Participants were recruited via online research panels operated by third-party research companies in Japan. The target population included men and women aged 20 to 60 years living in Japan who were registered with these research companies and were invited as part of an on-site experiment measuring brain activity. All participants had previously completed an online preliminary survey, and recruitment was designed to balance gender and age distributions as much as possible. Individuals who met the exclusion criteria were not recruited (see Table S1 in Multimedia Appendix 1). All 120 individuals who consented to participate were informed about the study procedures and provided consent for the onsite survey. The analysis ultimately included 100 participants (mean age 44.0, SD 8.8 y) who consented to the daily life survey and for whom online preliminary survey data were available. Participants who did not participate in the daily life survey or who lacked online preliminary survey data were excluded. The daily life surveys were conducted between November 29, 2023, and April 15, 2024.

This exploratory study was not designed to test a specific effect size; therefore, no a priori sample size calculation was performed. We pragmatically sampled approximately 120 adults. To assess robustness, we conducted sensitivity analyses and internal validation via a cluster bootstrap at the participant level (see Methods, Internal Validity Assessment and Sensitivity Analyses).

An online screening survey was conducted before enrollment, and participants answered all 13 questions from the SoAS (SoAS-13) [22,23], the full set of items in this scale to measure GenExp-SoA, and questions on occupation, exercise habits (defined as exercising more than 1 h/wk), and Big Five personality traits (extraversion, agreeableness, conscientiousness, neuroticism, and openness) [34,35]. Age and gender information was obtained from the registration records of the third-party research companies.

Participants received instructions before the survey and installed the custom-built app developed by our team on their smartphones. The app consisted of the following tasks, which were conducted in sequence as a single set of surveys: (1) the SoAS questionnaire, (2) the Libet clock task, (3) a visual analog scale for 7 emotions, (4) the cognitive task, and (5) Russell’s circumplex model plot. Despite collecting data from all parts, only the data from parts (1) and (2) were analyzed in this study. Data from parts (3) through (5) were obtained for other research purposes and were not included in the present analysis. The participants were asked to complete each survey daily as it was delivered to their smartphones for 3 months starting from a specified date. During the experimental period, the surveys were delivered randomly 3 times a day between waking and sleeping hours. The participants were asked to complete these surveys at least once a day within one hour of delivery. Participant compensation was based on the app-based survey recording rate, as described in the Ethical Considerations section.

Although 13 questions from the SoAS were used in the preliminary survey, the number of items was reduced in the mobile app-based survey to minimize participant burden from completing tasks 3 times daily. Following Legaspi et al [33], three items assessing positive agency were selected and validated from the SoAS: (1) “I am in full control of what I do”; (2) “My behavior is planned by me from the very beginning to the very end”; and (3) “I am completely responsible for everything that results from my actions.” Participants responded to these items on the app using a 15-point visual analog scale ranging from “strongly agree” (15) to “strongly disagree” (1), with a default value of 8. The GenExp-SoA in this study was evaluated as the total score of these 3 questions per participant (range: 3‐45 points). The internal consistency of the 3 items was excellent, with a Cronbach α of 0.92 (95% CI 0.916 to 0.921) calculated across all the responses. The correlations of each item with the total score ranged from 0.90 to 0.95, indicating high reliability (Table S2 in Multimedia Appendix 1).

In this study, the Libet clock displayed a clock-like dial on the app screen with 24 ticks and numbers every 2 ticks, along with a dot rotating clockwise in a 2560 ms cycle. The participants performed 2 sensory tasks using the Libet clock method: one without a behavioral task (baseline) and one with a task (operant). In these tasks, participants entered the time displayed at the moment the color of the Libet clock changed (Figure 1). Effect binding, calculated by subtracting the operant outcome judgment error from the baseline outcome judgment error, was used as the measure of LocalImp-SoA (range −30 to 30 points). The differences were calculated as signed shortest circular differences on the clock scale. Therefore, the following formula was used: (1) baseline outcome judgment error=baseline true time−baseline participant recorded value; (2) operant outcome judgment error=operant true time−operant participant recorded value; and (3) effect binding=baseline outcome judgment error−operant outcome judgment error. A positive effect binding indicates shortened time perception due to the behavioral task, suggesting an SoA, whereas a negative effect binding indicates decreased SoA.

Regarding baseline session, (1) first, participants were instructed to enter the time in numerical values when the clock’s color changed. (2) The dot began rotating without the participant’s operation. (3) The color of the clock changed during the rotation of the dot. (4) Participants recorded the time at the moment the clock color changed. Regarding operant session, (1) participants were instructed to click the start button to start the clock and then enter the numerical value when the clock color changed. (2) The dot began rotating immediately after the start button was clicked. (3) The color of the clock changed during the rotation of the dot. (4) The participants recorded the time at the moment the clock color changed. Thus, the major difference between the 2 sessions is whether the clock rotation begins without participant action (baseline session) or is triggered by participant action (operant session).

To promote data quality and consistency, several procedures were implemented.

Task delivery times in the app were customized by each participant to align with their daily routines, thereby enhancing compliance and ecological validity. During data collection, the research team monitored the participants’ app usage and survey completion rates. The team provided technical support and reminders to participants who had markedly low response rates or experienced other issues. All data entries were automatically assigned timestamps to track response timing and identify irregularities. In the analytic dataset, 39 participants had observation windows (first to last completed day +1) longer than 90 days (maximum 116 d), reflecting missed days and occasional app reinstalls or resets. Days with ≥4 completed survey sets occurred in 535 of 6210 (8.62%) participant-day observations with at least one completed survey. These issues can occur in routine smartphone use and were treated as part of real-world use conditions. For the measurements of SoA, outliers were defined as observations exceeding ±3 SD from the pooled mean across all participants and days and were excluded. In total, 71 (0.5%) observations were excluded for GenExp-SoA, and 352 (2.5%) observations were excluded for LocalImp-SoA as outliers. The analytic dataset comprised 13,391 paired observations in which both measures were available, and all analyses used these paired records.

Missing data occurred across days and within days. No variables were completely missing across all time points. Missing values were not imputed.

This research received ethical approval from the Ethics Committee of the National Institute of Information and Communications Technology (N230192300). All study procedures involving human participants were conducted in accordance with the principles outlined in the Declaration of Helsinki and its subsequent revisions. Before participation, individuals received a detailed explanation of this study’s purpose and procedures, and written informed consent was obtained using documents approved by the ethics committee. All the participants voluntarily agreed to participate in this study after fully understanding the information provided.

Participant records were pseudonymized using hashed IDs, and the analytic dataset contained no direct personal identifiers. Data were stored and shared via access-controlled research servers with IP-based access restrictions and 2-factor authentication.

Participants received cash-equivalent points based on their app-based survey recording rate, according to a prespecified stepwise schedule approved by the ethics committee and explained before enrollment. Compensation ranged from JP ¥10,000 to JP ¥15,000 per participant (approximately US $67 to $100), with the minimum provided for a recording rate of ≤30% and the maximum for >90% to ≤100%.

Table 1 presents the basic characteristics of the participants, with the main values summarized. Compared with the lower quartiles (Q1 and Q2) of GenExp-SoA as measured using the 13-item SoAS, participants in the higher quartiles (Q3 and Q4) were older and had higher scores for extraversion, agreeableness, conscientiousness, and openness, with lower neuroticism. A greater proportion of males was observed in Q1. These are descriptive comparisons without statistical testing. The corresponding 95% CIs are reported in Table S3 in Multimedia Appendix 1.

Table 2 shows the measurement characteristics of the 2 forms of the SoA in daily life. We also report completion-based adherence metrics in Table 2, including the response day rate and the number of completed survey sets per observation day. The corresponding 95% CIs are provided in Table S4 in Multimedia Appendix 1. The average number of measurements per day increased in the latter half of the survey period, whereas the daily average GenExp-SoA decreased. The daily average baseline outcome judgment error and operant outcome judgment error decreased in the latter half of the survey period; however, no clear change was observed for LocalImp-SoA at the group level. The within-day variance in the daily average of the explicit measure was relatively stable across periods, whereas the between-day variance appeared higher in the latter half of the survey period. On the other hand, no clear change was observed in either the within-day or between-day variance for the implicit measure. To visualize measurement frequency over time, we present the mean number of completed survey sets per participant per day in Figure S3 in Multimedia Appendix 1.

Table 3 shows longitudinal changes and within-individual and between-individual variances in GenExp-SoA according to individual characteristics and measurement timing on the basis of the main mixed-effects model analysis. The fixed-effect coefficients for elapsed days represent the expected change per 1 day in the original scale (possible range: 3‐45 points). The bootstrap estimates and 95% CIs for internal validation are presented in Tables S5 and S6 in Multimedia Appendix 1. In the main analysis, elapsed days were associated with a decrease in GenExp-SoA (β=−0.008, P<.001), with more pronounced decreases among women (β=−0.019, P<.001) and among those with extraversion scores higher than the median (β=−0.019, P<.001), while those aged 50‐59 years showed an increase (β=0.023, P<.001). In the bootstrap analysis, these effect sizes were generally consistent, and the 95% CIs for women (β=−0.020, 95% CI −0.041 to 0.003) and for those aged 50‐59 years (β=0.024, 95% CI 0.007 to 0.039) were consistent with the main results. For the other subgroups, no substantial associations were observed. In a sensitivity analysis of the primary mixed-effects model, replacing the linear time term with a natural spline (ns (time, df=3)) under the same random intercept structure, the model fit did not improve (P=.80). Variance component estimates were similar between analyses. The within-individual variance was 20.2, which was greater on Sundays (24.2) and lower among those aged 50‐59 years (14.5) and during the time period of 12 AM to 5:59 AM (16.8). The between-individual variance was 66.3, which was higher among women (73.2) and those with extraversion scores below the median (83.0) and lower among those with openness scores higher than the median (49.1).

Similar analyses for LocalImp-SoA are summarized in Table 4. The fixed-effect coefficients for elapsed days represent the expected change per 1 day in the original scale (possible range −30 to 30). The bootstrap estimates and 95% CIs for internal validation are presented in Tables S7 and S8 in Multimedia Appendix 1. In the main analysis, elapsed days were associated with a decrease in LocalImp-SoA (β=−0.003, P<.05). Similarly, replacing the linear time term with a natural spline (ns (time, df=3)) did not improve model fit for LocalImp-SoA (P=.14). Among the Big Five traits, participants whose agreeableness scores were higher than the median showed a greater decrease (β=−0.005, P<.01). During the time period of 3:00 PM to 5:59 PM, a relatively large decrease was observed (β=−0.014, P=.06, P-interaction=.07). In the bootstrap analysis, these associations were generally consistent. The 95% CI for all participants (β=−0.003, 95% CI −0.005 to −0.0003), for women (β=−0.004, 95% CI −0.008 to −0.0005), and for those with agreeableness scores above the median (β=−0.005, 95% CI −0.009 to −0.002) were consistent with the main analysis. The other subgroups did not show notable associations. Variance component estimates showed similar patterns in both analyses. The within-individual variance was 8.7, which was greater among those aged 50‐59 years (12.2) and the time period 3:00 PM to 5:59 PM (12.7) and lower among those aged 20‐39 years (6.7). The between-individual variance of the implicit measure was 0.14.

Table 5 shows the results of fitting LocalImp-SoA to GenExp-SoA using a linear mixed model. No significant associations were observed in either the crude model or the model adjusted for multiple variables. Bootstrap analysis also revealed no association between these parameters (Table S9 in Multimedia Appendix 1). In sensitivity analyses, no association was observed when all observations were used without outlier exclusion or when random intercept and random slope models were applied to the primary dataset. Similarly, no association was observed when the implicit measure was classified into 3 qualitative groups or when analyses were stratified by the median of the total value of baseline and operant outcome judgment errors (Table S10 in Multimedia Appendix 1). The generalized additive mixed model did not clearly support the average nonlinear association (Figure S1 in Multimedia Appendix 1). The slopes specific to each participant were better (ΔAIC [Akaike information criterion]=29), even though the average smoothing was not significantly curvilinear (P=.14). Comprehensive details are provided in Tables S11-S13 in Multimedia Appendix 1.

The estimated population mean values of each parameter for the latent variables of GenExp-SoA (eta1) and LocalImp-SoA (eta2) obtained from continuous-time structural equation modeling are shown in Figure 2. The autoregressive coefficient of eta1 in discrete time was 0.88 (95% credible interval 0.86 to 0.90). For eta2, the persistence was minimal (0.03, 95% credible interval 0.002 to 0.098). The cross-lagged coefficients were −0.035 (eta1 to eta2; 95% credible interval −0.070 to 0.003) and −0.015 (eta2 to eta1; 95% credible interval −0.037 to 0.007). Both measures were strongly influenced by system noise (ie, unpredictable external influences and random variations), with diffusion coefficients of 1.61 (95% credible interval 1.49 to 1.74) and 2.38 (95% credible interval 2.25 to 2.52), respectively.

The autoregressive and cross-lagged coefficients in discrete time as a function of elapsed days are shown in Figure 3, with the system noise standardized to 1.0 for visualization. The persistence of GenExp-SoA (eta1 to eta1) remained but gradually decreased over 10 days, whereas the very small persistence for LocalImp-SoA (eta2 to eta2) dissipated within approximately 2 days.

The estimated posterior correlation between the 2 SoA measures was –0.13 (95% credible interval –0.39 to 0.13), indicating uncertainty about the association (Table S14 in Multimedia Appendix 1). Summary statistics of the posterior distributions at the individual level for each period (<30, 30‐59, and ≥60 d) and each SoA are presented in Table S15 in Multimedia Appendix 1. The observed values mostly fell within the model’s predictive intervals, although these intervals were wide throughout this study (±18 for GenExp-SoA and ±5 for LocalImp-SoA). For GenExp-SoA, the estimated SD of the random effects (population SD 7.9) was similar to that of the observed data (SD 8.8), whereas for LocalImp-SoA, the estimated measurement error variance was unstable (mean 10.3, SD 25.8). Both SoA measures exhibited large diffusion coefficients, suggesting that the wide predictive intervals were attributable mainly to the large data variance, individual differences, and high levels of noise. The residual autocorrelation plots revealed no significant autocorrelation, suggesting that the model adequately captured the temporal structure (Figure S2 in Multimedia Appendix 1). However, the width of the credible intervals remained relatively large (approximately ±0.05 to ±0.1).

This current study tracked the 2 forms of SoA over 3 months in the daily life of 100 participants, using a longer duration and larger sample size than a previous study [33], which had a duration of approximately one month with fewer than 50 participants. In contrast to the previous study conducted under conditions involving goal pursuits, our study assessed the SoA without direct experimental manipulation, allowing observation under conditions closer to daily life. We observed small declines in the overall mean across all participants, stronger between- than within-individual variance, and high daily persistence for GenExp-SoA, but high within-individual variance and minimal persistence for LocalImp-SoA. No clear association was observed between the 2 measures in linear, nonlinear, and lagged effect analyses. Sensitivity analyses and continuous-time structural equation modeling supported these patterns.

These results inform hypothesis generation, study design, estimation of sample sizes, and other considerations in population science targeting the SoA and may guide strategies to evaluate the SoA in digital health interventions. Our findings support a 2-layer account in naturalistic settings. In free-living settings without an assigned goal or prescribed goal-pursuit activities, unmeasured day-to-day changes in attention and context may contribute to fluctuations in sensorimotor coupling (ie, eta2), which may partly explain the observed dissociation from reflective self-appraisals. Eta1 is slower and more stable, while eta2 is more context-dependent and more noise-sensitive.

The within- and between-individual variations in the 2 SoA measures differed in this study. GenExp-SoA, which has a high ICC, appears to adequately evaluate cross-sectional between-individual differences. In contrast, LocalImp-SoA, which has a very low ICC, suggests that the within-individual variation is very large, which may make it challenging to evaluate cross-sectional between-individual differences (Tables 3 and 4). The continuous-time structural equation modeling results support this difference. GenExp-SoA exhibited substantial autoregressive persistence, with a slower decay over time (high daily persistence; attenuation over ≈10 d), whereas LocalImp-SoA showed near-zero persistence (attenuation within ≈2 d), greater diffusion, and wide posterior predictive intervals, as well as a near-zero posterior correlation with the explicit measure (Figures 2 and 3, Tables S14 and S15 in Multimedia Appendix 1). In the continuous-time structural equation modeling, the 2 latent states were anchored by the observed measures of eta1 and eta2. The eta1 likely reflects a slowly varying explicit appraisal of control grounded in general agency beliefs, with daily adjustments influenced by mood, motivation, and contextual predictability. The eta2 likely indexes sensorimotor coupling between actions and outcomes at a short time-scale; is sensitive to attentional load, fatigue, adherence, and task valence and freedom of choice; and includes notable system noise. Accordingly, under free-living conditions, we expected the 2 measures to show dissociable dynamics, with eta2 being more contextually labile (ie, sensitive to contextual variation) than explicit metacognitive appraisals. Consequently, GenExp-SoA may be used to compare between-individual differences even with lower-frequency assessment, whereas LocalImp-SoA is better captured through higher-frequency, within-individual designs in free-living settings.

These patterns are also consistent with hierarchical computational accounts of SoA in predictive coding and active inference [17,21]. In brief, these accounts describe SoA as arising from hierarchical inference with 2 coupled levels: higher-level reflective beliefs about control and lower-level sensorimotor predictions (and prediction errors) that link actions to outcomes. In these accounts, GenExp-SoA, which is based on reflective self-reports, may reflect more stable metacognitive appraisals and higher-level beliefs about control, consistent with its high ICC and strong persistence. By contrast, LocalImp-SoA, indexed by outcome (temporal) binding, may depend more strongly on lower-level sensorimotor inference and short time-scale fluctuations that are sensitive to contextual variability, attentional state, fatigue, and uncertainty in temporal judgments, consistent with its low ICC, minimal persistence, and large diffusion. The near-zero association between the 2 measures may suggest that, in free-living settings, each measure is influenced by partly distinct components and measurement noise, even if the underlying levels are coupled in principle. This interpretation is tentative because prediction errors and relevant contextual or physiological covariates were not directly measured. Nevertheless, it provides a concise link between the observed temporal dynamics and computational accounts of agency.

Although LocalImp-SoA showed large diffusion and wide predictive intervals, these unstable observations were not only due to system noise. They may also reflect genuine short-time-scale variance in implicit sensorimotor coupling in mobile environments. Accordingly, improving measurement stability in mobile, free-living settings will likely require increasing the number of trials per session and measuring and compensating for latencies specific to the device and the operating system. Temporal binding is associated with dispositional traits such as self-transcendence [42], and LocalImp-SoA likely reflects a dynamic sensorimotor process with psychosocial relevance. Therefore, it is preferable to assess this with a multimethod battery rather than a single proxy. Analytically, nonlinear and time-varying models may better capture real-life changes and test just-in-time adaptations triggered by momentary states in free-living settings [43-46], with SoA as a candidate trigger to be validated.

In this study, a decline in the mean of each SoA measure across the cohort of participants was observed (Tables 3 and 4). Although the average change in both SoAs among participants was small, SoA can plausibly have cumulative implications for motivation and behavioral regulation because affective processing partially mediates the pathway from agency to action regulation [20], and agency functions as a dynamic psychological resource in health-related meaning-making and adjustment [19]. From a population-science perspective, even modest shifts in the cohort mean may still be relevant because they indicate a systematic group-level change rather than isolated individual fluctuations. Moreover, cohort-level averages can mask heterogeneity in individual trajectories and context-dependent fluctuations. This is consistent with the observed variance structure (high ICC for GenExp-SoA and low ICC for LocalImp-SoA) and the minimal persistence and large diffusion observed for LocalImp-SoA. In line with these observations, the decrease in GenExp-SoA reflects changes in participants’ awareness, whereas the decline in LocalImp-SoA indicates lower persistence, which is consistent with labile implicit coupling. Previous studies have reported associations between SoA and health, psychological, social, and environmental factors [14-16,47-49]. Thus, the observed decline in SoA may be attributable to unmeasured factors related to health, psychology, the environment, and social status. Such complex contextual changes may act at both the population and individual levels.

The observed decline in the SoA may also reflect seasonality. We note that environmental and seasonal covariates (eg, weather, illness seasonality, and work cycles) were not directly measured in this study. To our knowledge, no studies have reported seasonal variations in SoA. However, seasonal variations have been reported for indicators of health, psychological state, environmental factors, and social status, which are associated with the SoA [50-56], many of which have been linked in prior studies to processes that are relevant to agency [18,47-49]. In this study period (late fall to spring), SoA decreased on average. Previous studies have reported increases in weight [51], the prevalence of metabolic syndrome [50], and possible increases in depression in winter [52], as well as poorer scores on neuropsychological test batteries, such as visual and verbal attention, working memory, verbal ability, verbal fluency, and executive functioning [54]. Such seasonal changes across multiple factors may have contributed to the observed decline in the SoA and to differences in individual characteristics in this study. The contribution of seasonality may be nontrivial and it cannot be disentangled without explicitly modeling environmental covariates (eg, weather, illness seasonality, and work cycles). More broadly, seasonality may covary with reductions in multisensory congruence, constraints on autonomy, or disruptions of environmental regularities. These conditions can diminish perceived control and performance [14-18]. Taken together, these seasonality-related interpretations are speculative and hypothesis-generating.

In contrast to the results of the present study, Legaspi et al [33] reported an increase in GenExp-SoA from November 20, 2021, to December 31, 2021. Both studies measured GenExp-SoA using a mobile app in daily life. However, their observation window was late fall to early winter, whereas our observations spanned late fall to spring. In addition, the study by Legaspi et al [33] instructed participants to perform goal pursuit tasks. This contrast supports the idea that mastery experiences, autonomy, and structured feedback can increase eta1. Future studies should test this by explicitly manipulating and modeling these contextual factors in real-life settings [14,15,18].

For GenExp-SoA, a slight increase was observed among men, whereas a decrease was observed among women (Table 3). With respect to personality, women are known to have higher levels of agreeableness and neuroticism [57,58]. In the present study, women had higher scores on neuroticism than men did, but the agreeableness scores were similar (data not shown); participants scoring above the median on these traits showed a smaller decrease in the explicit measures than those scoring below the median did (Table 3). These findings suggest that social, environmental, and biological factors related to sex may be associated with perceptions and self-disclosure of the SoA and that higher agreeableness and neuroticism may attenuate these decreases.

With respect to age groups, GenExp-SoA decreased in the 20‐39 age group, slightly decreased in the 40‐49 age group, and increased in the 50‐59 age group (Table 3). A prior large longitudinal study reported that extraversion, conscientiousness, and openness decrease across adulthood [59]. In our study, the older age groups had higher mean scores for extraversion, agreeableness, and conscientiousness but lower scores for neuroticism (data not shown). The GenExp-SoA decreased less among groups with high agreeableness, conscientiousness, and neuroticism scores than among those with low scores. By contrast, participants with high extraversion showed a greater decrease (Table 3). Reorganization of roles, socioeconomic demands, and regulatory styles, all linked to age, has the potential to influence SoA trajectories through motivational, cognitive, and affective routes [18-20,47]. The existence of known interrelations among sex, age, and personality suggests the necessity of conducting mediation and moderation tests based on a priori hypotheses. These assessments should include affective processing as a potential mediator [20]. In addition, the contextual moderators described in the Introduction should be considered, including autonomy vs coercion and environmental regularities [14-16].

Although no linear, nonlinear, or cross-lagged associations were observed between the 2 SoA measures (Table 5, Figure S1 in Multimedia Appendix 1, and Figure 2), the nonlinear model suggested a slight increase in GenExp-SoA at high LocalImp-SoA, but this was not statistically significant and requires replication using larger, more diverse samples. This observation is consistent with the results of previous studies, demonstrating they are independent components in both experimental [60] and daily environments [33]. Existing theories suggest that eta1 reflects conscious self-control, whereas eta2 is related to unconscious behavioral control [48]. This dissociation is consistent with the distinct construct coverage of the 2 indicators. The SoAS depends on general agency beliefs that are independent of context, whereas outcome binding reflects local coupling between actions and effects that are specific to the task [18,22,24]. These measures are therefore complementary rather than interchangeable, and combining them provides a more comprehensive perspective. Prereflective states can inform explicit appraisals, and explicit appraisals can modulate prereflective coupling. Both associations are context-dependent and noise-sensitive. Several studies have indicated that eta2, as indexed by binding, is sensitive to outcome valence and the degree of choice freedom [8,15,61].

In digital health, eta1 can function as a tailoring variable for intervention delivery [62]. In the event of a downturn, interventions can accentuate recent successes, divide tasks into more manageable components, and provide precise and timely feedback [62,63]. These actions can support motivation, behavioral regulation, and adherence [18,20,62,63]. These findings also have implications for AI-driven adaptive digital interventions that tailor the timing and content of support to fluctuations in eta1 and eta2 in real time. In all the analyzed models of this study, no association was found between the 2 forms of SoA. However, modeling slopes that are specific to each participant improved the model fit, highlighting the importance of within-individual patterns (Tables S11 and S12 in Multimedia Appendix 1). This is a hypothesis-generating proposition: instability in LocalImp-SoA may be a marker of acute cognitive load or fatigue when considered together with a brief vigilance test or HRV. This is consistent with evidence that brief mobile vigilance tests and wearable HRV are sensitive indicators of acute fatigue and stress in applied and real-world settings [64,65]. In such instances, timely adaptation can postpone demanding inputs and rely more on passive sensing. This approach has the potential to promote sustained engagement [62,63]. However, to our knowledge, no study has jointly measured intentional and temporal binding with HRV and vigilance testing, which warrants further research. The passive sensing of behavioral data, including heart rate, activity, smartphone and app use, and GPS, can be integrated with brief active probes to estimate changes in the SoA and to trigger support in real time that is contingent on agency [30]. However, these models must undergo rigorous internal and external validation, periodic reanchoring to ground-truth SoA assessments, and fairness checks before deployment.

This study is the first to detail voluntary measurements of the 2 forms of SoA over 3 months using smartphones in free-living environments, with a sample balanced for age and sex, a larger sample size, and a longer observation period than prior mobile SoA research. This study combines multilevel and continuous-time modeling with internal validation.

However, there are several limitations. First, this study has a descriptive, exploratory, and observational design. Therefore, causal inference is constrained, and observed associations may include unmeasured confounding, including unmodeled environmental and seasonal covariates such as weather, illness seasonality, and work cycles. Second, because data collection using mobile apps depended on participants’ smartphone usage habits, environmental conditions, and adherence, it is not possible to exclude self-selection bias, self-report bias, behavioral bias, and environmental bias with network factors. In addition, because participant compensation was linked to the app-based survey recording rate, compensation may have influenced input behavior, including recording frequency, continuation, and response accuracy. As compensation was not contingent on the content or values of SoA responses, it was unlikely to directly incentivize particular SoA reports. Additionally, device and operating system heterogeneity, occasional app reinstallations, and connectivity issues may have introduced nonclassical measurement error, meaning state-dependent or device-dependent errors (ie, systematic, nonrandom errors) and technical disruptions that may lead to skipped days or irregular completion patterns. However, these factors are inherent to routine real-world app deployment. Accordingly, evaluating the SoA under such constraints provides real-life valid and implementation-relevant estimates for digital health. Third, generalizability is restricted to adults aged 20‐60 years living in Japan, and it may not extend to other age groups or to patients with serious mental illness or sensory impairment. In addition, transposability to settings outside Japan, including those with different languages, health systems, or cultural contexts, remains unknown. Fourth, LocalImp-SoA exhibited large within-individual variability. Trial counts at the session level may have been insufficient, and using outcome judgment errors as a compliance proxy is imperfect. The collective consideration of these limitations indicates a direction for improvement. In addition, the 3-month high-frequency mobile assessment with internal validation provides reliable empirical and practical information for future SoA research and digital health applications.

Future studies could integrate affective measures for mediation, contextual moderators such as autonomy and coercion or environmental regularities, and multimethod implicit batteries. They could also extend to diverse populations while leveraging computational accounts to guide feedback design aligned with intended and sensed outcomes. These steps may help clarify mechanisms, improve construct validity and generalizability, and inform the timing and content of adaptive interventions.

This study used a mobile app to measure the 2 forms of SoA in daily life over 3 months, without specific tasks or goals. The temporal dynamics of the SoA were characterized, as well as their relationships with individual characteristics. GenExp-SoA showed greater between-individual than within-individual variation, whereas LocalImp-SoA showed very high within-individual variation. The mean explicit measures decreased over time across all participants, with greater declines observed among women and participants with high extraversion scores, while the decline was attenuated among participants with high conscientiousness scores. Conversely, GenExp-SoA tended to increase among participants aged 50‐59 years. The mean implicit measures also showed a slight decrease, but associations with individual characteristics were limited, and no association was observed between the 2 forms of SoA. These findings suggest that epidemiological studies and digital health interventions should account for the distinct measurement properties of the 2 forms of SoA, use lower frequency assessments for eta1, use high frequency designs at the individual level for eta2, and consider individual characteristics when interpreting SoA dynamics.