Spatial and Spatio-temporal Epidemiology 39 (2021) 100441

Contents lists available at ScienceDirect

Spatial and Spatio-temporal Epidemiology

journal homepage: www.elsevier.com/locate/sste

Spatio-temporal models of bovine tuberculosis in the Irish cattle

population, 2012-2019
Jamie M. Madden 1, *, Guy McGrath 1, James Sweeney 2, Gerard Murray 3, Jamie A. Tratalos 1,
Simon J. More 1
1
Centre for Veterinary Epidemiology and Risk Analysis (CVERA), School of Veterinary Medicine, University College Dublin, Dublin, Ireland
2
Department of Mathematics and Statistics, University of Limerick, Limerick, Ireland
3
Department of Agriculture, Food and Marine, Drumshanbo Regional Veterinary Office, Derryhallagh, Drumshanbo, Co. Leitirm, Ireland

A R T I C L E I N F O A B S T R A C T

Keywords: Introduction: Bovine tuberculosis (bTB) is an important zoonotic disease which has serious and sometimes fatal
spatio-temporal effects on both human and non-human animals. In many countries it is endemic in the cattle population and has a
spatial statistics considerable economic impact through losses in productivity and impacts on trade. The incidence rate in Ireland
bovine tuberculosis
varies by herd and location and it is hoped that statistical disease-mapping models accounting for both spatio-
standardised incidence ratio
temporal correlation and covariates might contribute towards explaining this variation.
disease mapping
Methods: Ireland was divided into equally sized hexagons for computational efficiency (n = 997). Different
spatio-temporal random-effects models (e.g. negative binomial Besag-York-Mollié) were explored, using
comprehensive data from the national bTB eradication programme to examine the association between cova­
riates and the number of bTB cattle. Leveraging a Bayesian framework, model parameter estimates were obtained
using the integrated nested Laplace approximation (INLA) approach. Exceedance probabilities were calculated to
identify spatial clusters of cases.
Results: Models accounting for spatial correlation significantly improved model fit in comparison to non-spatial
versions where independence between regions was assumed. In our final model at hexagon level, the number of
cattle (IR = 1.142, CrI: 1.108 – 1.177 per 1000), the capture of badgers (IR = 5.951, CrI: 4.482 – 7.912), per­
centage of forest cover (IR = 1.031, CrI: 1.020 – 1.042) and number of farm fragments (IR = 1.012, CrI: 1.009 –
1.015 per 10 fragments) were all associated with an increased incidence of bTB. Habitat suitability for badgers,
percentage of dairy herds and the number of cattle movements into the herd were not. As an epidemiological tool
and to suggest future work, an interactive online dashboard was developed to monitor disease progression and
disseminate results to the general public.
Conclusion: Accounting for spatial correlation is an important consideration in disease mapping applications and
is often ignored in statistical models examining bTB risk factors. Over time, the same regions in Ireland generally
show highest incidences of bTB and allocation of more resources to these areas may be needed to combat the
disease. This study highlights national bTB incidence rates. Shifting from national level analysis to smaller
geographical regions may help identify localised high-risk areas.

Introduction
Bovine tuberculosis (bTB) is an infectious disease caused by infection
with Mycobacterium bovis (M. bovis) [1]. bTB is prevalent in the cattle
populations of many countries and has economic impacts through loss of
productivity and international trade restrictions, as well as posing a
health risk for humans. Ireland has had a national bTB eradication
programme in place since the late 1950s, when approximately 80 % of
all herds were estimated to be infected [2,3]. Since then, significant
advances have been made and herd-level incidence was approximately
3.2 %, during 2019 [4]. However, further progress towards bTB eradi­
cation, and an officially bTB (OTF) free status for Ireland, is proving to
be particularly challenging. bTB in Ireland is a complex disease of cattle
and can infect many species, including wildlife. In particular, badgers
(Meles meles) have been shown to facilitate bTB spread between cattle in
both the UK and Ireland [2]. Similarly, deer are suspected of playing a

* Corresponding author

https://doi.org/10.1016/j.sste.2021.100441
Received 15 January 2021; Received in revised form 8 July 2021; Accepted 19 July 2021
Available online 22 July 2021
1877-5845/© 2021 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
J.M. Madden et al.
role in transmission of the infection. The significance of this role is
currently the centre of much debate [5]. The bTB eradication pro­
gramme in Ireland is a statutory programme which employs the Single
Intradermal Comparative Tuberculin Test (SICTT) which is applied at
least annually to all cattle in Irish herds to identify infected cattle. An­
imals which are determined to have developed a positive reaction to the
SICTT are termed “reactors” are removed for slaughter, and the herd is
subject to trade restrictions, which remain in place, until the herd has
recorded two negative consecutive SICTT tests at a minimum of 60 days
apart. Routine slaughterhouse surveillance of bovine carcases by vet­
erinary inspectors is also an important element of the bTB eradication
programme and suspect lesions identified on post-mortem examination,
are routinely sent to the laboratory for bTB confirmation. In recent
years, badger vaccination has been introduced into the programme
seeking to limit the transmission of infection within badger populations
thus reducing the potential onward spread from badgers to cattle [2,6,
7].
Over many decades, a substantial body of research has been under­
taken examining the pathology and epidemiology of bTB and this has
contributed substantially to our understanding of the disease. The
pathways for bTB introduction and persistence within cattle populations
are well understood, and include animal movement, residual infection,
and the environment (including wildlife and neighbouring farms); their
relative importance is less clear and likely varies in time and space [3].
Furthermore, a number of risk factors for increased bTB herd-level risk
in Ireland are recognised, including herd size and herd bTB history [1,3,
8]. However, translating this information into operational policy can
prove challenging.
Proximity between infected and susceptible hosts is an important
consideration in infectious disease control, yet relatively few bTB
research studies have used this methodology to explicitly consider
spatial, or indeed temporal, correlation. Previous research in this area in
an Irish context has used experimental design to aggregate out any
spatial association, or used spatial proxies, as opposed to addressing
spatial correlation explicitly. Due to the nature of the spread of the in­
fectious diseases, it is plausible to hypothesize that regions located near
one another should be more correlated to each other when compared to
areas which are dispersed. Failure to account for spatial correlation (also
known as spatial autocorrelation) may lead to poor model fit and un­
reliable parameter estimates, particularly if there is strong correlation
present [9,10]. Hierarchical spatial models are an appropriate approach
to account for such correlation where local infection risk is modelled
using a set of spatial random-effects, which borrow information from
neighbouring areas to get more reliable region-specific estimates [9].
These models, known as disease-mapping models, can adjust for con­
founders while also accounting for spatial and temporal correlation, and
can provide us with a critically important epidemiological tool to
monitor bTB progression and visualise areas or clusters at increased
high-risk of disease across a country. Importantly, these models (by
accounting for spatial correlation), can help identify regions where
disease is occurring, but where we are unsure of the mechanism driving
this.
Thus, the aims of this study were (i) to implement a spatio-temporal
random-effects analysis in a Bayesian framework, using comprehensive
data from the national bTB eradication programme to examine the as­
sociation between various covariates and the number of bTB positive
cattle in Ireland, (ii) to use the model to identify geographical clusters of
cases and, (iii) to implement the model as an epidemiological tool to
monitor bTB outbreaks at a local level. This approach has the potential
to be used by both policy-makers and field workers for local planning
and disease surveillance purposes.
2
Spatial and Spatio-temporal Epidemiology 39 (2021) 100441
Methods
Study design and outcome
The study period was from 1st January 2012 until 31st December
2019 and we utilised data from the national bTB eradication programme
(Animal Health Computer System (AHCS) [11], Animal Identification
and Movement System (AIM) [12], Laboratory Information Manage­
ment System (LIMS)), which included results from herd-level skin tests,
animal-level laboratory tests (from animals with suspect bTB lesions)
and animal-level gamma interferon (IFN-γ) blood testing. An animal is
deemed a reactor using the SICTT in accordance with the Annex B of
Directive 64/432/EEC as amended section 2.2 [13] if the bovine skin
reaction is 4 mm or greater than the avian skin reaction or if there is
oedema, exudative necrosis, heat and/or pain at the bovine tuberculin
injection site and/or swelling of the related prescapular lymph node
[14]. Separately, we had access to the Land Parcel Identification System
(LPIS) which is a spatial database which identifies the boundaries of
farms [15]. Ideally, we would like to model each farm fragment as a
separate spatial unit but because we are dealing with the entire country
of Ireland, this approach would not be feasible as the computational
complexity would be too large (approximately 570, 000 individual
fragments). To overcome this issue, we divided Ireland into separate
geographical regions (areal units) in the shape of hexagons (n = 997,
area = c. 84 km2). This size was chosen to capture localised disease risk
which may be useful for local veterinary inspectors to monitor. In
contrast to point locations, aggregation of data into these hexagons
simplifies the spatial correlation structure with adjacent cells either
sharing a border or not. From the LPIS data, we assigned each herd an
Irish National Grid (TM65) x and y coordinates calculated at the
centroid of their largest fragment of land, with the exception of
approximately 5% of herds which are not in this database [12]. LPIS was
originally developed to assist farmers with agricultural subsidies and so,
to receive payments from Department of Agriculture, Food and the
Marine (DAFM), farmers need to register their land and it is for this
reason that there is a low proportion of herds which are not included in
the database [16].
For each hexagon, we calculated the number of bTB positive cattle
(also termed “bTB cases”; including all SICTT reactors that were deemed
positive, all slaughtered cattle with laboratory confirmed M. bovis
infection, and cattle positive to the IFN-γ tests) amongst herds whose x, y
coordinates were located within it, to represent our outcome “count”
variable per year. Similarly, where appropriate, we aggregated herd-
level variables to give a hexagon-level covariate value. So, for
example, if there were two herds within a hexagon with a herd size of 50
and 20 respectively, the overall number of cattle in that hexagon was 70
(50 + 20). Using this approach, we calculated a number of hexagon-level
variables per year that are known bTB risk factors at herd-level
including total cattle (based on the average number of cattle at herd
SICTT) [1,17] and total number of bovine movements (counts) into the
herd [2,12]. To incorporate information related to badger abundance [1,
6,18], two variables were created, the first was a time-varying dichot­
omous variable indicating if any badgers were caught or not within each
hexagon based on data from the Wildlife Unit of DAFM. An alternative,
the number of badgers caught, was not considered appropriate as tar­
geted badger capture is often conducted following large bTB outbreaks,
in cases where an epidemiological investigation has suggested that
badger presence may be a contributing factor [19]. As a means of
adjusting for variation in badger population density, the second badger
variable created was predicted badger social group density, based on
landscapes metrics (average of values in a hexagon) [20]. It ranges be­
tween zero and one, where values closer to one represent high proba­
bility of main sett occurrence within the hexagon. In some studies, dairy
herds have been shown to be at greater risk of bTB outbreaks than other
herds [8], and therefore the percentage of dairy herds per hexagon was
calculated. As it has been argued that deer may play a role in bTB
J.M. Madden et al.
transmission [5,21], the percentage of forestry cover was calculated
using the Co-ORdinated INformation on the Environment (CORINE)
land cover classification system [22], based on an assumption that the
abundance of deer would be positively correlated with forest cover. The
number of herds and fragments per hexagon were also calculated [23]. It
has been hypothesized that highly fragmented regions may facilitate the
spread of infection, as cattle from different herds may be more likely to
come in contact with one another. As our unit of interest is geographical
location, we believe examining fragments in this way is justified.
Statistical analysis
Through the Irish bTB eradication programme, we have almost full
coverage of the bTB infection status of the entire cattle population on an
annual basis, which allows us to examine population level metrics. At an
areal level, standardized incidence ratios (SIR) are a crude, but very
useful, measure of bTB risk, defined as the ratio of observed to expected
number of cases for a given time frame. The number of expected cases of
bTB, Ei for hexagon i, represents the total number of cases that one
would expect to find in that area, if it behaved like the overall standard
population, which can be estimated using indirect standardisation:
∑ this BYM model, smoothed SIRs can be considered the posterior mean of
Eit = rt pit
t=2012
where rt is the rate (bTB cases/population) at timepoint t in the standard
population, and pit is the population at timepoint t for area i. SIRs are a
good starting point to provide a quick snapshot of how bTB has spread
across regions but can give misleading results in regions with small
populations or where the expected counts are very low [24]. Addition­
ally, this approach does not account for spatial correlation [9]. It is
worth noting that there are other methods to return robust SIR estimates
when population sizes are small using empirical Bayes techniques
(where priors are estimated from the data itself) which are an approxi­
mation to more exact Bayesian methods [25].
Observed counts of cases, Yit in hexagon i, for year t, were modelled
using a negative binomial distribution:
Yit ∼ NegBin(Eit θit )
where Eit , as before, is the expected count and is included as an offset
term in a model for standardisation purposes, and θit , which is the
relative risk in hexagon i for year t.A common way to parameterize a
negative binomial distribution is to see it arising as a distribution of the
number of failures k, before the rth success with its probability mass
function given as:
( )
r+k− 1 r
f (kr, p) = P(y = kr, p) =
k
p (1 − p)k
where r is the number of successes, k is the number of failures, and p is
the probability of success, see Chambers et al. for further discussion
[26]. There are various spatial structures used in disease mapping but
one of the most widely used for areal data is the Besag-York-Mollié
(BYM) model [27], which can be incorporated with or without a tem­
poral component, to produce different models:
Spatial model : g(θit ) = log(Eit ) + α + Xit β + f (year) + (ui + υi )
Temporal model : g(θit ) = log(Eit ) + α + Xit β + f (year) + (γ t + ϕt )
Spatio − temporal model : g(θit )
= log(Eit ) + α + Xit β + (ui + υi ) + (γt + ϕt )
where g(.) is a link function (log function) defined according to the
conditions on θit , α denotes the intercept or overall risk in the country,
Xit is the design matrix associated with the covariates with β coefficients,
f(year) is a smooth function of time (cubic regression spline), ui a spatial
3
Spatial and Spatio-temporal Epidemiology 39 (2021) 100441
correlated random-effect specific to region i to model spatial depen­
dence between the relative risks (RRs), and υi an unstructured
exchangeable component that models uncorrelated noise which follows
a normal distribution, defined as N(0, σ2v ). γt is a smoother (rw2, random
walk 2nd order which models the difference between the two consecu­
tive differences) for year and ϕt is an extra independent and identically
distributed noise term. Together, these two random-effect terms, γt and
ϕt define the temporal term [24,28-30]. This will result in 997 values of
ui s and υi s and 8 values of γt s and ϕt s. The BYM model assumes the
spatial random-effect, ui , follows a Conditional Autoregressive (CAR)
distribution:
( )
∑N
ui |u− i ∼ N cij uj , σ 2
j=1
where cij is an adjacency matrix, with cij = 1 if area i and j are neighbours
and 0 otherwise. This is a conditional distribution where the ui |u− i no­
tation indicates that we are modelling the distribution of the random-
effect ui given all the other ui s, except ui itself. Hence, the conditional
mean of the random-effect for hexagon i, ui , given all the other random-
effects, is the average of the neighbouring hexagon random-effects. In
the RR [31].
Convergence issues can often be a problem when attempting to solve
these models in a traditional maximum likelihood estimation (MLE)
framework, owing to their high dimensionality in terms of parameters
requiring estimation. Bayesian approaches, however, are often able to
overcome these limitations via prior specification, although this comes
at the computational cost of using simulation based methods for
parameter estimation [9]. The Integrated nested Laplace approximation
(INLA) approach of Rue et al. [32] is a relatively new Bayesian estima­
tion procedure used as an alternative to Markov chain Monte Carlo
(MCMC) methods. They have the benefit of significantly reducing
computation burden and hence computation time by using very sparse
matrices [32,33]. For these reasons, a Bayesian approach using INLA
was considered for parameter estimation.
Vague prior densities are often assigned to the fixed parameters and
here we give them a normal distribution with mean 0 and precision τ =
0.001. As τ = 1/σ2 , this translates to σ = 31.6, so the fixed parameter
priors are of the form β ∼ N(0, 31.62 ). Similarly, for the variance pa­
rameters or hyperparameters; τu , τv , τγ and τϕ we chose vague priors
which all followed: ∼ LogGamma(1, 0.00005).
From a policy perspective, it would be useful to calculate the prob­
abilities of RR estimates in a region being greater than a given threshold
value. These probabilities are called exceedance probabilities and are
used to identify unusually elevated bTB risk. The probability that the RR
for hexagon i is higher than a threshold c can be calculated as:
P(θit > c) = 1 − P(θit ≤ c)
where P(θit ≤ c) is calculated from the marginal distribution of θit . Here
we use the final set of RR values (year 2019) to estimate the probability
of RR ≥ 1.5 for hexagon i, to identify high risk areas for 2020.
Model selection
To detect collinearity, correlation coefficients between covariates
were calculated and, if required, variables were removed. Alternative
models were compared to choose one best describing the data. This
process involved fitting different distributions to our count outcome
(Poisson, negative binomial and zero-inflated versions of these) and
formally comparing models using the deviance information criterion
(DIC) and widely applicable information criterion (WAIC), where lower
values indicate better fit [9,29]. Additionally, leave-one out cross vali­
dation predictive checks were also implemented where two quantities,
conditional predictive ordinate (CPO) and probability integral transform
J.M. Madden et al. Spatial and Spatio-temporal Epidemiology 39 (2021) 100441

Fig. 1. The number of bTB positive cattle and the corresponding percentage of herds per year in a breakdown in Ireland over time (2012 – 2019). A herd enters a bTB
breakdown and their trading status suspended if they have at least one bTB cow and normal trading cannot be returned until two clear tests are achieved.

Table 1
Progression of bTB programme with herds included in the study for years 2012-2019
Variable 2012 2013 2014 2015 2016 2017 2018 2019

No. of herds 110652 110373 109317 109614 108868 108154 106577 105195
No. of cattle in hexagon 6406 (1847- 6372 (1825- 6229 (1810- 6472 (1786- 6570 (1850- 6597 (1792- 6586 (1756- 6460 (1756-
(median, IQR) 10143) 10462) 10308) 10546) 10853) 10994) 11086) 10829)
No. of movements in hexagon 1833 (477- 1912 (511- 1972 (497- 2116 (484- 2133 (502- 2143 (512- 2069 (484- 2067 (459-
(median, IQR) 3189) 3381) 3488) 3721) 3872) 3905) 3642) 3632)
Dairy herds (%) 14.4 14.4 14.5 14.5 14.6 14.6 14.7 14.9
bTB incidence (animal level) 0.00317 0.00268 0.00274 0.00254 0.00274 0.00273 0.00279 0.00282

Fig. 2. Histogram (binwidth = 5) of the number of bTB positive cattle at a hexagon level for years 2012-2019.

4
J.M. Madden et al.
Fig. 3. Standardised incidence ratios (SIR) by hexagon in Ireland, at two-yearly intervals between 2013 and 2019. These are standardised using data for the relevant
year only, as opposed to standardisation over the entire time period, illustrating how individual hexagons performed for that year. SIR maps for all years, 2012-2019,
can be found in the Supplementary Material (Fig. 2).
(PIT), were used for evaluating the goodness of fit for these Bayesian
models [34]. Uniformity of PIT values indicate that the predictive dis­
tributions match the observations, suggesting a well-fitted model. The
product of all the CPO values can be considered a pseudo marginal
likelihood which gives a cross-validatory summary measure of fit. The
log pseudo marginal likelihood (LPML) is the log of this measure and is
often used as an alternative measure for DIC. Unlike DIC, however, high
LPML values suggest the model is better supported by the data [35]. We
also explored model fit by scaling CPOs (dividing by their maximum)
where low values for particular observations (< 0.01) will show outlying
observations where the model is not fitting well [36]. As a final vali­
dation step and to evaluate the predictive accuracy, we compared the
empirical distribution of our outcome to the distributions of simulated
draws from the posterior predictive distribution of our models [37,38].
All analysis and mapping were conducted using R (R Core Team, 2020).
INLA models were fitted using the R-INLA package [39]. inla.pmarginal
function within R-INLA is used to compute the cumulative probability
function for exceedance probabilities. To help visualise and disseminate
our results an accompanying, online, an interactive dashboard was
developed using R Shiny [40].
5
Spatial and Spatio-temporal Epidemiology 39 (2021) 100441
Results
For the study period, 2012-2019, the overall number of bTB positive
cattle and the percentage of herds per year in a bTB breakdown in
Ireland is presented in Fig. 1. This shows that bTB cases have generally
plateaued in the last eight years (a similar plot for 2008-2019 is pre­
sented as Supplementary Material Figure 1 to show the longer-term
trend and context of progression of the bTB programme). Table 1 pro­
vides an overview of the bTB programme for herds included in the study
for all years. The number of herds have declined over this period while
the percentage of dairy herds is slowing increasing. The distribution of
the number of bTB cases at a hexagon level, which represents the
outcome variable in the study, is given in Fig. 2. Note that it is highly
skewed with a large proportion of zeros (n = 1738, 22%) motivating the
exploration of zero-inflated models.
The spatial distribution of crude SIR values for selected years are
mapped in Fig. 3. These are standardised per year by the national
animal-level incidence rate and, allowing for variation between years, it
is notable that some regions are consistently above the annual average
rate e.g. County Dublin and County Wicklow (see Thomas et al. [41],
J.M. Madden et al. Spatial and Spatio-temporal Epidemiology 39 (2021) 100441

Table 2
Comparison of models with and without accounting for spatial and temporal correlation while applying different distributions
Model Spatial Correlation Temporal Correlation DIC WAIC LPML Dispersion statistic

Crude Poisson - - 217571 217978 -108946 37.35

Poisson (iCAR correlation) ✓ - 137363 154784 -71649 15.72
Poisson (CAR correlation) ✓ - 137362 154787 -71654 15.72
Poisson (BYM correlation) ✓ - 137362 154792 -71656 15.72
Poisson (BYM2 correlation) ✓ - 193747 198905 -95544 28.98
ZIP (BYM2 correlation) ✓ - 132686 149620 -69484 5.55
Crude NB - - 55680 55686 -27843 2.00
ZINB (BYM2 correlation) ✓ - 51787 51986 -26924 1.02
ZINB (BYM correlation) ✓ - 51787 51985 -26907 1.02
NB (BYM correlation)* ✓ - 51784 51983 -26894 0.92
ZINB (BYM & temporal correlation) ✓ ✓ 51783 51981 -26893 1.03
NB (BYM & temporal correlation) ✓ ✓ 51781 51979 -26896 0.92

ZIP: zero-inflated Poisson; NB: negative binomial; ZINB: zero-inflated negative binomial; BYM: Besag,. York, and Mollié model; CAR: conditionally autoregressive; DIC:
deviance information criterion; WAIC: widely available information criterion; LPML: log pseudo marginal likelihood
*
final model chosen

Table 2. It is clear that the Poisson models offer poor fits to the data. In
Table 3
contrast, a crude negative binomial model where spatial correlation is
Posterior estimates from negative binomial BYM model
not accounted for offers a substantial improvement in fit over Poisson
Variable Median Mean(SD) IR(95% Crl) models. Incorporation of spatial correlation structures in a negative
No. of animals (per 1000) 0.13 0.13 1.14(1.11 - binomial model further improves the model fit. Based on the metrics in
(0.02) 1.18) Table 2, the best fit to the data was a model that accounted for both
No. of movements in (per 1000) 0.02 0.02 1.02(0.99 -
spatial (BYM) and temporal correlation (through rw2). However, there
(0.02) 1.06)
Dairy herds (%) 0.01 0.01 1.01(1.00 -
was little difference among the spatial negative binomial models. The
(0.00) 1.02) same two conclusions were drawn when PIT histograms were examined
Badgers caught (no/yes) 1.78 1.78 5.95(4.48 - where negative binomial spatial models were approximately uniform
(0.14) 7.91)a (Supplementary Material Figure 3). Again, similar conclusions were
Badger suitability 0.01 0.01 1.01(1.00 -
drawn from the final model validation step when we examined the
(0.00) 1.02)
Forest cover (%) 0.03 0.03 1.03(1.02 - distributions of our outcome compared to simulated data from the
(0.01) 1.04) models (Supplementary Material Figure 4). These simulated plots sug­
Fragments of farming land (per 0.01 0.01 1.01(1.01 - gest that there is miscalibration in all models and no model fully cap­
10) (0.00) 1.01) tures the zero-inflated effect but indicate that the incorporation of the
Spline term 1 0.00 0.00 1.00(0.93 -
(0.03) 1.07)
spatial random-effects yield better calibrated models. Taking all of this
Spline term 2 -0.14 -0.14 0.87(0.82 - into consideration, and seeking to select the most parsimonious model,
(0.03) 0.93) we chose the negative binomial BYM model with spatial random-effects
Spline term 3 -0.14 -0.14 0.87(0.81 - as our final model. However, for interested readers we have included
(0.03) 0.93)
results from other models in the Supplementary Material (Table 1) along
Spline term 4 -0.15 -0.15 0.86(0.80 -
(0.03) 0.92) with more discussion on model selection. For the final model, the ma­
Spline term 5 -0.02 -0.02 0.98(0.91 - jority of scaled CPO values (81%) were >0.01 indicating a relatively
(0.04) 1.05) good fit.
Spatial hyperparameters Mean(95% Crl)
σui 1.19(1.07 - Hierarchical spatial Bayesian model
1.33)
σvi 0.19(0.07 - The posterior means and 95% credible intervals (CrIs) for the pa­
0.31)
rameters from our final model are provided in Table 3. The number of
IR: incidence ratio (exp(coef)). animals in a hexagon was associated with a significantly higher inci­
a
Interpretation example: After adjustment for all other covariates, the inci­ dence, incidence ratio (IR) = 1.142 (CrI: 1.108 – 1.177) for an increase
dence of bTB where badgers were caught was 5.95(4.48 - 7.91) times that where of 1000 animals. The capture of badgers in a hexagon in a year (yes vs
badgers were not caught.
no) was strongly associated with higher bTB incidence (IR = 5.951, CrI:
4.482 – 7.912) although badger habitat suitability was not. The pro­
Fig. 2, for map of Ireland). Similarly, on the west of the country (County portion of forest cover in a hexagon (IR = 1.031, CrI: 1.020 – 1.042) and
Clare) there has been a consistently higher than average incidence rate the number of farm fragments (IR = 1.012, CrI: 1.009 – 1.015 per 10
although it has not spread to surrounding regions over time to the same fragments) were both risk factors for bTB incidence. Neither the per­
extent as on the east. There are also persistently higher than average centage of dairy herds within a hexagon nor the number of cattle
regions to the north of the country near the Northern Ireland border. movements into the herd were significantly associated with bTB
incidence.
Model structure and selection The posterior mean of σ ui is 1.190, meaning that the majority of the
spatially correlated random-effects ui lie between -3.15 and 3.15 (1.190
The count of the number of herds in a hexagon were highly corre­ ± 1.96) while the additional noise term σvi is only 0.193, suggesting it is
lated with the number of fragments (r = 0.88) and it was decided to keep contributing much less to the model than σui . The combined spatial ef­
the latter in the model along with the other covariates for the multi­ fect of these is presented in Fig. 4, which highlights regions with a higher
variate analysis as there were no other strong correlations. Various incidence of bTB. Higher values indicate problematic areas.
distributions and different spatio-temporal models were fit to the data The resulting RRs from this BYM model for bTB incidence were
and their associated measures of fit (DIC, WAIC, LPML) are compared in mapped for Ireland (see Fig. 5). These are equivalent to the crude SIR

6
J.M. Madden et al.
Fig. 4. Spatially correlated and non-spatially correlated random-effects from negative binomial BYM model. Presented here are the posterior means of ui + vi ’s.
values but have been smoothed to account for spatio-temporal correla­
tions and hexagon-level covariates with the aim of eliminating random
noise. These maps strongly correlate with the crude SIR maps presented
in Fig. 3. The smoothed estimates broadly identify the same regions as
being at increased risk that were highlighted previously. However, using
this plot alone it is difficult to differentiate between years while
providing a suitable scale as estimates have shrunk. As an example, a
hexagon could have an RR changing from 1 to 1.9 between consecutive
years and this may be missed. In reality, however, the absolute differ­
ence of 0.9 or 90% is substantial. To identify such variation between
years, we calculated the percentage change in RR between two
consecutive years, shown in Fig. 6. This plot shows that there was a
decrease in risk from 2013 but that there was then a shift, with an in­
crease in 2019 compared to 2017. In order not to over interpret findings
from one map, we believe both should be viewed in tandem (Fig. 5 and
Fig. 6).
As an aid to bTB risk management planning, we calculated exceed­
ance probabilities based on 2019 RR estimates (for RR ≥ 1.5). This
identified locations at elevated risk for 2020 (Fig. 7). This static graph is
useful but it is difficult to know where county and local boundaries are.
The accompanying interactive dashboard provides tools to view and
explore hexagon-specific information more closely, which may be useful
for those working in the field and act as a template for how data from the
bTB programme could be visualised interactively and disseminated. The
app can be accessed here: https://cvera-jmadden.shinyapps.io/Spatial-b
TB-model/. Unfortunately, due to proprietary and GDPR concerns, we
are limited in the information that can accompany the public version of
the dashboard. Model estimates are included but raw data e.g. the
7
Spatial and Spatio-temporal Epidemiology 39 (2021) 100441
number of bTB cases per hexagon, sett locations etc. have been removed.
Discussion
In this study, we have applied a Bayesian hierarchical spatial model
using random-effects to examine the association between a number of
risk factors and bTB incidence over time, where crucially, the unit of
interest is an area of geographical space, as opposed to an individual
herd. We demonstrated that a model accounting for spatial correlation
significantly improved our model fit in comparison to one in which it
was ignored, and independence assumed. We found an association be­
tween the number of cattle, capture of badgers, forest cover, the number
of farm fragments within a hexagon; and elevated bTB incidence, while
accounting for spatial correlations.
The identification of the number of cattle and the presence of badger
capture as risk factors in this study is not surprising and reflects previous
findings in several reviews and their associated sources [8,42]. How­
ever, our finding that forest cover is associated with bTB incidence is
novel. Forestry has previously been associated with deer presence [43]
and also with badger habitat [44] and it has been suggested that deer
may be an important source of bTB infection in cattle although inves­
tigation of this transmission link is in its infancy [5,21] and further
studies which better account for forestry presence are needed to clarify
the role played by deer in the epidemiology of M. bovis in Ireland [2].
No association was observed between the percentage of dairy herds
within a hexagon and bTB infection, which is at odds with previous
findings. In Spain, in a study of over 1700 herds over 12 years, Alvarez et
al. [45] found that dairy herds were associated with increased
J.M. Madden et al.
Fig. 5. Relative risk (RR) estimates by hexagon in Ireland, at two-yearly intervals between 2013 and 2019 (for clarity). RR maps for all years can be found in the
Supplementary Material Figure 6. Also see Fig. 6 for reference.
within-herd transmission and increased time to recovery from an
outbreak. There are several possible reasons for this disparity. Our herd
classification variable on the AHCS system could have incorrectly clas­
sified herds, as we know it is based on historical records that have not
been updated for some time although the degree to which they are
misclassified is thought to be small. Although both related, it could be
the case that dairy herds are not associated with overall incidence of bTB
although they could have increased within-herd transmission when the
disease finally resides in a herd, reflecting the fact that between-herd
transmission can be different to within-herd rates. The authors did
note that there were shorter intervals between tests in dairy herds
compared to others suggesting discrepancies may be partly due to dif­
ference in definitions and management. However, it must be remem­
bered that this study is a grid-based rather than a herd-based model.
Dairy herds are on average much larger than beef, and larger numbers of
animals typically translates into larger outbreaks (hence a longer time to
recover) in a dairy herd, but it may be the case that this is balanced by
the fact that, ceteris paribus, for a given density of cattle across a
hexagon there will be fewer herds if these are dairy than if they were
beef.
We found no association with the number of inward movements, but
8
Spatial and Spatio-temporal Epidemiology 39 (2021) 100441
we acknowledge that we have used a very crude measure (a simple count
of movements). It is important to note that there is very substantial
animal movement in the Irish trading network [11], and that this
movement essentially occurs without regard to bTB risk. This informa­
tion, coupled with the known problem of residual infection [3] and the
release of herds following two clear tests [46] would suggest that animal
movement is an important contributor to ongoing bTB problems na­
tionally. In the UK, where there are defined areas of differing bTB risk,
the role of cattle movements in bTB spread has been shown to vary by
country (or area) and infection level. In low incidence areas of Scotland
[47] and herds in England with no bTB infection history [48] it was
concluded that the likely source of bTB was cattle movements from
high-risk areas. Within GB, variables linked to cattle movements were
identified in 2005 as the main predictors of bTB occurrence, supporting
the case for movement controls, especially from higher to lower risk
locations, as a disease control measure [49]. In recent work, connection
to more farms in high risk areas up to three movements away from the
root farm increased the odds of a bTB incident [50]. Further, measures to
mitigate spread through animal movement have been central to suc­
cessful national programmes, including in Australia [51] and New
Zealand [52]. Previous research in Ireland has suggested that a
J.M. Madden et al. Spatial and Spatio-temporal Epidemiology 39 (2021) 100441

Fig. 6. Mapped percentage change in relative risk (RR) estimates from two years previous (e.g. map for 2017 represents the percentage change in the RR estimates
for 2017 compared to 2015). See Fig. 5 for reference.

Fig. 7. Map of exceedance probabilities (RR > 1) for Ireland based on 2019 data from the negative binomial BYM model.

9
J.M. Madden et al.
relatively low percentage of bTB restrictions can be attributed to the
introduction of animals into the herd [53,54]. However, this research
must be interpreted with care, as this attribution was determined on the
basis of analysis of the bTB test history of recently introduced bTB
reactor animals. Network analyses can consider both direct and indirect
cattle movements, offering the potential to better quantify the overall
importance of cattle movement on national bTB epidemiology, and this
is being explored in a separate ongoing study by our co-authors.
Disease mapping is a key tool for epidemiologists and policy makers
to monitor and manage the spread of diseases. Hierarchical model-based
approaches that borrow information from neighbouring areas via CAR
random-effects on the local disease rates have recently gained a lot of
popularity, primarily through the increased availability of suitable
software packages [55]. In particular, INLA and the approach used in
this study has been implemented for a range of topics, including infec­
tious diseases e.g. incidence of COVID [56,57], violent crimes [58] and
dengue fever [59]. The key premise of these models is to provide
shrinkage and spatial smoothing of the raw RR estimates that would
otherwise be computed separately in each region [9,60]. This shrinkage
will result in a more stable estimate of the underlying pattern of disease
risk than that provided by the raw estimates. The raw estimates, which
are driven by the population size at risk, can lead to a noisy and blurred
picture of the true unobserved risk [60], for example in a region with
very few animals but a high number of bTB cases. On the other hand, we
also need to be careful when interpreting the shrunken estimates where
there is debate about the performance of these models in recovering the
true risk surface and the degree of smoothing that is performed on the
risks [60]. In our study, there was substantial spatial smoothing per­
formed and the calculation of RR differences help to explore timepoint
variations. In this regard, we explored the inclusion of space-time
interaction terms [30] but this did not improve model fit. We believe
that this reflects the progression of the bTB eradication programme in
Ireland in recent years. Up to 2014/2015, there was a steady decline in
bTB incidence followed by a shift in trends suggestive of a rising tra­
jectory [2], as we have shown in this study. As the true underlying risk
pattern is not known, it is difficult to know model performance or the
sensitivity and specificity of the model i.e. the ability of our model to
detect true patterns of heterogeneity of risk and the ability to discard
false patterns created by noise, respectively. We have illustrated though,
that our model which accounts for spatial correlations provides a
significantly better fit than crude generalised linear models which
ignore this dependency. In addition, as explained in the supplementary
material, we believe that, through the inclusion of this spatial structure,
variable estimates are more suitably estimated from this model.
Richardson et al. [60] undertook an extensive simulation (so that
true underlying risk is known) to examine the performance of these
types of spatial models. They found that Bayesian disease-mapping
models are generally conservative, having a low sensitivity for detect­
ing elevated risk areas that only have a small excess risk but that those
high-risk areas identified, are, generally, specific. They strongly advo­
cate advise against using RRs in isolation but rather to use them together
with exceedance probabilities. They argue that using the whole poste­
rior distribution of the RRs can be better exploited to discriminate be­
tween regions displaying elevated risk and regions with RR close to
background level. As with all statistical models, caution must be taken
when interpreting results.
There are several avenues that could be explored to build on this
study. Our analyses could be extended to an approach which is con­
ducted at a finer resolution, using smaller hexagons, to elucidate more
localised disease transmission, perhaps with a separate model for each of
the Irish counties. In addition, we could interactively visualise regions
and disease outbreaks in near real-time if access to live data was pro­
vided. In this study, we have been concerned with methods describing
the spatial distribution of bTB cases aggregated over small regions but a
different approach could be the exploration of point locations involving
the use of more detailed spatial models [9].
10
Spatial and Spatio-temporal Epidemiology 39 (2021) 100441
Our current study has several strengths. To the best of our knowl­
edge, this is the first study to implement spatio-temporal models through
INLA on bTB data, although MCMC estimation has been implemented
previously [61]. Additionally, in the Irish context, it is the first disease
mapping model to examine the association between potential risk fac­
tors and bTB while accounting for the spatial and temporal correlation in
the data. Considering the importance of “proximity” in relation to in­
fectious diseases, it is surprising that spatial correlation is often ignored.
Specifying the correct correlation structure and model implementation
may be barriers (despite increased availability of suitable software).
Access to data from a large-scale, and high quality bTB testing pro­
gramme, and its associated spatially referenced relational database, is
another strength. Outputs from this study can be used to inform ongoing
disease eradication strategies at a local level. The accompanying inter­
active shiny app offers insights into the potential uses of the rich data the
eradication programme has collated. Finally, the majority of bTB risk
factor studies attempt to incorporate variables on the prior bTB history
of the herd but in our approach, this is inherently captured within the
model as it takes into account the temporal autocorrelation.
The study is not without its limitations. Disease mapping (for areal
data), by its definition, aggregates measures of disease transmission and
their covariates into spatially distinct land areas and area-specific var­
iables are typically constructed. By using aggregated data we may
therefore be masking variation at finer spatial scales. In addition, we are
limited to variables that can be calculated at an areal level. For example,
we did not adjust for deer population in our model as such data is not
available to us at such a fine spatial scale. Badger vaccination was also
not included in our model as there is insufficient longitudinal data to
support its inclusion at this early stage. This was only introduced into
national policy in May 2018 and it is progressively being rolled out
across the country. No model is perfect but, despite these limitations, we
believe the Bayesian model presented in this study offers a novel
approach to monitor bTB progression in space and time. INLA offers an
efficient approach to conduct spatial and spatio-temporal epidemiologic
research.
Data availability
The datasets analysed during this study are available from Depart­
ment of Agriculture, Food and the Marine, but are subject to data pro­
tection regulations and limitations (https://www.agriculture.gov.ie/
dataprotection/).
Compliance with ethical standards
Funding: This research received no specific grant from any funding
agency in the public, commercial or not-for-profit sectors.
Declaration of Competing Interest
None declared.
Supplementary materials
Supplementary material associated with this article can be found, in
the online version, at doi:10.1016/j.sste.2021.100441.
References
Clegg, T.A., Good, M., Hayes, M., Duignan, A., McGrath, G., More, SJ., 2018. Trends and
predictors of large tuberculosis episodes in cattle herds in Ireland. Front Vet. Sci. 5,
86.
More, SJ., 2019. Can bovine TB be eradicated from the Republic of Ireland? Could this be
achieved by 2030? Ir. Vet. J. 72, 3.
More, S.J., Good, M., 2015. Understanding and managing bTB risk: perspectives from
Ireland. Vet. Microbiol. 176 (3-4), 209–218.
Central Statistics Office https://data.cso.ie/.
J.M. Madden et al.
Crispell, J., Cassidy, S., Kenny, K., McGrath, G., Warde, S., Cameron, H., et al., 2020.
Mycobacterium bovis genomics reveals transmission of infection between cattle and
deer in Ireland. Microb. Genom.
Aznar, I., Frankena, K., More, S.J., O’Keeffe, J., McGrath, G., de Jong, MCM., 2018.
Quantification of Mycobacterium bovis transmission in a badger vaccine field trial.
Prev. Vet. Med. 149, 29–37.
Martin, S.W., O’Keeffe, J., Byrne, A.W., Rosen, L.E., White, P.W., McGrath, G., 2020. Is
moving from targeted culling to BCG-vaccination of badgers (Meles meles)
associated with an unacceptable increased incidence of cattle herd tuberculosis in
the Republic of Ireland? A practical non-inferiority wildlife intervention study in the
Republic of Ireland (2011–2017). Prev. Vet. Med. 179, 105004.
Skuce, R.A., Allen, A.R., McDowell, SW., 2012. Herd-level risk factors for bovine
tuberculosis: a literature review. Vet. Med. Int. 2012, 621210.
Lawson, AB., 2008. Bayesian Disease Mapping: Hierarchical Modeling in Spatial
Epidemiology, 3rd Ed. CRC Press, Boca Raton, pp. 1–344.
Blangiardo, M., Cameletti, M., 2015. Spatial and Spatio-Temporal Bayesian Models with
R-INLA. John Wiley & Sons, Hoboken, NJ, USA.
McGrath, G., Tratalos, J.A., More, SJ., 2018. A visual representation of cattle movement
in Ireland during 2016. Ir. Vet. J. 71, 18.
Tratalos, J.A., Madden, J.M., McGrath, G., Graham, D.A., Collins, Á.B., More, SJ., 2020.
Spatial and network characteristics of Irish cattle movements. Prev. Vet. Med. 183,
105095.
EU Trade Directive (1964). Consolidated text: council directive of 26 june 1964 on
animal health problems affecting intra-community trade in bovine animals and
swine (64/432/EEC). http://eur-lex.europa.eu/legal-content/EN/TXT/PDF/?
uri=CELEX:01964L0432-20150527&rid=1.
McGrath, G., Clegg, T.A., More, SJ., 2014. Recent spatial changes in bovine tuberculosis
in the Republic of Ireland. Vet. Rec. 175 (2), 45.
Zimmermann, J., Fealy, R.M., Lydon, K., Mockley, E.M., O’Brien, P., Packham, I.,
Smith, G., Green, S., 2016. The Irish land-parcels identification system (LPIS)–
Experiences in ongoing and recent environmental research and land cover mapping.
Biol. Environ. Proc. R. Ir. Acad. 116B, 53–62.
Jesko, Z., xe, MF, K.L., Eva, M.M., Phillip, O.B., et al., 2016. The Irish land-parcels
identification system (LPIS)-experiences in ongoing and recent environmental
research and land cover mapping. Biol. Environ. Proc. R. Ir. Acad. 116B (1), 53–62.
Byrne, A.W., Barrett, D., Breslin, P., Madden, J.M., O’Keeffe, J., Ryan, E., 2020. Post-
mortem surveillance of bovine tuberculosis in Ireland: herd-level variation in the
probability of herds disclosed with lesions at routine slaughter to have skin test
reactors at follow-up test. Vet. Res. Commun. 1–6.
Houtsma, E., Clegg, T.A., Good, M., More, SJ., 2018. Further improvement in the control
of bovine tuberculosis recurrence in Ireland. Vet. Rec. 183 (20), 622.
Byrne, A.W., Kenny, K., Fogarty, U., O’Keeffe, J.J., More, S.J., McGrath, G., et al., 2015.
Spatial and temporal analyses of metrics of tuberculosis infection in badgers (Meles
meles) from the Republic of Ireland: Trends in apparent prevalence. Prev. Vet. Med.
122 (3), 345–354.
Byrne, A.W., White, P.W., McGrath, G., O’Keeffe, J., Martin, SW., 2014. Risk of
tuberculosis cattle herd breakdowns in Ireland: effects of badger culling effort,
density and historic large-scale interventions. Vet. Res. 45 (1), 109.
VerCauteren, K.C., Lavelle, M.J., Campa, H., 2018. Persistent spillback of bovine
tuberculosis from white-tailed deer to cattle in Michigan, USA: status, strategies, and
needs. Front Vet. Sci. 5, 301.
M. Bossard; J. Feranec; J. Otahel CORINE Land Cover Technical Guide; European
environmental agency technical report No. 40; European Topic Centre on Land
Cover: Copenhagen, Denmark, 2000; pp. 633–638.
Byrne, A.W., Barrett, D., Breslin, P., Madden, J., O’Keeffe, J., Ryan, E., 2020. Bovine
tuberculosis (mycobacterium bovis) outbreak duration in cattle herds in Ireland: a
retrospective observational study. Preprints, 2020080634. https://doi.org/
10.20944/preprints202008.0634.v1.
Moraga, P., 2019. Geospatial Health Data: Modeling and Visualization with R-INLA and
Shiny. Chapman & Hall/CRC Biostatistics Series, Boca Raton, Florida.
Bivand, R.S., Pebesma, E., Gómez-Rubio, V., 2013. Applied Spatial data Analysis with R.
Springer, New York. https://doi.org/10.1007/978-1-4614-7618-4_9.
Chambers, R., Dreassi, E., Salvati, N., 2014. Disease mapping via negative binomial
regression M-quantiles. Stat. Med. 33 (27), 4805–4824.
Besag, J.J.Y., Mollié, A., 1991. Bayesian image restoration with two applications in
spatial statistics Ann. Inst. Stat. Math. 43, 1–59.
Arab, A., 2015. Spatial and spatio-temporal models for modeling epidemiological data
with excess zeros. Int. J. Environ. Res. Public Health 12 (9), 10536–10548.
Asmarian, N., Ayatollahi, S.M.T., Sharafi, Z., Zare, N., 2019. Bayesian spatial joint model
for disease mapping of zero-inflated data with R-INLA: a simulation study and an
application to male breast cancer in Iran. Int. J. Environ. Res. Public Health 16 (22).
Zuur, A.F., Ieno, E.N., Saveliev, A.A., 2017. Beginner’s Guide to Spatial, Temporal and
Spatial-Temporal Ecological Data Analysis With R-INLA. Highland Statistics,
Newburgh, UK.
Sharafi, Z., Asmarian, N., Hoorang, S., Mousavi, A., 2018. Bayesian spatio-temporal
analysis of stomach cancer incidence in Iran, 2003–2010. Stoch. Environ. Res. Risk
Assess. 32, 2943–2950.
Rue, H., Martino, S., Chopin, N., 2009. Approximate Bayesian inference for latent
Gaussian models by using integrated nested Laplace approximations (with
discussion). J. R. Statist. Soc. B 71, 319–392.
11
Spatial and Spatio-temporal Epidemiology 39 (2021) 100441
Gómez-Rubio, V., 2020. Bayesian Inference with INLA. CRC Press.
L. Held, B. Schrödle, H. Rue, editors. Posterior and cross-validatory predictive checks: a
comparison of MCMC and INLA 2010.
Faraway, J., Wang, X., Yue, Y., 2018. Bayesian Regression Modeling with INLA.
Chapman & Hall, p. 304.
Congdon, P., 2005. Bayesian models for categorical data. Wiley Series in Probability and
Statistics. Wiley-Blackwell, Chichester.
Gelman, A., Hwang, J., Vehtari, A., 2014. Understanding predictive information criteria
for Bayesian models. Stat. Comput. 24 (6), 997–1016.
Morris, M., Wheeler-Martin, K., Simpson, D., Mooney, S.J., Gelman, A., DiMaggio, C.,
2019. Bayesian hierarchical spatial models: implementing the besag york mollié
model in stan. Spat. Spatiotemp. Epidemiol. 31, 100301.
Lindgren, F., Rue, H., 2015. Bayesian spatial modeling with R-INLA. J. Stat. Softw. 63
(19), 1–25.
W. Chang, J. Cheng, J.J. Allaire, Y. Xie, and J. McPherson. 2019. “Shiny: web application
framework for R.” R package version 1.3.2.
Thomas, S., Foley, C., Kane, B., Johnston, B.M., Lynch, B., Smith, S., et al., 2019.
Variation in resource allocation in urgent and emergency care systems in Ireland.
BMC Health Serv. Res. 19 (1), 657.
Allen, A.R., Skuce, R.A., Byrne, AW., 2018. Bovine Tuberculosis in Britain and Ireland-a
perfect storm? The confluence of potential ecological and epidemiological
impediments to controlling a chronic infectious disease. Front Vet. Sci. 5, 109.
Gerhardt, P., Arnold, J.M., Hackla¨nder, K., Hochbichler, E., 2013. Determinants of deer
impact in European forests–a systematic literature analysis. Forest. Ecology. Manag.
310, 173–186.
Hammond, R.F., McGrath, G., Martin, SW., 2001. Irish soil and land-use classifications as
predictors of numbers of badgers and badger setts. Prev. Vet. Med. 51 (3-4),
137–148.
Alvarez, J., Perez, A.M., Bezos, J., Casal, C., Romero, B., Rodriguez-Campos, S., et al.,
2012. Eradication of bovine tuberculosis at a herd-level in Madrid, Spain: study of
within-herd transmission dynamics over a 12 year period. BMC Vet. Res. 8, 100.
More, SJ., 2009. What is needed to eradicate bovine tuberculosis successfully: an Ireland
perspective. Vet. J. 180 (3), 275–278.
Gates, M.C., Volkova, V.V., Woolhouse, ME., 2013. Risk factors for bovine tuberculosis in
low incidence regions related to the movements of cattle. BMC Vet. Res. 9, 225.
Carrique-Mas, J.J., Medley, G.F., Green, LE., 2008. Risks for bovine tuberculosis in
British cattle farms restocked after the foot and mouth disease epidemic of 2001.
Prev. Vet. Med. 84 (1-2), 85–93.
Gilbert, M., Mitchell, A., Bourn, D., Mawdsley, J., Clifton-Hadley, R., Wint, W., 2005.
Cattle movements and bovine tuberculosis in Great Britain. Nature 435 (7041),
491–496.
Fielding, H.R., McKinley, T.J., Delahay, R.J., Silk, M.J., McDonald, RA., 2020. Effects of
trading networks on the risk of bovine tuberculosis incidents on cattle farms in Great
Britain. R. Soc. Open Sci. 7 (4), 191806.
More, S.J., Radunz, B., Glanville, RJ., 2015. Lessons learned during the successful
eradication of bovine tuberculosis from Australia. Vet. Rec. 177 (9), 224–232.
Enticott, G., 2014. Biosecurity and the bioeconomy: the case of disease regulation in the
UK & New Zealand. In: Morley, A., Marsden, T. (Eds.), Researching Sustainable
Food: Building the New Sustainability Paradigm. Earthscan, London, UK,
pp. 122–142. https://doi.org/10.4324/9780203083499.
Clegg, T.A., Good, M., More, SJ., 2015. Future risk of bovine tuberculosis recurrence
among higher risk herds in Ireland. Prev. Vet. Med. 118 (1), 71–79.
Clegg, T.A., More, S.J., Higgins, I.M., Good, M., Blake, M., Williams, DH., 2008. Potential
infection-control benefit for Ireland from pre-movement testing of cattle for
tuberculosis. Prev. Vet. Med. 84 (1-2), 94–111.
Vranckx, M., Neyens, T., Faes, C., 2019. Comparison of different software
implementations for spatial disease mapping. Spat. Spatiotem. Epidemiol. 31,
100302.
Rohleder, S., Bozorgmehr, K., 2021. Monitoring the spatiotemporal epidemiology of
COVID-19 incidence and mortality: a small-area analysis in Germany. Emerg. Th.
Epidemiol. https://doi.org/10.21203/rs.3.rs-34572/v1. In Press.
Neyens, T., Faes, C., Vranckx, M., Pepermans, K., Hens, N., Van Damme, P., et al., 2020.
A spatial model to optimise predictions of COVID-19 incidence risk in Belgium using
symptoms as reported in a large-scale online survey. medRxiv,
2020.05.18.20105627.
Vicente, G., Goicoa, T., Ugarte, M.D., 2020. Bayesian inference in multivariate spatio-
temporal areal models using INLA: analysis of gender-based violence in small areas.
Stoch. Environ. Res. Risk Assess. https://doi.org/10.1007/s00477-020-01808-x (IF
2.351) Pub Date: 2020-05-22.
Adin, A., Martínez-Bello, D.A., López-Quílez, A., Ugarte, MD., 2018. Two-level resolution
of relative risk of dengue disease in a hyperendemic city of Colombia. PLoS One 13
(9), e0203382.
Richardson, S., Thomson, A., Best, N., Elliott, P., 2004. Interpreting posterior relative risk
estimates in disease-mapping studies. Environ. Health Perspect. 112 (9), 1016–1025.
Bekara, M.E., Azizi, L., Bénet, J.J., Durand, B., 2016. Spatial-temporal variations of
bovine tuberculosis incidence in France between 1965 and 2000. Transbound.
Emerg. Dis. 63 (1), 101–113.