People who fill in daily mood or sleep questionnaires on their phones do not answer at the moment the thing they describe happened. A student reports last night's sleep at 9 a.m. one day and 6 p.m. the next. In between she may have spent four hours scrolling her phone in bed or an afternoon outdoors with friends. Memory research suggests these differences matter. Recall is rebuilt, not replayed, and the mood a person is in when answering colors what they remember. Mobile mental-health studies nonetheless store each answer as a single number for that day. This proposal asks whether the timing pattern between a behavior and the report about it predicts how wrong the report is. The timing pattern means how long it was, and what sequence of events, separated by what gaps, filled that time.
The test is possible because some self-reported items have a sensor counterpart. Self-reported sleep can be compared with sleep inferred from the phone. The sensor's error is fixed when the night happens, so it cannot depend on when the person later chooses to answer. Any systematic dependence of the report-minus-sensor gap on what happened afterward is therefore report-side error. We build a model of that error on items that can be checked. We then ask whether applying it to items that cannot be checked, such as mood and stress, makes those reports more useful for predicting later clinical depression scores.
Core Claim
Self-report error in ecological momentary assessment (EMA, short phone surveys prompted during daily life) has a time-varying component beyond a fixed personal offset. That component is predicted by the interval-annotated sequence of sensed events between the reported-on behavior and the report. A model of this component, learned where sensor checks exist, transfers to unverifiable items. Concretely, for person $i$ and report $e$, write the discrepancy as
$$d_{i,e} = r_{i,e} - s_{i,e} = b_i + f(W_{i,e}) + \epsilon_{i,e},$$
where $r$ is the self-reported value, $s$ the sensor estimate of the same quantity, $b_i$ a person-specific constant offset, $W_{i,e}$ the window of sensed events between the behavior and the report (each event tagged with its type, duration and the gap preceding it), $f$ a learned function, and $\epsilon$ noise that includes sensor error. The proposal's content lies in two things. The first is whether $f$ is non-trivial once $b_i$ is removed. The second is whether $f$ depends on the structure of $W$ beyond its total length.
What This Builds On and Where It Differs
The interval side supplies the representation. Efficient Mining of Temporally Annotated Sequences formalized patterns of the form "A, then B after a typical interval Δt," and we use that form for the behavior-to-report window. Matching State-Based Sequences with Rich Temporal Aspects defines a sequence distance that separates event order, how long each state lasts, and gap length. Its ablation showed that order alone gave 87.50% precision, versus 94.00% with all three on AT&T face. That result motivates our bet that gaps carry signal, though it was obtained on image-derived sequences, not on human behavior. TimelyGPT conditions a transformer directly on elapsed time between irregular clinical events. We reuse that "feed the exact gap to the model" mechanism as one choice for $f$. TimelyGPT only ever predicts objective records and has no notion of a biased reporter.
Two papers constrain the design. Effective Anomaly Detection in Smart Home by Integrating Event Time Intervals found that timing-anomaly accuracy fell to 45% for a high-variance routine. Naive fixed-interval profiles therefore will not survive human variability, which is why we model gaps relative to each person's own distribution. Temporal Deviations on Event Sequences builds an all-pairs signature that grows with the square of the vocabulary size. We avoid that blow-up by encoding only the window, not all pairs.
The self-report side supplies the problem and the closest competitors. Recall bias in retrospective symptom and diet reporting is documented in the RECOVER-Adult post-COVID ME/CFS report and the ultra-processed food and mental disorders study, but only at the scale of months and without any timing model. Three sensing studies pair EMA with passive data but treat the EMA value as a fixed daily quantity:
- Personalised depression forecasting on MAIKI removes a static per-person offset by subject-dependent standardization. That is exactly $b_i$ above; it has no $f$.
- The StudentLife spatiotemporal graph model concatenates EMA into daily node features and treats it as truth.
- The Intern Health Study micro-randomized trial works with weekly averages.
None of them measures report error, and none looks inside the day.
Research Program
The first stage is construction. On StudentLife, we pair each self-reported sleep item with the phone-inferred sleep estimate for the same night. For each report we extract the window from inferred wake time to report submission. The window contains phone-use sessions, screen-lock periods, stationary versus moving bouts, and conversation or call episodes, each tagged with duration and preceding gap. We normalize gaps and durations within person, so that "a long gap" means long for that individual. We build the same windows where MAIKI or the Intern Health Study data permit. The Intern Health Study has Fitbit sleep but only mood EMA, so it enters only in the transfer stage. MAIKI's sleep behavior features may allow a second calibration set if access is granted.
The second stage fits $f$ at three levels of richness, all after removing $b_i$:
- Total window length as a single number, the recall-delay hypothesis.
- The order of events in the window without gaps.
- The full interval-annotated window, encoded either by a gap-conditioned sequence model in the style of TimelyGPT or by nearest-neighbor regression under the GSM distance.
The comparison between these levels is the scientific question.
The third stage is interpretation. We use TaMIPER, which mines rules leading to a user-chosen target event at exact time offsets, with the target set to "large discrepancy." The aim is to recover readable precursor patterns such as "more than 90 minutes of in-bed phone use within 2 hours before the report." This is descriptive and not the main evidence.
The fourth stage is transfer. We apply the learned $f$ to mood and stress EMA on the same people, adjusting each report by its predicted window-dependent error. We then test whether the adjusted reports predict later outcomes better than the raw ones. On StudentLife the outcome is the end-of-term PHQ-9 (a standard 9-item depression questionnaire). On MAIKI it is next-day PHQ-2, a 2-item short form. The transfer assumption is that timing-driven distortion is shared across items. That assumption may fail, and stage four is where we find out.
The main identification threat is that people who report late differ from those who report early. We address it with within-person estimation, so person effects absorb who reports late. Where prompt delivery times vary for reasons unrelated to the person's state, we use that variation directly. A negative control checks that the sensor estimate's own error, measured as agreement between two sensor-side sleep proxies where available, does not depend on the later window.
Baselines
The naive reference is the uncorrected self-report, $f = 0$ and $b_i = 0$. For discrepancy prediction this means predicting the population mean gap. For downstream use it means feeding raw EMA into the forecaster.
The strongest existing method for removing self-report skew is subject-dependent standardization from the MAIKI forecasting study. That paper reports an MAE of 0.801 for same-day PHQ-2, against 1.062 non-personalized. For one-day-ahead PHQ-2 its best result was an MAE of 1.349 via transfer learning, against a 1.539 baseline. We implement standardization as the $b_i$-only model and fit it on the same burn-in data our method gets.
For downstream StudentLife prediction, the strongest reported model is the spatiotemporal graph network from the StudentLife graph study, with PHQ-9 RMSE 3.5 ± 3.1 using all features. We rerun it with raw versus window-adjusted EMA features under identical splits and tuning budgets. This keeps the comparison about the input, not the architecture.
For discrepancy prediction itself, no reported result was found on Lacuna for predicting EMA-minus-sensor error on any of these datasets. That baseline therefore has to be established here.
The key ablations are the three richness levels of $f$. The full model must beat total-window-length by a margin larger than the spread across person-level cross-validation folds. It must also beat order-only. A gap-shuffled control keeps event order and multiset of durations but permutes the gaps. If it performs as well as the full model, the claim that gaps matter is false.
All models get the same per-person burn-in length, the same hyperparameter search budget, and leave-person-out evaluation. "Beating" means two things:
- lower held-out discrepancy MAE than the $b_i$-only model;
- downstream gains on PHQ outcomes that exceed fold-to-fold standard deviation.
Gains on discrepancy alone do not count.
Why not just…?
Why not just subject-dependent standardization? It is the right fix for a constant personal offset, and it is built into our model as $b_i$. The question is whether anything remains after it. If the residual discrepancy shows no relation to window structure after standardization, our claim fails and standardization is sufficient. The experiment reports this directly. Standardization also needs weeks of burn-in, per the MAIKI study. A window model trained across people could apply from the first report, and we test that separately with zero-burn-in evaluation.
Why not just add "hours since waking" as a covariate? This is the real threat. A single delay number in a gradient-boosted model may capture most of the recall effect. It is our first ablation level, not a strawman. If it matches the full interval sequence, we report that recall delay, not window content, drives report error. That is still a new measurement result for EMA, but it would demote the sequence-representation contribution. The interval-sequence part of the proposal earns its place only if level 3 beats level 1.
Why not just use daily or weekly averages? The Intern Health Study trial found clear moderation effects from weekly aggregates. However, averaging removes report timing by construction, so it cannot test, let alone correct, timing-dependent error. We compare against daily-aggregate features in the downstream task. If within-day windows add nothing there, the practical case for them is weak even if they explain discrepancy.
Why not just drop self-reports and use sensors? WiFi features alone matched the full StudentLife model on PHQ-9 RMSE (3.5 ± 3.4) in the graph study, so this is plausible. On the other hand, the MAIKI study found EMA added substantially beyond passive data. We include a sensors-only arm in every downstream comparison. Corrected EMA must beat both raw EMA plus sensors and sensors alone.
Decisive Checks
Four results settle the proposal, all on held-out people:
- After removing person offsets, does window structure predict self-reported-minus-sensed sleep better than window length alone, and better than a gap-shuffled control?
- Does the sensor-side negative control show no such dependence?
- Does applying the sleep-trained error model to mood and stress EMA lower PHQ-9 RMSE on StudentLife?
- Does the same correction lower next-day PHQ-2 MAE on MAIKI, relative to standardization alone, the published graph model, and sensors alone?
A positive first result with a negative third and fourth would mean timing-driven error is real but item-specific. That would still change how EMA sleep items should be collected. Positive results on all four would mean EMA reports should be stored and modeled together with the event history between behavior and response, not as one number per day.
References
- Efficient Mining of Temporally Annotated Sequences
- Matching State-Based Sequences with Rich Temporal Aspects
- TimelyGPT: Recurrent Convolutional Transformer for Long Time-series Representation
- Effective Anomaly Detection in Smart Home by Integrating Event Time Intervals
- Temporal Deviations on Event Sequences
- Incidence and Prevalence of Post-COVID-19 Myalgic Encephalomyelitis: A Report from the Observational RECOVER-Adult Study
- Associations of Ultra-Processed Food Intake and Its Circulating Metabolomic Signature with Mental Disorders in Middle-Aged and Older Adults
- Personalised depression forecasting using mobile sensor data and ecological momentary assessment
- Monitoring Behavioral Changes Using Spatiotemporal Graphs: A Case Study on the StudentLife Dataset
- Assessing Real-Time Moderation for Developing Adaptive Mobile Health Interventions for Medical Interns: Micro-Randomized Trial
- Targeted Mining Precise-Positioning Episode Rules