Waagepetersen-Guan (2009) Two-Step Estimation for Inhomogeneous Spatial Point Processes

Summary

Waagepetersen and Guan give a computationally tractable two-step procedure for fitting inhomogeneous spatial point processes in which the intensity varies with covariates and the points also exhibit residual clustering. Step 1 estimates the intensity-regression (first-order) parameters $\beta$ by a Poisson-likelihood score estimating function; Step 2 takes $\hat\beta$ as given and estimates the clustering (second-order) parameters by minimum-contrast fitting of the inhomogeneous $K$-function. They establish asymptotic normality of both stages under mixing conditions, and apply the method to tropical-rainforest tree data to disentangle habitat- versus dispersal-driven aggregation.

Key Claims

• Intensity regression. The point process $X$ on $\mathbb R^2$ has intensity $\rho_\beta(u)=\rho(z(u)^\top\beta)$ (e.g. log-linear) in spatial covariates $z(u)$ — soil, topography — capturing inhomogeneity (first-order structure).
• Step 1: estimating the intensity by a Poisson score. $\beta$ is estimated by solving a Poisson-likelihood score estimating function — the same estimating equation one would use if the points were an inhomogeneous Poisson process. It is consistent for $\beta$ even when the process is not Poisson (i.e. even when the points cluster), because the first-order (intensity) property does not depend on the higher-order dependence — a key robustness point.
• Step 2: clustering by minimum contrast. Residual dependence (clustering beyond that induced by the covariates) is captured by the inhomogeneous $K$-function (second-order reduced moment). Its parameters (e.g. those of a log-Gaussian Cox or Neyman-Scott cluster process) are estimated by minimum-contrast — minimizing an integrated squared discrepancy between the empirical and model $K$-functions — plugging in $\hat\beta$.
• Asymptotics. Both $\hat\beta$ and the minimum-contrast clustering estimates are asymptotically normal under mixing (weak-dependence) conditions on the point process over an increasing observation window.
• Rainforest application. Fitted to individual tree locations with soil/terrain covariates, the method separates aggregation attributable to habitat/niche (via the intensity covariates) from aggregation attributable to seed dispersal (via the residual clustering), addressing diversity-coexistence hypotheses in ecology.

Concepts Introduced or Extended

• Spatial Point Process

Entities Mentioned

• Rasmus Waagepetersen
• Yongtao Guan

Quotes

"Regression parameters are estimated by using a Poisson likelihood score estimating function and in the second step minimum contrast estimation is applied for the residual clustering parameters."

"Asymptotic normality of parameter estimates is established under certain mixing conditions and we exemplify how the results may be applied in ecological studies of rainforests."

My Take

The elegant idea is the separation of first- and second-order structure: because the intensity (where points tend to be) is a first-moment property while clustering (how points co-occur) is a second-moment property, you can estimate them in sequence — a Poisson score for the covariate effects that is valid despite the clustering, then a $K$-function minimum-contrast fit for the dependence — instead of a single intractable joint likelihood. That is both computationally practical for the huge point patterns modern spatial data produce and conceptually clarifying, and it is exactly what the ecological question needs: attribute tree aggregation to habitat (intensity covariates) versus dispersal (residual clustering). For the wiki it anchors a new spatial point process concept — the continuous-space, event-location counterpart to the areal CAR and temporal counting-process machinery already present — and its transferable lesson is the composite-likelihood / estimating-function strategy for models whose full likelihood is unavailable.