Design

We constructed a population-based retrospective cohort of all patients discharged alive after an emergency admission for CHD to any VHA hospital in 2008. Patients were followed throughout the health information systems for 9 months after hospital discharge to assess their adherence patterns over time to four therapeutic groups: antiplatelet agents, beta-blockers, drugs acting on the renin-angiotensin system (angiotensin-converting enzyme inhibitors-ACEI- or angiotensin receptor blockers-ARB), and statins.

Setting

The study was conducted in the Valencia Community, an autonomous region in Spain with 5 million inhabitants in 2008. The Valencia Health Agency covers about 97% of the regional population. During the study period, medical care was free of charge with coverage extending to substantial pharmaceutical benefits: all medicines prescribed to pensioners (eligible because of either age or disability) and underprivileged groups were free of charge. The remaining population paid 40% of the cost of medication (i.e. antiplatelet agents, statins, ACEI or ARB in fixed-dose combinations), but only 10% of the cost with a ceiling of €2.45 (≈$3 USD) per package for some long-term treatments (i.e. beta-blockers, ACEI or ARB).

Population

The selection and characteristics of the general cohort have been fully described elsewhere [18]. In brief, all patients of both sexes aged 35 years and over, admitted through the Emergency Department and discharged in any of the 24 VHA hospitals with a main diagnosis of acute coronary syndrome (International Classification of Diseases 9th revision Clinical Modification, ICD9CM = 410.xx and 411.xx) or other forms of coronary heart disease (413.xx and 414.xx) between 1 January and 31 December 2008 were included. We excluded deaths in the 30 days following hospital discharge (n = 111), duplicate cases (if the patient had more than one CHD admission, only the first one was accounted for), some government employees whose prescriptions are reimbursed by civil service insurance mutualities not included in the pharmacy databases of the VHA (n = 65), and patients not registered in the municipal census, who left the region or who were discontinued from VHA coverage for other causes (n = 231), because of limitations on follow-up. We also excluded patients who had no contact (no visits, prescriptions, or any other contact) with the VHA within the 9 months after hospital discharge (n = 635) because we suspect that most of those patients could be temporary residents with limitations on follow-up. From the resulting cohort of patients (n = 7462) we created 4 subcohorts with at least one prescription written (regardless of whether the patient filled the prescription or not) from each of the four therapeutic groups within the first 3 months after discharge.

Data Sources

The main source of data was the VHA ambulatory electronic medical record for ambulatory care, the so-called ABUCASIS system, which among other information includes demographic and clinical data (e.g. diagnoses, personal and family history, lifestyle habits, lab tests/results, etc.), and information on physician prescriptions (what the doctor prescribed) and dispensations from pharmacy claims (what the patient fills from the pharmacy). The information on hospitalizations was based on the Minimum Basic Dataset (MBDS) at hospital discharge, a synopsis of clinical and administrative information on all hospital discharges, including diagnoses and procedures (all electronic health systems in the VHA use the ICD9CM). Mortality, VHA coverage (including dates and causes of VHA discharge) and some demographic characteristics (e.g. age, gender, country of origin, copayment status, etc.) were obtained from the Population Information System.

Main endpoint

The main outcome measure was monthly adherence to four therapeutic groups (antiplatelet agents, including acetylsalicylic acid (ASA) at doses of 100 mg and clopidogrel, beta-blockers, ACEI/ARB, and statins) based on pharmacy claims (dispensations) according to an ascertainment period of 9 months after hospitalization for CHD. The index date was defined as the date of discharge. We evaluated adherence by creating a “supply diary” for the 270 days (9 months) after the index date. The supply diary was calculated by linking all pharmacy claims based on the dispensing date and the days’ supply [7]. The days’ supply was estimated using the dosing schedule specified in the prescription (one tablet every 8, 12 or 24 hours) and the number of pills per package/prescribed. When this did not agree with standard dosing (around 1% of cases) we rounded up/down the dosing prescribed to the closest frequency (8, 12 or 24 h). Overall, in the Valencia Health System, most of the prescriptions provide 30 days’ supply (around 97.5% of prescriptions provide 30 days supply or less, and the rest provide 60 days supply). If a dispensing occurs before the previous dispensing should have run out, we assume that the new dispensing begins the day after the end of the old dispensing, and days with a drug supply are accumulated. Monthly adherence was defined as having ≥24 days covered out of 30 (≥80% covered), leading to a repeated binary outcome measure at 9 time periods for each drug group and patient, there being 512 theoretically possible adherence patterns. In case of dual antiplatelet therapy, we used a non-restrictive approach, considering patients as adherent to antiplatelet agents if they were adherent to at least one of them (ASA or clopidogrel). Hospitalization days were not taken into account to estimate adherence, as they have been shown to have little impact on adherence to these medications in patients after a myocardial infarction, and are expected to have lower impact in short follow-up periods. [19]

Covariates

We searched socio-demographic and clinical data (inpatient and outpatient) from the various electronic datasets. We identified the following variables at the time of discharge: age at hospital admission, main admission diagnosis (acute myocardial infarction, angina pectoris, other acute coronary heart disease and other chronic coronary heart disease), gender, country of birth (coded as Spain or other), copayment status, and presence of chronic conditions (hypertension, hyperlipidemia, diabetes, smoking, arrhythmias, congestive heart failure, chronic obstructive pulmonary disease (COPD), chronic renal failure, peripheral vascular disease, cerebrovascular disease, dementia and cancer).

Ethics

The study was approved by the Ethics and Clinical Trials Committee of the Public Health General Directorate and the Center for Public Health Research. All patient data were transferred to the research team anonymized and de-identified prior to analysis. The Regulatory Commission of Access to Ambulatory Care Information Access of the Valencia Health Authority approved the cession of anonymised data.

Analysis

First, socio-demographic and clinical characteristics are shown as percentages for the four therapeutic subcohorts. Second, adherence patterns to essential medications after hospitalization for CHD over time were identified using group-based trajectory modeling. Monthly trajectories were modeled with quadratic polynomial functions of time. Each subpopulation is characterized by an intercept, a lineal and quadratic slope allowing for curved developmental patterns. Variance and covariances of these parameters are fixed at zero. The heterogeneity (i.e. the distribution of individual differences within the data) is summarized by a finite set of unique polynomial functions, each corresponding to a discrete trajectory [20].

In the modeling procedure, seven trajectory models were estimated and compared in a stepwise procedure. The selection of the number of trajectory groups or latent classes that best represents the heterogeneity in adherence was based on: 1) model fit indices: Bayesian information criteria (BIC) [21], Akaike information criterion (AIC) [22] where a lower index value indicates a better model fit, and the Lo-Mendell-Rubin likelihood ratio test (LMR-LRT) [23], which tests the improvement in fit between neighboring class models (comparing k-1 and the k class models). LMR-LRT test provides a p-value that can be used to determine if there is a statistically improvement in fit for the inclusion of one more class. A p-value >0.05 was used to reject a new class model; 2) a minimum proportion of the study sample in a class: 5%; and 3) entropy, which indicates uncertainty in the classification of the model and is a measure of how well—or how precisely—study participants are classified into their most likely class. The cut-off value used was a probability of ≥0.7. In summary, the criteria for rejecting k class models (and, thus, selecting k-1 class models) was the presence of some of the following criteria: BIC or AIC score higher; LMR-LRT p-value>0.5; entropy (minimum membership probability) <0.7; and minimum sample size <5%. Once the final number of latent classes was decided, individuals were classified into their most likely class (i.e. the one with the highest membership probability) resulting in an observed (categorical) variable denoting class membership. Patients who died were censored at the date of death; from that date on, they had missing values, contributing to the trajectories and being classified into the different groups based on their adherence behavior during the period they were alive. As a sensitivity analysis, the 9-month PDC (as continuous and binary measure) has been described for each trajectory or latent class in each of the therapeutic subcohorts.

Additionally, we assessed trajectories of adherence to three or more index medications after hospitalization for CHD over time using group-based trajectory modeling as described above. Also, a multiple correspondence analysis (MCA) was used to explore the relationship between the adherence trajectories of the different medications. MCA allows the detection and graphical representation of the underlying structure (dimensions) in a given data set. Interpretation of the results provided by MCA is done based on the graphics; the relative position of the category points indicates the level of association between the categories. The closer the points are, stronger is the relationship between the categories.

Finally, we studied, for each therapeutic group, possible predictors of latent class membership (i.e. factors that characterize the adherence trajectories from each other) using multinomial logistic regression analyses. Their coefficients are the increase in the log-odds of being in a particular trajectory group, versus the nearly-always adherence group, for a one-unit increase in the predictor variable. These were transformed to odds ratios (OR) and their respective 95% confidence intervals. All the analyses were made using Mplus version 6.12 [24] and R [25].

Limitations

Some potential limitations of this study should be addressed. Regarding the study design, the possible sources of bias include: 1) our study has evaluated adherence in patients with at least one prescription in the 90 days after hospital discharge, thus not taking into account that for some patients the doctor could have suspended the prescription afterwards, due to side effects or for other reasons; 2) we used pharmacy claims for measuring adherence, but patients do not necessarily consume all the drugs they fill. Nevertheless, several studies have shown a high consistency between dispensation and patient consumption [29,30]; 3) ASA is an over-the-counter and low cost drug that could be acquired without prescription and thus could not be reliably captured in pharmacy claims; 4) dispensations obtained from pharmacies outside the Valencia region (e.g. on holidays outside the region) or during hospitalization (drugs are provided at no cost by the Hospital Pharmacy Service) have not been recorded in our study. 5) Refills previous to the index admission and hospitalization days after the index event were not taken into account, however, the latter have been shown that have little impact on adherence in post-MI patients to these medications, even in short follow-up periods [19]. Some of these design problems tend to overestimate nonadherence while others tend to underestimate it, but overall no major impact is expected on the identification of adherence patterns.

Regarding the analysis, some limitations should be mentioned with respect to latent class models [31–34]: 1) the modeling process includes several choices and each choice influences the final number of classes and/or their characteristics. For example, if the within-class variance parameters are fixed at zero, as in GBTM, then the final number of classes will often be larger than in models allowing freely estimated within-class variance parameters. In the same way, several features of the study design, such as sample size and the number of measurements, could also influence the number and characteristics of classes identified in the final model [23,35,36]. Moreover, the inconsistency of the model fit indices adds extra challenges in determining the final number of classes. Although simulation studies [37] suggest the relative advantage of some model fit indices, clear guidelines for building an appropriate latent class model are lacking; 2) the identification of distinct classes is no proof of the true existence of multiple subpopulations [20,23,38,39]. As in all statistical models, the final model is just a simplification of a complex reality, an issue that should be kept in mind when conducting latent class models.

Regarding adherence predictors, as stated previously, information from administrative databases and ambulatory electronic medical records do not include some of the variables that could influence adherence (i.e. a patient’s motivation or certain side effects such as fatigue or sexual dysfunction), and other relevant covariates such as procedures during the index event or prior ischemic events were not available. However, we were able to identify some important characteristics related to nonadherence. Finally, regarding the external validity of our findings, the present study was conducted in a specific region. Given the potential differences with other regions or countries, the generalization of our findings to other settings should be made prudently, as should the generalization to other drugs or types of patients.

Conclusions

We were able to identify, through group-based trajectory models, distinct adherence patterns over time beyond the classic approaches, which often overlook the dynamic nature of adherence. We also identified some predictors of these patterns, which could enable clinicians and researchers to design more appropriate and effective improvement interventions in patients with poor or intermediate adherence patterns.