Digital mental health studies track people's mood by asking them, several times a day, to rate how they feel on a short numbered scale. A recent benchmark found that the textbook linear model for such data does worse than simply predicting that the next rating will equal the last one. Vector autoregression, which predicts each variable as a weighted sum of recent values, fails in this way, while a piecewise-linear recurrent neural network forecasts well (Computational network models for forecasting and control of mental health trajectories in digital applications). The natural reading is that psychological dynamics are strongly nonlinear. That paper, however, feeds the raw ratings into its models as if they were the psychological state, or a noisy linear copy of it. Real ratings are bounded, coarse, and ordered, and people use the scale in their own way. Their personal cut-points drift as they tire of the study, and they skip prompts precisely when they feel worst. Each of these properties can make a perfectly linear hidden process look nonlinear on the page.
This note proposes a controlled test of where the nonlinearity lives. We fit a two-by-two family of state-space models, which are models with a hidden state that evolves over time and emits observations. One factor is linear versus piecewise-linear hidden dynamics. The other is a naive Gaussian readout versus an explicit self-report model: ordinal thresholds, person-specific and slowly drifting cut-points, and missingness that depends on the hidden state. Passive wearable signals serve as an anchor that does not share the self-report distortions. Randomized micro-interventions supply the variation needed to separate dynamics from measurement. The answer matters beyond forecasting accuracy. Impulse-response maps, intervention "tipping points," and network-control conclusions drawn from these models are only trustworthy if the nonlinearity they encode belongs to the person and not to the rating scale.
Core Claim
A substantial share of the forecasting gap between linear and nonlinear models of momentary self-report can be recovered by keeping the hidden dynamics linear and modelling the reporting process correctly. Our hypothesis is that the linear-dynamics, biased-observation model will close most of the gap to the piecewise-linear network on held-out ratings. We expect the nonlinear model's extra gain to shrink once it is given the same observation model. This is a claim that can fail, and the design treats failure as informative. If linear-latent models remain far behind even with a faithful reporting model, that is the strongest evidence yet that nonlinearity in psychological dynamics is real rather than a measurement artifact.
The proposal differs from the closest work in a specific way. Computational network models for forecasting and control of mental health trajectories in digital applications varies the dynamics and holds the naive observation model fixed. We vary both factors in a crossed design and ask how much of the observed gain each one explains. That paper motivates the study and supplies the strongest baseline.
Model
Let $x_t \in \mathbb{R}^d$ be a person's hidden affective state at prompt time $t$ and $u_t$ the randomized intervention indicator. The dynamics are either linear, $x_{t+1} = A x_t + B u_t + w_t$, or piecewise-linear, $x_{t+1} = A x_t + W \max(0, x_t) + B u_t + w_t$, where $w_t$ is Gaussian noise. The second form is the piecewise-linear recurrent network used in the benchmark above.
The self-report observation model links the state to the $k$-level rating $y_{t,j}$ on item $j$ through ordered cut-points:
$$ P(y_{t,j} \le k \mid x_t) = \Phi\!\left(\frac{\tau^{(i)}_{j,k}(t) - c_j^\top x_t - b^{(i)}(t)}{s^{(i)}}\right). $$
Here $\Phi$ is the standard normal distribution function and $c_j$ is the item loading. The term $\tau^{(i)}_{j,k}(t)$ is person $i$'s cut-point between levels $k$ and $k+1$, which may drift slowly, as a random walk with small variance, to capture fatigue or scale reinterpretation. $b^{(i)}(t)$ is a slowly varying response offset, for example a social-desirability shift after enrollment, and $s^{(i)}$ is a person-specific response spread.
Missingness is modelled as $P(\text{prompt answered} \mid x_t) = \sigma(\alpha^{(i)} + \gamma^\top x_t)$, with $\sigma$ the logistic function, so that skipping can depend on how the person feels. Passive signals such as nightly sleep duration and daily steps enter through a separate linear-Gaussian channel with no drift terms. That channel is what pins down the hidden scale.
Drift in cut-points and drift in the hidden state are not separable from ratings alone, because both shift the ratings. Two sources identify them here. The first is the sensor anchor. The second is randomized interventions, which generalize the argument of An Interventional Perspective on Identifiability in Gaussian LTI Systems with Independent Component Analysis. That paper shows that varied inputs make linear-Gaussian system matrices identifiable, but it assumes a linear-Gaussian readout. We borrow its logic and state plainly that our extension to ordinal readouts is argued and tested in simulation, not proven. Inference uses a variational smoother with a learned encoder. MissNet's regime-switching Kalman smoother (Mining of Switching Sparse Networks for Missing Value Imputation in Multivariate Time Series) is a simpler Gaussian reference point for the imputation behaviour.
Research Program
The first stage is synthetic and decides whether the study can work at all. We simulate linear hidden dynamics observed through drifting ordinal scales with state-dependent skipping, at the sampling density and missingness levels of the real studies. On these data we check two things: that vector autoregression on raw ratings loses to last-value persistence, and that a naive-readout piecewise-linear network appears to find nonlinearity. We then simulate genuinely nonlinear dynamics seen through a clean Gaussian readout. The linear-latent ordinal model must not falsely win there. If the crossed design cannot tell these two worlds apart in simulation, the empirical study is uninterpretable and we stop.
The second stage fits all four model cells to the AI4U living-lab micro-randomized trials. These consist of three 40-day samples, 145 participants in total, with 15 well-being items and 8 micro-interventions. We use the original paper's train and test splits.
The third stage reanalyses the 2018 Intern Health Study micro-randomized trial (Assessing Real-Time Moderation for Developing Adaptive Mobile Health Interventions for Medical Interns: Micro-Randomized Trial). It has 1,565 interns, daily 1–10 mood ratings, and Fitbit sleep and steps. The authors report that notifications raise mood below a rating of about 7 and lower it at high mood. We re-estimate this moderation on the inferred hidden scale.
A caveat governs the interpretation. Ceiling compression on a bounded scale can shrink a positive effect at high mood but cannot by itself flip its sign. If the negative effect survives on the hidden scale, the reanalysis confirms that the harm is real. What the reanalysis can change is the tipping point's location and the size of the effect, and those are what an adaptive intervention policy would use.
The fourth stage uses GLOBEM (GLOBEM Dataset: Multi-Year Datasets for Longitudinal Human Behavior Modeling Generalization). Its weekly depression questionnaire scores come with continuous passive sensing over four yearly cohorts. The question is how much of the documented cross-year collapse comes from drift in how students use the questionnaire rather than drift in behaviour. We fit cohort-specific cut-points with shared dynamics and sensor loadings, and ask whether cross-year prediction of the hidden state recovers.
Baselines
The naive reference point is last-value persistence together with the per-person running mean, run on the same held-out prompts. Computational network models for forecasting and control of mental health trajectories in digital applications showed that vector autoregression often fails to beat persistence. Any model in our design that does not beat it is uninformative about dynamics.
The strongest existing method is that paper's piecewise-linear recurrent network with its standard readout. It reports forecasting mean absolute error of 0.831 on Sample 2 and 0.795 on Sample 3. We will first reproduce these numbers with their code and hyperparameter protocol, and then give every cell of our design the same latent dimension range, training budget, and number of tuning trials. Beating it means our linear-latent ordinal model matches or improves held-out error on these samples. The match is judged by paired per-participant differences with bootstrap intervals, not by pooled means.
Linear vector autoregression and a linear Kalman filter on raw ratings form the linear-naive cell. The paper benchmarks them, but no reported result found on Lacuna for their numeric error values, so we report our own reruns.
The key ablations isolate the reporting model piece by piece. We fit the linear-latent model with ordinal thresholds but no drift, then add drift but not state-dependent missingness, then run the full version. Only this sequence says which reporting mechanism carries the effect.
For the Intern Health Study, the baseline is the authors' weighted, centered least-squares moderation estimate. For mood they report a moderation estimate of −0.052 (P=.001) and a 0.12-point drop at a mood of 9.
For GLOBEM, the reference is the reported within-year balanced accuracy of 0.691 and the 15.9% cross-dataset gain of the Reorder method, evaluated under the benchmark's leave-one-year-out protocol.
Why not just…?
Why not just use the piecewise-linear network, since it already wins? It wins at forecasting on the observed scale, but its impulse-response maps are read as statements about how one feeling drives another. If the nonlinearity it learns is the scale's floor and ceiling, those maps describe the questionnaire. The fourth cell of our design, the piecewise-linear network with the ordinal readout, tests this directly. If its learned nonlinear terms shrink toward zero once the readout is corrected, the original interpretation was partly an artifact. If they persist, the interpretation stands.
Why not just rank-transform each person's ratings to normal scores and run vector autoregression? This is the most serious threat. A per-person Gaussian copula handles individual scale use and boundedness at almost no cost. It extends the subject-dependent standardization that cut error by 25% in Personalised depression forecasting using mobile sensor data and ecological momentary assessment. We include it as a baseline and expect it to recover part of the gap.
The copula cannot represent cut-points that drift within a person, and it treats skipped prompts as ignorable. Our ablation ladder therefore predicts exactly where it should fall short: in late-study weeks, and around runs of missed prompts. If it matches the full model everywhere, we will report that the cheap fix suffices. That is still a useful answer to the central question.
Why not just add the lagged outcome and prior mood to the moderation regression in the Intern Health Study? Because treatment is randomized, the authors' estimator is already unbiased for effects on the observed scale. Adding covariates does not change which scale is measured. The question we ask is whether the tipping point sits at the same place on the hidden scale, and that requires a model of how hidden mood maps to ratings.
Proposed Experiments and Decisive Checks
The primary metric is held-out mean absolute error of the predictive median on the original rating scale, so that every cell is scored exactly like the published baseline. We add the ranked probability score to reward calibrated ordinal forecasts, and held-out sensor prediction error to check that the hidden state tracks something outside the questionnaire.
There are three decisive checks.
The first is the share of the gap between linear-naive and nonlinear-naive models that is recovered by the linear-ordinal model on AI4U. A share above one half supports the claim. A share near zero refutes it.
The second is whether the nonlinear model's advantage over the linear model narrows when both use the ordinal readout. This is the interaction term of the crossed design.
The third concerns intervention response. The original study predicted intervention responses correctly in 70–81% of participants. Using the same criterion, we ask whether linear-latent ordinal models predict responses to the randomized micro-interventions as well. This matters because the controllability and impulse-response analyses that motivate linear network control, as in Emergence of bimodality in controlling complex networks, are far more tractable under linear dynamics. Showing that linear dynamics are adequate once measurement is handled would make those tools usable again in this domain. Showing that they are not would settle, for now, that the nonlinearity is psychological.
References
- Computational network models for forecasting and control of mental health trajectories in digital applications
- An Interventional Perspective on Identifiability in Gaussian LTI Systems with Independent Component Analysis
- Mining of Switching Sparse Networks for Missing Value Imputation in Multivariate Time Series
- Assessing Real-Time Moderation for Developing Adaptive Mobile Health Interventions for Medical Interns: Micro-Randomized Trial
- GLOBEM Dataset: Multi-Year Datasets for Longitudinal Human Behavior Modeling Generalization
- Personalised depression forecasting using mobile sensor data and ecological momentary assessment
- Emergence of bimodality in controlling complex networks