LoginSign Up
← Back to all sparks
S

spatstat.model

ANALYTICS
Velocity2.5

Sub-package of spatstat containing functionality for parametric modelling and inference

github.com/spatstat/spatstat.model ↗

spatstat's inference layer builds out determinantal and cluster process fitting

spatial-statisticspoint-processesmodel-fittingr-packagediagnostics
◆Current state
spatstat.model fits point process models and provides the diagnostics that go with them. The recent window is dominated by determinantal point process work — a variance-covariance matrix and more diagnostics in 3.7-2, additional `intensity` and `repul` methods in 3.7-1, and ROC curves for determinantal models in 3.5-0. Cluster and Cox process inference has advanced in parallel, with Waagepetersen's composite likelihood arriving in 3.6-1.
◆Where it's heading
The pattern is that model classes enter the package as fitting machinery first and only later gain the apparatus that makes them usable in practice — standard errors, diagnostics, residuals, model checking. Determinantal processes are visibly midway through that progression, reaching variance-covariance estimation only in the most recent release. Around this, the package has been broadening where models can be fitted at all: replicated point patterns on linear networks in 3.5-0, extended spatial logistic regression, and conversion of recursively partitioned models to tessellations.
◆Prediction
Expect determinantal model support to keep filling out along the same path other model classes took, since variance estimation has only just arrived and partial residuals already exist for the cluster and Cox families. The entries do not signal a move into three dimensions here, unlike the geometry and simulation packages.

◆Recent moves

  1. 18d ago

    Variance-covariance and diagnostics for determinantal models

    Determinantal point process models gain a variance-covariance matrix and more diagnostics, plus variable plots across all model kinds. Variance estimation is what turns a fitted determinantal model into one you can put standard errors on, so this closes a real gap rather than adding a convenience.

    View source ↗
  2. 2mo ago

    More intensity and repul methods; boundary-aware predictions

    More methods for `intensity` and `repul`, and improved handling of boundary pixels in model predictions. The boundary pixel work mirrors the same fix in the geometry package, applied where predictions are evaluated on a discretised window.

    View source ↗
  3. 6mo ago

    Composite likelihood for cluster processes

    Waagepetersen's composite likelihood for cluster processes arrives, alongside conversion of a recursively partitioned point process model to a tessellation, an `rmax` argument on diagnostic functions, and cluster strength and spatial persistence indices. The composite likelihood gives cluster process fitting an estimator with established theory behind it.

    View source ↗
  4. 8mo ago

    Replicated network models and partial residuals

    Models for replicated point patterns on linear networks — fitting across several observed patterns on a network rather than one — plus ROC curves for determinantal models, partial residuals for cluster and Cox processes, and extended spatial logistic regression. Three separate model families gain checking apparatus in one release.

    View source ↗
  5. 10mo ago

    intensity.ppm improvements for Geyer models

    Improvements to `intensity.ppm` for Geyer models and better support for simulating a fitted model in a different window. Refinements to existing fitting paths with no new model class or diagnostic.

    View source ↗
  6. 1y ago

    ROC curve support substantially extended

    Substantial extensions to ROC curves, the release that begins the run of model-checking work continuing through 3.5-0's determinantal ROC curves and 3.7-2's diagnostics. The one-line entry understates a change the following releases keep building on.

    View source ↗