The phone buzzes while life is still happening
At 10:17, a participant is irritated after a meeting. At 13:42, they are absorbed in a task. At 18:06, they are tired on a train. A questionnaire months later might ask how emotional they tend to be. Experience sampling tries to catch smaller pieces before memory compresses them into a story.
That proximity is the method's great advantage. It is also where the methodological work begins.
Experience Sampling Method, often abbreviated ESM, and the closely related family of Ecological Momentary Assessment methods collect repeated reports from people in their daily environments. When applied to emotion, they can examine momentary affect, discrete emotions, regulation strategies, goals, context, social company, events, and change over time within the same person.
It is tempting to call this “real-life data” and stop there. A serious methods page cannot.
What experience sampling measures
Experience sampling is not one questionnaire. It is a research design family. Investigators define prompts, schedules, response windows, items, study duration, devices, and analytic models according to the question being asked.
An emotion study might ask participants several times per day how angry, sad, anxious, calm, interested, or joyful they feel. Another might ask what happened since the last prompt, what regulation strategy was used, whether the participant was alone, and how well the strategy seemed to work. A third might trigger reports after a specific event.
Because observations are repeated within persons, ESM can separate within-person variation from between-person differences. That distinction is central. A finding that people who are generally more anxious use more suppression does not necessarily mean that a particular person suppresses more at the moments when their anxiety rises.
Sampling designs
Common designs include signal-contingent sampling, in which prompts occur at scheduled or semi-random times; event-contingent sampling, in which participants report when a defined event occurs; interval-contingent sampling, in which reports are completed after fixed periods; and mixed designs.
Each creates different data.
Random or stratified prompts can reduce predictable timing but may interrupt activities. Fixed prompts may improve compliance but become anticipated. Event-contingent designs can target meaningful episodes but depend on participants noticing and classifying events consistently. Long response windows increase convenience while weakening momentary proximity.
Study duration matters too. A dense three-day protocol and a lighter thirty-day protocol do not answer the same temporal question. The correct design depends on expected timescale, participant burden, and the stability of the phenomenon.
Why proximity to the moment helps
Retrospective reports are shaped by memory, aggregation, salience, current mood, and beliefs about the self. Asking close to an experience can reduce some forms of recall bias and preserve contextual detail that later summaries lose.
This is especially valuable for emotion dynamics. Researchers can examine variability, inertia, instability, reactivity, recovery, co-occurrence, and context sensitivity. They can ask whether a strategy that looks adaptive between people is actually useful within a person in particular situations.
But “closer to the moment” is not identical to direct access. A participant still interprets a question, selects a response category, and may summarize the preceding minutes or hours. The measure remains self-report unless combined with other data.
The nine methodological questions that refuse to disappear
Arthur Stone, Stefan Schneider, and Joshua Smyth's 2023 review of ecological momentary assessment organized several persistent issues. Among them were whether reports are truly momentary, how participants interpret questions, what comparison standards they use, whether reporting periods can stretch beyond the moment, how participants should be trained, selection bias, missing assessments, reliability, and the conditions under which EMA can be considered a gold standard.
These are not minor technicalities. They determine what a datapoint means.
If one participant interprets “right now” as the last five seconds and another as the whole morning, a shared response scale does not create identical observations. If a person rates stress relative to their own usual day while another compares themselves with other people, the response contains different reference frames.
Compliance and missingness
Repeated prompting creates missing data almost by definition. People sleep, drive, teach, argue, care for children, silence phones, lose connectivity, or simply stop responding.
Missingness can be informative. The moments when someone cannot or does not respond may be the very moments with highest emotional intensity. A dataset with 80 percent compliance can still be biased if the missing 20 percent is systematically different.
Researchers should report compliance transparently, define valid response windows, examine patterns of missingness, and justify exclusion rules. Treating every missing prompt as random convenience can make daily-life data look cleaner than daily life actually is.
Reactivity: measurement can enter the phenomenon
Being asked repeatedly about emotion may change attention to emotion. A participant who receives six prompts a day about anger could become more observant, more annoyed, more reflective, or more likely to regulate it. This is measurement reactivity.
The existence and size of reactivity varies by design and construct. It should not be assumed away. In intervention studies, repeated measurement can even become a low-level intervention, complicating the interpretation of change.
The ethical side matters too. Frequent prompts about distress can burden participants. Protocols need risk procedures appropriate to the population without implying that a research app provides clinical monitoring unless it actually does.
Reliability in intensive longitudinal data
Traditional internal consistency metrics were developed for different measurement structures. ESM data contain repeated observations nested within people and sometimes within days or contexts. Reliability can differ at within-person and between-person levels.
A measure that reliably distinguishes people from one another may perform poorly at detecting moment-to-moment change inside a person. Researchers need reliability estimates suited to the level of inference they intend to make.
The same issue applies to emotion-dynamics metrics. Variability, inertia, and network measures can be sensitive to number of observations, spacing, time trends, and measurement error. A sophisticated metric does not compensate for a weak sampling design.
The field is methodologically diverse
A 2026 systematic review by Lisa Peeters and colleagues examined 150 articles describing 162 experience-sampling studies in psychology and psychiatry. The review found substantial variation across the entire workflow, from design through analysis and reporting. The median study duration was ten days, convenience samples were common, and multilevel modelling was widely used.
The review's importance is not that ten days is a recommended universal design. It is almost the opposite. Describing what is common exposes how much methodological choice remains and why transparent reporting is necessary.
For a developing field, methodological variation can support innovation. It also makes synthesis harder when studies use the same label for very different protocols.
Analysis: the person is not a pile of independent rows
Repeated observations from one participant are statistically dependent. Analyses therefore need to account for nesting and time. Multilevel models, time-series approaches, dynamic structural equation models, generalized additive methods, and other intensive-longitudinal techniques may be appropriate depending on the question.
Researchers must also separate person-mean effects from momentary deviations when relevant. Without that decomposition, a model can blend stable individual differences with within-person processes and produce interpretations that sound dynamic but are actually cross-sectional.
Temporal order matters for causal language. Even when one variable precedes another within the day, observational time-series data do not automatically establish causation. Unmeasured events and reciprocal processes remain possible.
What the evidence does not show
Experience sampling does not provide unfiltered access to “true emotion.” It does not eliminate self-report bias, guarantee ecological validity, or make causal inference automatic. More prompts do not always mean better data. High compliance does not prove missingness is harmless. Smartphone delivery does not make a measure objective.
ESM findings from one protocol also do not generalize automatically to another with different prompt timing, response windows, items, study duration, or population.
Institute synthesis: capacity has a timescale
The following is Institute synthesis.
The strongest reason for the Institute to care about experience sampling is that capacity is temporal. A person can describe themselves accurately in general and still behave differently at 16:40 on a hard Tuesday. They can know a strategy in theory and fail to access it under pressure. They can recover quickly in one relationship and remain activated for hours in another.
Retrospective questionnaires compress those differences. ESM can reopen them.
For Inner Technology research, this creates the possibility of studying not only whether a person “has” a capacity but when it becomes available, under what conditions, and how quickly it returns after disruption. That is an Institute framing, not a standard ESM construct.
The method also introduces an ethical design question. If digital systems can prompt people at precisely the moment a capacity is likely to fail, measurement and intervention begin to approach one another. Researchers need to state clearly whether a prompt is observing, coaching, or both.
Practical design checklist
Define the temporal process before choosing the prompt schedule. State whether sampling is signal-, interval-, event-contingent, or mixed. Specify response windows and timeframes in items. Pilot burden. Track compliance and missingness. Report participant training. Choose reliability estimates appropriate to within- or between-person inference. Model nesting and time. Pre-specify primary outcomes when possible. Document changes to the protocol. Distinguish observation from intervention.
Experience sampling is powerful because it gets closer to life. Its credibility depends on respecting everything that still stands between the prompt and the phenomenon.
Designing the item is designing the phenomenon
A repeated prompt needs to be short enough for daily life, but brevity can hide ambiguity. “How anxious do you feel?” might be interpreted as bodily arousal, worry, fear, or a global self-judgment. “Did you regulate your emotion?” assumes participants share a definition of regulation.
ESM therefore benefits from cognitive interviewing and pilot work before launch. Researchers can ask participants what they think each item means, how they interpret time windows, and whether response options fit the lived range of the experience. A protocol that works beautifully in software can still fail semantically.
Item count also creates a trade-off. Longer prompts allow richer constructs but increase burden. Shorter prompts preserve compliance but may reduce reliability. The optimal balance is empirical and population-specific.
Timing is part of causal plausibility
Intensive longitudinal data are attractive because measurements have time stamps. Yet temporal order only helps if the sampling interval matches the process. If anger rises and falls within ten minutes, four prompts per day may miss the sequence entirely. If burnout develops over months, minute-level sampling can generate enormous data without capturing the relevant timescale.
Researchers should therefore begin with a theory of temporal dynamics. How quickly can the predictor change? How quickly can the outcome respond? What lag is plausible? What carryover is expected? Sampling density should follow from those questions.
This is also why “more data points” is not a universal quality metric. A thousand observations at the wrong timescale can be less informative than fifty well-timed ones.
Personalized models require more than many rows
ESM is often promoted as a route to idiographic or person-specific models. In principle, repeated observations can reveal individual patterns obscured by group averages. In practice, person-specific inference requires enough observations, sufficiently variable constructs, appropriate models, and careful validation.
A network drawn from one person's phone data can look impressively individualized while remaining statistically unstable. Forecasting accuracy should be tested out of sample when prediction is the goal. Exploratory person-specific graphs should be labeled as exploratory.
The human interpretation also matters. A statistical association between loneliness and sadness within one participant does not explain the relationship. It may reflect time of day, weekends, conflict, illness, or another unmeasured process.
Sensors are complements, not truth machines
Smartphones and wearables can add location, activity, heart rate, sleep estimates, screen use, or communication metadata. These data can reduce reliance on memory for some behaviors. They also introduce sensor error, proprietary algorithms, missingness, privacy risk, and interpretive ambiguity.
Heart rate is not an emotion label. GPS does not reveal meaning. Screen time does not disclose whether a person was working, connecting with a friend, or compulsively scrolling. Combining passive sensing with self-report can be powerful precisely because the two sources answer different questions.
The governance burden rises with the richness of the data. Emotion research can become surveillance if collection outruns necessity, consent, and protection.
Reporting standards as cumulative infrastructure
Methodological diversity is not inherently a problem. Poor reporting is. If studies clearly document prompt schedules, items, response windows, compliance, exclusions, missingness, participant incentives, device rules, and analysis decisions, later researchers can understand why findings differ.
The 2026 methodological review makes this cumulative-science issue visible. An ESM literature becomes more useful when protocols can be compared, reproduced, and meta-analyzed rather than when every study invents a new opaque format.
Participant burden is a validity variable
Burden is often discussed as an ethical or logistical issue, but it also affects validity. A demanding protocol can selectively lose participants with less flexible schedules, greater distress, caregiving responsibilities, or lower digital access. Within retained participants, late-study responses may become faster and less considered.
Compensation, prompt frequency, quiet hours, snooze options, and study length therefore shape the sample and the data. Reporting these choices lets readers judge whose daily life the protocol was realistically able to observe.
Event-contingent protocols need an event ontology
When participants are told to report “after conflict,” “when you regulate,” or “after a craving,” the study depends on a hidden classification task. Participants must decide what counts as the event. One person may report a sharp disagreement; another may include mild tension. One may recognize regulation only when deliberate; another may include automatic distraction.
Researchers can improve consistency by defining events behaviorally, training participants with examples, and recording borderline cases during piloting. Even then, event-contingent data are partly data about detection and interpretation.
Feedback, dashboards, and intervention contamination
Some platforms show participants graphs of their own data during collection. That can improve engagement and offer ethical reciprocity. It can also change the process being observed. A person who sees that anxiety peaks after certain meetings may begin avoiding those meetings or regulating differently.
If feedback is part of the design, it should be treated as an intervention component and described explicitly. If the goal is pure observation, delayed feedback may be more appropriate. The choice is not only technical; it changes the causal system.
Privacy as a measurement condition
Participants may answer differently when prompts concern intimate emotion, relationships, self-harm, or work conflict and the phone screen is visible to others. Perceived privacy affects disclosure. Data governance can therefore influence measurement validity as well as ethics.
Protocols should minimize unnecessary collection, explain access clearly, encrypt sensitive data where appropriate, and separate research requirements from data that are merely convenient to collect. Rich longitudinal data can become identifying even after obvious names are removed.
Research references
Stone, A. A., Schneider, S., & Smyth, J. M. (2023). Evaluation of Pressing Issues in Ecological Momentary Assessment. Annual Review of Clinical Psychology, 19, 107–131. https://doi.org/10.1146/annurev-clinpsy-080921-083128
Peeters, L., Van den Noortgate, W., Blanchard, M. A., Suenaert, M., Eisele, G., Kirtley, O. J., Artner, R., & Lafit, G. (2026). Mapping Methodological Variation in Experience-Sampling Research From Design to Data Analysis: A Systematic Review. Advances in Methods and Practices in Psychological Science. https://doi.org/10.1177/25152459261453933
Wrzus, C., & Neubauer, A. B. (2023). Ecological momentary assessment: A meta-analysis on designs, samples, and compliance across research fields. Assessment. https://doi.org/10.1177/10731911211067538