A review on longitudinal data analysis with random forest.

In longitudinal studies variables are measured repeatedly over time, leading to clustered and correlated observations. If the goal of the study is to develop prediction models, machine learning approaches such as the powerful random forest (RF) are often promising alternatives to standard statistical methods, especially in the context of high-dimensional data. In this paper, we review extensions of the standard RF method for the purpose of longitudinal data analysis. Extension methods are categorized according to the data structures for which they are designed. We consider both univariate and multivariate response longitudinal data and further categorize the repeated measurements according to whether the time effect is relevant. Even though most extensions are proposed for low-dimensional data, some can be applied to high-dimensional data. Information of available software implementations of the reviewed extensions is also given. We conclude with discussions on the limitations of our review and some future research directions.

Semiparametric marginal regression for clustered competing risks data with missing cause of failure.

Clustered competing risks data are commonly encountered in multicenter studies. The analysis of such data is often complicated due to informative cluster size (ICS), a situation where the outcomes under study are associated with the size of the cluster. In addition, the cause of failure is frequently incompletely observed in real-world settings. To the best of our knowledge, there is no methodology for population-averaged analysis with clustered competing risks data with an ICS and missing causes of failure. To address this problem, we consider the semiparametric marginal proportional cause-specific hazards model and propose a maximum partial pseudolikelihood estimator under a missing at random assumption. To make the latter assumption more plausible in practice, we allow for auxiliary variables that may be related to the probability of missingness. The proposed method does not impose assumptions regarding the within-cluster dependence and allows for ICS. The asymptotic properties of the proposed estimators for both regression coefficients and infinite-dimensional parameters, such as the marginal cumulative incidence functions, are rigorously established. Simulation studies show that the proposed method performs well and that methods that ignore the within-cluster dependence and the ICS lead to invalid inferences. The proposed method is applied to competing risks data from a large multicenter HIV study in sub-Saharan Africa where a significant portion of causes of failure is missing.

Two-Stage TMLE to reduce bias and improve efficiency in cluster randomized trials.

Cluster randomized trials (CRTs) randomly assign an intervention to groups of individuals (e.g., clinics or communities) and measure outcomes on individuals in those groups. While offering many advantages, this experimental design introduces challenges that are only partially addressed by existing analytic approaches. First, outcomes are often missing for some individuals within clusters. Failing to appropriately adjust for differential outcome measurement can result in biased estimates and inference. Second, CRTs often randomize limited numbers of clusters, resulting in chance imbalances on baseline outcome predictors between arms. Failing to adaptively adjust for these imbalances and other predictive covariates can result in efficiency losses. To address these methodological gaps, we propose and evaluate a novel two-stage targeted minimum loss-based estimator to adjust for baseline covariates in a manner that optimizes precision, after controlling for baseline and postbaseline causes of missing outcomes. Finite sample simulations illustrate that our approach can nearly eliminate bias due to differential outcome measurement, while existing CRT estimators yield misleading results and inferences. Application to real data from the SEARCH community randomized trial demonstrates the gains in efficiency afforded through adaptive adjustment for baseline covariates, after controlling for missingness on individual-level outcomes.

Soft classification and regression analysis of audiometric phenotypes of age-related hearing loss.

Age-related hearing loss has a complex etiology. Researchers have made efforts to classify relevant audiometric phenotypes, aiming to enhance medical interventions and improve hearing health. We leveraged existing pattern analyses of age-related hearing loss and implemented the phenotype classification via quadratic discriminant analysis (QDA). We herein propose a method for analyzing the exposure effects on the soft classification probabilities of the phenotypes via estimating equations. Under reasonable assumptions, the estimating equations are unbiased and lead to consistent estimators. The resulting estimator had good finite sample performances in simulation studies. As an illustrative example, we applied our proposed methods to assess the association between a dietary intake pattern, assessed as adherence scores for the dietary approaches to stop hypertension diet calculated using validated food-frequency questionnaires, and audiometric phenotypes (older-normal, metabolic, sensory, and metabolic plus sensory), determined based on data obtained in the Nurses' Health Study II Conservation of Hearing Study, the Audiology Assessment Arm. Our findings suggested that participants with a more healthful dietary pattern were less likely to develop the metabolic plus sensory phenotype of age-related hearing loss.

Approximate balancing weights for clustered observational study designs.

In a clustered observational study, a treatment is assigned to groups and all units within the group are exposed to the treatment. We develop a new method for statistical adjustment in clustered observational studies using approximate balancing weights, a generalization of inverse propensity score weights that solve a convex optimization problem to find a set of weights that directly minimize a measure of covariate imbalance, subject to an additional penalty on the variance of the weights. We tailor the approximate balancing weights optimization problem to the clustered observational study setting by deriving an upper bound on the mean square error and finding weights that minimize this upper bound, linking the level of covariate balance to a bound on the bias. We implement the procedure by specializing the bound to a random cluster-level effects model, leading to a variance penalty that incorporates the signal-to-noise ratio and penalizes the weight on individuals and the total weight on groups differently according to the the intra-class correlation.

The effect of number of clusters and magnitude of within-cluster homogeneity in outcomes on the performance of four variance estimators for a marginal multivariable Cox regression model fit to clustered data in the context of observational research.

Researchers often estimate the association between the hazard of a time-to-event outcome and the characteristics of individuals and the clusters in which individuals are nested. Lin and Wei's robust variance estimator is often used with a Cox regression model fit to clustered data. Recently, alternative variance estimators have been proposed: the Fay-Graubard estimator, the Kauermann-Carroll estimator, and the Mancl-DeRouen estimator. Using Monte Carlo simulations, we found that, when fitting a marginal Cox regression model with both individual-level and cluster-level covariates: (i) in the presence of weak to moderate within-cluster homogeneity of outcomes, the Lin-Wei variance estimator can result in estimates of the SE with moderate bias when the number of clusters is fewer than 20-30, while in the presence of strong within-cluster homogeneity, it can result in biased estimation even when the number of clusters is as large as 100; (ii) when the number of clusters was less than approximately 20, the Fay-Graubard variance estimator tended to result in estimates of SE with the lowest bias; (iii) when the number of clusters exceeded approximately 20, the Mancl-DeRouen estimator tended to result in estimated standard errors with the lowest bias; (iv) the Mancl-DeRouen estimator used with a t-distribution tended to result in 95% confidence that had the best performance of the estimators; (v) when the magnitude of within-cluster homogeneity in outcomes was strong or very strong, all methods resulted in confidence intervals with lower than advertised coverage rates even when the number of clusters was very large.

Trivariate Joint Modeling for Family Data with Longitudinal Counts, Recurrent Events and a Terminal Event with Application to Lynch Syndrome.

Trivariate joint modeling for longitudinal count data, recurrent events, and a terminal event for family data has increased interest in medical studies. For example, families with Lynch syndrome (LS) are at high risk of developing colorectal cancer (CRC), where the number of polyps and the frequency of colonoscopy screening visits are highly associated with the risk of CRC among individuals and families. To assess how screening visits influence polyp detection, which in turn influences time to CRC, we propose a clustered trivariate joint model. The proposed model facilitates longitudinal count data that are zero-inflated and over-dispersed and invokes individual-specific and family-specific random effects to account for dependence among individuals and families. We formulate our proposed model as a latent Gaussian model to use the Bayesian estimation approach with the integrated nested Laplace approximation algorithm and evaluate its performance using simulation studies. Our trivariate joint model is applied to a series of 18 families from Newfoundland, with the occurrence of CRC taken as the terminal event, the colonoscopy screening visits as recurrent events, and the number of polyps detected at each visit as zero-inflated count data with overdispersion. We showed that our trivariate model fits better than alternative bivariate models and that the cluster effects should not be ignored when analyzing family data. Finally, the proposed model enables us to quantify heterogeneity across families and individuals in polyp detection and CRC risk, thus helping to identify individuals and families who would benefit from more intensive screening visits.

Marginal structural models for multilevel clustered data.

Marginal structural models (MSMs), which adopt inverse probability treatment weighting in the estimating equations, are powerful tools to estimate the causal effects of time-varying exposures in the presence of time-dependent confounders. Motivated by the Conservation of Hearing Study (CHEARS) Audiology Assessment Arm (AAA) where repeated hearing measurements were clustered by study participants, time, and testing sites, we propose two methods to account for the multilevel correlation structure when fitting the MSMs. The first method directly models the covariance of the repeated outcomes when solving the weighted generalized estimating equations for MSMs, while the second two-stage analysis approach fits cluster-specific MSMs first and then combines the estimated parameters using mixed-effects meta-analysis. Finite sample simulation results suggest that our methods can obtain less biased and more efficient estimates of the parameters by accounting for the multilevel correlation. Moreover, we explore the effects of using fixed- or mixed-effects model to estimate the treatment probability on the parameter estimates of the MSMs in the presence of unmeasured cluster-level confounders. Lastly, we apply our methods to the CHEARS AAA data set, to estimate the causal effects of aspirin use on hearing loss.

Analyzing clustered continuous response variables with ordinal regression models.

Continuous response data are regularly transformed to meet regression modeling assumptions. However, approaches taken to identify the appropriate transformation can be ad hoc and can increase model uncertainty. Further, the resulting transformations often vary across studies leading to difficulties with synthesizing and interpreting results. When a continuous response variable is measured repeatedly within individuals or when continuous responses arise from clusters, analyses have the additional challenge caused by within-individual or within-cluster correlations. We extend a widely used ordinal regression model, the cumulative probability model (CPM), to fit clustered, continuous response data using generalized estimating equations for ordinal responses. With the proposed approach, estimates of marginal model parameters, cumulative distribution functions , expectations, and quantiles conditional on covariates can be obtained without pretransformation of the response data. While computational challenges arise with large numbers of distinct values of the continuous response variable, we propose feasible and computationally efficient approaches to fit CPMs under commonly used working correlation structures. We study finite sample operating characteristics of the estimators via simulation and illustrate their implementation with two data examples. One studies predictors of CD4:CD8 ratios in a cohort living with HIV, and the other investigates the association of a single nucleotide polymorphism and lung function decline in a cohort with early chronic obstructive pulmonary disease.

A nonparametric test of group distributional differences for hierarchically clustered functional data.

Biological sex and gender are critical variables in biomedical research, but are complicated by the presence of sex-specific natural hormone cycles, such as the estrous cycle in female rodents, typically divided into phases. A common feature of these cycles are fluctuating hormone levels that induce sex differences in many behaviors controlled by the electrophysiology of neurons, such as neuronal membrane potential in response to electrical stimulus, typically summarized using a priori defined metrics. In this paper, we propose a method to test for differences in the electrophysiological properties across estrous cycle phase without first defining a metric of interest. We do this by modeling membrane potential data in the frequency domain as realizations of a bivariate process, also depending on the electrical stimulus, by adopting existing methods for longitudinal functional data. We are then able to extract the main features of the bivariate signals through a set of basis function coefficients. We use these coefficients for testing, adapting methods for multivariate data to account for an induced hierarchical structure that is a product of the experimental design. We illustrate the performance of the proposed approach in simulations and then apply the method to experimental data.

Rank intraclass correlation for clustered data.

Clustered data are common in biomedical research. Observations in the same cluster are often more similar to each other than to observations from other clusters. The intraclass correlation coefficient (ICC), first introduced by R. A. Fisher, is frequently used to measure this degree of similarity. However, the ICC is sensitive to extreme values and skewed distributions, and depends on the scale of the data. It is also not applicable to ordered categorical data. We define the rank ICC as a natural extension of Fisher's ICC to the rank scale, and describe its corresponding population parameter. The rank ICC is simply interpreted as the rank correlation between a random pair of observations from the same cluster. We also extend the definition when the underlying distribution has more than two hierarchies. We describe estimation and inference procedures, show the asymptotic properties of our estimator, conduct simulations to evaluate its performance, and illustrate our method in three real data examples with skewed data, count data, and three-level ordered categorical data.

Defining and estimating effects in cluster randomized trials: A methods comparison.

Across research disciplines, cluster randomized trials (CRTs) are commonly implemented to evaluate interventions delivered to groups of participants, such as communities and clinics. Despite advances in the design and analysis of CRTs, several challenges remain. First, there are many possible ways to specify the causal effect of interest (eg, at the individual-level or at the cluster-level). Second, the theoretical and practical performance of common methods for CRT analysis remain poorly understood. Here, we present a general framework to formally define an array of causal effects in terms of summary measures of counterfactual outcomes. Next, we provide a comprehensive overview of CRT estimators, including the t-test, generalized estimating equations (GEE), augmented-GEE, and targeted maximum likelihood estimation (TMLE). Using finite sample simulations, we illustrate the practical performance of these estimators for different causal effects and when, as commonly occurs, there are limited numbers of clusters of different sizes. Finally, our application to data from the Preterm Birth Initiative (PTBi) study demonstrates the real-world impact of varying cluster sizes and targeting effects at the cluster-level or at the individual-level. Specifically, the relative effect of the PTBi intervention was 0.81 at the cluster-level, corresponding to a 19% reduction in outcome incidence, and was 0.66 at the individual-level, corresponding to a 34% reduction in outcome risk. Given its flexibility to estimate a variety of user-specified effects and ability to adaptively adjust for covariates for precision gains while maintaining Type-I error control, we conclude TMLE is a promising tool for CRT analysis.

Multilevel joint frailty model for hierarchically clustered binary and survival data.

Hierarchical data arise when observations are clustered into groups. Multilevel models are practically useful in these settings, but these models are elusive in the context of hierarchical data with mixed multivariate outcomes. In this article, we consider binary and survival outcomes and assume the hierarchical structure is induced by clustering of both outcomes within patients and clustering of patients within hospitals which frequently occur in multicenter studies. We introduce a multilevel joint frailty model that analyzes the outcomes simultaneously to jointly estimate their regression parameters and explicitly model within-patient correlation between the outcomes and within-hospital correlation separately for each outcome. Estimation is facilitated by a computationally efficient residual maximum likelihood method that further predicts cluster-specific frailties for both outcomes and circumvents the formidable challenges induced by multidimensional integration that complicates the underlying likelihood. The performance of the model and estimation procedure is investigated via extensive simulation studies. The practical utility of the model is illustrated through simultaneous modeling of disease-free survival and binary endpoint of platelet recovery in a multicenter allogeneic bone marrow transplantation dataset that motivates this study.

The weighted log-rank tests based on stratified clustered survival data: saddle-point p-values and confidence intervals.

Clinical studies sometimes provide clustered data with censored failure times. A crucial factor of the randomized design that lessens selection bias is the random allocation rule. Given this, the weighted rank tests' p-values for stratified survival clustered sampling based on the random allocation rule are approximated using the double saddle-point approximation technique. For tests of significance and confidence intervals for the treatment effect, this approximation can be utilized. Through simulation experiments, the accuracy of the saddle-point approximation is examined by comparing saddle-point and normal approximations to the exact underlying permutation distribution.

Saddlepoint approximation for weighted log-rank tests based on block truncated binomial design.

Clustered data frequently occur in biomedical research fields and clinical trials. The log-rank tests are widely used for two-independent samples of clustered data tests. The randomized block design and truncated binomial design are used for forcing balance in clinical trials and reducing selection bias. In this paper, survival clustered data are randomized by generalized randomized block, and subsequently clustered data in each block are randomized by truncated binomial design. Consequently, the p-values of the null permutation distribution of log-rank tests for clustered data are approximated via the double saddlepoint approximation method. Comprehensive numerical studies are carried out to assess the accuracy of the saddlepoint approximation. This approximation has a great accuracy over the asymptotic normal approximation.

Sample size calculation for one-armed clinical trials with clustered data and binary outcome.

The formula of Fleiss and Cuzick (1979) to estimate the intraclass correlation coefficient is applied to reduce the task of sample size calculation for clustered data with binary outcome. It is demonstrated that this approach reduces the complexity of sample size calculation to the determination of the null and alternative hypothesis and the formulation of the quantitative influence of the belonging to the same cluster on the therapy success probability.

Improving marginal hazard ratio estimation using quadratic inference functions.

Clustered and multivariate failure time data are commonly encountered in biomedical studies and a marginal regression approach is often employed to identify the potential risk factors of a failure. We consider a semiparametric marginal Cox proportional hazards model for right-censored survival data with potential correlation. We propose to use a quadratic inference function method based on the generalized method of moments to obtain the optimal hazard ratio estimators. The inverse of the working correlation matrix is represented by the linear combination of basis matrices in the context of the estimating equation. We investigate the asymptotic properties of the regression estimators from the proposed method. The optimality of the hazard ratio estimators is discussed. Our simulation study shows that the estimator from the quadratic inference approach is more efficient than those from existing estimating equation methods whether the working correlation structure is correctly specified or not. Finally, we apply the model and the proposed estimation method to analyze a study of tooth loss and have uncovered new insights that were previously inaccessible using existing methods.

Estimation and testing for clustered interval-censored bivariate survival data with application using the semi-parametric version of the Clayton-Oakes model.

The Kaplan-Meier estimator is ubiquitously used to estimate survival probabilities for time-to-event data. It is nonparametric, and thus does not require specification of a survival distribution, but it does assume that the risk set at any time t consists of independent observations. This assumption does not hold for data from paired organ systems such as occur in ophthalmology (eyes) or otolaryngology (ears), or for other types of clustered data. In this article, we estimate marginal survival probabilities in the setting of clustered data, and provide confidence limits for these estimates with intra-cluster correlation accounted for by an interval-censored version of the Clayton-Oakes model. We develop a goodness-of-fit test for general bivariate interval-censored data and apply it to the proposed interval-censored version of the Clayton-Oakes model. We also propose a likelihood ratio test for the comparison of survival distributions between two groups in the setting of clustered data under the assumption of a constant between-group hazard ratio. This methodology can be used both for balanced and unbalanced cluster sizes, and also when the cluster size is informative. We compare our test to the ordinary log rank test and the Lin-Wei (LW) test based on the marginal Cox proportional Hazards model with robust standard errors obtained from the sandwich estimator. Simulation results indicate that the ordinary log rank test over-inflates type I error, while the proposed unconditional likelihood ratio test has appropriate type I error and higher power than the LW test. The method is demonstrated in real examples from the Sorbinil Retinopathy Trial, and the Age-Related Macular Degeneration Study. Raw data from these two trials are provided.

Probability of superiority for comparing two groups of clusters.

The probability of superiority (PS) has been recommended as a simple-to-interpret effect size for comparing two independent samples-there are several methods for computing the PS for this particular study design. However, educational and psychological interventions increasingly occur in clustered data contexts; and a review of the literature returned only one method for computing the PS in such contexts. In this paper, we propose a method for estimating the PS in clustered data contexts. Specifically, the proposal addresses study designs that compare two groups and group membership is determined at the cluster level. A cluster may be: (i) a group of cases with each case measured once, or (ii) a single case with each case measured multiple times, resulting in longitudinal data. The proposal relies on nonparametric point estimates of the PS coupled with cluster-robust variance estimation, such that the proposed approach should remain adequate regardless of the distribution of the response data. Using Monte Carlo simulation, we show the approach to be unbiased for continuous and binary outcomes, while maintaining adequate frequentist properties. Moreover, our proposal performs better than the single extant method we found in the literature. The proposal is simple to implement in commonplace statistical software and we provide accompanying R code. Hence, it is our hope that the method we present helps applied researchers better estimate group differences when comparing two groups and group membership is determined at the cluster level.

Tweedie Compound Poisson Models with Covariate-Dependent Random Effects for Multilevel Semicontinuous Data.

Multilevel semicontinuous data occur frequently in medical, environmental, insurance and financial studies. Such data are often measured with covariates at different levels; however, these data have traditionally been modelled with covariate-independent random effects. Ignoring dependence of cluster-specific random effects and cluster-specific covariates in these traditional approaches may lead to ecological fallacy and result in misleading results. In this paper, we propose Tweedie compound Poisson model with covariate-dependent random effects to analyze multilevel semicontinuous data where covariates at different levels are incorporated at relevant levels. The estimation of our models has been developed based on the orthodox best linear unbiased predictor of random effect. Explicit expressions of random effects predictors facilitate computation and interpretation of our models. Our approach is illustrated through the analysis of the basic symptoms inventory study data where 409 adolescents from 269 families were observed at varying number of times from 1 to 17 times. The performance of the proposed methodology was also examined through the simulation studies.