When synthetic biologists insert an engineered gene into a cell, the gene does not run in isolation. It competes with the cell's own DNA for a finite pool of polymerases and ribosomes, and this competition — called metabolic burden — distorts both the engineered circuit and the host cell's behavior. Adding more of the gene-carrying plasmid can even lower total output, because the measurement and production machinery are the same physical stuff Characterization and mitigation of gene expression burden in mammalian cells. Behavioral science has a structurally identical problem that no one has modeled as such: a questionnaire measures a person's state using the person's own finite stock of attention, memory, and willingness to disclose. Every probe consumes some of the very resource it is trying to read. Survey researchers call the resulting distortions self-report bias — social desirability, recall error, response fatigue — but they treat these as fixed psychological foibles to be corrected with statistical patches, not as the predictable output of a resource-competition process.
The proposal here is to import the mathematics of metabolic burden into the measurement of human states. Three literatures sit close to this intersection without touching it. Comparative correlation analysis supplies the statistical machinery — descriptive profiling, group comparison, correlation, and longitudinal panel models The Descriptive Comparative Correlational Framework — and knows its Achilles' heel: shared-instrument artifacts inflate correlations, and cross-sectional designs cannot fix direction. Self-report bias research catalogs the distortions: underreporting of addictive behavior to protect self-image Measurement Invariance of the Bergen Social Media Addiction Scale Across Genders, retrospective symptom recall skewed by the respondent's acute physical state Incidence and Prevalence of Post-COVID-19 Myalgic Encephalomyelitis: A Report from the Observational RECOVER-Adult Study, and fatigue-driven dropout in daily probing Assessing Real-Time Moderation for Developing Adaptive Mobile Health Interventions for Medical Interns: Micro-Randomized Trial. And metabolic-burden research supplies exactly the dynamical models needed: mass-conservation equations for shared machinery A resource dependent protein synthesis model for evaluating synthetic circuits, "effective rate" models where every process slows as the shared pool depletes Characterization and mitigation of gene expression burden in mammalian cells, and closed-form stability analysis showing systems collapse onto a shifting capacity manifold rather than a fixed threshold A resource dependent protein synthesis model for evaluating synthetic circuits. The one existing paper that joins all three themes superficially, linking self-reported diet to plasma metabolites in 30,059 UK Biobank participants Associations of Ultra-Processed Food Intake and Its Circulating Metabolomic Signature with Mental Disorders in Middle-Aged and Older Adults, treats the metabolic side as a static biomarker score and never models the act of reporting at all.
Core Claim
Self-report bias is a resource-competition phenomenon, and therefore predictable and partially correctable with conservation-law models. If each questionnaire item or probe seizes a share of a finite respondent capacity, then four signatures must appear in existing datasets: total information gathered per day should be non-monotonic in probe frequency (the survey analog of the non-monotonic plasmid dose-response curve in mammalian cells Characterization and mitigation of gene expression burden in mammalian cells); spurious inter-item correlation should grow with position in the instrument and with respondent depletion, rather than acting as the single uniform bias factor that standard corrections assume; recall distortion should scale with the respondent's state at reporting time, as observed for post-COVID symptom recall Incidence and Prevalence of Post-COVID-19 Myalgic Encephalomyelitis: A Report from the Observational RECOVER-Adult Study; and intervention thresholds should be state-dependent manifolds rather than the static cutoffs (mood below 7, sleep below 6.9 hours) reported in the intern micro-randomized trial Assessing Real-Time Moderation for Developing Adaptive Mobile Health Interventions for Medical Interns: Micro-Randomized Trial.
The model is deliberately minimal. Let $C_{\mathrm{free}}(t)$ be the respondent's uncommitted reporting capacity at time $t$, depleted by active probe demands $u_j(t)$ with per-probe costs $k_j$, and recovering between probes:
$$C_{\mathrm{free}}(t) = C_{\mathrm{tot}} - \sum_j u_j(t)\, k_j, \qquad \hat{x}_i(t) = f_i\!\big(x_i(t),\, C_{\mathrm{free}}(t)\big) + \varepsilon_i.$$
Here $\hat{x}_i$ is the self-reported answer, $x_i$ the true state, $f_i$ a reporting function whose distortion grows as free capacity shrinks (mirroring the effective reaction rates of Frei et al.), and $\varepsilon_i$ ordinary noise. The structure — a shared, conservable pool coupling otherwise independent processes — is taken directly from the ribosome-flow conservation laws of Halter et al. A resource dependent protein synthesis model for evaluating synthetic circuits and the ribosome/RNase redistribution model MIRELLA MIRELLA: a mathematical model explains the effect of microRNA-mediated synthetic genes regulation on intracellular resource allocation, whose open codebase implements exactly this class of ODEs. Two mechanisms are inherited, not assumed: coupling through the shared pool (which generates spurious correlation — the biological twin of common method variance), and depletion dynamics (which generate fatigue, missingness, and non-monotonic information yield).
Research Program
The work has four parts. First, formalize the respondent resource model: derive from the conservation equations how inter-item correlation inflation, response latency, and missingness depend on survey length, item order, and time since the last probe. Second, build a correction procedure: given an observed correlation matrix or panel model, estimate the pool parameters and subtract the coupling-induced component of shared variance, yielding a resource-adjusted correlation estimate. Third, fit and test the model on three real datasets described below. Fourth, invert the model into a probing controller: if bias is a function of free capacity, then an adaptive schedule that probes only when $C_{\mathrm{free}}$ is high maximizes information per intrusion — the direct analog of the antithetic feedback controllers that adjust translation demand to free ribosome pools A coarse-grained bacterial cell model for resource-aware analysis and design of synthetic gene circuits.
A known constraint from the burden literature shapes the design: coarse-grained resource models break down under extreme stress, where alternative shock responses activate A coarse-grained bacterial cell model for resource-aware analysis and design of synthetic gene circuits. The human analog — crisis states that suspend ordinary reporting behavior — is exactly the regime where self-reports are least trustworthy, so the model must be validated in mild-to-moderate stress regimes first, and its failure mode under extreme depletion must be characterized rather than hidden.
Baselines
The do-nothing reference point is the raw self-report, uncorrected: in the intern study, daily 1–10 mood ratings are the operational measure, and any correction must demonstrably improve on them against an external criterion. The strongest existing method for the intervention-threshold question is the centered and weighted least-squares moderation estimator of NeCamp et al. Assessing Real-Time Moderation for Developing Adaptive Mobile Health Interventions for Medical Interns: Micro-Randomized Trial, which on 1,565 interns with Fitbit trackers found that mood negatively moderated notification effect with an estimate of $-0.052$ ($P=.001$) — at mood 3 notifications raised mood by 0.19 points, at mood 9 they lowered it by 0.12 points, with a beneficial cutoff below 7 — and analogous cutoffs of 10,614 steps and 6.9 hours of sleep. It will be re-run on the same data with the same sample, the same nightly EMA cadence, and the same tuning budget, differing only in that our model replaces the static cutoff with a state-and-history-dependent receptivity function; beating it means significantly better out-of-sample prediction of when a notification helps or harms, judged by held-out intervention outcomes. For the correlation-correction question, the strongest competitor is the original cross-lagged panel model of the Chinese college-student study, which achieved CFI = 0.999, TLI = 0.986, RMSEA = 0.032, SRMR = 0.010 on N = 1,181 Longitudinal relationships among problematic mobile phone use, bedtime procrastination, sleep quality and depressive symptoms in Chinese college students: a cross-lagged panel analysis, together with a standard latent marker variable correction applied to the same data; our claim is not better fit — near-perfect fit is the problem, not the goal — but that our correction changes substantive path conclusions in a direction validated on data with objective anchors. Finally, the key ablation isolates the shared pool: the identical dynamic model with the coupling term set to zero, so items deplete capacity independently. If the coupled and uncoupled versions perform equally, the burden analogy is doing no work and the proposal should be abandoned.
Why not just…?
Why not just apply standard psychometric corrections — Harman's single-factor test or marker variables? These methods assume the bias is one uniform, linearly additive factor across all items. The evidence against that assumption is direct: distortion is non-linear and state-dependent, with acute physical state bending retrospective symptom recall Incidence and Prevalence of Post-COVID-19 Myalgic Encephalomyelitis: A Report from the Observational RECOVER-Adult Study, and social-desirability underreporting varying by construct and gender Measurement Invariance of the Bergen Social Media Addiction Scale Across Genders. Our model makes a sharper, falsifiable prediction they cannot: bias magnitude should vary systematically with item position and probe history. The panel re-analysis tests this head-to-head on identical data.
Why not just use wearables and skip self-report modeling? NeCamp et al. already paired Fitbit tracking with daily EMAs Assessing Real-Time Moderation for Developing Adaptive Mobile Health Interventions for Medical Interns: Micro-Randomized Trial, and their own results show the limits: high non-compliance and missingness exactly under high stress, and no objective anchor at all for the mood construct the thresholds depend on. This is a real threat — wearables solve part of the problem, and our proposal does not replace them but builds on them as the anchor that makes self-report correction measurable. The decisive experiment is whether probe-load-aware modeling of the self-reports predicts divergence and missingness that wearable data alone cannot, and whether adaptive probing recovers information that fixed nightly schedules lose.
Why not just run elastic-net regression on biomarkers, as Yuan et al. already did? The UK Biobank analysis Associations of Ultra-Processed Food Intake and Its Circulating Metabolomic Signature with Mental Disorders in Middle-Aged and Older Adults is the closest existing work and, honestly, a strong baseline for association discovery — it linked 24-hour dietary recall to 251 plasma metabolites and reported a hazard ratio of 1.05 per 10% increase in ultra-processed food intake for mental disorder risk. But it is a static black box: it cannot predict how reporting accuracy degrades with load, when a measurement regime exits its linear operating range, or any non-monotonic behavior — precisely the regime where resource-competition models have proven linear methods fail A coarse-grained bacterial cell model for resource-aware analysis and design of synthetic gene circuits. The head-to-head is concrete: predicting the self-report–metabolite divergence vector (the per-person delta between Oxford WebQ recall and the metabolomic signature, which Yuan et al. compute for mediation but never isolate as an outcome) from participant covariates and probe context, elastic net versus our model, same folds. If elastic net matches us there, the dynamic machinery is unjustified — though it would still be needed for the longitudinal questions elastic net cannot address at all.
Decisive Checks
Three experiments, in increasing order of ground-truth quality. First, the divergence analysis on the UK Biobank subset (N = 30,059): construct the per-participant delta between self-reported intake and the objective metabolomic signature, and test whether the resource-aware model predicts this delta out of sample better than elastic net; this is the cleanest self-report-versus-objective comparison available, though its single-snapshot design cannot test dynamics, which is an admitted limitation. Second, the intern-study re-analysis (N = 1,565, six months of daily data): test the three dynamic predictions — information yield per probe declines with cumulative probe history; missingness is predicted by modeled depletion; and the notification tipping point is a function of objective state and recent probe load rather than the static mood-below-7 cutoff, with prediction of held-out intervention outcomes as the metric. Third, a simulation study built on the MIRELLA codebase MIRELLA: a mathematical model explains the effect of microRNA-mediated synthetic genes regulation on intracellular resource allocation: generate synthetic respondents with known true states and known pools, then ask whether resource-adjusted correlation estimates recover the true correlations where Harman's test and marker-variable corrections fail, and whether pool parameters are identifiable at all from realistic survey data. No reported performance number exists on Lacuna for the MIRELLA codebase itself, so the simulation supplies the first quantified evaluation of this model class as a measurement model. The single decisive criterion across all three: the coupled model must beat both the uncoupled ablation and the strongest existing method on held-out data in at least two of the three, or the measurement-burden hypothesis — however attractive the analogy — is wrong.
References
- Characterization and mitigation of gene expression burden in mammalian cells
- Measurement Invariance of the Bergen Social Media Addiction Scale Across Genders
- Incidence and Prevalence of Post-COVID-19 Myalgic Encephalomyelitis: A Report from the Observational RECOVER-Adult Study
- Assessing Real-Time Moderation for Developing Adaptive Mobile Health Interventions for Medical Interns: Micro-Randomized Trial
- A resource dependent protein synthesis model for evaluating synthetic circuits
- Associations of Ultra-Processed Food Intake and Its Circulating Metabolomic Signature with Mental Disorders in Middle-Aged and Older Adults
- MIRELLA: a mathematical model explains the effect of microRNA-mediated synthetic genes regulation on intracellular resource allocation
- A coarse-grained bacterial cell model for resource-aware analysis and design of synthetic gene circuits
- Longitudinal relationships among problematic mobile phone use, bedtime procrastination, sleep quality and depressive symptoms in Chinese college students: a cross-lagged panel analysis