COMPARING METHODS FOR CALCULATING CONFIDENCE INTERVALS FOR VACCINE EFFICACY

A method is introduced for computing a Bayesian 95 per cent posterior probability region for vaccine efficacy. This method assumes independent vague gamma prior distributions for the incidence rates on each arm of the trial, and a Poisson likelihood for the counts of incident cases of infection. The approach is similar in spirit to the Bayesian analysis of the binomial risk ratio described by Aitchison and Bacon-Shone. However, the focus of our interest is not on incorporating prior information into the design of trials for efficacy, but rather on evaluating whether or not the Bayesian approach with vague prior information produces comparable results to a frequentist approach. A review of methods for constructing exact and large sample intervals for vaccine efficacy is provided as a framework for comparison. The confidence interval methods are assessed by comparing the size and power of tests of vaccine efficacy in proposed intermediate sized randomized double blinded placebo controlled trials.     

A marginalized zero‐inflated Poisson regression model with overall exposure effects

The zero‐inflated Poisson (ZIP) regression model is often employed in public health research to examine the relationships between exposures of interest and a count outcome exhibiting many zeros, in excess of the amount expected under sampling from a Poisson distribution. The regression coefficients of the ZIP model have latent class interpretations, which correspond to a susceptible subpopulation at risk for the condition with counts generated from a Poisson distribution and a non‐susceptible subpopulation that provides the extra or excess zeros. The ZIP model parameters, however, are not well suited for inference targeted at marginal means, specifically, in quantifying the effect of an explanatory variable in the overall mixture population. We develop a marginalized ZIP model approach for independent responses to model the population mean count directly, allowing straightforward inference for overall exposure effects and empirical robust variance estimation for overall log‐incidence density ratios. Through simulation studies, the performance of maximum likelihood estimation of the marginalized ZIP model is assessed and compared with other methods of estimating overall exposure effects. The marginalized ZIP model is applied to a recent study of a motivational interviewing‐based safer sex counseling intervention, designed to reduce unprotected sexual act counts. Copyright © 2014 John Wiley & Sons, Ltd.     

Mixed effect Poisson log-linear models for clinical and epidemiological sleep hypnogram data

Bayesian Poisson log-linear multilevel models scalable to epidemiological studies are proposed to investigate population variability in sleep state transition rates. Hierarchical random effects are used to account for pairings of subjects and repeated measures within those subjects, as comparing diseased with non-diseased subjects while minimizing bias is of importance. Essentially, non-parametric piecewise constant hazards are estimated and smoothed, allowing for time-varying covariates and segment of the night comparisons. The Bayesian Poisson regression is justified through a re-derivation of a classical algebraic likelihood equivalence of Poisson regression with a log(time) offset and survival regression assuming exponentially distributed survival times. Such re-derivation allows synthesis of two methods currently used to analyze sleep transition phenomena: stratified multi-state proportional hazards models and log-linear generalized estimating equations (GEE) models for transition counts. An example data set from the Sleep Heart Health Study is analyzed. Supplementary material includes the analyzed data set as well as the code for a reproducible analysis.     

Analytical approximation of exact Poisson-lognormal likelihood functions

Simple analytical approximations of exact Poisson-lognormal likelihood functions are obtained numerically. The Poisson-lognormal statistical model describes counting measurements with lognormally distributed normalization factors. The analytical expressions for the likelihood function allow maximum likelihood data fitting using nonlinear-least-squares-minimization computer programs. A computer program that converts count data into analytical approximation parameters as described in this note is freely available for downloading from the Los Alamos Bayesian Web site (www.lanl.gov/bayesian).     

Comparing hierarchical models via the marginalized deviance information criterion

Hierarchical models are extensively used in pharmacokinetics and longitudinal studies. When the estimation is performed from a Bayesian approach, model comparison is often based on the deviance information criterion (DIC). In hierarchical models with latent variables, there are several versions of this statistic: the conditional DIC (cDIC) that incorporates the latent variables in the focus of the analysis and the marginalized DIC (mDIC) that integrates them out. Regardless of the asymptotic and coherency difficulties of cDIC, this alternative is usually used in Markov chain Monte Carlo (MCMC) methods for hierarchical models because of practical convenience. The mDIC criterion is more appropriate in most cases but requires integration of the likelihood, which is computationally demanding and not implemented in Bayesian software. Therefore, we consider a method to compute mDIC by generating replicate samples of the latent variables that need to be integrated out. This alternative can be easily conducted from the MCMC output of Bayesian packages and is widely applicable to hierarchical models in general. Additionally, we propose some approximations in order to reduce the computational complexity for large‐sample situations. The method is illustrated with simulated data sets and 2 medical studies, evidencing that cDIC may be misleading whilst mDIC appears pertinent.     

Marginalized zero‐inflated negative binomial regression with application to dental caries

The zero‐inflated negative binomial regression model (ZINB) is often employed in diverse fields such as dentistry, health care utilization, highway safety, and medicine to examine relationships between exposures of interest and overdispersed count outcomes exhibiting many zeros. The regression coefficients of ZINB have latent class interpretations for a susceptible subpopulation at risk for the disease/condition under study with counts generated from a negative binomial distribution and for a non‐susceptible subpopulation that provides only zero counts. The ZINB parameters, however, are not well‐suited for estimating overall exposure effects, specifically, in quantifying the effect of an explanatory variable in the overall mixture population. In this paper, a marginalized zero‐inflated negative binomial regression (MZINB) model for independent responses is proposed to model the population marginal mean count directly, providing straightforward inference for overall exposure effects based on maximum likelihood estimation. Through simulation studies, the finite sample performance of MZINB is compared with marginalized zero‐inflated Poisson, Poisson, and negative binomial regression. The MZINB model is applied in the evaluation of a school‐based fluoride mouthrinse program on dental caries in 677 children. Copyright © 2015 John Wiley & Sons, Ltd.     

How does Poisson kriging compare to the popular BYM model for mapping disease risks?

Background

Geostatistical techniques are now available to account for spatially varying population sizes and spatial patterns in the mapping of disease rates. At first glance, Poisson kriging represents an attractive alternative to increasingly popular Bayesian spatial models in that: 1) it is easier to implement and less CPU intensive, and 2) it accounts for the size and shape of geographical units, avoiding the limitations of conditional auto-regressive (CAR) models commonly used in Bayesian algorithms while allowing for the creation of isopleth risk maps. Both approaches, however, have never been compared in simulation studies, and there is a need to better understand their merits in terms of accuracy and precision of disease risk estimates.

Results

Besag, York and Mollie's (BYM) model and Poisson kriging (point and area-to-area implementations) were applied to age-adjusted lung and cervix cancer mortality rates recorded for white females in two contrasted county geographies: 1) state of Indiana that consists of 92 counties of fairly similar size and shape, and 2) four states in the Western US (Arizona, California, Nevada and Utah) forming a set of 118 counties that are vastly different geographical units. The spatial support (i.e. point versus area) has a much smaller impact on the results than the statistical methodology (i.e. geostatistical versus Bayesian models). Differences between methods are particularly pronounced in the Western US dataset: BYM model yields smoother risk surface and prediction variance that changes mainly as a function of the predicted risk, while the Poisson kriging variance increases in large sparsely populated counties. Simulation studies showed that the geostatistical approach yields smaller prediction errors, more precise and accurate probability intervals, and allows a better discrimination between counties with high and low mortality risks. The benefit of area-to-area Poisson kriging increases as the county geography becomes more heterogeneous and when data beyond the adjacent counties are used in the estimation. The trade-off cost for the easier implementation of point Poisson kriging is slightly larger kriging variances, which reduces the precision of the model of uncertainty.

Conclusion

Bayesian spatial models are increasingly used by public health officials to map mortality risk from observed rates, a preliminary step towards the identification of areas of excess. More attention should however be paid to the spatial and distributional assumptions underlying the popular BYM model. Poisson kriging offers more flexibility in modeling the spatial structure of the risk and generates less smoothing, reducing the likelihood of missing areas of high risk.     

Bayesian analysis of the differences of count data

Paired count data usually arise in medicine when before and after treatment measurements are considered. In the present paper we assume that the correlated paired count data follow a bivariate Poisson distribution in order to derive the distribution of their difference. The derived distribution is shown to be the same as the one derived for the difference of the independent Poisson variables, thus recasting interest on the distribution introduced by Skellam. Using this distribution we remove correlation, which naturally exists in paired data, and we improve the quality of our inference by using exact distributions instead of normal approximations. The zero-inflated version is considered to account for an excess of zero counts. Bayesian estimation and hypothesis testing for the models considered are discussed. An example from dental epidemiology is used to illustrate the proposed methodology.     

Are marginalized two-part models superior to non-marginalized two-part models for count data with excess zeroes? Estimation of marginal effects,model misspecification,and model selection

The marginalized two-part models, including the marginalized zero-inflated Poisson and negative binomial models, have been proposed in the literature for modelling cross-sectional healthcare utilization count data with excess zeroes and overdispersion. The motivation for these proposals was to directly capture the overall marginal effects and to avoid post-modelling effect calculations that are needed for the non-marginalized conventional two-part models. However, are marginalized two-part models superior to non-marginalized two-part models because of their structural property? Is it true that the marginalized two-part models can provide direct marginal inference? This article aims to answer these questions through a comprehensive investigation. We first summarize the existing non-marginalized and marginalized two-part models and then develop marginalized hurdle Poisson and negative binomial models for cross-sectional count data with abundant zero counts. Our interest in the investigation lies particularly in the (average) marginal effect and (average) incremental effect and the comparison of these effects. The estimators of these effects are presented, and variance estimators are derived by using delta methods and Taylor series approximations. Though the marginalized models attract attention because of the alleged convenience of direct marginal inference, we provide evidence for the impact of model misspecification of the marginalized models over the conventional models, and provide evidence for the importance of goodness-of-fit evaluation and model selection in differentiating between the marginalized and non-marginalized models. An empirical analysis of the German Socioeconomic Panel data is presented.     

Signal detection in FDA AERS database using Dirichlet process

In the recent two decades, data mining methods for signal detection have been developed for drug safety surveillance, using large post‐market safety data. Several of these methods assume that the number of reports for each drug–adverse event combination is a Poisson random variable with mean proportional to the unknown reporting rate of the drug–adverse event pair. Here, a Bayesian method based on the Poisson–Dirichlet process (DP) model is proposed for signal detection from large databases, such as the Food and Drug Administration's Adverse Event Reporting System (AERS) database. Instead of using a parametric distribution as a common prior for the reporting rates, as is the case with existing Bayesian or empirical Bayesian methods, a nonparametric prior, namely, the DP, is used. The precision parameter and the baseline distribution of the DP, which characterize the process, are modeled hierarchically. The performance of the Poisson–DP model is compared with some other models, through an intensive simulation study using a Bayesian model selection and frequentist performance characteristics such as type‐I error, false discovery rate, sensitivity, and power. For illustration, the proposed model and its extension to address a large amount of zero counts are used to analyze statin drugs for signals using the 2006–2011 AERS data. Copyright © 2015 John Wiley & Sons, Ltd.     

Using full probability models to compute probabilities of actual interest to decision makers

The objective of this paper is to illustrate the advantages of the Bayesian approach in quantifying, presenting, and reporting scientific evidence and in assisting decision making. Three basic components in the Bayesian framework are the prior distribution, likelihood function, and posterior distribution. The prior distribution describes analysts' belief a priori, the likelihood function captures how data modify the prior knowledge; and the posterior distribution synthesizes both prior and likelihood information. The Bayesian approach treats the parameters of interest as random variables, uses the entire posterior distribution to quantify the evidence, and reports evidence in a "probabilistic" manner. Two clinical examples are used to demonstrate the value of the Bayesian approach to decision makers. Using either an uninformative or a skeptical prior distribution, these examples show that the Bayesian methods allow calculations of probabilities that are usually of more interest to decision makers, e.g., the probability that treatment A is similar to treatment B, the probability that treatment A is at least 5% better than treatment B, and the probability that treatment A is not within the "similarity region" of treatment B, etc. In addition, the Bayesian approach can deal with multiple endpoints more easily than the classic approach. For example, if decision makers wish to examine mortality and cost jointly, the Bayesian method can report the probability that a treatment achieves at least 2% mortality reduction and less than $20,000 increase in costs. In conclusion, probabilities computed from the Bayesian approach provide more relevant information to decision makers and are easier to interpret.     

Multiscale analysis of neural spike trains

This paper studies the multiscale analysis of neural spike trains, through both graphical and Poisson process approaches. We introduce the interspike interval plot, which simultaneously visualizes characteristics of neural spiking activity at different time scales. Using an inhomogeneous Poisson process framework, we discuss multiscale estimates of the intensity functions of spike trains. We also introduce the windowing effect for two multiscale methods. Using quasi‐likelihood, we develop bootstrap confidence intervals for the multiscale intensity function. We provide a cross‐validation scheme, to choose the tuning parameters, and study its unbiasedness. Studying the relationship between the spike rate and the stimulus signal, we observe that adjusting for the first spike latency is important in cross‐validation. We show, through examples, that the correlation between spike trains and spike count variability can be multiscale phenomena. Furthermore, we address the modeling of the periodicity of the spike trains caused by a stimulus signal or by brain rhythms. Within the multiscale framework, we introduce intensity functions for spike trains with multiplicative and additive periodic components. Analyzing a dataset from the retinogeniculate synapse, we compare the fit of these models with the Bayesian adaptive regression splines method and discuss the limitations of the methodology. Computational efficiency, which is usually a challenge in the analysis of spike trains, is one of the highlights of these new models. In an example, we show that the reconstruction quality of a complex intensity function demonstrates the ability of the multiscale methodology to crack the neural code. Copyright © 2013 John Wiley & Sons, Ltd.     

Bayesian noninferiority test for 2 binomial probabilities as the extension of Fisher exact test

Noninferiority trials have recently gained importance for the clinical trials of drugs and medical devices. In these trials, most statistical methods have been used from a frequentist perspective, and historical data have been used only for the specification of the noninferiority margin Δ>0. In contrast, Bayesian methods, which have been studied recently are advantageous in that they can use historical data to specify prior distributions and are expected to enable more efficient decision making than frequentist methods by borrowing information from historical trials. In the case of noninferiority trials for response probabilities π ₁,π ₂, Bayesian methods evaluate the posterior probability of H ₁:π ₁>π ₂?Δ being true. To numerically calculate such posterior probability, complicated Appell hypergeometric function or approximation methods are used. Further, the theoretical relationship between Bayesian and frequentist methods is unclear. In this work, we give the exact expression of the posterior probability of the noninferiority under some mild conditions and propose the Bayesian noninferiority test framework which can flexibly incorporate historical data by using the conditional power prior. Further, we show the relationship between Bayesian posterior probability and the P value of the Fisher exact test. From this relationship, our method can be interpreted as the Bayesian noninferior extension of the Fisher exact test, and we can treat superiority and noninferiority in the same framework. Our method is illustrated through Monte Carlo simulations to evaluate the operating characteristics, the application to the real HIV clinical trial data, and the sample size calculation using historical data.     

Multiple indicator hidden Markov model with an application to medical utilization data

Monthly counts of medical visits across several years for persons identified to have alcoholism problems are modeled using two-state hidden Markov models (HMM) in order to describe the effect of alcoholism treatment on the likelihood of persons to be in a 'healthy' or 'unhealthy' state. The medical visits can be classified into different types leading to multivariate counts of medical visits each month. A multiple indicator HMM is introduced, which simultaneously fits the multivariate Poisson counts by assuming a shared hidden state underlying all of them. The multiple indicator HMM borrows information across different types of medical encounters. A univariate HMM based on the total count across types of medical visits each month is also considered. Comparisons between the multiple indicator HMM and the total count HMM are made, as well as comparisons with more traditional longitudinal models that directly model the counts. A Bayesian framework is used for the estimation of the HMM and implementation is in Winbugs.     

Bayesian latent variable modelling in studies of air pollution and health

This paper describes the use of Bayesian latent variable models in the context of studies investigating the short‐term effects of air pollution on health. Traditional Poisson or quasi‐likelihood regression models used in this area assume that consecutive outcomes are independent (although the latter allows for overdispersion), which in many studies may be an untenable assumption as temporal correlation is to be expected. We compare this traditional approach with two Bayesian latent process models, which acknowledge the possibility of short‐term autocorrelation. These include an autoregressive model that has previously been used in air pollution studies and an alternative based on a moving average structure that we describe here. A simulation study assesses the performance of these models when there are different forms of autocorrelation in the data. Although estimated risks are largely unbiased, the results show that assuming independence can produce confidence intervals that are too narrow. Failing to account for the additional uncertainty which may be associated with (positive) correlation can result in confidence/credible intervals being too narrow and thus lead to incorrect conclusions being made about the significance of estimated risks. The methods are illustrated within a case study of the effects of short‐term exposure to air pollution on respiratory mortality in the elderly in London, between 1997 and 2003. Copyright © 2010 John Wiley & Sons, Ltd.     

Advanced methods in meta-analysis: multivariate approach and meta-regression

This tutorial on advanced statistical methods for meta-analysis can be seen as a sequel to the recent Tutorial in Biostatistics on meta-analysis by Normand, which focused on elementary methods. Within the framework of the general linear mixed model using approximate likelihood, we discuss methods to analyse univariate as well as bivariate treatment effects in meta-analyses as well as meta-regression methods. Several extensions of the models are discussed, like exact likelihood, non-normal mixtures and multiple endpoints. We end with a discussion about the use of Bayesian methods in meta-analysis. All methods are illustrated by a meta-analysis concerning the efficacy of BCG vaccine against tuberculosis. All analyses that use approximate likelihood can be carried out by standard software. We demonstrate how the models can be fitted using SAS Proc Mixed.     

Adaptive Bayesian variable clustering via structural learning of breast cancer data

The clustering of proteins is of interest in cancer cell biology. This article proposes a hierarchical Bayesian model for protein (variable) clustering hinging on correlation structure. Starting from a multivariate normal likelihood, we enforce the clustering through prior modeling using angle-based unconstrained reparameterization of correlations and assume a truncated Poisson distribution (to penalize a large number of clusters) as prior on the number of clusters. The posterior distributions of the parameters are not in explicit form and we use a reversible jump Markov chain Monte Carlo based technique is used to simulate the parameters from the posteriors. The end products of the proposed method are estimated cluster configuration of the proteins (variables) along with the number of clusters. The Bayesian method is flexible enough to cluster the proteins as well as estimate the number of clusters. The performance of the proposed method has been substantiated with extensive simulation studies and one protein expression data with a hereditary disposition in breast cancer where the proteins are coming from different pathways.     

Bayesian spatial modeling of HIV mortality via zero‐inflated Poisson models

In this paper, we investigate the effects of poverty and inequality on the number of HIV‐related deaths in 62 New York counties via Bayesian zero‐inflated Poisson models that exhibit spatial dependence. We quantify inequality via the Theil index and poverty via the ratios of two Census 2000 variables, the number of people under the poverty line and the number of people for whom poverty status is determined, in each Zip Code Tabulation Area. The purpose of this study was to investigate the effects of inequality and poverty in addition to spatial dependence between neighboring regions on HIV mortality rate, which can lead to improved health resource allocation decisions. In modeling county‐specific HIV counts, we propose Bayesian zero‐inflated Poisson models whose rates are functions of both covariate and spatial/random effects. To show how the proposed models work, we used three different publicly available data sets: TIGER Shapefiles, Census 2000, and mortality index files. In addition, we introduce parameter estimation issues of Bayesian zero‐inflated Poisson models and discuss MCMC method implications. Copyright © 2012 John Wiley & Sons, Ltd.     

Permutation and Bayesian tests for testing random effects in linear mixed-effects models

In many applications of linear mixed-effects models to longitudinal and multilevel data especially from medical studies, it is of interest to test for the need of random effects in the model. It is known that classical tests such as the likelihood ratio, Wald, and score tests are not suitable for testing random effects because they suffer from testing on the boundary of the parameter space. Instead, permutation and bootstrap tests as well as Bayesian tests, which do not rely on the asymptotic distributions, avoid issues with the boundary of the parameter space. In this paper, we first develop a permutation test based on the likelihood ratio test statistic, which can be easily used for testing multiple random effects and any subset of them in linear mixed-effects models. The proposed permutation test would be an extension to two existing permutation tests. We then aim to compare permutation tests and Bayesian tests for random effects to find out which test is more powerful under which situation. Nothing is known about this in the literature, although this is an important practical problem due to the usefulness of both methods in tackling the challenges with testing random effects. For this, we consider a Bayesian test developed using Bayes factors, where we also propose a new alternative computation for this Bayesian test to avoid some computational issue it encounters in testing multiple random effects. Extensive simulations and a real data analysis are used for evaluation of the proposed permutation test and its comparison with the Bayesian test. We find that both tests perform well, albeit the permutation test with the likelihood ratio statistic tends to provide a relatively higher power when testing multiple random effects.     

Modulation models for seasonal time series and incidence tables

We model monthly disease counts on an age-time grid using the two-dimensional varying-coefficient Poisson regression. Since the marginal profile of counts shows a very strong and varying annual cycle, sine and cosine regressors model periodicity, but their coefficients are allowed to vary smoothly over the age and time plane. The coefficient surfaces are estimated using a relatively large tensor product B-spline basis. Smoothness is tuned using difference penalties on the rows and columns of the tensor product coefficients. Heavy over-dispersion occurs, making it impossible to use Akaike's information criterion or Bayesian information criterion based on a Poisson likelihood. It is handled by selective weighting of part of the data and by the use of extended quasi-likelihood. Very efficient computation is achieved with fast array algorithms. The model is applied to monthly deaths due to respiratory diseases, for U.S. females during 1959-1998 and for ages 51-100.