Terminal Anchors: Borrowing Backward Decoding to Identify Multi-Year Spatial Leakage

When a conservation or pollution policy works locally, part of the damage it prevents often reappears next door, and it may take several years to arrive. In Mato Grosso, soybean farming expands onto old cattle pasture, and the displaced cattle then clear forest further out on the frontier. Estimates of how much deforestation a policy really avoids therefore depend on how this delayed displacement, or "leakage", spreads across neighbouring municipalities over time. Most spatial panel models estimate the spillover as happening within the same year, or add lags that run forward in time. Neither approach says when a panel actually contains enough information to separate this year's leakage from leakage left over from earlier years. This note proposes to answer that question with a structural lesson from communications theory.

Backward decoding is a strategy for relay and interference networks in which each transmission block carries a new message on top of help relayed about the previous block. Because the help for block $b$ arrives in block $b+1$, the receiver stores the whole sequence and decodes from the last block back to the first. This works only because the final block carries no new message and can be decoded on its own. Our central claim is that lagged spatial spillovers have the same layered structure. A year's outcome in a municipality combines its own new pressure with pressure displaced from neighbours in earlier years. The lag profile of leakage becomes identifiable when the panel contains anchor periods in which the direct driver has stopped changing while displacement is still arriving. In such periods only "relayed" signal is present. We will turn this into a checkable identification condition, an estimator diagnostic, and a rule for staggering policies so that their leakage can be measured. We will test all three on the Mato Grosso deforestation panel and on Chinese air-pollution panels.

Core Claim

The obvious combination of these two fields would be a "backward decoder for spillover trajectories": an algorithm that runs from the last year to the first and fills in intermediate displacement using terminal observations. We are not proposing that. A standard state-space smoother already does it, as discussed below. The research file finds no published work that joins these two fields at all. The closest cousins are the spatiotemporal forecasting models it surveys, and those run strictly forward.

What backward decoding does contribute is the reason it succeeds. In Two-way interference channel capacity: How to have the cake and eat it too, receivers decode later symbols first and use them to cancel interference in earlier slots. In Capacity-Achieving Rateless Polar Codes, bits decoded in a later block become known ("frozen") bits for an earlier block, so decoding can proceed backward. In the multiple-access schemes of Multi-layer Gelfand-Pinsker strategies for the generalised multiple-access channel and the relay schemes of On Joint Source-Channel Coding for Correlated Sources Over Multiple-Access Relay Channels, the backward chain starts at a terminal block with no fresh message. These papers supply the structure we reuse: a superposition of new and relayed signal, an anchor block, and successive cancellation.

To write the spatial analogue, let $y_t$ be the vector of deforestation, one entry per municipality in year $t$. Let $\Delta x_t$ be the vector of cropland expansion, and let $W$ be the neighbour weight matrix. A minimal lagged-leakage model is

$$ y_t = \beta\,\Delta x_t + \sum_{k=0}^{K} \theta_k\, W \Delta x_{t-k} + u_t , $$

where $\beta$ is the direct effect, $\theta_k$ is the leakage that arrives $k$ years after a neighbour's expansion, and $u_t$ is noise.

When $\Delta x_t$ trends smoothly, which is typical in a commodity boom, the terms $W\Delta x_{t-k}$ are nearly collinear. Contemporaneous leakage and lagged leakage then cannot be told apart, however the model is estimated. An anchor period is a stretch where $\Delta x$ is near zero for a set of units while their neighbours' earlier expansion is non-zero. In that stretch $y_t$ reflects only the $\theta_{k\ge1}$ terms. Peeling backward from the anchors then recovers the $\theta_k$ one lag at a time. The claim is about identification, which is why it matters for policy. The 1,122 km² of avoided deforestation that Spatial spillover effects from agriculture drive deforestation in Mato Grosso, Brazil reports for a 10% cropland reduction comes from a model whose spillovers are contemporaneous. Whether multi-year leakage would revise that figure up or down is an open question, and it can only be answered where anchors exist.

Research Program

The first component is theoretical. We will derive the identification condition for the model above, extended to two-step displacement chains (soy pushes out cattle, which push into forest) through powers of $W$. We will state the condition as a rank requirement on the stacked "relayed-only" design matrix from anchor periods. We will also derive the Fisher information for the $\theta_k$ as a function of how many anchor unit-years there are and where they sit. Here the communications analogy becomes concrete. As in block-Markov schemes, each lag can be recovered only if some period exposes it without fresh interference. The theory should therefore predict how many anchors each lag requires.

The second component is an estimator and a diagnostic. For the estimator, we will use joint Bayesian estimation of the full panel, which is the soft, all-blocks equivalent of decoding every block together. We will run backward peeling alongside it only as an interpretable check that each lag is pinned down by a specific set of anchor periods. For the diagnostic, we will report, for any panel and weight matrix, which leakage lags the data can identify and which lags rest entirely on the prior or on functional form.

The third component is design. If a regulator rolls out a policy in waves, the order of treatment determines where anchors fall. We will derive rollout orders that maximise identifiability of the lag profile. These are analogous to the way incremental freezing schedules retransmissions so that later blocks unlock earlier ones.

Baselines

The naive reference point is the aspatial panel regression. On Mato Grosso, Spatial spillover effects from agriculture drive deforestation in Mato Grosso, Brazil reports that this model attributes only 165 km² of avoided deforestation to a 10% cropland reduction. Any spatial method should clear this bar easily. Its role is to show the scale of the error at stake.

The strongest existing published method is the static Spatial Durbin Model from the same paper and from Unveiling Drivers of Deforestation: Evidence from the Brazilian Amazon. This is a panel regression that splits effects into direct effects and spillovers through $W$, and it reports 1,122 km² for the same counterfactual. It is the strongest competitor because it is the accepted leakage estimate on the very panel we use. We will refit it on the identical 141 municipalities, years 2006–2017, and covariates. "Beating" it does not mean producing a larger number. It means two things: in calibrated simulations, our approach recovers the true cumulative leakage with lower error; and on real data, it gives better held-out prediction of frontier-municipality deforestation in the final years.

The second strong competitor is a dynamic Spatial Durbin Model with distributed lags, estimated jointly over all years. This is the Spatial Durbin Model extended with the forward-lag structure that Water pollution, industrial agglomeration and economic growth: Evidence from China uses in its distributed-lag analysis of the Yangtze River Economic Belt. We found no reported result on Lacuna for a lagged spatial model on Mato Grosso, so we will fit one ourselves. It will use the same lag depth $K$, the same $W$, and the same tuning budget for lag selection.

For forecasting comparisons, we include the CNN-GRU-LSTM-Attention hybrid of A STUDY OF PROVINCIAL CARBON EMISSIONS IN CHINA'S CONSTRUCTION INDUSTRY USING THE STIRPAT AND DEEP LEARNING MODEL. It reports an R² of 0.993 on provincial emissions, but that is predictive fit, not identification of leakage. We will run a bidirectional variant with an equal hyperparameter search budget.

The key ablation removes or masks the anchor periods. In simulation, we generate matched panels that are identical except that the anchors are replaced by continued smooth expansion. We then compare how tightly the posterior pins down each $\theta_k$. If masking anchors does not widen the posteriors, the central claim fails.

Why not just…?

Why not just fit a Spatial Durbin Model with distributed lags? This is the estimator we will in fact use inside our approach, so the question is fair. The point is that fitting the model does not tell you whether its lag coefficients are identified. On a smoothly trending panel, it will return confident-looking $\theta_k$ that are really driven by collinearity. Our contribution is the anchor condition and the diagnostic that says when those coefficients mean something. If the Mato Grosso panel turns out to have plenty of anchors, the plain distributed-lag model will match us. We will report that outcome as validating its use there, and it is a real possibility.

Why not just use a bidirectional LSTM or graph neural network? These models condition on future years and can fit very well, as the R² of 0.993 shows. However, they report no structural parameters, so they cannot answer "how much leakage arrives after two years". Like any regression, they cannot create identifying variation that the data lack. We expect them to match us on in-sample fit and to fail the anchor-masking test in the same way, giving confident but wrong counterfactuals. The simulation will check this directly, because the true lag profile is known there.

Why not just use a Kalman smoother? We agree that a forward-backward smoother is the right computational tool, and it is equivalent to our joint estimator in the linear case. Proposing "backward decoding" as a new algorithm would therefore add nothing. A smoother cannot make an unidentified lag profile identified, however. What it cannot provide, and what we provide, is a statement of which periods carry the identifying information and a rollout design that creates them.

Why not just use a causal sliding-window filter? Leakage assessments are retrospective over multi-year policy periods, so storing the full panel costs nothing. A forward-only filter discards exactly the anchor periods that follow a shock. This parallels the result in On Joint Source-Channel Coding for Correlated Sources Over Multiple-Access Relay Channels that sequential forward decoding is strictly weaker than simultaneous backward decoding in its relay setting.

Proposed Experiments

The first experiment is a calibrated simulation. We will build synthetic panels on the real Mato Grosso adjacency of 141 municipalities over 12 years, with direct effects scaled to the published Spatial Durbin estimates. We will set known lag kernels, including two-hop cattle displacement, and vary both the number and the placement of anchor periods. For every method, we will measure error in the $\theta_k$, error in cumulative avoided deforestation, and the coverage of uncertainty intervals.

The decisive check is that posterior width for the lagged terms grows as anchors are removed, roughly in the way the derived Fisher information predicts. The model is linear and Gaussian, so the information should scale with the anchor-period design matrix. A placebo check uses randomly chosen pseudo-anchor years, which should give no identification gain.

The second experiment works on the real Mato Grosso panel. We will map where anchors exist, which lags are identified, and whether lagged leakage revises the 1,122 km² figure. Held-out evaluation uses the years 2015–2017.

The third experiment transfers the approach to air pollution. In the Yangtze River Delta panel of Urban Sprawl in the Yangtze River Delta: Spatio-Temporal Characteristics and Impacts on PM2.5, the 2013 air-pollution measures act as a shock, and monthly fine-particulate data give many short blocks. We will also use the satellite nitrogen-dioxide record in Rapid growth in nitrogen dioxide pollution over Western China, 2005–2013, which reports Western growth of 8.6 ± 0.9% yr⁻¹ alongside −3.3% yr⁻¹ in the Pearl River Delta. Using these two records, we will ask whether the timing of relocation lags after Eastern regulation is identified or only assumed.

The fourth experiment tests the rollout-design rule in simulation. We will compare anchor-maximising rollout orders against random staggered orders, measuring how well each lets us estimate the leakage lag profile.