This study was a retrospective cohort study. AARDEX Group’s database, the Medication Event Monitoring System (MEMS; AARDEX Group) Adherence Knowledge Center was used as the data source. This database contains anonymized electronic medication adherence data from ambulant participants enrolled in clinical studies that ran between 1989 and 2016. These participants’ medication adherence was electronically monitored using the MEMS. The following selection criteria were used: (1) once-daily medication with (2) a follow-up longer than 14 days. This second criterion was required because 14 days is the minimum duration needed to compute the weekly cross-correlation. These selection criteria matched data from 15,818 participants totaling 3,053,779 medication intakes.

For each participant, 2 objective habit metrics were computed. The first objective habit metric is the SD of the day-to-day hour of intake. The hours and minutes are extracted for each medication intake timestamp, irrespective of the date of intake, and the SD of the resulting list of hours and minutes is computed. The lower the SD, the more consistent a person is in the timing of their medication intakes. In the extreme case, when SD equals 0 h, all intakes occur at exactly the same time of the day.

The second objective habit metric is the “weekly cross-correlation” introduced by Phillips et al [11]. This metric compares each day of a given week with the corresponding day of the next week and quantifies whether intakes occurred around the same time on these 2 days. To do so, medication intakes are represented in a 2D matrix, denoted A, of dimensions 24 × N_days, with N_days being the number of follow-up days. The matrix is initially filled with zeros. Each medication intake occurring on day l at hour k is translated as a value of 1 for element A[k,l]. Then, each element A[k,l] of the matrix is multiplied by its matching element from the previous week, A[k,l – 7]. If intakes occur at the same time from week to week, the product will be close to 1. Otherwise, it will be close to zero. Finally, the weekly cross-correlation is equal to the sum of all element-wise products divided by the Euclidean norm of A. More details and code explaining how to compute these metrics are presented in Multimedia Appendix 1. As the pragmatic habit index incorporates adherence, it was not included in the present analysis.

Medication adherence, more precisely its implementation component [1], was quantified as the proportion of correct days, that is, the proportion of days with exactly 1 intake. A second implementation metric was used: the proportion of doses taken. The proportion of doses taken was computed as the ratio between the total number of doses taken and the prescribed number of doses. In this work, since participants were on a once-daily regimen, the prescribed number of doses was equal to the number of follow-up days. If a participant discontinued treatment too early, the analysis was limited to the period during which the participant was on treatment.

The hypotheses were that (1) SD of the day-to-day hour of intake would be negatively correlated to objective adherence (the more variable the hour of intake, the worse the adherence); (2) the weekly cross-correlation would be positively correlated to objective adherence (the more consistent the pattern of intakes, the better the adherence); and (3) the weekly cross-correlation would be more strongly correlated to objective adherence than the SD of the hour of intake.

All dosing history data were compiled while patients were engaged with the monitored medication in the trial. Therefore, noninitiation and nonpersistence are not part of the adherence evaluation in this research.

Continuous variables were reported using medians and first and third quartiles. Categorical variables were reported using counts and proportions. Pairwise correlations between continuous variables were assessed using the Spearman rank correlation coefficient. Correlation coefficients were converted using Fisher Z-transformation to obtain CIs [25,26]. Correlation coefficients were compared by performing a z-test on the difference between their Fisher Z-transformations [26,27]. The proportion of variance in adherence explained by habit was quantified using the Pearson correlation coefficient, squared [26]. Analyses were performed using Python 3 (Python Software Foundation) [28].

This work showed that medication-taking habit is positively correlated to medication adherence in a very large, cross-pathology population of participants enrolled in clinical studies. Medication-taking habit was quantified using 2 measures of the consistency in the timing of medication intakes: day-to-day and week-to-week. The implementation component of medication adherence was quantified using 2 measures: the proportion of correct days and the proportion of doses taken. The conclusion was the same irrespective of the habit or adherence measure: the stronger the habit, the better the medication adherence.

The results showed no clear superiority of any of the 2 objective habit metrics per the strength of their correlation with adherence. The correlation of objective habit metrics with self-reported habit metrics, which are the most frequently used habit metrics, might be used to differentiate the 2 objective habit metrics. Previous studies exploring this correlation are discussed in the next section.

The link between objective habit metrics and medication adherence has been investigated in very few other studies. The first one [11], performed by the authors on a much smaller dataset, reported very similar results: (1) a negative correlation between a measure of day-to-day variability and the proportion of correct days and (2) a positive correlation between the weekly cross-correlation and the proportion of correct days. The second one [13], previously discussed, used a habit metric incorporating adherence and reported an expected correlation between these 2 quantities.

Over 700 determinants of adherence have been reported [29]. Among all these determinants, objective habit strength seems to be an important one, as it explained over 30% of the variance in medication adherence in this work. This observation empirically justifies studying objective medication-taking habit.

A question related to the present work is whether people use time-based cues for taking their medications [10]. To answer this question, the proportion of participants having 95% of their intakes in a 1-hour window was computed. Only 3.6% of participants met this criterion. In a previous study [10], 21.7% of participants reported taking their medication at a specific time of the day. These numbers mean that a minority of patients rely on cues triggered by the clock and take medication at the same time every day. This finding might explain why reminders for medication adherence sent at a specific time of the day can be ineffective [30].

Few studies have studied how well objective habit metrics for medication adherence correlate with conventional self-reported habit metrics. SD or other metrics of day-to-day consistency were found to correlate with self-reported habit in 2 studies [17,18], but not in a third one [11]. Weekly cross-correlation was found to correlate with self-reported habits in 1 study [11].

First, the 2 objective habit metrics used in this study are measures of consistency in timing, day-to-day, or week-to-week. However, not all habits translate into consistency in timing, and consistency in timing can originate from other factors than habit [9]. In such cases, the objective habit metrics will not reliably quantify the presence or absence of habit. On the other hand, the objective consistency metrics have the advantage that they do not endure biases associated with self-reported habit metrics, such as recall bias.

Second, the dataset used in this paper was very large, so the findings are likely to be generalizable to the whole population of participants enrolled in clinical trials. On the other hand, no conclusion about a specific disease, population, or time can be drawn, because there would be a large confounding influence of the underlying study design and specific medication characteristics. The main result is that, in general, adherence is related to objective habit.

Third, in this study, habit and medication adherence were computed over a participant’s whole follow-up, until treatment discontinuation. However, for a single participant, habit strength may have changed throughout the duration of this study, for instance, because of the introduction of a new device for medication adherence monitoring. The evidence on the effects that electronic adherence monitoring causes on adherence itself is unclear [22,31]. The effects of electronic adherence monitoring on medication-taking habit have never been studied, to the best of the authors’ knowledge.

Another reason that may have caused individual changes in habit strength is that some of the studies analyzed were testing habit-building interventions. In such settings, the objective habit metrics used in this paper can be computed over shorter periods to capture the dynamics of habit formation and maintenance. For instance, in the case of a specific habit-building intervention, Pironet et al [32] observed that habit strength increased after a 6-month intervention and remained stable 6 months after the intervention was stopped.

If a patient’s implementation adherence is measured electronically, their habit can be computed continuously from the recorded data and serve as a support for an intervention, in-person or through a mobile app. The intervention could focus on providing support to build a better medication-taking habit. An example of such a mobile intervention is presented in the study by Stawarz et al [33].

This work investigated the relationship between habit and the implementation component of medication adherence. Other, evenly important questions are whether objective habit indices predict the quality of implementation over the longer term, and if objective habit indices predict early treatment discontinuation.

In this study, the correlation between objective habit metrics, computed from electronic medication intake data, and objective medication adherence was assessed in a large database of participants enrolled in past clinical trials. The 2 objective habit metrics tested in this work were correlated to the 2 objective adherence metrics. These 4 pairwise correlations all imply that a lower variability in the pattern of medication intakes, be it day-to-day or week-to-week, is correlated with higher medication adherence. However, no variability is not the norm.

Objective habit metrics allow us to better identify patients who might benefit from habit-building support. If a patient uses electronic monitoring, the habit metrics can even be automated and computed in real time, directly reflecting habit changes and allowing for timely intervention. In addition, some effective interventions for medication adherence are based on developing or improving habits [5]. In such settings, electronic monitoring can be used to assess the effect of the habit-building intervention in real time [32].