Journal of Statistical Research

Optimal allocation schemes in mixed ANCOVA models for longitudinal data

Xiaojian Xu, Sanjoy K Sinha — Sun, 09 Jul 2023 00:00:00 +0000

We discuss the construction of optimal allocation schemes for the linear mixed model with clustered outcomes or repeated measurements often encountered in longitudinal studies. We consider both treatment and covariate effects in the mixed model, where latent pro- cesses are used to describe random cluster or subject effects. A goal of optimal design schemes is to determine proportions of sample units allocated to each treatment for a given total sample size. We develop the optimal designs in a general setting using both D- and A- optimal design criteria. Specifically, we propose a two-stage design approach to deal with unknown parameters in the linear mixed model, where the variances of the random effects across the treatment groups are considered different. We study the empirical properties of the proposed designs using Monte Carlo simulations. An application is also provided using actual clinical data from a longitudinal study.

Journal of Statistical Research, Vol 56, No 2, p101-114

Debias random forest regression predictors

Lihua Chen, Prabhashi Withana Gamage, John Ryan — Sun, 09 Jul 2023 00:00:00 +0000

The random forest can reduce the variance of regression predictors through bagging while leaving the bias mostly unchanged. In general, the bias is not negligible and consequently bias correction is necessary. The default bias correction method implemented in the R package randomForest often works poorly. Several approaches have been developed which in general outperform the R default. However, little work has been done to com- prehensively evaluate the performance of these methods and thus guide users to select an appropriate method for bias correction. This paper fills this gap by providing an informa- tive ranking of these bias correction methods based on an extensive numerical study. We further offered practical suggestions on the application of the winner of these methods and suggested a visualization technique to help users decide when bias correction is needed.

Journal of Statistical Research 2022, Vol. 56, No. 2, pp. 115-131

Properties of inverse probability of adherence weighted estimator of the per-protocol effect for sustained treatment strategies under different data-generating mechanisms and adherence patterns

Lucy Mosquera, Mohammad Ehsanul Karim, Md Belal Hossain — Sun, 09 Jul 2023 00:00:00 +0000

Inverse Probability (of Adherence) Weighted per-protocol (IPW-PP) estimators are get- ting popular in addressing medication non-adherence while analyzing pragmatic trial data. However, their finite sample properties under different data generating mechanisms (DGMs) have not been investigated comprehensively. In the current work, we investigated the finite sample performances of such estimators in the context of a pragmatic random- ized controlled trial. We compared the performances of IPW-PP estimators with commonly used naive and baseline-adjusted per-protocol estimators, under different DGMs emulating pragmatic trials, comparing two sustained treatment strategies, possibly with a non-null effect. DGMs include (i) different roles of a baseline variable; whether future time-varying prognostic factors are impacted by past adherence; and whether the baseline variable is measured, (ii) whether adherence patterns observed in two arms are differential, and when we have access to measurements of adherence and confounders that are recorded infre- quently (sparsely). When baseline confounders are adjusted, we generally obtain unbiased estimates, but if some necessary variables are not measured, the IPW-PP estimator may still be preferable. High non-adherence patterns might negatively impact IPW-PP effect estimators, particularly when DGMs include confounding that may be influenced by previ- ous adherence history. We used the above estimators to analyze a case study from the Lipid Research Clinics Coronary Primary Prevention Trial data in the presence of non-adherence.

Journal of Statistical Research 2022, Vol. 56, No. 2, pp.134-154

Approximate methods for analyzing semiparametric longitudinal models with nonignorable missing responses

Najla Aloraini, Sanjoy Sinha — Sun, 09 Jul 2023 00:00:00 +0000

We often encounter missing data in longitudinal studies. When the missingness in longitudinal data is nonignorable, it is necessary to incorporate the missing data mechanism into the observed data likelihood function for a valid statistical inference. In this article, we propose and explore a novel semiparametric approach to estimating the regression parameters and variance components using a partially linear mixed model with nonignorable and nonmonotone missing responses. The ﬁnite sample properties of the proposed method are studied using Monte Carlo simulations, where our method is found to be very effective in capturing any curvilinear pattern in the mean response. The method is also illustrated using some actual longitudinal data obtained from a public health survey, referred to as the Health and Retirement Study (HRS).

Journal of Statistical Research 2022, Vol. 56, No. 2, pp. 155-183

Penalized logistic normal multinomial factor analyzers for high dimensional compositional data

Wangshu Tu, Sanjeena Subedi — Sun, 09 Jul 2023 00:00:00 +0000

Model-based clustering utilizes a finite mixture model to identify underlying patterns or clusters across samples. A finite mixture model is a convex combination of two or more distributions, where appropriate distributions are chosen depending on the type of the data. Recently, there has been a great interest in clustering human microbiome data. Microbiome data are compositional (yielding relative abundance) and are high-dimensional. Previously, a family of logistic normal multinomial factor analyzers (LNM-FA) for model-based clus- tering of high-dimensional microbiome data was proposed via a factor analyzer structure. This reduced the number of parameters and computation overhead compared to a traditional mixtures of logistic normal multinomial models. Here, we propose a penalized LNM-FA (PLNM-FA) model by utilizing lasso regularization to each entry of the loading matrix. This introduces further parsimony compared to LNM-FA and also estimates the number of latent factors simultaneously. Parameter estimation is done using a variational variant of the alternating expectation conditional maximization algorithm to maximize the penalized maximum likelihood. The performance of proposed algorithm is evaluated using simula- tion studies and real data.

Journal of Statistical Research 2022, Vol. 56, No. 2, pp.185-216