Zoning-Proof Hotspot Maps from Boundary-Adaptive Density Bands

When analysts map where events happen (taxi pickups, thefts, pollution exposures), they usually count events inside human-drawn zones such as census tracts or grid cells and then rank the zones. The conclusions depend on where the lines were drawn, a long-known difficulty called the modifiable areal unit problem. Mitigating allocative tradeoffs and harms in an environmental justice data tool shows what this costs in practice. In California's CalEnviroScreen, reasonable changes to aggregation and indicators alter the "disadvantaged" status of 16.1% of census tracts, and a tract at the funding cutoff can move by up to 44 percentile ranks. We propose to stop treating the zones as the object of inference. Instead, we estimate the continuous event density on the true, irregular study region (coastlines, rivers, parks) and put a simultaneous confidence band around it: an envelope that contains the whole true surface at once with probability 95%. A band over the surface automatically bounds the average density of every zone anyone could draw, including zones chosen after seeing the data. That band therefore certifies which areas are above or below a policy threshold under all zonings simultaneously, and it honestly marks the rest as ambiguous.

Two obstacles have kept this from being routine. First, ordinary smoothing leaks probability mass across hard geographic borders and biases the surface exactly where many zones sit. Second, the density-estimation tools that do fix boundary bias with valid uniform bands, lpdensity and lpcde, work on intervals, not on non-convex planar regions. The contribution is to carry their construction into two dimensions on arbitrary polygons and to turn the resulting band into a zoning-robust decision certificate.

Core Claim

A boundary-adaptive local polynomial density estimator can be built on irregular planar domains by fitting a local polynomial to the two-dimensional empirical distribution function inside a window that respects the region's connectivity. With robust bias-corrected centering and a multiplier-bootstrap critical value, this estimator gives a simultaneous band with near-nominal coverage right up to curved and non-convex borders. Let $f$ be the true density on region $D$, $\hat f$ the bias-corrected estimate, $\hat\sigma(s)$ its standard error at location $s$, and $c_\alpha$ the bootstrap quantile of the supremum of the standardized error over $D$. Then with probability at least $1-\alpha$, simultaneously for every measurable zone $A \subseteq D$,

$$\left|\frac{1}{|A|}\int_A \hat f(s)\,ds-\frac{1}{|A|}\int_A f(s)\,ds\right| \le \frac{c_\alpha}{|A|}\int_A \hat\sigma(s)\,ds .$$

This follows directly from integrating the pointwise-in-$s$ envelope over $A$. It holds for all zonings at once, at no extra multiplicity cost. A zone is certified above a threshold $\tau$ if its lower bound exceeds $\tau$, certified below if its upper bound falls short, and ambiguous otherwise. The testable claim is that certified labels flip across zoning schemes at a rate no higher than $\alpha$, while raw zone rankings flip often. We also claim that boundary adaptivity is what keeps this guarantee honest in the border zones where competing smoothers fail.

The idea has a known weakness, and the method is designed around it rather than hidden. For large zones, simply counting events inside the zone is unbiased and has small variance. Integrating a surface band is then conservative compared with inference on the counts directly. We expect a crossover in zone size relative to the bandwidth, and the program measures it.

Research Program

The estimator generalizes the construction of Simple Local Polynomial Density Estimators and Local Regression Distribution Estimators, which fit a polynomial to the step-function empirical distribution and read the density off its derivative. The fit uses only data that exist near the evaluation point, so it extrapolates to an edge instead of being pulled toward zero. In the plane we use the orthant distribution function $F(s)=P(S_1\le s_1, S_2\le s_2)$. Its mixed second derivative is the density, and it stays smooth up to a smooth boundary of $D$ because data exist only inside $D$. A quadratic-in-each-coordinate local fit then inherits the automatic boundary correction.

The genuinely new difficulty is non-convexity. A Euclidean window centred near a bay or river can straddle two shores. Across the water the distribution function is piecewise smooth rather than one polynomial, so a joint fit is biased. We therefore restrict each local fit to points within geodesic distance $h$ inside $D$, meaning distance measured along paths that stay in the region. This keeps the single-bandwidth tuning that makes the local polynomial approach attractive. The mean-squared-error bandwidth selector is re-derived with the window's effective area.

Inference follows the robust bias-corrected recipe of the two packages: centre at a higher-order fit, estimate the variance, and simulate the supremum of the Gaussian approximation. Bootstrap practice for uniform bands comes from Haoze Hou et al. for the multiplier version and from the wild-bootstrap functional work of Linglong Kong et al.. The multidimensional band results of Statistical Inference for Functional Data over Multi Dimensional Domain supply the multidimensional-domain precedent, but they assume regular domains without border correction.

Real event data are not independent: taxi pickups cluster by day and hour. Revisiting the Modifiable Areal Unit Problem in Deep Traffic Prediction with Visual Analytics shows that dependence structure shifts with aggregation. For that reason, multipliers are drawn per time block (per day), not per point. Because the band lives on the continuous surface, its critical value does not depend on any zoning. This is the structural reason we expect it to avoid the coverage collapse that zone-level bands would suffer when re-partitioned.

Our position relative to the MAUP literature is as follows. SPAD, in Fairness by "Where", treats partitions as a distribution to sample during training. It is limited to regular grids of up to 5×5 and reports scalar disparity metrics, not confidence statements. A Unified Model for Spatio-Temporal Prediction Queries with Arbitrary Modifiable Areal Units answers arbitrary region queries with deterministic predictions assembled from quad-tree blocks, including subtraction, which can yield negative values in sparse areas. Legally-Compliant Spatial Fairness Framework runs permutation tests on pre-aggregated region pairs. All three take zones as primitive. We take the continuous surface as primitive and derive every zone's bound from it.

Baselines

The naive reference is the common practice itself: count events per zone, rank the zones, and apply the threshold, with or without pointwise 95% intervals at each zone or grid location. It is run on exactly the same point data and zonings. It sets the volatility level we must reduce, and its pointwise intervals show the multiplicity failure that a simultaneous band exists to fix.

The strongest smoothing competitor is two-dimensional kernel density estimation with the standard edge correction that divides by the kernel mass falling inside $D$, given a sup-norm band from the same block multiplier bootstrap. This removes the level bias at borders but not the slope bias, which the local quadratic fit removes. It is a real threat in regions with gentle density gradients. It gets the same bandwidth-selection budget: a cross-validated bandwidth, plus an undersmoothed variant so that its bands are not unfairly biased. Beating it means holding boundary-strip coverage within two points of 95% where the corrected KDE falls measurably short, at comparable band width. For this 2D irregular-domain setting we found no reported result on Lacuna. The closest reported evidence is one-dimensional: lpdensity reports coverage "near the nominal 95% target at the boundary (x=0)" over 2,000 Monte Carlo repetitions, against KernSmooth, ks, and np.

The strongest non-smoothing competitor is direct simultaneous inference on zone counts. It takes each candidate zone's empirical share, standardizes it, and uses a max-type bootstrap critical value over the finite union of all zonings under study. It is unbiased and, for coarse zones, likely tighter than our integrated band. Beating it means narrower or equally valid certificates for zones below the crossover size. It also means validity for zones not in its pre-registered family. It does not beat us only if the crossover lies at sizes smaller than most policy zones, and in that case we report so.

The heuristic competitor is the ensembling of Huynh et al., which takes the union of designations across two or three plausible model versions. They report that this reduced designation sensitivity by 40.7% and 71.0% respectively on CalEnviroScreen. We apply the same union rule over two and three of our zonings. Beating it means a comparable or larger drop in flip rate together with an explicit error guarantee, with the price reported as the fraction of zones labelled ambiguous.

Two ablations isolate the key ingredients. The first uses the same local polynomial estimator with Euclidean rather than geodesic windows, which isolates the non-convex correction. The second uses the local polynomial fit with ordinary centering at the mean-squared-error bandwidth instead of bias-corrected centering, which isolates the bias-correction step.

Why not just…?

Why not just use edge-corrected kernel density estimation? It is the most serious alternative. It needs one line of code, and near straight borders with flat density it may perform about as well. Its correction only rescales mass, so where the density slopes toward a border (a waterfront pickup corridor, a dense street that ends at a park) it keeps a first-order bias. That bias does not shrink with the bandwidth the way the local polynomial's does, and a sup-norm band centred on a biased estimate under-covers there. Boundary-strip coverage in the simulation and in the held-out taxi check settles this directly. If corrected KDE matches our coverage, the planar local polynomial machinery is not earning its complexity, and we will say so.

Why not just do simultaneous inference on zone counts? For large zones this is the better tool, and we expect it to win there. It fails in two cases. The first is small zones, where counts are too sparse for tight bounds. The second is any zone outside the family fixed in advance, which is precisely the gerrymandering-style re-drawing at the heart of the zoning effect. Our band covers arbitrary zones by construction. The crossover experiment reports where each approach wins. The recommended practical output splits $\alpha$ between the two certificates and intersects them.

Why not just ensemble several zonings, as done for CalEnviroScreen? Taking unions measurably damps volatility, as the 40.7% and 71.0% figures show. However, it only enlarges the designated set, has no error rate, and gives no rule for how many or which zonings to include. We run it head-to-head on identical zonings. If its flip-rate reduction matches ours at a lower ambiguous fraction, our method's only remaining advantage is the formal guarantee, and we will report that plainly.

Why not just use SPAD or One4All-ST? Both produce predictions, not confidence statements, and both live on rectangular grids. They answer a different question: fitting a model that behaves well across partitions, rather than stating which areas are provably high. We therefore do not rank them on coverage. We cite One4All-ST's reported 10.6% RMSE gain on Taxi NYC only as context for the dataset.

Proposed Experiments

Simulations come first. We place known densities on three domains: a Manhattan-shaped polygon, a region with a narrow inlet, and a star-shaped domain with sharp corners. Density gradients either run parallel or perpendicular to the borders, and sample sizes range from 2,000 to 200,000 points, with and without day-level clustering. We measure whole-domain and boundary-strip coverage of the 95% band, average band width, and the flip rate of certified threshold labels across 50 zonings. These zonings mix grids of five resolutions with random offsets and random Voronoi partitions. Critical values use 2,000 bootstrap draws, matching the lpdensity default.

The real-data study uses Manhattan pickups from the Taxi NYC data used by One4All-ST, 36 million records on a domain bounded by rivers and interrupted by Central Park. It also connects to the zoning sensitivity documented for Manhattan mobility in Discovering and Characterizing Mobility Patterns. Because the data are so large, we can check coverage empirically. We build bands from a random sample of one or several weeks, then test whether they contain a high-precision estimate computed from the remaining months, both over the whole surface and in a 200-metre riverfront strip. We then impose a CalEnviroScreen-style top-quartile rule on zones drawn from the official taxi zones, grids, and Voronoi partitions. For each method we report the fraction of zones whose designation changes across zonings, the analogue of the 16.1% figure, alongside the ambiguous fraction.

Three results would decide the proposal. The band must hold boundary-strip coverage within two points of nominal, both in simulation and against the held-out taxi surface. Certified designations must flip across zonings at no more than 5%, while raw rankings flip far more. The geodesic ablation must lose coverage in the inlet domain and along the East River, while the full method does not. Failing the first check would mean the boundary machinery does not transfer to the plane. Failing the third would mean the non-convex correction is unnecessary and the method reduces to a simpler planar version of lpdensity.