Jonathan Wakefield

HIBAG—HLA genotype imputation with attribute bagging

Genotyping of classical human leukocyte antigen (HLA) alleles is an essential tool in the analysis of diseases and adverse drug reactions with associations mapping to the major histocompatibility complex (MHC). However, deriving high-resolution HLA types subsequent to whole-genome single-nucleotide polymorphism (SNP) typing or sequencing is often cost prohibitive for large samples. An alternative approach takes advantage of the extended haplotype structure within the MHC to predict HLA alleles using dense SNP genotypes, such as those available from genome-wide SNP panels. Current methods for HLA imputation are difficult to apply or may require the user to have access to large training data sets with SNP and HLA types. We propose HIBAG, HLA Imputation using attribute BAGging, that makes predictions by averaging HLA-type posterior probabilities over an ensemble of classifiers built on bootstrap samples. We assess the performance of HIBAG using our study data (n=2668 subjects of European ancestry) as a training set and HLA data from the British 1958 birth cohort study (n≈1000 subjects) as independent validation samples. Prediction accuracies for HLA-A, B, C, DRB1 and DQB1 range from 92.2% to 98.1% using a set of SNP markers common to the Illumina 1M Duo, OmniQuad, OmniExpress, 660K and 550K platforms. HIBAG performed well compared with the other two leading methods, HLA*IMP and BEAGLE. This method is implemented in a freely available HIBAG R package that includes pre-fit classifiers for European, Asian, Hispanic and African ancestries, providing a readily available imputation approach without the need to have access to large training data sets.

Long-term associations of outdoor air pollution with mortality in Great Britain

Background: Recent studies have indicated long-term effects on mortality of particulate and sulphur dioxide (SO2) pollution, but uncertainties remain over the size of any effects, potential latency and generalisability. Methods: A small area study was performed across electoral wards in Great Britain of mean annual black smoke (BS) and SO2 concentrations (from 1966) and subsequent all-cause and cause-specific mortality using random effect models within a Bayesian framework adjusted for social deprivation and urban/rural classification. Different latencies and changes in associations over time were assessed. Results: Significant associations were found between BS and SO2 concentrations and mortality. The effects were stronger for respiratory illness than other causes of mortality for the most recent exposure periods (shorter latency times) and most recent mortality period (lower pollutant concentrations). In pooled analysis across four sequential 4 year mortality periods (1982–98), adjusted excess relative risk for respiratory mortality was 3.6% (95% CI 2.6% to 4.5%) per 10 μg/m3 BS and 13.2% (95% CI 11.5% to 14.9%) per 10 ppb SO2, and in the most recent period (1994–8) it was 19.3% (95% CI 5.1% to 35.7%) and 21.7% (95% CI 2.9% to 38.5%), respectively. Conclusions: These findings add to the evidence that air pollution has long-term effects on mortality and point to continuing public health risks even at the relatively lower levels of BS and SO2 that now occur. They therefore have importance for policies on public health protection through regulation and control of air pollution.

An investigation of the association between traffic exposure and the diagnosis of asthma in children.

This study investigated whether proximity to traffic at residence location is associated with being diagnosed with asthma as a young child. A survey of parents of children (aged 5–7) in kindergarten and first-grade in 13 schools was completed in Anchorage, Alaska, and Geographical Information System (GIS) mapping was used to obtain an exposure measure based on traffic density within 100 m of the cross streets closest to the childs residence. Using the range of observed exposure values, a score of low, medium or high traffic exposure was assigned to each child. After controlling for individual level confounders, relative to the low referent group, relative risks (95% confidence intervals) of 1.40 (0.77, 2.55) and 2.83 (1.23,6.51) were obtained in the medium and high exposure groups, respectively. For the null hypothesis of no difference in risk, a significance level of 0.056 was obtained, which suggests that further investigation would be worthwhile. Children without a family history of asthma were more likely to have an asthma diagnosis if they resided in a high traffic area than children who had one or more parents with asthma. The relative risk for children without a family history of asthma is 2.43 (1.12, 5.28) for medium exposure and 5.43 (2.08, 13.74) for high exposure. For children with a family history of asthma, the relative risk is 0.66 (0.25, 1.74) for medium exposure and 0.67 (0.12, 3.69) for high exposure. The P-value for the overall “exposure-effect” (i.e. both main effects AND interaction terms) is 0.0097.

A critique of statistical aspects of ecological studies in spatial epidemiology

In this article, the mathematical assumptions of a number of commonly used ecological regression models are made explicit, critically assessed, and related to ecological bias. In particular, the role and interpretation of random effects models are examined. The modeling of spatial variability is considered and related to an underlying continuous spatial field. The examination of such a field with respect to the modeling of risk in relation to a point source highlights an inconsistency in commonly used approaches. A theme of the paper is to examine how plausible individual-level models relate to those used in practice at the aggregate level. The individual-level models acknowledge confounding, within-area variability in exposures and confounders, measurement error and data anomalies and so we can examine how the area-level versions consider these aspects. We briefly discuss designs that efficiently sample individual data and would appear to be useful in environmental settings.

Sources of bias in ecological studies of non-rare events

Ecological studies investigate relationships at the level of the group, rather than at the level of the individual. Although such studies are a common design in epidemiology, it is well-known that estimates may be subject to ecological bias. Most discussion of ecological bias has focused on rare disease events, where the tractability of the loglinear model allows some characterization of the nature of different biases. This paper concentrates on non-rare events, where the Poisson approximation to the binomial distribution is not appropriate. We limit the discussion to bias that arises from within-area variability in exposures and confounders. Our aims are to investigate the likely sizes and directions of bias and, where possible, to suggest methods for controlling the bias or for addressing the sensitivity of inference to assumptions on the nature of the bias. We illustrate that for non-rare events it is much more difficult to characterize the direction of bias than in the rare case. A series of simple numerical examples based on a chronic study of respiratory health illustrate the ideas of the paper.

Ecological Inference: Ecological Inference Incorporating Spatial Dependence

Ecological inference for a series of 2 × 2 tables suffers from an inherent lack of identifiability. Any attempt at a solution to this inferential problem must either (a) incorporate additional information or (b) make assumptions. Without further information and given the data in the margins alone, critical assumptions, such as that of no contextual effects, remain untestable . This suggests a strategy of reporting a series of models based on a range of plausible assumptions, and thus performing a sensitivity analysis with respect to untestable assumptions. The work of this paper is motivated by a voter-registration example from the U.S. state of Louisiana in 1990, where each 2 × 2 table represents one of 64 parishes. When aggregation is on the basis of geography, as in our example, it is intuitive that spatial effects may have a role in an ecological inference analysis. Thus far, such a role has received little attention in the literature. In this paper, we draw on the spatial epidemiological and biostatistical literature and consider the inclusion of a hierarchical spatial model into a sensitivity analysis for ecological inference. We outline issues regarding specification, interpretation, and computation for this particular model when applied to the Louisiana example. A small simulation study suggests that, in the presence of spatial effects, traditional approaches to ecological inference may suffer from incorrect estimation of variability, while models that explicitly allow for spatial effects have generally better performance. 12.

Environmental Transmission of Typhoid Fever in an Urban Slum

Background Enteric fever due to Salmonella Typhi (typhoid fever) occurs in urban areas with poor sanitation. While direct fecal-oral transmission is thought to be the predominant mode of transmission, recent evidence suggests that indirect environmental transmission may also contribute to disease spread. Methods Data from a population-based infectious disease surveillance system (28,000 individuals followed biweekly) were used to map the spatial pattern of typhoid fever in Kibera, an urban informal settlement in Nairobi Kenya, between 2010–2011. Spatial modeling was used to test whether variations in topography and accumulation of surface water explain the geographic patterns of risk. Results Among children less than ten years of age, risk of typhoid fever was geographically heterogeneous across the study area (p = 0.016) and was positively associated with lower elevation, OR = 1.87, 95% CI (1.36–2.57), p <0.001. In contrast, the risk of typhoid fever did not vary geographically or with elevation among individuals less than 6b ten years of age. Conclusions Our results provide evidence of indirect, environmental transmission of typhoid fever among children, a group with high exposure to fecal pathogens in the environment. Spatially targeting sanitation interventions may decrease enteric fever transmission.

Ecological Inference: Prior and Likelihood Choices in the Analysis of Ecological Data

A general statistical framework for ecological inference is presented, and a number of previously proposed approaches are described and critiqued within this framework. In particular, the assumptions that all approaches require to overcome the fundamental nonidentifiability problem of ecological inference are clarified. We describe a number of three-stage Bayesian hierarchical models that are flexible enough to incorporate substantive prior knowledge and additional data. We illustrate that great care must be taken when specifying prior distributions, however. The choice of the likelihood function for aggregate data is discussed, and it is argued that in the case of aggregate 2 × 2 data, a choice that is consistent with a realistic sampling scheme is a convolution of binomial distributions, which naturally incorporate the bounds on the unobserved cells of the constituent 2 × 2 tables. For large marginal counts this choice is computationally daunting, and a simple normal approximation previously described by Wakefield (2004) is discussed. Various computational schemes are described, ranging from an auxiliary data scheme for tables with small counts, to Markov chain Monte Carlo algorithms that are efficient for tables with larger marginal counts. We investigate prior, likelihood, and computational choices with respect to simulated data, and also via registration–race data from four southern U.S. states.

PREDICTIVE MODELING OF CHOLERA OUTBREAKS IN BANGLADESH

Despite seasonal cholera outbreaks in Bangladesh, little is known about the relationship between environmental conditions and cholera cases. We seek to develop a predictive model for cholera outbreaks in Bangladesh based on environmental predictors. To do this, we estimate the contribution of environmental variables, such as water depth and water temperature, to cholera outbreaks in the context of a disease transmission model. We implement a method which simultaneously accounts for disease dynamics and environmental variables in a Susceptible-Infected-Recovered-Susceptible (SIRS) model. The entire system is treated as a continuous-time hidden Markov model, where the hidden Markov states are the numbers of people who are susceptible, infected, or recovered at each time point, and the observed states are the numbers of cholera cases reported. We use a Bayesian framework to fit this hidden SIRS model, implementing particle Markov chain Monte Carlo methods to sample from the posterior distribution of the environmental and transmission parameters given the observed data. We test this method using both simulation and data from Mathbaria, Bangladesh. Parameter estimates are used to make short-term predictions that capture the formation and decline of epidemic peaks. We demonstrate that our model can successfully predict an increase in the number of infected individuals in the population weeks before the observed number of cholera cases increases, which could allow for early notification of an epidemic and timely allocation of resources.

A common framework for ecological inference in epidemiology, political science and sociology

Ecological studies arise within many different disciplines. This chapter describes common approaches to ecological inference in an environmental epidemiology setting, and compares these with traditional approaches in political science and sociology. These approaches vary considerably, both in their use of terminology and notation, and in the relative importance of the various issues that make ecological analyses problematic. The aims of this chapter are twofold. Firstly, we describe ecological inference in an epidemiology setting, where the interest is in the relationship between disease status and exposure to some potential risk factor. We concentrate on those issues which are of particular concern in epidemiology, for example the presence of additional (possibly unmeasured) covariates, termed confounders. Secondly, we seek to unite the current work in epidemiology, political science, and sociology by clarifying differences in terminology, by describing commonly used approaches within a common statistical framework, and by highlighting similarities and differences between these approaches. Often different models can be attributed to different sets of underlying assumptions; we emphasize that such assumptions are crucial in the conclusions drawn from ecological data, and their appropriateness should be carefully considered in any specific situation. Combining approaches from all three disciplines gives a broad range of possible assumptions and available techniques from which to choose.