ............................................................................................. 74
Box B.4 Data required to produce incidence estimates ........................................................... 75
Box B.5 Data required to produce survival estimates.............................................................. 76
Box B.6 Probability distributions used in epidemiology ......................................................... 76
Box B.7 The incidence model .................................................................................................. 78
Box B.8 The relative survival model ....................................................................................... 79
Box B.9 Prior distributions for the random effects .................................................................. 79

List of Tables
Table 1.1 Examples of the types of data used in spatial epidemiological studies ..................... 1
Table 2.1 Common mapping, GIS and statistical software and capabilities ........................... 32
Table 2.2 Examples of software used by method .................................................................... 33
Table 5.1 Summary of key methods ........................................................................................ 49

vii
List of Figures
Figure 1.1 Components of a spatial statistical analysis ............................................................. 2
Figure 2.1 Spatial correlation patterns ....................................................................................... 8
Figure 2.2 Semivariogram components ................................................................................... 17
Figure 4.1 Schematic representation of a hierarchical model .................................................. 40
Figure B.1 The representation of neighbourhood structure of area i. ...................................... 77
Figure B.2 Bayesian smoothed estimate of RER ..................................................................... 83
Figure B.3 Uncertainty of Bayesian smoothed estimate of RER............................................. 84
Figure B.4 Uncertainty of Bayesian smoothed estimate of RER............................................. 84
Figure B.5 Distribution of smoothed RER estimates............................................................... 85
Figure B.6 Using posterior probabilities to classify risk ......................................................... 86
Figure B.7 Thematic map depicting the probability of RER exceeding 1 and 1.2 .................. 87

viii
List of Abbreviations
APC Age-Period-Cohort
ATA Area-to-area
ATP Area-to-point
BLUE Best Linear Unbiased Estimator
BLUP Best Linear Unbiased Predictor
BUGS Bayesian inference Using Gibbs Sampling
BYM Besag, York and Mollié
CAR Conditional Autoregressive
CI Confidence interval
DSR Directly Standardised Rate
EB Empirical Bayes
EBLUP Empirical Best Linear Unbiased Predictor
ESRI Environmental Systems Research Institute
GIS Geographic Information System
GLM Generalised Linear Model
GLMM Generalised Linear Mixed Model
G-NAF Geocoded National Address File
GNU GNU’s Not Unix
GPS Global Positioning System
GRASS Geographic Resources Analysis Support System
HB Hierarchical Bayes
INLA Integrated Nested Laplace Approximation
JAGS Just Another Gibbs Sampler
LISA Local Indicators of Spatial Association
MCMC Markov chain Monte Carlo
MEET Maximised Excess Events Test
M-H Metropolis-Hastings
MLE Maximum Likelihood Estimation
MRF Markov Random Field
MSE Mean Squared Error
NHPA National Health Performance Authority
REML Restricted Maximum Likelihood
RER Relative Excess Risk
SA1 Statistical Area 1
SA2 Statistical Area 2
SAS Statistical Analysis System
SEBLUP Spatial Empirical Best Linear Unbiased Predictor
SIR Standardised Incidence Ratio

ix
SLA Statistical Local Area
SMR Standardised Mortality Ratio
STARS Space-Time Analysis of Regional Systems
STIS Space-Time Information System
UK United Kingdom
USA United States of America

x
1. Introduction
“Knowing where things are, and why, is
essential to rational decision making.”
~ Jack Dangermond,
Environmental Systems Research Institute (ESRI)

1.1 Background

Place affects health (1). Spatial epidemiology aims to quantify and explain geographic
variation in diseases and their relationship with environmental, demographic, behavioural,
socioeconomic, genetic and infectious disease factors (2, 3). As such, disease mapping is an
integral component of spatial epidemiology (3). Disease mapping can explain and predict
patterns of diseases outcomes across geographical areas, identify areas of increased risk, and
assist in understanding the causes of diseases (4).

Data used for spatial epidemiological analyses require information on the disease of interest,
as well as a geographic location (Table 1.1) (5). This geographic location may be available at
either the point or area level. Point level data refers to having the exact geocoded locations
available, while area-level, or areal, data are only available for a region. Areal data are
considered to have a constant estimate over the entire region, but commonly this is an
aggregate measure such as the number of counts. Areas may consist of a regular lattice, or
they may consist of irregular shapes.

Table 1.1 Examples of the types of data used in spatial epidemiological studies
Data Description
Health or disease Vital statistics, notifiable diseases, patient registries, and health
surveys from various international or government agencies.
[Location is usually based on residential address]

Field epidemiology Surveyed data on disease occurrences with location coordinates

collected via GPS.
Spatially referenced base Digital cartographic data available from various international or
government agencies. [Often includes contours, rivers, and built
environment features]
Remotely sensed Land cover, elevation, soil type as reflected by satellite images.
Environmental and natural resources Interpreted data on land use, water quality, air quality, climate,
geology, etc.
Census or demographic Sociodemographic and economic data.
Note: Modified version of Table 2.2 in (5), page 23. GPS=Global Positioning System.

The observed data represents one of three components involved in a spatial statistical analysis
(Figure 1.1). Any one of these components may drive subsequent development of the other
two, and often multiple circuits will occur before the process is complete (6). An existing
map could inform data collection which then determines the appropriate statistical analysis.

1
Or a model could be developed, data collected and a map produced. This report concentrates
on the analysis and modelling element. As areal data are commonly used by NHPA, this
report focuses exclusively on methods for analysing area-level data. These data are also more
readily available due to fewer data privacy restraints than point data. Another associated
report “Communicating statistical outputs through maps” discusses the mapping component
in detail. Cancer is a collection of diseases that are increasingly being mapped at the small-
area level, and the third associated report “Grey Literature Review: Internet Published Cancer
Maps” focuses on online, interactive cancer maps.

Figure 1.1 Components of a spatial statistical analysis

Data

Analysis Maps

Note: Modified from Figure 12.4 in (6), page 179.

Many areal disease mapping methods initially arose from approaches to restoring images
(that is, undoing image defects, such as motion blur, noise and/or camera misfocus), but there
are important differences in public health data (7). First, areas are often irregular in shape and
size (8), in contrast to the regular gridded lattice used in image restoration. There are
commonly far fewer areas in public health data than pixels in an image, and these have
varying rather than constant numbers of neighbours. Adjustment for important variables in
the public health context is likely to influence results, whether that be population size or age,
comorbidities or disease stage, while image restoration is primarily about the visual structure
alone. Finally, any true boundaries in underlying risk are likely to be obscured by random
noise in public health data, whereas images typically have clearly defined boundaries and
multiple consecutive pixels with an identical colour (7).

1.2 Definition of small area/spatial

A small-area is defined in this report as an area that has a small population, and is not
necessarily associated with their geographical size. What is a small population? This is
determined by the disease of interest. A common disease, with a rate of 50%, could be quite
well approximated with a population of 100. A less common disease, such as cancer, with a
rate of 0.5%, would need a population of 10,000 to obtain a comparable number of cases.

2
Debate continues over the appropriate scale and definition of an area for a spatial analysis (9).
There are no clear rules: selecting an appropriate spatial scale depends on the objective of the
analysis, and (to a slightly lesser extent) data availability (9, 10). Public health estimates
commonly require population data, which is often only available for administrative
boundaries, and it is unusual for population estimates to be disaggregated to very fine
resolution, especially by age and sex breakdowns.

Note that the terms small-area analysis and spatial analysis are used interchangeably in this
report.

1.3 Aims

The aims of this report are threefold. First, to provide an accessible overview of methods
used in analysing spatial public health data, ranging from raw (unsmoothed) estimates
through to complex Bayesian hierarchical models. Secondly, to outline the practical
computational implementation of these methods. Finally, by comparing the advantages and
disadvantages of these methods, to provide general guidelines and recommendations for their
use.

This report is not designed to be a comprehensive review, but instead seeks to broadly
examine the key methods and models appropriate for areal data. Cancer is a complex group
of diseases whose outcomes are increasingly the subject of small-area analyses. Cancer is
used throughout this report to illustrate methods used in analysing spatial variation in disease
outcomes.

1.4 Structure of report

This report is structured as follows. Technical details are provided in Boxes throughout the
report for those interested. Additional details are also available in the Appendices, including a
glossary of terms, a tutorial on Bayesian disease mapping, details on computational packages
and software available, as well as further recommended reading.

In Chapter 2, methods used for analysing spatial data are presented, including unsmoothed
estimates, direct smoothing methods as well as model-based smoothing approaches. No data
are perfect, and this chapter outlines several statistical inference approaches to enable
learning from these data. This encompasses examination of correlation (also known as cluster
or hot-spot analysis), unsmoothed estimation, direct smoothing approaches such as locally-
weighted averages or kernel smoothers, model-based smoothing, and computational
approaches.

Chapter 3 then focuses on one group of diseases – cancer – and discusses the methods that
have been applied to generate published small-area cancer estimates for screening,
incidence/mortality, and survival data.

3
Chapter 4 presents a discussion of Bayesian approaches to spatial modelling, with four
specific topics considered: (i) multiple nested geographies, (ii) combined remote and
socioeconomic categories, (iii) survey data, and (iv) spatio-temporal data.

Finally, recommendations for when to use smoothing methods, and the types of smoothing or
modelling methods likely to be the most appropriate, both generally and specifically for
cancer atlases, are presented along with concluding remarks in Chapter 5.

4
2. Calculating small-area health estimates
“Everything is related to everything else, but
near things are more related than far things.”
~ Waldo Tobler (First Law of Geography) (11)

Spatial data has specific characteristics that must be considered. Indeed, Fischer and Wang
(12) titled their discussion around these issues as “the tyranny of spatial data”.

First, data quality is critical (12). If data are geocoded, understanding the accuracy of the
process is important. Although different types of geocoding exist, the most pertinent in
spatial epidemiology is the geocoding of residential addresses (13). Potential errors in
geocoding at the street number level include a low match rate (the completeness of geocoding
to the street number level), positional error (geocoded point is not near the ‘true’ location),
and low concordance (assignment to the correct geographic unit) (13). Geocoding is only
reliable if the output is of high quality and repeatable (13). Note that repeatability can be
influenced by variations in the reference data, the matching algorithms used by the geocoding
software, as well as the skills and experience of geocoding personnel (13). Recently,
Australia released the Geocoded National Address File (G-NAF), which is updated quarterly,
although the level of uptake from health agencies is unclear.

2.1 Analysing spatial correlation

Beyond data quality, there are important inferential issues for spatial data. Perhaps the most
important of these involve spatial correlation, clearly expressed in Tobler’s first law of
geography (11). This law posits that areas closer together are more similar than those further
apart. Spatial correlation implies correlation among the same measure from different
locations (14). Where spatial correlation is present, the assumptions that data are independent
and identically distributed (the backbone of most traditional regression analyses) are violated
(15, 16). Ignoring these spatial properties can result in false conclusions (17, 18). Any
statistical techniques that assume data are independent are therefore not valid when spatial
correlation is present.

A large range of options for testing for spatial correlation are available, and many GIS
packages as well as standard statistical packages include the ability to conduct several tests.
Popular options for area-level data include Moran’s I (19), Geary’s C (20), and the Localised
Indicators of Spatial Association (LISA) (21) (Boxes 2.1 and 2.2). Newer options that have
been shown to perform well (22) include Tango’s Maximised Excess Events Test (MEET)
(23) and the spatial scan statistic (24) (Boxes 2.1 and 2.2). These may assess spatial
correlation throughout the entire study region (called global clustering, such as Moran’s I), or
may detect localised correlation (also called local clustering, or hot-spot analysis, such as the
LISA).

5
Box 2.1 Global measures of spatial correlation

Moran’s I (19) can be defined as:

𝐼 ∑𝐼𝑖=1 ∑𝐼𝑗=1 𝑤𝑖𝑗 (𝑧𝑖 − 𝑧̅)(𝑧𝑗 − 𝑧̅)
Moran′ s 𝐼 = ×
∑𝐼𝑖=1(𝑧𝑖 − 𝑧̅)2 ∑𝐼𝑖=1 ∑𝐼𝑗=1 𝑤𝑖𝑗

where 𝑧𝑖 is the observed value at each area i=1,…I areas, 𝑧̅ is the mean value, 𝐼 is the number
of areas and 𝑤𝑖𝑗 are the weights indicating which areas are adjacent/close together.

The input observed values 𝑧𝑖 may be the original observations, or some standardisation to
avoid scale dependence, such as the deviations from the mean (21). Standardised values are
generally preferable.

The values of Moran’s I generally span from -1 (dispersed) to +1 (clustering). Values around
0 indicates no spatial correlation.

Geary’s C (20) is similar to Moran’s I, as can be seen from the following definition using the
same notation:
𝐼−1 ∑𝐼𝑖=1 ∑𝐼𝑗=1 𝑤𝑖𝑗 (𝑧𝑖 − 𝑧𝑗 )2
Geary ′ s 𝐶 = 𝐼 ×
∑𝑖=1(𝑧𝑖 − 𝑧̅)2 2 ∑𝐼𝑖=1 ∑𝐼𝑗=1 𝑤𝑖𝑗

The values of Geary’s C typically range between 0 and 2. A value of 1 means no spatial
correlation, <1 indicates positive spatial correlation, and >1 indicates negative spatial
correlation (25). Geary’s C is considered more sensitive to local spatial correlation than
Moran’s I (5).

Tango’s MEET (23) is based on the calculation of the excess events test (EET) (26), which is
a weighted sum of the excess number of events (observed minus expected), with higher
weights when areas are proximal, as follows:

2
4𝑑𝑖𝑗
− 𝑝𝑖 𝑂𝑇𝑂𝑇 𝑝𝑗 𝑂𝑇𝑂𝑇
𝐸𝐸𝑇 = ∑ ∑ 𝑒 𝜆2 (𝑂𝑖 − ) (𝑂𝑗 − )
𝑃𝑇𝑂𝑇 𝑃𝑇𝑂𝑇
𝑖 𝑗

where 𝑜𝑖 and 𝑝𝑖 are the observed count and population, respectively, in each area, 𝑂𝑇𝑂𝑇 and
𝑃𝑇𝑂𝑇 are the overall count and population, and 𝑑𝑖𝑗 represents the distance between area i and
area j. The choice of 𝜆 can influence the outcome, with large values of 𝜆 increasing
sensitivity to detecting large geographical clusters, while small 𝜆 increases the sensitivity to
small clusters.

Tango’s MEET overcomes this by considering multiple versions of 𝜆 up to a pre-determined

value. This enables clustering to be detected irrespective of geographical scale. A small p-
value indicates clustering is present.

6
Box 2.2 Local measures of spatial correlation

One form of the LISA (Local Indicators of Spatial Association) can be considered the local
equivalent of Moran’s I, and expressed as:

𝑛
𝐼𝑖 = (𝑧𝑖 − 𝑧̅) × ∑ 𝑤𝑖𝑗 (𝑧𝑖 − 𝑧̅)2
𝑗∈𝐽𝑖

It is also possible to have a LISA version of other common global indicators, including
Geary’s C. The LISA for each area indicates the extent of significant spatial clustering
around that area, and the sum of LISAs for all areas is in proportion to the corresponding
global statistic (27).

Disadvantages of the LISA include multiple testing issues as a separate statistical test is
conducted for each region (14). These regions are also small, and rates are unstable, risking
spurious significance. Although a Bonferroni adjustment is often used to account for multiple
tests, the correlation between neighbouring LISAs (as they share some of the same
observations), would cause this adjustment to be very conservative (14).

The spatial scan statistic (24) considers a large number of overlapping circles of assorted
sizes and locations. Specialised software has been developed and is freely available to
implement this method (SaTScan) (28). This method can be used for a range of data (count,
ordinal, binomial, even multinomial and survival) and can also be adjusted for covariates (29,
30). This method of detecting clusters is based on maximising the likelihood ratio.

The spatial scan statistic is proportional to

𝑂𝐼𝑁 𝑂𝐼𝑁 𝑂𝑂𝑈𝑇 𝑂𝑂𝑈𝑇

max ( ) ( )
𝐸𝐼𝑁 𝐸𝑂𝑈𝑇

where 𝑂𝐼𝑁 and 𝑂𝑂𝑈𝑇 are the observed counts inside and outside the circle, respectively, and
𝐸𝐼𝑁 and 𝐸𝑂𝑈𝑇 are the respective expected counts inside and outside the circle (14).

Although the spatial scan statistic has been reported as performing well in comparison to
other methods (22), others have raised concerns about the large size of the clusters detected
and difficulties in detecting cluster shapes other than circles (31).

While the aim of global clustering methods is to determine if there is clustering throughout
the region, the precise location of any clustering is not important (22). Instead, results may
provide a general indication of overall patterns, such as whether any correlation is positive
(similar values are clustered together), or negative (dissimilar values are together) (Figure
2.1).

7
In contrast, local clustering methods seek to detect the location of statistically significant
spatial clusters and outliers in disease risk (32). This is achieved by comparing the value at
one location with values at nearby locations, up to a specified threshold distance (32).

Figure 2.1 Spatial correlation patterns

Positive correlation None Negative correlation

Clustered Random Dispersed
Moran’s I ~ 1 Moran’s I = 0 Moran’s I ~ -1
Geary’s C ~ 0 Geary’s C = 1 Geary’s C ~ 2

Notes: Modified from Figure 3.6 in Lai et al. (33).

2.2 Defining the neighbourhood

One way of accounting for spatial correlation in the data is by defining a neighbourhood as
part of the model. A neighbourhood is composed of surrounding areas that are considered to
exert influence on the observations of an area (12).

The definition of an area-based neighbour may be based on spatial adjacency, such as those
sharing a boundary, or instead may be based on the distance between the centroids (14). Here,
if the distance between two area centroids is below a certain threshold distance, they are
considered to be neighbours. Ways to measure the distance between the centroids include
straight-line distances (the shortest distance between the two coordinates assuming they are
on a flat surface), great circle distances (determining the length of the arc of the earth’s
surface between the two points), or using a Geographic Information System (GIS) to
calculate travel distances or times (12).

Note that when there is great variation in the size of the areas, determining a suitable
threshold distance value is difficult. Even just allowing for the largest areas to have at least
one neighbour may result in far too many neighbours for smaller areas (34). Options for
overcoming this problem include assigning a fixed number of neighbours for each area (k-
nearest neighbours) (Box 2.3).

8
Box 2.3 Weighting mechanism examples (12, 35, 36)

Distance-based weights
1 if 𝑑𝑖𝑗 < 𝛿
𝑤𝑖𝑗 = {
0 otherwise

where 𝑑𝑖𝑗 is the distance between the centroids of regions 𝑖 and 𝑗, and 𝛿 is a given critical
value.

Distance-based simultaneously weighted by population

𝑒𝑖 𝑒𝑗 /𝑑𝑖𝑗 if 𝑑𝑖𝑗 < 𝛿
𝑤𝑖𝑗 = {
0 otherwise

where 𝑒𝑖 is the standardised population for an area and 𝑒𝑗 is the standardised population for
its neighbour in area 𝑗. These can be standardised against the mean and standard deviation.

Distance-based simultaneously weighted by distance

E.g. Based on the inverse distance function
𝑑𝑖𝑗 −𝛾 if 𝑑𝑖𝑗 < 𝛿
𝑤𝑖𝑗 = {
0 otherwise

where the parameter 𝛾 specifies the declining rate of the weight, and can be set a priori or
estimated. Common choices for 𝛾 are the values of one or two.

k-nearest neighbours (Note that 𝑤𝑖𝑗 might not be equal to 𝑤𝑗𝑖 )

1 if centroid of 𝑗 is one of the 𝑘 nearest to centroid 𝑖
𝑤𝑖𝑗 = {
0 otherwise

Adjacency-based neighbours
1 if regions 𝑖 and 𝑗 share a boundary
𝑤𝑖𝑗 = {
0 otherwise

Adjacency-based weighted by the fraction of a shared border

𝑙𝑖𝑗
𝑤𝑖𝑗 = { 𝑙𝑖 if regions 𝑖 and 𝑗 share a boundary
0 otherwise

where 𝑙𝑖𝑗 is the length of shared common boundary between regions 𝑖 and 𝑗, and 𝑙𝑖 is the
perimeter of region 𝑖.

9
Part of assigning neighbours involves applying a measure of weighting to indicate the extent
to which the information from an area’s neighbours impacts on the observed estimate for that
area. A weight of zero indicates no relationship, while a weight above zero indicates that
areas 𝑖 and 𝑗 are considered to be neighbours, and some influence is expected. The weights
(𝑤𝑖𝑗 ) are placed into a matrix with dimensions of the number of areas. When calculating the
similarity with nearby regions, the diagonal (𝑤𝑖𝑖 ) is generally set to 0 as an area is not
considered to be a neighbour of itself. (This differs from situations when the proportion is
averaged over areas (see Section 2.4). Here, data in the area should be included, so 𝑤𝑖𝑖 = 1
(37).) The greater the weight, the more resistant they are to their neighbour’s influence (35).

The derivation of this weight can be based on a range of options (Box 2.3). The weighting
can also be modified according to distance (the weight decreases for more distant
neighbours), or based on the population size of neighbours (larger populations receive greater
weight). Often if little is known about the assumed spatial pattern, a binary weighting is
assigned with 1 for neighbours and 0 otherwise (35) This is often then standardised by
dividing by the number of neighbours so that the rows sum to 1. When areas are irregularly
shaped, this standardised weight matrix is generally not symmetric (14). As each neighbour
receives the same proportional weight, interpretation is simple as it becomes a weighted
average of neighbouring values (12).

The same spatial arrangement can lead to many different neighbourhood definitions. Key
considerations when selecting an approach to assigning neighbours include whether areas are
regular or irregular shapes and sizes and how localised spatial dependencies are. Given the
influence a neighbourhood structure can exert on a spatial analysis, checking the
appropriateness of this choice and its impact on the conclusions is important (12, 38).

2.3 Unsmoothed estimates

The simplest of all techniques for generating values for small-areas are to calculate and map
unsmoothed, or ‘raw’ estimates.

Counts may be displayed as dots on a map, randomly allocated within the area supplied.
Although mapping the counts can be useful and appropriate if the aim is to inform service
provision, often understanding the disease risk is of interest, and this requires some form of
adjustment for population size and structure. Commonly this is achieved by using rates as a
reflection of risk (14).

There are several types of rates commonly calculated. Crude rates adjust only for population
size, but not structure (Box 2.4). Proportions and percentages are other commonly used
measures for a crude rate. The assumption is that the risk remains constant over all age and
sex categories (37), but most diseases disproportionately affect specific age groups (14).
When comparing crude rates between areas, observed differences for a disease that varies
with age may reflect differences in the age distribution alone.

10
Box 2.4 Crude rates

A crude rate for the 𝑖 th area (i=1,…I) can be calculated as:

𝑂𝑖
CR 𝑖 =
𝑃𝑖

where 𝑂𝑖 are the observed counts in area 𝑖 and 𝑃𝑖 are the number of people residing in area 𝑖.
Commonly small crude rates are multiplied by a constant and expressed as per constant (e.g.
per 1,000).

An alternative approach, which also adjusts for population structure, is to calculate age-
standardised rates by considering the counts of disease and the expected counts using some
standard population (39). These may be either directly or indirectly standardised rates. It is
also possible to further adjust for specific area-level variables, such as socioeconomic status.

Box 2.5 Directly age-standardised rates

Directly standardised rates represent the rate these areas would have if their age distribution
matched that of the standard population (14).

Excluding sex for simplicity, and assuming the 𝑚 age groups (m=1,…M, e.g. M=18 five-year
age-groups (0-4, 5-9,…, 85+)), the directly age-standardised rate (DSR) for the 𝑖 th area
(i=1,…I) can be calculated as (40):

𝑀
𝑂𝑖𝑚
DSR 𝑖 = ∑ 𝜋𝑚
𝑃𝑖𝑚
𝑚=1

where 𝜋𝑚 is the proportion of people in age group 𝑚 from the standard population, 𝑂𝑖𝑚 are
the observed disease counts (number of cases for incidence, number of deaths for mortality)
in area 𝑖 and age group 𝑚, and 𝑃𝑖𝑚 are the number of residents in area 𝑖 and age group 𝑚.

Direct standardisation focuses on estimating the number of cases/deaths that would be

observed in the standard population if the observed age-specific rates of disease applied (41).
This is achieved by weighting the age and sex-specific rates for each small area so they
correspond to the age distribution of a single standard population (Box 2.5). This method
enables comparison between areas, but does require age (and sex) specific counts and
populations for each area, which may not be available, or may be very unstable (5, 14). Also,
the standard population is arbitrarily defined, and estimates may differ substantially between
different standard population definitions (42).

11
In contrast, indirect standardisation focuses on estimating the number of cases/deaths that
would be expected if the study population contracted/died from the disease at the same rate as
the standard population (41). Indirectly standardised rates thus multiply the stratified
population of each small area by the known stratified disease rates of some reference
population (Box 2.6). This process produces a standardised morbidity (if using incidence
data) or mortality (if using death data) ratio (SMR). This estimator is very popular, and only
requires the population at risk in each age-sex group and area, as well as the total counts in
each area. It also has a lower standard error (33), and is useful for small areas with unstable
rates (5). However, in contrast to directly standardised rates, weights differ for each area
considered, and bias can potentially result if the age-distributions differ between the areas
being compared (43). Therefore indirectly standardised rates tend to not be directly
comparable between different geographical regions (14).

Box 2.6 The Standardised Morbidity/Mortality Ratio (SMR)

Indirect standardisation reflects whether the number of cases in an area are higher or lower
than expected, given the population size and structure for that area.

The definition of an SMR for the 𝑖 th area (i=1,…I) is:

𝑂𝑖
SMR 𝑖 =
𝐸𝑖

where 𝑂𝑖 are the observed counts in area 𝑖 and 𝐸𝑖 are the expected counts in area 𝑖, applying
the overall (reference) disease rates to the age-specific population structure 𝑃 of area 𝑖 and
summing across all 𝑚 age groups (with a maximum of M age groups), and excluding sex, as
follows:

𝑀
𝑂ref𝑚
𝐸𝑖 = ∑ × 𝑃𝑖𝑚
𝑃ref𝑚
𝑚=1

Where 𝑂ref𝑚 represents the observed disease counts (either cases or deaths) in the reference
population in age group 𝑚, and 𝑃ref𝑚 is the reference population in age group 𝑚.

√𝑂𝑖
The corresponding standard error for the 𝑖 th area is: 𝐸𝑖

These simple estimates are commonly displayed using a choropleth map. The
colouring/shading of areas in a choropleth map uses a discrete scale based on the values of
the estimate. Any kind of choropleth map implicitly smooths the display of results, and the
fewer the number of categories, the greater this visual smoothing (44). However, the
assumption of spatial independence inherent in choropleth mapping could be misleading, and
caution in using with ‘raw’ (unsmoothed) estimates is advised (45).

12
Further, as area size diminishes, the use and interpretation of these unsmoothed estimates
become increasingly difficult due to the greater variance (Box 2.7) associated with them (46).
These estimates can also be prone to substantial fluctuation from year to year without there
necessarily being a change in the underlying rate for a specific area. Other concerns include
that when an area has no counts, the estimate is zero, regardless of denominator size (47). As
such, these ‘unsmoothed’ methods are perhaps most useful for preliminary investigation to
guide further analyses, rather than being an end in themselves.

Box 2.7 Variance and covariance

Variance can be defined as:

𝑛
1
var(𝑋) = 𝜎 = ∑(𝑋𝑖 − 𝑋̅)2
2
𝑛
𝑖=1

In other words, variance can be visualised as lines describing how far away each observation
is from the mean on a scatter plot.

Covariance is a measure of how much two variables change together, so can be defined as:

𝑛
1
cov(𝑋, 𝑌) = ∑(𝑋𝑖 − 𝑋̅) (𝑌𝑖 − 𝑌̅)
𝑛
𝑖=1

Covariance can be visualised as rectangles describing how far away pairs of observations are
from the mean on a scatter plot.

2.4 Direct smoothing

The objective of disease mapping is to produce an accurate estimate of the underlying rate in
different areas, with noise removed (48). This ‘noise’ is simply additional variation in the
data, and a major source is often unmeasured variables that affect the outcome (49).
Smoothing methods aim to remove or minimise this noise by incorporating neighbouring
information in a flexible way (40). When information from geographical neighbours are
included, the information for the region is artificially inflated. This provides greater stability
for the specific area as well as between areas.

2.4.1 Locally-weighted average/median

A straightforward smoothing method is averaging the values associated with neighbouring

areas. First, the neighbours and appropriate weights must be selected (see Section 2.2).

13
Different forms of smoothing result from different weighting choices (14). If binary weights
of zero and one are used, then any difference in the precision of the rates is ignored.
Weighting neighbours by their population is one way to incorporate the precision (14).

Although the measures weighted are usually based on the mean of either the crude rate or the
SMR, the resulting sensitivity to extreme outliers has led to extensions based on the median
instead (37). To allow for differences in precision, the median crude rate could be weighted
by the population, while the median SMR could be weighted by the inverse standard error
(37). Here the original values are sorted then matched with both the weight (e.g. population
size) and a cumulative sum of the weights (50). Whichever weight has a cumulative sum of
more than half the total cumulative sum is used (50).

Despite the simplicity and range of options available for these locally-weighted methods,
there are some key disadvantages. Firstly, the number and regularity of locations may
influence the efficiency of the algorithm (5). There is also a risk that because neighbouring
areas are “forced” to have some numerical association with each other, this method may
induce spatial structure, even when the data are completely random (51).

2.4.2 Kernel smoothers

A general, non-parametric approach to smoothing rates while differentially weighting

neighbours is to use a two-dimensional kernel function (37). A kernel function decreases with
increasing distance (distance decay function), and the rate and range of decay is modified by
the functional form of the kernel, as well as the threshold beyond which the kernel is set to
zero (bandwidth) (37).

Although more commonly applied in geostatistical (point) data, kernel smoothers can be
applied to areal data at a specified moving window size (25). A range of kernel smoothers
exist, but one of the most commonly used is the Nadaraya-Watson kernel estimator (Box 2.8)
(52, 53). Although this is a weighted average of the neighbourhood observations, the weights
are controlled by the kernel function, so all observations are not treated equally (49).

Depending on the application, disadvantages of kernel smoothing can include estimates

resulting in different totals across areas than in the original data (45). Boundary effects cause
problems for kernel smoothers (49), while the use of cross-validation has been shown to
induce over-smoothing (47). A comparison of several smoothing methods concluded that
kernel smoothers performed poorly when spatial correlation was present, and suggested only
using them for exploratory data analysis (47).

14
Box 2.8 The Nadaraya-Watson kernel estimator (47, 49, 52, 53)

This kernel smoother is simply a weighted average, so if applied to an indirectly standardised

ratio such as an SMR:

𝜃̂𝑖 = ∑ 𝑤𝑖𝑗 SMR𝑗

𝑗≠𝑖

The weights are functions of neighbouring values:

𝐾((SMR 𝑖 − SMR𝑗 )/ℎ)

𝑤𝑖𝑗 =
∑𝑗 𝐾((SMR 𝑖 − SMR𝑗 )/ℎ)

The function 𝐾(∙) is the kernel function: a smooth probability density function symmetric
−1
around 0 and nondecreasing on [ , 0] and ℎ is the bandwidth and selected by minimising
2
some goodness-of-fit measure, such as cross-validation (where data are partitioned into
testing and validation sub-samples to see how it generalises to an independent dataset).

2.5 Model-based smoothing

Two standard paradigms for statistical models are:

1. The sampling model, where population characteristics are estimated from a sample or
subset of the population, and
2. The measurement error model, where the focus is on estimating an underlying pattern,
but the data are measured with error. This model is also applicable when complete
data are observed (54).

Although these differ, in practice both approaches can be combined (54). Most models
discussed in this section are measurement error models. Section 4.3 focuses on sampling
models in the context of survey data.

2.5.1 Foundational approaches

In public health applications, the most widely used regression models are the generalised
linear models (GLMs) and the generalised linear mixed models (GLMMs) (Box 2.9),
particularly using the Poisson distribution for count data (14).

The Poisson model is appropriate when there are low disease counts and comparatively large
populations in each small area (55). The counts are assumed to follow a Poisson distribution

15
which defines the mean and variance. The mean is dependent on two components: 1) the
expected count, generally obtained through indirect standardisation, and 2) the excess risk,
which is the SMR, also often referred to as the relative risk in this context (Box 2.10).

Box 2.9 GLMs and GLMMs

A generalised linear model (GLM) involves:

◦ A data vector 𝑂 = (𝑂1 , 𝑂2 , … , 𝑂𝐼 )
◦ Predictors 𝑋 and coefficients 𝛽, to give the linear predictor 𝑋𝛽
◦ A link function g, that links the linear predictor to a nonlinear transformation of the
expected response (e.g. logarithm)
◦ An assumed data distribution (e.g. Poisson)
◦ Potentially other parameters involved in the predictors, link function and data
distribution, such as variances, overdispersions and cutpoints (54).

A generalised linear mixed model (GLMM) involves the same components as the GLM, with
the addition of random effects. The ‘mixed’ in the name thus refers to the model containing
both fixed and random effect terms. Defining whether a term is fixed or random is seldom
straightforward, but perhaps the cleanest approach is that fixed effects are constant if they are
identical for all groups, while random effects are allowed to vary between groups (56).

The binomial distribution is sometimes preferred for small areas due to the Poisson
distribution having some probability of obtaining more counts than persons at risk in each
area (14). However, this is extremely unlikely for rare diseases, where the practical difference
between Poisson and binomial distributions is negligible (14). The Poisson distribution also
constrains the mean to be equal to the variance, but as small areas often have a variance
greater than the mean (termed over-dispersion), some prefer to use the negative binomial
distribution instead. However, the use of a fully Bayesian formulation where a prior
distribution is placed on the SMR/relative risk can accommodate some over-dispersion (55).

Box 2.10 The Poisson model

The disease counts 𝑂 in each of i areas (i=1,…I) is assumed to have a mean dependent on the
expected count 𝐸𝑖 and the SMR/relative risk 𝜃𝑖 as follows:
𝑂𝑖 ~Poisson(𝐸𝑖 𝜃𝑖 )

The main interest is normally in modelling 𝜽𝒊 in the ith area. A logarithmic link is often
assumed (which ensures estimates are positive) to a linear predictor model (Box 2.9), as
follows:
log 𝜃𝑖 = 𝜂𝑖 = 𝑋𝑖 𝛽

16
 2.5.2 Poisson Kriging

Kriging was originally developed to estimate attribute values from a limited set of sampled
data over a continuous spatial region (57). The weights used in kriging incorporate distance
measures, as well as spatial correlation (58). Although this increases the complexity, it also
increases the flexibility of the method and the reliability of predictions (59). Areal disease
mapping often doesn’t require the interpolative ability of kriging, but a specific variant of
kriging known as Poisson kriging has been developed and is becoming increasingly used for
disease maps (57, 60).

Box 2.11 The Semivariogram (5)

The semivariogram displays the semivariance, which is a measure of the level of spatial
correlation, against distance or lag (Figure 2.2).

Figure 2.2 Semivariogram components

Sill
Nugget

Range

Distance/Lag

Note: Modified from the Figure in Explanation Box 6.1 in (5), page 107.

The distance at which the model first levels out is called the range. Areas separated by
distances greater than this are not considered to influence each other, whereas areas separated
by distances within the range are spatially correlated.

17
There are several variants of Poisson kriging, but as we are exploring methods appropriate for
areal data, our focus is on area-to-area (ATA) and area-to-point (ATP) Poisson kriging.
Goovaerts (61) extended the work of Kyriakidis (45) to introduce these. While ATA kriging
can be used when both the observations and the desired predictions are over areas, ATP
kriging predicts point values from areal data (45). An alternative is to simply collapse the
data on the centroids, but this is not considered appropriate when the shape and/or size of
areas are irregular (62).

In both ATA and ATP Poisson kriging, the risk over an area is estimated as a weighted linear
combination of the rate observed for that area and neighbouring areas (61). Areas with
smaller populations receive less weight. These weights are solved from a system of linear
equations, but requires either the point-support covariance of the risk, or the equivalent point-
support semivariogram (Box 2.11), to be modelled. This point-support model is where the
spatial correlation is included.

Developing a semivariogram structure that accounts for irregularly shaped areas and varying
distributions is relatively complex, and an iterative procedure is recommended (62).
Approaches range from solving a set of integral equations (63), to iteratively re-weighted
generalised least squares methods (64), to simulated grids within the regions of interest (62).

While kriging is a useful filter of noise, and produces uncertainty estimates, it is not designed
to estimate the risk within each area (51). Nonetheless, a comparison against the popular
Bayesian Besag, York & Mollié (BYM) model (see Box 2.14) found that Poisson kriging
gave better discrimination between areas with high and low risks, and more precise and
accurate probability intervals (65). Poisson kriging was also found to out-perform simple
population-weighted averages and empirical Bayesian smoothers (see Section 2.5.3) (51).

2.5.3 Empirical Bayes

Bayesian methods differ from other statistical approaches as they consider both the data and
the parameters to be random variables (37). Inference under a Bayesian approach requires
specific items (Box 2.12), with the most controversial element being the selection of the prior
distribution.

Some practical suggestions when selecting a prior distribution include graphing it, to ensure
the shape is plausible, as well as potentially calculating the effective sample size of the prior
(Box 2.13) (66). Choosing particular families of prior distributions (called ‘conjugate priors’)
may assist in solving the posterior distribution without resorting to complicated integrations.
For further details on the Bayesian approach, refer to Appendix B.

Empirical Bayes (EB) methods use the data to estimate the unknown information on the prior
and conditional distributions (67, 68). Spatial disease patterns were initially explored using
EB methods by Clayton and Kaldor (69), Cressie and Read (70) and Cressie (46).

18
Box 2.12 Bayesian inference

To make inference about the unknown parameter 𝜃 from the data 𝑂 requires the following
(71):
1. A model 𝑓(𝑂|𝜃); the likelihood
2. A distribution for 𝜃. This is called a prior distribution as it is determined before seeing
the data.

The combination of the likelihood and prior distribution(s) by Bayes’ rule gives the posterior
distribution:
Posterior ∝ Prior × Likelihood
from which subsequent model-based inferences are drawn.

Empirical Bayes predictors have attractive statistical properties, provided the model is
appropriate, including being the best linear unbiased predictor (BLUP) (46). The associated
uncertainty for each estimate is also available. Ironically, the greatest criticism of EB models
are against the uncertainty measures. Since they do not account for the additional variability
in estimating the parameter values, the resulting variance tends to be too small (68, 72-74).
Although methods are available to adjust the variance estimates (68, 75), fully Bayesian
methods have many of the same desirable properties as an EB estimate, while adequately
representing the distribution of underlying rates (72).

Box 2.13 Calculating the effective sample size of the conjugate prior distribution (66)

The conjugate distribution to the Poisson is the gamma. Say the gamma distribution is
𝑚2 𝑚
expressed as gamma(𝑟, 𝑣) where 𝑟 = and 𝑣 = 𝑠2 where 𝑚 is the prior mean and 𝑠 is the
𝑠2
prior standard deviation, and that 𝑂1 , … 𝑂𝑛 is a random sample of observed counts from a
𝜇
Poisson distribution, Poisson(𝜇), so that the expected value of 𝑂 has mean 𝜇 and variance 𝑛.

To check the amount of prior information entering the model, the equivalent sample size can
be calculated by solving the following for 𝑛𝐸𝑆𝑆 :

𝜇 𝑟
=
𝑛𝐸𝑆𝑆 𝑣2

𝑟
If the mean is set equal to the prior mean, i.e. 𝜇 = 𝑣 then under the gamma(𝑟, 𝑣) prior 𝑛𝐸𝑆𝑆 =
𝑣. This value represents the size of a random sample from the Poisson(𝜇) that is equivalent
to your prior knowledge of 𝜇. If 𝑛𝐸𝑆𝑆 seems too large, increase the prior standard deviation 𝑠
and recalculate.

19
 2.5.4 Fully Bayesian

In contrast to Empirical Bayes where some unknown parameters are assigned a point
estimate, a fully Bayesian model results when prior distributions are placed on all unknown
parameters. Beyond simple models with few parameters, this results in a hierarchical model
structure, where each layer defines a relationship between the observed data and/or unknown
parameters. This general class of model is commonly referred to as a Bayesian hierarchical
model.

There are several advantages to using Bayesian hierarchical models, including the ability to
structure very complicated models from a succession of relatively simple components (76),
good performance and ease of implementation (72, 77-79). They are also a natural approach
to model spatially misaligned data, as occurs when the exposure and response are measured
at different levels of aggregation (80).

The fully Bayesian approach enables complex, realistic models to be developed with reliable
disease rates in low population areas, clear summaries of spatial and temporal correlation,
precise and easily interpretable confidence intervals, and more comprehensive accounting of
sources of uncertainty (77). The Bayesian approach also has excellent flexibility in handling
changing inferential goals, such as obtaining smoothed risk maps as well as identifying
motivating predictors of disease such as ethnicity or socioeconomic status (77).

2.5.4.1 Incidence/mortality data

The most popular Bayesian hierarchical model for disease mapping is the BYM model (Box
2.14) (81), which further developed the model of Clayton and Kaldor (69). The key feature
of this model are the two random effects: one which is spatially structured, so smooths locally
(towards the values of nearby areas), and one which is unstructured, so smooths globally,
towards the overall average (82, 83).

Box 2.14 The BYM model

The BYM model can be expressed as follows:

𝑂𝑖 ~Poisson(𝐸𝑖 𝜃𝑖 )

log(𝜃𝑖 ) = 𝛼 + 𝑢𝑖 + 𝑣𝑖
where 𝑂𝑖 is the number of disease events in the ith region, 𝐸𝑖 is the expected number of cases,
𝜃𝑖 is the standardised incidence ratio and 𝛼 is the intercept. The model incorporates extra-
Poisson variability by including two spatial random effects: 𝑣𝑖 allows for inter-area
heterogeneity, while 𝑢𝑖 is structured and represents the spatial component (84).

20
Commonly, the prior distributions placed on the structured spatial component is a
Conditional Autoregressive (CAR) distribution. The CAR prior is characterised by an
adjacency matrix, which defines the geographical neighbours of each area. The intrinsic
Gaussian CAR prior (81) assumes this matrix is binary, where immediately adjacent
neighbours are given the value 1, and all other pairs of areas are given the value 0. In this
case, the estimated random effects for each area are smoothed towards the average of the
random effects for the neighbours. The resulting simple functions of the neighbouring values
and number of neighbours, 𝑛𝑖 , equates to the following conditional distribution:

2
𝜔𝑢
𝑢𝑖 |𝐮−𝐢 ~Normal (𝜇̅𝑖 , ) where 𝜇̅𝑖 is the average of the neighboring regions of area i and the
𝑛𝑖
variance term 𝜔𝑢2 represents a conditional variance (so is deliberately not portrayed as 𝜎𝑢2 )
𝐮−𝐢 = (𝑢1 , … 𝑢𝑖−1 , 𝑢𝑖+1 , … , 𝑢𝐼 )

1
This can be expressed jointly as 𝐮~Normal𝐼 (𝟎, 𝜔2 (𝐃 − 𝐖)−1 )
𝑢
where 𝐃 is a diagonal matrix with 𝐷𝑖𝑖 = 𝑛𝑖 , 𝐖 is a spatial weight matrix of dimension 𝑁 × 𝑁
with diagonal elements 𝑤𝑖𝑖 = 0 and off-diagonal elements 𝑤𝑖𝑗 = 1 if regions 𝑖 and 𝑗 share a
boundary, and 0 otherwise.

The intrinsic CAR distribution is restricted to specifying prior distributions, as the pairwise
difference joint specification results in an improper joint distribution. The computational ease
is a key advantage of the intrinsic CAR formulation (85, 86).

The inter-area heterogeneity effect commonly has a vague normal prior distribution:
𝑣𝑖 ~Normal(0, 𝜎𝑣2 ) or expressed jointly as 𝐯~Normal𝐼 (0, 𝜎𝑣2 𝐈) where 𝐈 is the identity
matrix (diagonals are set to 1, 0 otherwise).

Under the Bayesian hierarchical formulation, the variance components for both the CAR and
the normal distributions will also receive prior distributions (termed ‘hyperpriors’).

Common choices include a vague gamma distribution on the inverse variance, or a uniform
distribution on the standard deviation, e.g.
𝜎𝑢 ~Uniform(0,10)
𝜎𝑣 ~Uniform(0,10)

The BYM model has been shown to produce robust estimates, but results may be sensitive to
the choice of priors, particularly the choice of hyperpriors (47, 65, 87). The intrinsic Gaussian
CAR prior results in a spatially smooth risk surface, which has the advantage of using
information from multiple areas to estimate the random effects, but is not ideal if the aim is to
identify clusters of high-risk areas (87). This is because a cluster of high risk areas may have
low-risk neighbours, and therefore the estimated risk for these areas becomes less distinct
when geographical smoothing is used (88). Identifiability is also a concern, as the one
residual component is split into two independent, additive components (89). Very sparse data,

21
as would be seen for smaller geographical areas such as Statistical Area 2 (SA2) or SA1, may
cause difficulties when applying these models, particularly if there are also none or very few
neighbours, such as on a coastline or an island (90). Finally, the spatial correlation may
inflate the variance of the 𝛽 components when covariates are included (91, 92).

Alternative approaches have suggested modifications to try to overcome the issues with
identifiability. The Leroux CAR prior (93) (Box 2.15) outperformed the BYM model in a
comparison of methods (94). MacNab’s alternative convolution prior (95) (Box 2.15) also
overcomes identifiability issues, at the cost of greater model complexity (89).

Box 2.15 Alternative priors to the BYM

The Leroux prior (93)

Instead of the 𝑣𝑖 + 𝑢𝑖 components in the BYM model, the Leroux prior has the one term
modelled by the multivariate normal prior, as follows:

𝐛 ~ Normal𝐼 (𝟎, Σ𝑏 )

1 1
= 2 (𝜆(𝐃𝑢 − 𝐖) + (1 − 𝜆)𝐈𝐼 )
Σ𝑏 𝜎𝑏

Where 𝝀 is between 0 and 1, and referred to as the spatial correlation parameter, as it reflects
the proportion of excess Poisson variation explained by spatial dependencies, D is a diagonal
matrix with 𝑫𝒊𝒊 = 𝒏𝒊 , W is the spatial weight matrix and I is the identity matrix (see Box 2.14
for further details)

MacNab’s alternative convolution prior (95)

This prior facilitates identifiability of the spatial and unstructured random effects. Using the
same notation as above, this can be expressed as:

𝐛 ~ Normal𝐼 (𝟎, Σ𝑏 )

𝜆𝜎𝑏2
Σ𝑏 = + (1 − 𝜆)𝜎𝑏2 𝐈𝐼
(𝐃 − 𝐖)

Allowing discontinuous risks between areas

Other approaches have focused on allowing discrete changes between areas. Some of the
current approaches deal with this issue by defining the adjacency matrix such that the
elements are random quantities to be estimated (96). In this way, boundaries between clusters
of areas can be identified when the adjacency matrix elements for neighbouring pairs are
estimated to be near zero. There are two main problems with this methodology. First, using

22
random quantities in an adjacency matrix of size 𝐼 × 𝐼 (where I represents the number of
areas) means an additional 𝐼 2 parameters need to be estimated, which is usually more
parameters than can be estimated reliably. And secondly, there is no constraint on boundary
segments to enclose an area or cluster of areas (88).

Box 2.16 Lawson and Clark’s model

Here, an intrinsic CAR prior and a difference prior act as a ‘mixture of priors’ (97).

Instead of the 𝑣𝑖 + 𝑢𝑖 components in the BYM model, this model has the following terms:
𝑣𝑖 + 𝑝𝑖 𝑢𝑖 + (1 − 𝑝𝑖 )𝑤𝑖
where 𝒑𝒊 is interpreted as the strength of support for spatial smoothing in the ith area, 𝒖𝒊 and
𝒗𝒊 represent the spatially structured and inter-area heterogeneity terms, respectively, as in
Box 2.14, and 𝑤𝑖 represents the jump component. Note that the above equation defaults to the
BYM model if 𝑝𝑖 = 1, but as 𝑝𝑖 approaches 0, the jump component is preferred.

The prior distributions on 𝒖𝒊 and 𝒗𝒊 are as previously defined (Box 2.14), while the prior on
𝑤𝑖 is intended to measure spatial rates of change in risk. Although a range of options is
possible, the suggested approach was:
1 1
𝐰∝ exp (− ∑ |𝑤𝑖 − 𝑤𝑗 |)
√𝜆 𝜆
𝑖~𝑗

where 𝜆 acts as a constraining term, and 𝑗 represents areas that are neighbours.

As there are only two mixing probabilities, a standard Beta distribution is used,
𝑝𝑖 ~beta(𝛼, 𝛼)
For higher dimensions a Dirichlet prior could be used.

New methodology to address these problems was proposed by Anderson et al. (98). This
approach consists of two stages. In the first stage, a set of candidate cluster configurations
are identified by using what is called a ‘modified hierarchical agglomerative clustering
algorithm’. Initially, each area is considered to be a cluster, and clusters are combined
together sequentially based on how similar they are according to some metric applied to the
spatial data until all areas are combined into one cluster, resulting in a total of 𝐼 cluster
configurations. The term ‘modified’ is used because the usual clustering algorithm does not
necessarily produce spatially contiguous clusters, but this is enforced by only allowing
clusters to be combined if the clusters share a common border. The spatial data used in this
stage should not be the study data used in the second stage of the model, but should be a
similar dataset, such as data on the same disease for a previous time period, data on a similar
disease, or even covariate information. Each clustering configuration has a corresponding
adjacency matrix, with each matrix resulting in a different degree of spatial smoothing. In

23
the second stage, a separate Bayesian hierarchical model is fit to the data for each of the 𝐼
cluster configurations. These 𝐼 models can be compared using model goodness-of-fit
criterion (88, 98).

Anderson et al. (88) refine this methodology by fitting a single model in the second stage
which is capable of estimating the cluster structure and disease risk simultaneously. This is
achieved by specifying the prior for the random effects as a mixture of 𝐼 CAR priors, where
each mixture component has a different adjacency matrix with a corresponding prior weight.
The most appropriate number of clusters can then be selected using the posterior mode or
median. This updated methodology improves computational efficiency because little time is
spent estimating the model parameters for those mixture components which correspond to an
untenable cluster configuration. However, note that these methods have only been applied to
diseases such as chronic obstructive pulmonary disease (COPD), which tends to be more
common than diseases such as cancer.

Alternative approaches include Lawson and Clark’s (99) weighted sum of spatial priors
which has no single global smoothing, so the underlying risk is free to either be smoothed or
to ‘jump’ between areas (Box 2.16). In contrast, several of the semi-parametric mixture
models force the risk surface to be discontinuous, including marginal mixture models and
spatial partition models (99, 100). Although some of the semi-parametric mixture models
allow for a smooth underlying risk, Green and Richardson’s hidden Markov model (101)
smoothed the data more than the BYM model when the data had insufficient evidence to
create a higher-risk group (102). Also, identified clusters might not be spatially contiguous
under this model (88).

Mixture models are prone to less computational stability than the BYM, the risk of label
switching and component identifiability difficulties, as well as requiring greater care in
covariate selection due to their influence on risk label categorisation (103). Often the mixture
models also require greater programming skills, relying on GNU and/or Fortran to run the
models (101, 104).

Identifying clusters

Another research problem concerned with clustering is image segmentation, in which the
goal is to classify image pixels (which represent arbitrary areas defined by a grid) into well-
defined clusters (105). Hidden Potts-Markov random field (MRF) models are commonly
used in Bayesian image segmentation methods (Box 2.17). However, making inferences on
these types of models is difficult, hence current image segmentation methods typically rely
on approximate estimators (106). A major limitation of such approaches is that they are
supervised, meaning that the regularisation parameter of the Potts model must be specified a
priori. Selecting an appropriate regularisation parameter a priori can be difficult since they
can be highly dependent on the image. Unsupervised approaches which self-adjust the
regularisation parameter are currently possible, but at an enormous computational cost (107).

The recently proposed methodology of Pereyra and McLaughlin (107) permits approximate
inference on hidden Potts MRFs which is unsupervised and also computationally fast. The

24
crux of their approach involves dividing the problem into two simpler problems, both of
which can be solved easily and relatively quickly.

Box 2.17 Hidden Potts-Markov random field model

Let 𝑦𝑖 is an element of 𝒚 (𝑦𝑖 ∈ 𝒚) be observations with latent labels 𝑧𝑖 ∈ 𝒛. For example, 𝑦𝑖

might represent the intensity of the 𝑖 th pixel in a greyscale image, while 𝑧𝑖 identifies a
segment of the image to which the 𝑖 th pixel belongs, for a finite set of segments, {1, … , 𝐾}.
Given the segment identifier 𝑧𝑖 , the intensity of the pixels in that segment will be similar. For
example, if the intensities were assumed to be Gaussian, pixels in the 𝑘 th segment might have
the same mean and variance,
𝑝(𝑦𝑖 |𝑧𝑖 = 𝑘) = Normal(𝜇𝑘 , 𝜎𝑘2 ).

The unobserved random variables 𝒁 = {𝑍𝑖 } represent nodes in a hidden Markov random
field, each node corresponding to a pixel 𝑦𝑖 . For each 𝑍𝑖 , a neighbourhood is defined such
that 𝑍𝑖 only depends on those nodes in the neighbourhood, and is conditionally independent
of other nodes (the Markov property):

𝑝(𝑧𝑖 |𝒛\𝑖 ) = 𝑝(𝑧𝑖 |𝒛𝑗 , 𝑗~𝑖),

where 𝒛\𝑖 denotes all values of 𝒛 except 𝑧𝑖 , and 𝑗~𝑖 denotes nodes 𝑖 and 𝑗 are in the same
clique. The Hammersley-Clifford theorem states that this conditional probability distribution
has the form:
1
𝑝(𝑧𝑖 |𝒛𝑗 , 𝑗~𝑖) = exp(−𝛽𝐻(𝑧𝑖 ))
𝑊𝑖

where 𝑊𝑖 is the partition function, 𝛽 is the regularisation parameter, and 𝐻(𝑧𝑖 ) is the energy
function of 𝑧𝑖 . In the case of the Potts MRF model, the energy function is chosen to be
∑𝑖~𝑗 𝕝(𝑧𝑖 = 𝑧𝑗 ), leading to

exp{−𝛽 ∑𝑖~𝑗 𝕝(𝑧𝑖 = 𝑧𝑗 )}

𝑝(𝑧𝑖 |𝒛𝑗 , 𝑗~𝑖) = .
∑𝐾
𝑘=1 exp{−𝛽 ∑𝑖~𝑗 𝕝(𝑧𝑗 = 𝑘)}

Pereyra and McLaughlin (107) compare the results obtained from this method against four
state-of-the-art supervised segmentation algorithms, and one unsupervised MCMC algorithm,
each applied to three different datasets of varying complexity. Visually, the resulting image
segmentation of the proposed method is comparable to all five other methods. In terms of
computational efficiency, the proposed method is at least one thousand times faster than the
unsupervised MCMC algorithm, and only two or three times slower than the supervised
algorithms. Only two or three clusters were used in these tests, so how well this proposed
method works for larger numbers of clusters is yet to be quantified. However, these results
indicate that the proposed method is very promising.

25
Change of Support

One limitation that persists in spatial modelling is the difficulty in making statistical
inference on spatial support points which differ to the support points provided by the data.
For example, data may be collected for each SA1, and predictions/estimates of a particular
variable can be obtained from an appropriate model for these areas quite easily. But making
inferences for postal areas, for example, or some arbitrarily defined area is not so
straightforward. This is known as a change-of-support problem. This problem may arise
when the desired support points for inference cannot be foreseen or constrained to one
support type at the time of modelling, or when the support points are limited by the data
available (108).

Current methods have provided solutions to this problem when the underlying data is
assumed to be Gaussian. Yet disease mapping studies often have count data and are typically
modelled by a Poisson or Binomial distribution. Even for this type of data, methods have
been developed, such as simple areal interpolation (109), whereby inferences may be made
on target support points by imputing values from surrounding data support points. However,
when the data are not recorded without error, such as with survey data, the uncertainty of the
estimates at target support points is unknown which limits their inferential usefulness (108).

Bradley et al. (108) proposed new methodology to address these problems by incorporating
the estimated variance of the data in the model. The areal count data are interpreted as an
aggregation of events from a latent, unobserved, spatial point process. This latent process is
modelled using a Bayesian hierarchical GLMM. Specifically, the latent spatial process is
modelled as a combination of additive covariate and spatial basis function effects, and the
observed count data, conditional on the latent spatial process, are modelled by a Poisson
distribution. Estimates of the variances of the observed data values are modelled jointly with
the observed data. By estimating the model parameters at the point level, estimates can be
obtained for any desired support point by aggregating the latent process. Including variance
estimates in the model is not necessary, but doing so provides more precise estimates.

2.5.4.2 Survival data

Survival can be measured in several different ways. All-cause, or overall, survival captures
all deaths regardless of cause. Often net survival is of more relevance, as it aims to capture
only deaths resulting from the disease of interest. Net survival can be approximated by either
cause-specific or relative survival. Information on the recorded cause of death is required for
a cause-specific analysis, whereas for relative survival, deaths due to any cause among cancer
patients are compared against background population mortality rates.

The Cox proportional hazards model is the most widely used survival model (110, 111). This
model is applied to either overall or cause-specific survival, and has no assumptions
regarding the nature or shape of the underlying survival distribution. Correlation can be
incorporated between areas by including random effects termed ‘frailties’ (112, 113).

26
However, it does assume a proportional (multiplicative) relationship between the hazard and
the log-linear covariate function (111), and this assumption is often violated (114).

Some Bayesian spatial cancer survival analyses have preferred to use either overall or cause-
specific survival analyses, and have based their models on variants of the Cox proportional
hazards model (115), or parametric models such as accelerated failure-time models (116-
118). Survival for multiple cancers has also been jointly modelled with spatial frailties (119).
However, these either assume accurate cause of death data (if a cause-specific analysis),
which is a key disadvantage for population-based cancer data, while overall survival analyses
may have confounding from unrelated differences in mortality between areas.

There have been a few different types of relative survival models incorporating spatial
components within a fully Bayesian framework. Fairley et al. (120) expanded the additive
hazard model recommended by Dickman et al. (121) to incorporate spatial and unstructured
random effect components similar to the BYM model (Box 2.18).

Box 2.18 Bayesian spatial relative survival model

Fairley et al. (120) introduced the following relative survival model within a fully Bayesian
context:
𝑑𝑖𝑗𝑘 ~ Poisson(𝜇𝑖𝑗𝑘 )
∗
log(𝜇𝑖𝑗𝑘 − 𝑑𝑖𝑗𝑘 ) = log(𝑦𝑖𝑗𝑘 ) + 𝛼𝑗 + x𝑖𝑗𝑘 β𝑘 + 𝑢𝑖 + 𝑣𝑖

where 𝑑𝑖𝑗𝑘 represents the number of deaths resulting from any cause in the ith area, jth follow
∗
up time from diagnosis interval, and kth age group, 𝑦𝑖𝑗𝑘 is the person-time at risk, 𝑑𝑖𝑗𝑘 is the
expected number of deaths due to causes other than the cancer of interest, 𝛼𝑗 is the intercept
(which varies by follow-up year), x is the predictor variable vector (although proportional
excess hazards are assumed, interactions can be accommodated), 𝑢𝑖 is the spatial component
assigned an intrinsic CAR prior and 𝑣𝑖 is the unstructured component with a normal prior
centred on 0.

∗
The term log(𝜇𝑖𝑗𝑘 − 𝑑𝑖𝑗𝑘 ) is a non-standard link function representing the log excess deaths,
or the deaths considered to result from the disease of interest (121). Follow-up intervals can
be of any duration, but often annual intervals are used.

Many of the advantages for this approach are similar to that of the Cox proportional hazards
model (indeed, if using cause-specific or overall survival with time split at each event, the
Poisson piecewise model equates to the Cox proportional hazard model (122)), including no
assumption of the baseline survival shape. However, the disjointed piecewise process is
biologically implausible, and covariates such as age cannot be included as continuous
variables without the model becoming too cumbersome.

27
A similar Poisson piecewise model was combined with Bayesian geoadditive models so the
baseline hazard was modelled using penalized splines (123). This flexible semiparametric
model overcomes many of the limitations of the Poisson piecewise approach and incorporates
a spatial effect, random effects and fixed effects, but is computationally intensive.

An alternative approach is to combine a parametric formulation with splines for flexibility in

modelling the baseline hazard, producing flexible parametric survival models (122). Cramb et
al. (124) extended Nelson’s relative survival version (125) to propose the Bayesian spatial
flexible parametric relative survival model. This approach combines the benefits of flexible
parametric models: the smooth, well-fitting baseline hazard functions and predictive ability,
with the Bayesian benefits of robust and reliable small-area estimates. Both spatially
structured (with an intrinsic CAR prior) and unstructured frailty components are included.
Advantages of this approach include the ease of including additional complexity, the use of
individual-level input data, and the capacity to conduct overall, cause-specific and relative
survival analysis within the same framework (124).

2.6 Computation

The practical aspects of producing small-area estimates is discussed in this section, both in
calculating estimates and available software.

2.6.1 Numerical

Unsmoothed estimates can be calculated in any statistical software package, or in a

spreadsheet.

Direct smoothing approaches such as locally-weighted averages/medians and kernel

smoothing are available in many GIS packages, including commercial packages such as
ArcGIS and MapInfo as well as freely-available programs such as GRASS GIS and GeoDa,
among others (Appendix C).

There are also a range of options for obtaining model-based results, and most statistical
software will perform these. Parameter estimates from GLMs such as simple forms of the
Poisson model (see Box 2.10) can be obtained via best linear unbiased estimator (BLUE)
analyses (Box 2.19). The corresponding approach for GLMMs, composed of both fixed and
random effects (see Box 2.9), is via best linear unbiased predictor (BLUP) estimation (126).
If the variances and covariances of random effects are estimated and used in a BLUP
estimator, then it is referred to as empirical BLUP, or EBLUP (126). Including spatial
structure within the random effects can improve the EBLUP estimator even further,
becoming the spatial EBLUP, or SEBLUP, estimator (127).

Both empirical Bayes and EBLUPs use a similar process of estimation. First, the variance
components are assumed to be known, and BLUPs or EB predictors are obtained for the

28
unknown parameters (128). Then, the variances and covariances are estimated by the method
of fitting constants/moments, or if normality is assumed, then via maximum likelihood (Box
2.19) or restricted maximum likelihood methods (126). For further details on these and
similar methods for computing empirical Bayes estimates, refer to Meza (75).

However, as complexity in the Poisson model increases, alternative methods are often
required, and this can range from approximating the likelihood via ‘quasi’ or ‘pseudo’
likelihoods (14), through to sampling from the posterior distribution of a Bayesian model.

Although the BLUP and Bayes approaches theoretically produce identical point estimates for
small-areas (126), in certain circumstances fully Bayesian estimates were shown to have
smaller MSEs than the corresponding BLUP (129).

Specific software has been developed to enable Poisson kriging estimates to be easily
calculated. Centroid-based Poisson kriging can be calculated using the freely available
poisson-kriging.exe (51), which was written using Fortran 77. BioMedware’s SpaceStat
software (130) is also able to conduct Poisson kriging, including ATP Poisson kriging (in
addition to many tests for spatial correlation). This software replaces the space-time
information system (STIS) (131).

Box 2.19 BLUE, Maximum likelihood and Least Squares

The least squares estimate is the value that minimises the sum of squared errors (54). More
formally, for the model 𝑦𝑖 = 𝑋𝑖 𝛽 + 𝜀𝑖 , the least squares estimate is the 𝛽̂ that minimises
∑𝑛𝑖=1(𝑦𝑖 − 𝑋𝑖 𝛽̂ )2. This is also the best linear unbiased estimator (BLUE) if the variance-
covariance matrix of any linear unbiased estimator 𝛽̃ is greater than or equal to the variance-
covariance matrix of 𝛽̂ (132). If the errors 𝜀𝑖 are independent with equal variance and
normally distributed, then the least squares estimate is also the maximum likelihood estimate
(54).

Under maximum likelihood, the probability of the likelihood (which is the joint distribution)
of all observations is maximised in regards to several relevant parameters (12). Maximum
likelihood estimation has several desirable attributes, including consistency and efficiency, as
well as being able to handle small departures from the normality assumption (133)

Bayesian hierarchical models containing spatially structured components generally cannot be

solved via numeric integration, but an alternative approach that is well suited to these models
was developed in the 1950s, although it wasn’t until the 1990s that this became widely
applied in statistics (134). This method is Markov chain Monte Carlo (MCMC).

29
 2.6.2 Markov chain Monte Carlo

Methods such as approximating large-sample exact solutions (asymptotic approximations),

traditional numerical approaches and non-iterative Monte Carlo methods are likely to either
be infeasible or produce results with low accuracy when applied to complex statistical
models, many of which are Bayesian (135). MCMC methods (Box 2.20) are able to reduce
complex multidimensional problems to a series of lower-dimensional problems, while not
requiring conjugate structure between the likelihood and the prior distribution (135). MCMC
samples from the posterior distribution of Bayesian models and has dramatically expanded
the potential scope of statistical models, thanks to modern computing power (136).

Box 2.20 MCMC

A Markov chain has been described as a frog jumping on a set of lily pads (137). Assuming it
must always land on a lily pad, the probability of jumping onto another (or even the same)
lily pad depends only on the lily pad it is currently on. Likewise, the future behaviour of a
Markov chain is dependent only its present state (137).

Provided the Markov chain has converged, the desired summary of the posterior distribution
is approximated by MCMC, which are simulated random processes conditional on the
previous value. A range of MCMC algorithms are available, but currently the most popular
for disease mapping applications is the Gibbs sampler.

The Gibbs sampler (138, 139) is an algorithm that samples from each of the full conditional
distributions 𝑝(𝜃𝑖 |𝜽𝑗≠𝑖 , 𝒚) in the model. A single new value of 𝜃𝑖 is generated at each
iteration, conditional on all other 𝜃’s, as all proposals are accepted in Gibbs sampling (55).

The Gibbs sampler algorithm proceeds as follows for k parameters, given a set of starting
(0) (0)
values {𝜃1 , … , 𝜃𝑘 } :
(𝑡) (𝑡−1) (𝑡−1) (𝑡−1)
1. Draw 𝜃1 from 𝑝(𝜃1 |𝜃2 , 𝜃3 , … , 𝜃𝑘 , 𝒚)
(𝑡) (𝑡) (𝑡−1) (𝑡−1)
2. Draw 𝜃2 from 𝑝(𝜃2 |𝜃1 , 𝜃3 , … , 𝜃𝑘 , 𝒚)
⋮
(𝑡) (𝑡) (𝑡) (𝑡)
k. Draw 𝜃𝑘 from 𝑝(𝜃𝑘 |𝜃1 , 𝜃2 , … , 𝜃𝑘−1 , 𝒚)

Concerns have been raised with regards to assessing convergence, selecting starting values,
and the length (and necessity) of burn-in periods for MCMC analyses (140). The
computational resources required is perhaps their greatest disadvantage, with alternative
methods seeking to provide good approximations in a drastically reduced timeframe (141).
However, their ability to directly approximate probabilities (142), and answer a broad range
of questions (143) remains unsurpassed.

30
Programs have been developed to assist in conducting MCMC-based analyses, including
BUGS (Bayesian inference using Gibbs sampling) software (91, 144), Stan (which uses
Hamiltonian Monte Carlo) (145), MLwiN (146), JAGS (Just Another Gibbs Sampler) (147)
and the R package MCMCpack (148).

2.6.3 MCMC approximation methods

The computational requirements and time needed to conduct MCMC analyses can be off-
putting to those considering a fully Bayesian analysis. More recently, a range of
approximation methods have become available as an alternative to MCMC, with the benefit
of a reduced computational burden.

The most popular of these within the disease mapping context is integrated nested Laplace
approximation (INLA), and this is available in an R package (www.r-inla.org/). The
approximation is broken down into smaller sub-problems, and a method of approximation
known as Laplace approximation is applied when the densities are near-normal (Box 2.21)
(149). A wide range of models can be approximated by INLA, including most GLMs, and it
has been shown to produce good approximations to output from MCMC for cancer
(simulated and real) data, provided the disease is not incredibly rare (150).

The key advantages of INLA are its speed and flexible model specification (55).
Disadvantages in its current form are the somewhat restricted range of prior distributions and
an inability to handle: models not expressible in log-linear form, mixture distributions, as
well as certain types of missing data/measurement errors (55).

Box 2.21 INLA (149)

Critical assumptions in INLA are that:

1. The number of hyperparameters is small, and does not exceed 20. (Typically this is
between two and five.)
2. The distribution of the latent field is Gaussian. When the dimension is high (104-105),
this is either a Gaussian Markov random field, or close to one.
3. Each observation only depends on one component of the latent field.

2.6.4 Creating the neighbourhood matrix

Several methods are possible to create a neighbourhood matrix. Adjacency-based neighbours

can be assigned using any GIS package, provided the polygon arrangement and relationships
are clean. Sometimes the shapefile will have small artefacts where boundaries do not meet
precisely, which would require intervention, such as ‘snapping’ vertices within a threshold
distance together (151).

31
 Some statistical programs, such as R, offer several options for creating neighbour matrices,
including contiguity or distance-based options (including k-nearest neighbour and threshold
distance). GeoDa also offers a wide range of options for creating and visualising
neighbourhood matrices (Table 2.1). Most packages offer several export options for the
resulting neighbourhood matrix, but it is worth ensuring that an appropriate format for the
software used in further analyses exists.

2.6.5 Software

Tables 2.1 and 2.2 summarise the broad capabilities of common software and specific
examples of software used by method, respectively. Refer to Appendix C for further details
on software.

Table 2.1 Common mapping, GIS and statistical software and capabilities
Analyses
Smoothing Model-based smoothing
Software Type Visualise Neighbour- Spatial Raw Locally- Kernel Spatial Poisson EB HB
maps hood matrix correlation estimates weighted regression kriging

Open source
Bing Maps Map Y
BUGS Stat Y Y Y Y Y
Epi Info Tools Y Y
GeoDa Tools Y Y Y Y Y Y Y Y
GRASS GIS Y
Google Earth Map Y
JAGS Stat Y
NIMBLE Stat Y
PySAL Tools Y Y Y Y Y Y Y Y
R Stat Y Y Y Y Y Y Y Y Y
SaTScan Tools Y Y Y
Stan Stat Y*
Commercial
ArcGIS GIS Y Y Y Y Y Y
MapInfo GIS Y Y Y Y Y
MLwiN Stat Y Y*
SAS Stat Y Y Y Y Y Y* Y Y
S-Plus Stat Y Y Y Y
SpaceStat Tools Y Y Y Y
Stata Stat Y Y Y Y Y Y Y Y
TerrSet GIS Y Y Y Y Y Y
Abbreviations: Stat=Statistical software, Map=Mapping software, GIS=Geographic Information Systems software,
EB=Empirical Bayes, HB=Hierarchical Bayes, Y=Yes.
* Indicates limited functionality, such as lacking programmed CAR prior distributions. Note that often software can interface
with other software to either provide greater functionality (e.g. between statistical packages and GIS software), or to enable
programming within the language of convenience (e.g. Stan and JAGS can interface with R).
Software is considered able to perform a hierarchical Bayes analysis if a random effects term for each area can be modelled.

32
 Table 2.2 Examples of software used by method

Analysis Example of software used

Spatial correlation Global Moran’s I (19) GeoDa (152)

Geary ’s C (20) GeoDa (153)
Tango ’s MEET (23) S+ code in R (154)
Local LISA (21) GeoDa (152)
Spatial scan statistic (24) SaTScan (155)
Unsmoothed Crude/standardised rates Unstated, statistical (43)
estimates
Direct smoothing Locally-weighted STARS (156) – now in PySAL
average/median
Kernel smoother R (157)
Model-based Poisson kriging SpaceStat, ArcGIS (60)
smoothing Empirical Bayes SAS (158)

33
Fully Bayesian – Incidence etc. BYM (81) WinBUGS (159)
Leroux CAR prior (93) WinBUGS (160)
MacNab alternative convolution prior (95) Unstated, BLUE+REML (95)
Anderson spatial pattern & cluster model (88) R (88)
Lawson & Clark mixture model (99) M-H algorithm, self-coded (99)
Spatial partition models (100) MCMC, self-coded (100, 161)
Green & Richardson hidden Markov model (101) MCMC coded in Fortran (101)
Hidden Potts-Markov random field model (107) Self-coded, unstated (107)
Bradley latent spatial process (108) Self-coded, unstated (108)
Fully Bayesian - Survival Cox proportional hazards model SPSS, BUGS (162)
Bayesian spatial relative survival model (piecewise) Stata, WinBUGS (120)
Bayesian spatial flexible parametric relative survival Stata, WinBUGS, MapInfo
model (124)
34
3. Current approaches for small-area cancer estimates
“I often say that when you can measure what you are speaking about,
and express it in numbers, you know something about it; but when you
cannot express it in numbers, your knowledge is of a meagre and
unsatisfactory kind; it may be the beginning of knowledge, but you
have scarcely, in your thoughts, advanced to the stage of science,
whatever the matter may be.”
~ William Thomson, 3 May 1883,
‘Electrical Units of Measurement’ lecture

The data used to generate cancer atlases may relate to cancer incidence, mortality, survival,
or screening data. Methods used to generate published estimates are examined in this chapter.

3.1 Small-area cancer screening estimates

Cancer screening is the application of a test to an apparently cancer-free group to identify

those people likely to have the disease (163). Cancer screening programs involve large
numbers of people. As such, most studies of small-area variation used unsmoothed
percentages, and occasionally additional tests were applied to determine areas that showed
statistically significant evidence of higher/lower outcomes.

Cervical cancer screening was examined in small-areas of Rotterdam, although areas with
<2000 residents were excluded from the analysis to prevent unstable results (164).
Percentages were calculated for uptake of screening, and the association with the proportion
of migrants and specific marital statuses considered.

Cervical cancer screening, breast cancer screening, and bowel cancer screening (faecal occult
blood testing/colonoscopy) were examined across 205 small-areas of Peel region in Ontario,
Canada (165). The average population in each area was 4,000 people (range: 2,500 to 8,000).
Maps were overlaid with the proportion of South Asian people, and also used LISA (166) to
objectively identify areas of extreme variation. The authors noted that they deliberately did
not choose a smaller level of resolution due to several issues including the potential for
unstable rates (165). Another Ontario-based analysis examined the uptake of cancer screening
tests in conjunction with screening for glucose and cholesterol across 18,950 small areas
(167), with funnel plots used to identify abnormal areas falling outside the 95% or 99% CI
for Ontario’s screening rate.

Another descriptive analysis examining a small region of Florida, USA, while not mapping
screening rates, mapped the ethnicity of each small area (expressed proportionally), and
showed the location of colonoscopy services on the map (168).

35
3.2 Small-area cancer incidence/mortality estimates

Compared to the numbers involved in a population screening program, the number of people
who are diagnosed with cancer or who die from cancer in a given time period is relatively
few. Small numbers mean that spatial analyses of cancer incidence and cancer mortality often
require some form of smoothing being employed for small-area studies. Nonetheless,
unsmoothed estimates were mapped for small-area cancer atlases in Canada (169), India
(170), New York (USA) (171), Pennsylvania (USA) (172), South Australia (Australia) (173),
Sweden (174), New Hampshire (USA) (175) and the USA (176). Details on some of these are
available in the associated report: “Grey Literature Review: Internet Published Cancer
Maps”.

Poisson kriging was used to examine cervical cancer mortality rates in 118 counties across
four states in Western USA (177). ATP Poisson kriging was used in another study to examine
age-standardised lung and cervical cancer mortality for two different areas of the USA – one
with 92 counties of reasonably similar shape and size, and another area of 118 counties with
varying size and shape (61). ATA Poisson kriging was used to examine age-standardised
oesophageal cancer incidence over 336 areas in Iran (60), and Poisson kriging has also been
used to examine lung cancer incidence around Perth in Western Australia (57).

Empirical Bayes methods have been used in several small-area cancer incidence/mortality
analyses, including:

 Endemic Burkitt’s lymphoma among children in Kenya. This modelled 272 cases
identified from hospital data between 1999-2004 across 324 regions (178).
 Breast cancer mortality on the island of Sardinia (covering 22 regions across 1983-
1987) (179).
 Pleural cancer mortality was modelled to approximate asbestos exposure in north-
western Italy, across 1,209 areas during 1980-1992 (180). Poisson regression was
then used to check for an association with lung cancer mortality across the areas
(180).
 Lung cancer mortality in Missouri (1972-1981), across 115 areas and 4 age groups
(45-54, 55-64, 65-74, 75+) (181).
 Lung cancer mortality ratios among women in 287 central Italian regions (182).
 Gastric cancer mortality in Hungary investigating an association with nitrate exposure
over 192 settlements with regularly maintained nitrate records. Proxy information was
used to adjust for dietary habits, smoking prevalence and socioeconomic status (183).

Fully Bayesian models have also been employed. Geographical variation in mortality from
haematological tumours (leukaemia, non-Hodgkin’s lymphoma and multiple myeloma)
(184), thyroid cancer (2) and pleural cancer (185) was examined over 8,077 areas in Spain
using the BYM model.

Explanatory covariates are often included in models. An area-level measure of sunlight

exposure was included when modelling lip cancer incidence in Scotland (83). The model

36
used was similar to BYM, but had only one random effect term which was spatially
structured (83). Age, sex, and age-sex interactions were incorporated into a model examining
lung cancer mortality in Missouri (186). Atmospheric pollutants and lung cancer mortality in
Tuscany, Italy, 1995-1999, were modelled using BYM with a nested latent factor model
(187). An exploration of late stage breast and colorectal cancer incidence during 1995-1997
across 87 counties in Minnesota, USA also adjusted for a range of environmental effects
(188).

The inverse distance of each census tract centroid from the nearest hazardous waste site was
included when modelling leukaemia incidence in upstate New York (67). Similarly, the effect
of industrial pollution on lung cancer and lymphohaematopoietic cancers in Northern Italy
was explored (189, 190). Cervical cancer inequalities in stage at diagnosis in the former
German Democratic Republic was modelled to examine disparities in Papanicolaou testing
uptake (191).

Although less common, a few analyses have also used different forms of Bayesian
hierarchical models aimed at enabling disparate changes to be detected. An extension of
hidden Markov models (which can be considered a generalization of a mixture model) was
applied to larynx cancer mortality in France (101). Leukaemia incidence in New York was
modelled using Bayesian spatial partition models (100).

3.3 Small-area cancer survival estimates

Survival measures the proportion of people expected to remain alive for a given length of
time after diagnosis, and the calculations often require individual-level data. Survival is a
useful measure for exploring and comparing the impact of the healthcare system over time
and place (162). When analysing cancer survival estimates across small-areas, it is
recognised that estimates will be unstable if the resolution is too fine (192). As survival
calculations focus on the deaths within a specified time from diagnosis, numbers tend to be
smaller than for either incidence or mortality analyses.

Relatively few small-area survival analyses have been performed, and these have often
utilised a Bayesian approach. An exception is Huang et al.’s (193) analysis of lung cancer
and late-stage colorectal cancer survival across small areas in California (and then a more
detailed analysis of Los Angeles areas). Here the 5-year and 3-year survival estimates were
mapped, but also the adjusted survival time was calculated for each region and then a spatial
scan statistic applied to determine areas with higher/lower survival (193).

Empirical Bayes methods were used to model leukaemia under proportional hazards (194).
Osnes and Aalen expanded a Bayesian Cox proportional hazards model using components
from the BYM model, to explore regional differences in survival for breast cancer and
melanoma patients in Norway (162). Acute myeloid leukaemia was modelled in northwest
England across 24 districts under proportional hazards model using a range of possible
correlation structures (58). Breast cancer survival in France across 377 areas was modelled

37
using Hennerfeind’s flexible continuous time geoadditive model (195), using metastasis as a
proxy for staging information (113).

The Bayesian relative survival models incorporating spatial components are growing in
popularity. Fairley et al. (120) used their Bayesian spatial relative survival model to explore
variation in prostate cancer survival across 44 regions in Northern and Yorkshire England
(average population size was ~ 150,000, ranging from <70,000 to 307,000). This same model
was used to examine geographical variation in breast cancer survival in Catalonia, Spain over
several different area definitions, down to the level of the census tract (average population of
just 604 women aged over 15 years, and a standard deviation of 302) (196). This model was
also applied to cancer data to examine small-area variation across 478 areas of Queensland,
Australia by Cramb et al. (154, 197) and modified forms were used in detailed analyses of
breast cancer relative survival across Queensland by Hsieh et al. (198, 199).

Breast cancer relative survival was examined across north-eastern France using the Bayesian
geoadditive model proposed by Hennerfeind et al. (123), while Cramb et al. (124)
demonstrated their proposed Bayesian spatial flexible parametric relative survival model on
breast, colorectal and lung cancer in Queensland.

3.4 Summary and conclusion

Cancer is a relatively rare disease. When cancer measures are mapped, it is important these
estimates are reliable. For cancer outcomes such as incidence, mortality and survival, most
analyses use some form of smoothing. Models based on GLMMs are often employed, and
these also have the advantages of the ease of incorporating covariates, considering
interactions and examining model fit.

Even for cancer screening data, where numbers are exponentially higher, producing
unsmoothed estimates often constrains the level of resolution possible, or necessitates
excluding some of the areas. Smoothing could potentially be useful for screening data as
well, depending on the level of resolution of areas.

38
4. Further topics in Bayesian models
“The most important questions of life are indeed,
for the most part, really only problems of probability.”
~ Pierre Simon Laplace
Théorie Analytique des Probabilités, 1812

4.1 Inclusion of multiple nested geographies

Models formulated within the Bayesian framework are naturally hierarchical, given their
expression of a statistical model as a series of related layers. For this reason, they are an
appealing choice for the analysis of nested data structures.

The analysis of health outcomes often involves the integration of data from multiple sources,
observed at different scales. In a spatial setting, it is common for these scales to be
embedded in one another – for example, individuals within regions or regions within a state –
resulting in a hierarchical or nested data structure.

Bayesian hierarchical models can be developed to take account of this structure, to (i) allow
for spatial correlation between effects defined at the same spatial scale; and (ii)
relate/compare effects defined at different levels. The latter form of inference can be
achieved through careful consideration of prior distributions and, for the most part, their
specification should be guided by the comparative inferences the analyst wishes to draw
(200). An example of this modelling approach is provided in Box 4.1.

In this example, the defined model allows for three main inferences:
1. The comparison of state estimates (𝜶) relative to the overall average estimate (𝛾).
2. The comparison of statistical division estimates (𝜷𝑘 ) within each state (𝑘 = 1, … 𝐾),
relative to the overall estimate for state 𝑘.
3. The comparison of statistical subdivision estimates (𝜽𝑗𝑘 ) within each statistical
division (𝑗 = 1, … 𝐽𝑘 ), relative to the overall estimate for statistical division 𝑗.

Bayesian hierarchical models are often referred to as multilevel (54) or multiscale (55)
models. In the non-Bayesian paradigm, multilevel models are a popular model class for
analysing data of the aforementioned form. Common among these methodologies is the aim
of apportioning variation in the outcome to different levels of the hierarchy (Figure 4.1).
When formulated within the Bayesian setting, a multilevel model can itself be re-expressed as
a hierarchical model, through the use of hierarchical centring (201); for this reason, these
terms are often used interchangeably.

39
Box 4.1 Nested geographies as a hierarchical model

Figure 4.1 Schematic representation of a hierarchical model

Statistical Subdivision 1

…
Statistical Division 1 Statistical Subdivision i
…

… Statistical Subdivision Ijk

State k Statistical Division j

Statistical Division Jk

Let 𝑦𝑖𝑗𝑘 = The number of cancer cases in subdivision 𝑖, within division 𝑗, within state 𝑘.

𝑂𝑖𝑗𝑘 ~Poisson(𝐸𝑖𝑗𝑘 𝜃𝑖𝑗𝑘 )

1
log(𝜽𝑗𝑘 )~Normal𝐼𝑗𝑘 (𝛽𝑗𝑘 , 2 (𝑫𝜃𝑗𝑘 − 𝑾𝜃𝑗𝑘 )−1 )
𝜎𝜃
1
𝜷𝑘 ~Normal𝐽𝑘 (𝜶𝑘 , 2 (𝑫𝛽𝑘 − 𝑾𝛽𝑘 )−1 )
𝜎𝛽

1
𝜶~Normal𝐾 (𝛾𝟏, (𝑫 − 𝑾𝜶 )−1 )
𝜎𝛼2 𝜶

𝛾~𝑝(𝛾)

The matrices 𝑫.. and 𝑾.. encode spatial correlation by defining the neighbourhood structure
among geographic units defined as the same spatial scale. Each variance component is
assigned a prior distribution, similar to the earlier Bayesian models in this report. The overall
intercept (𝛾) is also assigned a prior distribution, denoted generically as 𝑝(𝛾). Examples of
prior distributions include a Uniform distribution, 𝛾~Uniform(−1000,1000) or a Normal
distribution with large variance, 𝛾~Normal(0,1000).

40
In Australia, Turrell et al. (202) proposed a multilevel model with five levels: individuals
nested in statistical local areas, statistical subdivisions, statistical divisions and States, for
associating socioeconomic disadvantage with all-cause mortality (135). In the Atlas of
cancer mortality in the European Union (203), Poisson regression was used to attribute
variation in cancer mortality rates to age groups, countries and regions nested within
countries. Models for both applications were developed in a non-Bayesian setting and
correlation among spatially indexed effects was not accounted for.

The extension of these models to the Bayesian framework to allow for spatial smoothing is
relatively straightforward. Lawson (55) used a Bayesian hierarchical model to analyse oral
cancer incidence across the state of Georgia, USA, including both public health districts and
nested counties plus the contextual effects of district on county. In this example, the joint
model was slightly preferred over separate models for district and county. Another example
is provided by Louie and Kolaczyk (204), who exploited the Bayesian approach to detect
areas with significantly increased risk. They analysed aggregated count data across the three
nested levels of region (one area), province (nine areas) and municipality (287 areas).
Although their focus was not on estimation, it would be straightforward to combine this to
produce a disease mapping approach with testing aspects. Bayesian multiscale analyses
require careful consideration of prior selection, but have many advantages.

4.2 Inclusion of remoteness and area-level socioeconomic status

In Australia, there is substantial interaction between the geographic remoteness and

socioeconomic level of an area, with more remote areas often being more socioeconomically
disadvantaged as well as having higher levels of poverty (205). As such, understanding the
differences between different combinations, such as comparing urban very advantaged areas
to very remote very disadvantaged areas, may be desirable. Common approaches to
modelling this scenario includes either including an interaction term between the levels of
remoteness and socioeconomic disadvantage, or creating a composite variable and either
stratifying the analysis on this term, or including the levels of the composite term as dummy
variables in the model (without any remoteness/socioeconomic main effects).

Advantages of stratifying the data are the simplicity, and results can be intuitively easier for
non-statisticians to grasp. Disadvantages include an inability to compare between the areas as
thoroughly as when they are in the same model.

Advantages of including a composite variable is the simplicity of calculating parameter

estimates, however, the disadvantages include the inability to untangle main effects, which
requires careful interpretation of results (206). Previous examination of a model with main
effects would be recommended before using this approach.

The key advantage of including an interaction between the levels of the variables is the
flexibility. Interactions with other variables of interest, such as age groups, could also be
incorporated easily, while still measuring the impact of the main effects.

41
The above advantages and disadvantages are true even if a Bayesian approach is not used.
Using a Bayesian regression model would additionally require considering the prior choice
on parameters with care, ensuring convergence and identifiability of all parameters of
interest, and assessing the model via sensitivity checks (207). If spatial correlation is desired
to be included in the analysis through using a structured prior, such as the CAR prior
distribution, stratification would require a separate adjacency matrix to be generated for each
combination, due to the varying number of included areas. The Bayesian approach
additionally facilitates comparison of non-nested models, so can assist in choosing between
model options.

4.3 Use with survey data

Sample surveys are commonly used to obtain a variety of information over time for both the
total population, as well as a variety of subpopulations (126). Although these subpopulations
can be any domain, such as sociodemographic groups, our focus in this section will be on
geographic subpopulations.

Due to cost, as well as unanticipated uses of survey data, often a sample size is not
sufficiently large to enable reliable estimates for all domains. Spatial subpopulations thus
require the use of small-area estimation methods, which may involve statistical models (126).
Note that in contrast to our earlier definition of small-areas (Section 1.2), which was based on
population size, a small-area for survey data is based on having a small (and insufficient)
sample size, regardless of population (208).

The focus of small-area estimation is on producing reliable estimates of means, counts,

quantiles, as well as the associated error, for areas with limited/no sample data (209). When
outcome data are lacking, auxiliary covariate information (such as obtained from censuses or
disease registries) with good predictive power, becomes critical (209). Auxiliary variables are
thus used to ‘borrow strength’ (208).

A recent comparison of a range of procedures, spanning from weighted ‘raw’ estimates

through to models with random effect components, found that model-based estimates were
generally the ‘more effective’ approach (208). In practice, and especially in the Australian
context, direct estimates are often suppressed for at least some areas due to small numbers
and high uncertainty.

Providing the model is appropriate and the sampling is robust, there are several advantages to
modelling small-area estimates from survey data, such as (126, 210):
1. The assumed model allows ‘optimal’ estimators to be obtained
2. Each estimator can have area-specific measures of variability
3. Sample data can be used to validate models
4. Complicated data structures (such as spatial correlation) can be examined by a variety
of models.

42
When modelling survey data, the predominant paradigm employed is that of the sampling
model (see Section 2.5) (54). Most models for survey data are mixed effects models built on
the model developed by Fay and Herriot (211). It is now standard practice to include not just
the variation in auxiliary variables across small areas, but to also add random area effects to
further account for between area variability (210). Linear estimators in GLMMs can be
estimated by EBLUP, empirical Bayes, or hierarchical Bayesian models. Both empirical and
hierarchical Bayesian methods are also appropriate for a broader range of modelled
outcomes, whether binary or count data, and alternate model structures (126).

Hierarchical Bayes approaches are now extensively utilised for small-area estimation (126).
In addition to the advantages mentioned in Section 2.5.4, benefits within the sampling context
include obtaining smaller coefficient of variations for direct estimates, especially for areas
with smaller populations (126). They also avoid the problems that EBLUP or EB can have if
the restricted maximum likelihood (REML) model estimate variance is estimated to be
̂ 𝒊 being given a weight of zero (212). Any
around zero, which results in all the estimates of 𝜽
small-area estimation model can be expanded to the hierarchical Bayesian context (Box 4.2).
This extends to unmatched sampling and linking models, or incorporating spatial correlation.

Box 4.2 The basic area-level model (126)

This model can be expressed as:

𝜃̂𝑖 = 𝐳𝑖𝑇 𝜷 + 𝑏𝑖 𝑣𝑖 + 𝑒𝑖
where 𝜃̂𝑖 is an estimate of the ith area parameter 𝜃𝑖 = 𝑔(𝑌𝑖 ), 𝑧𝑖 is a vector of area-level
covariates, 𝑏𝑖 is a known positive constant, 𝑣𝑖 are area effects that are considered to be
independent and identically distributed with a mean and variance of (0, 𝜎 2 ) and are
independent of the sampling errors 𝑒𝑖 which are independently distributed with a mean of 0
and known variance 𝜓𝑖 .

The hierarchical Bayes version of this model has the addition of priors on the following
levels, for instance:
𝜃̂𝑖 ~Normal(𝜃𝑖 , 𝜓𝑖 )
𝜃𝑖 ~Normal(𝐳𝑖𝑇 𝜷, 𝑏𝑖2 𝜎𝑣2 )
𝜎𝑣2 ~Uniform(−∞, ∞)

Note that using the flat prior shown on 𝜎𝑣2 may not be ideal when the sampling variances
differ substantively over areas.

The main disadvantage of using a Bayesian analysis is that it should be conditional on all
variables that affect the probability of inclusion and non-response, and this can rapidly result
in extremely complicated models, especially when aiming to produce population estimates
from sample survey data that are not representative of the population (213). Although
weighting is often used in this situation, producing appropriate weights can be difficult, and
empty cells can cause additional difficulties for weighting in the small-area context (213).

43
Suggestions have included using multiple Bayesian hierarchical models and then averaging
over the posterior distribution, although this remains an area of active research (214).

4.4 Spatio-temporal data

Spatio-temporal data consists of data points which are stratified both by space and time. The
rate at which spatio-temporal data is generated and collected is ever-increasing, and new
methods are persistently being developed to deal with this type of data. Spatio-temporal
models can be seen as a natural extension of spatial models, but this extension increases the
complexity, both in terms of notation and computation, and introduces new complications to
be addressed, such as how to account for interactions between space and time. Moreover,
difficulties in spatial modelling, such as the handling of missing data, are exacerbated in
spatio-temporal modelling (44, 215).

Nonetheless, spatio-temporal models have many benefits in interpretation of overall patterns

of risk and dynamics, as well as improved accuracy compared with purely spatial models (216-
218).

Naturally, much of the earliest work on Bayesian spatio-temporal models focused on

extending the BYM model. The CAR prior used in the BYM model can define
neighbourhood structures across space and time, so that an area’s neighbours includes spatial
neighbours as well as its own value in the previous and following time periods (77).

One of the earliest Bayesian approaches was the Bernardinelli space-time model (Box 4.3)
(219). This has been applied to diseases such as insulin-dependent diabetes mellitus (220),
and leishmaniasis (221). Covariates have been included (82, 222), and in some cases, errors
in the estimates of indirectly observed covariates (such as, for example, estimating cigarette
smoking prevalence from survey data) have also been incorporated (222, 223).

Box 4.3 The Bernardinelli spatio-temporal model

Let 𝑂𝑖𝑡 denote observations from area 𝑖 = 1, … , 𝐼 at time 𝑡 = 1, … , 𝑇. The Bernardinelli

model can then be expressed as follows:
𝑂𝑖𝑡 ~ Poisson(𝐸𝑖𝑡 𝜃𝑖𝑡 )
log(𝜃𝑖𝑡 ) = 𝛼 + 𝑢𝑖 + 𝛾𝑡 + 𝛿𝑖𝑡
where 𝑌𝑖𝑡 are the observed cases for the ith area and tth time interval, 𝐸𝑖𝑡 are the expected
number of cases, 𝜃𝑖𝑡 are the underlying relative risks, α is the mean log-rate over all areas, 𝑢𝑖
represents the area effect and follows an intrinsic CAR distribution, 𝛾𝑡 is the mean linear time
trend over all areas and 𝛿𝑖𝑡 represents the difference between the area-specific trend and the
mean trend 𝛽𝑡 (219). In this model, the intercept is the sum of 𝛼 + 𝑢𝑖 , while the trend is the
sum of 𝛾𝑡 + 𝛿𝑖𝑡 (219). The prior for 𝛿𝑖𝑡 was a modified CAR distribution that allowed for
correlation between the intercept and trend.

44
However, the restriction to linear trends over time in the Bernardinelli model was an
important limitation (219). Further extensions have been proposed to overcome this,
including using quadratic instead of linear time trends (221, 224). In contrast, Waller et al.
(225) applied the BYM model to each time point separately. Although this allowed the spatial
structure to evolve over time, it essentially treated time as exchangeable (225, 226). This may
not be ideal for modelling a disease such as cancer since it would be unlikely to have a
separate spatial distribution within each time period (227).

Spatio-temporal interactions have also been incorporated. Sun et al. (227) and Kim et al.
(228) included random spatial and spatio-temporal interaction effects when modelling cancer
mortality in Missouri, but the temporal component was still restricted to a linear form (229).
Abellan (216) included a space-time interaction term in a BYM-type model to capture any
departure from predictable patterns based on the overall time trend and the overall spatial risk
surface. Further extensions allowed for random spatial, temporal and spatio-temporal
interaction terms, and was used to examine prostate cancer incidence in Iowa over six time
periods of 5-year groupings (229).

Mixture models have also been extended to a spatio-temporal formulation, which were
applied to lung cancer incidence and mortality in Germany for 30 years (divided into three
time periods) across 215 counties (230).

The BYM model has also been combined with dynamic models (231). Dynamic models
allow estimates to ‘borrow’ strength from adjacent timepoints, so do not assume linearity or
stationarity, but instead enable non-parametric estimation of temporal trends (226, 231). This
means time-changing effects of covariates can be included (231). In principle this model
allows for estimation of any age-period interaction, including cohort effects (224). This
model was demonstrated on Ohio lung cancer mortality data, stratified by age, gender, race
for each year (of 21 years) and each county (of 88 counties) (231).

Specific age-period-cohort (APC) Bayesian hierarchical spatio-temporal models have also

been proposed as a method to jointly study the spatial pattern of disease risk and evolution in
time (232). Generally the BYM model again forms the basis, with additional time main
effects defining age, period and cohort specific parameters; space-time interactions as
specified in Knorr-Held (226); or cohort effects (232). Time effects are assumed to vary
smoothly over space (232). These models have been applied to lung cancer in Tuscany
(232), and stomach cancer in Germany (224). A broader version of this model was proposed
which incorporated age-area and age-time effects (233). However, the inclusion of cohort
effects increases model complexity, and cohort effects in small areas may be tenuous,
particularly if there are high rates of migration between areas which would dilute cohort by
birthplace effects (233).

While methods for modelling spatio-temporal data have only transpired in the last few decades,
a plethora of spatio-temporal models now exist and continue to grow in number. The complex
nature of spatio-temporal data and the underlying processes that give rise to such data
necessitates complex models. Bayesian hierarchical models are particularly well-suited for this
task, as they provide a flexible way to describe and relate model parameters. The use of prior

45
distributions also makes it easy to account for spatial and/or temporal heterogeneity (i.e.
autocorrelation and/or clustering), as well as uncertainty and expert knowledge (87, 215).

46
5. Recommendations
“Any approach to scientific inference which seeks to
legitimize an answer in response to complex
uncertainty is, for me, a totalitarian parody of a
would-be rational learning process.”
~ Adrian F. M. Smith, in (234)

This section summarises the issues and outlines some approaches for determining an
appropriate method of analysis of spatial data in a given situation.

5.1 When should smoothing/modelling replace direct estimation?

If a raw, unsmoothed estimate possesses a sufficient level of reliability for the desired
purpose then more detailed methods may not be necessary. Nonetheless, the definition of
‘statistical reliability’ varies between different agencies and countries, even when used for
similar purposes (208). Often the suppression of estimates is dependent on both the
underlying counts as well as the uncertainty in the estimate (47), and attempts to increase
counts to sufficient levels may involve aggregating over the regions of interest.

The key advantages of smoothing/modelling are that rates can be stabilised at the resolution
of interest, and noise in the rates resulting from differences in population size is reduced (47).

Waller and Gotway (47) suggested that smoothing should be considered when:
1. The addition of one event (disease case/death), or one more person at risk, results in a
large difference (such as 25% or more) in at least one area’s rates.
2. The number of events (rate numerator) is less than three for at least one area.
3. The population at risk per area is small (for instance, less than 500 people), and these
numbers vary by an order of magnitude across the areas.

Even if the raw estimate meets confidentiality/reliability/precision guidelines, modelling is

recommended when it is desirable to:
 Include covariates
 Understand the underlying pattern of risks.

Validation of results (either external or internal) is important regardless of the method chosen
(126).

5.2 What type of smoothing/modelling should be used?

The accuracy of the method of smoothing – whether model-based or not – is critical. Areas of
high and low risk should be correctly identified, while artificially elevated, unstable rates

47
should be reduced (235). No trends or patterns should be induced by the method. Uncertainty
should also be quantified.

Building on suggested practical guidelines from Griffith (236) as well as Waller and Gotway
(14) regarding the choice of spatial proximity:
1. Using any reasonable method for modelling spatial correlation is preferable to
assuming the data are independent.
2. Exploratory spatial data analysis can ensure the choice of spatial dependence is
supported by the data.
3. Comparing the results from several different types of spatial models is also useful.
4. Spatial correlation reduces the amount of information – the effective sample size. A
very rough rule of thumb is to assume it will halve the information contained in the
data. So, if 30 data values are needed assuming independent and identically
distributed data, 60 correlated values should be used.
5. It is vital that the method used accounts for population heterogeneity.
6. Parsimony is still important. Choose the simplest model that adequately describes the
data without compromising interpretation.

When the aim is to explore the data, simplicity, speed and ease of use is preferable (47).
When the aim is to perform more detailed inferential analyses involving adjustment for
confounders, hypothesis tests, and/or ranking of areas, Bayesian methods offer several
advantages (47). Although no method perfectly compensates for small counts (237), some
approaches perform better than others. Table 5.1 provides an overview of the main
approaches discussed in this report.

Note that it is impossible to select the ideal model prior to examining the data. The amount of
smoothing that occurs is dependent on both the model and the data (86). Generally, smaller
counts will result in greater smoothing, and vice versa.

5.3 What methods should be used for a cancer atlas?

A cancer atlas may be purely descriptive or it may have a purpose or goal specific to a
particular group of users. Therefore, the first step is to obtain input from potential end users
to identify their requirements in using the maps (238). Specific methods may be better suited
to different purposes, whether for providing an accurate overview, guiding further
epidemiological studies, uncovering cancer hot-spots, or comparing regions.

However, in most situations we recommend the use of Bayesian hierarchical models for the
reason that their output is useful in decision-making (239). A Bayesian model is able to rank
estimates, compare between regions, and provide robust, reliable estimates with associated
uncertainty (239). These models also have more flexibility in adjusting to changing purposes
and aims.

48
 Table 5.1 Summary of key methods
Robust risk Can include Identifies Quantifies
Method Data privacy estimates covariates high-risk areas uncertainty Recommended for

Spatial
Global Moran's I Exploratory/Significance of results
correlation
Geary's C Exploratory/Significance of results

Tango's MEET Exploratory/Significance of results

Local LISA Exploratory

SaTScan Exploratory

Unsmoothed estimates Count Exploratory

Crude rate Exploratory

ASR Exploratory

SMR Exploratory

Direct smoothing Locally-weighted average Exploratory

49
Locally-weighted median Exploratory

Kernel smoothers Exploratory

Model-based
Poisson kriging Final results
smoothing
EBLUP Final results

Empirical Bayes Final results

Fully
BYM Final results
Bayesian
Anderson’s spatial pattern
Final results
& cluster model
Mixture models Final results

Spatial partition models Final results

Legend Yes Somewhat No

5.4 Conclusion

There is no universal approach to analysing spatial data. The characteristics of the data, the
presence of spatial correlation, and the purpose of the analysis are all important
considerations.

In public health, spatial analyses are increasing in importance and popularity. Increasingly,
decisions regarding resource and service allocation are influenced by mapped estimates. It is
therefore vital that the small-area estimates used are robust, accurate and reliable.

Our recommendations are for unsmoothed estimates as well as directly smoothed estimates to
be calculated and mapped as part of the exploratory data analysis. For producing final
estimates, modelling that incorporates smoothing has many advantages.

50
References
1. Poland B, Lehoux P, Holmes D, Andrews G. How place matters: unpacking technology and
power in health and social care. Health & Social Care in the Community. 2005;13(2):170-180.
2. Lope V, Pollan M, Perez-Gomez B, Aragones N, Ramis R, Gomez-Barroso D, et al.
Municipal mortality due to thyroid cancer in Spain. BMC Public Health. 2006;6.
3. Elliott P, Wartenberg D. Spatial epidemiology: current approaches and future challenges.
Environmental Health Perspectives. 2004;112(9):998-1006.
4. Shen W, Louis TA. Triple-goal estimates for disease mapping. Statistics in Medicine.
2000;19(17-18):2295-2308.
5. Lai P-C, So F-M, Chan K-W. Spatial Epidemiological Approaches in Disease Mapping and
Analysis. Baton Rouge: CRC Press; 2008.
6. Ord JK. Spatial Autocorrelation: A Statistician’s Reflections. In: Anselin L, Rey JS, editors.
Perspectives on Spatial Data Analysis. Berlin, Heidelberg: Springer Berlin Heidelberg; 2010. p. 165-
180.
7. Ma H, Carlin BP, Banerjee S. Hierarchical and Joint Site-Edge Methods for Medicare
Hospice Service Region Boundary Analysis. Biometrics. 2010;66(2):355-364.
8. Zhang J, Atkinson P, Goodchild MF. Lattice data and scale models. In: Zhang J, Atkinson P,
Goodchild MF, editors. Scale in Spatial Information and Analysis. Boca Raton, FL: CRC Press; 2014.
9. Vázquez EF, Morollón FR. Preface. In: Vázquez EF, Morollón FR, editors. Defining the
Spatial Scale in Modern Regional Analysis: New Challenges from Data at Local Level. Berlin:
Springer; 2012.
10. Kang SY, McGree J, Baade P, Mengersen K. An investigation of the impact of various
geographical scales for the specification of spatial dependence. Journal of Applied Statistics. 2014;
41(11):2515-2538.
11. Tobler WR. A Computer Movie Simulating Urban Growth in the Detroit Region. Economic
Geography. 1970;46:234-240.
12. Fischer MM, Wang J. Spatial Data Analysis: Models, Methods and Techniques. Berlin:
Springer; 2011.
13. Zandbergen PA. Geocoding Quality and Implications for Spatial Analysis. Geography
Compass. 2009;3(2):647-680.
14. Waller LA, Gotway CA. Applied Spatial Statistics for Public Health Data. Chichester: John
Wiley & Sons, Inc; 2004.
15. Demsar U, Harris P, Brunsdon C, Fotheringham AS, McLoone S. Principal component
analysis on spatial data: An overview. Annals of the Association of American Geographers. 2013;
103(1):106-128.
16. Anselin L, Griffith DA. Do spatial effects really matter in regression analysis? Papers in
Regional Science. 1988;65(1):11-34.
17. Choo L, Walker SG. A new approach to investigating spatial variations of disease. Journal of
the Royal Statistical Society: Series A (Statistics in Society). 2008;171(2):395-405.
18. Wakefield J, Elliott P. Issues in the statistical analysis of small area health data. Statistics in
Medicine. 1999;18(17-18):2377-2399.
19. Moran PAP. Notes on continuous stochastic phenomena. Biometrika. 1950;37(1-2):17-23.
20. Geary RC. The Contiguity Ratio and Statistical Mapping. The Incorporated Statistician. 1954;
5(3):115-146.
21. Anselin L. Local Indicators of Spatial Association—LISA. Geographical Analysis. 1995;
27(2):93-115.
22. Kulldorff M, Song C, Gregorio D, Samociuk H, DeChello L. Cancer map patterns: are they
random or not? American Journal of Preventive Medicine. 2006;30(2 Suppl):S37-S49.
23. Tango T. A test for spatial disease clustering adjusted for multiple testing. Statistics in
Medicine. 2000;19(2):191-204.
24. Kulldorff M. A spatial scan statistic. Communications in Statistics - Theory and Methods.
1997;26(6):1481-1496.

51
25. Sankey TT. Statistical Descriptions of Spatial Patterns. In: Shekhar S, Xiong H, editors.
Encyclopedia of GIS. Boston, MA: Springer US; 2008. p. 1135-1141.
26. Tango T. A class of tests for detecting 'general' and 'focused' clustering of rare diseases.
Statistics in Medicine. 1995;14(21-22):2323-2334.
27. Oyana TJ, Margai F. Spatial Analysis : Statistics, Visualization, and Computational Methods.
Boca Raton: CRC Press; 2015.
28. Kulldorff M, Information Management Services Inc. SaTScan v9.4.2: Software for the
spatial, temporal and space-time scan statistics. www.satscan.org/; 2015.
29. Jung I, Kulldorff M, Richard OJ. A spatial scan statistic for multinomial data. Statistics in
Medicine. 2010;29(18):1910-1918.
30. Huang L, Kulldorff M, Gregorio D. A Spatial Scan Statistic for Survival Data. Biometrics.
2007;63(1):109-118.
31. Tango T. A Spatial Scan Statistic with a Restricted Likelihood Ratio. Japanese Journal of
Biometrics. 2008;29(2):75-95.
32. Samarasundera E, Walsh T, Cheng T, Koenig A, Jattansingh K, Dawe A, et al. Methods and
tools for geographical mapping and analysis in primary health care. Primary Health Care Research &
Development. 2012;13(1):10-21.
33. Inskip H, Beral V, Fraser P, Haskey J. Methods for age-adjustment of rates. Statistics in
Medicine. 1983;2(4):455-466.
34. Anselin L. Under the hood. Issues in the specification and interpretation of spatial regression
models. Agricultural Economics. 2002;27(3):247-267.
35. Bavaud F. Models for Spatial Weights: A Systematic Look. Geographical Analysis. 1998;
30(2):153-171.
36. Cliff AD, Ord JK. Spatial Processes: Models and Applications. London: Pion; 1981.
37. Anselin L, Lozano N, Koschinsky J. Rate transformations and smoothing. Urbana, IL:
Department of Geography, University of Illinois, 2006.
38. Earnest A, Morgan G, Mengersen K, Ryan L, Summerhayes R, Beard J. Evaluating the effect
of neighbourhood weight matrices on smoothing properties of Conditional Autoregressive (CAR)
models. International Journal of Health Geographics. 2007;6:54.
39. Lawson AB, Williams FLR. An Introductory Guide to Disease Mapping. New York: John
Wiley & Sons, Ltd; 2001.
40. Kafadar K. Smoothing geographical data, particularly rates of disease. Statistics in Medicine.
1996;15(23):2539-2560.
41. Mausner JS, Kramer S. Epidemiology: An Introductory Text. Philadelphia: W. B. Saunders;
1985.
42. Krieger N, Williams DR. Changing to the 2000 Standard Million: Are Declining
Racial/Ethnic and Socioeconomic Inequalities in Health Real Progress or Statistical Illusion?
American Journal of Public Health. 2001;91(8):1209-1213.
43. Semenciw RM, Le ND, Marrett LD, Robson DL, Turner D, Walter SD. Methodological
issues in the development of the Canadian Cancer Incidence Atlas. Statistics in Medicine. 2000;
19(17-18):2437-2449.
44. Banerjee S, Carlin BP, Gelfand AE. Hierarchical Modeling and Analysis for Spatial Data,
Second Edition. Boca Raton, FL, USA: Chapman and Hall/CRC; 2014.
45. Kyriakidis PC. A Geostatistical Framework for Area‐to‐Point Spatial Interpolation.
Geographical Analysis. 2004;36(3):259-289.
46. Cressie N. Smoothing Regional Maps Using Empirical Bayes Predictors. Geographical
Analysis. 1992;24(1):75-95.
47. Lawson AB, Biggeri AB, Boehning D, Lesaffre E, Viel JF, Clark A, et al. Disease mapping
models: an empirical evaluation. Statistics in Medicine. 2000;19(17-18):2217-2241.
48. Lawson AB. Statistical Methods in Spatial Epidemiology, Second Edition. West Sussex:
Wiley; 2006.
49. Qiu P. Basic Statistical Concepts and Conventional Smoothing Techniques. In: Qiu P, editor.
Image Processing and Jump Regression Analysis. Chichester, UK: John Wiley & Sons, Inc.; 2005. p.
13-54.

52
50. Mungiole M, Pickle LW, Simonson KH. Application of a weighted head-banging algorithm
to mortality data maps. Statistics in Medicine. 1999;18(23):3201-3209.
51. Goovaerts P. Geostatistical analysis of disease data: estimation of cancer mortality risk from
empirical frequencies using Poisson kriging. International Journal of Health Geographics. 2005;4:31.
52. Nadaraya EA. On Estimating Regression. Theory of Probability & Its Applications. 1964;
9(1):141-142.
53. Watson GS. Smooth regression analysis. Sankhya (Series A). 1964;26:359-372.
54. Gelman A, Hill J, Alvarez RM, Beck NL, Wu LL. Data Analysis Using Regression and
Multilevel/Hierarchical Models. Cambridge: Cambridge University Press; 2006.
55. Lawson AB. Bayesian Disease Mapping: Hierarchical Modeling in Spatial Epidemiology.
Second edition. Boca Raton: CRC Press; 2013.
56. Gelman A. Analysis of variance - why it is more important than ever. The Annals of
Statistics. 2005;33(1):1-53.
57. Shao C. Approaches to the spatial modelling of cancer incidence and mortality in
metropolitan Perth, Western Australia, 1990-2005. [PhD Thesis]. Perth, WA: Edith Cowan
University; 2011.
58. Henderson R, Shimakura S, Gorst D. Modeling spatial variation in leukaemia survival data.
Journal of the American Statistical Association. 2002;97:965-972.
59. Montero J-M, Fernández-Avilés G, Mateu J. Spatial and Spatio-Temporal Geostatistical
Modeling and Kriging. Chichester, UK: John Wiley & Sons, Ltd; 2015.
60. Asmarian NS, Ruzitalab A, Amir K, Masoud S, Mahaki B. Area-to-Area Poisson Kriging
analysis of mapping of county- level esophageal cancer incidence rates in Iran. Asian Pacific Journal
of Cancer Prevention. 2013;14(1):11-13.
61. Goovaerts P. Geostatistical analysis of disease data: accounting for spatial support and
population density in the isopleth mapping of cancer mortality risk using area-to-point Poisson
kriging. International Journal of Health Geographics. 2006;5(1):1-31.
62. Goovaerts P. Kriging and Semivariogram Deconvolution in the Presence of Irregular
Geographical Units. Mathematical Geology. 2008;40(1):101-128.
63. Mockus A. Estimating Dependencies from Spatial Averages. Journal of Computational and
Graphical Statistics. 1998;7(4):501-513.
64. Gotway CA, Young LJ. A geostatistical approach to linking geographically aggregated data
from different sources. Technical report # 2004-012. Gainesville, FL: Department of Statistics,
University of Florida, 2004.
65. Goovaerts P, Gebreab S. How does Poisson kriging compare to the popular BYM model for
mapping disease risks? International Journal of Health Geographics. 2008;7:6.
66. Bolstad WM. Bayesian Inference for Poisson. Introduction to Bayesian Statistics. Hoboken,
NJ: John Wiley & Sons, Inc.; 2007. p. 183-198.
67. Ghosh M, Natarajan K, Waller LA, Kim D. Hierarchical Bayes GLMs for the analysis of
spatial data: An application to disease mapping. Journal of Statistical Planning and Inference. 1999;
75(2):305-318.
68. Devine OJ, Louis TA, Halloran ME. Empirical Bayes methods for stabilizing incidence rates
before mapping. Epidemiology. 1994;5(6):622-630.
69. Clayton D, Kaldor J. Empirical Bayes estimates of age-standardised relative risks for use in
disease mapping. Biometrics. 1987;43(3):671-681.
70. Cressie N, Read TRC. Spatial data analysis of regional counts. Biometrics Journal. 1989;6:
699-719.
71. Ghosh JK, Delampady M, Samanta T. An Introduction to Bayesian Analysis: Theory and
Methods. New York, NY: Springer; 2006.
72. Louis TA, Shen W. Innovations in Bayes and empirical Bayes methods: Estimating
parameters, populations and ranks. Statistics in Medicine. 1999;18(17-18):2493-2505.
73. Maiti T. Hierarchical Bayes estimation of mortality rates for disease mapping. Journal of
Statistical Planning and Inference. 1998;69(2):339-348.
74. Lahiri P, Rao JNK. Robust estimation of mean squared error of small area estimators. Journal
of the American Statistical Association. 1995;90:758-766.

53
75. Meza JL. Empirical Bayes estimation smoothing of relative risks in disease mapping. Journal
of Statistical Planning and Inference. 2003;112(1-2):43-62.
76. Mugglin AS, Cressie N, Gemmell I. Hierarchical statistical modelling of influenza epidemic
dynamics in space and time. Statistics in Medicine. 2002;21(18):2703-2721.
77. Carlin BP, Xia H. Assessing environmental justice using Bayesian hierarchical models: two
case studies. Journal of Exposure Analysis and Environmental Epidemiology. 1999;9(1):66-78.
78. Dunson DB. Commentary: Practical advantages of Bayesian analysis of epidemiologic data.
American Journal of Epidemiology. 2001;153(12):1222-1226.
79. Thompson JA, Carozza SE, Zhu L. An evaluation of spatial and multivariate covariance
among childhood cancer histotypes in Texas (United States). Cancer Causes & Control. 2007;18(1):
105-113.
80. Greco FP, Lawson AB, Cocchi D, Temples T. Some interpolation estimators in
environmental risk assessment for spatially misaligned health data. Environmental and Ecological
Statistics. 2005;12(4):379-395.
81. Besag J, York J, Mollie A. Bayesian image restoration, with two applications in spatial
statistics. Annals of the Institute of Statistical Mathematics. 1991;43:1-59.
82. Kim H, Sun DC, Tsutakawa RK. Lognormal vs. gamma: Extra variations. Biometrical
Journal. 2002;44(3):305-323.
83. Bell BS, Broemeling LD. A Bayesian analysis for spatial processes with application to
disease mapping. Statistics in Medicine. 2000;19(7):957-974.
84. Ocana-Riola R. The misuse of count data aggregated over time for disease mapping. Statistics
in Medicine. 2007;26(24):4489-4504.
85. Gelfand AE, Vounatsou P. Proper multivariate conditional autoregressive models for spatial
data analysis. Biostatistics. 2003;4(1):11-25.
86. Waller LA, Carlin BP. Disease mapping. In: Gelfand AE, Diggle PJ, Guttorp P, Fuentes M,
editors. Handbook of spatial statistics. Boca Raton: CRC Press; 2010.
87. Best N, Richardson S, Thomson A. A comparison of Bayesian spatial models for disease
mapping. Statistical Methods in Medical Research. 2005;14(1):35-59.
88. Anderson C, Lee D, Dean N. Bayesian cluster detection via adjacency modelling. Spatial and
Spatiotemporal Epidemiology. 2016;16:11-20.
89. MacNab YC. On Gaussian Markov random fields and Bayesian disease mapping. Statistical
Methods in Medical Research. 2011;20(1):49-68.
90. Kokki E, Ranta J, Penttinen A, Pukkala E, Pekkanen J. Small area estimation of incidence of
cancer around a known source of exposure with fine resolution data. Occupational and Environmental
Medicine. 2001;58(5):315-320.
91. Hughes J, Haran M. Dimension reduction and alleviation of confounding for spatial
generalized linear mixed models. Journal of the Royal Statistical Society: Series B (Statistical
Methodology). 2013;75(1):139-159.
92. Clayton DG, Bernardinelli L, Montomoli C. Spatial correlation in ecological analysis.
International Journal of Epidemiology. 1993;22(6):1193-1202.
93. Leroux BG, Lei X, Breslow N. Estimation of disease rates in small areas: a new mixed model
for spatial dependence. In: Halloran ME, Berry D, editors. Statistical models in epidemiology, the
environment and clinical trials. New York: Springer; 2000. p. 135-178.
94. Lee D. A comparison of conditional autoregressive models used in Bayesian disease
mapping. Spatial and Spatiotemporal Epidemiology. 2011;2(2):79-89.
95. MacNab YC, Dean CB. Parametric bootstrap and penalized quasi-likelihood inference in
conditional autoregressive models. Statistics in Medicine. 2000;19(17-18):2421-2435.
96. Lu H, Reilly CS, Banerjee S, Carlin BP. Bayesian areal wombling via adjacency modeling.
Environmental and Ecological Statistics. 2007;14(4):433-452.
97. White NM. Review of statistical methods for disease mapping. Available online:
http://eprints.qut.edu.au/56859/. 2012.
98. Anderson C, Lee D, Dean N. Identifying clusters in Bayesian disease mapping. Biostatistics.
2014;15(3):457-469.
99. Lawson AB, Clark A. Spatial mixture relative risk models applied to disease mapping.
Statistics in Medicine. 2002;21(3):359-370.

54
100. Denison DGT, Holmes CC. Bayesian partitioning for estimating disease risk. Biometrics.
2001;57(1):143-149.
101. Green PJ, Richardson S. Hidden Markov models and disease mapping. Journal of the
American Statistical Association. 2002;97(460):1055-1070.
102. Richardson S, Thomson A, Best N, Elliott P. Interpreting posterior relative risk estimates in
disease-mapping studies. Environmental Health Perspectives. 2004;112(9):1016-1025.
103. Hossain MM, Lawson A, B. Mixtures and Latent Structure in Spatial Epidemiology. In:
Lawson AB, Banerjee S, Haining RP, Ugarte MD, editors. Handbook of Spatial Epidemiology.
Chapman & Hall/CRC Handbooks of Modern Statistical Methods: Chapman and Hall/CRC; 2016. p.
349-361.
104. Fernández C, Green PJ. Modelling spatially correlated data via mixtures: a Bayesian
approach. Journal of the Royal Statistical Society: Series B (Statistical Methodology). 2002;64(4):
805-826.
105. Cremers D. Image Segmentation with Shape Priors: Explicit Versus Implicit Representations.
In: Scherzer O, editor. Handbook of Mathematical Methods in Imaging. New York, NY: Springer
New York; 2015. p. 1909-1944.
106. Zach C, Häne C, Pollefeys M. What Is Optimized in Convex Relaxations for Multilabel
Problems: Connecting Discrete and Continuously Inspired MAP Inference. IEEE Transactions on
Pattern Analysis and Machine Intelligence. 2014;36(1):157-170.
107. Pereyra M, McLaughlin S. Fast unsupervised Bayesian image segmentation with adaptive
spatial regularisation. E-print on Arxiv (arxiv.org/abs/1502.01400v3). 2016.
108. Bradley JR, Wikle CK, Holan SH. Bayesian Spatial Change of Support for Count-Valued
Survey Data with Application to the American Community Survey. Journal of the American
Statistical Association. 2015:1-43.
109. Flowerdew R, Green M. Areal interpolation and types of data. In: Fotheringham S, Rogerson
P, editors. Spatial Analysis and GIS. London: Taylor and Francis; 1994. p. 121-145.
110. Dunson DB, Herring AH. Bayesian model selection and averaging in additive and
proportional hazards models. Lifetime Data Analysis. 2005;11(2):213-232.
111. Omurlu IK, Ozdamar K, Ture M. Comparison of Bayesian survival analysis and Cox
regression analysis in simulated and breast cancer data sets. Expert Systems with Applications. 2009;
36(8):11341-11346.
112. Banerjee S, Carlin BP. Semiparametric spatio-temporal frailty modeling. Environmetrics.
2003;14(5):523-535.
113. Sauleau EA, Hennerfeind A, Buemi A, Held L. Age, period and cohort effects in Bayesian
smoothing of spatial cancer survival with geoadditive models. Statistics in Medicine. 2007;26(1):212-
229.
114. Banerjee T, Chen MH, Dey DK, Kim S. Bayesian analysis of generalized odds-rate hazards
models for survival data. Lifetime Data Analysis. 2007;13(2):241-260.
115. Henderson R, Shimakura S, Gorst D. Modeling spatial variation in leukemia survival data.
Journal of the American Statistical Association. 2002;97:965-972.
116. Zhang J, Lawson AB. Bayesian parametric accelerated failure time spatial model and its
application to prostate cancer. Journal of Applied Statistics. 2011;38(2):591-603.
117. Li L, Hanson T, Zhang J. Spatial extended hazard model with application to prostate cancer
survival. Biometrics. 2015;71(2):313-322.
118. Wang S, Zhang J, Lawson AB. A Bayesian normal mixture accelerated failure time spatial
model and its application to prostate cancer. Statistical Methods in Medical Research. 2016;25(2):
793-806.
119. Diva U, Dey DK, Banerjee S. Parametric models for spatially correlated survival data for
individuals with multiple cancers. Statistics in Medicine. 2008;27(12):2127-2144.
120. Fairley L, Forman D, West R, Manda S. Spatial variation in prostate cancer survival in the
Northern and Yorkshire region of England using Bayesian relative survival smoothing. British Journal
of Cancer. 2008;99(11):1786-1793.
121. Dickman PW, Sloggett A, Hills M, Hakulinen T. Regression models for relative survival.
Statistics in Medicine. 2004;23(1):51-64.

55
122. Royston P, Lambert PC. Flexible parametric survival analysis using Stata: beyond the Cox
model. College Station, Texas: StataCorp LP; 2011.
123. Hennerfeind A, Held L, Sauleau EA. A Bayesian analysis of relative cancer survival with
geoadditive models. Statistical Modeling. 2008;8(2):117-139.
124. Cramb SM, Mengersen KL, Lambert P, Ryan L, Baade PD. A flexible parametric approach to
examining spatial variation in relative survival In: Cramb S, editor. Spatio-temporal modelling of
cancer data in Queensland using Bayesian methods PhD by Publication. Brisbane: Queensland
University of Technology; 2015.
125. Nelson CP, Lambert PC, Squire IB, Jones DR. Flexible parametric models for relative
survival, with application in coronary heart disease. Statistics in Medicine. 2007;26(30):5486-5498.
126. Rao JNK, Molina I. Small Area Estimation, Second Edition. New Jersey: John Wiley & Sons,
Inc; 2015.
127. Pratesi M, Salvati N. Small area estimation: the EBLUP estimator based on spatially
correlated random area effects. Statistical Methods and Applications. 2008;17(1):113-141.
128. Datta GS, Ghosh M. Bayesian prediction in linear models: applications to small area
estimation. The Annals of Statistics. 1991;19(4):1748-1770.
129. Arora V, Lahiri P. On the superiority of the Bayesian method over the BLUP in small area
estimation problems. Statistica Sinica. 1997;7(4):1053-1063.
130. BioMedware. Spacestat v4.0. Ann Arbor, MI: BioMedware, 2014.
131. AvRuskin GA, Jacquez GM, Meliker JR, Slotnick MJ, Kaufmann AM, Nriagu JO.
Visualization and exploratory analysis of epidemiologic data using a novel space time information
system. International Journal of Health Geographics. 2004;3(1):26.
132. Konishi S. Introduction to Multivariate Analysis : Linear and Nonlinear Modeling. Bosa
Roca: CRC Press; 2014.
133. LeSage JP, Pace RK. Introduction. In: LeSage JP, Pace RK, editors. Spatial and
Spatiotemporal Econometrics. Amsterdam: Elsevier; 2004. p. 1-32.
134. Brooks S, Gelman A, Jones G, Meng X. Handbook of Markov Chain Monte Carlo. Boca
Raton, FL: Chapman & Hall/CRC; 2011.
135. Cowles MK, Carlin BP. Markov Chain Monte Carlo Convergence Diagnostics: A
Comparative Review. Journal of the American Statistical Association. 1996;91(434):883-904.
136. Gelfand AE. Gibbs sampling. Journal of the American Statistical Association. 2000;95(452):
1300-1304.
137. Howard RA. Dynamic Probabilistic Systems. Volume I: Markov models. Mineola, NY:
Dover Publications, Inc; 2007.
138. Geman S, Geman D. Stochastic relaxation, Gibbs distributions, and the Bayesian restoration
of images. IEEE Transactions on Pattern Analysis and Machine Intelligence. 1984;PAMI-6(6):721-
741.
139. Gelfand AE, Smith AFM. Sampling-based approaches to calculating marginal densities.
Journal of the American Statistical Association. 1990;85(410):398-409.
140. Geyer CJ. Introduction to Markov Chain Monte Carlo. In: Brooks S, Gelman A, Jones G,
Meng X, editors. Handbook of Markov Chain Monte Carlo. Boca Raton, FL: Chapman & Hall/CRC;
2011.
141. Taylor BM, Diggle PJ. INLA or MCMC? A tutorial and comparative evaluation for spatial
prediction in log-Gaussian Cox processes. Journal of Statistical Computation and Simulation.
2013;84(10):2266-2284.
142. Besag J, Green P, Higdon D, Mengersen K. Bayesian Computation and Stochastic Systems.
Statistical Science. 1995;10(1):3-41.
143. Haran M. Gaussian Random field models for spatial data. In: Brooks S, Gelman A, Jones G,
Meng X, editors. Handbook of Markov Chain Monte Carlo. Boca Raton, FL: Chapman & Hall/CRC;
2011.
144. Lunn DJ, Thomas A, Best N, Spiegelhalter DJ. WinBUGS - a Bayesian modelling
framework: concepts, structure, and extensibility. Statistics and Computing. 2000;10:325-337.
145. Stan Development Team. Stan Modeling Language User's Guide and Reference Manual.
Version 2.6.1. 2015.

56
146. Browne WJ. MCMC Estimation in MLwiN v2.1. Bristol: Centre for Multilevel Modelling,
University of Bristol, 2009.
147. Plummer M. rjags: Bayesian Graphical Models using MCMC. R package version 4-6, 2016.
148. Martin AD, Quinn KM, Park JH. MCMCpack: Markov Chain Monte Carlo Package. R
package version 1.3-4: 2014.
149. Rue H, Riebler A, Sørbye SH, Illian JB, Simpson DP, Lindgren FK. Bayesian Computing
with INLA: A Review. arXiv:1604.00860v1, 2016.
150. De Smedt T, Simons K, Van Nieuwenhuyse A, Molenberghs G. Comparing MCMC and
INLA for disease mapping with Bayesian hierarchical models. Archives of Public Health. 2015;
73(Suppl 1):O2-O2.
151. Bivand RS, Pebesma E, Gómez-Rubio V. Areal Data and Spatial Autocorrelation. In: Bivand
RS, Pebesma E, Gómez-Rubio V, editors. Applied Spatial Data Analysis with R. New York: Springer;
2008. p. 237-272.
152. Sadeq M. Spatial patterns and secular trends in human leishmaniasis incidence in Morocco
between 2003 and 2013. Infectious Diseases of Poverty. 2016;5:48.
153. Chaikaew N, Tripathi NK, Souris M. Exploring spatial patterns and hotspots of diarrhea in
Chiang Mai, Thailand. International Journal of Health Geographics. 2009;8:36.
154. Cramb SM, Mengersen KL, Baade PD. Atlas of Cancer in Queensland: geographical variation
in incidence and survival, 1998 to 2007. Brisbane: Viertel Centre for Research in Cancer Control,
Cancer Council Queensland, 2011.
155. Ozdenerol E, Williams BL, Kang SY, Magsumbol MS. Comparison of spatial scan statistic
and spatial filtering in estimating low birth weight clusters. International Journal of Health
Geographics. 2005;4(1):1-10.
156. Rey SJ, Janikas MV. STARS: Space–Time Analysis of Regional Systems. Geographical
Analysis. 2006;38(1):67-86.
157. Nguyen P, Brown PE, Stafford J. Mapping Cancer Risk in Southwestern Ontario with
Changing Census Boundaries. Biometrics. 2012;68(4):1228-1237.
158. Baade PD, Fritschi L, Aitken JF. Geographical differentials in cancer incidence and survival
in Queensland: 1996 to 2002. Brisbane: Viertel Centre for Research in Cancer Control, Queensland
Cancer Fund, 2005.
159. Cramb SM, Mengersen KL, Baade PD. Developing the atlas of cancer in Queensland:
methodological issues. International Journal of Health Geographics. 2011;10:9.
160. Congdon P. Bayesian models for area health and mortality variations: an overview [ppt
presentation]. Lodon, UK: Centre for Statistics and Department of Geography, Queen Mary
University of London, 2007.
161. Costain DA. Bayesian partitioning for mapping disease risk using a matched case-control
approach to confounding. Biostatistics. 2013;14(1):99-112.
162. Osnes K, Aalen OO. Spatial smoothing of cancer survival: A Bayesian approach. Statistics in
Medicine. 1999;18(16):2087-2099.
163. Moss S. General Principles of Cancer Screening. In: Chamberlain J, Moss S, editors.
Evaluation of Cancer Screening. London: Springer; 1996. p. 1-13.
164. Kreuger FA, van Oers HA, Nijs HG. Cervical cancer screening: spatial associations of
outcome and risk factors in Rotterdam. Public Health. 1999;113(3):111-115.
165. Lofters AK, Gozdyra P, Lobb R. Using geographic methods to inform cancer screening
interventions for South Asians in Ontario, Canada. BMC Public Health. 2013;13(1):1-8.
166. Anselin L. Local indicators of spatial association - LISA. Geographical Analysis. 1995;27.
167. Fernandes KA, Sutradhar R, Borkhoff CM, Baxter N, Lofters A, Rabeneck L, et al. Small-
area variation in screening for cancer, glucose and cholesterol in Ontario: a cross-sectional study.
CMAJ Open. 2015;3(4):E373-381.
168. Gwede CK, Ward BG, Luque JS, Vadaparampil ST, Rivers D, Martinez-Tyson D.
Application of geographic information systems and asset mapping to facilitate identification of
colorectal cancer screening resources. Online Journal of Public Health Informatics. 2010;
2(1):ojphi.v2i1.2893.
169. Le N, Marrett LD, Robson DL, Semenciw RM, Turner D, Walter SD. Canadian Cancer
Incidence Atlas. Ottawa: Ministry of Supply and Services, 1995.

57
170. Nandakumar A, Gupta PC, Gangadharan P, Visweswara RN. Development of an Atlas of
Cancer in India: First all India report 2001-2002. Bangalore: National Cancer Registry Programme
(ICMR), 2004.
171. New York State Cancer Registry. Maps of Cancer Incidence by County. New York: New
York State Department of Health; 2016. Available from:
http://www.health.state.ny.us/statistics/cancer/registry/cntymaps/index.htm.
172. Penn State Health. Pennsylvania Cancer Atlas: a CDC/GeoVISTA prototype. Pennsylvania:
Penn State GeoVISTA. Available from: http://www.geovista.psu.edu/grants/CDC.
173. SA Department of Health. The Geography of Cancer in South Australia,1991-2000. Adelaide:
SA Department of Health, 2005.
174. Swedish Oncological Centres. Atlas of Cancer Incidence in Sweden. Stockholm: Swedish
Oncological Centres, 1995.
175. University of New Hampshire's Institute for Health Policy and Practice. MapNH Health:
Insights for our future. Available at: mapnhhealth.org/. Concord, NH: University of New Hampshire,
2014.
176. Devesa SS, Grauman DJ, Blot WJ, Pennello G, Hoover RN, Fraumeni JFJ. Atlas of cancer
mortality in the United States, 1950-94. Washington: US Government Printing Office, 1999.
177. Goovaerts P. Geostatistical Analysis of Health Data: State-of-the-Art and Perspectives. In:
Soares A, Pereira MJ, Dimitrakopoulos R, editors. geoENV VI – Geostatistics for Environmental
Applications: Proceedings of the Sixth European Conference on Geostatistics for Environmental
Applications. Dordrecht: Springer Netherlands; 2008. p. 3-22.
178. Rainey JJ, Omenah D, Sumba PO, Moormann AM, Rochford R, Wilson ML. Spatial
clustering of endemic Burkitt's lymphoma in high-risk regions of Kenya. International Journal of
Cancer. 2007;120(1):121-127.
179. Militino AF, Ugarte MD, Dean CB. The use of mixture models for identifying high risks in
disease mapping. Statistics in Medicine. 2001;20(13):2035-2049.
180. Martuzzi M, Comba P, De Santis M, Iavarone I, Di Paola M, Mastrantonio M, et al.
Asbestos-related lung cancer mortality in Piedmont, Italy. American Journal of Industrial Medicine.
1998;33(6):565-570.
181. Lu WS, Tsutakawa RK. Analysis of mortality rates via marginal extended quasi-likelihood.
Statistics in Medicine. 1996;15(13):1397-1407.
182. Chellini E, Gorini G, Martini A, Giovannetti L, Costantini AS. Lung cancer mortality patterns
in women resident in different urbanization areas in central Italy from 1987-2002. Tumori. 2006;
92(4):271-275.
183. Sandor J, Kiss I, Farkas O, Ember I. Association between gastric cancer mortality and nitrate
content of drinking water: Ecological study on small area inequalities. European Journal of
Epidemiology. 2001;17(5):443-447.
184. Prieto RR, Garcia-Perez J, Pollan M, Aragones N, Perez-Gomez B, Lopez-Abente G.
Modelling of municipal mortality due to haematological neoplasias in Spain. Journal of Epidemiology
and Community Health. 2007;61(2):165-171.
185. Lopez-Abente G, Hernandez-Barrera V, Pollan M, Aragones N, Perez-Gomez B. Municipal
pleural cancer mortality in Spain. Occupational and Environmental Medicine. 2005;62(3):195-199.
186. Ghosh P, Huang L, Yu BB, Tiwari RC. Semiparametric Bayesian approaches to joinpoint
regression for population-based cancer survival data. Computational Statistics & Data Analysis.
2009;53(12):4073-4082.
187. Biggeri A, Bonannini M, Catelan D, Divino F, Dreassi E, Lagazio C. Bayesian ecological
regression with latent factors: Atmospheric pollutants emissions and mortality for lung cancer.
Environmental and Ecological Statistics. 2005;12(4):397-409.
188. Thomas AJ, Carlin BP. Late detection of breast and colorectal cancer in Minnesota counties:
an application of spatial smoothing and clustering. Statistics in Medicine. 2003;22(1):113-127.
189. Parodi S, Stagnaro E, Casella C, Puppo A, Daminelli E, Fontana V, et al. Lung cancer in an
urban area in Northern Italy near a coke oven plant. Lung Cancer. 2005;47(2):155-164.
190. Parodi S, Vercelli M, Stella A, Stagnaro E, Valerio F. Lymphohaematopoietic system cancer
incidence in an urban area near a coke oven plant: an ecological investigation. Occupational and
Environmental Medicine. 2003;60(3):187-194.

58
191. Knorr-Held L, Raer G, Becker N. Disease mapping of stage-specific cancer incidence data.
Biometrics. 2002;58(3):492-501.
192. Rachet B, Coleman MP. Cancer survival indicators for the National Health Service in
England: exploration of alternative geographic units of analysis – the Primary Care Organisation and
Strategic Health Authority. London: London School of Hygiene and Tropical Medicine, 2004.
193. Huang L, Pickle LW, Stinchcomb D, Feuer EJ. Detection of spatial clusters: application to
cancer survival as a continuous outcome. Epidemiology. 2007;18(1):73-87.
194. Chen DG, Lio YL. Comparative studies on frailties in survival analysis. Communications in
Statistics - Simulation and Computation. 2008;37(8):1631-1646.
195. Hennerfeind A, Brezger A, Fahrmeir L. Geoadditive survival models. Journal of the
American Statistical Association. 2006;101:1065-1075.
196. Li Y, Brown P, Rue H, al-Maini M, Fortin P. Spatial modelling of lupus incidence over 40
years with changes in census areas. Journal of the Royal Statistical Society: Series C (Applied
Statistics). 2012;61(1):99-115.
197. Cramb SM, Mengersen KL, Turrell G, Baade PD. Spatial inequalities in colorectal and breast
cancer survival: premature deaths and associated factors. Health Place. 2012;18(6):1412-1421.
198. Hsieh JCF, Cramb SM, McGree JM, Dunn NAM, Baade PD, Mengersen KL. Does
geographic location impact the survival differential between screen- and interval-detected breast
cancers? Stochastic Environmental Research and Risk Assessment. 2016;30(1):155-165.
199. Hsieh JC-F, Cramb SM, McGree JM, Baade PD, Dunn NAM, Mengersen KL. Bayesian
Spatial Analysis for the Evaluation of Breast Cancer Detection Methods. Australian & New Zealand
Journal of Statistics. 2013;55(4):351-367.
200. Walsh CD, Mengersen KL. Ordering of Hierarchies in Hierarchical Models: Bone Mineral
Density Estimation. In: Alston CL, Mengersen KL, Pettitt AN, editors. Case Studies in Bayesian
Statistical Modelling and Analysis. Chichester, UK: John Wiley & Sons, Ltd; 2012. p. 159-170.
201. Papaspiliopoulos O, Roberts GO, Sköld M. A general framework for the parametrization of
hierarchical models. Statistical Science 2007;22(1):59-73.
202. Turrell G, Kavanagh A, Draper G, Subramanian SV. Do places affect the probability of death
in Australia? A multilevel study of area-level disadvantage, individual-level socioeconomic position
and all-cause mortality, 1998-2000. Journal of Epidemiology and Community Health. 2007;61(1):13-
19.
203. Robertson C, Mazzetta C, D'Onofrio A. Regional variation and spatial correlation. In: Boyle
P, Smans M, editors. Atlas of cancer mortality in the European Union and the European Economic
Area 1993-1997 IARC Scientific Publication No 159. Lyon, France: International Agency for
Research on Cancer; 2008.
204. Louie MM, Kolaczyk ED. Multiscale detection of localized anomalous structure in aggregate
disease incidence data. Statistics in Medicine. 2006;25(5):787-810.
205. National Rural Health Alliance, Australian Council of Social Service. A snapshot of poverty
in rural and regional Australia. Canberra: NRHA and ACOSS, 2013.
206. Ntzoufras I. Bayesian Modeling Using WinBUGS. Hoboken, NJ: John Wiley & Sons, Inc.;
2008.
207. Congdon P. Bayesian methods and Bayesian estimation. Applied Bayesian Modelling,
Second edition. Chichester: John Wiley & Sons, Ltd; 2014. p. 1-33.
208. Hindmarsh DM. Small area estimation for health surveys [PhD thesis]. Wollongong: School
of Mathematics and Applied Statistics, University of Wollongong; 2013.
209. Pfeffermann D. New important developments in small area estimation. Statistical Science.
2013;28(1):40-68.
210. Chambers R, Clark R. An Introduction to Model-Based Survey Sampling with Applications.
Oxford: Oxford University Press; 2012.
211. Fay RE, Herriot RA. Estimates of income for small places: An application of James-Stein
procedures to census data. Journal of the American Statistical Association. 1979;74:269–277.
212. Bell WR. Accounting for uncertainty about variances in small area estimation. Bulletin of the
International Statistical Institute. 1999;52nd session, Helsinki.
213. Gelman A. Struggles with survey weighting and regression modeling. Statistical Science.
2007;22(2):153-164.

59
214. Gelman A. Rejoinder: Struggles with survey weighting and regression modeling. 2007:184-
188.
215. Cressie N, Wikle CK. Statistics for spatio-temporal data. Hoboken, NJ: Wiley; 2011.
216. Abellan JJ, Richardson S, Best N. Use of space-time models to investigate the stability of
patterns of disease. Environmental Health Perspectives. 2008;116(8):1111-1119.
217. Christakos G, Lai JJ. A study of the breast cancer dynamics in North Carolina. Social Science
& Medicine. 1997;45(10):1503-1517.
218. Marshall RJ. A review of methods for the statistical analysis of spatial patterns of disease.
Journal of the Royal Statistical Society Series A (Statistics in Society). 1991;154(3):421-441.
219. Bernardinelli L, Clayton D, Pascutto C, Montomoli C, Ghislandi M, Songini M. Bayesian
analysis of space-time variation in disease risk. Statistics in Medicine. 1995;14(21-22):2433-2443.
220. Songini M, Bernardinelli L, Clayton D, Montomoli C, Pascutto C, Ghislandi M, et al. The
Sardinian IDDM study: 1. Epidemiology and geographical distribution of IDDM in Sardinia during
1989 to 1994. Diabetologia. 1998;41(2):221-227.
221. Assuncao RM, Reis IA, Oliveira CD. Diffusion and prediction of Leishmaniasis in a large
metropolitan area in Brazil with a Bayesian space-time model. Statistics in Medicine. 2001;20(15):
2319-2335.
222. Bernardinelli L, Pascutto C, Best NG, Gilks WR. Disease mapping with errors in covariates.
Statistics in Medicine. 1997;16(7):741-752.
223. Xia H, Carlin BP. Spatio-temporal models with errors in covariates: Mapping Ohio lung
cancer mortality. Statistics in Medicine. 1998;17(18):2025-2043.
224. Schmid V, Held L. Bayesian extrapolation of space-time trends in cancer registry data.
Biometrics. 2004;60(4):1034-1042.
225. Waller L, Carlin BP, Xia H, Gelfand AE. Hierarchical spatio-temporal mapping of disease
rates. Journal of the American Statistical Association. 1997;92(438):607-617.
226. Knorr-Held L. Bayesian modelling of inseparable space-time variation in disease risk.
Statistics in Medicine. 2000;19(17-18):2555-2567.
227. Sun DC, Tsutakawa RK, Kim H, He ZQ. Spatio-temporal interaction with disease mapping.
Statistics in Medicine. 2000;19(15):2015-2035.
228. Kim H, Sun DC, Tsutakawa RK. A bivariate bayes method for improving the estimates of
mortality rates with a twofold conditional autoregressive model. Journal of the American Statistical
Association. 2001;96(456):1506-1521.
229. Kim H, Oleson JJ. A Bayesian dynamic spatio-temporal interaction model: An application to
prostate cancer incidence. Geographical Analysis. 2008;40(1):77-96.
230. Bohning D, Dietz E, Schlattmann P. Space-time mixture modelling of public health data.
Statistics in Medicine. 2000;19(17-18):2333-2344.
231. Knorr-Held L, Besag J. Modelling risk from a disease in time and space. Statistics in
Medicine. 1998;17(18):2045-2060.
232. Lagazio C, Biggeri A, Dreassi E. Age-period-cohort models and disease mapping.
Environmetrics. 2003;14(5):475-490.
233. Congdon P. A model framework for mortality and health data classified by age, area, and
time. Biometrics. 2006;62(1):269-278.
234. Smith AFM. Present Position and Potential Developments: Some Personal Views: Bayesian
Statistics. Journal of the Royal Statistical Society Series A (General). 1984;147(2):245-259.
235. Kafadar K. Choosing among two-dimensional smoothers in practice. Computational Statistics
and Data Analysis. 1994;18(4):419-439.
236. Griffith DA. Some guidelines for specifying the geographic weights matrix contained in
spatial statistical models. In: Arlinghaus SL, editor. Practical Handbook of Spatial Statistics. Boca
Raton, FL: CRC Press; 1996. p. 65–82.
237. Jerrett M, Gale S, Kontgis C. Spatial Modeling in Environmental and Public Health Research.
International Journal of Environmental Research and Public Health. 2010;7(4):1302-1329.
238. Pickle LW. A history and critique of U.S. mortality atlases. Spatial and Spatiotemporal
Epidemiology. 2009;1(1):3-17.

60
239. Kang SY, Cramb SM, White NM, Ball SJ, Mengersen KL. Making the most of spatial
information in health: a tutorial in Bayesian disease mapping for areal data. Geospatial Health.
2016;11(2):428.
240. Rytkönen MJP. Not all maps are equal: GIS and spatial analysis in epidemiology.
International Journal of Circumpolar Health. 2004;63(1):9-24.
241. Burrough PA, McDonnell R. Principles of Geographical Information Systems. Oxford:
Oxford University Press; 1998.
242. Elliot P, Wakefield JC, Best NG, Briggs DJ. Spatial Epidemiology: Methods and
Applications. Oxford: Oxford University Press; 2000.
243. English PB. An Introductory Guide to Disease Mapping. American Journal of Epidemiology.
2001;154(9):881-882.
244. López-Abente G, Aragonés N, García-Pérez J, Fernández-Navarro P. Disease mapping and
spatio-temporal analysis: importance of expected-case computation criteria. Geospatial Health.
2014;9:27-35.
245. Catelan D, Biggeri A. Multiple testing in disease mapping and descriptive epidemiology.
Geospatial Health. 2010;4:219-229.
246. Goovaerts P. Geostatistical analysis of disease data: visualization and propagation of spatial
uncertainty in cancer mortality risk using Poisson kriging and p-field simulation. International Journal
of Health Geographics. 2006;5:7.
247. International Agency for Research on Cancer. World Cancer Report 2014. Geneva: IARC,
World Health Organization, 2014.
248. Torre LA, Bray F, Siegel RL, Ferlay J, Lortet-Tieulent J, Jemal A. Global cancer statistics,
2012. CA: A Cancer Journal for Clinicians. 2015;65(2):87-108.
249. Australian Institute of Health and Welfare. Cancer in Australia: an overview. Cat. no. CAN
88. Canberra: AIHW, 2014.
250. Wilkinson D, Cameron K. Cancer and Cancer Risk in South Australia: What Evidence for a
Rural–Urban Health Differential? Australian Journal of Rural Health. 2004;12(2):61-66.
251. Woods LM, Rachet B, Coleman MP. Origins of socio-economic inequalities in cancer
survival: a review. Annals of Oncology. 2006;17(1):5-19.
252. Ernst J, Zenger M, Schmidt R, Schwarz† R, Brähler E. [Medical and psychosocial care needs
of cancer patients: a systematic review comparing urban and rural provisions]. Dtsch med
Wochenschr. 2010;135(31/32):1531-1537.
253. Mason TJ, McKay FW, Hoover R, Blot WJ, Fraumeni JF. Atlas of Cancer Mortality for US
Counties: 1950-1969. Washington: US Govt. Printing Office, 1975.
254. Koch T. Disease Maps: Epidemics on the Ground. Chicago: University of Chicago Press;
2011.
255. Borrell C, Marí-Dell’Olmo M, Serral G, Martínez-Beneito M, Gotsens M. Inequalities in
mortality in small areas of eleven Spanish cities (the multicenter MEDEA project). Health & Place.
2010;16(4):703-711.
256. Clayton DG, Kaldor J. Empirical Bayes estimates of age-standardised relative risks for use in
disease mapping. Biometrics. 1987;43.
257. Cressie N, Chan NH. Spatial Modeling of Regional Variables. Journal of the American
Statistical Association. 1989;84(406):393-401.
258. Ancelet S, Abellan JJ, Del Rio Vilas VJ, Birch C, Richardson S. Bayesian shared spatial-
component models to combine and borrow strength across sparse disease surveillance sources.
Biometrical Journal. 2012;54(3):385-404.
259. Bernardo J, Smith A. Bayesian Theory. Chichester, UK: John Wiley & Sons, Ltd, 2000.
260. Johnson GD. Small area mapping of prostate cancer incidence in New York State (USA)
using fully Bayesian hierarchical modelling. International Journal of Health Geographics. 2004;
3(1):29.
261. Gelman A, Carlin JB, Stern HS, Rubin DB. Bayesian Data Analysis. Third edition. Boca
Raton, FL: Chapman & Hall/CRC; 2014.
262. Gurrin LC, Kurinczuk JJ, Burton PR. Bayesian statistics in medical research: an intuitive
alternative to conventional data analysis. Journal of Evaluation in Clinical Practice. 2000;6(2):193-
204.

61
263. Wakefield J. Disease mapping and spatial regression with count data. Biostatistics. 2007;8(2):
158-183.
264. ApSimon HM, Warren RF, Kayin S. Addressing uncertainty in environmental modelling: a
case study of integrated assessment of strategies to combat long-range transboundary air pollution.
Atmospheric Environment. 2002;36(35):5417-5426.
265. Li Y, Brown P, Gesink DC, Rue H. Log Gaussian Cox processes and spatially aggregated
disease incidence data. Statistical Methods in Medical Research. 2012;21(5):479-507.
266. Gotway CA, Young LJ. Combining Incompatible Spatial Data. Journal of the American
Statistical Association. 2002;97(458):632-648.
267. Yu B. A Class of Transformation Covariate Regression Models for Estimating the Excess
Hazard in Relative Survival Analysis. American Journal of Epidemiology. 2013;177(7):708-717.
268. Congdon P. Estimating diabetes prevalence by small area in England. Journal of Public
Health. 2006;28(1):71-81.
269. Thomas DC. Statistical Methods in Environmental Epidemiology. Oxford: Oxford University
Press; 2014.
270. Wakefield J. Sensitivity analyses for ecological regression. Biometrics. 2003;59(1):9-17.
271. Wakefield J. Ecological inference for 2 × 2 tables (with discussion). Journal of the Royal
Statistical Society: Series A (Statistics in Society). 2004;167(3):385-445.
272. Gardner W, Mulvey EP, Shaw EC. Regression analyses of counts and rates: Poisson,
overdispersed Poisson, and negative binomial models. Psychological Bulletin. 1995;118(3):392-404.
273. Browning CR, Cagney KA, Wen M. Explaining variation in health status across space and
time: implications for racial and ethnic disparities in self-rated health. Social Science & Medicine.
2003;57(7):1221-1235.
274. Fahrmeir L, Kneib T. Bayesian Smoothing and Regression for Longitudinal, Spatial and
Event History Data. Oxford: Oxford University Press; 2011.
275. Wang F. Quantitative Methods and Applications in GIS. Boca Raton, FL: CRC Press; 2006.
276. Paciorek CJ. Spatial models for point and areal data using Markov random fields on a fine
grid. Electronic Journal of Statistics. 2013;7:946-972.
277. Lawson AB, Browne WJ, Vidal Rodeiro CL. Disease Mapping with WinBUGS and MLwiN.
Chichester: John Wiley & Sons, Ltd; 2004.
278. Clayton D, Bernardinelli L. Bayesian methods for mapping disease risk. In: Elliott P, Cuzick
J, English D, Stern R, editors. Geographical and Environmental Epidemiology: Methods for Small
Area Studies. Oxford: Oxford University Press; 1996. p. 205-220.
279. Mollié A. Bayesian mapping of disease. In: Gilks WR, Richardson S, Spiegelhalter DJ,
editors. Markov Chain Monte Carlo in Practice. London: Chapman & Hall; 1996. p. 359–379.
280. Wakefield JC, Best NG, Waller LA. Bayesian approaches to disease mapping. In: Elliot P,
Wakefield JC, Best NG, Briggs DJ, editors. Spatial Epidemiology: Methods and Applications.
Oxford: Oxford University Press; 2001. p. 104–127.
281. Assunção R, Krainski E. Neighborhood Dependence in Bayesian Spatial Models. Biometrical
Journal. 2009;51(5):851-869.
282. Junaidi, Stojanovski E, Nur D, editors. Prior sensitivity analysis for a hierarchical model.
Proceedings of the Fourth Annual ASEARC Conference, 17–18 February 2011; 2011; University of
Western Sydney, Paramatta, Australia.
283. Crainiceanu CM, Ruppert D, Wand MP. Bayesian Analysis for Penalized Spline Regression
Using WinBUGS. Journal of Statistical Software. 2005;14:1-24.
284. Lunn D, Jackson C, Best N, Thomas A, Spiegelhalter D. The BUGS book: A Practical
Introduction to Bayesian Analysis. Boca Raton, FL: Chapman and Hall/CRC Press; 2012.
285. Lykou A, Ntzoufras I. WinBUGS: a tutorial. Wiley Interdisciplinary Reviews: Computational
Statistics. 2011;3(5):385-396.
286. Spiegelhalter D, Thomas A, Best N, Lunn D. WinBUGS User Manual, version 1.4. London:
MRC Biostatistics Unit, Cambridge and Imperial College School of Medicine, 2003.
287. The R Core Team. R: A language and environment for statistical computing. Vienna, Austria:
R Foundation for Statistical Computing, 2012.

62
288. Rue H, Martino S, Chopin N. Approximate Bayesian inference for latent Gaussian models by
using integrated nested Laplace approximations. Journal of the Royal Statistical Society: Series B
(Statistical Methodology). 2009;71(2):319-392.
289. Schrödle B, Held L. A primer on disease mapping and ecological regression using INLA.
Computational Statistics. 2011;26(2):241-258.
290. Schrödle B, Held L. Spatio-temporal disease mapping using INLA. Environmetrics.
2011;22(6):725-734.
291. Blangiardo M, Cameletti M, Baio G, Rue H. Spatial and spatio-temporal models with R-
INLA. Spatial and Spatiotemporal Epidemiology. 2013;4:33-49.
292. Rue H MS, Lindgren F. The R-INLA Project. www.r-inla.org. 2012.
293. Anselin L, Syabri I, Kho Y. GeoDa: An Introduction to Spatial Data Analysis. In: Fischer
MM, Getis A, editors. Handbook of Applied Spatial Analysis: Software Tools, Methods and
Applications. Berlin, Heidelberg: Springer Berlin Heidelberg; 2010. p. 73-89.
294. Bivand R, Altman M, Anselin L, Assunção R, Berke O, Bernat A, et al. spdep: Spatial
dependence: weighting schemes, statistics and models. R package version 0.6-5. 2011.
295. Athens JK, Catlin BB, Remington PL, Gangnon RE. Using Empirical Bayes Methods to Rank
Counties on Population Health Measures. Preventing Chronic Disease. 2013;10:E129.
296. Jaccard J, Becker MA, Wood G. Pairwise multiple comparison procedures: A review.
Psychological Bulletin. 1984;96(3):589-596.
297. Brewer CA, Pickle L. Evaluation of Methods for Classifying Epidemiological Data on
Choropleth Maps in Series. Annals of the Association of American Geographers. 2002;92(4):662-681.
298. Lawson AB, Biggeri AB, Böhning D, Lesaffre E, Viel J-F, Bertollini R, editors. Disease
Mapping and Risk Assessment for Public Health. West Sussex: John Wiley & Sons; 1999.
299. McNamee R. Regression modelling and other methods to control confounding. Occupational
and Environmental Medicine. 2005;62(7):500-506.
300. Statewide Health Service Strategy and Planning Unit. Cancer care services statewide health
service strategy 2014. Brisbane: Queensland Health, 2014.

63
64
Appendix A Glossary
Asymptotic Here, referring to large sample.

Bandwidth In kernel smoothing, it refers to the maximum distance from an

area that its influence is expected to extend. Beyond this, the
kernel is set to zero (see Box 2.8).

Boundary effects Areas on the ‘edge’ of the analysis region have fewer neighbours
as neighbours beyond the boundary are excluded.

Choropleth map Displays the value of interest on a set of regions within the study
area.

Convergence In MCMC analysis, the point at which it is reasonable to believe

that samples are truly representative of the underlying stationary
distribution of the Markov chain.

Covariance A measure of the dependence between two random variables,

and how they change together (see Box 2.7).

Covariate In statistics, a covariate is a variable that is possibly predictive of

the outcome under study. A covariate may be of direct interest or
it may be a confounding or interacting variable.

Cross-validation A model validation technique assessing how the results from a

statistical analysis will generalise to an independent data set.

Direct method of Apply stratum-specific rates observed in the populations of

standardisation interest to a standard population. The ratio of two directly
standardised rates is called the comparative incidence ratio (see
Box 2.5).

Gaussian An alternative term for the Normal distribution, which is a

symmetrical bell-shaped curve.

Geostatistical data Point-referenced data.

Gibbs sampling See Box 2.20. A common form of MCMC sampling.

Hamiltonian Monte Original name is Hybrid Monte Carlo, as this is a hybrid

Carlo between traditional dynamical simulation and the Metropolis
algorithm. Used by Stan software.

65
Hidden Markov model Represents probability distributions over sequences of
observations. Assumes that the state of the process generating
the data is hidden, and that the current state is independent of all
others except for the one immediately prior to it.

Hierarchical model A model written in a hierarchical form or in terms of sub-

models.

Hyperparameter A parameter in a prior distribution.

Hyperprior distribution A prior distribution on a hyperparameter, i.e., on a parameter of

a prior distribution.

Incidence A measure of the risk of developing a disease within a specified

period of time.

Indirect method of Apply stratum-specific reference rates to the populations of

standardisation interest. The ratio of two indirectly standardised rates is called
the SMR (see Box 2.6).

Kernel function A kernel is a weighting function used in non-parametric

estimation techniques. Common types of kernel functions
include uniform, triangular, Gaussian, quadratic and cosine (see
Box 2.8).

Likelihood The probability of the evidence given the parameters. It is the

probability of a given sample being randomly drawn regarded as
a function of the parameters of the population.

Marginal mixture A model based on the assumption that the total space can be split
models into local regions where the responses come from the same
distribution. Similar to spatial partition models.

Markov chain A mechanism for generating plausible parameter value, whereby

the value to be drawn depends on the previously drawn value.

Markov chain Monte A class of algorithms for sampling from probability distributions
Carlo (MCMC) by constructing a Markov chain that has the desired distribution
as its equilibrium distribution (see Box 2.20).

Monte Carlo methods A broad class of computational algorithms that use repeated
random sampling to obtain numerical results.

66
Non-parametric model The structure of the model is not fixed, but determined by the
data. The number and nature of parameters are flexible.
(Compare against parametric.)

Over-dispersion In the statistical context, the presence of greater variability in a

data set than expected using a given statistical model.

Parameter A value used to represent a certain population characteristic

which is usually unknown and therefore has to be estimated.

Parametric model Assumes there is an underlying probability distribution based on

a fixed set of parameters.

Posterior distribution A probability distribution on the values of an unknown

parameter that combines prior information about the parameter
contained in the observed data to give a composite picture of the
final judgements about the values of the parameter.

Predictor A predictor variable is also known as an independent variable.

Prevalence The number or proportion of cases or events or conditions in a

given population.

Prior distribution A probability distribution that represents the uncertainty about

the parameter before the current data are examined.

Random effects Effects that account for differences among the individual
observational units in the sample, which are randomly sampled
from the population. These effects usually conform to a
specified distribution (typically a Normal distribution) and have
a mean of zero.

Regression A statistical technique for estimating the relationships among

variables.

Relative survival A standard estimate of net survival (measuring survival from the
disease of interest) in population based disease survival studies
(see Box 2.18).

Risk factors An aspect of personal behaviour or lifestyle, an environmental

exposure, or an inborn or inherited characteristic that is
associated with an increased occurrence of disease or other
health-related event or condition.

67
 Semi-parametric model A model containing parametric and nonparametric components.
Often the nonparametric components are not of interest, such as
the baseline hazard in the Cox proportional hazards model.

Sensitivity checks These check the influence of model inputs (such as prior
distributions) on the variation in model output.

Spatial partition models A model based on the assumption that the total space can be split
into local regions where the responses come from the same
distribution. Similar to marginal mixture models.
Variance A measure of how far values are spread out from their mean.
The variance is the square of the standard deviation and the
covariance of a random variable with itself (see Box 2.7).

Note: Several of these definitions were obtained from or based on the glossary in (239).

68
Appendix B Bayesian disease mapping tutorial
The following is from (239):

Making the most of spatial information in health: A tutorial in Bayesian disease

mapping for areal data

Abstract

Disease maps are effective tools for explaining and predicting patterns of disease outcomes
across geographical space, identifying areas of potentially elevated risk, and formulating and
validating aetiological hypotheses for a disease. Bayesian models have become a standard
approach to disease mapping in recent decades. This article aims to provide a basic
understanding of the key concepts involved in Bayesian disease mapping methods for areal
data. It is anticipated that this will help in interpretation of published maps, and provide a
useful starting point for anyone interested in running disease mapping methods for areal data.
The article provides detailed motivation and descriptions on disease mapping methods by
explaining the concepts, defining the technical terms, and illustrating the utility of disease
mapping for epidemiological research by demonstrating various ways of visualising model
outputs using a case study. The target audience includes spatial scientists in health and other
fields, policy or decision makers, health geographers, spatial analysts, public health
professionals, and epidemiologists.

Introduction

Disease mapping is a flourishing field due to the growing amount of routinely collected
health information worldwide (240). Advances in geographic information systems have
greatly aided the analytical manipulation and visual representation of spatial data (241).
Spatial information in health is especially useful for informing the locations of disease
occurrences and the onus is on making the best possible use of this information.

Some excellent introductory guides for disease mapping are available in the literature.
Nonetheless, many of these are either not intended for non-statistical audiences, or lack
specific details. For instance, Elliot et al. (242) present a comprehensive review of the recent
developments in spatial epidemiology but the statistical methods require a level of
background knowledge which may not be suitable for beginners. Marshall (218) covers a
broad range of methods for the analysis of the geographical distribution of disease, rather
than upskill the reader in using particular methods. Lawson and Williams (39) provide a
broad overview of the issues concerning disease mapping but is short on specifics (243).
Banerjee et al. (44) presents a fully model-based approach to all types of spatial data,
including point level, areal, and point pattern data. Cramb et al. (159) offer insight into the
decisions made in generating a health atlas, but is not intended as an entry-level article for a

69
non-statistical audience. This article fills the niche by providing motivation, definition and
description at a general level, and illustrating these ideas via a substantive case study.
Although disease mapping has been undertaken in various forms for over 100 years, the
opportunity now exists to use model-based maps that acknowledge uncertainty in inputs and
outputs (244, 245), take account of the spatial nature of the data to ‘borrow strength’ from
neighbouring areas in order to improve small area estimates, and can provide probability
statements (246). In this article, we describe Bayesian disease mapping for areal data (39, 55)
as an approach that addresses these issues. We focus on a running example of mapping
cancer, although the methods are applicable to other diseases.

The primary purpose of this article is to provide a basic understanding of the key concepts
involved in Bayesian statistical models for disease mapping of areal data. We commence with
a discussion of why disease model-based mapping methods are required. Background on
Bayesian methods typically used for disease mapping is then provided, and then some of the
cartographic outputs commonly used are discussed, including methods for indicating
statistical uncertainty in relative risk of disease.

Case Study: Cancer in Australia

Cancer is now the world’s and Australia’s biggest killer (247). The number of cases
diagnosed continues to increase worldwide due to population growth and aging, with the
increasing prevalence of physical inactivity, poor diet and reproductive changes (such as later
parity) also contributing (248). In Australia, cancer accounts for almost one-fifth (19%) of the
total disease burden (249).

Disparities in cancer outcomes across broad socioeconomic status and urban/rural categories
have been reported internationally (250-252). Within Australia, there are disparities in cancer
outcomes with respect to geographic remoteness and socioeconomic status (249). Cancers
such as cervical and lung had higher incidence and mortality as remoteness or area-level
disadvantage increased. Furthermore, the five-year relative survival from all cancers
combined decreased with greater remoteness and greater socioeconomic disadvantage.

Understanding disparities in these broad areas, while useful, is unlikely to accurately reflect
the heterogeneity in outcomes at the local level. Efforts to monitor and reduce cancer
disparities can benefit greatly from quantifying variation across population groups and
pertinent, small geographical areas. An understanding of the geographic patterns of cancer
enables health decision-making by health service planners, clinicians, epidemiologists and
industry groups to be more accurate and effective, for example by targeting policy
development and resource allocation at areas of greater need (22, 253).

Cramb et al. (154) produced the first Atlas of Cancer in Queensland to describe geographical
variation in cancer incidence and survival across small areas in Queensland, using routinely-
collected health information from the Queensland Cancer Registry. For the first time,

70
Bayesian model-based cancer incidence and survival maps for Queensland were
systematically presented at a comprehensive level. The Atlas significantly contributed to the
understanding of geographical variation of cancer incidence and survival across Queensland,
and subsequently influenced government policy decisions.

Methods

Disease maps are a visual representation of disease outcomes. The use of disease maps to aid
decision making in epidemiological and medical research is well recognized (254). Disease
maps are effective tools for explaining and predicting patterns of disease outcomes across
geographical space, identifying areas of potentially elevated risk, and formulating and
validating aetiological hypotheses for a disease (4). They are able to uncover local-level
inequalities frequently masked by health estimates from large areas such as states, regions or
cities (255), enabling the development of disease reduction and prevention programs
targeting high-risk populations, see for instance, Mason et al. (253) and Kulldorff et al. (22)
who have used cancer maps to depict the geographic patterns of cancer outcomes.

Disease mapping encompasses small area studies that use data aggregated over small areas
and take into account local spatial correlation, see for example, Clayton and Kaldor (256);
Cressie and Chan (257); Besag et al. (81) and Bernardinelli et al. (222). Data sparseness is
common in small area analyses, especially when working with less common diseases. A
small number of observed and expected disease occurrences leads to unstable risk estimates
(258).

The problem of potentially unstable risk estimates for sparse spatial data needs to be
mitigated to obtain reliable estimates. In practice, this is achieved by implementing spatial
smoothing techniques. Spatial smoothing effectively “borrows strength” across small areas,
so that the disease rate estimated for an area with a small population denominator would be
weighted towards the estimated disease rate of neighbouring areas that have larger
denominators. The estimates obtained by smoothing information from neighbouring areas are
more reliable and robust due to the increased precision in the risk estimates in areas with few
observations (258). In the context of disease mapping for small areas, the implementation of
spatial smoothing is commonly achieved via the incorporation of a conditional autoregressive
prior distribution for the spatial effects (see Lee (94) and the “Bayesian Spatial Statistical
Models” section for details).

A disease mapping model is essentially a regression model that links a disease outcome to a
set of risk factors. An important concept in disease mapping models (which is common to
many other regression models) is the use of random effects. In this context, random effects
provide a way of estimating variation in disease risk between areas that is not otherwise
captured by known risk factors (e.g. age, sex, socioeconomic status, etc.).

71
Why Bayesian?

Bayesian statistics takes its name from the English clergyman Thomas Bayes (1702-1761),
although the key concepts were also contemporaneously established by Laplace and
embedded in the general view of ‘inverse probability’ at that time (259). It is an approach to
data analysis that focuses on relating observed and unknown quantities using conditional
probabilities, which are measures of the probability of an event given that another event has
occurred.

In a Bayesian model (Box B.1), an unknown parameter is represented using a distribution

rather than a single point estimate (260). The model parameters have distributions and are
probabilistic (e.g. parameters representing coefficients associated with covariates in a
regression model might be given a Normal distribution (Box B.2)). These distributions are
known as prior distributions. These prior distributions can be considered as representing the
uncertainty about the parameter before the data are seen. The parameters in the prior
distributions (e.g. the mean and variance of the prior on a regression coefficient) can also
have distributions which are known as hyperprior distributions. Again, these distributions
also represent uncertainty about our knowledge of these values.

The combination of the prior information and the data results in a posterior distribution. The
posterior distribution can be thought of as a probability distribution on the values of an
unknown parameter that combines prior knowledge about the parameter and the observed
data. The Bayesian model thus consists of parameters related to one another in the form of a
hierarchy. The complex nature of spatial data can be captured using this hierarchical structure
(4, 87).

Box B.1 Bayesian model

Given Bayes’ theorem (261),

𝑷(𝑨|𝑩) ∝ 𝑷(𝑨)𝑷(𝑩|𝑨)
The posterior distribution (𝑷(𝑨|𝑩)) is proportional to the prior distribution for parameters
(𝑷(𝑨)) multiplied by the data-based distribution given parameters (also known as the
likelihood, 𝑷(𝑩|𝑨)).
◦ Posterior estimates (model output) are a combination of the prior information and the data.
◦ Parameters in the model are assigned prior distributions.
◦ A prior distribution is the probability distribution that represents the uncertainty about the
parameter before the current data are examined.
◦ Parameters in the prior distribution can also be assigned distributions.
◦ Parameters in the prior distribution (called ‘hyperparameters’) can also be assigned
distributions.

72
Random effects are generally included in these models. Typically, a random effect is
specified as being normally distributed, whereby a few areas are allowed to have a disease
incidence much lower than expected based on these risk factors, a few areas much higher, but
most are close to expected (following a bell curve). For spatial data, we assume that sites
closer to each other are more similar, so we can use information from neighbouring sites to
obtain better estimates of disease risk. Hence, when we fit a spatially-correlated random
effect, the variation at a particular site is normally distributed relative to the mean of its
neighbours. These random effects thus relate disease risk estimates to neighbouring
estimates, producing a ‘smoothing’ effect across the area of interest.

Box B.2 Normal distribution

A distribution contains information on every possible observation and its associated

probability. For instance, a Normal distribution is a continuous distribution that is “bell-
shaped”, at which data are most likely to be distributed around the mean and are less likely to
be farther away from the mean.

A Normal distribution is often specified in terms of its mean (𝜇) and variance (𝜎 2 ) and can
be written in the form of Normal(𝜇, 𝜎 2 ). A parameter can be assigned a Normal distribution
with mean 0 and variance 100 which can be denoted as Parameter~Normal(0,100).

Alternatively, instead of specifying the values (0,100), uncertainty about these parameters
can also be described probabilistically. For example, instead of specifying ‘100’ for the
variance, the prior distribution could be written as Normal(𝜇, 𝜎02 ), and 𝜇 set to 0 while 𝜎02 is
described by another probability distribution. Here 𝜎02 is termed a hyperparameter and the
distribution on 𝜎02 a hyperprior distribution.

There are many reasons why the Bayesian approach is a useful framework for disease
mapping. Firstly, Bayesian smoothing methods produce robust and reliable estimation of
health outcomes of interest in a small area, even when based on small sample sizes (258).
Within these small areas, the sample sizes are sometimes too small to yield estimates with
adequate precision and reliability. Bayesian smoothing techniques improve the estimation by
using information from neighbouring areas.

Secondly, the use of prior distributions (usually based on existing knowledge or expert
opinion) in disease mapping models helps strengthen inferences about the true value of the
parameter and ensures that all relevant information is included (262). These can be
‘uninformative’ (e.g. set to be Normally distributed with a mean of zero and a very large
variance) or ‘informative’ if there is other information about the effect of this risk factor
(given the other risk factors in the model). Thirdly, the Bayesian approach allows for
quantification of the uncertainty related to the health estimates from the posterior
distributions (67, 263). Spatial uncertainties added to the resulting risk maps depict local

73
details of the spatial variation of the risk and provide valuable information for policy makers
to make decisions about thresholds and public health (246, 260, 264).

Lastly, direct probabilistic statements can be made about the underlying and unobserved
parameters of interest using their posterior probability distributions. In disease mapping, it
might be of interest to make probability statements about areas of high risk for a disease. For
instance, computing and mapping probabilities that the risk in an area exceeds certain
thresholds can be done using the posterior probability distributions (101). This probability of
exceedance can then be used to decide whether an area should be classified as having excess
risk of a disease (102). It is straightforward to make these kind of statements in a Bayesian
context, since they are directly obtained from the corresponding posterior distribution.

Box B.3 Selecting regional scale

Important questions to consider when deciding on an appropriate area scale to conduct the
analysis include:

1. Is there a risk of patient confidentiality being compromised?

2. Are population data available at the same scale as disease occurrences?
3. Will boundaries change over time? If so, what options are possible for keeping your data
consistent?
4. Is there a digital boundary file available?
5. Will areas have a practical and relevant interpretation?
6. How does the size of the areas compare relative to the spatial pattern of the variation? If
there is a lot of variation in an environmental effect within areas, this will limit the scope to
measure the effect.
7. How many areas will there be? This affects computational time.
8. Are some areas likely to have zero population? This is likely to cause difficulties in
modelling and estimation, e.g., zero denominator causes difficulties when using a Poisson
distribution.
9. What scale have other similar studies used?
10. What spatial scale is available for covariate data? If spatial variation that takes fixed
effects into account is of interest, it is not necessary to have a spatial scale finer than the
available covariate data.

Data

Often health data are only available with location data supplied as a small area (known as
areal data), rather than a street address geocoded to a latitude/longitude point. Determining
the most appropriate region size to use involves several considerations (Box B.3). This article

74
focuses on the application of disease mapping methods for areal data aggregated over small
areas and omits the discussion of other forms of spatial data such as geostatistical and point
patterns data. As an alternative, health outcome data may also be analysed at the individual
level, while incorporating spatial information at any geographical scale such as a point or an
area.

The data described in the Atlas (154) focused on Queensland cancer data aggregated to the
SLA level, which was the smallest area with annual population data available. However,
consistent with most administrative regions, the areas are of varying sizes, and larger areas
tend to dominate the map. An alternative approach is to aggregate disease data with
continuous coordinate information to regular grid cells; see Li et al. (196, 265) and Kang et
al. (10). Such an approach allows modelling of disease data at a fine spatial scale,
independent of administrative boundaries while preserving patient confidentiality. Using this
approach, the spatial scale can be manipulated to a practically, geographically and
computationally sensible scale. It does, however, require individual level geocoded data,
which may not be accessible due to confidentiality concerns. Spatial data may also be
available at various geographical scales and hence there is a need to combine information
from multiple sources (see Gotway and Young (266) for further details).

Box B.4 Data required to produce incidence estimates

Given a disease of interest, the information required to produce incidence estimates includes:
◦ Number of disease cases among people within a certain time period for each small area
◦ Estimated population counts by age group, sex, year and small area of residence − this is
used as the denominator for calculating rates and for age-standardisation
◦ Geographical boundaries − this is used to compute the adjacency matrix required for spatial
smoothing
◦ Optional: any desired small area level covariates (if available) such as rurality and
socioeconomic status

Cramb et al. (154) mapped two health outcome measures in the Atlas, namely the incidence
estimates and the relative survival estimates (discussed in the following Section). Incidence is
a measure of the risk of developing a disease within a specified period of time. Relative
survival is the standard measure of survival from a disease in population-based disease
survival studies (267). Each of these outcomes require specific input data (refer to Boxes B.4
and B.5).

Although other estimates of disease, such as prevalence, are beyond the scope of this article,
Bayesian mapping approaches are described in Congdon (268).

75
Box B.5 Data required to produce survival estimates

To produce relative survival estimates of a disease of interest, the input data required include:
◦ From the patients with the disease of interest (if not available for each individual then
aggregated over each small area, any covariates and follow-up time intervals):
− The observed number of deaths (from any cause) within a certain time period
− Person-time at risk (the length of time between diagnosis and either death or censoring)
◦ General population mortality data used to calculate the expected number of deaths, which
represents deaths due to causes other than the disease of interest for each small area, sex and
broad age group
◦ Geographical boundaries − this is used to compute the adjacency matrix required for spatial
smoothing
◦ Optional: individual or area-level covariates, including age, tumour stage, or area rurality
and socioeconomic status

Bayesian Spatial Statistical Models

A response variable is the event studied and expected to vary whenever the independent
variable is altered. It is also known as a dependent variable. Here we consider two response
variables, namely the number of cancers diagnosed (incidence model) and the number of
deaths within x years of a cancer diagnosis (relative survival model). Because both response
distributions are counts, and the disease is less common, a Poisson distribution is used to
model them (Box B.6).

Box B.6 Probability distributions used in epidemiology

For common diseases, the Binomial distribution models the number of disease occurrences in
a sample size n from a population size N. The Binomial distribution is also commonly used in
the analysis of disease prevalence data and case-control studies (269).
◦ When the disease is rare or less common (i.e., the probability of a disease is small), the
Poisson distribution is used as an approximation to a Binomial distribution (270, 271). A
Poisson distribution expresses the probability of a given number of events occurring in a fixed
interval of time and/or space.
◦ For over-dispersed count distributions (where the data admit more variability than expected
under the assumed distribution), a Negative Binomial distribution may be appropriate (272).
◦ For empirical data that show more zeroes than would be expected, zero-inflated models may
be employed (272).

76
The resulting estimate for the incidence of a disease is known as the standardised incidence
ratio or SIR, which is an estimate of relative risk within each area based on the population
size, that compares the observed incidence against the expected incidence. The SIR explains
if the observed incidence in a particular area is higher or lower than the average across all
areas included, given the age and sex distribution and population size of the area.

The relative survival of a disease is modelled using an excess mortality model that contrasts
the mortality in the background population with disease mortality. The survival model results
in an excess hazard, which is called the relative excess risk (RER). The RER informs the
relative survival of a disease within each area, by reporting the risk of death within a certain
number of years of diagnosis after adjusting for broad age groups, compared to the average.

Small-area disease data typically exhibit spatial correlation due to spatial structure in the
unknown risk factors. The presence of spatial correlation can be caused by a combination of
socio-demographic clustering and environmental effects (273). Traditional regression models
assume independence of random effects and so ignore the potential presence of spatial
correlation. This may lead to false conclusions regarding covariate effects and unstable risk
estimates (274).

The spatial correlation can be accounted for using spatial smoothing techniques, by
estimating the effect of interest at a location using the effect values at nearby locations (275).
Spatial smoothing approaches based on neighbourhood dependence are widely employed in
disease mapping where areas with a common boundary are treated as neighbours (276). By
accounting for the spatial correlation, model inference, prediction and estimation can be
improved (143). The effect of the arbitrary geographical boundaries can also be reduced via
spatial smoothing. Other smoothing techniques include interpolation methods, kernel
regression, kriging and partition methods (61, 277).

Figure B.1 The representation of neighbourhood structure of area i.

Note: Based on the Rook method, neighbours for area i include areas 2, 4, 6 and 8, while the Queen method
defines regions 1 − 8 as neighbours of area i.

77
Two popular ways of defining a neighbourhood structure for the modelling of spatial
correlation are the Queen definition and the Rook definition. The Rook method defines that
two areas are considered neighbours if they share a common boundary whereas the Queen
method specifies that two areas are termed neighbours if they share a common boundary or
vertex. Following Earnest et al. (38), the illustration of these two methods for defining a
neighbourhood structure is given in Figure B.1. Such information can be used to calculate the
average of spatially correlated random effects of neighbours for area i.

The following Bayesian spatial models take the spatial correlation into account by
incorporating spatially correlated random effects. Both the incidence and relative survival
models assume a Poisson distribution for the observed data and contain spatial and
unstructured (non-spatial) random effects. The well-known Bayesian BYM model (81) is
widely used to model disease incidence (Box B.7) as it has desirable properties for disease
mapping, particularly in modelling the geographical dependence between neighbouring areas
(87). The incidence model can also be used to model mortality.

Box B.7 The incidence model

Given a set of n areas, the model for area i (i = 1,, n) can be written as follows:
Observed counts in area i ∼ Poisson(expected counts of area i × SIR of area i)
log(SIR of area i) = intercept term + coefficient × predictor variable vector for area i +
spatial random effect of area i + unstructured random effect of area i.
Apply stratum-specific reference rates to the populations of interest.
The ratio of two indirectly standardised rates is called the SIR.

With regard to relative survival, the excess mortality can be modelled via a GLM, using exact
survival times (121). The excess mortality is the mortality that is attributable to a particular
disease. It is a measure of the deaths which occur over and above those that would be
expected for a given population. Such a Bayesian relative survival model (Box B.8) has been
used by Fairley et al. (120) and Cramb et al. (154). See Boxes 2.14 and 2.18 for the statistical
models for incidence and relative survival, respectively.

In both models, the spatial random effect is the component that accounts for spatial
correlation between neighbouring areas. The unstructured or non-spatial random effect
accounts for the unexplained variation in the model.

In a Bayesian analysis, it is assumed that all parameters arise from a probability distribution.
As such, distributions representing the likely spread of values are placed on each parameter.
Commonly, a vague Normal distribution such as one with mean 0 and variance 1.0 × 106 or
Normal(0, 1.0 × 106 ) is used for the intercept or coefficients of predictor terms. Vague priors

78
refer to distributions with high spread, such as a Normal distribution with extremely large
variance. Such a distribution gives similar prior value over a large range of parameter values.

Box B.8 The relative survival model

The model can be written as below, where for area i, follow-up interval j, and age group k,
Number of deaths𝑖𝑗𝑘 ~Poisson(expected number of deaths𝑖𝑗𝑘 )
log(expected number of deaths𝑖𝑗𝑘 − expected number of deaths due to other causes𝑖𝑗𝑘 )
= log(person time at risk 𝑖𝑗𝑘 ) + intercept 𝑗 + coefficient 𝑘
× predictor variable vector + spatial random effect of area𝑖
+ unstructured random effect of area𝑖

Generally, the unstructured (non-spatial) random effects and the spatial random effects are
both assigned a prior distribution with additional hyperparameters (Box B.9). To allow for
spatial correlation, commonly an intrinsic conditional autoregressive (CAR) distribution is
used. The CAR prior models the spatial dependence in a study region by effectively
borrowing information from neighbouring areas than from distant areas and smoothing local
rates toward local, neighbouring values. The method provides some shrinkage and spatial
smoothing of the raw relative risk estimates (69). This results in a more stable estimate of the
pattern of the underlying disease risk than that provided by the raw estimates. Consequently,
the variance in the associated estimates is reduced and the spatial effect of geographical
differences can be identified. This prior has been widely employed in disease mapping to
study the geographical variation of disease risk (278-280), and works particularly well to
smooth out variability not relevant to the underlying risk (281).

Box B.9 Prior distributions for the random effects

Unstructured
The unstructured random effects are assumed to follow a Normal distribution with mean zero
and a hyperparameter for variance.

Unstructured random effect of area i ∼ Normal(0, variance hyperparameter).

Spatial
The spatial random effects are assumed to follow a CAR prior (81) with some
hyperparameters, as follows:
Spatial random effect of area i ∼ Normal (average of spatial effects of neighbours of area i,
variance hyperparameter / number of neighbours of area i).

79
Commonly, both of the precision (inverse of the variance) hyperparameters are assigned a
Gamma distribution. Alternative hyperprior distributions may include placing either a
Uniform or half-Normal distribution on the standard deviation (square root of the variance)
(54).

The prior distributions used for the parameters may influence the results and therefore should
be carefully considered and compared. There are two issues to consider when deciding on a
prior distribution (54): (a) what information is going into the prior distribution; and (b) the
impact on the resulting posterior distribution. A sensitivity analysis (282) can be used to
investigate the dependence of the posterior distribution on prior distributions by comparing
posterior inferences under different reasonable choices of prior distribution. A literature
review is usually helpful to determine the prior distributions being used in similar Bayesian
models.

Computation

The complexity of these models mean they cannot be solved analytically. Instead, some
method of approximation is required. One approach is to use Markov chain Monte Carlo
(MCMC) methods, which sample from the posterior distribution. A variety of software is
available to conduct MCMC, including BUGS (Bayesian inference Using Gibbs Sampling),
JAGS (Just Another Gibbs Sampler), Stan and BACC (Bayesian Analysis, Computation &
Communication). WinBUGS is one of the most popular options (134) that provides great
flexibility in Bayesian modelling, has a simple programming language (283) and interfaces
with multiple statistical software, including R, Matlab, Stata and SAS. See Additional
Information B.1 for the WinBUGS code for the discussed models. Some useful resources to
help learn WinBUGS include Lawson et al. (277), Lunn et al. (284), Ntzoufras (206), Lykou
and Ntzoufras (285), and Spiegelhalter (286).

Bayesian computation for the above models can also be conducted in R (287), by calling the
inla program and adopting the integrated nested Laplace approximation (INLA) approach
proposed by Rue et al. (288). The INLA approach performs Bayesian inference for spatial
models and is able to return accurate parameter estimates in a much shorter time than
MCMC. The use of R-INLA for statistical analysis in various disciplines is increasingly
common in recent years, including disease mapping. Additional Information B.3 provides R-
INLA code to perform computation for the discussed models. Some useful resources for
getting started with R-INLA include Schrödle and Held (289, 290), Blangiardo et al. (291),
and Rue et al. (292).

To incorporate neighbourhood dependence into the Bayesian models, a neighbourhood

matrix is required. The neighbourhood matrix contains a list of neighbours for an area. Freely
available software programs that will calculate a neighbourhood matrix include GeoDa (293),
the spdep R package (294), or within WinBUGS.

80
Making Decisions

Perhaps the greatest advantage of Bayesian methods is the diversity of options available to
assist in the decision making process. Communicating results in a way that is easily
interpretable and accurate enables informed decisions to be made. Here we outline some of
the ways modelled estimates can be used and visualized.

The SIR and RER estimates produced using the methods described in the previous sections
are two commonly seen measures of disease risk. The estimates produced by Bayesian
models give great flexibility in reporting results, including comparison of the risk estimates
against the average, ranking estimates, and/or examining the uncertainty around the
estimates.

Ranking of disease estimates ensures that public health investigations or interventions are
prioritized correctly (4). In the Bayesian context, the posterior distributions of health outcome
measures (such as SIR and RER) allow for the calculation of rank estimates of each area (47,
256). For instance, Athens et al. (295) use five health outcome measures to obtain county
rank estimates for a composite health outcome measure. The five health outcome measures
are converted to a score, and then ranked by weighted means. The ranking of health outcomes
is useful for representing health performance of each area which can then be used to inform
health decision making.

Moreover, comparison between two areas can be made easily in the Bayesian framework.
Outside of Bayesian methods, it may be difficult and problematic to conduct a large number
of pairwise comparisons for all areas using post-hoc tests (296). The problem is that by
conducting so many comparisons, the probability of finding some of the differences
statistically significant by chance alone increases. The Bayesian context eliminates this issue
with pairwise comparisons of the posterior distributions.

Bayesian methods produce measures of uncertainty for each modelled estimate. The
uncertainty attached to the spatial distribution of risk values across the study region can be
known as spatial uncertainty (246). It is valuable to visualize spatial uncertainty as it provides
local details of the spatial variation of the risk, as well as an input to resource allocation,
management and policy strategies. Several methods have been proposed to describe the
uncertainty attached to the smoothed rates, including mapping the 95% credible interval of
the posterior distribution of smoothed rates (260) and the probability that the risk in each
small area exceeds a certain threshold (102).

Under the Bayesian paradigm, there is great flexibility in communicating and visualising
results. Options include maps or graphs of the smoothed estimates, their associated
uncertainty, or the probabilities of being above/below certain values. Mapping of disease
rates or outcomes facilitates comparison of spatial patterns in disease rates between males
and females, between age groups, between races, over time, and motivates comparison with
patterns of potential causes (297). By comparing disease rates of different areas, clues to

81
possible causation may be found and this serves as a starting point for further investigation.

The purpose of this Section is to showcase various visualisations that can be produced using
the outputs obtained from Bayesian modelling techniques and the associated interpretation.
This is demonstrated on a common cancer with poor survival: male lung cancer in
Queensland. Figures B.2 to B.7 present an array of maps or plots based on the results from
modelled survival (RER of death within 5 years of diagnosis) for each SLA that are useful for
communicating the results of statistical analysis via the Bayesian paradigm. The RER
expresses the risk of cancer patients dying from their cancer within five years of diagnosis in
an SLA compared to the Queensland average (RER = 1), and therefore should not be directly
compared between two SLAs. The figures were produced using R software, package
maptools.

Figure B.2 maps the posterior distribution of SLA-level RER and provides a picture of the
spatial pattern of the underlying risk. Figure B.3 depicts the uncertainty associated with the
Bayesian estimates of RER by mapping the 95th percentile range of the 10,000 values
sampled from the posterior distribution of RER for each SLA. A graph showing the ranked
RER with the associated 95% credible interval for each SLA is provided in Figure B.4.
Horizontal box plots of the RER estimates by socioeconomic status and rurality are provided
in Figure B.5 to provide additional information about where the extent of variability across
the Queensland state. Figure B.6 maps the SLAs having a 90% probability of RER being
higher than the Queensland average (RER = 1) (highlighted in red) and the SLAs having at
least a 90% probability of RER being lower than the Queensland average (RER = 1)
(highlighted in blue). Figure B.7(a) depicts the probability of the SLAs having RER
exceeding 1 and Figure B.7(b) depicts the probability of the SLAs having RER exceeding
1.2.

82
Figure B.2 Bayesian smoothed estimate of RER

Notes: To show the spatial pattern of the underlying risk, the median of the posterior distribution of SLA-level
RER is mapped. An inset of South-East Queensland is provided for greater detail as this region has a large number
of SLAs. Thematic categories are based on fixed breaks method.

83
Figure B.3 Uncertainty of Bayesian smoothed estimate of RER

th
Notes: This map depicts the uncertainty associated with the estimates of relative risk. The 95 percentile range
(97.5 minus the 2.5 percentile) of the 10,000 values sampled from the posterior distribution of RER for each SLA
is mapped here. An inset of South-East Queensland is provided for greater detail as this region has a large number
of SLAs. Thematic categories are based on quintiles.

Figure B.4 Uncertainty of Bayesian smoothed estimate of RER

Notes: The 95% credible interval (97.5 − 2.5 percentile) of the 10,000 values sampled from the posterior
distribution of RER for each SLA is plotted here. This plot shows how much reliance can be placed on the
estimates. The black line is the median RER for each SLA. The blue vertical lines are the 95% credible intervals,
and indicate the amount of uncertainty associated with each estimate. The red line shows the Queensland average
(set to 1).

84
Figure B.5 Distribution of smoothed RER estimates according to (a) Socioeconomic
status (b) Rurality
(a)

(b)

Notes: The distributional plots reflect the general patterns in the smoothed RER estimates across the area-based
categories of socioeconomic status and rurality. These plots show the proportion of RER estimates that are above
or below the Queensland average (vertical red line) within each of the area-based categories. The plots only
present the range of point estimates, and so do not take the amount of uncertainty associated with each SLA-
specific estimate into account.

85
Figure B.6 Using posterior probabilities to classify risk

Notes: In the Bayesian paradigm, the SLAs highlighted in red have a 90% probability of RER being higher than
the Queensland average (RER = 1). This means that the lower 10th percentile of the posterior distribution of RER
exceeds 1. The SLAs highlighted in blue express at least a 90% probability of RER being lower than the
Queensland average (RER = 1). This means that the upper 90 th percentile of the posterior distribution of RER is
less than 1. The density plots show the posterior distribution of RER for four randomly chosen SLAs where the
x-axis is the RER values. The two density plots on the left show that there is more than 90% chance for the RER
to be higher than 1. The two density plots on the right show that there is more than 90% chance for the RER to be
lower than 1. The percentage of low risk or high risk for each SLA is also given in each density plot. An inset of
South-East Queensland is provided for greater detail as this region has a large number of SLAs.

Discussion

In this article we have outlined the benefits of Bayesian models for both analysis and
visualization. The public health arena regularly makes practical decisions affecting people’s
health. To facilitate decisions, it is vital that the analysis is conducted appropriately, and
results are communicated effectively.

Bayesian methods are increasingly being used to analyse routinely collected data. The
Bayesian framework is now the tool of choice in many applied statistical areas, including
disease mapping (298). In small area studies, Bayesian methods often have better model fit
than non-Bayesian smoothing methods (47). Greater flexibility in distributional assumptions
is possible under Bayesian methods than in traditional regression models (14).

86
Figure B.7 (a) Thematic map depicting the probability of RER exceeding 1, (b)
Thematic map depicting the probability of RER exceeding 1.2

(a)

(b)
Notes: The threshold 1.2 was chosen to reflect high risk as it lies in the fifth quintile. Four SLAs are chosen to
demonstrate how the probabilities change when the thresholds change. An inset of South-East Queensland is
provided for greater detail as this region has a large number of SLAs.

87
Whether to standardise response rates depends on the study objectives. For the cancer atlas, it
was desirable to remove the influence of age, so that differences were not due to different age
structures between areas. For incidence, we used the standardised incidence ratio (SIR),
which adjusts for the area-specific age and sex structure. An alternative method to
standardisation for dealing with confounders is via the use of regression models (299). These
can be particularly useful when multiple confounders need to be controlled for
simultaneously. For relative survival, we included age in the regression equation to remove
its influence on the results. However, if the purpose of a study is to identify where the highest
rates of disease are, such as for service provision, then there is no need to standardise (or
otherwise adjust) the incidence rates. This is because the cause of the variation (whether sex,
age or other factors), is inconsequential.

Visualising disease patterns through maps remains an effective method to convey a large
amount of information in an engaging way. Few modern day visualisations include
uncertainty measures, yet this greatly assists in decision making. Online, interactive
visualisations can dynamically link maps (e.g. Figure B.2 showing the smoothed Bayesian
RER), with plots of the uncertainty (e.g. Figure B.3 showing the 95% credible interval for
each area). Selecting an area would then highlight the corresponding region in both plots,
providing much greater information to the user.

There are limitations associated with using routinely collected data. Determining the direction
of causation may not be possible. Often there is a lag time between exposure and disease
detection, and patients may move during this time. Bayesian methods also have certain
limitations, including greater computational time if using Markov chain Monte Carlo
approaches, and requiring sensitivity analyses to ensure priors are not exerting undue effect.
With regard to computation using R-INLA, models must be expressible in the linear model
format and there are restrictions on the types of prior distributions that can be assumed.

However, we believe the advantages outlined in this article outweigh any limitations.
Routinely collected data exist to enable disease monitoring and control. Appropriate analyses
convert this data into information, which once communicated, enables action. Bayesian
methods not only enable appropriate analyses to be performed, they also provide greater
flexibility in visual communications.

Can descriptive studies really influence government policy? The disparities identified in the
cancer atlas resulted in the Queensland government including a specific objective aimed at
reducing the geographic disparities in cancer outcomes in their Strategic Directions (300).
Results were also used in lobbying to increase the amount of financial assistance the
government provided to remote patients to offset travel and accommodation costs while
obtaining treatment away from home, and the amount provided was subsequently increased.
Our experience is that routinely collected data, when appropriately analysed and
communicated, facilitate appropriate government action.

We hope this article will enable greater understanding, and potentially uptake, of Bayesian

88
methods in disease mapping, along with available options for communicating estimates and
their uncertainty.

Additional Information

B.1 WinBUGS code

WinBUGS code for the incidence model

Model
{
for (i in 1 : N) {
# Likelihood
O[i] ~ dpois(mu[i])
Opred[i] ~ dpois(mu[i])
log(mu[i]) <- log(E[i]) + alpha + u[i] + v[i]
# Area-specific relative risk (for maps)
RR[i] <- exp(alpha + u[i] + v[i])
# Prior distribution for the uncorrelated heterogeneity
v[i] ~ dnorm(0, tauv)
}
# CAR prior distribution for spatial random effects
u[1 : N] ~ car.normal(adj[], weights[], num[], tauu)
for(k in 1:sumNumNeigh) {
weights[k] <- 1
}
# Other priors:
alpha ~ dflat()
# Hyperpriors on precisions
tauu ~ dgamma(0.1, 0.1)
tauv ~ dgamma(0.001, 0.001)
sigmau <- sqrt(1 / tauu)
sigmav <- sqrt(1 / tauv)
#Standard deviations
sdv <- sd(v[]) #marginal SD of heterogeneity
sdu <- sd(u[]) #marginal SD of clustering
}

WinBUGS code for the relative survival model

Model
{
# Likelihood
for (i in 1 : datarows) {
d[i] ~ dpois(mu[i])
mu[i]<-d_star[i] + excessd[i]
log(excessd[i]) <- log(y[i])+ alpha[RiskYear[i]] + beta[1]*agegp2[i]
+ beta[2]*agegp3[i]+ u[slaNo[i]] + v[slaNo[i]]

89
for (j in 1:N_RiskYear){
alpha[j] ~ dnorm (0, 0.001)
}
}
# CAR prior for spatial effects
u[1:Nsla] ~ car.normal(adj[], weights[], num[], tauu)
for (k in 1:sumNumNeigh) {weights[k] <- 1 }
for (i in 1:Nsla) {
# Prior distribution for the uncorrelated heterogeneity
v[i] ~ dnorm(0, tauv)
logRER[i]<-u[i]+v[i]
RER[i]<-exp(logRER[i])
}
# Other priors
tauu ~ dgamma(0.5, 0.001)
tauv ~ dgamma(0.5, 0.001)
varv <- 1/tauv
varu_con <-1/tauu
varu_marg<-sd(u[])*sd(u[])
}

90
B.2 R-INLA code

R-INLA code for the incidence model

Assume that data are available for a set of areas as {yi,ei,x1i,x2i} for i = 1,...,n, where yi is a
count, ei is an expected count, and x1i and x2i are two predictors/covariates. These data should
be read into R as vectors and can be held in a list. In the code below, n represents the number
of areas, obs represents disease count, expe represents expected count, cov1 and cov2
represent the covariates, u represents the spatial random effects, and v represents the
unstructured (non-spatial) random effects.

u=seq(1:n)
v=seq(1:n)
data.incid = list(obs=obs, expe=expe, cov1=cov1, cov2=cov2, u=u, v=v)
formula1 = obs ~ cov1 + cov2
+ f(u, model="besag", graph="queensland.graph", param=c(0.1, 0.1))
+ f(v, model="iid", param=c(0.001, 0.001))
result1 = inla(formula1, family="poisson", data=data.incid,
control.compute=list(dic=TRUE, cpo=TRUE, mlik=TRUE), E=expe)
summary(result1)

R-INLA code for the relative survival model

In the code below, n represents the number of areas, d represents the number of deaths (dijk),
d_star represents the expected number of deaths due to causes other than the disease of
interest (d∗ijk), y represents the person-time at risk (yijk), cov1 and cov2 represent the
covariates, u represents the spatial random effects, and v represents the unstructured (non-
spatial) random effects.
u=seq(1:n)
v=seq(1:n)
data.surv = list(d=d, d_star=d_star, y=y, cov1=cov1, cov2=cov2, u=u, v=v)
formula2 = d ~ offset(d_star) + cov1 + cov2
+ f(u, model="besag", graph="queensland.graph", param=c(0.5, 0.001))
+ f(v, model="iid", param=c(0.5, 0.001))
result2 = inla(formula2, family="poisson", data=data.surv,
control.compute=list(dic=TRUE, cpo=TRUE, mlik=TRUE), E=y)
summary(result2)

91
92
Appendix C Computational software
Free resources

BUGS (includes WinBUGS and OpenBUGS), available from: http://www.mrc-

bsu.cam.ac.uk/software/bugs/ Enables running of Bayesian models using (predominately)
Gibbs sampling. The built-in GeoBUGS can be used to generate neighbourhood matrices.

GeoDa Easy-to-use software featuring various smoothing and regression models, as well as
generation of various types of neighbourhood matrices. Available from:
https://geodacenter.asu.edu/software/downloads.

JAGS Has a cross-platform engine for the BUGS language, but also allows users to write
their own distributions, functions etc. Available from: http://mcmc-jags.sourceforge.net/.

The National Cancer Institute has developed several resources, all of which are freely
available at gis.cancer.gov/tools/nci_tools.html including:

o Plug-ins for using with ESRI ArcGIS ArcMap include, among others:
o ColorTool (Assists in using ColorBrewer colours for chloropleth maps)
o Head-Bang (Smooths data within ArcMap using the Head-Bang
smoothing algorithms. These are semi-related to the locally-weighted
median discussed in Section 2.4.1.)

o Linked MicroMaps (a graphing program written in Java, allowing easy

comparison of statistics across regions and time. Multiple variables can be
examined interactively)

o HD*Calc (statistical software for evaluating health disparities. Originally

developed for cancer data, so can be used as an extension of SEER*Stat
software, but also with any dataset. Generates tables and/or graphs containing
calculated summary measures of disparities.)

o SaTScan (aims to detect clusters in spatial, temporal., or spatio-temporal data

using scan statistics and evaluate their significance.) www.satscan.org/

NIMBLE http://r-nimble.org/ Can be used as an extension of the BUGS language to write

flexible statistical models, or can also be used without BUGS models as a way to compile
simple code similar in form to R into C++, which is then compiled and loaded into R.

PySAL www.pysal.org An open source library of spatial analysis functions written in

Python.

93
R R is statistical software, available from: https://www.r-project.org/ A myriad of
packages enable spatial analysis within R, including methods appropriate for point- and area-
level data. Useful packages for areal data could include:
o bdsmatrix: routines for block diagonal symmetric matrices
o CARBayes: spatial GLMMs for areal data
o CARBayesST: spatio-temporal GLMMs for areal data
o coda: output analysis and diagnostics for MCMC
o colorspace: maps between a variety of colour spaces (e.g. RGB, HSV,
CIELAB)
o DCluster: detection of spatial clusters of diseases
o epitools: for epidemiology data and graphics
o fields: curve, surface and function fitting with an emphasis on splines, spatial
data and spatial statistics
o gdistance: calculates distances and routes on geographic grids
o glmmBUGS: pass spatial models to WinBUGS
o geoR: geostatistical analysis
o geosphere: computes distances and related measures for geocoordinates
o geospacom: generates distance matrices from shape files and plots data on
maps
o gwrr: fits geographically weighted regression models with diagnostic tools
o INLA (available from www.r-inla.org/, not CRAN): Integrated Nested
Laplace Approximation
o INLABMA: Bayesian model averaging with INLA
o lmtest: testing linear regression models
o locfit: local regression, likelihood and density estimation
o maps: draw geographical maps
o maptools: tools for reading and handling spatial objects
o Matrix: sparse and dense matrix classes and methods
o MCMCpack: functions to perform Bayesian inference using posterior
simulation for a number of statistical models
o McSpatial: nonparametric spatial data analysis
o mgcv: mixed generalised additive model with multiple smoothing parameter
estimation
o nlme: linear and nonlinear mixed effects models
o pixmap: import, export and other functions of bitmapped images
o plotGoogleMaps: plot spatial or spatio-temporal data over Google maps
o PReMiuM: for profile regression (a Dirichlet process Bayesian clustering
model)
o raster: enables many GIS methods
o R2BayesX: interfaces R with BayesX (performs Bayesian inference in
structured additive regression models
o R2WinBUGS: interfaces R with WinBUGS
o RandomFields: simulation and analysis of Gaussian fields, as well as extreme
value random fields
o RColorBrewer: provides colour schemes for maps as described at
colorbrewer2.org

94
 o RPyGeo: ArcGIS processing in R via Python
o sandwich: robust covariance matrix estimators
o shapefiles: read and write ESRI shapefiles
o sp: classes and methods for spatial data
o spacetime: classes and methods for spatio-temporal data
o spaMM: spatial GLMMs
o sparr: estimates kernel-smoothed relative risk and subsequent inference
o SparseM: Basic linear algebra for sparse matrices
o spatcounts: Spatial count regression via customised MCMC
o SpatialEpi: cluster detection and disease mapping functions, including
Bayesian cluster detection
o spatsurv: Bayesian inference for parametric proportional hazards spatial
survival models
o spBayes: Univariate and multivariate spatio-temporal models with MCMC
o spBayesSurv: Bayesian modelling and analysis of spatially correlated survival
data
o spdep: useful functions to create spatial weights matrix objects from polygon
contiguities, and various tests for global and spatial correlation
o spgrass6: interfaces R with GRASS 6+ GIS
o sphet: Estimation of spatial autoregressive models with and without
heteroskedastic innovations
o tmap: thematic maps

Stan Can be used for Bayesian modelling with either MCMC or approximate Bayesian
inference, or penalised MLE. Available from: http://mc-stan.org/.

Commercial software

ArcGIS Comprehensive GIS software from ESRI. Further details at: www.arcgis.com/.

BoundarySeer Statistical analysis software from BioMedware that enables detection and
analysis of geographic boundaries. Further details at:
www.biomedware.com/?module=Page&sID=boundaryseer-overview

ClusterSeer Statistical analysis software from BioMedware that enables detection and
analysis of event clusters. Further details at:
www.biomedware.com/?module=Page&sID=clusterseer-overview

MapInfo Comprehensive GIS software from Pitney Bowes. Further details at:
www.mapinfo.com.

MLwiN Statistical software for fitting multilevel (hierarchical) models via either
maximum likelihood estimation or MCMC methods. Further details at:
www.bristol.ac.uk/cmm/software/mlwin/

95
SAS (Statistical Analysis System). This is a software suite developed by SAS Institute for
advanced data analysis and management. Further details at: www.sas.com. The SAS-ESRI
bridge enables ArcGIS functionality. Other useful commands include:
o Proc mapimport – converts a shapefile to a dataset
o Proc gmap – for creating maps (includes choropleth maps)
o WinBUGSio – A user-written macro for interfacing SAS with WinBUGS

Stata Comprehensive software for data analysis and statistical analyses developed by
StataCorp. Further details at: www.stata.com/. User-written programs for spatial analyses
include:
o geocode3 – Using Google geocoding can either geocode addresses into
coordinates or reverse geocode coordinates into addresses to examine the
quality of geocoding
o shp2dta – imports .shp data to stata formats
o spatgsa – calculates global spatial autocorrelation measures
o spatlsa – calculates local spatial autocorrelation measures
o spatwmat – generates a matrix of weights
o spmap – generates a large variety of thematic maps
o spgrid – generates two-dimensional grids
o spkde – uses datasets generated by spgrid to perform a variety of kernel
estimators
o traveltime3 – uses Google Distancematrix to retrieve distance and travel time
between two locations (either geocoded coordinates or addresses)
o winbugs – A suite of commands starting with “wb” that allow Stata to
interface with WinBUGS.

SpaceStat Statistical analysis software from BioMedware that enables visualisation,

analysis, modelling and exploration of spatiotemporal data. Further details at:
www.biomedware.com/?module=Page&sID=spacestat-features

96
Appendix D Recommended further reading
Lai P-C, So F-M, Chan K-W. Spatial Epidemiological Approaches in Disease Mapping and
Analysis. Baton Rouge: CRC Press; 2008.

Lawson AB. Bayesian Disease Mapping: Hierarchical Modeling in Spatial Epidemiology.

Second edition. Boca Raton: CRC Press; 2013.

Lawson AB. Statistical Methods in Spatial Epidemiology, Second Edition. West Sussex:
Wiley; 2006.

Lawson AB, Browne WJ, Vidal Rodeiro CL. Disease Mapping with WinBUGS and MLwiN.
Chichester: John Wiley & Sons, Ltd; 2004.

Lawson AB, Williams FLR. An Introductory Guide to Disease Mapping. New York: John
Wiley & Sons, Ltd; 2002.

Lunn D, Jackson C, Best N, Thomas A, Spiegelhalter D. The BUGS book: A Practical

Introduction to Bayesian Analysis. Boca Raton, FL: Chapman and Hall/CRC Press; 2012.

Mengersen KL. Bayes for Beginners. Brisbane: Queensland University of Technology; 2011.

Ntzoufras I. Bayesian Modeling Using WinBUGS. Hoboken, NJ: John Wiley & Sons, Inc.;
2008.

Waller LA, Gotway CA. Applied Spatial Statistics for Public Health Data. Chichester: John
Wiley & Sons, Inc; 2004.

97