Most studies that measure how much people from different income groups actually encounter each other in a city now use smartphone location traces instead of home addresses. They share one hidden choice: every encounter counts the same. Two people within 50 meters of each other for five minutes form one "exposure." That holds whether they passed on a sidewalk or sat in the same restaurant for two hours. We propose to measure segregation as a function of how long people are together. Each co-presence becomes an episode with a start time and a duration, and we estimate an income-segregation curve over duration. A naive version is easy and probably wrong. Phone location pings arrive sparsely and irregularly, so short episodes are systematically missed, and the miss rate itself differs across income groups. Our central contribution is therefore a duration-resolved segregation estimator that corrects for duration-dependent detection. We validate it by planting known curves in simulated and thinned real traces.
Why this matters: the headline finding of the field is that exposure segregation is 67% higher in the largest U.S. metros than in small towns. It comes from Nilforoshan et al., and it is computed at a single 5-minute scale. If large cities are well mixed in brief encounters but sorted in long ones, the policy conclusions change. More transit hubs would do little, and the long-stay venues are what would need to change. This question cannot even be posed with existing indices.
Core Claim
We claim that experienced income segregation is strongly duration-dependent, rising with encounter length. We further claim that current single-scale indices mix two regimes: near-integrated fleeting co-presence and highly sorted sustained co-presence. We also predict that the city-size gradient and the post-pandemic diversity loss sit mostly in the long-duration regime. Part of this rests on measurement: without correcting for how detection probability depends on duration, the shape of the curve is an artifact of ping density. We will show this directly.
The idea from the temporal-scale literature that we borrow is a representation, not a neural network. Li et al. map video moments onto a two-dimensional grid of start time against duration. They show that a single-scale view either washes out short events or misses the context of long ones. We index every co-presence episode the same way and compute segregation per cell of that grid. This makes the "dilution versus context-blindness" tension, described for fixed windows in video work such as Benmokhtar, an explicit object of measurement rather than a modelling nuisance.
Research Program
Episode extraction. From individual GPS traces we first detect stays, then form co-presence episodes. An episode is a maximal interval during which two users' stays overlap at the same venue (or within 50 m when no venue match exists). It records start time $s_e$ and duration $d_e$. This replaces the fixed 5-minute path-crossing of Nilforoshan et al. and the flat six-month dwell-time sums of Moro et al.. Following both papers, socioeconomic status is imputed from the home location's income quartile. We inherit their ecological error and treat it as a stated limitation rather than solving it.
Duration-resolved segregation. For a duration bin $b$, we define $\tau_{iq}(b)$ as the share of person $i$'s episode-time in bin $b$ spent with members of income quartile $q$. The person-level segregation at that duration follows the form used by Moro et al., which is 0 for a perfect 25% split and 1 for complete homogeneity:
$$S_i(b) = \tfrac{2}{3}\sum_{q=1}^{4}\left|\tau_{iq}(b) - \tfrac14\right|.$$
The curve $S(b)$, averaged over people, is the main object we estimate. The same quantity indexed by both start-time cell and duration bin gives the two-dimensional map.
Detection correction. An episode of true duration $d$ between users whose ping rates are $r_i$ and $r_j$ is observed with a probability $\pi(d, r_i, r_j)$ that rises steeply with $d$. Ping rate varies with device type, which plausibly correlates with income. We estimate $\pi$ empirically. We take the densest-sampled users, treat their episodes as ground truth, and thin their pings to match the empirical distribution of sampling rates. We then weight each observed episode by $1/\hat\pi$ when computing $\tau_{iq}(b)$:
$$\hat\tau_{iq}(b) = \frac{\sum_{e \in E_i(b),\, q(e)=q} d_e/\hat\pi_e}{\sum_{e \in E_i(b)} d_e/\hat\pi_e},$$
where $E_i(b)$ is the set of $i$'s observed episodes in bin $b$ and $q(e)$ is the partner's quartile. This is ordinary inverse-probability weighting. Its unbiasedness rests on the thinning experiment reproducing real sampling, which we test rather than assume. Nilforoshan et al. already had to correct attenuation bias across city sizes at one fixed scale. We extend the correction to a whole spectrum of scales.
Decomposition. A rising $S(b)$ could arise because long stays happen at more segregated venue types, which is a between-venue effect. It could also arise because, within a venue, people who stay together longer are more similar, which is a within-venue effect. Two examples of the latter are sections of a stadium and lunch-hour versus evening clienteles. We decompose the curve by recomputing it after reweighting each venue category's duration distribution to the city-wide one. The category segregation levels reported by Nilforoshan et al. supply a prior expectation for this decomposition: 0.16 at performing arts centers, 0.17 at stadiums, and 0.42 at golf courses and country clubs. The decomposition tells us whether those "integration hubs" stay integrated at long durations.
Data. We use individual-level GPS of the kind analysed by Moro et al. (4.5M users, 11 U.S. cities, Foursquare venues) and by Yabe et al.. The latter is the Spectus panel covering Boston, Dallas, Los Angeles and Seattle from 2019 to 2021, with ACS income quartiles. We will use Spectus as the primary panel because it allows the pandemic comparison. Aggregated foot-traffic data such as the SafeGraph data used by Yu et al. lacks pairwise co-presence and cannot support this analysis. GreenMove from Feng et al. discards stays under 10 minutes, so it serves only as an illustration of how threshold choice truncates the short end of the curve.
Baselines
The naive reference point is the single-scale exposure index of Nilforoshan et al.: co-location within 50 m inside a 5-minute window, every crossing weighted equally. We compute it on exactly the same panel, city set and income imputation as our estimator. It anchors comparisons with the published city-size result of 67% higher segregation in the ten largest metros and a Spearman correlation of 0.62 with population. We do not expect to reproduce those numbers exactly on a different panel. We will report how close we come as a check that our pipeline behaves sensibly.
The strongest existing measurement is the dwell-time-weighted experienced segregation of Moro et al.. It already lets long visits count more, and it reports that home-tract income predicts experienced segregation only weakly ($\rho \approx -0.17$). It is the strongest competitor because a reviewer can argue that duration weighting already captures what matters. We run it on the same episodes with the same stay detection. Beating it does not mean a higher score on some target, since no reported result on Lacuna exists for duration-resolved segregation. It means one of two things. Either the duration curve has a slope significantly different from zero after correction and decomposition, which a single weighted average cannot represent. Or the curve explains variation in the external validity outcomes below that the weighted average does not. The time-of-day binning of Wang et al., which found experienced isolation of 0.46 against residential isolation of 0.61, is a third baseline along the orthogonal axis of the two-dimensional map.
Two ablations isolate our key ingredients. The first is the uncorrected duration curve, a plain threshold sweep without inverse-probability weights. This isolates the contribution of the detection correction. The second is the curve without the venue-mix decomposition. This isolates whether the slope is merely a venue-type effect. All variants share one stay-detection setting. We additionally rerun everything under the 300 m / 10 min stay rule of Feng et al. and two looser rules.
Why not just…?
Why not just recompute the existing index at several co-presence thresholds? This is the real threat, and in spirit it is what we do. A threshold sweep is the uncorrected ablation above. If thinning experiments show that detection probability depends on duration but not on users' income or city, the sweep's shape would be roughly right and our correction would add little. In that case the contribution reduces to the empirical curve and the decomposition. We report this outcome honestly. We expect otherwise, because the 5-minute index already needed attenuation corrections across city sizes (Nilforoshan et al.), and those corrections must grow at short durations. The simulation experiment settles the question directly.
Why not just weight each visit by its dwell time? Linear weighting, as in Moro et al., assumes that the demographic mix of exposure is the same at every duration. A sixty-minute stay is then twelve five-minute stays of the same composition. Our curve tests that assumption directly. A flat corrected curve would vindicate weighting. A rising one shows that weighting averages two different regimes into one number.
Why not just bin by time of day or weekday versus weekend? Wang et al. did this and found that isolation varies through the day. Within a midday bin, however, a passing encounter and a long lunch are still pooled. Our two-dimensional map contains their binning as its marginal over duration, so the comparison is nested and exact.
Why not just train a multi-scale neural model, such as a boundary-sensitive detector or a hierarchical recurrent network? The boundary-proposal network of Lin et al. and the hierarchical LSTM of Ramaswamy et al. gain accuracy on labelled video. The gains are 37.46% average recall on THUMOS14, and 82.4% versus 63% frame-mAP when the hierarchy is removed on UCF101D. Mobility data has no labels for true encounters, and our question is measurement, not prediction. A learned stay detector could be a later refinement. We bound its potential effect through the stay-rule sensitivity analysis instead of building it.
Proposed Experiments
The first experiment is recovery under ground truth. We simulate city populations with a Social-EPR mobility model in the form of Yabe et al., extended so that stay durations and partner composition depend on duration. We plant three known segregation curves: flat, rising and U-shaped. We then sample pings at empirical rates, with ping density correlated with income. The decisive check is whether the corrected estimator recovers each planted curve within its bootstrap interval while the uncorrected sweep does not. We repeat this on thinned real traces of dense-ping users.
The second experiment is the empirical curve and its decomposition on the Spectus panel for 2019 across four cities, reported per city with bootstrap intervals over users. The hypothesis is falsified if the corrected, venue-mix-adjusted slope is indistinguishable from zero.
The third experiment asks where the city-size gradient sits. Across the available metros, we regress segregation on population separately in each duration bin. The prediction is that the gradient is weak for episodes under about 15 minutes and strong beyond an hour. Too few cities may limit power, and we will report that limitation if it arises.
The fourth experiment asks where the pandemic loss sits. Yabe et al. report that encounter diversity fell 15–30% and stayed about 10% below baseline into late 2021, with reduced social exploration explaining roughly 45% of the unexplained loss. We compute the duration curve for 2019 and late 2021 and test whether the persistent loss concentrates in long episodes. If it is instead spread evenly, then "less social exploration" operates at all durations, which is informative in itself.
The final experiment is external validity. Wang et al. link experienced segregation to Chetty et al. income-mobility data, though no reported numeric correlation was found on Lacuna. We test whether long-duration segregation tracks that outcome more closely than short-duration segregation or the single-scale index. This is the clearest evidence that resolving duration measures something that matters rather than a statistical curiosity.
References
- Human mobility networks reveal increased segregation in large cities
- Multi-scale 2D Representation Learning for weakly-supervised moment retrieval
- INRIA-WILLOW at TRECVID 2010 : Surveillance Event Detection
- Mobility patterns are associated with experienced income segregation in large US cities
- Behavioral changes during the COVID-19 pandemic decreased income diversity of urban encounters
- InclusiViz : Visual Analytics of Human Mobility Data for Understanding and Mitigating Urban Segregation
- Dataset for visitations of public green spaces in Shanghai, China
- Estimating experienced racial segregation in US cities using large-scale GPS data
- BSN: Boundary Sensitive Network for Temporal Action Proposal Generation
- Spatio-temporal action detection and localization using a hierarchical LSTM