When the Cell Tower Is the Address: Berkson-Corrected Hazard Models for Coarse Mobility Data

Researchers increasingly want to ask whether where people actually spend their time changes their risk of a health event: an asthma attack, a hospitalization, an infection. The richest population-scale record of where people go comes from mobile phone networks. These networks do not store a person's coordinates. They store which cell tower handled the phone, and one tower covers anything from a few hundred meters in a city center to several kilometers in the countryside. The usual practice is to give each person the average pollution, green space, or traffic level of the tower's coverage area. That value is then passed to a Cox proportional hazards model, the standard regression for time until an event.

This note argues that this practice has a specific, fixable flaw. Assigning the tower-area average creates what statisticians call Berkson error: the true exposure scatters around the assigned value, rather than the assigned value scattering around the truth. In a Cox model, Berkson error has a different effect from the textbook "noisy measurement" case. Because risk enters the model through an exponential, the correct risk for someone in a given tower area is the average of the exponentiated exposure over the places they might be inside that area. The exponential of the average exposure is not the same thing.

The gap between the two grows with how much exposure varies within the tower area. That variation tracks tower density, which tracks urban versus rural residence, which tracks age, income, and access to care. So the error does not just shrink the pollution effect. It can leak into the coefficients of other covariates and distort risk comparisons between urban and rural people. We propose a Cox likelihood that computes the within-cell average of exponentiated exposure directly, by Monte Carlo over plausible within-cell positions. We will test it on semi-synthetic cohorts built from real trajectories whose true exposures are known.

Core Claim

Write the true exposure at time $t$ as $x(L(t))$, the value of a known exposure map $x(\cdot)$ at a person's true location $L(t)$. The analyst sees only the serving cell $c(t)$. If the Cox model holds for the true exposure, the hazard given what we observe is

$$h(t \mid c) = h_0(t)\, \mathbb{E}\big[\exp(\beta\, x(L(t))) \mid c(t), \text{history}\big],$$

where $h_0(t)$ is the unspecified baseline hazard and $\beta$ the log hazard ratio per unit exposure. Suppose the within-cell exposure is roughly Gaussian with mean $\mu_c$ and variance $\sigma_c^2$. Then the expectation is approximately $\exp(\beta\mu_c + \beta^2\sigma_c^2/2)$; this form comes from the Gaussian moment generating function. The standard "cell-mean" analysis keeps only $\exp(\beta\mu_c)$ and silently drops a term that varies from cell to cell.

Three consequences follow. First, the naive analysis is misspecified exactly where cells are large and heterogeneous. Second, the needed correction can often be computed rather than estimated. The exposure map inside each coverage polygon is known, so the only unknown is where the person is inside it. Third, the correction must model time jointly. A person who stays still keeps the same unknown position across many records, so errors are correlated within a dwell and do not average away over cumulative exposure.

Relation to Prior Work

The measurement-error side supplies the tool but points the wrong way. An Exact Corrected Log‑Likelihood Function for Cox's Proportional Hazards Model under Measurement Error and Some Extensions builds an exact corrected Breslow likelihood for classical additive error $W = X + U$, using the error's moment generating function. We reuse its central insight: an exponential-family correction through a moment generating function. Our error structure is the reverse, Berkson rather than classical, so applying its correction as-is would overcorrect. In our setting the relevant moment generating function comes from the within-polygon exposure distribution, not from a validation study.

Modeling the probability distribution of positional errors incurred by residential address geocoding. shows that spatial position errors are heavy-tailed and aligned with street grids, and it fits mixtures of bivariate $t$-distributions. It stops before any outcome model. We use it as a constraint: it rules out Gaussian within-cell position models. We also use it as a source of realistic within-cell position distributions for the residential variant of our experiments.

Three papers motivate the proposal by showing the gap in practice. Preliminary investigation of the association between air pollution exposure and childhood asthma hospitalizations from 2015 to 2018 in East China fits time-dependent Cox models on land-use-regression exposures built from nine monitors and static home locations. It reports a PM2.5 hazard ratio of 1.01 (p = 0.599). We cannot say whether that null is real or produced by spatial error. That uncertainty is precisely the question our method is built to answer. Using Geographic Location-based Public Health Features in Survival Analysis takes the opposite route and aggregates to U.S. states. It gains 0.01 to 0.02 in concordance but gives up all within-state variation and treats state membership as error-free. Mobile Network Data for Public Health: Opportunities and Challenges describes Voronoi approximations of tower coverage (each tower's area taken as the region closer to it than to any other tower). It also notes the scale mismatch between kilometer-wide cells and local health processes, but its epidemiological models are aggregate, not individual time-to-event models.

On the mobility side, Outdoor Position Recovery From Heterogeneous Telco Cellular Data (PRNet+) recovers positions from cellular measurement records to a 15.2 m median error in its Jiading-1 2G setting, using neural sequence models and a speed-consistency penalty tied to inferred travel mode. It outputs point estimates. We treat it as a competitor, and optionally as a source of within-cell position distributions. Inferring modes of transportation using mobile phone data documents the "handover blind spot": walking trips too short to switch towers go unrecorded, so pedestrian share is inferred at 2.37% against about 10% in surveys. This constrains our design. Within-cell movement is real and must enter the position model rather than be assumed away. Finally, The impact of social segregation on human mobility in developing and industrialized regions notes cell oscillation, where a stationary phone near a boundary flips between towers. Our dwell-persistent position model absorbs this naturally: an oscillating sequence is most plausibly one fixed position near the shared boundary.

For simulation, Generating Survival Times Using Cox Proportional Hazards Models with Cyclic and Piecewise Time-Varying Covariates gives closed-form ways to generate event times under cyclic time-varying covariates with Weibull or Gompertz baselines. We use it to simulate outcomes from daily commuting exposure. DeepSurv: personalized treatment recommender system using a Cox proportional hazards deep neural network replaces the linear risk score with a neural network. Our correction applies inside such models too, but we keep the linear score so that bias in $\beta$ can be measured.

Research Program

The method has three parts. The first is a within-cell position model: a distribution over a person's true location given the sequence of serving cells. We will compare three versions of increasing strength. The simplest is uniform over the Voronoi polygon. The second is weighted by population or building density. The third is a probabilistic localizer trained on the Shanghai measurement records with GPS ground truth. Positions persist within a dwell, move under a speed limit between dwells, and are resolved jointly across an oscillation episode.

The second part is a Monte Carlo corrected likelihood. For each person-interval we draw $M$ position paths and compute exposures from the map. The risk term in the Breslow likelihood becomes a log-mean-exp over draws. We then maximize, with risk-set subsampling so that cohorts of hundreds of thousands stay tractable.

The third part is analysis. We will derive the bias of the cell-mean estimator as a function of $\beta$, the spread of $\sigma_c$ across cells, and the correlation between $\sigma_c$ and other covariates. We will also check the Monte Carlo bias of log-mean-exp. It shrinks as $1/M$ by a delta-method argument, and we will verify that it is negligible at practical $M$.

Baselines

The oracle fits the same time-varying Cox model on exposures computed from the true GPS positions. It defines the achievable target for bias and interval coverage.

The do-nothing reference assigns the exposure at the tower location or polygon centroid, which is what most pipelines effectively do.

The strongest existing competitor is regression calibration: replace exposure with the cell-mean exposure $\mu_c$ computed over the polygon. Under Berkson error this is the textbook fix, and it is unbiased in linear models. It must be beaten on bias of $\beta$ and of the urban/rural covariate coefficients, and on 95% interval coverage, in regimes where $\beta\sigma_c$ is not small.

The second-order ablation adds $\sigma_c^2/2$ times $\beta^2$ as a plug-in adjustment. This is the Gaussian approximation to our correction and isolates what the full Monte Carlo treatment buys under heavy tails and skew.

The two-stage pipeline runs PRNet+ localization and plugs its point estimates into Cox. It represents the strongest localization-first approach, and its reported localization figure is 15.2 m median error (Jiading-1 2G, PRNet+).

We also include SIMEX, which estimates bias by adding extra noise and extrapolating back, run with a Berkson-appropriate simulation. Finally, we include coarse district aggregation with a shared frailty term (a random per-district risk multiplier), in the spirit of the state-level feature of Using Geographic Location-based Public Health Features in Survival Analysis.

All methods receive identical cell sequences, exposure maps, simulated outcomes, and baseline hazard handling. Tuned components share the same budget: bandwidths, the SIMEX grid, and PRNet+ training on the same Shanghai split.

For the question we study, no reported result was found on Lacuna for bias or coverage of Cox coefficients under cell-tower coarsening. That comparison is what this work would supply. Beating a baseline means lower absolute bias and nominal coverage in the high-heterogeneity regime, without worse performance in the low-heterogeneity regime.

Why not just…?

Why not just use regression calibration with the cell-mean exposure? This is the real threat, and the proposal is designed to settle it rather than assume it fails. When $\beta\sigma_c$ is small, the dropped term $\beta^2\sigma_c^2/2$ is tiny, and regression calibration should match our method. The experiments sweep effect size and tower density to locate the boundary. If the boundary lies outside realistic exposure maps and tower layouts, we will report that the simple fix suffices. That would itself be a useful negative result for mobility epidemiology.

Why not just add within-cell exposure variance as an extra covariate? This is the second-order ablation above, and it may recover most of the gain. It fails when within-cell exposure is skewed or heavy-tailed. Position errors are heavy-tailed and grid-aligned (geocoding error study), and pollution surfaces near roads are sharply peaked. It also ignores dwell correlation in cumulative exposure. The ablation measures exactly how much these matter.

Why not just localize first with PRNet+ and treat the output as truth? Point estimates carry no uncertainty forward, so intervals will be too narrow. PRNet+ also needs measurement records with signal features, which many call-record datasets lack. We test the pipeline directly on bias and coverage.

Why not apply the existing exact corrected Cox likelihood? It corrects classical error (Exact corrected log-likelihood). Applied to Berkson error, it pushes estimates in the wrong direction. We include it to show the overcorrection empirically.

Proposed Experiments

The primary testbed is semi-synthetic, because no public cohort links cellular traces to health outcomes. We take real GPS trajectories from GeoLife and the GPS-matched Shanghai measurement records used by PRNet+. We coarsen them into Voronoi cells under synthetic tower layouts, ranging from dense urban to sparse rural, and inject oscillation and handover-threshold dropout at rates matching the Santiago walking undercount. We then overlay exposure maps that are smooth, road-peaked, and patchy.

Event times are simulated with the cyclic time-varying generator under Weibull and Gompertz baselines, at known $\beta$. We add a rurality-correlated covariate so that bias leakage into other coefficients can be measured. Primary metrics are bias and 95% coverage for $\beta$ and for the secondary coefficient. Secondary metrics are hazard calibration by urban/rural stratum and concordance.

A residential variant replaces tower cells with geocoding error drawn from bivariate $t$ mixtures, calibrated to the 168 m versus 44 m median errors reported by the geocoding study. This mirrors the static-address design of the Nanjing asthma study.

Two outcomes would decide the question. The idea is supported if the corrected likelihood keeps coverage near 95% while regression calibration drops well below it in rural, road-peaked settings, and if the secondary coefficient's bias under the naive analysis grows with the rurality correlation. The idea is refuted if regression calibration stays within Monte Carlo error of the oracle across realistic settings.