Skip to Content
You are currently on the new version of our website. Access the old version .
MDPIMDPI
PharmacyPharmacy
PDF
  • Article
  • Open Access

9 April 2025

Investigating the Time-Varying Nature of Medication Adherence Predictors: An Experimental Approach Using Andersen’s Behavioral Model of Health Services Use

,
,
,
,
,
and
1
Department of Pharmacotherapy and Outcomes Science, VCU School of Pharmacy, Richmond, VA 23298, USA
2
Center for Pharmacy Practice Innovation, VCU School of Pharmacy, Richmond, VA 23298, USA
3
Department of Statistical Sciences and Operations Research, VCU College of Humanities & Sciences, Richmond, VA 23220, USA
4
College of Pharmacy, University of Michigan College of Pharmacy, Ann Arbor, MI 48109, USA
Pharmacy2025, 13(2), 53;https://doi.org/10.3390/pharmacy13020053
This article belongs to the Topic Optimization of Drug Utilization and Medication Adherence
Version Notes
Order Reprints

Abstract

Medication adherence is a crucial factor for managing chronic conditions, especially in aging adults. Previous studies have identified predictors of medication adherence. However, current methods fail to capture the time-varying nature of how risk factors can influence adherence behavior. This objective of this study was to implement multitrajectory group-based models to compare a time-varying to a time-fixed approach to identifying non-adherence risk factors. The study population comprised 11,068 Medicare beneficiaries aged 65 and older taking select medications for hypertension, high blood cholesterol, and oral diabetes medications, between 2008 and 2016. Time-fixed predictors (e.g., sex, education) were examined using generalized multinomial logistic regression, while time-varying predictors were explored through multitrajectory group-based modeling. Several predisposing, enabling, and need characteristics were identified as risk factors for following at least one non-adherence trajectory. Time-varying predictors displayed an alternative representation of those risk factors, especially depression symptoms. This study highlights the dynamic nature of medication adherence predictors and the utility of multitrajectory modeling. Findings suggest that targeted interventions can be developed by addressing the key time-varying factors affecting adherence.

1. Introduction

Non-adherence to medications is a major barrier to achieve desired outcomes, improve clinical outcomes, and improve health status [1,2,3,4,5,6,7,8]. It is estimated that only 50% of patients with chronic conditions are adherent to their treatment plan, which results in significant clinical and economic burdens. Every year in the United States, non-adherence to medications results in 125,000 deaths and health care costs up to USD 289 billion [9,10,11,12]. Given the tremendous impact in outcomes, the Centers for Medicare and Medicaid Services (CMS) administers the Star Ratings program, which rewards health plans based on their performance in health plan metrics, including medication adherence. The drugs considered in this quality metric program include renin–angiotensin system antagonists for hypertension, statins for hypercholesterolemia, and diabetes drugs excluding insulin (i.e., biguanides, sulfonylureas, thiazolidinediones, dipeptidyl peptidase-4 inhibitors, incretin mimetics, meglitinides, or sodium-glucose cotransporter 2 inhibitors).
In 2003, the World Health Organization issued a report highlighting the multifactorial causes of non-adherence, including socioeconomic, health care team and health system, disease-related, therapy-related, and patient-related factors [8,13]. These factors align with Andersen’s Behavior Model of Health Services Use (ABM), a widely used theoretical framework in health service research. Originally developed to study the family health service use, ABM is now used to explain interactions with medication use [14,15]. Its dimensions—predisposing characteristics (e.g., socio-demographic, social structure, and health beliefs), resources (personal, family, and community), and need (health status, comorbidities, treatment complexity, and patient’s independence)—overlap conceptually with the WHO’s non-adherence factors. Despite the seemingly broad agreement to using this ABM as a theoretical framework for studying health service use, the operationalization of specific items included in predisposing characteristics, enabling factors, and need resources has been inconsistent in previous studies [16].
In order to determine patterns of medication adherence behavior and predictors of poor performance, researchers increasingly use group-based trajectory modeling (GBTM) to analyze patterns of medication adherence across various prescription drugs [17,18,19,20,21,22,23]. Unlike categorizing patients as adherent or non-adherent, GBTM identifies similar adherence trajectories over time [24]. Previous studies primarily focused on predisposing characteristics, such as education, sex, ethnicity and race, or a single need characteristic, like comorbidities. However, adherence trajectories and their predictors can change over time, as risk factors for non-adherence do not occur in isolation or simultaneously. Traditional methods identify predictors by estimating their effects while holding other variables constant, focusing on fixed aspects influencing behavior over time [25,26]. Yet, time-varying factors like income, Medicaid eligibility, family support (e.g., spouse loss, household changes), and ambulatory independence may have a dynamic combined effect on medication adherence. Multitrajectory group-based modeling, an extension of GBTM, examines how such dynamic factors contemporaneously influence outcomes. This conceptual study used multitrajectory group-based models to describe the time-varying predictors of medication adherence trajectories following the ABM theoretical framework. This research addresses a fundamental gap in the medication adherence literature concerning the dynamic effect of risk factors of non-adherence (Figure 1).
Figure 1. Study conceptual framework.

2. Methods

2.1. Group-Based Trajectory Models (GBTMs) of Medication Adherence

This study ran a post hoc analysis of GBTM models previously estimated from monthly measurements of the proportion of days covered (PDC) to describe the longitudinal patterns of medication adherence of Medicare beneficiaries ≥65 years old between January 2008 and December 2016 [27]. The GBTM models were derived from participants from the Health and Retirement Study (HRS), which is a longitudinal panel study with a nationally representative sample of approximately 20,000 people in the United States sponsored by the National Institute of Aging (grant number NIA U01AG009740) and is conducted by the University of Michigan. Patients were taking select antihypertensives, including renin–angiotensin–aldosterone system inhibitors (RAASs), HMG Co-A reductase inhibitors (statins), or oral diabetes medications during the follow-up period [27]. The inclusion and exclusion criteria are described elsewhere, as well as a complete list of drugs included in the GBTM models [27]. GBTM models were estimated for antihypertensive drugs, statins, and diabetes medications (excluding insulin). Briefly, medication adherence was quantified using monthly measurements of PDC at the drug class level. This approach ensured that adherence measures were not influenced by drug within-class or impacts from brand-to-generic switches. In short, the GBTM models yielded three different models based on the drug class: select antihypertensives, with 3 trajectories (high-to-very high adherence, slow decline, and rapid decline); statins, yielding 5 trajectories (high-to-very high adherence, slow decline, low then increasing adherence, moderate decline, and rapid decline); and oral antidiabetics, which revealed 6 trajectories (high-to-very high adherence, slow decline, high then increasing adherence, low then increasing, moderate decline, and rapid decline). This study was approved by the Virginia Commonwealth University Internal Review Board (IRB) and the University of Michigan. HRS-linked administrative health claims data files from Medicare were obtained through CMS’s 3rd party data provider Medicare Research Information Center (MedRIC).

2.2. Predictors of Medication Adherence Trajectories

The conceptual framework of this study was based on ABM. However, it was important to ensure consistency with previous research. Therefore, the operationalization of variables for each of the dimensions of the ABM theoretical framework followed those as prescribed in Babitsch and colleagues (Table 1) [16].
Table 1. Operationalization of the dimensions of two conceptual frameworks: the Andersen’s Behavioral Model of Health Services Use and the causes of non-adherence summarized by the World Health Organization.
Measurements informing the covariates were obtained from the RAND longitudinal data file of the HRS public survey, including Sections A through K, respective to the period of analysis of medication adherence: 2008–2016 [28]. The relationship between predisposing, antecedents, enabling, and need characteristics (Appendix A) was examined in two ways: Firstly, using a time-stable approach, in which the last observation of each characteristic was investigated in the appropriate risk factor variation regression model. Secondly, through a time-varying approach in which repeated measures of each characteristic were explored in a multitrajectory group-based method. To minimize the impact of missing data, the last observation was carried forward in the time-varying approach. Only complete cases were considered in the final analysis. All statistical analyses were conducted using STATA MP 17 [29].

2.3. Time-Stable Predictors

A risk factor variation was implemented in the GBTM medication adherence model to examine which non-modifiable covariate was associated with membership to medication adherence trajectories for each of the 3 pharmacotherapeutic drug classes. This was achieved by performing a generalized logistic regression to each of the group-based trajectory models, in which time-stable covariates were tested for their ability to change group-membership probability [30]. A generalized logistic regression is an optimal approach because the parameters for each trajectory θ j of the multinomial logistic regression are able to denote the probability (P) of an individual i’s membership in group ( π j ( x i ) ), given the vector of variables that determine trajectory group membership (xi) (Equation (1)) [25,31,32]. Effectively, the effect of each vector of non-modifiable risk factor over time is modeled without loss of generality, where θ1 = 0 (Equation (2)) [30].
P Y i = ∑ j i e x i θ j ∑ j e x i θ j P j Y i
Equation (1)—Unconditional probability of i (individual) in any of the group-based trajectory model.
π j x i = e x i θ j ∑ j e x i θ j
Equation (2)—Vector of each time-stable risk factor over time.
Each covariate was investigated individually, followed by an adjusted model including all covariates found to be statistically significant in predicting membership for at least one medication adherence trajectory. Regression estimates, odds ratios, standard errors, and p-values were estimated to demonstrate the strength of association between each covariate and trajectory membership. To examine the potential for multicollinearity, variance inflation factor (VIF) was computed to determine by how much each risk factor estimate is increased because of the high correlation with other risk factors. When VIF is equal to 1, the coefficient of determination (R2) = 0, which means that the risk factor is not linearly related to other variables [33]. Therefore, a VIF greater than 5 was considered to be an indication of multicollinearity [34].

2.4. Time-Varying Predictors

A multitrajectory group-based model was used to examine how time-varying covariates influence membership probabilities across medication adherence trajectories for hypertension, hypercholesterolemia, and diabetes. This model incorporates the previously identified adherence trajectories while simultaneously plotting changes in time-varying predictors [25]. Similar methods have been applied in studying chronic kidney disease [35]. Unlike standard GBTM, this approach calculates conditional probabilities of trajectory membership for additional predictors beyond the first, allowing for a more comprehensive description of multiple risk factors [25,30]. Finally, since measurements of the predictors of medication adherence were obtained every two years, each annual measurement was matched with every 24 months of medication adherence follow-up period.

3. Results

In total, 11,068 participants were included in this post hoc analysis as those identified taking RAAS, statins, or oral diabetes medications between 2008 and 2016. The predisposing, enabling, and need characteristics are described in Table 2. Missingness was noteworthy in all the characteristics. The number of observations (n) indicated for each characteristic and the proportion of missingness are represented in Table 2.
Table 2. Study sample of sociodemographic, enabling, and need characteristics.

3.1. Time-Fixed Predictors of Medication Adherence Trajectories

A risk factor variation implemented in each group-based trajectory model of medication adherence is estimated elsewhere [27]. All risk factors included in each trajectory model displayed a VIF < 5, suggesting negligible evidence of multicollinearity (Appendix B). The risk factor variation was achieved by performing a generalized logistic regression with each of the group-based trajectory models, in which time-stable covariates are tested for their ability to change group-membership probability, considering the high-to-very adherence trajectory group as reference. Regression estimates, adjusted odds ratios (aORs), standard errors, and p-values were estimated to demonstrate the strength of association between each predisposing, enabling, or need characteristic and the likelihood of medication adherence trajectory membership, assuming the high-to-very high adherence trajectory as the reference group in each model (Table 3, Table 4, Table 5, Table 6 and Table 7).
Table 3. Time-fixed predictors of the rapid decline trajectory of the select antihypertensives, statins, and diabetes medications in medication adherence trajectory models.
Table 4. Time-fixed predictors of the slow decline trajectory of the select antihypertensives, statins, and diabetes medications in medication adherence trajectory models.
Table 5. Time-fixed predictors of the moderate decline trajectory of the select antihypertensives, statins, and diabetes medications in medication adherence trajectory models.
Table 6. Time-fixed predictors of the low then increasing adherence trajectory of the statins, and diabetes medications in medication adherence trajectory models.
Table 7. Time-fixed predictors of the high then increasing adherence trajectory of the diabetes medications in medication adherence trajectory models.

3.2. Time-Varying Predictors of Medication Adherence Trajectories

A multi-group-based trajectory analysis was implemented to investigate if and to what extent each of the time-varying enabling and need characteristics are associated with changes in medication adherence trajectories in each medication adherence model. Figure 2, Figure 3 and Figure 4 describe the multi-group-based trajectory models for the select antihypertensives, statins, and diabetes medications, respectively.
Figure 2. Multitrajectory model of enabling characteristics and select antihypertensive medication adherence trajectory.
Figure 3. Multitrajectory model of enabling characteristics and statin medication adherence trajectory.
Figure 4. Multitrajectory model of enabling characteristics and diabetes drug medication adherence trajectory.
  • Enabling characteristics
  • Self-reported health status
In the antihypertensives model, better health status correlated with high adherence, with minimal shifts across trajectories. Slow decline showed worse health than high adherence but better than rapid decline (Figure 2, Figure 3, Figure 4, Figure 5, Figure 6 and Figure 7). For statins and diabetes models, health status remained stable across trajectories, ranging from Good to Fair (Figure 3 and Figure 4).
Figure 5. (a) Multitrajectory model of need characteristics and select antihypertensive medication adherence trajectory. (b) Multitrajectory model of need characteristics and select antihypertensive medication adherence trajectory (continued).
Figure 6. (a) Multitrajectory model of need characteristics and statin medication adherence trajectory. (b) Multitrajectory model of need characteristics and statin medication adherence trajectory (continued).
Figure 7. (a) Multitrajectory model of need characteristics and diabetes drug medication adherence trajectory. (b) Multitrajectory model of need characteristics and diabetes drug medication adherence trajectory (continued).
  • Depression symptoms
In the antihypertensives model, depression increased in rapid decline, remained stable in slow decline, and declined sharply in high adherence (Figure 2). Statins and diabetes models showed similar patterns, with low, stable depression in high adherence, and sharp increases in moderate and rapid decline trajectories (Figure 3 and Figure 4).
  • Life satisfaction
Low adherence groups improved in life satisfaction over time, while slow decline showed stability (Figure 2). Statins showed the sharpest decline in moderate increase trajectories (Figure 3). Diabetes models exhibited no major changes, with all groups scoring as very satisfied (Figure 4).
  • Retirement satisfaction
No significant trends emerged across trajectories in any model, with retirement satisfaction remaining constant for every trajectory in all three models (Figure 2, Figure 3, Figure 4, Figure 5, Figure 6 and Figure 7).
  • Limitations in work due to health
High-to-very high adherence groups showed declining limitations, while slow decline and other trajectories increased (Figure 2). Trajectories in the statin model saw rising limitations overall, with high adherence starting from the lowest baseline (Figure 3). The diabetes models displayed similar increases across all trajectory groups (Figure 4).
2.
Need characteristics
  • Household income below poverty threshold
The select antihypertensive model, the high-to-very high adherence group, displayed a clear decline in the probability of living below the poverty threshold, even though other trajectory groups exhibited lower probabilities of living below the poverty threshold throughout the period of analysis (Figure 5a). The high-to-very high and low then increasing trajectories of the statin model display sharp decreases in the likelihood of living below the poverty threshold (Figure 6a). The rapid decline trajectory of the statins exhibited a slight decrease, although the likelihood of living below the poverty threshold was minimal at the beginning of this study (Figure 6a). Additionally, the slow decline trajectory displayed a slight increase in the likelihood of living below the poverty threshold (Figure 6a). In the diabetes medication model, all trajectories exhibit a constant low probability of living below the poverty threshold throughout the follow-up period (Figure 7a).
  • Marital status (loss of spouse)
The select antihypertensive model showed that the probability of living without a spouse decreased in the high-to-very high adherence group and for the slow decline group (Figure 5a). Contrastingly, patients in the rapid decline group exhibited the growing probability of being without a spouse (Figure 5a). The statin model showed no differences between trajectory groups, as all reported slight increases in the probability of losing a spouse over time (Figure 6a). In the diabetes medication model, all but the slow decline trajectories display increasing chances of losing a spouse during the follow-up period (Figure 7a). The sharpest increase in the probability of losing a spouse was observed in the moderate decline trajectory (Figure 7a). Notably, the small increase was observed in the high-to-very high adherence (“inverted U” shaped curve) and the low then increasing adherence trajectories (Figure 7a). Even though the slow decline trajectory of the diabetes medication model exhibited a decrease in the likelihood of losing a spouse, the probability of living without a spouse at the baseline and end of the follow-up period was one the highest (Figure 7a).
  • Living with resident children
The results show that the probability of living with resident children in the household remained stable throughout the follow-up period with no clear trends or shifts in the select antihypertensive model (Figure 5a). In the statin model, all trajectories exhibited the declining probability of residing with children in the household (Figure 6a). The high-to-very high trajectory in the statin model exhibits the largest probability of living with children at the beginning of this study and also the sharpest decline throughout the follow-up period, followed by the slow decline trajectory group (Figure 7a). Like the select antihypertensives, the diabetes medication model displayed no clear trends with all trajectories displaying a low probability of participants living with their children (Figure 7a).
  • Medicaid beneficiary
The probability of being a Medicaid beneficiary was consistently low across all trajectories in the select antihypertensive model, with minimal variation throughout the follow-up (Figure 5a). In the statin model, the high-to-very high adherence trajectory initially had the highest probability, followed by the sharpest decline (Figure 6a). The lower then increasing trajectory showed a smaller decline, while the slow decline trajectory exhibited a notable increase in likelihood (Figure 6a). The moderate and rapid decline trajectories maintained consistently low probabilities (Figure 6a). Similarly, all trajectories in the diabetes medication model displayed consistently low probabilities of Medicaid beneficiary status (Figure 7a).
  • Additional health coverage
In the antihypertensive model, the high-to-very high adherence trajectory showed the steepest decline in additional health insurance benefits, with slow decline following a similar but less pronounced pattern (Figure 5a). The rapid decline group remained stable, with minimal benefits (Figure 5a). In the statin model, high adherence showed a notable increase in additional coverage, while slow decline and lower then increasing trajectories decreased (Figure 6a). Moderate and rapid decline groups had consistently low, stable probabilities (Figure 6a). In the diabetes model, additional health insurance benefits declined overall, with high adherence maintaining the highest probability at both baseline and follow-up, while rapid decline showed the lowest (Figure 7a).
  • Smoking status
The high-to-very high adherence in the select antihypertensive model displayed the sharpest decline in the likelihood of being a smoker, while the remaining trajectories of this model exhibited sustained a low probability of being smokers (Figure 5b). In the statins, all trajectories displayed a very small and constant probability of being smokers throughout the follow-up period (Figure 6b). The same was observed in the diabetes medication model (Figure 7b).
  • Number of drinking days/week
The number of drinking days per week was overall low in all trajectories of the select antihypertensive, statin, and diabetes medication models, with all trajectories exhibiting a constant measure of no more than 1 drinking day per week (Figure 5b, Figure 6b, Figure 7b).
  • Instrumental activities of daily living (IADLs)
Difficulty with instrumental activities of daily living seem to generally increase with time, with the rapid decline trajectory exhibiting the sharpest surge in the select antihypertensive model (Figure 5b). Similarly, in the statin model, all but the lower then increasing adherence and rapid decline trajectories exhibit an increase in difficulty with instrumental activities of daily (Figure 6b). The lower then increased trajectory of the statins remained constant throughout the follow-up period, whereas the rapid decline trajectory seems to report slightly less difficulty with instrumental activities of daily during the follow-up period (Figure 6b). Nevertheless, the baseline score of IADLs of the rapid decline in the statin model was the highest compared to all other trajectories in the model (Figure 6b). The diabetes medication model exhibited similar results as the select antihypertensive model, except for the high-to-very high adherence and slow decline trajectories (Figure 7b).
  • Activities of daily living (ADLs)
In the select antihypertensive model, the high-to-very high adherence and rapid decline trajectories showed slight decreases in difficulty with activities of daily living (ADLs), while the slow decline trajectory exhibited an increase in tasks requiring assistance (Figure 5b). In the statin model, all trajectories except rapid decline showed increases in ADL difficulty, with the high-to-very high adherence group having the lowest baseline score and smallest increase (Figure 6b). ADL difficulty decreased over time in the statin rapid decline trajectory but started with the highest baseline score (Figure 6b). In the diabetes medication model, ADL scores generally increased, with sharpest rises in the slow decline and high then increasing trajectories (Figure 7b). The rapid decline group, despite improvement, had the highest baseline difficulty at the start of follow-up (Figure 7b).

4. Discussion

The purpose of this study was to examine the time-varying nature of risk factors of medication adherence trajectories of aging adults taking chronic medications for hypertension (RASA), hypercholesterolemia (statins), and diabetes (except for insulin). This study applied a multitrajectory group-based model, guided by the ABM framework, to analyze how predisposing, enabling, and need characteristics influence membership in medication adherence trajectory groups. Unlike prior models based solely on administrative claims, this study used HRS survey data to capture enabling and need characteristics. Predictors of medication adherence trajectories were assessed using two approaches: a time-fixed risk model examining the association between predictors and trajectory membership, and a multitrajectory model exploring how adherence trajectories align with changes in time-varying need and enabling characteristics. Importantly, the dynamic trajectory of each risk factor exhibited inconsistent shapes when examined individually for each medication adherence trajectory, while the time-fixed risk factor models exhibited consistency with the predictors of non-adherence to chronic medications [1,2,36,37].
The numerous recent studies examining medication adherence patterns using GBTM is proof that research recognizes that medication adherence is a dynamic behavior that can change with time [27,37,38,39,40,41,42,43,44,45,46,47]. Nevertheless, if one recognizes that medication adherence can change with time, the same can be said about the factors that influence it. Recent studies implementing a risk model based on multinomial logistic regressions do not allow this type of characterization [38,39,40,42]. This is because the traditional approach is limited to reporting adjusted odds ratios, representing the association between predictors and trajectory memberships, all else equal [48,49].
Time-fixed models found several risk factors associated with non-adherence, including predisposing characteristics such as being female, foreign-born, or non-white. These results align with previous studies linking non-adherence to demographic factors [7,23,50,51,52,53]. Even though college education was not found to be a significant risk factor for belonging to at least one non-adherent trajectory in any of the three models, a similar study linking administrative health care claims to a population-level survey from Australia reported similar findings when education was adjusted for covariates similar to ones considered in this study [50].
The multitrajectory model revealed that enabling characteristics like self-reported health, depression symptoms, and life satisfaction significantly predicted non-adherence. While time-fixed models linked non-adherence to depression, smoking, and drinking, the multitrajectory analysis showed stable probabilities for smoking and drinking but highlighted dynamic shifts in Medicaid eligibility, additional health coverage, and independence levels (IADL/ADL). Notably, additional health coverage, non-significant in time-fixed models, was strongly linked to high adherence in the multitrajectory analysis. It is important to clarify that variations in these characteristics do not imply a causal relationship but rather a longitudinal description of how each adherence trajectory and covariate trajectory progressed with time.
In essence, the time-fixed approach exhibited inconsistency in identifying which predictors were statistically significant factors of each medication adherence trajectory across pharmacotherapy classes. If researchers use only a time-fixed approach, results can exhibit statistical significance or not, like in this study. In case of non-statistical significance, the strength of evidence to guide actual practice innovations could be hampered. However, using the time-varying approach, researchers can look at the trajectory of individual predictors and determine if there is an actual variation over time that could be clinically meaningful. Practitioners can then investigate whether those predictor variations over time are worth tackling in practice to improve medication adherence.
This study emphasizes the value of multitrajectory modeling in identifying predictors of non-adherence linked to significant changes over time. This approach helps health care providers pinpoint key aspects of a patient’s life requiring intervention, such as the loss of a caregiving spouse, secondary health coverage, or autonomy. Health care providers and pharmacists could proactively assess the patient’s circumstance and identify causes of non-adherence (e.g., frailty, changes in household support). This would require building data-reporting systems that include contextual information about what the patient is going through that could be identified as a risk for non-adherence. Additionally, health care systems could implement data systems that allow for longitudinal measurements of medication adherence for individual patients, instead of measuring the PDC dichotomously during predetermined periods (i.e., the annual PDC). By characterizing these predictors throughout time, a multitrajectory analysis can guide targeted interventions and referrals, tailoring care to the specific needs of the patient population.

Limitations

Several limitations exist. Risk factors were drawn from the HRS using the ABM framework, but the HRS was not specifically designed to measure medication adherence predictors. Firstly, this study used an initial model of medication adherence trajectories that were estimated from administrative health care claims. While this approach for measuring adherence has been validated extensively, including by Grymonpre et al. and Galozy and colleagues, there are recognizable pitfalls associated with using these data [54,55]. These include patients filling prescriptions without insurance, via cash purchases or using promotional coupons (e.g., GoodRx or equivalent) that might not be recorded as claims and therefore indicate non-adherence, when, in reality, the patient had medication in hand. Despite the potential pitfalls of the PDC as a proxy measure for medication adherence, PDC estimates from administrative data sources have shown to correlate well with other direct observation methods to inform medication adherence [56]. High rates of missing data may have affected significance in time-fixed models and biased multitrajectory analysis. Additionally, there was a mismatch in measurement periods—adherence was estimated monthly using the PDC, while risk factors were measured biennially in the HRS. Despite these limitations, this study highlights the multitrajectory analysis as a promising method for exploring the impact of time-varying predictors on adherence. Moreover, this study did not examine the impact of adverse events in the trajectories of medication adherence, such as myocardial infarction or stroke. Such events have been previously described as predictors of poor medication adherence. Future studies following a quasi-experimental approach could explore the impact of acute negative events in the trajectories of medication adherence to chronic medications. Finally, this study included data obtained from HRS, which were obtained via surveys, which could be subject to potential recall bias.

5. Conclusions

This study demonstrated the potential of multitrajectory modeling to identify time-varying risk factors for non-adherence. Unlike traditional multinomial regression, this approach identifies both static and dynamic predictors, offering insights into which factors meaningfully change over time. Such methods can guide targeted interventions, improve medication adherence, and better support at-risk patient populations.

Author Contributions

Conceptualization, V.M.P., K.B.F. and D.A.H.; methodology, V.M.P., J.A.P., N.V.C., K.B.F. and D.A.H.; validation, V.M.P., K.B.F., D.L.D. and D.A.H.; formal analysis, V.M.P.; investigation, V.M.P., K.B.F., N.V.C. and D.A.H.; resources, D.A.H. and K.B.F.; data curation, V.M.P., J.A.P. and K.B.F.; writing—original draft preparation, V.M.P.; writing—review and editing, J.A.P., N.V.C., D.L.D., D.M., K.B.F. and D.A.H.; visualization, V.M.P.; supervision, K.B.F. and D.A.H.; project administration, V.M.P., K.B.F. and D.A.H.; funding acquisition, V.M.P. and D.A.H. All authors have read and agreed to the published version of the manuscript.

Funding

This research was partly funded by the PhRMA Foundation through the Predoctoral Fellowship in Value Assessment & Health Outcomes Research.

Institutional Review Board Statement

This study was conducted in accordance with the Declaration of Helsinki and approved by the Institutional Review Board of the Virginia Commonwealth University (protocol code HM20020850, approved on 23 October 2020).

Data Availability Statement

Data pertaining to the medication adherence trajectories is considered restricted data and therefore cannot be made available publicly, according to the Data Use Agreement with National Institute of Aging. Nevertheless, the data pertaining to the predictors of medication adherence trajectories utilized in this study was obtained from the public survey of the Health and Retirement Study, available at https://hrsdata.isr.umich.edu/data-products/public-survey-data.

Conflicts of Interest

Dave L. Dixon received grant funding from Boehringer Ingelheim. All other authors declare no relevant conflicts of interest or financial relationships. The funders had no role in the design of this study; in the collection, analyses, or interpretation of data; in the writing of the manuscript; or in the decision to publish the results.

Abbreviations

The following abbreviations are used in this manuscript:
GBTMGroup-based trajectory modeling
PDCProportion of days covered
HRSHealth and Retirement Study
VIFVariance inflation factor
ADLsActivities of Daily Living
IADLsInstrumental Activities of Daily Living

Appendix A

Table A1. Operationalization of enabling and need characteristics covariates.
Table A1. Operationalization of enabling and need characteristics covariates.
CharacteristicCovariatesMeasurement Approach
Enabling characteristicsSelf-reported health status5-point scale:
1—Excellent
2—Very good
3—Good
4—Fair
5—Poor
Depression symptomsCES-D 8-Item Scale. Per Steffick and colleagues, a score > 3 is indicative of clinical depression [24]
0—No depression symptoms (CES-D score ≤ 3)
1—With depression symptoms (CES-D score > 3)
Life satisfaction5-point scale:
1—Completely satisfied
2—Very satisfied
3—Somewhat satisfied
4—Not very satisfied
5—Not at all satisfied
Retirement satisfaction3-point scale:
1—Very satisfying
2—Moderately satisfying
3—Not at all satisfying
Limitations in work due to healthYes (1)/No (0)
Need characteristicsPoverty thresholdBelow (1)/Above (0)
Family structure
  • Loss of spouse
Yes (1)/No (0)
  • Living with resident children
Yes (1)/No (0)
Medicaid beneficiaryYes (1)/No (0)
Additional health coverageYes (1)/No (0)
Substance abuse
  • Smoking status
Yes (1)/No (0)
  • Alcohol consumption
Number of drinking days/week
Assistance with activities
  • Instrumental activities of daily living (IADLs)
  • Activities of daily living (ADLs)
Number of activities requiring assistance/cannot perform
CES-D: The 8-item Center for Epidemiological Studies Depression Scale [57].

Appendix B

Given the possibility of multicollinearity, the variance inflation factor (VIF) was computed to determine by how much each risk factor estimate is increased because of high correlation with other risk factors. VIF and R2 were computed to examine the presence of multicollinearity for the risk factors in each adjusted group-based trajectory model. In general, a VIF greater than 5 is indicative of multicollinearity.
Table A2. Multicollinearity Analysis.
Table A2. Multicollinearity Analysis.
GBTM ModelSelect HypertensivesStatinsDiabetes
CovariateVIFR2VIFR2VIFR2
Predisposing and antecedents
Sex: Female1.1700.1441.1500.1301.2200.179
Birthplace: Foreign-born1.4300.2991.4700.3201.5900.372
Race: Non-white1.2000.1651.1800.1531.1900.162
Ethnicity: Hispanic1.5300.3471.5500.3571.7400.425
Education: Not college-educated1.8200.4511.8500.4591.7900.440
Enabling characteristics
Self-reported Health Status1.5500.3551.5000.3321.4700.319
Depression Symptoms1.9300.4822.0100.5021.9600.490
Life Satisfaction1.2800.2161.2600.2041.2600.205
Retirement Satisfaction1.3100.2371.3200.2421.2400.191
Limitations in Work Due to Health1.2700.2111.2900.2241.3000.232
Need characteristics
Household income below poverty index1.3400.2521.3300.2511.3400.256
Marital spouse: Loss of spouse1.2200.1821.2000.1691.2800.218
Number of resident children1.0800.0741.0800.0751.0600.057
Medicaid eligibility1.3200.2451.3200.2411.3600.263
Additional health coverage1.1300.1181.1200.1101.1800.155
Smoking status: Smoker1.0500.0511.0600.0541.0300.029
Number of drinking days/week1.1400.1191.1000.0931.0800.070
Instrumental activities of daily living1.3600.2631.4100.2891.5500.356
Activities of daily living1.5100.3361.5100.3391.7600.432
Mean VIF1.381 1.389 1.410

References

  1. Arlt, S.; Lindner, R.; Rosler, A.; von Renteln-Kruse, W. Adherence to medication in patients with dementia: Predictors and strategies for improvement. Drugs Aging 2008, 25, 1033–1047. [Google Scholar] [CrossRef] [PubMed]
  2. Bowry, A.D.; Shrank, W.H.; Lee, J.L.; Stedman, M.; Choudhry, N.K. A systematic review of adherence to cardiovascular medications in resource-limited settings. J. Gen. Intern. Med. 2011, 26, 1479–1491. [Google Scholar] [CrossRef] [PubMed]
  3. Coletti, D.J.; Stephanou, H.; Mazzola, N.; Conigliaro, J.; Gottridge, J.; Kane, J.M. Patterns and predictors of medication discrepancies in primary care. J. Eval. Clin. Pract. 2015, 21, 831–839. [Google Scholar] [CrossRef] [PubMed]
  4. Gard, P.R. Non-adherence to antihypertensive medication and impaired cognition: Which comes first? Int. J. Pharm. Pract. 2010, 18, 252–259. [Google Scholar] [CrossRef]
  5. Hudani, Z.K.; Rojas-Fernandez, C.H. A scoping review on medication adherence in older patients with cognitive impairment or dementia. Res. Soc. Adm. Pharm. 2016, 12, 815–829. [Google Scholar] [CrossRef]
  6. Lenti, M.V.; Selinger, C.P. Medication non-adherence in adult patients affected by inflammatory bowel disease: A critical review and update of the determining factors, consequences and possible interventions. Expert Rev. Gastroenterol. Hepatol. 2017, 11, 215–226. [Google Scholar] [CrossRef]
  7. Warren, J.R.; Falster, M.O.; Fox, D.; Jorm, L. Factors influencing adherence in long-term use of statins. Pharmacoepidemiol. Drug Saf. 2013, 22, 1298–1307. [Google Scholar] [CrossRef]
  8. World Health Organization. Adherence to Long-Term Therapies: Evidence for Action; World Health Organization: Geneva, Switzerland, 2020; Available online: https://apps.who.int/medicinedocs/pdf/s4883e/s4883e.pdf (accessed on 20 January 2020).
  9. McCarthy, R. The price you pay for the drug not taken. Bus. Health 1998, 16, 27–28, 30, 32–33. [Google Scholar]
  10. Osterberg, L.; Blaschke, T. Adherence to medication. N. Engl. J. Med. 2005, 353, 487–497. [Google Scholar] [CrossRef]
  11. New England Healthcare Institute. Thinking Outside the Pillbox: A System-Wide Approach to Improving Patient Medication Adherence for Chronic Disease. Available online: https://www.nehi.net/writable/publication_files/file/pa_issue_brief_final.pdf (accessed on 14 March 2020).
  12. Viswanathan, M.; Golin, C.E.; Jones, C.D.; Ashok, M.; Blalock, S.J.; Wines, R.C.; Coker-Schwimmer, E.J.; Rosen, D.L.; Sista, P.; Lohr, K.N. Interventions to Improve Adherence to Self-administered Medications for Chronic Diseases in the United States. Ann. Intern. Med. 2012, 157, 785–795. [Google Scholar] [CrossRef]
  13. Brown, M.T.; Bussell, J.K. Medication adherence: WHO cares? Mayo Clin. Proc. 2011, 86, 304–314. [Google Scholar] [CrossRef] [PubMed]
  14. Andersen, R.M. Revisiting the behavioral model and access to medical care: Does it matter? J. Health Soc. Behav. 1995, 36, 1–10. [Google Scholar] [CrossRef] [PubMed]
  15. Andersen, R.M.; Davidson, P.L.; Baumeister, S.E. Improving access to care in America. In Changing the US Health Care System: Key Issues in Health Services Policy and Management, 3rd ed.; Jossey-Bass: Hoboken, NJ, USA, 2007; pp. 3–31. [Google Scholar]
  16. Babitsch, B.; Gohl, D.; von Lengerke, T. Re-revisiting Andersen’s Behavioral Model of Health Services Use: A systematic review of studies from 1998–2011. Psychosoc. Med. 2012, 9, Doc11. [Google Scholar] [CrossRef]
  17. Alhazami, M.; Pontinha, V.M.; Patterson, J.A.; Holdford, D.A. Medication Adherence Trajectories: A Systematic Literature Review. J. Manag. Care Spec. Pharm. 2020, 26, 1138–1152. [Google Scholar] [CrossRef]
  18. Ajrouche, A.; Estellat, C.; De Rycke, Y.; Tubach, F. Trajectories of Adherence to Low-Dose Aspirin Treatment Among the French Population. J. Cardiovasc. Pharmacol. Ther. 2020, 25, 37–46. [Google Scholar] [CrossRef] [PubMed]
  19. Dillon, P.; Stewart, D.; Smith, S.M.; Gallagher, P.; Cousins, G. Group-Based Trajectory Models: Assessing Adherence to Antihypertensive Medication in Older Adults in a Community Pharmacy Setting. Clin. Pharmacol. Ther. 2018, 103, 1052–1060. [Google Scholar] [CrossRef]
  20. Feldman, C.H.; Collins, J.; Zhang, Z.; Subramanian, S.; Solomon, D.H.; Kawachi, I.; Costenbader, K.H. Dynamic patterns and predictors of hydroxychloroquine nonadherence among Medicaid beneficiaries with systemic lupus erythematosus. Semin. Arthritis Rheum. 2018, 48, 205–213. [Google Scholar] [CrossRef]
  21. Franklin, J.M.; Krumme, A.A.; Shrank, W.H.; Matlin, O.S.; Brennan, T.A.; Choudhry, N.K. Predicting adherence trajectory using initial patterns of medication filling. Am. J. Manag. Care 2015, 21, e537–e544. [Google Scholar]
  22. Franklin, J.M.; Shrank, W.H.; Pakes, J.M.; Sanfélix-Gimeno, G.; Matlin, O.S.; Brennan, T.A.M.; Choudhry, N.K. Group-based trajectory models: A new approach to classifying and predicting long-term medication adherence. Med. Care 2013, 51, 789–796. [Google Scholar] [CrossRef]
  23. Hernandez, I.; He, M.; Chen, N.; Brooks, M.M.; Saba, S.; Gellad, W.F. Trajectories of Oral Anticoagulation Adherence Among Medicare Beneficiaries Newly Diagnosed With Atrial Fibrillation. J. Am. Heart Assoc. 2019, 8, e011427. [Google Scholar] [CrossRef]
  24. Nagin, D.S.; Jones, B.L.; Passos, V.L.; Tremblay, R.E. Group-based multi-trajectory modeling. Stat. Methods Med. Res. 2018, 27, 2015–2023. [Google Scholar] [CrossRef] [PubMed]
  25. Nagin, D.S. Group-Based Modeling of Development; Harvard University Press: Cambridge, MA, USA, 2005. [Google Scholar]
  26. Nagin, D.S. Group-based trajectory modeling: An overview. Ann. Nutr. Metab. 2014, 65, 205–210. [Google Scholar] [CrossRef]
  27. Pontinha, V.M.; Patterson, J.A.; Dixon, D.L.; Carroll, N.V.; Mays, D.; Barnes, A.; Farris, K.B.; Holdford, D.A. Longitudinal medication adherence group-based trajectories of aging adults in the US: A retrospective analysis using monthly proportion of days covered calculations. Res. Soc. Adm. Pharm. 2023, 20, 363–371. [Google Scholar] [CrossRef]
  28. RAND Center for the Study of Aging. RAND HRS Longitudinal File 2016 (V2) Documentation. Available online: https://hrsdata.isr.umich.edu/sites/default/files/documentation/codebooks/randhrs1992_2016v2.pdf (accessed on 3 May 2020).
  29. StataCorp LLC. Stata Statistical Software: Release 17; StataCorp LLC: College Station, TX, USA, 2021. [Google Scholar]
  30. Jones, B.L.; Nagin, D.S. A Note on a Stata Plugin for Estimating Group-based Trajectory Models. Sociol. Methods Res. 2013, 42, 608–613. [Google Scholar] [CrossRef]
  31. Maddala, G.S. Qualitative and Limited Dependent Variable Models in Econometrics; Cambridge University Press: Cambridge, UK, 1983; p. 498. [Google Scholar]
  32. Hsiao, C. Logit and probit models. In The Econometrics of Panel Data; Springer: Berlin/Heidelberg, Germany, 1996; pp. 410–428. [Google Scholar]
  33. Montgomery, D.C. Introduction to Linear Regression Analysis, 5th ed.; Wiley: Hoboken, NJ, USA, 2013. [Google Scholar]
  34. Kutner, M.H.; Nachtsheim, C.; Neter, J. Applied Linear Regression Models, 4th ed.; McGraw-Hill/Irwin: Boston, MA, USA; New York, NY, USA, 2004. [Google Scholar]
  35. Burckhardt, P.; Nagin, D.S.; Padman, R. Multi-Trajectory Models of Chronic Kidney Disease Progression. AMIA Annu. Symp. Proc. 2016, 2016, 1737–1746. [Google Scholar] [PubMed]
  36. Durand, H.; Hayes, P.; Harhen, B.; Conneely, A.; Finn, D.P.; Casey, M.; Murphy, A.W.; Molloy, G.J. Medication adherence for resistant hypertension: Assessing theoretical predictors of adherence using direct and indirect adherence measures. Br. J. Health Psychol. 2018, 23, 949–966. [Google Scholar] [CrossRef] [PubMed]
  37. Stentzel, U.; Berg, N.v.D.; Schulze, L.N.; Schwaneberg, T.; Radicke, F.; Langosch, J.M.; Freyberger, H.J.; Hoffmann, W.; Grabe, H.-J. Predictors of medication adherence among patients with severe psychiatric disorders: Findings from the baseline assessment of a randomized controlled trial (Tecla). BMC Psychiatry 2018, 18, 155. [Google Scholar] [CrossRef]
  38. Woolpert, K.M.; Schmidt, J.A.; Ahern, T.P.; Hjorth, C.F.; Farkas, D.K.; Ejlertsen, B.; Collin, L.J.; Lash, T.L.; Cronin-Fenton, D.P. Clinical factors associated with patterns of endocrine therapy adherence in premenopausal breast cancer patients. Breast Cancer Res. 2024, 26, 59. [Google Scholar] [CrossRef]
  39. Fatima, B.; Mohan, A.; Altaie, I.; Abughosh, S. Predictors of adherence to direct oral anticoagulants after cardiovascular or bleeding events in Medicare Advantage Plan enrollees with atrial fibrillation. J. Manag. Care Spec. Pharm. 2024, 30, 408–419. [Google Scholar] [CrossRef]
  40. Chang, C.-Y.; Jones, B.L.; Hincapie-Castillo, J.M.; Park, H.; Heldermon, C.D.; Diaby, V.; Wilson, D.L.; Lo-Ciganic, W.-H. Association between trajectories of adherence to endocrine therapy and risk of treated breast cancer recurrence among US nonmetastatic breast cancer survivors. Br. J. Cancer 2024, 130, 1943–1950. [Google Scholar] [CrossRef]
  41. Wabe, N.; Timothy, A.; Urwin, R.; Xu, Y.; Nguyen, A.; Westbrook, J.I. Analysis of Longitudinal Patterns and Predictors of Medicine Use in Residential Aged Care Using Group-Based Trajectory Modeling: The "MEDTRAC-Cardiovascular" Longitudinal Cohort Study. Pharmacoepidemiol. Drug Saf. 2024, 33, e5881. [Google Scholar] [CrossRef] [PubMed]
  42. Schmidt, J.A.; Woolpert, K.M.; Hjorth, C.F.; Farkas, D.K.; Ejlertsen, B.; Cronin-Fenton, D. Social Characteristics and Adherence to Adjuvant Endocrine Therapy in Premenopausal Women With Breast Cancer. J. Clin. Oncol. 2024, 42, 3300–3307. [Google Scholar] [CrossRef]
  43. Mohan, A.; Chen, H.; Deshmukh, A.A.; Wanat, M.; Essien, E.J.; Paranjpe, R.; Fatima, B.; Abughosh, S. Group-based trajectory modeling to identify adherence patterns for direct oral anticoagulants in Medicare beneficiaries with atrial fibrillation: A real-world study on medication adherence. Int. J. Clin. Pharm. 2024, 46, 1525–1535. [Google Scholar] [CrossRef] [PubMed]
  44. Ishiwata, R.; AlAshqar, A.; Miyashita-Ishiwata, M.; Borahay, M.A. Dispensing patterns of antidepressant and antianxiety medications for psychiatric disorders after benign hysterectomy in reproductive-age women: Results from group-based trajectory modeling. Womens Health 2024, 20, 17455057241272218. [Google Scholar] [CrossRef]
  45. Huang, W.; Ahmed, M.M.; Morris, E.J.; Yang, L.; O’Neal, L.; Hernandez, I.; Bian, J.; Kimmel, S.E.; Smith, S.; Guo, J. Trajectories of Sacubitril/Valsartan Adherence Among Medicare Beneficiaries With Heart Failure. JACC Adv. 2024, 3, 100958. [Google Scholar] [CrossRef] [PubMed]
  46. Abegaz, T.M.; Shehab, A.; Gebreyohannes, E.A.; Bhagavathula, A.S.; Elnour, A.A. Nonadherence to antihypertensive drugs: A systematic review and meta-analysis. Medicine (Baltimore) 2017, 96, e5641. [Google Scholar] [CrossRef]
  47. Ruksakulpiwat, S.; Schiltz, N.K.; Irani, E.; Josephson, R.A.; Adams, J.; Still, C.H. Medication Adherence of Older Adults with Hypertension: A Systematic Review. Patient Prefer. Adherence 2024, 18, 957–975. [Google Scholar] [CrossRef]
  48. Cummings, P. The relative merits of risk ratios and odds ratios. Arch. Pediatr. Adolesc. Med. 2009, 163, 438–445. [Google Scholar] [CrossRef]
  49. Greenland, S. Interpretation and choice of effect measures in epidemiologic analyses. Am. J. Epidemiol. 1987, 125, 761–768. [Google Scholar] [CrossRef]
  50. Park, K.H.; Tickle, L.; Cutler, H. Identifying temporal patterns of adherence to antidepressants, bisphosphonates and statins, and associated patient factors. SSM Popul. Health 2022, 17, 100973. [Google Scholar] [CrossRef]
  51. Wang, C.-H.; Huang, L.C.; Yang, C.-C.; Chen, C.-L.; Chou, Y.-J.; Chen, Y.-Y.; Yang, W.-C.; Chen, L. Short- and long-term use of medication for psychological distress after the diagnosis of cancer. Support. Care Cancer 2017, 25, 757–768. [Google Scholar] [CrossRef] [PubMed][Green Version]
  52. Hsu, Y.H.; Mao, C.L.; Wey, M. Antihypertensive medication adherence among elderly Chinese Americans. J. Transcult. Nurs. 2010, 21, 297–305. [Google Scholar] [CrossRef] [PubMed]
  53. Bird, G.C.; Cannon, C.P.; Kennison, R.H. Results of a survey assessing provider beliefs of adherence barriers to antiplatelet medications. Crit. Pathw. Cardiol. 2011, 10, 134–141. [Google Scholar] [CrossRef]
  54. Grymonpre, R.; Cheang, M.; Fraser, M.; Metge, C.; Sitar, D.S. Validity of a prescription claims database to estimate medication adherence in older persons. Med. Care 2006, 44, 471–477. [Google Scholar] [CrossRef]
  55. Galozy, A.; Nowaczyk, S.; Sant’Anna, A.; Ohlsson, M.; Lingman, M. Pitfalls of medication adherence approximation through EHR and pharmacy records: Definitions, data and computation. Int. J. Med. Inform. 2020, 136, 104092. [Google Scholar] [CrossRef] [PubMed]
  56. Farley, J.F.; Urick, B.Y. Is it time to replace the star ratings adherence measures? J. Manag. Care Spec. Pharm. 2021, 27, 399–404. [Google Scholar] [CrossRef]
  57. Turvey, C.L.; Wallace, R.B.; Herzog, R. A revised CES-D measure of depressive symptoms and a DSM-based measure of major depressive episodes in the elderly. Int. Psychogeriatr. 1999, 11, 139–148. [Google Scholar] [CrossRef]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Article Metrics

Citations

Article Access Statistics

Journal Statistics
Multiple requests from the same IP address are counted as one view.
Impact of Removing Race Coefficient from Glomerular Filtration Rate Estimation Equations on Antidiabetics Among Black Patients
Previous
Evaluating Pharmacy Student Perspectives and Attitudes Towards Compliance Aids and Devices Through Health Disparity Simulation
Next