Search

We construct flexible spatio-temporal models through stochastic partial differential equations (SPDEs) where both diffusion and advection can be spatially varying. Computations are done through a Gaussian Markov random field approximation of the solution of the SPDE, which is constructed through a finite volume method. The new flexible non-separable model is compared to a flexible separable model both for reconstruction and forecasting and evaluated in terms of root mean square errors and continuous rank probability scores. A simulation study demonstrates that the non-separable model performs better when the data is simulated with non-separable effects such as diffusion and advection. Further, we estimate surrogate models for emulating the output of a ocean model in Trondheimsfjorden, Norway, and simulate observations of autonoumous underwater vehicles. The results show that the flexible non-separable model outperforms the flexible separable model for real-time prediction of unobserved locations.

Anonymizing the GPS locations of observations can bias a spatial model's parameter estimates and attenuate spatial predictions when improperly accounted for, and is relevant in applications from public health to paleoseismology. In this work, we demonstrate that a newly introduced method for geostatistical modeling in the presence of anonymized point locations can be extended to account for more general kinds of positional uncertainty due to location anonymization, including both jittering (a form of random perturbations of GPS coordinates) and geomasking (reporting only the name of the area containing the true GPS coordinates). We further provide a numerical integration scheme that flexibly accounts for the positional uncertainty as well as spatial and covariate information. We apply the method to women's secondary education completion data in the 2018 Nigeria demographic and health survey (NDHS) containing jittered point locations, and the 2016 Nigeria multiple indicator cluster survey (NMICS) containing geomasked locations. We show that accounting for the positional uncertainty in the surveys can improve predictions in terms of their continuous rank probability score., Comment: main manuscript: 31 pages, 6 figures, 2 tables; supplemental materials: 10 pages, 4 figures, 7 tables

The R-package GeoAdjust https://github.com/umut-altay/GeoAdjust-package implements fast empirical Bayesian geostatistical inference for household survey data from the Demographic and Health Surveys Program (DHS) using Template Model Builder (TMB). DHS household survey data is an important source of data for tracking demographic and health indicators, but positional uncertainty has been intentionally introduced in the GPS coordinates to preserve privacy. GeoAdjust accounts for such positional uncertainty in geostatistical models containing both spatial random effects and raster- and distance-based covariates. The R package supports Gaussian, binomial and Poisson likelihoods with identity link, logit link, and log link functions respectively. The user defines the desired model structure by setting a small number of function arguments, and can easily experiment with different hyperparameters for the priors. GeoAdjust is the first software package that is specifically designed to address positional uncertainty in the GPS coordinates of point referenced household survey data. The package provides inference for model parameters and can predict values at unobserved locations.

Obtaining high-resolution maps of precipitation data can provide key insights to stakeholders to assess a sustainable access to water resources at urban scale. Mapping a nonstationary, sparse process such as precipitation at very high spatial resolution requires the interpolation of global datasets at the location where ground stations are available with statistical models able to capture complex non-Gaussian global space-time dependence structures. In this work, we propose a new approach based on capturing the spatial dependence of a latent Gaussian process via a locally deformed Stochastic Partial Differential Equation (SPDE) with a buffer allowing for a different spatial structure across land and sea. The finite volume approximation of the SPDE, coupled with Integrated Nested Laplace Approximation ensures feasible Bayesian inference for tens of millions of observations. The simulation studies showcase the improved predictability of the proposed approach against stationary and no-buffer alternatives. The proposed approach is then used to yield high resolution simulations of daily precipitation across the United States.

Isotropic covariance structures can be unreasonable for phenomena in three-dimensional spaces such as the ocean. In the ocean, the variability of the response may vary with depth, and ocean currents may lead to spatially varying anisotropy. We construct a class of non-stationary anisotropic Gaussian random fields (GRFs) in three dimensions through stochastic partial differential equations (SPDEs) where computations are done using Gaussian Markov random field approximations. The approach is proven in a simulation study where the amount of data required to estimate these models is explored. Then, the method is applied to construct a GRF prior on an ocean mass outside Trondheim, Norway, based on simulations from the complex numerical ocean model SINMOD. This GRF prior is compared to a stationary anisotropic GRF using in-situ measurements collected with an autonomous underwater vehicle where our approach outperforms the stationary anisotropic GRF for real-time prediction of unobserved locations.

Fine-scale covariate rasters are routinely used in geostatistical models for mapping demographic and health indicators based on household surveys from the Demographic and Health Surveys (DHS) program. However, the geostatistical analyses ignore the fact that GPS coordinates in DHS surveys are jittered for privacy purposes. We demonstrate the need to account for this jittering, and we propose a computationally efficient approach that can be routinely applied. We use the new method to analyse the prevalence of completion of secondary education for 20--49 year old women in Nigeria in 2018 based on the 2018 DHS survey. The analysis demonstrates substantial changes in the estimates of spatial range and fixed effects compared to when we ignore jittering. Through a simulation study that mimics the dataset, we demonstrate that accounting for jittering reduces attenuation in the estimated coefficients for covariates and improves predictions. The results also show that the common approach of averaging covariate values in windows around the observed locations does not lead to the same improvements as accounting for jittering.

Spatial aggregation with respect to a population distribution involves estimating aggregate quantities for a population based on an observation of individuals in a subpopulation. In this context, a geostatistical workflow must account for three major sources of `aggregation error': aggregation weights, fine scale variation, and finite population variation. However, common practice is to treat the unknown population distribution as a known population density and ignore empirical variability in outcomes. We improve common practice by introducing a `sampling frame model' that allows aggregation models to account for the three sources of aggregation error simply and transparently. We compare the proposed and the traditional approach using two simulation studies that mimic neonatal mortality rate (NMR) data from the 2014 Kenya Demographic and Health Survey (KDHS2014). For the traditional approach, undercoverage/overcoverage depends arbitrarily on the aggregation grid resolution, while the new approach exhibits low sensitivity. The differences between the two aggregation approaches increase as the population of an area decreases. The differences are substantial at the second administrative level and finer, but also at the first administrative level for some population quantities. We find differences between the proposed and traditional approach are consistent with those we observe in an application to NMR data from the KDHS2014., Comment: main manuscript: 33 pages, 5 figures, 5 tables; supplemental materials: 15 pages, 2 figures, 15 tables

Household survey data from the Demographic and Health Surveys (DHS) Program is published with GPS coordinates. However, almost all geostatistical analyses of such data ignore that the published GPS coordinates are randomly displaced (jittered). In this short report, we develop a geostatistical model that accounts for the positional uncertainty when analysing DHS surveys, and provide a fast implementation using Template Model Builder. The key focus is inference with Gaussian random fields under positional uncertainty, and our approach works for both Gaussian and non-Gaussian likelihoods. A simulation study with a binomial observation model shows that the new approach performs equally or better than the common approach of ignoring jittering, both in terms of more accurate parameter estimates and improved predictive measures. We demonstrate that the improvement would be larger under stronger jittering. An analysis of contraceptive use in Kenya shows that the approach is fast and easy to use in practice.

The emerging need for subnational estimation of demographic and health indicators in low- and middle-income countries (LMICs) is driving a move from design-based area-level approaches to unit-level methods. The latter are model-based and overcome data sparsity by borrowing strength across covariates and space and can, in principle, be leveraged to create fine-scale pixel level maps based on household surveys. However, typical implementations of the model-based approaches do not fully acknowledge the complex survey design, and do not enjoy the theoretical consistency of design-based approaches. We describe how spatial methods are currently used for prevalence mapping in the context of LMICs, highlight the key challenges that need to be overcome, and propose a new approach, which is methodologically closer in spirit to small area estimation. The main discussion points are demonstrated through a case study of vaccination coverage in Nigeria based on 2018 Demographic and Health Surveys (DHS) data. We discuss our key findings both generally and with an emphasis on the implications for popular approaches undertaken by industrial producers of subnational prevalence estimates., Comment: An extensive revision of the previous version of the preprint. The spatial aspects have been fleshed out more, and the temporal aspects have been removed

Priors allow us to robustify inference and to incorporate expert knowledge in Bayesian hierarchical models. This is particularly important when there are random effects that are hard to identify based on observed data. The challenge lies in understanding and controlling the joint influence of the priors for the variance parameters, and makemyprior is an R package that guides the formulation of joint prior distributions for variance parameters. A joint prior distribution is constructed based on a hierarchical decomposition of the total variance in the model along a tree, and takes the entire model structure into account. Users input their prior beliefs or express ignorance at each level of the tree. Prior beliefs can be general ideas about reasonable ranges of variance values and need not be detailed expert knowledge. The constructed priors lead to robust inference and guarantee proper posteriors. A graphical user interface facilitates construction and assessment of different choices of priors through visualization of the tree and joint prior. The package aims to expand the toolbox of applied researchers and make priors an active component in their Bayesian workflow., Comment: 43 pages, 15 figures

The increasing availability of complex survey data, and the continued need for estimates of demographic and health indicators at a fine spatial and temporal scale, which leads to issues of data sparsity, has led to the need for spatio-temporal smoothing methods that acknowledge the manner in which the data were collected. The open source R package SUMMER implements a variety of methods for spatial or spatio-temporal smoothing of survey data. The emphasis is on small-area estimation. We focus primarily on indicators in a low and middle-income countries context. Our methods are particularly useful for data from Demographic Health Surveys and Multiple Indicator Cluster Surveys. We build upon functions within the survey package, and use INLA for fast Bayesian computation. This paper includes a brief overview of these methods and illustrates the workflow of accessing and processing surveys, estimating subnational child mortality rates, and visualizing results with both simulated data and DHS surveys.

Current implementations of multiresolution methods are limited in terms of possible types of responses and approaches to inference. We provide a multiresolution approach for spatial analysis of non-Gaussian responses using latent Gaussian models and Bayesian inference via integrated nested Laplace approximation (INLA). The approach builds on `LatticeKrig', but uses a reparameterization of the model parameters that is intuitive and interpretable so that modeling and prior selection can be guided by expert knowledge about the different spatial scales at which dependence acts. The priors can be used to make inference robust and integration over model parameters allows for more accurate posterior estimates of uncertainty. The extended LatticeKrig (ELK) model is compared to a standard implementation of LatticeKrig (LK), and a standard Mat\'ern model, and we find modest improvement in spatial oversmoothing and prediction for the ELK model for counts of secondary education completion for women in Kenya collected in the 2014 Kenya demographic health survey. Through a simulation study with Gaussian responses and a realistic mix of short and long scale dependencies, we demonstrate that the differences between the three approaches for prediction increases with distance to nearest observation., Comment: main manuscript: 33 pages, 7 figures, 2 tables; supplemental materials: 9 pages, 3 figures, 5 tables

The need for rigorous and timely health and demographic summaries has provided the impetus for an explosion in geographic studies, with a common approach being the production of pixel-level maps, particularly in low and middle income countries. In this context, household surveys are a major source of data, usually with a two-stage cluster design with stratification by region and urbanicity. Accurate estimates are of crucial interest for precision public health policy interventions, but many current studies take a cavalier approach to acknowledging the sampling design, while presenting results at a fine geographic scale. In this paper we investigate the extent to which accounting for sample design can affect predictions at the aggregate level, which is usually the target of inference. We describe a simulation study in which realistic sampling frames are created for Kenya, based on population and demographic information, with a survey design that mimics a Demographic Health Survey (DHS). We compare the predictive performance of various commonly-used models. We also describe a cluster level model with a discrete spatial smoothing prior that has not been previously used, but provides reliable inference. We find that including stratification and cluster level random effects can improve predictive performance. Spatially smoothed direct (weighted) estimates were robust to priors and survey design. Continuous spatial models performed well in the presence of fine scale variation; however, these models require the most "hand holding". Subsequently, we examine how the models perform on real data; specifically we model the prevalence of secondary education for women aged 20-29 using data from the 2014 Kenya DHS., Comment: Main text: 35 pages, 5 figures, 2 tables. Supplementary materials: 63 pages, 5 figures, 21 tables

Variance parameters in additive models are typically assigned independent priors that do not account for model structure. We present a new framework for prior selection based on a hierarchical decomposition of the total variance along a tree structure to the individual model components. For each split in the tree, an analyst may be ignorant or have a sound intuition on how to attribute variance to the branches. In the former case a Dirichlet prior is appropriate to use, while in the latter case a penalised complexity (PC) prior provides robust shrinkage. A bottom-up combination of the conditional priors results in a proper joint prior. We suggest default values for the hyperparameters and offer intuitive statements for eliciting the hyperparameters based on expert knowledge. The prior framework is applicable for R packages for Bayesian inference such as INLA and RStan. Three simulation studies show that, in terms of the application-specific measures of interest, PC priors improve inference over Dirichlet priors when used to penalise different levels of complexity in splits. However, when expressing ignorance in a split, Dirichlet priors perform equally well and are preferred for their simplicity. We find that assigning current state-of-the-art default priors for each variance parameter individually is less transparent and does not perform better than using the proposed joint priors. We demonstrate practical use of the new framework by analysing spatial heterogeneity in neonatal mortality in Kenya in 2010-2014 based on complex survey data., Comment: 64 pages, 33 figures

Coming up with Bayesian models for spatial data is easy, but performing inference with them can be challenging. Writing fast inference code for a complex spatial model with realistically-sized datasets from scratch is time-consuming, and if changes are made to the model, there is little guarantee that the code performs well. The key advantages of R-INLA are the ease with which complex models can be created and modified, without the need to write complex code, and the speed at which inference can be done even for spatial problems with hundreds of thousands of observations. R-INLA handles latent Gaussian models, where fixed effects, structured and unstructured Gaussian random effects are combined linearly in a linear predictor, and the elements of the linear predictor are observed through one or more likelihoods. The structured random effects can be both standard areal model such as the Besag and the BYM models, and geostatistical models from a subset of the Mat\'ern Gaussian random fields. In this review, we discuss the large success of spatial modelling with R-INLA and the types of spatial models that can be fitted, we give an overview of recent developments for areal models, and we give an overview of the stochastic partial differential equation (SPDE) approach and some of the ways it can be extended beyond the assumptions of isotropy and separability. In particular, we describe how slight changes to the SPDE approach leads to straight-forward approaches for non-stationary spatial models and non-separable space-time models., Comment: Extensive update, restructuring of sections

Accurate estimates of the under-5 mortality rate (U5MR) in a developing world context are a key barometer of the health of a nation. This paper describes new models to analyze survey data on mortality in this context. We are interested in both spatial and temporal description, that is, wishing to estimate U5MR across regions and years, and to investigate the association between the U5MR and spatially-varying covariate surfaces. We illustrate the methodology by producing yearly estimates for subnational areas in Kenya over the period 1980 - 2014 using data from demographic health surveys (DHS). We use a binomial likelihood with fixed effects for the urban/rural stratification to account for the complex survey design. We carry out smoothing using Bayesian hierarchical models with continuous spatial and temporally discrete components. A key component of the model is an offset to adjust for bias due to the effects of HIV epidemics. Substantively, there has been a sharp decline in U5MR in the period 1980 - 2014, but large variability in estimated subnational rates remains. A priority for future research is understanding this variability. Temperature, precipitation and a measure of malaria infection prevalence were candidates for inclusion in the covariate model., Comment: 36 pages, 11 figures

Background: The direct modeling of water networks is not a common practice in modern epidemiology. While space often serves as a proxy, it can be problematic. There are multiple ways to directly model water networks, but these methods are not straightforward and can be difficult to implement. This study suggests a simple approach for modeling water networks and diseases, and applies this method to a dataset of self-reported gastrointestinal conditions from a questionnaire-based population health survey in central Norway. Method: Our approach is based on a standard conditional autoregressive (CAR) model. An inverse matrix was constructed, with nodes weighted based on the distance to neighboring nodes within the networks. This matrix was then fitted as a generic model. To illustrate its possible use, we utilized data taken from a questionnaire-based population health survey, the HUNT Study, to measure self-reported gastrointestinal complaints. For hypothesis testing, we used the deviance information criterion (DIC) and included variables in a stepwise manner. Results: The full model converged after six hours. We found no relation between the water networks and the health conditions of people whose residences connected to different parts of the network in the geographical area studied. Conclusion: All water network models are simplifications of the real networks. Nevertheless, we suggest a valid approach for distinguishing between the general spatial effect and the water network using a generic model.

Priors are important for achieving proper posteriors with physically meaningful covariance structures for Gaussian random fields (GRFs) since the likelihood typically only provides limited information about the covariance structure under in-fill asymptotics. We extend the recent Penalised Complexity prior framework and develop a principled joint prior for the range and the marginal variance of one-dimensional, two-dimensional and three-dimensional Mat\'ern GRFs with fixed smoothness. The prior is weakly informative and penalises complexity by shrinking the range towards infinity and the marginal variance towards zero. We propose guidelines for selecting the hyperparameters, and a simulation study shows that the new prior provides a principled alternative to reference priors that can leverage prior knowledge to achieve shorter credible intervals while maintaining good coverage. We extend the prior to a non-stationary GRF parametrized through local ranges and marginal standard deviations, and introduce a scheme for selecting the hyperparameters based on the coverage of the parameters when fitting simulated stationary data. The approach is applied to a dataset of annual precipitation in southern Norway and the scheme for selecting the hyperparameters leads to concervative estimates of non-stationarity and improved predictive performance over the stationary model.

A stationary spatial model is an idealization and we expect that the true dependence structures of physical phenomena are spatially varying, but how should we handle this non-stationarity in practice? We study the challenges involved in applying a flexible non-stationary model to a dataset of annual precipitation in the conterminous US, where exploratory data analysis shows strong evidence of a non-stationary covariance structure. The aim of this paper is to investigate the modelling pipeline once non-stationarity has been detected in spatial data. We show that there is a real danger of over-fitting the model and that careful modelling is necessary in order to properly account for varying second-order structure. In fact, the example shows that sometimes non-stationary Gaussian random fields are not necessary to model non-stationary spatial data., Comment: Minor change from previous version. arXiv admin note: text overlap with arXiv:1306.0408

A non-stationary spatial Gaussian random field (GRF) is described as the solution of an inhomogeneous stochastic partial differential equation (SPDE), where the covariance structure of the GRF is controlled by the coefficients in the SPDE. This allows for a flexible way to vary the covariance structure, where intuition about the resulting structure can be gained from the local behaviour of the differential equation. Additionally, computations can be done with computationally convenient Gaussian Markov random fields which approximate the true GRFs. The model is applied to a dataset of annual precipitation in the conterminous US. The non-stationary model performs better than a stationary model measured with both CRPS and the logarithmic scoring rule.

Gaussian random fields (GRFs) constitute an important part of spatial modelling, but can be computationally infeasible for general covariance structures. An efficient approach is to specify GRFs via stochastic partial differential equations (SPDEs) and derive Gaussian Markov random field (GMRF) approximations of the solutions. We consider the construction of a class of non-stationary GRFs with varying local anisotropy, where the local anisotropy is introduced by allowing the coefficients in the SPDE to vary with position. This is done by using a form of diffusion equation driven by Gaussian white noise with a spatially varying diffusion matrix. This allows for the introduction of parameters that control the GRF by parametrizing the diffusion matrix. These parameters and the GRF may be considered to be part of a hierarchical model and the parameters estimated in a Bayesian framework. The results show that the use of an SPDE with non-constant coefficients is a promising way of creating non-stationary spatial GMRFs that allow for physical interpretability of the parameters, although there are several remaining challenges that would need to be solved before these models can be put to general practical use.