Short-term behavioural consistency demonstrates a significantly stronger causal impact on daily well-being than long-term self-regulation, despite a near-zero correlation, and causally-informed feature selection significantly improves predictive accuracy of well-being in machine learning models compared to conventional methods.
Key Findings
Methods
The study developed a novel computational metric called the Behaviour Self-Regulation Score (BSRS) to quantify both trait-like (long-term) and state-like (short-term) behavioural consistency from daily reports of physical activity and sleep.
BSRS was derived from longitudinal self-report data over a 28-day period
The metric was designed to be used within a causal graphical modelling framework controlling for motivational and perceptual confounders
Methods
Of 141 enrolled participants, 94 were analysed after applying an a priori completeness threshold of at least 20 of 28 daily entries.
Total enrolled: N=141
Total analysed: N=94
Completeness threshold was set a priori at ≥20 of 28 daily entries
Data consisted of daily self-reports of physical activity and sleep over a 28-day period
Results
Long-term self-regulation (BSRS-L) demonstrated a stable positive causal effect on well-being.
The causal effect of BSRS-L was estimated using causal graphical models and propensity score methods
The effect was described as 'stable' and positive
Motivational and perceptual confounders were controlled for in the analysis
The finding distinguishes causal effect from mere correlation
Results
Short-term behavioural consistency (BSRS-S) demonstrated a significantly stronger causal impact on daily well-being compared to long-term self-regulation, despite having a near-zero correlation with well-being.
BSRS-S showed a 'significantly stronger causal impact on daily well-being' than BSRS-L
The correlation between BSRS-S and well-being was described as 'near-zero'
This finding illustrates a key discrepancy between correlation-based and causal analysis approaches
The result was obtained after controlling for motivational and perceptual confounders using propensity score methods
Results
Features selected via the causal framework significantly improved the predictive accuracy of well-being in machine learning models compared to conventional feature selection methods.
Causally-selected features outperformed conventionally selected features in machine learning prediction of well-being
The improvement was described as 'significant'
The comparison was made against 'conventional feature selection methods'
This finding supports the utility of causal inference for improving digital health prediction models
Discussion
The study demonstrates that causal inference methods can disentangle drivers of well-being from complex longitudinal self-report data where traditional correlation-based analyses may be misleading.
Traditional analyses were described as often conflating 'correlation with causation'
Causal graphical models and propensity score methods were employed to estimate causal effects
The near-zero correlation of BSRS-S with well-being despite its strong causal effect exemplifies this discrepancy
The authors argue this framework identifies 'more potent targets for digital health interventions'
What This Means
This research suggests that how consistently people regulate their daily behaviours — like physical activity and sleep — in the short term has a stronger influence on their well-being than their long-term behavioural patterns, even though short-term consistency showed almost no simple statistical correlation with well-being. The researchers developed a new scoring system called the Behaviour Self-Regulation Score (BSRS) to measure both day-to-day (short-term) and overall (long-term) behavioural consistency in 94 adults over a 28-day period. By using advanced causal analysis techniques rather than standard correlation methods, they were able to identify which behavioural factors actually drive well-being rather than just being associated with it.
A key finding is that short-term behavioural consistency (BSRS-S) had a much stronger causal effect on daily well-being than long-term consistency, yet if researchers had relied on traditional correlation statistics alone, they would have concluded it had no meaningful relationship with well-being at all. This highlights a major limitation of conventional data analysis approaches in health research, where correlation and causation are often conflated. The study also found that using causally-selected features in machine learning models produced significantly better predictions of well-being than using features chosen by standard methods.
This research suggests that digital health tools and interventions aimed at improving well-being may be more effective if they focus on helping people maintain consistent short-term behavioural routines rather than solely targeting long-term habit formation. It also provides a methodological template for researchers working with longitudinal health data who want to move beyond correlation and identify true causal drivers of health outcomes.
Wang J, Wang P, Woo W, Sakalidis K, Hettinga F, Rodrigues A, et al.. (2026). Quantifying the effect of Behaviour Self-Regulation on well-being through causal analysis: A methodological framework for longitudinal health data.. Journal of biomedical informatics. https://doi.org/10.1016/j.jbi.2026.104984