|
Abstracts of contributed talks |
![]() |
We have developed a Bayesian hierarchical model for the analysis of spatio-temporal data in the presence of interactions, in order to strengthen the interpretation of the spatial patterns of risk that are sustained over time and to pinpoint atypical areas showing evidence of unusual variability in the time pattern of the risk (Abellan et al, 2008). For the prior distribution of space-time interactions, a mixture model with two components is assumed. The first component reflects only residual noise, whereas the second one captures “true” departures from the space and time main effects. The posterior probabilities that each space-time interaction parameter comes from the second component are used to detect areas which do not follow the overall time trend. This approach has been compared with the space-time scan statistics (Kulldorff et al (1997)). As the assumed shape of the prior distribution of space-time interactions is likely to be influential, we have investigated alternative priors and used several Bayesian criteria for comparing these alternatives: DIC, posterior and mixed predictive p-values. Finally, we have applied these methods to analyse the spatio-temporal variations of the risk of bladder cancer in Utah (1973-2004). Our models, of increasing complexity, were used to highlight unusual risk patterns.
In this work we propose a framework for the detection of clusters of disease based on Bayesian Hierarchical Models which extends the spatial scan statistic to a more general case. This extension is established my means of including possible clusters as a dummy variables in a regression model and performing a variable selection to identify the most relevant clusters. Instead of fitting several models using MCMC, which would be very time consuming, we have used approximate methods to compute the marginals of the coefficients of the cluster variables and other parameters of interest. With these approximate methods we can explore the space of possible clusters in a similar way as the spatial scan statistic. Cluster selection is performed with the DIC, so that not only fixed effects models can be compared. We discuss both the Binomial and Poisson cases, as well as mixed-effects models that can cope with overdispersed data. Models for dealing with zero-inflation are also considered. Finally, we show the advantages of this approach using several case studies. In particular, we consider detection of clusters of a single disease in space and space-time, joint clusters of two or more diseases and detection of clusters in the presence of zero-inflation.
This work is motivated by an agricultural issue concerning sugarcane. Sugarcane can be infected with yellowing and stunting disease called sugarcane yellow leaf syndrome. The causal agent sugarcane yellow leaf virus (ScYLV) is transmitted by the aphid melanaphis sacchari. It is well-known that virus-free plants are quickly infected by proximity to other infected plants. We are interested in caracterizing the mechanisms which underlie the spread of the disease. We develop an approach based on survival analysis techniques by considering times to contamination and introducing a contamination factor depending on distance from infected area. A Weibull model is assumed, with a scale parameter depending on location and times to contamination. Maximum likelihood estimation are developed and a Bayesian approach is investigated to assess the contamination rate. We define and study the contamination risk factor. Results on real data are displayed. Simulation studies are conducted.
The mortality due to malignant diseases is extremely unfavourable in Hungary. The leading cause of cancer mortality in both sexes is that of the trachea and lungs. There is a considerable territorial inequality of mortality due to lung cancer inside the country, which was propped up by numerous epidemiological studies. The aim of our study was to reveal spatial distribution of SES adjusted premature mortality due to lung cancer (ICD-X: C33-C34) and to investigate the temporal changes of clusters between 1994 and 2007 years broken down into three years’ consecutive periods. The descriptive study was carried out by Rapid Inquiry Facility (RIF) and SaTScan software. Primary death place clusters were detected for females in and around the capital, Budapest during the whole investigation period. In case of men we did not find similar territorial accumulation of clusters unambiguously being restricted to an area as in case of women. However a fairly constant cluster of male cancer mortality can be identified in the Central-Eastern part of the country. From 2000 new clusters appeared in the North–Eastern, in the South- Eastern and Western part of the country. The results identified the population at risk in the examination period by using RIF. The risk analysis methodology can be applied in public health practice.
Among the specific objectives of the EUROHEIS 2 project is the inclusion of spatio-temporal methods for disease mapping in the Rapid Inquiry Facility (RIF). However, there is not a wide consensus on how to describe temporal and spatial evolution at the same time in a proper way. Although several spatio-temporal disease mapping techniques have been proposed recently, the implementation of these methods is not always easy or adequate for a quick response tool. In this talk we outline a general framework for spatio-temporal models, breaking them up into four stages: the probabilistic model for observations, the components of the linear predictor, the structures of the effects and the inference methodology. For each stage, the most commonly used alternatives in the literature are discussed, with special emphasis on the spatio-temporal interaction. We also review some of the most prominent proposals in the literature that have been used for spatio-temporal disease mapping. These models are classified according to the structure of the temporal trends that may arise, discussing their relative advantages and disadvantages. We conclude with a discussion on what we have identified as the most relevant aspects to be considered in the selection of a methodology for spatio-temporal disease mapping to be included in a RIF-like application.
Bayesian spatio-temporal models formulated in a hierarchical framework (Bernardinelli et al., 1995; Knorr-Held, 2000) are a useful tool for describing spatial and temporal patterns and identifying interactions between time and space within reported data of a disease. The presented application deals with case reporting data on several animal diseases provided by the Swiss federal veterinary office. While for some diseases active surveillance is carried out and reported cases are collected by surveillance programmes there are diseases which are monitored by passive surveillance only. Besides regional heterogeneity in prevalence of these diseases differences in disease occurrence may also be due to e.g. implications when a case is detected or the amount of knowledge on disease characteristics. Since the system of case registration for a region in Switzerland is highly connected with aliation to a canton importance of this factor also has to be investigated. Hence, several spatio-temporal models were fitted to the reporting data to detect unusual spatial and temporal trends. Conclusions on e.g. the sensitivity of the reporting data towards information campaigns, changes in disease awareness and veterinary political issues concerning a disease can be drawn from the obtained results. From an inferential point of view the usability of integrated nested Laplace approximations (Rue et al., 2009) as a tool for Bayesian inference in spatio-temporal models will be pointed out.
Nowadays the use of statistical techniques for small areas estimates are very popular in epidemiology. The approach of Besag, York and Mollié (BYM) is the most accepted model in spatial studies but it ignores the temporal evolution of risk in the study region. Our interest of this work is to obtain a spatio-temporal view of mortality in the municipalities of the Comunitat Valenciana (a Mediterranean region of Spain) with a yearly temporal aggregation and to elaborate a digital atlas of mortality. We use an approach in which the risk is defined by temporal concatenation, as if it were an autoregressive time series of order 1, of spatial and heterogeneous random effects. The consideration of an appropriate dependence structure allows disaggregating in very small spatio-temporal units because information is shared in space and time and so this permits to obtain more reliable estimates. This model was implemented at municipal level in the Comunitat Valenciana for a wide range of mortality causes (selected cancers, circulatory and respiratory diseases, diabetes, AIDS and Alzheimer's disease) in the period 1987-2006, with spatial patterns previously studied through the BYM proposal. We obtained the Spatio-temporal Smoothed Standardized Mortality Ratio for each municipality and year of study. A digital mortality atlas (a web-based system with Geographic Information System -GIS- utilities) was produced. A collection of 460 maps (20 periods x 23 death causes) for each sex (920 totals) are shown in the digital atlas. This has allowed knowing the risk mortality evolution for the studied period. The maps from lung cancer in women and prostate cancer in men are remarkable. From them we can generate hypotheses about risk factors that may have been involved in their mortality distribution. By contrast, the maps of leukaemia in both sexes are homogeneous for the whole period, which does not allow for stating assumptions about underlying factors. A digital atlas like this is a novelty in the field of space-time studies and it makes available the latest information on risk distribution for each of the causes.