Data for this analysis were provided by Calm, a popular meditation app that had >2 million paying subscribers at the time of data collection. Calm provided a sample of 14,958 randomly selected users who created their account in 2017 and used the app at least 8 times over the first 2 months after their account was created. This last criterion eliminated participants who did not establish an average of at least one use per week in the time immediately after their account was created. The sample was selected such that the timing during which participants last used the app was spread evenly over the 2 years after account creation. Sample demographics were not provided other than estimates of age and gender made via a Calm proprietary prediction algorithm, which were not used in our analyses.

The Chapman University Institutional Review Board provided an exemption from review (IRB-21-88) for this project.

Table 1 illustrates summary statistics for the variables used in subsequent analyses, with all measures averaged over each of the 14,958 participants. The highest proportion of meditation sessions occurred during the morning and evening TOD windows, and entropy was relatively high, meaning that participants tended to vary the time window during which their meditation occurred. Participants averaged 11 meditation sessions during the sixth and seventh months of their enrollment and had their last meditation session 656 days after their account was created.

Figure 1 illustrates the results of the GAMMs that estimated the probability over time of meditating on a given day. For all participants, there was just over a 50% probability of meditating immediately after creating their account, with this value quickly decreasing over time. This pattern was consistent for participants in the bottom quartile of the M67 and LS outcomes and in the top quartile of the LS outcome, albeit with a slower decay in meditation probability for the latter group. In contrast, the probability of daily meditation for participants in the top quartile of the M67 outcome was relatively constant over time. Notably, for all participants, including those in the top and bottom quartiles of the outcome measures, the probability of meditating immediately after creating the account was relatively consistent (between 50% and 60%) for the earliest days of enrollment.

Figure 1 also illustrates differences in meditation probability according to TOD window. For participants in the top quartile of both outcome measures, the probability of meditating was highest in the morning, followed by the evening, late night, and midday. Similar results were seen for participants in the bottom quartile of the outcomes, but with minimal differences between morning and evening probabilities and a convergence of probabilities for all TOD windows to approximately zero when examining later enrollment. The latter effect is expected for participants with relatively poor outcomes. Table 2 quantifies these effects by showing the percent change in probability from d=1 to d=180 for all outcomes and participants. The percent change for participants in the top quartile on both measures was considerably smaller for morning and late-night meditation than for meditation that occurred during midday or evening. For example, morning and late-night reductions were 8.4% and 2.3%, respectively, for M67 compared with 33.5% and 21.8%, respectively, for midday and evening meditation. In other words, participants who meditated at a higher frequency were more persistent in their morning and late-night meditation than they were at other TODs. A similar result was found for the LS outcome, but differing probability reductions by TOD were not observed for participants in the bottom quartile of either outcome.

Table 3 summarizes the results of multiple regression models used to assess the relationships between the number of meditation sessions, temporal consistency, and TOD preference on both outcomes. In the table, the morning, midday, evening, and late night columns each represent a single regression of the form , where is the M67 or LS outcome, n is the number of meditation sessions, H is the entropy, and T is 1 of the 4 TODs. As expected, there was a significant positive relationship between the number of meditation sessions over the first 180 days and both outcome measures, such that each additional meditation session was associated with approximately 0.4 more meditation sessions during months 6 and 7 and the last observed meditation session occurring approximately 4 days later. The effect of temporal consistency was reversed for the 2 outcomes, with more temporal consistency (ie, a lower value for H) producing better outcomes for M67 and worse outcomes for LS (both associations were statistically significant). A greater proportion of meditation in the morning was significantly associated with better short- and long-term outcomes. For example, a 10% increase in the proportion of meditation performed in the morning was associated with 0.28 more meditation sessions during months 6 and 7 and the last observed meditation session occurring approximately 5.5 days later. The opposite was true for late-night meditation, where a greater proportion of these sessions was associated with poorer meditation outcomes. The standardized regression coefficients indicate that for both outcomes, the associations with temporal consistency and TOD preferences had relatively similar magnitudes. However, the effect of the number of sessions was roughly 10 times as large for the M67 outcome and 4 to 5 times as large for the LS outcome, relative to temporal consistency and TOD preferences.

Based on the total sum of squared error (Figure 2 and Table 4), participants’ proportion of morning meditation was stratified into 4 clusters. Clusters 1 and 2 indicate a high proportion of morning meditation at the outset of enrollment, with this level being sustained over time in cluster 1 and decreasing with time in cluster 2. Cluster 3 represents a consistent, moderate rate of meditation over time, while cluster 4 is indicative of a low level of meditation throughout the observation period. The clusters identified in the other TOD windows were qualitatively similar to those shown in Figure 2. Table 4 shows the number of individuals assigned to each cluster for each TOD. Cluster 4 was by far the largest cluster, with other clusters containing roughly 5 to 10 times fewer participants.

Because clusters are consistent across TODs, a 4-component TOD trajectory profile can be built to summarize each participant’s TOD preferences over time. We denote each profile as m-d-e-l, where {m,d,e,l}∈{1..4} corresponds to one of the 4 cluster groups shown in Figure 2 for morning, midday, evening, and late night, respectively. For example, a 1-4-4-3 cluster profile corresponds to a participant who had a consistently high proportion of morning meditation, low levels of midday and evening meditation, and a consistently moderate proportion of late-night meditation. Table 5 illustrates the 10 TOD trajectory profiles (minimum 50 participants per profile) with the highest average value for both outcomes, which represents 1163 participants for the M67 metric and 1012 participants for the LS metric. For each TOD window, we also calculated the proportion of each cluster included in these top-performing profiles (Table 6). Notably, cluster 2 was severely underrepresented, indicating that participants who began with a high proportion of meditation within any TOD window and then dramatically dropped off did not fare well for either outcome measure. Additionally, for every outcome/TOD combination, a considerably higher proportion of participants had TOD trajectory profiles characterized by cluster 3 rather than cluster 1, indicating that sustaining a moderate frequency of meditation at a given TOD led to better outcomes than sustaining a high frequency.

In this paper, we examined the effects of TOD preferences and temporal consistency on both short-term and long-term indicators of sustaining an mHealth meditation practice. This analysis was undertaken to assess the appropriateness of common advice for establishing a meditation habit, namely, to meditate at the same time every day, preferably in the morning.

Each of the 3 statistical analyses presented herein supports the conclusion that meditating in the morning is beneficial for sustaining engagement with the mHealth app. First, we observed that the decrease in the probability of meditating over the first 180 days after creating an account was statistically smaller for morning meditation sessions (63.4% reduction) than for other times of day (reductions ranged from 67.8% to 74.5%). This was especially true for participants whose outcome measures were in the top quartile (ie, > the 75th percentile) of all measures. Second, when regressing both short-term and long-term outcome variables on TOD proportion, a 10% increase in the proportion of all meditation sessions performed in the morning was associated with 0.28 more meditation sessions during months 6 and 7 and the last observed meditation session occurring approximately 5.5 days later. Similar-magnitude decreases in these 2 outcomes were associated with increases in the proportion of meditation sessions performed in the late night, and no effect was found for the proportion of meditation sessions performed in midday or evening. Lastly, the positive impact of morning meditation was also seen in the cluster analysis, where the top 10 most-observed cluster profiles contained fewer instances of consistent low levels of meditation (cluster 4) and greater instances of sustained meditation (clusters 1 and 2) for the morning TOD window than for the other windows, particularly for the M67 outcome. This consistent finding across multiple analyses supports the commonly provided advice of establishing a consistent meditation practice in the morning and suggests that future meditation interventions should target morning meditation for helping users attain the corresponding health benefits.

Further, our findings show that a high intensity of meditation practice at the start of using the app was generally not sustainable. This can be seen by the lack of cluster 1 representation in the TOD trajectory profiles in Table 6, which indicates that people with the best short- and long-term outcomes were not able to sustain an exceptionally high level of meditation. Instead, they either sustained a more moderate probability of meditating in a given TOD window (cluster 3) or began with a high probability but then decayed to a probability near zero (cluster 2). Among users that started with high intensity (cluster 1), doing so in the morning was more sustainable than any other TOD. In other words, initial high intensity did not prove to be a good method for sustained app meditation practice, but if one were to meditate zealously, it would be better to do so in the mornings.

These findings should be expanded upon in future work by using ecological momentary assessments or environmental sensors to measure contextual cues that will help to better determine if and when meditation has become habitual.

There were several limitations to this study. First, generalizability may be an issue given our limited access to demographic and other contextual information, and our findings may not extend to other behaviors, since app-based meditators generally tend to be female with low racial diversity [28,50]. Second, our findings do not account for any meditation sessions not performed with the app, which could bias our results if more experienced meditators eventually find the app to be extraneous and unnecessary. Third, we do not account for other app features (eg, Sleep Stories) that may impact participants’ app usage. Lastly, we use temporal consistency as a proxy for routine app usage, but not all routines and habits will be performed at approximately the same time each day, as they may be contingent upon a time-varying cue (eg, meditating after a work shift that varies from day to day).

One possible explanation for the success of morning meditations could be the influence of the circadian rhythm of cortisol. In attempting to understand the development of healthy habits at various points throughout a circadian cycle, Fournier et al [36] found that morning practitioners were the quickest to habituate a targeted behavior. Further, cortisol levels played a significant role in determining the time for the healthy behavior to become a habit, with cortisol levels being higher in the morning than in the evening. According to Adan et al [51], preferences toward activity at certain times in the day indicate a circadian typology, and those with a preference for morning activity (ie, a morning typology) tend to exhibit healthier overall habits. This is concurrent with findings from Taylor et al [52], who observed better academic performance in college students with morning typology preferences. A suggested (but not yet tested) explanation for this phenomenon is that contexts and behaviors in the morning are more stable and, therefore, more conducive to habit development [35]. Beyond the human experience, even snails have been shown to learn more efficiently in the morning [53]. While our findings might support the conclusion that habit formation is best for those with a morning typology, TOD preferences should not be the only indicator, as temporal consistency should also be taken into account.

Contrary to popular advice to meditate at the same time every day, this study indicates that temporal consistency in meditation timing was inversely related to a long-term meditation practice, although it was beneficial for short-term outcomes. This finding is consistent with recent work indicating that flexibility, rather than routine, led to better exercise outcomes. Beshears et al [40] found that incentivizing gym attendance only during a participant-selected 2-hour window led to less gym attendance than when participants were incentivized to attend at any time. The researchers hypothesized that this finding could be due to a bias in their sample, which was composed of technology managers who likely had less routine built into their day. However, it is notable that only 22% of participants selected a 2-hour window in the morning. Therefore, the poorer outcomes for routine incentives may reflect the suboptimal performance of nonmorning TOD preferences, as discussed above. Additionally, Berardi et al [41] found that stochastic exercise performance over the first 100 days of an intervention was associated with increased exercise throughout the remainder of the 1-year trial. Meta-flexibility was hypothesized as a potential mechanism for this finding, whereby participants who had the resilience to adjust their behavior in response to changing life circumstances were ultimately more successful at sustaining their exercise behavior.

In conclusion, we have used novel analytic techniques to demonstrate that persistence in the use of an mHealth meditation app is most successful when meditation primarily occurs in the morning and when participants are not overly rigid to the point that no meditation occurs at other times of the day. These findings can be used by practitioners, researchers, and policy makers to develop behavioral interventions for meditation and other health-promoting behaviors that have the greatest chance of establishing healthy habits, thereby improving public health.