Ecological Momentary Assessment

Abstract

Ecological momentary assessment (EMA) is a type of psychological test: a family of methods that repeatedly sample a person's experience and behavior in real time, in the natural settings of daily life, to minimize the distortions of retrospective recall. This article treats EMA as a measurement instrument: how its defining commitments — real-world context, real-time capture, and repeated sampling — differ from a one-time questionnaire, how the three sampling designs decide when a report is collected, why a momentary report is less biased than a remembered one, and how compliance governs whether the resulting dataset is usable at all. Three interactive demonstrations model the sampling schedule across a day, the recall bias separating a remembered rating from the momentary average, and the compliance arithmetic that turns a protocol into a count of captured data points.

Keywords: momentary self-report, recall bias, experience sampling

Ecological momentary assessment is the method cognitive and clinical science reaches for when a single questionnaire will not do — when the quantity of interest is a mood, a craving, or a symptom that changes from hour to hour and that memory reconstructs badly. It is a type of psychological test, sharing that class's machinery of standardized items and quantified scores, but it inverts the usual design: instead of measuring a person once, in a clinic, about the past, it measures them many times, in situ, about the present moment. Following EMA from its defining goals to the compliance rates that make or break a study is a compact account of how self-report learned to trade the convenience of a single session for the fidelity of measuring life as it is lived.

What Ecological Momentary Assessment Is

Ecological momentary assessment is a family of methods in which participants report on their momentary states and behaviors repeatedly, in real time, and in their natural environments (Shiffman, Stone, & Hufford, 2008). The term was coined to name a set of commitments that a conventional questionnaire violates: the data should be ecological, gathered in the real-world settings where behavior actually occurs; momentary, describing the present moment rather than a summary of the past; and repeated, sampling the same person many times so that within-person change becomes visible (Stone & Shiffman, 1994). MeSH files ecological momentary assessment as a narrower descriptor under psychological tests, the general class of standardized psychological measurement of which it is one instance. Its defining move is temporal: where a standard scale measures a person once and asks them to average their own experience over weeks, EMA measures the experience itself as it unfolds, and leaves the averaging to the analyst (Bolger, Davis, & Rafaeli, 2003).

Two features mark the instrument. The first is repeated within-person sampling: because each participant supplies many observations, EMA data are intensively longitudinal and nested — moments within days within people — and are read with the multilevel models that such structure demands rather than with a single cross-sectional score (Bolger, Davis, & Rafaeli, 2003). The second is the deliberate minimization of recall: by asking about now, EMA sidesteps the memory processes that make a person's summary of last week diverge from the sum of its moments (Stone & Shiffman, 2002). As with any psychological measure, the value of a momentary dataset is judged by its methodological rigor — the sampling design, the prompt schedule, and above all the compliance rate — which the reporting guidelines for the method were written to make explicit (Stone & Shiffman, 2002).

From the Experience Sampling Method to EMA

The modern method begins with a pager. In the 1970s and 1980s Mihaly Csikszentmihalyi and Reed Larson developed the Experience Sampling Method (ESM), in which participants carried an electronic pager that signaled at random times through the day, each signal prompting them to record what they were doing and how they felt at that moment (Csikszentmihalyi & Larson, 1987). ESM established the core idea on which everything later rests: that subjective states must be measured in situ, as they occur, because a state sampled at the moment is a different and better datum than the same state reconstructed afterward. Csikszentmihalyi's theory of flow was built on exactly this kind of momentary data.

The paradigm was named and codified as EMA by Arthur Stone and Saul Shiffman, who set out its defining goals for behavioral medicine and, later, its reporting standards (Stone & Shiffman, 1994; Stone & Shiffman, 2002). Shiffman made event-contingent sampling central to addiction science, capturing craving and lapse in the moment of smoking rather than through the distorted lens of retrospective recall (Shiffman, 2009). The 2008 Annual Review by Shiffman, Stone, and Hufford consolidated the method's sampling designs and its compliance principles into a canonical statement (Shiffman, Stone, & Hufford, 2008). In parallel, Niall Bolger and colleagues situated EMA within the broader family of diary and intensive-longitudinal designs, supplying the multilevel-modeling framework the data require (Bolger, Davis, & Rafaeli, 2003). Figure 1 sets out these milestones.

Sampling Designs: Signal-, Interval-, and Event-Contingent

What most distinguishes one EMA protocol from another is the rule that decides when a report is collected. Three sampling designs are standard (Shiffman, Stone, & Hufford, 2008). In signal-contingent sampling the device prompts the participant at times the researcher controls — often random within blocks — and the participant reports on the moment of the signal; this is the direct descendant of the ESM pager and is the design of choice for characterizing the ordinary flow of experience, because the prompt catches the participant unawares. In interval-contingent (or time-contingent) sampling the participant reports on a fixed schedule, such as every evening or every three hours, which is simple but lets the participant anticipate the report. In event-contingent sampling the participant initiates a report whenever a defined event occurs — a cigarette, a panic attack, an argument — which is the only design that reliably captures rare or fleeting events that a fixed or random schedule would miss (Shiffman, 2009). Table 1 sets out the three designs, the rule that triggers a report in each, and what each is best suited to capture.

The designs are not interchangeable; each answers a different question, and many studies combine them (Shiffman, Stone, & Hufford, 2008). Signal-contingent sampling estimates how a state is distributed across time; event-contingent sampling estimates what surrounds a particular kind of moment. The choice also sets the sampling density — how many observations per day, and therefore how finely within-person dynamics can be resolved. The first demonstration makes the schedule manipulable: the reader chooses a design and a number of prompts across a waking window and watches the sampling times fall on a day timeline, with the average interval between prompts reported, because the interval is what determines how much of the day's variation the protocol can actually see.

Placing prompts across a day: the three sampling designs

Signal-contingent sampling prompts at random times the researcher controls, catching the participant unawares; it estimates how a state is distributed across the day. With 7 prompts over a 14-hour window the average spacing is 2.00 hours — denser sampling resolves finer within-day variation at the cost of greater burden.

Why Momentary Beats Retrospective: Recall Bias

The reason to bear the cost of repeated in-situ sampling is that the alternative — asking a person to remember — is systematically biased. Memory does not store and average every moment of an episode; it reconstructs a summary using heuristics, and two of them distort self-report in predictable directions (Stone & Shiffman, 2002). The peak-end rule, established in Redelmeier and Kahneman's studies of remembered pain, is that a remembered episode is evaluated mostly by its most intense moment and its final moment rather than by its duration or its average, so an episode with a high peak or a bad ending is remembered as worse than the moment-by-moment average would warrant (Redelmeier & Kahneman, 1996). Recency weights the most recent moments over earlier ones. A retrospective questionnaire therefore does not return the average of the moments it asks a person to summarize; it returns a memory-weighted distortion of them (Shiffman, Stone, & Hufford, 2008).

EMA is built to defeat exactly this. By asking about the present moment, it collects the raw moments themselves, and the analyst — not the participant's memory — computes whatever summary is wanted (Stone & Shiffman, 2002). The gap between the two is not hypothetical: studies that collect momentary reports and a retrospective report of the same period find that recalled and momentarily sampled quantities diverge, and diverge more for some constructs than others (Solhan et al., 2009). The second demonstration makes the bias visible: the reader enters a sequence of momentary intensity ratings, and the demo shows the true momentary mean beside a peak-end estimate — the average of the peak and the final rating — with the recall bias between them, the quantity EMA exists to eliminate.

The momentary mean versus the remembered episode

The five moments average to a momentary mean of 4.80, but memory following the peak-end rule weights the peak (9) and the final moment (3) into a remembered estimate of 6.0. The recall bias is +1.20 — the distortion EMA removes by keeping the moments and computing the mean itself.

Compliance and Data Quality

A momentary dataset is only as good as the fraction of prompts a participant actually answers. Compliance — completed prompts as a proportion of issued prompts — is the single most consequential quality metric of an EMA study, and it appears in the reporting guidelines for exactly that reason (Stone & Shiffman, 2002). Low compliance shrinks the dataset, but the deeper problem is non-random missingness: if participants skip prompts when they are busiest, most distressed, or intoxicated, then the moments that go unrecorded are precisely the ones of interest, and the captured data are biased in a way no amount of them can fix (Shiffman, Stone, & Hufford, 2008).

Compliance depends on design choices the researcher controls — prompt frequency, item burden, device, and population. Smartphone delivery has made EMA far easier to field, but the smartphone did not solve compliance; a meta-analysis of mobile-EMA studies in children and adolescents found that response rates vary widely and depend systematically on prompt design and study features (Wen et al., 2017). The methodological literature therefore treats a reported compliance rate, and the analysis of who missed which prompts, as a precondition for believing an EMA result at all (Stone & Shiffman, 2002; aan het Rot, Hogenelst, & Schoevers, 2012). The third demonstration makes the arithmetic concrete: the reader sets prompts per day, study length, and a compliance rate, and watches the protocol resolve into a count of captured versus lost momentary observations against a conventional benchmark, because that captured count — not the protocol on paper — is the real sample.

From a protocol to a captured sample: compliance

The protocol issues 42 prompts, but the real sample is the 34 a participant actually answers at 80% compliance, leaving 8 lost. The rate alone is not enough: whether those 8 missed prompts fell at random or at the moments of interest is what a compliance analysis must establish before the 34 captured reports can be trusted.

Ambulatory Assessment and Momentary Intervention

EMA is the self-report core of a wider program. Ambulatory assessment extends momentary measurement from questionnaire to sensor: alongside the participant's reports, wearable and phone-based devices record physiology, movement, and location continuously, so that a self-reported mood can be aligned in time with heart rate or activity (Trull & Ebner-Priemer, 2013). This is how EMA reaches constructs a questionnaire cannot see — affective instability, for instance, the moment-to-moment variability of mood that is a single number only when computed from a dense momentary series, and that distinguishes clinical groups such as borderline personality disorder (Ebner-Priemer & Trull, 2009).

The same momentary stream can also be turned around to deliver treatment. Ecological momentary intervention (EMI) uses a participant's real-time data to trigger support in the moment and context where it is needed — a coping prompt when stress rises, feedback when a target behavior occurs — rather than only in a weekly session (Colombo et al., 2019). In psychiatry this move from passive assessment toward real-time, in-context intervention, together with digital phenotyping from passive sensor data, is the frontier the experience-sampling tradition has grown into (Myin-Germeys et al., 2018). Across clinical and developmental research the method's reach now extends from mood disorders to child and adolescent psychology, wherever a construct changes faster than a clinic visit can capture (aan het Rot, Hogenelst, & Schoevers, 2012; Russell & Gajos, 2020).

Worked Example

Follow the three demonstrations through one coherent study, checking that the arithmetic on the page matches the arithmetic in the demos.

Start with the sampling design. A researcher runs a signal-contingent protocol over a waking window of 14 hours and issues 7 prompts across it. The average interval between prompts is the window divided by the number of prompts: 14 / 7 = 2 hours. A denser protocol of 14 prompts over the same window would halve that interval to 14 / 14 = 1 hour, resolving finer within-day variation at the cost of greater burden. The design and the prompt count together fix how much of the day's variation the study can see.

Now the recall bias. On one day a participant supplies five momentary anxiety ratings on a 0-10 scale: 2, 4, 9, 6, 3. The true momentary mean is (2 + 4 + 9 + 6 + 3) / 5 = 24 / 5 = 4.8. Later the participant is asked to remember the day, and their memory follows the peak-end rule: it weights the peak moment (9) and the final moment (3), giving a retrospective estimate of (9 + 3) / 2 = 6.0. The recall bias is 6.0 − 4.8 = +1.2 points: memory reports the day as more anxious than the moments, averaged, actually were. EMA keeps the five moments and computes the 4.8; the retrospective questionnaire would have recorded only the 6.0.

Finally compliance. The study issues 6 prompts per day for 7 days, so the protocol on paper is 6 × 7 = 42 prompts. The participant answers at a compliance rate of 0.80, so the captured sample is 42 × 0.80 = 33.6, which rounds to 34 completed momentary observations, leaving 42 − 34 = 8 lost. The 34 clears a conventional 80% benchmark, but the number that matters for bias is not the rate alone — it is whether those 8 missed prompts fell at random or at the day's most anxious moments, which is exactly what a compliance analysis must check before the 34 captured reports can be trusted.

Discussion

Ecological momentary assessment is to the measurement of dynamic states what the standardized test is to the measurement of stable traits: the instrument the field built when it accepted that some quantities cannot be measured once. Csikszentmihalyi and Larson's insight that experience must be sampled in situ, and Stone and Shiffman's codification of that insight into a named method with reporting standards, gave psychology a way to study mood, craving, and symptom as they move rather than as they are remembered (Csikszentmihalyi & Larson, 1987; Stone & Shiffman, 1994). The three sampling designs, the momentary report's advantage over memory, and the compliance rate that gates data quality are the working parts of that instrument.

The tensions that remain are the ones the design creates. Repeated in-situ sampling defeats recall bias but introduces its own threats: participant burden, reactivity to being measured, and above all non-random non-compliance, which can bias a momentary dataset as surely as the peak-end rule biases a remembered one (Shiffman, Stone, & Hufford, 2008; Stone & Shiffman, 2002). And the intensively longitudinal, nested data EMA yields cannot be read with the tools of cross-sectional measurement; they demand the multilevel models the diary-methods tradition supplies (Bolger, Davis, & Rafaeli, 2003). The momentary report is a more faithful datum than the remembered one, and it is a report collected under a schedule whose design and compliance the analyst must scrutinize before trusting the average built from it.

Current Directions

Two active lines of work are extending EMA beyond assessment. The first is the move from measuring a state to acting on it. Building on the smartphone's ability to both sample and respond, researchers are developing ecological momentary intervention, in which a participant's real-time reports trigger treatment content in the moment it is needed, and are mapping its current state and open problems for disorders such as major depression (Colombo et al., 2019). In psychiatry the same technical developments — dense sampling, passive sensing, digital phenotyping — are reframing experience sampling as a route to real-time, in-context care rather than only a research instrument (Myin-Germeys et al., 2018).

The second is the disciplining of the method's data quality as it scales to new populations and devices. As mobile EMA spreads into child and adolescent research, meta-analytic work is quantifying how compliance depends on prompt design and study features, so that protocols can be built to the response rates their questions require rather than discovering the shortfall after the fact (Wen et al., 2017; Russell & Gajos, 2020). Across both lines the through-line is the same: the momentary, in-situ commitment that Csikszentmihalyi and Larson established is now being pushed simultaneously toward richer data — sensors alongside self-report — and toward intervention, while the methodological literature works to keep compliance and analysis rigorous enough to bear the weight (Trull & Ebner-Priemer, 2013).

Key Researchers

Niall Bolger. Professor of psychology at Columbia University; a leading methodologist of diary and intensive-longitudinal designs, the analytic family EMA belongs to, whose work on capturing life as it is lived and on the multilevel modeling of within-person processes set the standards for reading repeated momentary reports. ORCID - Faculty Page

Mihaly Csikszentmihalyi (1934-2021). Psychologist at Claremont Graduate University; co-developer of the Experience Sampling Method, EMA's direct methodological ancestor, and originator of flow theory, whose pager-prompted sampling established that subjective states must be measured in situ rather than reconstructed after the fact. Wikipedia

Ulrich W. Ebner-Priemer. Professor at the Karlsruhe Institute of Technology; a central figure in ambulatory assessment who extended EMA from self-report to real-time physiological and movement sensing, and whose work on affective instability showed that momentary sampling can quantify moment-to-moment mood dynamics a single questionnaire cannot see. Faculty Page

Reed W. Larson. Professor at the University of Illinois Urbana-Champaign; co-developer with Csikszentmihalyi of the Experience Sampling Method and its foundational validity and reliability study, whose work applied momentary sampling to the daily emotional lives of adolescents, demonstrating the method's reach beyond the clinic. ORCID - Faculty Page

Inez Myin-Germeys. Professor at KU Leuven; the leading contemporary developer of experience sampling in psychiatry, applying momentary assessment to psychosis and the reactivity of daily life, and driving the field's move from passive assessment toward real-time, in-context intervention and digital phenotyping. ORCID - Faculty Page

Saul Shiffman. Professor of psychology at the University of Pittsburgh; co-originator of ecological momentary assessment and the researcher who made event-contingent sampling central to addiction science, capturing craving, lapse, and their antecedents in the moment of smoking rather than through retrospective recall. ORCID - Faculty Page

Arthur A. Stone. Professor at the University of Southern California; co-originator of ecological momentary assessment, who named the method and set out its reporting standards, and whose research on the science of self-report established why momentary data avoid the memory heuristics that bias retrospective questionnaires. ORCID - Faculty Page

Timothy J. Trull. Professor of psychological sciences at the University of Missouri; a leading developer of ambulatory assessment who brought EMA into clinical personality science, using momentary sampling to operationalize affective instability in borderline personality disorder and to show that recalled and momentarily sampled variability are not the same measurement. ORCID - Faculty Page

Glossary

Affective instability.

The moment-to-moment variability of mood, a dynamic construct that becomes a single quantity only when computed from a dense momentary series, and that distinguishes clinical groups such as borderline personality disorder.

Ambulatory assessment.

The wider program of which EMA is the self-report core, adding continuous physiological, movement, and location sensing so that a momentary self-report can be aligned in time with objective signals.

Compliance.

The proportion of issued prompts a participant actually completes; the single most consequential quality metric of an EMA study, because non-random missingness biases the captured dataset.

Ecological momentary assessment (EMA).

A family of methods that repeatedly sample a person's states and behaviors in real time and in natural settings, to minimize the distortions of retrospective recall; a type of psychological test.

Ecological momentary intervention (EMI).

The use of a participant's real-time momentary data to deliver treatment content in the moment and context where it is needed, turning the assessment stream around into a delivery channel.

Event-contingent sampling.

A design in which the participant initiates a report whenever a defined event occurs — a cigarette, a panic attack, an argument — the only design that reliably captures rare or fleeting events.

Experience Sampling Method (ESM).

The pager-based method developed by Csikszentmihalyi and Larson in which random signals prompt participants to record their momentary activity and feeling; EMA's direct methodological ancestor.

Intensive longitudinal data.

Data with many repeated observations per person, nested as moments within days within people, which EMA yields and which require multilevel rather than cross-sectional analysis.

Interval-contingent sampling.

A design in which the participant reports on a fixed schedule, such as every evening or every three hours; simple to run but predictable, so it lets the participant anticipate the report. Also called time-contingent sampling.

Momentary report.

A rating of a person's present-moment state or behavior, collected as it occurs rather than reconstructed from memory; the raw datum of EMA from which summaries are computed by the analyst.

Peak-end rule.

The memory heuristic by which a remembered episode is evaluated chiefly by its most intense moment and its final moment rather than by the average of all its moments, a principal source of retrospective bias.

Reactivity.

A change in the very behavior or state being measured that is caused by the act of repeated self-monitoring; a threat to validity specific to intensive momentary sampling.

Recall bias.

The systematic divergence between a retrospective summary and the moments it is meant to summarize, produced by memory heuristics such as the peak-end rule and recency; the distortion EMA is designed to eliminate.

Signal-contingent sampling.

A design in which the device prompts the participant at researcher-controlled, often random, times, and the participant reports on the moment of the signal; the direct descendant of the ESM pager.

Frequently Asked Questions

What is ecological momentary assessment?
It is a family of methods that repeatedly sample a person's states and behaviors in real time and in their natural environments, so that the data are ecological, momentary, and repeated. It is a type of psychological test designed to minimize the recall bias that distorts a one-time retrospective questionnaire (Stone & Shiffman, 1994; Shiffman, Stone, & Hufford, 2008).

How is EMA different from a normal questionnaire?
A standard questionnaire measures a person once and asks them to summarize weeks of experience from memory. EMA measures the experience itself many times, as it happens, and leaves the summarizing to the analyst, which removes the memory step where bias enters (Stone & Shiffman, 2002; Bolger, Davis, & Rafaeli, 2003).

What are the three sampling designs?
Signal-contingent sampling prompts the participant at researcher-controlled, often random, times; interval-contingent (time-contingent) sampling uses a fixed schedule; and event-contingent sampling has the participant report whenever a defined event occurs. Each answers a different question, and studies often combine them (Shiffman, Stone, & Hufford, 2008).

Why is a momentary report better than a remembered one?
Memory does not average every moment; it reconstructs an episode using heuristics such as the peak-end rule and recency, which weight the most intense and most recent moments. A retrospective report therefore returns a memory-weighted distortion, whereas a momentary report returns the moment itself (Stone & Shiffman, 2002; Solhan et al., 2009).

Why does compliance matter so much?
Compliance is the proportion of prompts a participant answers, and it governs both the size and the trustworthiness of the dataset. The danger is non-random missingness: if participants skip prompts when most distressed or busy, the unrecorded moments are the ones of interest, biasing the captured data (Stone & Shiffman, 2002; Shiffman, Stone, & Hufford, 2008).

Did smartphones solve the compliance problem?
No. Smartphone delivery made EMA far easier to field, but a meta-analysis of mobile-EMA studies in children and adolescents found that response rates still vary widely and depend on prompt design and study features, so compliance remains a design problem, not a solved one (Wen et al., 2017).

What is the difference between EMA and ambulatory assessment?
EMA is the self-report core; ambulatory assessment is the wider program that adds continuous physiological, movement, and location sensing alongside the reports, so a self-reported state can be aligned with objective signals such as heart rate or activity (Trull & Ebner-Priemer, 2013; Ebner-Priemer & Trull, 2009).

Can EMA be used to deliver treatment, not just measure?
Yes. Ecological momentary intervention uses a participant's real-time data to trigger support in the moment it is needed, such as a coping prompt when stress rises, and is an active frontier in the treatment of disorders such as major depression (Colombo et al., 2019; Myin-Germeys et al., 2018).

References