Routes as Witnesses: Measuring Recall Bias in Depression Self-Reports with Geometric Routine Deviation

Mobile mental-health studies train models on phone sensor data to predict answers to questionnaires such as the PHQ-9, a nine-item depression screener that asks people to rate how they felt "over the last two weeks." That question depends on memory, and memory is not a fair average. People tend to overweight the last few days and the mood they are in while answering. Current work either ignores this and treats the questionnaire as ground truth, or tries to avoid it with frequent in-the-moment surveys, which are themselves self-reports with their own biases and missing days. We propose to use a person's GPS trajectories as an independent witness. Each day we measure how far that day's route departs, in shape, from the person's own habitual routes, using distances that compare whole curves (Fréchet, Hausdorff, and segment-projection edit distance) rather than matching point to point. We then ask how a retrospective report weights the days it claims to summarize.

The central quantity is a recall-weight curve. It says how much each day in the recall window contributes to the reported score, estimated against a signal the reporter's memory did not produce. A flat curve means the person summarized the window faithfully. A curve piled onto the last days, or one that shifts with mood on the day of reporting, is a direct, per-cohort measurement of recall bias. The curve can then be used to correct the labels that sensing models are trained on.

Core Claim

Retrospective depression scores can be decomposed into a weighted memory of daily states plus a current-mood term. The weights are identifiable when two things are available: a daily behavioral signal that is not itself a self-report, and a set of in-the-moment reports that serve as a no-recall reference. The recall bias is the excess recency, relative to that reference, of the retrospective score's weights on the same behavioral witness.

We claim that routine deviation measured geometrically is a better witness for this purpose than the scalar mobility summaries the field uses now. By scalar summaries we mean location variance, distance traveled, and time at home, the features used by Depression Diagnosis and Forecast based on Mobile Phone Sensor Data and Personalised depression forecasting using mobile sensor data and ecological momentary assessment. The reason is that routine disruption is a comparison between trajectories, not a level of one trajectory. Two days with the same mileage and the same home time can follow the usual route or an entirely different one. The geometric claim is secondary and falsifiable: if scalar deviation measures yield equally reliable weight curves, we will report that, and the recall-bias measurement stands on its own.

This differs from the closest intersection work in two ways. The MAIKI studies and the StudentLife spatiotemporal graph model (Monitoring Behavioral Changes Using Spatiotemporal Graphs: A Case Study on the StudentLife Dataset) use mobility to predict the self-report and never question the label. Time Window Frechet and Metric-Based Edit Distance for Passively Collected Trajectories designs curve distances for exactly this kind of sparse phone data but evaluates nothing on behavioral cohorts. We use its distances to audit the label.

Research Program

Building the witness. For each participant we segment GPS into days, resample by arc length, and form a small set of personal habit prototypes. These are medoids of weekday and weekend days under the chosen distance. The split follows the spirit of the stable-habit versus transient-intent decomposition in Predicting Human Mobility via Self-supervised Disentanglement Learning, which learns this split for next-place recommendation but never links it to mental state. The daily witness $D_t$ is the distance from day $t$'s trajectory to its nearest prototype.

We compare three families of distance, each chosen to address a known weakness of the others:

  • Time-Window Fréchet (Ding et al.) respects visit order while tolerating the irregular time gaps of passive logging.
  • Edit Distance with Projections (EDwP) (Indexing and matching trajectories under inconsistent sampling rates) projects missing points onto segments, which suits the duty-cycled sampling phones produce.
  • Hausdorff distance ignores order and serves as the "same places, any sequence" contrast.

Quadratic Fréchet cost is not a practical obstacle here. Comparisons are within-person, against a handful of prototypes, on daily traces of a few hundred resampled points. Straight-segment ambiguity, a known failure of shape-only matching (A Shape-Based Map Matching Approach for Geographic Transferability of Discriminative Subtrajectories), matters less because we keep absolute coordinates within a person.

The recall model. Let $S_r$ be a retrospective score reported on day $r$, and $D_{r-k}$ the witness $k$ days earlier, for $k = 0,\dots,K-1$ over the recall window of length $K$. Let $m_r$ be the in-the-moment mood reported on day $r$. We fit

$$S_r = \alpha_i + \beta \sum_{k=0}^{K-1} w_k D_{r-k} + \gamma\, m_r + \varepsilon_r, \qquad \sum_k w_k = 1,\; w_k \ge 0,$$

where $\alpha_i$ is a per-person intercept and $w_k$ is the recall-weight curve. The weights are regularized to vary smoothly in $k$, and $\gamma$ captures mood-congruent shading.

The same curve is fitted with the retrospective score replaced by the mean of the daily in-the-moment reports over the same window. Call that reference curve $w^{\text{EMA}}_k$. This reference absorbs everything that is not memory: the lagged effect of behavior on mood, autocorrelation in the witness, and any tendency for disruption late in the window to matter more clinically. The recall bias is the difference between the two curves. We summarize it as the recency index, defined as the share of total weight in the last three days, computed for $w_k$ minus the same share for $w^{\text{EMA}}_k$. Mood-congruent shading shows up as a $\gamma$ that remains after the witness is included.

Using the curve. Given estimated weights, we construct recall-corrected targets. A corrected target re-expresses each retrospective score as the flat-weight average implied by the model. We then test whether sensing models trained on corrected targets predict held-out window averages of daily reports better than models trained on raw PHQ-9.

This also gives a concrete reading of a puzzle in Depression Diagnosis and Forecast based on Mobile Phone Sensor Data. There, next-week forecasting (RMSE 4.094) slightly beat concurrent diagnosis (RMSE 4.184), which the authors attribute to sensors acting as leading indicators. If retrospective scores lean heavily on the report day, sensor data from earlier in the window underfits the concurrent label, and the recall curve would predict exactly this pattern.

Data. The primary cohort is MAIKI, which pairs GPS with daily PHQ-2 reports and weekly PHQ-9 (Personalised depression forecasting using mobile sensor data and ecological momentary assessment). It therefore has both the retrospective target and the no-recall reference. StudentLife (StudentLife dataset) provides GPS, daily mood and stress reports, and PHQ-9. We will use it for replication at whatever PHQ-9 frequency the released data actually supports; if only pre/post scores exist, it serves only for validating the witness. The Intern Health Study (NeCamp et al.) has steps and sleep but no GPS or retrospective scale, so it is out of scope for the core test.

Baselines

Do nothing. The naive reference point treats PHQ-9 as ground truth, which implies a flat recall curve and $\gamma = 0$. Every downstream model in the file makes this assumption. If our estimated curve cannot be distinguished from flat under a bootstrap over participants, the proposal's central claim fails.

Strongest existing predictive systems. For the downstream comparison, we reimplement the scalar-feature LSTM of He et al., which reports PHQ-9 forecasting RMSE of 4.094 on MAIKI. We also use the subject-dependent standardization model of Personalised depression forecasting using mobile sensor data and ecological momentary assessment, which reports daily PHQ-2 MAE of 0.801 against a non-personalized MAE of 1.062. On StudentLife, the reference is the spatiotemporal graph model, which reports PHQ-9 RMSE of 3.5 ± 3.1 (Monitoring Behavioral Changes Using Spatiotemporal Graphs). Its standard deviation is almost as large as the error, which already suggests noisy labels.

These are the right competitors because they represent the field's best use of the same sensors under the uncorrected-label assumption. Each will be retrained on identical participant splits, subject-independent folds as in He et al., and with the same hyperparameter search budget. The only change is swapping raw targets for corrected ones. Beating them means lower error against held-out window averages of daily reports; a better fit to raw PHQ-9 is not the goal.

Witness ablations. These isolate the key ingredient. The scalar-deviation witness is the absolute daily departure of distance traveled, location variance, and home time from each person's median, fitted into the identical recall model. The point-wise witness is timestamp-interpolated Euclidean RMSE to the prototype. The discrete-edit witness uses DTW, EDR, and LCSS. For context on what to expect, EDwP is reported to be up to 5x more accurate than EDR and LCSS for trajectory matching on Beijing taxi data (Ranu et al.); no reported result was found on Lacuna for any of these distances on a mental-health cohort.

A geometric witness counts as better if it gives:

  • higher within-person validity, measured as the correlation of $D_t$ with same-day PHQ-2;
  • higher day-to-day reliability on held-out weeks; and
  • narrower bootstrap intervals on the recency index at equal sample size.

Why not just…?

Why not just estimate recall weights from the daily surveys alone? Regressing weekly PHQ-9 on the preceding daily PHQ-2 reports does estimate a recall curve. It is a genuine competitor, and we will run it and report it first. Its limitation is that the same mood that shades the retrospective report on day $r$ also shades that day's in-the-moment report. Mood-congruent bias therefore partly cancels, and $\gamma$ is not separately identifiable. Daily surveys are also missing on exactly the days people disengage (Predicting Participant Compliance With Fitness Tracker Wearing and Ecological Momentary Assessment Protocols in Information Workers), and prompts can change the states they measure (NeCamp et al.). The decisive check is whether the passive-witness model recovers a nonzero $\gamma$ and a recency excess where the survey-only model sees none. If both analyses give the same answer, we will say the passive witness added nothing for this question.

Why not just use scalar mobility summaries as the witness? They may well suffice, since mobility scalars already carry most of the passive signal on StudentLife (mobility-only RMSE of 3.5 equals all-features RMSE; Monitoring Behavioral Changes Using Spatiotemporal Graphs). Our argument that shape catches disruption that level does not is plausible but unproven. It is settled directly by the witness ablation above. If scalar deviation matches geometric deviation on validity, reliability, and interval width, the recall-bias finding stands and the geometric component is dropped.

Why not just train a sequence model on raw GPS and let it learn whatever matters? A learned model would be fitted to the biased label and would absorb the recall weighting rather than expose it. The point of the witness is to be defined without reference to the self-report. A learned encoder trained on location alone (for example, self-supervised as in Gao et al.) is admissible as an additional witness. We include it if compute allows, under the same no-label constraint.

Proposed Experiments

Witness validity. On MAIKI and StudentLife, we compute each witness per day and report within-person correlation with same-day mood and next-day mood. We also report week-to-week reliability and sensitivity to artificially thinned GPS sampling. This last test provides the first empirical check of the Time-Window Fréchet distance on real passive cohorts.

The main test. On MAIKI, we fit the recall model for weekly PHQ-9 and the reference model for window-averaged PHQ-2, using each witness. We report both weight curves, the recency index difference with participant-level bootstrap intervals, and $\gamma$. The claim is supported if the retrospective curve shows significantly more recency than the reference curve and $\gamma$ is positive after the witness is included. It is refuted if the curves coincide.

A placebo check guards against artifacts. We shift the witness window forward past the report day. Weights on days after the report are logically impossible for memory, so they should vanish; if they do not, the curve is picking up confounding by trend rather than recall.

Downstream correction. We retrain the He et al. LSTM and the personalized MAIKI model on corrected versus raw targets and score both against held-out window means of daily reports. We also check whether the forecast-versus-concurrent gap of 4.094 versus 4.184 shrinks under corrected targets, as the recall account predicts.

Heterogeneity. We test whether recall curves differ by gender. This mirrors the measurement-invariance question asked for questionnaire structure in Measurement Invariance of the Bergen Social Media Addiction Scale Across Genders, but here it is asked of memory weighting against a behavioral witness rather than of factor structure. The subgroup analysis is kept to one pre-registered comparison because of the small cohorts.