Elmer et al. BMC Medical Research Methodology (2018) 18:152 https://doi.org/10.1186/s12874-018-0620-9 TECHNICALADVANCE Open Access Using the Beta distribution in group-based trajectory models Jonathan Elmer1, Bobby L. Jones2 and Daniel S. Nagin3* Abstract Background: We demonstrate an application of Group-Based Trajectory Modeling (GBTM) based on the beta distribution. It is offered as an alternative to the normal distribution for modeling continuous longitudinal data that are poorly fit by the normal distribution even with censoring. The primary advantage of the beta distribution is the flexibility of the shape of the density function. Methods: GBTM is a specialized application of finite mixture modeling designed to identify clusters of individuals who follow similar trajectories. Like all finite mixture models, GBTM requires that the distribution of the data composing the mixture be specified. To our knowledge this is the first demonstration of the use of the beta distribution in GBTM. A case study of a beta-based GBTM analyzes data on the neurological activity of comatose cardiac arrest patients. Results: The case study shows that the summary measure of neurological activity, the suppression ratio, is not well fit by the normal distribution but due to the flexibility of the shape of the beta density function, the distribution of the suppression ratio by trajectory appears to be well matched by the estimated beta distribution by group. Conclusions: The addition of the beta distribution to the already available distributional alternatives in software for estimating GBTM is a valuable augmentation to extant distributional alternatives. Keywords: Group-based trajectory modeling, Beta distribution, Cardiac arrest Background of the specified distribution (e.g. mean and variance of a A trajectory describes the evolution of a behavior, normal distribution) are allowed to vary across trajectory biomarker, or some other repeated measure of interest groups. To our knowledge, previously published applica- over time. Group-based trajectory modeling (GBTM) tions have all specified the normal distribution, perhaps [1], also called growth mixture modeling [2], is a special- with censoring, the Poisson distribution, perhaps with ized application of finite mixture modeling designed to zero-inflation, or the binary logit function. Real-world identify clusters of individuals who follow similar trajec- continuous biomedical data are frequently not normally tories. Originally developed to study the developmental distributed even after allowing censoring. This is par- course of criminal behavior [3], GBTM is now widely ticularly true of biomarker data, which are generally applied in biomedical research in such diverse applica- positive, right skewed, and often zero-inflated. This tion domains as chronic kidney disease progression [4], creates a need for flexible alternatives to the Gaussian obesity [5, 6], pain [7], smoking [8], medication adoption distribution [12]. and adherence [9, 10], and concussion symptoms [11]. In this article, we demonstrate an application of Like all finite mixture models, GBTM requires that the GBTM based on the beta distribution. It is offered as an distribution of the data composing the mixture be alternative to the normal distribution for modeling con- specified, although there are no theoretical limits on the tinuous longitudinal data are poorly fit by other distribu- distributions that could be used. In GBTM, parameters tions. The primary advantage of the beta distribution is the flexibility of the shape of the density function. The normal density function, even in its censored form, must * Correspondence: [email protected] 3Heinz College, Carnegie Mellon University, Pittsburgh, PA 15206, USA follow some portion of it familiar bell-shaped form Full list of author information is available at the end of the article © The Author(s). 2018 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated. Elmer et al. BMC Medical Research Methodology (2018) 18:152 Page 2 of 5 whereas the shape of beta distribution is far less random vector Yi are assumed to be independent of one constrained. The disadvantage of beta distribution is that another. Thus, we may write: the data under study must be transformable to a 0–1 scale. YT Pk Y ijj; β j ¼ pk yit; j; β j ð2Þ Methods t¼1 The beta distribution can be parameterized in several different ways. One which is particularly useful for our While conditional independence is assumed at the purposes was proposed by [12]. Let y denote a beta level of the latent trajectory group, at the population distributed random variable: level outcomes are not conditionally independent because they depend on a latent construct, trajectory ΓðÞϕ PyðÞ¼; μ; ϕ yμϕ−1 group membership. See chapter 2 of [1] for a discussion ΓðÞμϕ ΓðÞðÞ1−μ ϕ of the conditional independence assumption. The GBTM modeling framework does not require μ ϕ where 0 < y < 1, 0 < < 1 and > 0. Under this that the random vector Yi be complete for all individ- μ Var y μ − μ ϕ parameterization E(y) = and ( )= (1 )/1 + ). uals. For the baseline GBTM specified above, missing ϕ The parameter is known as the precision parameter, values in Yi are assumed missing at random. However, μ ϕ because for any a larger value of results in a smaller for applications such as that described below where Var(y). measurement ends due to some external event—in We turn now to incorporating the beta distribution this case due to the death of the patient or the into GBTM. In describing a GBTM, we denote the patient awakening from coma—an extension of P Y distribution of trajectories by ( i), where the random GBTM described in [13] may be used to account for Y y y …y i’ vector i=( i1, i1, iT) represents individual s longitu- non-random dropout. dinal sequence of measurements over T measurement Detailed discussion of the methods to approach selec- occasions. The GBTM assumes that the population tion of J, the number of latent groups in the population, distribution of trajectories arises from a finite mixture and the order of the polynomial specifying each group’s J composed of groups. The likelihood for each individual trajectory are beyond the scope of this paper and have i J , conditional on the number of groups , may be written been previously described [1]. Briefly, no test statistics as: identifies the number of components in a finite mixture [14, 15]. Also, as argued in [1], in most application do- XJ ÀÁ mains of GBTM the population is not literally composed PYðÞ¼i π j∙PYijj; θ j ð1Þ j¼1 of a finite mixture of groups. Instead the finite mixture is intended to approximate an underlying unknown where πj is the probability of membership in group j, continuous distribution of trajectories for the purpose of and the conditional distribution of Yi given membership identifying and summarizing its salient features. As de- in j is indexed by the unknown parameter vector θj. scribed in [14, 16], finite mixture models are a valuable Typically, the trajectory is modeled by a polynomial tool for approximating an unknown continuous distribu- function of time (or age). For the case where P(Yi| j; θj) tion. In this paradigm, model selection is performed by is assumed to follow the beta distribution, its mean at combining test statistics such as AIC and BIC, which time t for group j, μjt, is linked to time as follows: can guide the statistician to identify which model best fits the data. This is combined with expert knowledge of μ ¼ β þ β t þ β t2…: jt 0 j 1 j 2 j which model best reveals distinctive trajectory groups that are substantively interesting. The order of the where, in principle, the polynomial can be of any order.1 polynomial used to model each group’s trajectory is Note that the parameters linking μjt to time are trajec- typically determined by starting with an assumed tory group specific, thus allowing the shapes of trajector- maximum order for each trajectory group then suc- ies to vary freely across group. Also associated with each cessively reducing the order if the highest order term trajectory group is a group specific precision parameter, is statistically insignificant. ϕj. The remaining components of θj pertain to the All models are estimated with software that is freely parameterization of πj, which in this case is specified to available at https://www.andrew.cmu.edu/user/bjones/.The follow a multinomial logistic function. maximization is performed using a general quasi-Newton For given j, conditional independence is assumed. In procedure [17, 18] and the variance-covariance of other words, except as explained by individual i’s parameter estimates are estimated by the inverse of the trajectory group membership, serial observations in the information matrix. Elmer et al. BMC Medical Research Methodology (2018) 18:152 Page 3 of 5 Results suppression ratio at hour 12. It has two spikes close to We demonstrate use of the beta distribution in a GBTM the minimum of 0 and the maximum of 1. In between, of data quantifying brain activity of 396 comatose pa- the suppression ratio is approximately uniformly distrib- tients resuscitated from cardiac arrest. The University of uted. The histogram bears no resemblance to the normal Pittsburgh Institutional Review Board approved all as- distribution. While it is possible for a mixture of cen- pects of this study.