Adherence Patterns 1 Year After Initiation of SGLT2 Inhibitors: Results of a National Cohort Study

ABSTRACT

Objectives: Adherence to medications is important for the management of chronic diseases. Although the proportion of days covered (PDC) is a common metric for measuring adherence, it may be insufficient to distinguish relevant differences in medication-taking behavior. Group-based trajectory models (GBTMs) have been used to better represent adherence over time. This study aims to examine adherence patterns 1 year after initiation among users of sodium-glucose cotransporter 2 (SGLT2) inhibitors using GBTMs and evaluate the ability of baseline characteristics to predict adherence trajectory.

Study Design: SGLT2 inhibitor new-user cohort study from 2014 to 2018.

Methods: We calculated 12-month PDC and categorized patients with PDC of 80% or greater as adherent. We performed multivariable logistic regression on adherence status controlling for baseline covariates. GBTMs were fit to identify adherence patterns 12 months following SGLT2 inhibitor initiation. Five multinomial logistic regression models including different subsets of predictors were used to predict adherence trajectory group assignment.

Results: In a cohort of 228,363 SGLT2 inhibitor users, the mean PDC was 57%, with 36% of the cohort being adherent. Overall, women and patients with anxiety or depression were less likely to be adherent. Six patterns of SGLT2 inhibitor adherence were identified with GBTMs: 1 fill (PDC = 0.08), early discontinuation (PDC = 0.22), consistently low adherence (PDC = 0.35), moderate adherence (PDC = 0.48), high adherence (PDC = 0.79), and near-perfect adherence (PDC = 0.95). All prediction models showed poor predictive accuracy (0.35).

Conclusions: We found wide variation in adherence patterns among SGLT2 inhibitor users in a national cohort. Predictors from a health care claims database were unable to accurately predict adherence trajectory.

Am J Manag Care. 2024;30(8):e226-e232. https://doi.org/10.37765/ajmc.2024.89591

Takeaway Points

This study examined adherence among users of sodium-glucose cotransporter 2 inhibitors and found a wide variation of medication patterns among the patients. We found that measured variables from claims data were insufficient to predict trajectory patterns.

• Researchers should consider using both proportion of days covered and trajectory pattern when describing medication-taking behavior.
• To accurately predict medication adherence patterns, it will be necessary to move beyond data included in medication claims.

Chronic diseases, such as diabetes, often require long-term medication treatment for effective control. Medication adherence can be a key factor in successful disease management.^1,2 However, medication adherence can be difficult to assess. Standard measures of adherence, such as the proportion of days covered (PDC) and the medication possession ratio (MPR), suffer from known limitations.³ For example, PDC and MPR are often dichotomized with a single cut point to define high vs low adherence, which ignores the dynamic nature of medication adherence. This approach may also not adequately capture the long-term behavior of patients. An alternative is group-based trajectory models (GBTMs) that use repeated adherence measures, such as PDC, to fit long-term trajectories of medication use. This approach may better capture the variability in long-term medication-taking behavior that is often not adequately captured in traditional adherence measurements.^3-5 GBTMs have been demonstrated to be useful for describing adherence in several chronic disease states.⁶

In this study, we sought to assess adherence to sodium-glucose cotransporter 2 (SGLT2) inhibitors. SGLT2 inhibitors are a class of oral antidiabetic medications initially approved by the FDA in March 2013 for patients with type 2 diabetes (T2D). In 2021, the American Diabetes Association included SGLT2 inhibitors as one of the treatment options following first-line treatment failure to improve glucose control.⁷ The use of SGLT2 inhibitors has increased in the US, Europe, and Japan.^8-12 With the growing utilization of SGLT2 inhibitors, understanding adherence to this class of medications is an important issue.

Several observational studies have evaluated adherence to SGLT2 inhibitors, but only 1 study used a GBTM approach.^13-17 A cohort study conducted by Hawley et al evaluated adherence to SGLT2 inhibitors and showed that almost half of the SGLT2 inhibitor users remained adherent (PDC ≥ 80%) 1 year after initiation. These authors reported 3 common adherence pattern trajectories: low, moderate, and high adherence.¹³ However, the study used Medicare data that included patients 66 years or older, limiting extrapolation of the results to a younger population who often have different adherence profiles. Additionally, the use of only 3 group-based trajectories may have oversimplified the complexity of SGLT2 inhibitor medication patterns.

The purpose of this study is to address limitations of the traditional dichotomous adherence measurements by examining the adherence to SGLT2 inhibitors using both PDC and GBTMs. Moreover, we expanded upon previous SGLT2 inhibitor trajectory studies by adding additional GBTM trajectory groups to characterize unique patterns that may otherwise be grouped together. The additional unique patterns can provide insight to providers as to what patterns of medication use might be expected when patients initiate treatment and also lead to earlier actions when patients begin to exhibit fill patterns indicative of one of the patterns. Additionally, this study furthers prior research by examining the use of baseline characteristics to predict the adherence trajectory of patients treated with SGLT2 inhibitors. Predicting the adherence trajectory group of an individual would allow clinicians to target tailored interventions based on their likely medication-taking behavior.⁴ Therefore, the primary objective of this study was to evaluate adherence to SGLT2 inhibitors 1 year after initiation and characterize common medication use trajectories. Secondly, we aimed to evaluate the performance of different sets of baseline variables for predicting adherence trajectory.

MATERIALS AND METHODS

Data Source and Study Cohort

This was a retrospective new-user cohort study using the Truven MarketScan data from January 2014 through December 2019. Truven MarketScan is an administrative claims database that contains patient demographics, outpatient drug dispensing, inpatient and outpatient claims, and health care utilization information for insured patients in the US.¹⁸

We included patients 18 years or older who initiated an SGLT2 inhibitor between January 2014 and December 2018. The first dispensing of an SGLT2 inhibitor during the study period was defined as the index date. We excluded patients who did not have continuous enrollment and pharmacy benefits in the 12 months prior to the index date and those who had received a type 1 diabetes diagnosis during that period.

Study Outcomes and Covariates

The main outcome of this study was adherence during the 12 months after initiation. SGLT2 inhibitor adherence was defined using PDC. We calculated the 12-month PDC by dividing the sum of days supplied for SGLT2 inhibitor dispensing over 12 months by 365 days. Patients with a 12-month PDC of 80% or greater were defined as adherent. We also calculated 12 monthly PDCs for group-based trajectory modeling.

Covariates were measured using information in the 365 days prior to and including the index date (baseline period). We preselected a set of baseline characteristics that, based on prior literature, were expected to be related to SGLT2 inhibitor adherence, including age, sex, comorbidities, and medication used.^13,17 Patients were considered to have the specified comorbidity if they had a corresponding inpatient or outpatient International Classification of Diseases (ICD), Ninth Revision or International Statistical Classification of Diseases, Tenth Revision code during the baseline period. A detailed list of codes can be found in eAppendix Table 1 (eAppendix available at ajmc.com). Additionally, we included indicators for the use of cardiovascular medications (antihypertensives, statins) and other antidiabetic medications during the baseline period. All the medications were identified using National Drug Codes (NDCs) and used as a dichotomous variable at the class level. Last, we calculated concurrent medication use by counting the unique NDC with days’ supplies that overlapped the index date.

Analysis

Baseline characteristics were summarized using counts and percentages. We performed multivariable logistic regression on adherence status adjusting for age, sex, comorbidities, and medication use. Estimates of associations were presented using ORs and 95% CIs. We estimated GBTMs using monthly PDC measures to identify common adherence trajectories for SGLT2 inhibitor users during the 1-year time frame after initiation. In GBTMs, clusters of individuals who follow similar trajectories over time are identified and grouped together. This approach has been previously described.^6,19,20 In brief, we first estimated a PDC for each patient for each month during the 1-year follow-up period. Using these 12 monthly PDCs, we modeled the patterns of medication use in linear GBTMs where patients were assigned to a trajectory group in which they had the highest probability of membership. We performed multiple models with 3, 4, 5, or 6 trajectory groups and selected the final model based on the lowest Bayesian information criterion (BIC) value.^21,22 The GBTM analyses were performed using an add-on package, PROC TRAJ, in SAS version 9.4 (SAS Institute Inc). We calculated summary statistics of the 12-month PDC for each trajectory group to characterize differences among groups. The significance level was set as a P value less than .05, and analyses described thus far were performed using SAS 9.4.

Prediction Models for Adherence Trajectory

To predict adherence trajectory, we performed a multinomial logistic regression using the full cohort. The dependent variable in each of the multinomial models was the adherence trajectory group based on the results from GBTM. We fit several multinomial logistic regression models to compare the predictive performance across different models (model 1 to model 3). A detailed description of each prediction model and its methods can be found in eAppendix Table 2.

To determine whether additional baseline factors can predict the adherence trajectory group, we used a data-driven approach similar to high-dimensional propensity score variable selection to better utilize information available in the claims data.²³ We extracted all baseline codes (ICD codes for inpatient and outpatient claims and NDC codes for drug dispensing) and selected the most frequently appearing 200 codes during baseline. We created 200 dichotomous indicator variables for each of the 200 codes and calculated the number of times each code appeared during baseline for 200 count variables that account for potential cumulative effect of each code. In total, we created 400 predictors where half were dichotomous variables and half were count variables of the 200 most frequently occurring codes during baseline. We repeated the same approach to create the prediction models using the 400 data-derived variables (model 4 to model 5).

To estimate the performance of the prediction models, we conducted 10-fold cross-validation using the full cohort.^24,25 The results of the cross-validation were compared with the actual trajectory group assignment to determine the predictive accuracy. Model accuracy was calculated based on the percentage of patients who were correctly assigned into the true trajectory group divided by all patients. Cohen κ values were calculated for each prediction model.^26,27 The κ value is a summary index that describes the strength of agreement between observed accuracy and expected accuracy accounting for random chance. Larger values for κ represent more agreement, with a value of 0 indicating the agreement was mainly due to chance.²⁸ A confusion matrix that displayed the distribution of predicted trajectory group and actual trajectory group from each model was created. All prediction model building and cross-validation analyses were done in R version 4.2.2 (R Foundation for Statistical Computing) using the CrossValidate package.

To account for patients with short-term use of SGLT2 inhibitors who may be different from patients with long-term SGLT2 inhibitor use, we performed sensitivity analysis on prediction models excluding patients in 2 trajectory groups with the smallest PDC. The purpose of the sensitivity analysis was to evaluate whether prediction accuracy can be improved when focusing solely on long-term SGLT2 inhibitor users. We repeated the same approach of the main prediction modeling without the model selection and calculated the prediction accuracy and κ values.

RESULTS

SGLT2 Inhibitor Cohort

We identified a total of 228,363 adult new users of SGLT2 inhibitors during the study period who met the inclusion and exclusion criteria (eAppendix Table 3). In the cohort, 44.9% were women and the mean (SD) age was 54.5 (9.8) years (Table 1). For baseline comorbidities, 10.9% had coronary artery disease, 74.3% had hypertension, 74.6% had hyperlipidemia, and 25.3% had obesity. For medication used, 79.5% had an antihypertensive dispensing, 65.7% had a statin dispensing, and 94.4% had an antidiabetic dispensing other than SGLT2 inhibitors during the baseline period.

Factors Associated With SGLT2 Inhibitor Adherence Status

The mean (SD) 12-month PDC was 0.57 (0.3), with 35.7% of patients being defined as adherent (eAppendix Table 4). Overall, women were less likely to be adherent (OR, 0.71; 95% CI, 0.70-0.72) and older individuals (OR, 1.01; 95% CI, 1.01-1.01) were more likely to be adherent (Table 2). We found most of the baseline comorbidities were negatively associated with being adherent. Among comorbidities, chronic kidney disease (OR, 0.79; 95% CI, 0.75-0.83) and Alzheimer disease/dementia (OR, 0.82; 95% CI, 0.76-0.89) had the largest effect on being less adherent. Patients with more concurrent medications (OR, 1.07; 95% CI, 1.06-1.07), antihypertensives (OR, 1.07; 95% CI, 1.03-1.11), or statins (OR, 1.07; 95% CI, 1.05-1.09) were more likely to be adherent to SGLT2 inhibitors. Patients who had insulin dispenses during baseline were less likely to be adherent (OR, 0.89; 95% CI, 0.87-0.91).

GBTM

Based on BIC value, we selected the GBTM with 6 trajectory groups (Figure). The identified 6 trajectory patterns of SGLT2 inhibitor adherence included 1 fill (11.3% of SGLT2 inhibitor cohort), early discontinuation (17.4%), consistently low adherence (5.9%), moderate adherence (14.9%), high adherence (33.8%), and near-perfect adherence (16.8%). Table 3 shows the summary statistics of PDC for each trajectory group. When comparing age and sex among trajectory groups, the near-perfect trajectory had the highest average age and the lowest percentage of women. The descriptive characteristics for each of the trajectory groups are shown in eAppendix Table 5.

Prediction Models

Table 4 presents the performance metric of each prediction model. All 5 models showed poor prediction accuracy of approximately 35% with no significant change after including or excluding more predictors. The κ values for all prediction models were close to 0, indicating the poor agreement between observed accuracy and expected accuracy. Based on the confusion matrix, the majority of prediction accuracy came from the model correctly predicting high adherence trajectory in all 5 models (eAppendix Table 6). Moreover, all models were not able to predict the group with consistently low adherence. Sensitivity analyses excluding patients with short-term SGLT2 inhibitor use showed similar results to the main prediction models. All models showed prediction accuracy of 48% with κ values that were close to 0 (eAppendix Table 7).

DISCUSSION

This cohort study of new users of SGLT2 inhibitors showed a mean 12-month PDC of 0.57, with 35.7% of SGLT2 inhibitor users being adherent. We found wide variation in adherence trajectories among SGLT2 inhibitor users 12 months after initiation of an SGLT2 inhibitor. Moreover, we found that the ability to accurately predict trajectory group assignment using baseline characteristics was poor regardless of the number of predictors.

Our results align with those of previous studies. Five observational studies have evaluated 12-month adherence to SGLT2 inhibitors.^13-17 We observed the lowest 12-month PDC (0.57) compared with results of the 5 previous studies. This may be explained by the differences in inclusion criteria for the SGLT2 inhibitor cohort. Three of the 5 previous studies applied postindex inclusion criteria that excluded patients who were not enrolled in the health care system 12 months after index.^14-16 The studies may include SGTL2 inhibitor users who interacted more frequently with the health care system, leading to higher PDC. One study excluded patients who only had their index SGLT2 inhibitor dispensing and resulted in the highest 12-month PDC of 0.79 among all 5 observational studies.¹⁷ Our results show slightly lower 12-month PDC (0.57 vs 0.63) compared with the study conducted by Hawley et al that included Medicare patients 66 years or older, whereas our analysis included adult patients 18 years or older.¹³ The mean age for our SGLT2 inhibitor user cohort is much younger compared with the study from Hawley et al (55 vs 72 years). Aligned with our findings, previous studies’ results have suggested that older patients were more adherent to SGLT2 inhibitors than younger patients.^15,16 The age difference between cohorts could contribute to the difference in 12-month SGLT2 inhibitor PDC across studies.

Prior studies have used GBTM methods to assess medication use patterns. In fact, a systematic review of chronic medication use measured with GBTM methods showed 5 trajectories were the most common medication-taking patterns.³ These patterns align with the SGLT2 inhibitor trajectories observed in our study.³ Consistent with previous studies, we found a wide variation of nonadherence patterns among SGLT2 inhibitor users.^3,29 Adherence patterns can be more informative as there might be various patterns for patients with similar PDCs. By using trajectory modeling, we may be able to speculate on the reasons for nonadherence, something dichotomous PDC cannot do. For example, 11% of SGLT2 inhibitor users in our cohort had only their initial 30 days of SGLT2 inhibitor dispense. This group of patients likely may have experienced adverse effects that led to discontinuing the SGLT2 inhibitor rather than having poor adherence.³ Future research can focus on the causes of trajectory patterns and how these models can provide insights for trajectory-specific intervention. Additionally, we observed the range of PDC among the group with consistently low adherence overlapped entirely with the PDC from the early discontinuation group (Table 3). This further validated that the dichotomous PDC cannot recognize longitudinal medication-taking behavior because patients with the same PDC may have different patterns.

Although GBTMs can be useful in better describing medication use, we were unable to accurately predict trajectory group assignment. The ability to accurately predict an adherence pattern when initiating a drug could lead to targeting a set of interventions to patients who might be at high risk for a suboptimal adherence pattern. However, regardless of the variables that were included in the model, there was no improvement in the predictive accuracy relative to including only age and sex. Our results demonstrate that the best predictive ability was among the high adherence trajectory group (eAppendix Table 6). This result is consistent with a previous study that used baseline characteristics to predict statin trajectory and observed better predictive ability for an extreme adherence or nonadherence trajectory.⁴ The accuracy in predicting assignment to other trajectories was poor.

Despite our attempt to predict group assignment through the addition of multiple variables, the predictive accuracy did not improve, which was similar to results from a prior study.³⁰ One potential explanation is that the variables available in a retrospective claims database may not be useful predictors for medication adherence.^31,32 Although we attempted to use a high-dimensional approach for inclusion of multiple variables and variable types in hopes that these may represent proxies for important predictor variables, it is clear that these variables are insufficient as predictors of medication-taking behavior. This is consistent with results of other studies demonstrating that retrospective information from health care interactions is not enough to predict long-term medication adherence. In attempting to identify patients who are most likely to follow a specific medication-taking trajectory, it will be important to consider alternative data sources.

Limitations

It is important to acknowledge some of the limitations of this study. Similar to other studies that used administrative claims data, the use of drug dispensing information may not represent actual medication-taking behavior. Additionally, we did not have information on clinical variables (diabetes duration, hemoglobin A_1C value, glucose level), payer formulary/policy, or socioeconomic-related variables (income, education level) that may be associated with medication-taking behavior.^33-35 This information may ultimately improve the ability to predict adherence trajectory. We understand that not including postindex enrollment may result in misclassification of individuals who disenrolled but remained on SGLT2 inhibitors under a separate benefit payment. However, we did not want to introduce any bias by requiring continuous enrollment during the follow-up period. Thus, we are trading off potential misclassification with avoiding immortal time bias. We did not require a T2D diagnosis for the SGLT2 inhibitor cohort, and patients without a baseline T2D diagnosis may have different SGLT2 inhibitor trajectory patterns compared with those who had a baseline T2D diagnosis. However, the class of medication we are interested in is quite specific to patients with T2D, and we anticipate that few patients are using this medication without a T2D diagnosis. Moreover, we want to capture the real-world SGLT2 inhibitor use among all patients who initiated an SGLT2 inhibitor regardless of the T2D diagnosis.

CONCLUSIONS

We found that PDC alone is not sufficient in describing medication-taking behavior. Patients defined as nonadherent based on PDC included those with a wide variation of medication use patterns among SGLT2 inhibitor users. Moreover, baseline characteristics were insufficient in predicting medication adherence. To accurately predict medication adherence patterns, it will be necessary to move beyond data included in medication claims.

Author Affiliations: Department of Pharmacy Systems Outcomes and Policy, College of Pharmacy, University of Illinois Chicago (HCH, DRT, GTS, TAL), Chicago, IL; Leslie Dan Faculty of Pharmacy, University of Toronto (MT), Toronto, Ontario, Canada; School of Public Health, University of Illinois Chicago (SA), Chicago, IL.

Source of Funding: None.

Author Disclosures: Dr Touchette has been a paid consultant to AstraZeneca, eMax Health, Horizon Pharmaceuticals, and Stage Analytics; has been the principal investigator on grants and contracts paid to University of Illinois Chicago by AbbVie, Institute for Clinical and Economic Review, and Takeda Pharmaceuticals; and has received royalties for economic models licensed to the Institute for Clinical and Economic Review. The remaining authors report no relationship or financial interest with any entity that would pose a conflict of interest with the subject matter of this article.

Authorship Information: Concept and design (HCH, MT, SA, TAL); acquisition of data (HCH, TAL); analysis and interpretation of data (HCH, DRT, MT, GTS, SA, TAL); drafting of the manuscript (HCH, TAL); critical revision of the manuscript for important intellectual content (HCH, DRT, MT, GTS, SA, TAL); statistical analysis (HCH, MT, SA, TAL); provision of patients or study materials (HCH); administrative, technical, or logistic support (GTS, TAL); and supervision (DRT, MT, TAL).

Address Correspondence to: Hsiao-Ching Huang, PhD, Department of Pharmacy Systems Outcomes and Policy, School of Pharmacy, University of Illinois Chicago, 833 S Wood St, Chicago, IL 60612. Email: hhuang80@uic.edu.

REFERENCES

1. De Geest S, Sabaté E. Adherence to long-term therapies: evidence for action. Eur J Cardiovasc Nurs. 2003;2(4):323. doi:10.1016/S1474-5151(03)00091-4

2. Chen CC, Tseng CH, Cheng SH. Continuity of care, medication adherence, and health care outcomes among patients with newly diagnosed type 2 diabetes: a longitudinal analysis. Med Care. 2013;51(3):231-237. doi:10.1097/MLR.0b013e31827da5b9

3. Alhazami M, Pontinha VM, Patterson JA, Holdford DA. Medication adherence trajectories: a systematic literature review. J Manag Care Spec Pharm. 2020;26(9):1138-1152. doi:10.18553/jmcp.2020.26.9.1138

4. Franklin JM, Krumme AA, Shrank WH, Matlin OS, Brennan TA, Choudhry NK. Predicting adherence trajectory using initial patterns of medication filling. Am J Manag Care. 2015;21(9):e537-e544.

5. Kronish IM, Ye S. Adherence to cardiovascular medications: lessons learned and future directions. Prog Cardiovasc Dis. 2013;55(6):590-600. doi:10.1016/j.pcad.2013.02.001

6. Nagin DS. Group-based trajectory modeling: an overview. Ann Nutr Metab. 2014;65(2-3):205-210. doi:10.1159/000360229

7. American Diabetes Association Professional Practice Committee. 9. pharmacologic approaches to glycemic treatment: Standards of Medical Care in Diabetes—2022. Diabetes Care. 2022;45(suppl 1):S125-S143. doi:10.2337/dc22-S009

8. Harris ST, Patorno E, Zhuo M, Kim SC, Paik JM. Prescribing trends of antidiabetes medications in patients with type 2 diabetes and diabetic kidney disease, a cohort study. Diabetes Care. 2021;44(10):2293-2301. doi:10.2337/dc21-0529

9. Curtis HJ, Dennis JM, Shields BM, et al. Time trends and geographical variation in prescribing of drugs for diabetes in England from 1998 to 2017. Diabetes Obes Metab. 2018;20(9):2159-2168. doi:10.1111/dom.13346

10. Wilkinson S, Douglas I, Stirnadel-Farrant H, et al. Changing use of antidiabetic drugs in the UK: trends in prescribing 2000-2017. BMJ Open. 2018;8(7):e022768. doi:10.1136/bmjopen-2018-022768

11. Yagi N, Komiya I, Arai K, et al; Japan Diabetes Clinical Data Management Study Group (JDDM study group). Current status of oral antidiabetic drug prescribing patterns based on the body mass index for Japanese type 2 diabetes mellitus patients and yearly changes in diabetologists’ prescribing patterns from 2002 to 2019 (JDDM61). J Diabetes Investig. 2022;13(1):65-73. doi:10.1111/jdi.13621

12. Kohro T, Yamazaki T, Sato H, et al. Trends in antidiabetic prescription patterns in Japan from 2005 to 2011. Int Heart J. 2013;54(2):93-97. doi:10.1536/ihj.54.93

13. Hawley CE, Lauffenburger JC, Paik JM, Wexler DJ, Kim SC, Patorno E. Three sides to the story: adherence trajectories during the first year of SGLT2 inhibitor therapy among Medicare beneficiaries. Diabetes Care. 2022;45(3):604-613. doi:10.2337/dc21-1676

14. Buysman EK, Anderson A, Bacchus S, Ingham M. Retrospective study on the impact of adherence in achieving glycemic goals in type 2 diabetes mellitus patients receiving canagliflozin. Adv Ther. 2017;34(4):937-953. doi:10.1007/s12325-017-0500-4

15. Cai J, Divino V, Burudpakdee C. Adherence and persistence in patients with type 2 diabetes mellitus newly initiating canagliflozin, dapagliflozin, DPP-4s, or GLP-1s in the United States. Curr Med Res Opin. 2017;33(7):1317-1328. doi:10.1080/03007995.2017.1320277

16. Cai J, Wang Y, Baser O, Xie L, Chow W. Comparative persistence and adherence with newer anti-hyperglycemic agents to treat patients with type 2 diabetes in the United States. J Med Econ. 2016;19(12):1175-1186. doi:10.1080/13696998.2016.1208208

17. Ofori-Asenso R, Liew D, Lalic S, et al. Adherence, persistence, and switching among people prescribed sodium glucose co-transporter 2 inhibitors: a nationwide retrospective cohort study. Adv Ther. 2019;36(11):3265-3278. doi:10.1007/s12325-019-01077-3

18. MarketScan Research Databases for life sciences researchers. Merative. 2023. Accessed July 10, 2024. https://www.merative.com/content/dam/merative/documents/brief/marketscan-research-databases-for-life-sciences-researchers.pdf

19. Nagin DS, Odgers CL. Group-based trajectory modeling in clinical research. Annu Rev Clin Psychol. 2010;6:109-138. doi:10.1146/annurev.clinpsy.121208.131413

20. Nagin DS, Odgers CL. Group-based trajectory modeling (nearly) two decades later. J Quant Criminol. 2010;26(4):445-453. doi:10.1007/s10940-010-9113-7

21. Schwarz G. Estimating the dimension of a model. Ann Statist. 1978;6(2):461-464. doi:10.1214/aos/1176344136

22. Neath AA, Cavanaugh JE. The Bayesian information criterion: background, derivation, and applications. Wiley Interdiscip Rev Comput Stat. 2012;4(2):199-203. doi:10.1002/wics.199

23. Schneeweiss S, Rassen JA, Glynn RJ, Avorn J, Mogun H, Brookhart MA. High-dimensional propensity score adjustment in studies of treatment effects using health care claims data. Epidemiology. 2009;20(4):512-522. doi:10.1097/EDE.0b013e3181a663cc

24. Jung Y, Hu J. A K-fold averaging cross-validation procedure. J Nonparametr Stat. 2015;27(2):167-179. doi:10.1080/10485252.2015.1010532

25. Steyerberg EW, Harrell FE Jr, Borsboom GJ, Eijkemans MJ, Vergouwe Y, Habbema JD. Internal validation of predictive models: efficiency of some procedures for logistic regression analysis. J Clin Epidemiol. 2001;54(8):774-781. doi:10.1016/s0895-4356(01)00341-9

26. McHugh ML. Interrater reliability: the kappa statistic. Biochem Med. 2012;22(3):276-282. doi:10.11613/BM.2012.031

27. Cohen J. A coefficient of agreement for nominal scales. Educ Psychol Meas. 1960;20(1):37-46. doi:10.1177/001316446002000104

28. Chmura Kraemer H, Periyakoil VS, Noda A. Kappa coefficients in medical research. Stat Med. 2002;21(14):2109-2129. doi:10.1002/sim.1180

29. Lo-Ciganic WH, Donohue JM, Jones BL, et al. Trajectories of diabetes medication adherence and hospitalization risk: a retrospective cohort study in a large state Medicaid program. J Gen Intern Med. 2016;31(9):1052-1060. doi:10.1007/s11606-016-3747-6

30. Franklin JM, Shrank WH, Lii J, et al. Observing versus predicting: initial patterns of filling predict long-term adherence more accurately than high-dimensional modeling techniques. Health Serv Res. 2016;51(1):220-239. doi:10.1111/1475-6773.12310

31. Steiner JF, Ho PM, Beaty BL, et al. Sociodemographic and clinical characteristics are not clinically useful predictors of refill adherence in patients with hypertension. Circ Cardiovasc Qual Outcomes. 2009;2(5):451-457. doi:10.1161/CIRCOUTCOMES.108.841635

32. Steiner JF. Can we identify clinical predictors of medication adherence... and should we? Med Care. 2010;48(3):193-195. doi:10.1097/MLR.0b013e3181d51ddf

33. Parra DI, Romero Guevara SL, Rojas LZ. Influential factors in adherence to the therapeutic regime in hypertension and diabetes. Invest Educ Enferm. 2019;37(3):e02. doi:10.17533/udea.iee.v37n3e02

34. Gonzalez JS, Tanenbaum ML, Commissariat PV. Psychosocial factors in medication adherence and diabetes self-management: implications for research and practice. Am Psychol. 2016;71(7):539-551. doi:10.1037/a0040388

35. Xie Z, Liu K, Or C, Chen J, Yan M, Wang H. An examination of the socio-demographic correlates of patient adherence to self-management behaviors and the mediating roles of health attitudes and self-efficacy among patients with coexisting type 2 diabetes and hypertension. BMC Public Health. 2020;20(1):1227. doi:10.1186/s12889-020-09274-4