Distribution Of Random Effects Research Articles

Prediction of Complex Traits: Robust Alternatives to Best Linear Unbiased Prediction.

A widely used method for prediction of complex traits in animal and plant breeding is “genomic best linear unbiased prediction” (GBLUP). In a quantitative genetics setting, BLUP is a linear regression of phenotypes on a pedigree or on a genomic relationship matrix, depending on the type of input information available. Normality of the distributions of random effects and of model residuals is not required for BLUP but a Gaussian assumption is made implicitly. A potential downside is that Gaussian linear regressions are sensitive to outliers, genetic or environmental in origin. We present simple (relative to a fully Bayesian analysis) to implement robust alternatives to BLUP using a linear model with residual t or Laplace distributions instead of a Gaussian one, and evaluate the methods with milk yield records on Italian Brown Swiss cattle, grain yield data in inbred wheat lines, and using three traits measured on accessions of Arabidopsis thaliana. The methods do not use Markov chain Monte Carlo sampling and model hyper-parameters, viewed here as regularization “knobs,” are tuned via some cross-validation. Uncertainty of predictions are evaluated by employing bootstrapping or by random reconstruction of training and testing sets. It was found (e.g., test-day milk yield in cows, flowering time and FRIGIDA expression in Arabidopsis) that the best predictions were often those obtained with the robust methods. The results obtained are encouraging and stimulate further investigation and generalization.

Assessing the robustness of estimators when fitting Poisson inverse Gaussian models

The generalized linear mixed model (GLMM) extends classical regression analysis to non-normal, correlated response data. Because inference for GLMMs can be computationally difficult, simplifying distributional assumptions are often made. We focus on the robustness of estimators when a main component of the model, the random effects distribution, is misspecified. Results for the maximum likelihood estimators of the Poisson inverse Gaussian model are presented.

The Impact of Misspecified Random Effect Distribution in a Weibull Regression Mixed Model

Mixed models are useful tools for analyzing clustered and longitudinal data. These models assume that random effects are normally distributed. However, this may be unrealistic or restrictive when representing information of the data. Several papers have been published to quantify the impacts of misspecification of the shape of the random effects in mixed models. Notably, these studies primarily concentrated their efforts on models with response variables that have normal, logistic and Poisson distributions, and the results were not conclusive. As such, we investigated the misspecification of the shape of the random effects in a Weibull regression mixed model with random intercepts in the two parameters of the Weibull distribution. Through an extensive simulation study considering six random effect distributions and assuming normality for the random effects in the estimation procedure, we found an impact of misspecification on the estimations of the fixed effects associated with the second parameter σ of the Weibull distribution. Additionally, the variance components of the model were also affected by the misspecification.

Correlation-Based Meta-Analytic Structural Equation Modeling: Effects of Parameter Covariance on Point and Interval Estimates

More and more researchers use meta-analysis to conduct multivariate analysis to summarize previous findings. In the correlation-based meta-analytic structural equation modeling (cMASEM), the average sample correlation matrix is used to estimate the average population model. Using a simple mediation model, we illustrated that random effects covariation in population parameters can theoretically bias the path coefficient estimates and lead to nonnormal random effects distribution of the correlations. We developed an R function for researchers to examine by simulation the impact of random effects in other models. We then reanalyzed two real data sets and conducted a simulation study to examine the magnitude of the impact on realistic situations. Simulation results suggest parameter bias is typically negligible (less than .02), parameter bias and root mean square error do not differ across methods, 95% confident intervals are sometimes more accurate for the two-stage structural equation modeling approach with a diagonal random effects model, and power is sometimes higher for the traditional Viswesvaran-Ones approach. Given the increasing popularity of cMASEM in organizational research, these simulation results form the basis for us to make several recommendations on its application.

On the formal specification of sum-zero constrained intrinsic conditional autoregressive models

We propose a formal specification for sum-zero constrained intrinsic conditional autoregressive (ICAR) models. Our specification first projects a vector of proper conditional autoregressive spatial random effects onto a subspace where the projected vector is constrained to sum to zero, and after that takes the limit when the proper conditional autoregressive model approaches the ICAR model. As a result, we show that the sum-zero constrained ICAR model has a singular Gaussian distribution with zero mean vector and a unique covariance matrix. Previously, sum-zero constraints have typically been imposed on the vector of spatial random effects in ICAR models within a Markov chain Monte Carlo (MCMC) algorithm in what is known as centering-on-the-fly. This mathematically informal way to impose the sum-zero constraint obscures the actual joint density of the spatial random effects. By contrast, the present work elucidates a unique distribution for ICAR random effects. The explicit expressions for the resulting unique covariance matrix and density function are useful for the development of Bayesian methodology in spatial statistics which will be useful to practitioners. We illustrate the practical relevance of our results by using Bayesian model selection to jointly assess both spatial dependence and fixed effects.

Bayesian Semiparametric Modelling of Contraceptive Behaviour in India Via Sequential Logistic Regressions

SummaryFamily planning has been characterized by highly different strategic programmes in India, including method-specific contraceptive targets, coercive sterilization and more recent target-free approaches. These major changes in family planning policies over time have motivated considerable interest towards assessing the effectiveness of the different planning programmes. Current studies mainly focus on the factors driving the choice among specific subsets of contraceptives, such as a preference for alternative methods other than sterilization. Although this restricted focus produces key insights, it fails to provide a global overview of the different policies, and of the determinants underlying the choices from the entire range of contraceptive methods. Motivated by this consideration, we propose a Bayesian semiparametric model relying on a reparameterization of the multinomial probability mass function via a set of conditional Bernoulli choices. This binary decision tree is defined to be consistent with the current family planning policies in India, and coherent with a reasonable process characterizing the choice between increasingly nested subsets of contraceptive methods. The model allows a subset of covariates to enter the predictor via Bayesian penalized splines and exploits mixture models to represent uncertainty in the distribution of the state-specific random effects flexibly. This combination of flexible and careful reparameterizations allows a broader and interpretable overview of the policies and contraceptive preferences in India.

Semiparametric regression analysis of clustered survival data with semi-competing risks

Analysis of semi-competing risks data is becoming increasingly important in medical research in which a subject may experience both nonterminal and terminal events, and the time to the intermediate nonterminal event (e.g. onset of a disease) is subject to dependent censoring by the terminal event (e.g. death) but not vice versa. Typically, both two types of events are dependent. In many applications, subjects may also be nested within clusters, such as patients in a multi-center study, leading to possible association among event times due to unobserved shared factors across subjects. To incorporate dependency within clusters and association between two types of event times, we propose a new flexible semiparametric modeling framework where a copula model is employed for the joint distribution of the nonterminal and terminal events, and their marginal distributions are modeled by Cox proportional hazards models with random effects. A nonparametric maximum likelihood estimation procedure is developed and implemented through a Monte Carlo EM algorithm. The proposed estimator is also shown to enjoy desirable asymptotic properties. Results from extensive simulation studies indicate that the proposed method performs very well in finite samples and is especially robust against misspecification of the random effects distribution. We further illustrate the practical utility of the method by analyzing data from a multi-institutional study of breast cancer.

Toward A Diagnostic Toolkit for Linear Models with Gaussian-Process Distributed Random Effects

Gaussian processes (GPs) are widely used as distributions of random effects in linear mixed models, which are fit using the restricted likelihood or the closely related Bayesian analysis. This article addresses two problems. First, we propose tools for understanding how data determine estimates in these models, using a spectral basis approximation to the GP under which the restricted likelihood is formally identical to the likelihood for a gamma-errors GLM with identity link. Second, to examine the data's support for a covariate and to understand how adding that covariate moves variation in the outcome y out of the GP and error parts of the fit, we apply a linear-model diagnostic, the added variable plot (AVP), both to the original observations and to projections of the data onto the spectral basis functions. The spectral- and observation-domain AVPs estimate the same coefficient for a covariate but emphasize low- and high-frequency data features respectively and thus highlight the covariate's effect on the GP and error parts of the fit, respectively. The spectral approximation applies to data observed on a regular grid; for data observed at irregular locations, we propose smoothing the data to a grid before applying our methods. The methods are illustrated using the forest-biomass data of Finley et al.(2008).

PFIM 4.0, an extended R program for design evaluation and optimization in nonlinear mixed-effect models

Background and ObjectiveNonlinear mixed-effect models (NLMEMs) are increasingly used for the analysis of longitudinal studies during drug development. When designing these studies, the expected Fisher information matrix (FIM) can be used instead of performing time-consuming clinical trial simulations. The function PFIM is the first tool for design evaluation and optimization that has been developed in R. In this article, we present an extended version, PFIM 4.0, which includes several new features. MethodsCompared with version 3.0, PFIM 4.0 includes a more complete pharmacokinetic/pharmacodynamic library of models and accommodates models including additional random effects for inter-occasion variability as well as discrete covariates. A new input method has been added to specify user-defined models through an R function. Optimization can be performed assuming some fixed parameters or some fixed sampling times. New outputs have been added regarding the FIM such as eigenvalues, conditional numbers, and the option of saving the matrix obtained after evaluation or optimization. Previously obtained results, which are summarized in a FIM, can be taken into account in evaluation or optimization of one-group protocols. This feature enables the use of PFIM for adaptive designs. The Bayesian individual FIM has been implemented, taking into account a priori distribution of random effects. Designs for maximum a posteriori Bayesian estimation of individual parameters can now be evaluated or optimized and the predicted shrinkage is also reported. It is also possible to visualize the graphs of the model and the sensitivity functions without performing evaluation or optimization. ResultsThe usefulness of these approaches and the simplicity of use of PFIM 4.0 are illustrated by two examples: (i) an example of designing a population pharmacokinetic study accounting for previous results, which highlights the advantage of adaptive designs; (ii) an example of Bayesian individual design optimization for a pharmacodynamic study, showing that the Bayesian individual FIM can be a useful tool in therapeutic drug monitoring, allowing efficient prediction of estimation precision and shrinkage for individual parameters. ConclusionPFIM 4.0 is a useful tool for design evaluation and optimization of longitudinal studies in pharmacometrics and is freely available at http://www.pfim.biostat.fr.

Conditional mix-GEE models for longitudinal data with unspecified random-effects distributions

ABSTRACTIn the longitudinal studies, the mixture generalized estimation equation (mix-GEE) was proposed to improve the efficiency of the fixed-effects estimator for addressing the working correlation structure misspecification. When the subject-specific effect is one of interests, mixed-effects models were widely used to analyze longitudinal data. However, most of the existing approaches assume a normal distribution for the random effects, and this could affect the efficiency of the fixed-effects estimator. In this article, a conditional mixture generalized estimating equation (cmix-GEE) approach based on the advantage of mix-GEE and conditional quadratic inference function (CQIF) method is developed. The advantage of our new approach is that it does not require the normality assumption for random effects and can accommodate the serial correlation between observations within the same cluster. The feature of our proposed approach is that the estimators of the regression parameters are more efficient than CQIF even if the working correlation structure is not correctly specified. In addition, according to the estimates of some mixture proportions, the true working correlation matrix can be identified. We establish the asymptotic results for the fixed-effects parameter estimators. Simulation studies were conducted to evaluate our proposed method.

Heteroscedastic nested error regression models with variance functions

The article considers a nested error regression model with heteroscedastic variance functions for analyzing clustered data, where the normality for the underlying distributions is not assumed. Classical methods in normal nested error regression models with homogenous variances are extended in the two directions: heterogeneous variance functions for error terms and non-normal distributions for random effects and error terms. Consistent estimators for model parameters are suggested, and second-order approximations of their biases and variances are derived. The mean squared errors of the empirical best linear unbiased predictors are expressed explicitly to second-order. Second-order unbiased estimators of the mean squared errors are provided analytically in closed forms. The proposed model and the resulting procedures are numerically investigated through simulation and empirical studies.

A More Robust Random Effects Model for Disease Mapping

Disease mapping studies have found wide applications within geographical epidemiology and public health and are typically analysed within a Bayesian hierarchical model formulation. The most popular disease mapping model is the Besag-York-Molli´e model. A distinguishing feature of this model is the use of two sets of random effects: one spatially structured to model spatial autocorrelation and the other spatially unstructured to describe residual unstructured heterogeneity. Very often the spatially unstructured random effect is assumed to be normally distributed. Under practical situations, this normality assumption is found to be over restrictive. In this study, we investigate a more robust spatially unstructured random effect distribution by considering the Inverse Gaussian (IG) distribution in the disease mapping problem. The distribution has the normal distribution as special case. The inferences under this model are carried out within a bayesian hierarchical model formulation implemented in WinBUGS. The IG distribution is introduced in WinBUGS using zero tricks. The usefulness of the proposed model is investigated with a simulation study and applied in real data; mapping HIV in Kenya. In this work we showed that the IG distribution can produce better results when the normality assumption is violated due to the skewness of the data. For the case of data in which the random effects are truly normal, the IG distribution adjusts to a normal distribution as dictated by the data itself. On the other hand, the spatially structured random effect is normally modelled using the intrinsic conditional autoregressive (iCAR) prior. This prior is improper and has the undesirable large scale property of leading to a negative pairwise correlation for regions located further apart. In addition, the BYM model presents some identifiability problems of the spatial and non-spatial effects. In this work, we consider Leroux CAR (named lCAR hereafter) prior, a less widely used prior in disease mapping, as the prior distribution for the spatially structured random effects.

Flexible linear mixed models with improper priors for longitudinal and survival data

We propose a Bayesian approach using improper priors for hierarchical linear mixed models with flexible random effects and residual error distributions. The error distribution is modelled using scale mixtures of normals, which can capture tails heavier than those of the normal distribution. This generalisation is useful to produce models that are robust to the presence of outliers. The case of asymmetric residual errors is also studied. We present general results for the propriety of the posterior that also cover cases with censored observations, allowing for the use of these models in the contexts of popular longitudinal and survival analyses. We consider the use of copulas with flexible marginals for modelling the dependence between the random effects, but our results cover the use of any random effects distribution. Thus, our paper provides a formal justification for Bayesian inference in a very wide class of models (covering virtually all of the literature) under attractive prior structures that limit the amount of required user elicitation.

Bias and Precision of Alternate Estimators in Meta-Analysis: Benefits of Blending Schmidt-Hunter and Hedges Approaches

We describe a new estimator (labeled Morris) for meta-analysis. The Morris estimator combines elements of both the Schmidt-Hunter and Hedges estimators. The new estimator is compared to (a) the Schmidt-Hunter estimator, (b) the Schmidt-Hunter estimator with variance correction for the number of studies (“ k correction”), (c) the Hedges random-effects estimator, and (d) the Bonett unit weights estimator in a Monte Carlo simulation. The simulation was designed to represent realistic conditions faced by researchers, including population random-effects distributions, numbers of studies, and skewed sample size distributions. The simulation was used to evaluate the estimators with respect to bias, coverage of the 95% confidence interval of the mean, and root mean square error of estimates of the population mean. We also evaluated the quality of credibility intervals. Overall, the new estimator provides better coverage and slightly better credibility values than other commonly used methods. Thus it has advantages of both commonly used approaches without the apparent disadvantages. The new estimator can be implemented easily with existing software; software used in the study is available online, and an example is included in the appendix in the Supplemental Material available online.

On the Use of Mixed Markov Models for Intensive Longitudinal Data

ABSTRACTMarkov modeling presents an attractive analytical framework for researchers who are interested in state-switching processes occurring within a person, dyad, family, group, or other system over time. Markov modeling is flexible and can be used with various types of data to study observed or latent state-switching processes, and can include subject-specific random effects to account for heterogeneity. We focus on the application of mixed Markov models to intensive longitudinal data sets in psychology, which are becoming ever more common and provide a rich description of each subject’s process. We examine how specifications of a Markov model change when continuous random effect distributions are included, and how mixed Markov models can be used in the intensive longitudinal research context. Advantages of Bayesian estimation are discussed and the approach is illustrated by two empirical applications.

Adaptive Hierarchical Bayes Estimation of Small Area Proportions

Unit-level logistic regression models with mixed effects have been used for estimating small area proportions in the literature. Normality is commonly assumed for the random effects. Nonetheless, real data often show significant departures from normality assumptions of the random effects. To reduce the risk of model misspecification, we propose an adaptive hierarchical Bayes estimation approach in which the distribution of the random effect is chosen adaptively from the exponential power class of probability distributions. The richness of the exponential power class ensures the robustness of our hierarchical Bayes approach against departure from normality. We demonstrate the robustness of our proposed model using both simulated and real data. The results suggest that the proposed model works reasonably well to incorporate potential kurtosis of the random effects distribution.

A random effects meta-analysis model with Box-Cox transformation

BackgroundIn a random effects meta-analysis model, true treatment effects for each study are routinely assumed to follow a normal distribution. However, normality is a restrictive assumption and the misspecification of the random effects distribution may result in a misleading estimate of overall mean for the treatment effect, an inappropriate quantification of heterogeneity across studies and a wrongly symmetric prediction interval.MethodsWe focus on problems caused by an inappropriate normality assumption of the random effects distribution, and propose a novel random effects meta-analysis model where a Box-Cox transformation is applied to the observed treatment effect estimates. The proposed model aims to normalise an overall distribution of observed treatment effect estimates, which is sum of the within-study sampling distributions and the random effects distribution. When sampling distributions are approximately normal, non-normality in the overall distribution will be mainly due to the random effects distribution, especially when the between-study variation is large relative to the within-study variation. The Box-Cox transformation addresses this flexibly according to the observed departure from normality. We use a Bayesian approach for estimating parameters in the proposed model, and suggest summarising the meta-analysis results by an overall median, an interquartile range and a prediction interval. The model can be applied for any kind of variables once the treatment effect estimate is defined from the variable.ResultsA simulation study suggested that when the overall distribution of treatment effect estimates are skewed, the overall mean and conventional I2 from the normal random effects model could be inappropriate summaries, and the proposed model helped reduce this issue. We illustrated the proposed model using two examples, which revealed some important differences on summary results, heterogeneity measures and prediction intervals from the normal random effects model.ConclusionsThe random effects meta-analysis with the Box-Cox transformation may be an important tool for examining robustness of traditional meta-analysis results against skewness on the observed treatment effect estimates. Further critical evaluation of the method is needed.

ROBUST MIXED EFFECTS MODEL FOR CLUSTERED FAILURE TIME DATA: APPLICATION TO HUNTINGTON'S DISEASE EVENT MEASURES.

An important goal in clinical and statistical research is properly modeling the distribution for clustered failure times which have a natural intraclass dependency and are subject to censoring. We handle these challenges with a novel approach that does not impose restrictive modeling or distributional assumptions. Using a logit transformation, we relate the distribution for clustered failure times to covariates and a random, subject-specific effect. The covariates are modeled with unknown functional forms, and the random effect may depend on the covariates and have an unknown and unspecified distribution. We introduce pseudovalues to handle censoring and splines for functional covariate effects, and frame the problem into fitting an additive logistic mixed effects model. Unlike existing approaches for fitting such models, we develop semiparametric techniques that estimate the functional model parameters without specifying or estimating the random effect distribution. We show both theoretically and empirically that the resulting estimators are consistent for any choice of random effect distribution and any dependency structure between the random effect and covariates. Last, we illustrate the method's utility in an application to a Huntington's disease study where our method provides new insights into differences between motor and cognitive impairment event times in at-risk subjects.

Random effects models for identifying the most harmful medication errors in a large, voluntary reporting database

Medical errors are a major source of preventable morbidity, mortality and healthcare costs. Voluntary reporting systems are useful data sources that collect detailed information on the circumstances of medical errors occurring in hospitals. Identifying the characteristics of errors that frequently result in patient harm when they occur would allow investigators to prioritize among the many sources of potential errors and design targeted prevention strategies. In this paper, we use data from MEDMARX, a large anonymous and voluntary reporting system for medication errors, to identify the combinations of error characteristics that are more likely to result in harm. To this end, we consider a Bayesian hierarchical model with crossed random effects and a flexible specification of the random effects distribution. We then provide a ranking of the errors using optimal Bayesian ranking based on their probability of harm. The use of optimal Bayesian ranking accounts for the varying amount of uncertainty across the random effects estimates. Finally, we examine the sensitivity of results to different specifications of the random effects distributions. The utility of flexible random effects assumptions is illustrated by empirically comparing results under several choices. We found that errors caused by mistakes in reconciling a patient’s current medication list with the medications prescribed at hospital discharge have an estimated 10.5% probability of harm. These errors had the highest rate of harm of errors that occur during the prescribing stage of medication use. In addition, we found that the results are sensitive to the random effects distribution used in estimation. Thus, an approach that explores this sensitivity is important for accurately comparing the relative harm across errors.

Two-Part and Related Regression Models for Longitudinal Data.

Statistical models that involve a two-part mixture distribution are applicable in a variety of situations. Frequently, the two parts are a model for the binary response variable and a model for the outcome variable that is conditioned on the binary response. Two common examples are zero-inflated or hurdle models for count data and two-part models for semicontinuous data. Recently, there has been particular interest in the use of these models for the analysis of repeated measures of an outcome variable over time. The aim of this review is to consider motivations for the use of such models in this context and to highlight the central issues that arise with their use. We examine two-part models for semicontinuous and zero-heavy count data, and we also consider models for count data with a two-part random effects distribution.