Correcting Under-Reporting at the Right Layer of Fleet Dispatch

City repair crews, inspectors and delivery fleets are usually dispatched in two steps. First the city is split into territories, one per vehicle. Then each vehicle plans its own route through the jobs inside its territory. Both steps take citizen complaints at face value, and citizen complaints are self-reports of the physical world. They are biased in two known and measurable ways. Some neighborhoods report fewer of the incidents that actually occur, and some report the incidents they do file much later. For identical tree damage, downtown Manhattan reports about three times faster than Queens (Quantifying spatial under-reporting disparities in resident crowdsourcing). Consider a dispatcher that sizes territories by report counts and ranks jobs by how recently they were reported. It gives low-reporting areas too few vehicles. Within their routes, it also treats their stale incidents as fresh.

The idea of this note fits in two sentences. Each of the two biases should be corrected in the dispatch step whose timescale matches it. Missing reports should be corrected where territories are drawn, because that step is slow and city-wide. Late reports should be corrected where routes are ordered, because that step is fast and local. We argue that this split also resolves a real architectural tension. Estimating under-reporting needs city-wide pooling across demographically similar but distant neighborhoods, while scalable routing needs only local information. We will test the claim in a semi-synthetic but data-grounded simulator of NYC and Chicago 311 service (311 is the city non-emergency complaint line). The test asks whether "right correction, right layer" serves more true incidents before they expire, with smaller gaps between demographic groups, than correcting everywhere, correcting nowhere, or correcting after the fact.

Core Claim

Partition-routing decoupling is the practice of splitting "who covers which area" from "in what order each agent visits its jobs." It is well established as the way to make multi-agent dispatch scale. Decentralized vehicle routing in a stochastic and dynamic environment with customer impatience moves territory centers by local gradient steps toward equal-area Voronoi cells, where a Voronoi cell is the set of points closer to one agent's center than to any other. Each vehicle then runs its own traveling-salesperson tour, and the paper proves the fleet needs at most about 2.6 times the theoretical minimum number of vehicles. Clustering Based Priority Queue Algorithm for Spatial Task Assignment in Crowdsourcing (TAMP) builds territories by spectral clustering of task density. Inside each territory it orders jobs with a score that mixes distance and deadline urgency. Both methods assume that the observed jobs are the true jobs and that the report time is the incident time.

Under-reporting research has recently become quantitative, but it stops before dispatch. The duplicate-report method in Quantifying spatial under-reporting disparities in resident crowdsourcing estimates each neighborhood's reporting rate from how often the same incident is reported again. It fits a zero-inflated Poisson model to census-tract features. A Bayesian Spatial Model to Correct Under-Reporting in Urban Crowdsourcing treats tracts with no reports as a mix of true negatives and missed events. It infers a latent flooding field with a spatial Ising prior, a model in which neighboring tracts tend to share the same state. It then chooses 100 tracts to inspect by posterior probability, with no travel times, capacities or routes.

We claim that the correct composition is asymmetric. Territories should be balanced on estimated true workload,

$$w(z) = \frac{\hat\lambda_{\text{obs}}(z)}{\hat\psi(z)},$$

where $\hat\lambda_{\text{obs}}(z)$ is the smoothed observed report rate at location $z$ and $\hat\psi(z)$ is the estimated probability that an incident there is ever reported. Routes should rank each reported job $j$ by its estimated remaining life rather than its apparent one,

$$\hat r_j = D_j - (t - t^{\text{rep}}_j) - \hat d(z_j),$$

where $D_j$ is the incident's service deadline measured from occurrence, $t^{\text{rep}}_j$ is its first report time, $t$ is the current time, and $\hat d(z_j)$ is the expected reporting delay at its location. The delay estimate comes from the same duplicate-report model. The rate $\hat\psi$ changes over weeks and needs global pooling. It therefore sits in the territory layer, which is recomputed rarely and can be computed centrally. The delay $\hat d$ is a per-job scalar that can be looked up once when the report arrives. The routing layer therefore stays fully local, as in Pavone et al.

Two things are new here: the layer assignment, and its test against the natural alternatives. Neither the posterior model nor the queueing rule is new. We borrow both.

Research Program

We will build three components. The first is a correction module. It fits the duplicate-report reporting-rate and delay model on NYC street-tree and Chicago 311 data, and the Ising latent-field model on NYC flooding data. Both produce $\hat\psi$ and $\hat d$ maps at tract resolution. We will use the published estimators as they are described rather than inventing new ones.

The second is a dispatcher with switchable correction points. The territory layer offers equal-area Voronoi, TAMP-style spectral clustering on observed density, and each of these reweighted by $w(z)$. For weighted Voronoi we reuse Pavone et al.'s gradient law, applied to a non-uniform density. The routing layer offers TAMP's mixed priority queue with apparent remaining life and with $\hat r_j$, followed by a traveling-salesperson tour. We keep TAMP's reported best mixing weight of about 0.65 as the starting point and retune it for every condition.

The third is a ground-truth simulator, which is necessary because no real dataset records unreported incidents. Latent incidents are drawn from fitted spatial fields. Each incident is reported with the fitted neighborhood-specific rate and delay, then expires after a deadline. Only reports are shown to the dispatcher. To avoid circularity, where the correction model grades itself, we include deliberately misspecified generators. One uses clustered, socially contagious reporting that violates the Poisson assumption. Another uses reporting rates that depend on unobserved factors not among the census covariates. A third lets incident deadlines correlate with neighborhood.

The self-report literature shapes how we interpret the results. Reports diverge systematically from objective state, as documented for surveys by Explainable AI for Population Mental Well-being Surveillance Using Community Health Survey Data and for momentary mood reports by Assessing Real-Time Moderation for Developing Adaptive Mobile Health Interventions for Medical Interns: Micro-Randomized Trial. That second paper found that prompts sent on the basis of reported mood harmed users who were already doing well. The civic lesson we draw is that correction should use a model of the reporting process, not a heuristic inflation of the reports.

Baselines

The naive reference point is TAMP run on raw reports with apparent deadlines, together with equal-area Voronoi plus traveling-salesperson tours as in Pavone et al. TAMP reports completion rates up to 90% on synthetic tasks, and a 5.9% gain from weighting urgency at about 0.65 rather than 0.5. Pavone et al. keep losses below a 5% design target under uniform Poisson demand. Neither result was obtained under under-reporting. We found no reported result on Lacuna for any routing method on 311 data with under-reporting correction, so every baseline will be rerun in our simulator. Each will use the same fleet size, the same incident streams and random seeds, and the same tuning budget: a grid over the priority weight and the number of territory updates, chosen on validation streams.

The strongest existing competitor is predict-then-route. Following the Bayesian Ising paper, we take the tracts with the highest posterior incident probability as proactive targets and solve a traveling-salesperson tour over them together with the reported jobs. This competitor uses the best available latent model and differs from our proposal only in its structure. It spends a fixed inspection budget instead of reshaping territories, and it ignores delay. We will match the proactive stop budget to the extra capacity our reweighted territories imply.

The key ablations isolate layer placement: correction only in territories, correction only in routing, both, and a "swapped" version. The swapped version applies $\hat\psi$ as a priority multiplier inside routes and uses delay-inflated counts to size territories.

Beating the baselines means a higher fraction of true incidents serviced before expiry, the primary metric, with a smaller gap in that fraction between the highest- and lowest-reporting demographic quintiles. It must also come without more than a stated travel increase, which we set in advance at 10%. The advantage must hold on held-out streams under at least one misspecified generator.

Why not just…?

Why not just reweight reports by inverse reporting propensity and feed them to TAMP? This is the most serious threat, and we have partly absorbed it. At the territory level, simple inverse-propensity weighting of the kind described in the file may balance workload about as well as the full Ising posterior, because territories average over many tracts. We therefore treat the choice of estimator for $\hat\psi$ as an ablation, not a contribution. If the reweighted and Ising-based maps give indistinguishable territories, we will report that and use the simpler one. The file notes that the Ising model beats a Gaussian-process baseline by 0.14 AUC, but that gain concerns tract-level recovery, not aggregate workload. The part reweighting cannot supply is delay correction in routing. Scaling report counts does not change how old a given reported incident is.

Why not just select proactive inspection targets from the posterior, as the Bayesian paper already does? Top-K selection answers "where are the missed incidents" but not "who goes there, and in what order relative to reported jobs with deadlines." The predict-then-route baseline tests exactly this. If it matches our method on the primary metric at equal budget, the layered design adds nothing, and we will say so.

Why not just sort jobs by time since report, or use one conservative global deadline? Pavone et al. use a single conservative critical time and acknowledge that it wastes capacity when urgency varies. Sorting by report age is correct only if reporting delay is uniform, which the threefold Manhattan–Queens difference contradicts. The routing-only ablation measures how much of the gain delay correction alone delivers. If that ablation captures nearly everything, the territory correction is unnecessary, and that outcome would be informative too.

Proposed Experiments

The first experiment tests validity. Before any dispatch results, the correction module must reproduce the published diagnostics. That means delay estimates tracking storm-damage timing (the paper reports Pearson r = 0.88) and the reported race gaps in flooding reporting (24% and 18%). It must also reproduce the transfer result in which the first four hours of Hurricane Ida reports predict which tracts will accumulate reports over the next seven days. If these fail, downstream dispatch comparisons are uninterpretable.

The second experiment is the main comparison. We run the full baseline and ablation grid on simulated NYC flooding (Ida, Henri and Ophelia fields), NYC street-tree incidents and Chicago 311 infrastructure incidents. We vary fleet size, deadline spread and misspecification type, and report the primary metric, the quintile gap and travel.

The third experiment is the decisive check of the core claim. The "both layers" condition must beat each single-layer version and the swapped version, and the ordering must persist under the non-Poisson and unobserved-covariate generators. If swapped placement performs as well as correct placement, the timescale argument is wrong.

The fourth experiment tests whether the locality argument is real. We measure how often territories must be recomputed to keep the gains as $\hat\psi$ is re-estimated. If weekly recomputation suffices while routing stays purely local, the claimed resolution of the locality-versus-pooling tension holds. Finally, we run a scalability check on the Gowalla-scale task counts used by TAMP, to confirm that the correction adds only per-report lookups to the routing step.