Group-Based Trajectory Modelling (GBTM) To Assess the Effect of Medication Adherence on Health-Related Outcomes: A Protocol for A Systematic Review

Background: The Group-based trajectory modelling (GBTM) method is increasingly used in pharmacoepidemiologic studies to describe medication adherence trajectories over time. However, assessing the effects of these medication adherence trajectories on health-related outcomes remains challenging. The purpose of this review is to describe studies assessing the effects of medication adherence trajectories estimated by the GBTM method on health-related outcomes. Methods: We will conduct a systematic review according to the recommendations of the Preferred Reporting Items for Systematic Review and Meta-analysis (PRISMA) guidelines. We will search in the following databases: PubMed, Embase, PsycINFO, Web of Science, CINAHL, and Cochrane database up to April 1st, 2021. Two reviewers will independently select articles and extract data. Discrepancies at every step will be resolved through discussion, and consensus will be reached for all disagreed articles. A third reviewer will act as a referee if needed. We will use tables to synthesize the modalities used to estimate medication adherence trajectories and the effect of adherence trajectories on health-related outcomes. We will identify the types of health-related outcomes studied and how they are dened, the statistical models used, the effect measure yield, and how medication adherence trajectories have been incorporated in the model. We will also review the limitations and biases reported by the authors and their attempts to mitigate them. We will provide a narrative synthesis. Discussion: This review will provide a clear view of the strategies and methods used in medication adherence research to estimate the effects of adherence trajectories on different health-related outcomes. A thorough exploration of how GBTM is used for this specic purpose could represent the rst crucial steps towards optimizing the utilization of this method in adherence studies. Systematic review registration: Prospero CRD42021213503.


Background
Medication adherence is a real challenge for healthcare professionals and patients. Sub-optimal adherence may be associated with poorer healthrelated outcomes and higher healthcare costs, depending on the disease. [1][2][3] Numerous adherence studies have been conducted to identify patients' characteristics with suboptimal adherence and the impact of non-adherence on health-related outcomes. [4] Several measures of adherence exist. When using medico-administrative databases as the source of drug information, researchers often use the proportion of days covered (PDC) or medication possession ratio (MPR), usually dichotomized with an 80% cut-off to distinguish adherent from non-adherent individuals. [4][5][6] Other adherence measures include self-reported questionnaires (e.g., Medication Adherence Report Scale (MARS)), tablet counts, electronic devices for vial caps) and more direct methods such as directly observing medication administration or dosage of drug metabolites in blood. [7][8][9][10] Most of these measures allow summarizing adherence over a de nite period, ranging from few weeks to years. However, an individual's drug adherence may vary considerably over time, and a given summarized measure may recover situations drastically different from each other.
Studying adherence dynamics over time in a population may re ect patient's adherence behaviors more accurately than summarizing adherence as a single average measure over time. The Group-based trajectory modelling (GBTM) is increasingly used to describe the dynamic and mutable nature of medication adherence behaviors. [11][12][13][14] It is a statistical method for modelling a variable of interest throughout time or age by identifying groups of individuals with similar pro les. [15] The method was initially developed by Nagin et al. to characterize developmental trajectories of criminal activities [16] and became prevalent for studying developmental trajectories in criminology, social science, and psychology. [17] GBTM is also increasingly used in medicine and clinical research to study the development of different mental health disorders, such as depression, attention de cit hyperactivity disorder, post-traumatic stress, or addiction, [17][18][19][20] and to capture heterogeneity in treatment response in clinical trials. [21] The GBTM method is part of the latent class growth analysis (LCGA) family and is a specialized application of the nite mixture models. [22][23][24] [15]As described by Nagin et al., the basic model yields two results: rst, it identi es individuals with similar trajectories (i.e., a group), and second, it estimates the individuals' probability of being part of each group. [25] For example, in medication adherence, the model allows to identify several adherence trajectories, each of them de ning a speci c adherence group (e.g., optimal adherence for the entire duration of the treatment, early discontinuation, suboptimal adherence throughout the treatment course, etc.). Each individual is then allocated to the trajectory that he/she has the highest probability to belong to. Once groups are identi ed, trajectories are plotted and presented in graphical form. Trajectories can also be used as the dependent variable to estimate characteristics associated with trajectory membership. Extensions of the classical GBTM model have been made, such as the possibility to account for risk and protective factors (e.g., age, sex, gender). [17] These analyses are relatively widespread as they are well described in the literature and bene t from software macros to help and guide researchers. [25,26] Medication adherence trajectories issued from GBTM could be bene cial for studying the associations between adherence and health-related outcomes, such as hospitalizations, death, or any critical clinical event. However, the GBTM literature does not provide speci c instructions for assessing these associations. To our knowledge, there is no software macro developed to perform analyses combining trajectories and the outcome in a joint model. Researchers generally proceed in two steps; identifying the adherence trajectories with GBTM and then using these trajectories as an observed variable in any suitable model to infer health-related outcomes.
Nonetheless, this approach presents statistical challenges, such as the uncertainty associated with group membership. The groups identi ed with the GBTM method are probabilistic. The group assignment may be considered a 100% imputation, possibly resulting in a not-quanti able uncertainty when inference about these groups is made through regression. Another concern is non-identi ability since GBTM imputes group membership on a not su ciently general model, resulting in attenuated estimates of the relationship between trajectories and health-related outcomes. All these problems are not speci c to the GBTM method but all latent class modeling. [27] These challenges have been described in particular by Bakk et al.[28] and Nylund-Gibson et al. [29], who have tried to nd solutions to limit these biases.
The lack of guidance on the modelling of adherence trajectories to estimate their effect on health-related outcomes may result in heterogeneity of the methods used. Thus, it is critical to investigate the different approaches existing in the literature, the biases, and the di culties encountered when applying adherence trajectory models to the study of distal outcomes. To our knowledge, there is no systematic review of studies that have used the GBTM method to measure the effects of medication adherence trajectories on health-related outcomes.
The purpose of this review is to identify and describe the studies assessing the effect of medication adherence trajectories, estimated by GBTM methods, on health-related outcomes. We will document the different types of study designs used, methods used to identify adherence trajectories, health-related outcomes studied, statistic modelling and parameters used, and limitations acknowledged by the studies' authors and how they were addressed.

Methods/design
This review will be conducted according to the recommendations of the Preferred Reporting Items for Systematic Review and Meta-analysis (PRISMA) guidelines. [30] For this protocol, we followed the recommendations of PRISMA-p. [31] The lled checklist is available in Appendix A.

Eligibility criteria
This systematic review will include all studies estimating medication adherence trajectories with the GBTM method [25] and evaluating the effects of these trajectories on speci ed health-related outcomes. We will not include studies on adherence to recommendations other than drug therapy (e.g., diet or exercise). Inclusion criteria are summarized in Table 1, available in Appendix B. Any study about medication adherence, estimated with any kind of method whether direct (e.g., pill count or plasmatic measurement) or indirect (e.g., questionnaire, adherence measure using medication database such as the proportion of days' covered medication possessing ratio, medication gap).

Groupbased trajectory
The group-based trajectory modeling a statistical methodology for modeling the evolution of medication adherence over time. This includes the based method described by Nagin et al. 19 and all its extensions.

Adherence trajectory
Adherence trajectory estimated with group-based trajectory as de ned by Nagin et al. and considered as an independent variable in any statistic modeling.

Health related outcomes
Health-related outcomes any health results measured following an intervention or behavior (e.g., surgery, treatment etc.) that describe a consequence of disease, treatment, or event for an individual. These health-related outcomes can be symptoms, hospitalizations, death, patient's quality of life, participation in activities, and social roles Population: We will not apply restrictions to the population: all studies measuring medication adherence trajectories in any human population will be considered. We will not apply restrictions based on age, race, sex or gender.
Intervention, exposure: We will consider any study in which medication adherence trajectories, de ned by the GBTM method, [25] are used as an exposure variable to explore or estimate the effect of adherence on any speci ed health-related outcome. Comparators: Depending on the study design, no comparator may be required. We will not exclude any of the studies based on the comparator. Outcomes: We will allow for any health-related measure described in the study as an endpoint outcome (dependent variable) in relation to adherence trajectories.

Study types:
We will include all original studies, observational studies, randomized trials, quasi-experimental studies, cohort or case-control studies. Conference abstracts, commentaries, letters to editors, and reviews will be excluded but retrieved to identify potentially eligible references.

Setting and time frame
No limit will be set for the study setting or time frame. We will retain all the original studies, including those conducted in clinical settings or those part of an intervention or a trial. Selected articles will enter the initial screening stage without a time limit for execution or publication.

Keys de nitions
Medication adherence: Medication adherence is de ned as "the extent to which patients follow the instructions given to them for prescribed treatments." [32] It is composed of three main concepts: 1) initiation (representing the extent to which a newly prescribed treatment is undertaken); 2) persistence (representing to what extent the treatment is taken for the recommended duration); and 3) implementation (representing to what extent the treatment is taken at the recommended doses and according to the recommended schedule). [33] We will consider all studies on medication adherence, whatever the type of adherence concept is measured.
Adherence trajectory: Adherence trajectories are de ned as descriptive longitudinal patterns of adherence over a de ned time set. De ned adherence metrics measured over a short time set are allowed to vary over a longer prede ned time horizon. Trajectories help to distinguish differences in patterns of adherence for individuals or groups of individuals over time. [11] Adherence trajectories model the evolution of adherence measures (for example, monthly PDC) over time and make it possible to identify people with similar adherence behaviours. [14] Health-related outcome [34] Health-related outcomes are results measured following an intervention or behavior (e.g., surgery, treatment) and describe a consequence of disease, treatment, or event for an individual. These health-related outcomes can be symptoms, hospitalizations, death, patient's quality of life, participation in activities, and social roles. In medication adherence, these outcomes allow us to measure the effect and consequences of different adherence pro les on individuals.

Information sources
We will search for relevant article references in the following databases: PubMed, Embase, PsycINFO, Web of Science, CINAHL, and Cochrane database up to April 1st, 2021. The search strategy has been developed and adapted for each database using the most sensible approach validated by a specialized librarian at Laval University. A complete description of the applied search strategies is described in Table 2 in Appendix B. Duplicate citations will be removed using EndNote and Covidence Solution software. [35] A nal manual revision of the database will be conducted to check for remaining duplicates. We will consider and include any additional articles not identi ed by our search strategy and brought to our attention by screening references of selected articles or relevant systematic reviews if they meet the inclusion criteria.
All identi ed studies will be compiled and kept with the full text (when needed) in a shared reference management software (i.e., EndNote)[36] and with Covidence, a web-based solution for systematic reviews. [35] Study selection Article selection will be performed using Covidence solution. [35] First, two reviewers will screen titles and abstracts independently, and articles will either be included, excluded, or categorized as unsure. Articles excluded by both reviewers will not be selected. Second, reviewers will discuss discrepancies to reach a consensus for every disagreed article. A pilot screening test will be conducted on a sample of a randomly selected 10% of the articles.
Likewise, the full text of selected articles will be reviewed independently by two reviewers, and articles will be included or excluded. The reason for exclusion will be documented. Discrepancies between the two reviewers will be resolved by discussion until an agreement is reached. A third reviewer will act as a referee if needed.

Data extraction
We will develop an extraction form using the Cochrane checklist of items to consider in data extraction. [37] The form will include the following elements: Study identi cation, including title, corresponding author's name and contact details, country, language, and publication date; Study design and objectives; Health domain, including diseases and medications of interest; Sample size; Medication adherence measure used, including description of data source, adherence measure (e.g., adherence questionnaire, adherence measured from health database, medication electronic monitoring system), and a full de nition of the variable and its operationalization (e.g., continuous, or dichotomized, threshold, scale, time frame); GBTM model software used, parameters, selection, and adequacy: link function, order, number of trajectories, statistics, and clinical criteria considered for the model selection.
Health-related outcome de nition: data source, variable de nitions and their operationalization (e.g., continuous, dichotomized, time frame); Approach used in modelling the health-related outcomes (one step, two steps, or three steps) and the rationale; Model used to assess the effect of medication adherence trajectories on health-related outcomes and its description (e.g., linear regression, logistic regression, Cox modelling) and methods to take into account missing data, lost to follow-up, and censoring; Exposure variable: how trajectories were introduced in the model (e.g., group membership, inverse probability weighting); Limitations and biases identi ed by the authors; How authors tried to mitigate identi ed biases.
The form will be tested on a random sample of 10% of the included studies. We will contact study authors (three attempts, two weeks apart) to request any relevant missing information. Data extraction will be conducted independently by two reviewers, and discrepancies checked for accurate extraction.
The data form with item de nitions is available in Table 3 in Appendix B.    The risk of bias and the quality of each study will be assessed using two checklists. For randomized trials, the risk of bias will be assessed with the Cochrane RoB 2.0 Tool.
[38] For observational studies, we will use the ROBINS-I tool. [39] As the review does not intend to estimate a global measure of effect, no studies will be excluded based on the quality assessment. Quality assessment will only serve for analysis purposes and discussion of ndings. This assessment will follow the same procedure as the data collection process. The quality assessment of each study will be done independently by two reviewers. Disagreements will be resolved by discussion between the reviewers or with a third reviewer as a referee.

Assessment of reporting
Included studies will be classi ed according to the quality of their reporting. The Detailed Guidelines for reporting on Latent Trajectory Studies (GRoLTS), [40] and the ESPACOMP Medication Adherence Reporting Guideline (EMERGE), [41] will be used to evaluate the studies. No study will be excluded during this step; instead, the reporting quality will be used for discussion purposes.
Analysis Data analysis will proceed in three phases. In the rst phase, we will describe selected studies with simple descriptive statistics and classify them in a table under the health domain studied (e.g., cancer, cardiovascular disease), medications used, population, and the studies' stated objective. In the second phase, we will summarize data according to the studies' method to estimate adherence trajectories. Studies will also be classi ed according to the source of data for medication adherence, medication adherence measure, the parameters used in GBTM, including the rationale behind parameter choice (e.g., statistics, clinical characteristics, the minimal number of patients included, number of groups). In the third phase, we will perform classi cation and narrative synthesis. We will review the choice and modalities used to estimate the effect of adherence trajectories on health-related outcomes, including the source of data, nature, and de nition of the health-related outcomes studied, the statistical model used, the effect measure yield, and how medication adherence measure has been incorporated in the model. We will also summarize limitations and biases reported by the authors and their attempt to mitigate them. Moreover, we will classify studies according to the quality of reporting and the study's overall quality in the synthesis. As the review does not aim to estimate a measured effect, we will not conduct a metaanalysis and assess studies' heterogeneity.
Discussion GBTM method has grown in popularity in adherence research over the last 20 years. [17] The method and its applications, the macroimplementations in the software, are well established and developed in many disciplines, such as pharmacoepidemiology. [42] Most of the studies on GBTM in this eld have primarily used trajectories to describe membership groups over time and associated factors. [43,44] To our knowledge, there are two systematic reviews on Latent Class Modelling approaches, including the GBTM method. [24,45] Nevertheless, they did not focus on the GBTM method to assess the association between adherence trajectories and health-related outcomes and related challenges. While the GBTM method provides a more re ned measure of medication adherence over time by identifying adherence trajectories, [46] it remains essential to study the effect of these trajectories on health-related outcomes. Despite the growing use of GBTM in adherence research and the availability of statistical tools, there is still considerable heterogeneity in how researchers use this method to model health-related outcomes. [47,48] This again leads to disparate and sometimes confusing ways of studying and reporting outcomes. However, modelling health outcomes according to pathway groups could help identify problematic groups and subsequently guide interventions and policies. It is, therefore, necessary to review how the method is used to model data and how results are reported.
The review will summarize the various strategies and methods used by authors to estimate the effect of adherence trajectories on health-related outcomes. Special attention will be paid to study designs, model parameter speci cations, and limitations. It will also document biases that could arise while using GBTM as an independent variable and how authors attempted to mitigate them. We will discuss how authors constructed the model, how they interpreted the results, and consider the effect of latent trajectories on health-related outcomes. Moreover, the review will also describe the studies' reporting quality with the two reporting guidelines speci c to latent class analysis and adherence studies. Therefore, this review could represent the rst crucial step towards developing a guide for the use of GBTM in medication adherence studies to infer health-