Maps of expanding cities show that development thins out as ground gets steeper. That pattern does not say why. A steep parcel may stay rural because building on it costs too much. It may stay rural because the state forbids building there. Or roads, and the capital that follows roads, may simply have gone to the flat valley next door first. Urban growth models blend these causes into one prediction of which cell turns urban next. This proposal borrows the design of clinical retrospective cohort studies to pull them apart. A retrospective cohort study rebuilds a group of subjects from historical records, fixes a common starting date, and follows each subject forward until an event happens. Here the subjects are undeveloped land parcels, the starting date is a regional development shock, and the event is conversion to built-up land.
The central idea is simple. Physical friction from slope rises smoothly as terrain steepens, while land policy in China switches on at fixed slope cutoffs. Following parcels forward from a shared start date lets us measure how fast they convert on either side of such a cutoff. That separates a policy wall from a physical wall. It also lets us test whether the apparent "terrain threshold" moves over time as flat land runs out. A cross-sectional map cannot answer that question.
Core Claim
Zhao et al. found that human land-use intensity on the Loess Plateau collapses once a composite elevation-and-slope score, the Terrain Niche Index, exceeds 0.57. They attribute this partly to physical barriers and partly to the Grain for Green program, which returns sloping cropland to forest and grassland. Their data are a single 2015 snapshot of four village transects, so the two explanations cannot be told apart.
We claim that three measurable quantities can be recovered from annual land-survey records, and that current growth models conflate them. The first is a smooth physical component of terrain friction. The second is a discontinuous policy component that appears at administrative slope cutoffs. The third is a time-varying shift in the friction threshold as a function of how much flat land remains in the surrounding cluster.
The new element is not survival analysis applied to land, which on its own is a relabelled panel regression. It is the combination of four design choices:
- an explicit time zero;
- eligibility defined only by pre-exposure state;
- competing events for ecological conversion;
- a smoothness-versus-jump contrast at policy cutoffs.
Research Program
The work builds a parcel cohort for Liangjiang New District in Chongqing from the National Annual Land Use Survey data used by He et al.. That source provides annual land-use classes, a digital elevation model, slope, and transport networks. Time zero is the district's designation as a national-level development area. This is a shock to development demand that is common to the whole district, and it plays the role that symptom onset plays in Sternlicht et al.. There, anchoring the clock at biological onset rather than hospital admission removed a survivor bias. Anchoring at the designation removes the analogous bias of starting to follow parcels only after some neighbours have already been developed.
The cohort includes every parcel that is non-urban at time zero. The exposure is terrain, measured as slope and Terrain Niche Index at the parcel. The event of interest is conversion to built-up land. Conversion to protected forest or grassland is a competing event, because a parcel reforested under policy can no longer be developed. We estimate the cumulative incidence of development by year using cause-specific hazards and a Fine–Gray model, which is a survival regression that handles outcomes that preclude one another. Following the multilevel logic of Wen et al., who nested 50,912 stroke patients within 375 hospitals, parcels are nested within the separated terrain-bounded clusters that Duan et al. identified in Chongqing. Their finding that inner clusters hold about 44% of population on 9.3% of land motivates the demand-pressure variable described below.
A design choice that growth models get wrong concerns which variables to adjust for. Roads are built in flat valleys because the valleys are flat. Distance to a road measured after time zero is therefore a consequence of terrain, not a confounder of it. Adjusting for it estimates only the part of terrain's effect that does not run through infrastructure.
We report two estimands:
- Total effect: adjusting only for conditions fixed at time zero, namely baseline land class, distance to the pre-existing centre, and baseline road network.
- Direct effect: additionally conditioning on road arrival as a time-varying covariate.
Propensity weighting in the style of Tewari et al. balances parcels of different terrain classes on the time-zero covariates. Tewari et al. balanced prostate cancer patients across treatments with propensity scores before fitting survival models.
The separation of policy from physics uses the slope eligibility cutoff of Grain for Green, set at 25 degrees in the national guidelines, together with any local slope limits in Chongqing's development rules. We model the log hazard of development $h$ at slope $s$ as
$$\log h(s) = f(s) + \delta \cdot \mathbf{1}[s \ge c] + \beta^\top x,$$
where $f$ is a smooth spline standing for physical friction, $c$ is the policy cutoff, $\delta$ is the jump attributable to policy, and $x$ collects the time-zero covariates.
Two predictions follow:
- If the terrain threshold is mostly physical, $\delta$ is near zero and $f$ bends gradually.
- If it is mostly policy, $\delta$ is large and negative while $f$ is shallow near $c$.
Finally, we let $f$ interact with a time-varying measure of remaining flat, unconverted land within the parcel's cluster. This tests whether the topographic wall retreats uphill under demand pressure.
The approach is validated in two further regions:
- Southern Auckland (1996–2006, also from He et al.): a different policy regime with no Grain for Green cutoff. We expect no jump at 25 degrees there.
- Jiaxing (Chen et al.): a flat plain where terrain varies little. This is a negative-control region in which estimated terrain effects should be close to zero.
Baselines
Naive reference. The simplest reference is the cross-sectional approach of Zhao et al.: the share of built-up land by terrain gradient in a single end-year snapshot, with a threshold read off where intensity bottoms out. We compute it on the Liangjiang end-year map with the same terrain variables. It defines what the cohort design must improve on. It should reproduce a threshold, but it gives no way to attribute that threshold to policy or physics, and no way to see it move.
Strongest existing method. This is the LSTM-coupled cellular automaton of He et al.. A cellular automaton here is a grid model in which each cell's chance of turning urban depends on its own attributes and its neighbours. He et al. reported a fuzzy Kappa of 0.48 on Liangjiang, against 0.36 and 0.32 for two perceptron variants. For the flat-region comparison, the Deep Belief Network automaton of Chen et al. reported Kappa between 77.109% and 78.366% on Jiaxing.
We rerun the LSTM automaton on identical parcels, years, and covariates, with the same hyperparameter search budget as our models. We compare on two axes:
- Held-out prediction. Fuzzy Kappa on held-out years, computed by converting our cumulative incidence into annual conversion maps. Our model counts as competitive if it stays within a few Kappa points. We do not expect it to beat a model built purely for pattern accuracy.
- Effect recovery. We extract the automaton's implied slope effect using partial dependence curves. Both methods are then scored on a semi-synthetic benchmark: real Liangjiang covariates, roads, and clusters, with outcomes regenerated from a known smooth terrain effect plus a known policy jump. Beating the automaton means lower bias and correct confidence-interval coverage for both the smooth component and the jump.
Ablations. These isolate the key ingredients. We will (a) drop the fixed time zero and instead enrol parcels whenever they appear non-urban, (b) treat reforestation as censoring rather than a competing event, (c) adjust for post-time-zero road distance in the total-effect model, and (d) remove the discontinuity term. Each should bias the semi-synthetic estimates in a predictable direction. If an ablation does no harm, the corresponding ingredient is not earning its place.
Why not just…?
Why not just use the cellular automaton models, which already use yearly data and slope? They are built to reproduce where cells turn urban, and they do so well. However, their slope input sits beside road distance, and road distance is itself caused by slope. The model therefore has no defined causal quantity to report. He et al. themselves attribute weaker Liangjiang performance to unmodelled, conflicting policy, which is exactly the confound we target. The semi-synthetic recovery test decides the question: if partial dependence from the automaton recovers the planted jump as well as our model does, the added design is unnecessary.
Why not just run a discrete-time logistic regression on the annual panel? This is a real threat. A discrete-time logit with slope splines is mathematically close to a discrete-time hazard model. We will fit it and expect similar coefficients when it is given the same time zero, eligibility rule, and covariate set. What would separate it from our approach is the design choices, not the estimator. Ablations (a) through (c) show what the panel logit gets wrong when fitted in the usual way: all parcel-years pooled, reforestation ignored, current road distance included. If a carefully specified logit matches our estimates, we will say so and present the design as the contribution.
Why not just run a cross-sectional spatial regression of built-up share on slope? It reproduces the threshold of Zhao et al. but cannot tell whether the threshold moves over time. It also cannot distinguish land that was never developable from land withdrawn by policy. The time-varying interaction with remaining flat land cannot be estimated from one snapshot. The countrywide natural experiment on walkability shows how cross-sectional built-environment associations had long been disputed until units were followed over time. That study followed people across relocations rather than parcels, and it did not treat terrain.
Decisive Checks
Four checks decide whether the claim holds:
- Cutoff specificity. The policy jump $\delta$ should appear in Liangjiang at the Grain for Green cutoff. It should be absent at placebo cutoffs of 20 and 30 degrees, and absent in Southern Auckland.
- Negative-control region. In Jiaxing, terrain effects should be statistically indistinguishable from zero.
- Placebo time zero. Moving time zero several years before the designation should flatten the rise in hazard among low-slope parcels.
- Threshold drift. If the threshold drift is real, the slope at which development hazard halves should rise over time in clusters where flat land is depleted. It should stay fixed in clusters with spare flat land.
There are no reported results on Lacuna for time-to-conversion hazards, discontinuity estimates at policy slope cutoffs, or effect-recovery benchmarks for urban growth models. The semi-synthetic benchmark and all hazard estimates here would therefore be new measurements rather than improvements on published numbers.
References
- Analyzing the characteristics of land use distribution in typical village transects at Chinese Loess Plateau based on topographical factors
- Integrating ANNs and Cellular Automata–Markov Chain to Simulate Urban Expansion with Annual Land Use Data
- Timing to intubation COVID-19 patients: can we put it off until tomorrow?
- Risk factors associated with 31-day unplanned readmission in 50,912 discharged patients after stroke in China
- Measuring polycentric urban structure using Easygo big data: A case study of Chongqing metropolitan area
- Long-Term Survival in Men With High Grade Prostate Cancer: A Comparison Between Conservative Treatment, Radiation Therapy and Radical Prostatectomy—A Propensity Scoring Approach
- Integrating Cellular Automata with the Deep Belief Network for Simulating Urban Growth
- Countrywide natural experiment links built environment to physical activity