Header

UZH-Logo

Maintenance Infos

Statistical analysis of spatio-temporal veterinary surveillance data: Applications of integrated nested Laplace approximations


Schrödle, B. Statistical analysis of spatio-temporal veterinary surveillance data: Applications of integrated nested Laplace approximations. 2011, University of Zurich, Faculty of Science.

Abstract

The surveillance of animal diseases is an important task of national veterinary authorities.
Major aims are the prevention of disease spread and, for zoonoses, the transmission of diseases
from animals to humans. Monitoring is mostly done by passive surveillance, where laboratory
confirmed cases have to be reported. However, such data are often biased due to reporting
delay or underreporting.
Since 1991 the Swiss federal veterinary office (BVET) collects data on about 80 notifiable diseases.
The week number of diagnosis and the location within one of 184 administrative regions
is known for each case. Additionally, data from the Principality of Liechtenstein are
available. The aim of this dissertation is to propose approaches for the statistical modelling of
spatio-temporal patterns based on the specific characteristics of a disease and to apply them
to selected data from the BVET database. If, e.g., regions with a high disease incidence are
identified by such a model, appropriate control measures can be initiated. A further emphasis
is on the presentation of user-friendly software and available model choice criteria.
Spatio-temporal count data are often analyzed using hierarchical Bayesian models. For inference
we propose integrated nested Laplace approximations (INLA) and show their versatile
applicability as regards space-time modelling. High usability is guaranteed by freely available
INLA software. Along with the deviance information criterion we discuss predictive scores,
which are provided by INLA for model choice and criticism. Such scores turn out to be useful
for the evaluation of cross-validatory as well as one-step-ahead forecasts.
We begin with an analysis of aggregated regional data of diseases with constant, endemic risk.
In addition to modelling spatial autocorrelation we describe the inclusion of a linear time trend
for a case study on Coxiellosis in cows. Furthermore, we discuss the appropriate specification
(linear, nonparametric) of a region-specific covariate. For an analysis of Salmonellosis cases
in cows we propose a nonparametric time trend and discuss various modelling options. A
further emphasis is on the versatile interpretation of spatio-temporal interaction terms and
the derivation of criteria to guarantee their identifiability. To analyze case reporting of Bovine
Viral Diarrhoea concerning the affiliation of a region to a Swiss canton, we expand these
models by a coarser, cantonal grid. A comparison with exclusively regional models using
cross-validated scores shows a biased case reporting in several Swiss cantons.
An active surveillance and vaccination program was launched for Bluetongue (BT) in 2008/09
within Switzerland. We perform a regression which assesses the association between individual
information on vaccination, surveillance and altitude and the occurrence of BT for each
farm. Additionally, a two-dimensional location effect on a regular lattice is included in the
model. The results indicate that a vaccination reduces the risk of a BT infection.
We propose a vector-autoregressive model for multivariate time series to model diseases with
local outbreaks. Furthermore, we show how information on networks between regions can directly
be related to observed disease counts. Using this methodology, a spatio-temporal spread
of Coxiellosis in cows between neighbouring regions and by cattle trade is detected. Comparing
one-step-ahead predictive scores it turns out that, for this case study, such a parameterdriven
approach exhibits a better predictive performance than so-called observation-driven
models, where actually observed previous cases govern the infection mechanism

Abstract

The surveillance of animal diseases is an important task of national veterinary authorities.
Major aims are the prevention of disease spread and, for zoonoses, the transmission of diseases
from animals to humans. Monitoring is mostly done by passive surveillance, where laboratory
confirmed cases have to be reported. However, such data are often biased due to reporting
delay or underreporting.
Since 1991 the Swiss federal veterinary office (BVET) collects data on about 80 notifiable diseases.
The week number of diagnosis and the location within one of 184 administrative regions
is known for each case. Additionally, data from the Principality of Liechtenstein are
available. The aim of this dissertation is to propose approaches for the statistical modelling of
spatio-temporal patterns based on the specific characteristics of a disease and to apply them
to selected data from the BVET database. If, e.g., regions with a high disease incidence are
identified by such a model, appropriate control measures can be initiated. A further emphasis
is on the presentation of user-friendly software and available model choice criteria.
Spatio-temporal count data are often analyzed using hierarchical Bayesian models. For inference
we propose integrated nested Laplace approximations (INLA) and show their versatile
applicability as regards space-time modelling. High usability is guaranteed by freely available
INLA software. Along with the deviance information criterion we discuss predictive scores,
which are provided by INLA for model choice and criticism. Such scores turn out to be useful
for the evaluation of cross-validatory as well as one-step-ahead forecasts.
We begin with an analysis of aggregated regional data of diseases with constant, endemic risk.
In addition to modelling spatial autocorrelation we describe the inclusion of a linear time trend
for a case study on Coxiellosis in cows. Furthermore, we discuss the appropriate specification
(linear, nonparametric) of a region-specific covariate. For an analysis of Salmonellosis cases
in cows we propose a nonparametric time trend and discuss various modelling options. A
further emphasis is on the versatile interpretation of spatio-temporal interaction terms and
the derivation of criteria to guarantee their identifiability. To analyze case reporting of Bovine
Viral Diarrhoea concerning the affiliation of a region to a Swiss canton, we expand these
models by a coarser, cantonal grid. A comparison with exclusively regional models using
cross-validated scores shows a biased case reporting in several Swiss cantons.
An active surveillance and vaccination program was launched for Bluetongue (BT) in 2008/09
within Switzerland. We perform a regression which assesses the association between individual
information on vaccination, surveillance and altitude and the occurrence of BT for each
farm. Additionally, a two-dimensional location effect on a regular lattice is included in the
model. The results indicate that a vaccination reduces the risk of a BT infection.
We propose a vector-autoregressive model for multivariate time series to model diseases with
local outbreaks. Furthermore, we show how information on networks between regions can directly
be related to observed disease counts. Using this methodology, a spatio-temporal spread
of Coxiellosis in cows between neighbouring regions and by cattle trade is detected. Comparing
one-step-ahead predictive scores it turns out that, for this case study, such a parameterdriven
approach exhibits a better predictive performance than so-called observation-driven
models, where actually observed previous cases govern the infection mechanism

Statistics

Downloads

1175 downloads since deposited on 17 Jan 2012
133 downloads since 12 months
Detailed statistics

Additional indexing

Item Type:Dissertation
Referees:Held L, Furrer R, Hector A
Communities & Collections:04 Faculty of Medicine > Epidemiology, Biostatistics and Prevention Institute (EBPI)
Dewey Decimal Classification:610 Medicine & health
Language:English
Date:2011
Deposited On:17 Jan 2012 21:39
Last Modified:07 Dec 2017 11:17

Download