09:30-10:15 Geostatistical Models for Environmental Epidemiology Applications
Recent models of stochastic simulations are presented for characterization of spatial relationship between environmental and ecological factors and public health indicators.
Two case studies illustrate the new methodologies:
1) Geostatistical stochastic simulation to measure associations between outdoor air quality and birth weight in portuguese Alentejo Litoral region (Ribeiro et al, 2013). Several studies suggest birth weight is associated with exposure to outdoor air pollution during gestation. Here we applied geostatistical methods to an environmental epidemiology study to measure associations between outdoor air quality and birth weight in a group of babies participating in Gestão Integrada Saúde e Ambiente (GISA) project (Portugal). We performed a semi-ecological analysis, collected individual data on birth weight and health covariates using a questionnaire. We also collected data on day and night place of residence during pregnancy. To collect air quality data we used a lichen diversity biomonitoring program and measured an air quality index value as a surrogate of exposure to outdoor air quality. We applied geostatistical simulation to air quality index values. The set of simulated values at each site provided a measure of personal exposure uncertainty. We used each simulation as input for statistical analysis with generalized linear models. After fitting all multivariate models, we built a mean and confidence interval of the empirical distribution of exposure parameter. Results showed a modest but significant association between exposure to air quality and birth weight among babies exposed to gestational tobacco smoke, even after controlling for significant covariates .
2)
Stochastic Simulation Model for Spatial Characterization of Lung Cancer Mortality Risk and Study of Environmental Factors (Oliveira et al, 2012). In this study, male lung cancer risk was characterized based on a stochastic simulation model of mortality rates. Block Sequential Simulation of mortality rates, measured in counties of different sizes, was implemented and applied to a high spatial resolution normal grid of continental Portugal. The uncertainty in the mortality rate measurements, directly related to differences in the population size of each county, was integrated in a block direct sequential simulation through Poisson kriging of local means and variances. This study addressed three age groups (50-59, 60-69, and 70-79 years). After continuous geographic patterns of lung cancer risk were obtained, several factors potentially associated with the main risky areas were analyzed for southern Portugal. Thus, a certain class of land use and dry events, related with airborne particulate matter, were found to be associated with high-risk areas, resulting in high local spatial correlation patterns in all three age groups.
12:15-13:00 Sampling design issues in environmental data
In environmental monitoring, in particular when aiming to control air pollution levels, the sampling locations are mostly found near their most likely pollution sources, typically in areas of high population density. Traditional models for environmental data do not allow for the fact that the locations chosen for the pollution monitors may depend on the hypothesised concentrations at these locations, a phenomenon known as preferential sampling. In this talk, we briefly describe recent advances to deal with the preferential sampling issue (e.g. Gelfand et al, 2012; Lee et al, 2011; Pati et al, 2011), emphasizing the model-based approach given in Diggle et al (2010). The latter considers log-Gaussian Cox processes to model the stochastic dependence of sampling locations on spatial variable under study. Related to this problem, the monitoring networks present clusters of sampling locations, not equally representative of the observation region. In Menezes et al (2008), a proposal is given to correctly model the spatial dependence structure under clustered sampling, while Margalho et al (2013) suggest to include this information into the prediction model as an explanatory variable. The methodological approaches, here discussed, will be strongly illustrated with their application to biomonitoring data from western Iberian Peninsula (Galicia and Portugal), where since the 90's moss has been used as biomonitor for air pollution to measure heavy metal concentration, with intensified sampling in large urban or industrial areas. The results presented give support to the importance of addressing design issues, such as clustering and preferential sampling, which has largely been ignored in environmental risk analysis.
15:15-16:00 Quantification of annual wildfire risk: A spatio-temporal marked point process approach
Policy responses for local and global fire management depend heavily on the proper understanding of the fire extent as well as its spatio-temporal variation across any given study area. Data which we base our studies and findings are annual satellite imagery data, which consist of the location of observed fire scars in Portugal and their sizes (area burned). Ideally, such data can be assumed to be generated by a nonhomogeneous spatio-temporal marked point process, discrete in time and continuous in space, where points are the centroids of the fire scars and the marks are the sizes of the respective fire scars. The inference primarily targets the nonhomogeneous intensity function of this point process. In this talk, we first look at the first and second order properties of the empirical marked point processes which give us valuable information on the type of models which can be employed to represent the data. The empirical studies indicate strong dependence between the (random) points and the respective marks, complicating the structure of the resulting models. Thus, at first stage we fit a spatio-temporal log Gaussian Cox process (a nonhomogeneous spatio-temporal Poisson process whose log-intensity function is a Gaussian field) to the spatial locations ignoring the marks. We follow the stochastic partial differential equation (SPDE) approach of Lindgren et al(2011). Lindgren et al study certain SPDEs, whose stationary solutions are the Matérn fields, allowing approximate the Gaussian fields by a Gaussian Markov random fields defined over irregular discrete grids. As a byproduct, it is possible to obtain the predictive distribution of the intensity function for future years, permitting us to make probabilistic statements regarding the fire risk in space. Further extensions and a critical evaluation of the shortcomings of the models and methods will also be given.
17:15-18:00 The future of statistical ecology
"We are all interested in the future, for that is where we will spend the rest of our lives" Jeron Criswell King
In this talk, I will give an overview of some recent hot topics in statistical ecology, before peering into my crystal ball and forecasting which areas are likely to become hot in future years. Clearly, the increasingly pressing biodiversity crisis will guide our research priorities from the application side. Continuing increases in computer power, achieved through massive parallelization and quantum computing, will greatly increase the scope of the possible. However, it will always (for the foreseeable future) be easy to construct a realistic model that is both over-parameterized and impossible to compute, and hence parsimony and approximation will be a feature of our lives for many years to come. Finding more accurate yet computable approximations to complex models is an important research priority.
18:00-18:30
09:15-09:30
10:15-11:00 Maximum likelihood estimation of general state-space models using hidden Markov-type approximations with applications in statistical ecology
In this talk, I will discuss an approximate maximum likelihood approach to estimation of general (i.e., possibly nonlinear and possibly non-Gaussian) state-space models (SSMs). The likelihood of an SSM is typically given by a high-order multiple integral that in general cannot be evaluated directly. Numerous approaches for fitting corresponding models have been developed, including, inter alia, Kalman filtering, MCMC, simulated maximum likelihood, INLA and ADMB. Here, I discuss the relative strengths and weaknesses of a very simple yet extremely flexible approach to estimation of SSMs, which exploits techniques developed for hidden Markov models (HMMs). The basic idea is to finely discretize the state space of the considered model, which corresponds to an approximation of the multiple integral appearing in the likelihood by a multiple sum. Brute force evaluation of the resulting multiple sum is usually infeasible, but one can rewrite it in a recursive scheme characteristic of HMMs, which makes numerical likelihood maximization feasible in a lot of scenarios. I will illustrate the methodology by means of simulation experiments and some real data applications, including earthquake count data and mark-recapture data.
11:00-11:30
11:30-12:15 Using effort, sightings, and body condition data to estimate survival and health of individuals and the entire right whale population
Natural and anthropogenic stressors are presumed to impact the health and survival of right whales, yet the effect of these stressors has proven difficult to quantify. To address this issue we built a Hierarchical Bayesian model for survival of individual right whales and fit this model to 30 years worth of sighting data. We assimilate the photographic evidence of condition for individuals as observations of true, but hidden, health. We use the model to make inference on movement, individual survival, and individual health. Estimates of individual health fluctuate across broad ranges, with a mean "healthy" score equaling 84 (on a 0 to 100) scale. In contrast, animals scored with body fat 3, have health values below 50, from which they do not recover. We are able to quantitatively link discrete health observations to underlying continuous states, though estimates are less certain for animals with sparse sighting histories. For individuals, discrete observations of poor skin and body condition in particular, appear to have a strong impact on health. At the population level, health is stable throughout much of the 1980's. Health values in the 1990's decline from a population average of 85 to a low of 72 in 1999. Population health stabilizes and increases in the 2000's, though the decade scale average (76) is lower than the 1980's. Consistent with the PCAD framework, we use these outputs to infer the differential risk associated with the major habitat zones on individual survival, however initial results suggest that coarser geographic partitioning is needed for better parameter estimation. In the future we hope to use these estimates of movement, survival and health to suggest possible management scenarios that increase survival among individuals and the population as a whole.
13:00-14:30
14:30-15:15 Statistical modelling for marine EIAs: some developments, issues and directions
Ecological statisticians in the UK are being increasingly asked to formally address difficult questions regarding the impact of a variety marine activities e.g. installation of off-shore renewables, seismic surveying, military SONAR. Defensible statistical assessments have become effectively a formal regulatory rquirement for the licencing of many of these activities, which have sharply increased in number due to increasing energy demands and the need to meet renewable energy targets. There have also been technological developments that have accentuated the potential for impacts in the marine environment. The UK has EIA obligations relating to these under its own legislature, as well as a variety of EU requirements. This talk will outline some research and experiences in the areas of off-shore turbine installations (current and wind), seismic surveying and SONAR activities. In terms of methods, the following are discussed: the analysis of spatially/temporally auto-correlated data, geodesic smoothing/smoothing on complex domains, agent-based/Monte-Carlo simulations, power analyses via simulation for complex cases, long-term simulations of animal populations. In terms of applications, the following are considered: impacts on abundance and distribution of marine mammals and birds from the installation activities for under-sea current turbines and off-shore wind farm developments; the physical and behavioural effects on marine mammals from military SONAR, seismic surveys and pile-driving for off-shore installations. Examples will reflect experiences in producing mitigation software for the navy and large numbers of commercial EIAs for commercial developments.
16:00-16:30
16:30-17:15 Towards a theory of land-use change impacts on biodiversity
Land-use change has been the main driver of biodiversity loss in terrestrial systems over the last few decades and it is expected to remain a major driver in this century. However, the majority of scenario projections of future biodiversity change have focused on climate change. Here I present our research on the development of an integrated theory of the impacts of land-use change on biodiversity. Many studies have reported the impacts of land-use change on species composition and abundance at the local scale, but the effects at larger scales remain more difficult to study. Until recently, the main large-scale models used in scenario analysis were species-area relationships. However, those models fail to capture the capacity for species to adapt to land-use changes, and I proposed as an alternative the countryside-species area relationship. We have successfully applied this model at the local scale, and we are now tackling the harder problem of predicting large-scale patterns, using regional distribution atlas. But species-area relationships only allow identifying overall changes in numbers of species. In order to be able to model the individual responses of species, we are developing species niche-base models for land-use change equivalent to the existing bioclimatic envelope models for climate change.We are using data from multiyear biodiversity surveys such as the Breeding Bird Survey in the USA to develop models for the relationship between species abundance and habitat cover at the regional scale. For the landcover datasets, we are using comparable datasets at different points in time, such as the NLCD 1992/2001 Retrofit Land Cover Change Product (USA). Finally, one of the problems with available scenario models is that they often ignore the feedbacks from ecosystems back to the socio-economic dynamics. This limits the realism and policy relevance of existing scenarios, as they ignore positive feedback dynamics and associated regime shifts that may occur in ecosystems with dramatic consequences for biodiversity and ecosystem services. I will present a couple socio-ecological model for farmland abandonment that illustrates how regime shifts lead to irreversible dynamics and how the socio-ecological coupling influences those shifts.