Onset Timing Separates Spatial Spillover from Spatial Confounding

When a retrospective study finds that places near a treated place also changed, two stories compete. The first is spillover: the treatment really leaked across the border. The second is spatial confounding: some unrecorded, geographically smooth factor, such as income, terrain or local policy culture, made both the treatment and the outcome more likely in that whole region. Standard practice cannot tell these apart, because it first sorts events into common time bins (months, years) and then reads off cross-sectional correlations. This proposal argues that the timing that binning throws away is the most useful evidence available. In real registries, treatments start at scattered, unsynchronized moments across places. A genuine spillover from place $j$ to place $i$ has to show up after $j$'s onset, with a lag that grows with distance. A slowly varying hidden confounder has no reason to produce a response locked to the exact onset time of a particular neighbor.

The idea comes from wireless communications. There, receivers decoding many uncoordinated devices learned to stop treating arrival-time offsets as noise. They model each offset as a hidden variable to be estimated (Bian et al.), and in some cases use it as a fingerprint that pulls overlapping signals apart (FTrack). We transfer that logic to observational spatial causal inference. The result is an estimator that recovers spillover strength and propagation delay from unbinned onset times, together with a permutation diagnostic. The diagnostic shuffles onset times while keeping the spatial pattern of who was treated fixed, which tells us how much of an apparent spillover could be confounding.

Core Claim

Asynchronous treatment onsets are an identifying signal for spillover. They are not a nuisance to be averaged away. Suppose unit $i$ has outcome series $y_i(t)$ and treatment onset time $s_i$, and its spatial neighbors $j$ sit at distance $d_{ij}$. The working model is

$$ y_i(t) = \alpha_i + g(U_i, t) + \beta\, k(t - s_i) + \sum_{j} \theta(d_{ij})\, k\big(t - s_j - \tau_{ij}\big) + \varepsilon_i(t). $$

The terms are as follows. $\alpha_i$ is a unit-specific level. $U_i$ is an unmeasured, spatially smooth confounder that shifts both onset times and the outcome trajectory $g$. $k$ is a response kernel, meaning the shape of the outcome's reaction after an onset. $\beta$ is the direct effect. $\theta(d)$ is the spillover strength at distance $d$. $\tau_{ij}$ is an unknown propagation delay, modeled as a distance-dependent mean plus random variation in arrival time. $\varepsilon_i(t)$ is noise.

The identifying assumption is a separation in time scale. The confounder can be correlated with when neighbors adopt, but its effect on $i$'s outcome, $g(U_i,t)$, must be smoother in time than the kernel. Equivalently, it must not be locked to the delayed neighbor onsets $s_j + \tau_{ij}$.

Under this assumption, the onset-locked part of the co-movement between neighbors identifies $\theta$ and $\tau$. Binning onsets into windows wider than the delays destroys exactly this part. The assumption is substantive and can fail. A shared shock, such as one storm, could trigger both nearby onsets and outcome jumps. Much of the evaluation below is built around measuring what happens when it fails.

Research Program

The work has three parts.

The first part is the estimator. For each unit, we build a dictionary of candidate regressors. Each regressor is the kernel triggered by one neighbor's onset and shifted by one candidate delay. Estimating spillover then becomes a sparse recovery problem with shifted, correlated columns. This is structurally the same problem as the shifted pilot matrices in asynchronous random access, where delays destroy the independence assumptions that ordinary sparse solvers rely on (Bian et al.). We borrow their remedy. Delays are treated as latent variables shared across a group, here a distance band rather than a set of antennas. Delays and amplitudes are fit jointly, with each delay constrained to grow with distance. We start with group-sparse regression over the delay grid, and move to a message-passing solver only if the group-sparse version proves too slow at SpaCE scale.

The smooth confounder term $g$ is absorbed in one of two ways. The simple option is a low-frequency temporal basis for each unit. The full option uses a substitute confounder in the style of the Spatial Deconfounder, which reconstructs hidden spatial factors with a conditional variational autoencoder (a neural model that learns a compressed hidden representation) by reading neighboring treatments as several noisy views of the hidden field. Our change is that the autoencoder sees onset times as continuous inputs rather than binned treatment indicators.

The second part is the diagnostic. We hold the spatial set of treated units fixed and permute onset times within them. Confounding that works through who is treated and through smooth trends survives the shuffle. Onset-locked spillover does not. The gap between the estimated $\theta$ and its permutation distribution therefore measures the spillover we can attribute to timing. This builds on the Fisher-style randomization inference that Aronow et al. use for randomized spatial experiments. Their setting has randomized assignment, and they leave observational data as future work. We swap the assignment randomization for a timing permutation, which is valid only under the separation assumption.

The third part is a benchmark. SpaCE supplies 124 semi-synthetic spatial datasets with masked confounders and ground-truth counterfactuals, but they are static snapshots. We add time. Onset times are generated as a function of the masked confounder plus arrival-time jitter, and outcomes are generated as time series under the model above, with distance-dependent delays and known $\theta$. Three knobs are varied: delay spread relative to bin width, the strength of the link between confounder and onset time, and a "shared shock" setting that deliberately breaks the separation assumption.

Each direction plays a different role. Asynchrony supplies the mechanism. Spatial spillover is the estimand. Retrospective bias, namely unmeasured smooth confounding and irregular logging, defines the regime the method must survive.

Baselines

The naive reference is binning plus a Spatial Durbin Model, a standard spatial regression that includes neighbors' outcomes and neighbors' covariates as predictors. We fit it with unit fixed effects at the bin width a practitioner would choose, and also at the finest feasible width. It shares our data and gets the same covariates. Beating it means lower spillover RMSE whenever the delay spread is comparable to the bin width, and no worse performance when delays are much shorter than a bin.

The strongest existing method for confounded spatiotemporal spillover is GST-UNet. It combines a U-Net (a convolutional network that processes a grid at several scales) with ConvLSTM recurrent layers and iterative G-computation to handle time-varying confounding and interference on a regular time lattice. Its paper reports reducing RMSE by 21% to 38% relative to the best non-recursive baseline in synthetic simulations with feedback confounding. We run GST-UNet on our time-extended SpaCE data rasterized at several bin widths, give it the same hyperparameter search budget and GPU hours as our method, and select by the same held-out geographic splits that SpaCE recommends. Beating it means lower spillover-effect RMSE and correct delay recovery at matched compute, specifically in the regime where delays vary across pairs. We expect parity when delays are fixed and known.

The strongest static deconfounder is the Spatial Deconfounder, run on binned treatments. Its paper reports posterior predictive check p-values near 0.5 for confounder reconstruction on interference-augmented SpaCE. That figure validates the reconstruction, not effect error, and I found no reported spillover-effect error on Lacuna for it. Among spatial regressions, SpaCE reports that a spatial-spline method was the top performer for binary treatments and a spatial error model for continuous ones, but gives no numbers I can quote. Both are included.

Two ablations isolate the key ingredient. The first keeps our full estimator but replaces true onset times with bin midpoints, which tests whether the gain comes from timing at all. The second runs the estimator without the substitute confounder, which tests whether the timing signal alone handles confounding.

Why not just…?

Why not just bin finely and run GST-UNet or a fine-grained Durbin model? This is the most serious threat. If the bins are much narrower than the delays and the data are large, a flexible discrete-time model can in principle learn any lag structure. Our estimator differs in two ways. It pools across pairs by forcing delays to depend on distance. It also includes a timing permutation that separates confounding from spillover without depending on the architecture. The bin-width sweep in our experiments answers the question directly. If GST-UNet at the finest bins matches our RMSE and delay recovery on the time-extended SpaCE data at equal compute, the estimator is redundant. In that case the contribution shrinks to the diagnostic and the benchmark, and we will report it that way.

Why not just run a staggered-adoption event study with neighbor-exposure indicators? Event-study difference-in-differences already uses onset timing, and it is the natural econometric competitor. It does not appear in the research file, so we have no reported numbers for it. It works when every pair shares one known lag. Its usual safeguard against confounding, parallel trends, is exactly what breaks when a smooth hidden factor drives both adoption time and trend. We implement it with distance-band exposure indicators. If it matches us when delays vary across pairs and confounding is strong, the delay-estimation machinery is unnecessary.

Why not just apply the Spatial Deconfounder and ignore timing? It handles the static confounding well. But on binned data it cannot tell a confounded neighbor correlation from a lagged causal one, because the bins erase their ordering. The first ablation and the timing-permutation test measure how much of its estimated spillover vanishes once onset order is scrambled.

Proposed Experiments and Decisive Checks

On the time-extended SpaCE data, the primary metrics are RMSE of direct and spillover effects, error in recovered delays, interval coverage, and the type I error and power of the permutation test. The decisive synthetic check has two halves. With spillover set to zero and confounding strongly tied to onset times, the test should reject at its nominal rate, while binned methods report spurious spillover. When the separation assumption is broken by shared shocks, we report how fast our bias grows, so readers can see where the method stops working.

Two real retrospective datasets follow. The first is hourly PM2.5 across 10 California counties during the 2019 Kincade Fire, where Bayesian spillover graphs found Sonoma accounted for over 40% of system-wide spillover. Our recovered delays should increase with distance from Sonoma, and the Sonoma-origin spillover should survive the timing permutation. The second is daily county respiratory hospitalizations during the 2018 Camp Fire. GST-UNet estimated about 4,650 excess hospitalizations over the 10-day peak. Treating each county's smoke-arrival date as its onset, we test whether accounting for asynchronous arrival changes that total, and whether county-level effects are onset-locked.

The Uganda forest experiment is static, so it is not used for timing. Uganda and the random-access simulation of Bian et al. serve only as sanity checks on known ground truth: the first for the distance-band estimand, the second for the shifted-dictionary solver.

The whole project hinges on one outcome. If timing carries no separating signal beyond what fine binning captures, the gap between our method and fine-binned GST-UNet will close, and that answer is worth reporting too.