Do Residents Live Where They Say? Testing the Amenity Shift Against Behavior and Across Regions

A common story in post-pandemic urban economics says that remote work has made transit-centric housing models obsolete. The older models priced homes by their distance to downtown jobs and train stations. In the new story, people instead value what they can walk to: cafes, restaurants, parks. Much of the evidence for this rests on residents rating their own neighborhood satisfaction in a single very dense city. Those ratings may reflect what people feel they should value rather than what they do. A Tokyo city-wide pattern may also say little about a smaller, car-oriented region. We propose to state the amenity-shift claim as short, human-readable rules, then check each rule twice. The first check asks whether it holds in passively recorded smartphone location data as well as in survey answers. The second asks whether it still holds when a model learned in Tokyo is applied to a lower-density Japanese region.

The method comes from clinical model interpretability. In that field, glass-box models (models whose decision logic can be read directly, such as shallow decision trees turned into IF–THEN rules with stated support and confidence) have been used to find people whose self-reports contradict objective risk markers. We bring that auditing habit to urban location choice and add one diagnostic borrowed from cross-city recommendation research. In a "swap test," a zone's amenity features are replaced with those of another zone that has the same overall density. If the prediction does not change, the supposed amenity rule is really a density rule in disguise. The output is a map of which parts of the amenity-shift story hold up under three conditions: across stated and revealed data, inside and outside the city that produced them, and after density is held fixed.

Core Claim

We expect the amenity-over-transit pattern to be real but narrower than the survey literature suggests. Specifically, it should hold in revealed behavior for remote-capable residents of non-central Tokyo. Stated satisfaction should overstate it for some identifiable groups: we expect central single-family residents and parents to be among them. Part of what looks like amenity preference should turn out to be density or popularity, and that part should not carry over to a lower-density region. The new element is not the rule learner, and not any single data source. It is a test protocol that requires each interpretable claim about the end of transit-centric housing to agree across measurement modes and across regions before it is accepted. Where it fails, the protocol reports exactly which rule failed and for whom.

What Prior Work Establishes and Leaves Open

The motivating evidence comes from The Relationship between Geographic Accessibility to Neighborhood Facilities, Remote Work, and Changes in Neighborhood Satisfaction after the Emergence of the COVID-19 Pandemic. That study fit ordinal logistic regression to 9,123 Tokyo survey responses. It found that remote work raised the odds of increased satisfaction more in non-central areas (odds ratio 1.50) than in central ones (1.31), and that access to eating facilities was a positive factor (odds ratios 1.01 to 1.06). Its outcome is entirely self-reported and retrospective, and it covers one metropolis. Measuring latent housing preference during the early stage of the COVID-19 pandemic reports latent demand for natural amenities far from transit. Both papers motivate the claim, but neither checks it against behavior.

The Death and Life of Great Italian Cities: A Mobile Phone Data Perspective shows that passive phone activity can replace surveys for measuring urban vitality. Third places and intersection density explain most of the variance in that work (R² = 0.77). However, its data predate the pandemic and it uses a single global regression. Uncovering Parental Urban Mobility and Amenity Visits through Large-scale GPS Data supplies our behavioral data source, Blogwatcher smartphone GPS for Tokyo (2022) and Miyagi (2019) matched to Zenrin points of interest. It also supplies a warning: parents in the same amenity-dense areas make fewer and narrower visits, so amenity counts are not amenity use.

The constraining evidence comes from two places. Assessing Spatial Heterogeneity in Urban Park Vitality for a Sustainable Built Environment: A Case Study of Changsha uses multiscale geographically weighted regression (a regression whose coefficients vary smoothly over space) and finds that transit facilities suppress park vitality in commuter districts but not elsewhere. A single global "transit is obsolete" coefficient is therefore unlikely to be meaningful. An Empirical Evaluation of Cross-City POI Recommendation on a Large-Scale Benchmark shows that supposedly "hometown-aware" models ignore user history. Swapping in a stranger's history leaves performance unchanged, and a plain popularity baseline (F1 0.1069) beats the best LLM agent (F1 0.0820). We reuse that swap logic as our density placebo.

The method comes from Explainable AI for Population Mental Well-being Surveillance Using Community Health Survey Data. That work extracts C4.5 rules with support and confidence from 108,252 Canadian survey records and uses them to flag people whose self-reported low stress contradicts long work hours and poor sleep. It has no spatial data and no cross-region test. We take its idea of rules that expose subjective–objective gaps and apply it to neighborhoods. We also borrow from A reluctant additive model framework for interpretable nonlinear individualized treatment rules. That paper allows nonlinear terms only when they clearly beat linear ones, and reports that ordinary cross-validation over-selects nonlinearity. We adopt its stricter complexity criterion when choosing rule depth.

Research Program

The unit of analysis is a residential zone. We use the finest areal unit at which the Tokyo survey records home location; ward or station-area is the likely resolution, and we will confirm it before analysis. Each zone receives the same feature set from both sources. The features are distance to the nearest rail station and to the main business district, counts of eating, drinking and park points of interest within walking distance (from Zenrin and park data), building type mix, and overall density.

Each zone also receives two outcomes. The stated outcome is the share of surveyed residents reporting increased satisfaction, split by remote-work status as in the Tokyo study. The revealed outcome comes from GPS and is the local amenity share. To compute it, we infer each device's home location from overnight stays and its work location from repeated weekday daytime stays. Devices that stay home on weekdays are flagged as remote-capable. Local amenity share is then the fraction of a device's non-home, non-work stay time spent at eating, drinking or park points of interest within walking distance of home. We compare this against stay time near rail hubs or in the business district.

We then fit shallow rule models (C4.5 trees converted to rule lists) separately to the stated and revealed outcomes. Depth is chosen by the reluctant criterion, not by cross-validation. Each rule carries a support set of zones and resident types, a direction, and a confidence. Rules are then audited in three ways.

The first audit is agreement. A stated rule is confirmed only if the revealed model, restricted to the same support set, shows an effect of the same sign. Rules that disagree are reported as "said-not-done" or "done-not-said."

The second audit is the density swap. For a fitted model $f$ and zone $z$, let $\tilde z$ be $z$ with its amenity features replaced by those of a randomly chosen zone in the same density decile. We report the swap sensitivity

$$\Delta = \mathbb{E}_z \, | f(z) - f(\tilde z) |,$$

relative to its value under a full random swap. If $\Delta$ is near zero, the model's apparent amenity effect is carried by density alone.

The third audit is transfer. Tokyo-fitted rules are applied unchanged to Miyagi. We measure the loss in predictive accuracy and the fraction of rules that keep their sign.

The transfer audit has a known weakness. In the data we have, Miyagi is observed in 2019, before the pandemic, and Tokyo in 2022. A Tokyo-to-Miyagi gap therefore mixes the effects of place and era. We handle this in two ways. First, we run a within-era geographic test inside Tokyo 2022, training on central zones and testing on non-central zones and the reverse. This mirrors the survey paper's own split. Second, we will request a post-2020 Miyagi slice from the same provider. If that slice is unavailable, every Miyagi result will be labeled as a joint place-and-era shift, not a pure geographic transfer. Using Miyagi 2019 as a pre-pandemic reference still has value: if a rule already holds there, it is not a remote-work effect.

Baselines

The naive reference is a density-only model. It predicts each outcome from the walkable count of eating, drinking and park points of interest plus total density, with no subgroup structure. This is the urban analogue of the popularity baseline that beat complex models on Trip World. Any rule model that does not improve on it in held-out accuracy and in swap sensitivity has learned nothing beyond "busy places are busy."

The strongest existing method on this question is the ordinal logistic regression of the Tokyo study. Its specification, with remote-work interactions and a central/non-central split, is the published evidence for the amenity shift. Its reported odds ratios are 1.50 for remote work in non-central areas, 1.31 in central areas, and 1.01–1.06 for eating-facility access. We refit exactly that specification on our zone features, once on the stated outcome and once on the revealed outcome. Both it and the rule models use identical zones, folds and feature sets, and each receives a matched tuning budget of about fifty configurations. The rule approach beats this baseline under two conditions. It must match the regression's held-out accuracy within noise, and it must identify at least one stated–revealed disagreement or density-carried effect that the regression's odds ratios conceal.

The spatial competitor is multiscale geographically weighted regression, as used in the Changsha study, fit to the same outcomes and data. No reported result was found on Lacuna for its fit metrics on the Changsha data, so we have no external number to match. We compare it against our models on our own data instead. The comparison covers held-out accuracy under the central/non-central split and the stability of the model's conclusions when bootstrap samples are drawn by zone.

For reference, the CCHS rule models report 95.76% accuracy for work stress and 82.1% for general stress. Those figures come from a different task and are not comparable targets here.

Two ablations isolate the key ingredients. The first trains rules on the stated outcome only and scores them against revealed behavior. This shows what a survey-only interpretable model would wrongly certify. The second removes the density-swap audit. This shows how many rules a standard pipeline would report that the swap test rejects. A third, smaller ablation replaces the reluctant complexity criterion with five-fold cross-validation and measures how many extra rules appear without any gain in accuracy.

Why not just…?

Why not just fit the ordinal logit to both data sources and compare coefficients? This is the most serious threat. If the amenity effect is roughly uniform within the central and non-central groups, comparing the two regressions' coefficients would answer the agreement question with less machinery. The parental GPS evidence suggests the effect is not uniform within areas: parents and non-parents in the same zones behave differently. The CCHS work also found its subjective–objective gaps in multi-attribute subgroups that a main-effects regression does not isolate. The baseline comparison settles this directly. If the rule models find no disagreement that the regression with remote-work interactions misses, we will report that the simpler tool suffices and drop the rule component.

Why not just use geographically weighted regression, which already handles spatial variation? It does handle smooth spatial variation, and it may beat our rules on accuracy within Tokyo. However, it gives a continuous surface of coefficients, not a small set of auditable claims. It also has no built-in check for whether a local amenity coefficient is really density. Its local coefficients are also fit to one region and cannot be applied unchanged to Miyagi, so it cannot answer the transfer question. We run it as a competitor and apply the same density-swap test to its predictions.

Why not just drop the surveys and trust the GPS? GPS shows what people do, not what they value, and it has its own biases. Location panels skew toward certain demographics, and the 125-meter location blurring in the Blogwatcher data adds noise to short walking trips. The research question is precisely whether stated and revealed evidence agree. GPS alone would replace one unchecked measurement with another. We report the share of stay time that falls within one blurring radius of a point of interest, so readers can judge how much the revealed outcome depends on that noise.

Why not ask an LLM to reason about which neighborhoods people prefer? On Trip World, LLM agents scored below a popularity baseline and recommended places that sound attractive but are rarely visited. That is exactly the gap between stated and revealed preference we are trying to measure, so an LLM-based approach would add the bias rather than detect it.

Decisive Checks

The project produces a clear answer whichever way the evidence falls.

The amenity-shift claim is supported if three things hold together. First, the rule "remote-capable residents of non-central zones spend more out-of-home time at local amenities" holds in GPS behavior with the same sign as the survey. Second, the effect survives the density swap, meaning $\Delta$ stays well above zero. Third, the rule keeps its sign in the within-Tokyo split.

The claim is undermined in either of two cases. One is that stated rules frequently lack revealed counterparts, especially among central single-family residents, where the survey odds ratio was 0.59 for remote workers. The other is that most of the amenity effect disappears under the density swap.

The geographic-scope question is answered by the fraction of Tokyo rules that keep their sign in Miyagi. That answer is credible only if the post-2020 Miyagi slice can be obtained. Without it, the Miyagi results are reported as a joint place-and-era shift.