Source-linked AI summary
INLA or MCMC? A Tutorial and Comparative Evaluation for Spatial Prediction in log-Gaussian Cox Processes
Benjamin M. Taylor, Peter J. Diggle
TL;DR
For spatial log-Gaussian Cox process inference, the paper evaluates the trade-off between MCMC’s computational and convergence challenges and INLA’s approximation error. It compares the methods for predictive inference and finds that MCMC can provide more accurate predictive probabilities, depending on settings, while remaining comparatively quick.
Problem
MCMC can be computationally burdensome and difficult to validate for convergence, whereas INLA is faster but may introduce errors in posterior probabilities; the trade-off matters when predictive probabilities inform potentially costly public-health interventions.
Source: §1 · pp. 2–3Method
The paper provides a tutorial and critically compares MCMC and INLA for predictive Bayesian inference in spatial log-Gaussian Cox processes.
Source: §1 · p. 2, §7 · p. 28Results
MCMC can yield more accurate predictive-probability estimates than INLA-based methods, depending on the chosen settings, and can do so comparatively quickly with less bias.
Source: §7 · p. 28Takeaways & Limitations
The authors advise appropriate caution when using INLA for this model class and raise the utility of gradient-based MCMC as an open question.
Source: §7 · p. 29The evaluation is restricted to spatial rather than spatio-temporal prediction, primarily for ease of exposition; pilot runs found INLA quite slow for spatio-temporal prediction on a regular grid.
Source: §7 · p. 28
Abstract
from arXiv · showhide
We investigate two options for performing Bayesian inference on spatial log-Gaussian Cox processes assuming a spatially continuous latent field: Markov chain Monte Carlo (MCMC) and the integrated nested Laplace approximation (INLA). We first describe the device of approximating a spatially continuous Gaussian field by a Gaussian Markov random field on a discrete lattice, and present a simulation study showing that, with careful choice of parameter values, small neighbourhood sizes can give excellent approximations. We then introduce the spatial log-Gaussian Cox process and describe MCMC and INLA methods for spatial prediction within this model class. We report the results of a simulation study in which we compare MALA and the technique of approximating the continuous latent field by a discrete one, followed by approximate Bayesian inference via INLA over a selection of 18 simulated scenarios. The results question the notion that the latter technique is both significantly faster and more robust than MCMC in this setting; 100,000 iterations of the MALA algorithm running in 20 minutes on a desktop PC delivered greater predictive accuracy than the default \verb=INLA= strategy, which ran in 4 minutes and gave comparative performance to the full Laplace approximation which ran in 39 minutes.
1 Introduction
The article compares MCMC and INLA for Bayesian inference in spatial log-Gaussian Cox processes, while also providing a tutorial on the computation and inference involved. The comparison is motivated by the need for accurate predictive probabilities in spatial health surveillance.
1 Introduction
- The article compares relatively slow but asymptotically exact MCMC with faster but approximate INLA for Bayesian inference in spatial log-Gaussian Cox processes.It also aims to provide a tutorial on technical aspects of computation and inference for these models.Source: §1 · p. 2
- Predictive probabilities that disease risk exceeds a public-health threshold matter because high probabilities may activate costly interventions.The authors therefore compare methods using a metric of predictive-probability accuracy.Source: §1 · p. 2
- MCMC can be computationally burdensome and its validity depends on chain convergence, while INLA typically computes faster but may introduce approximation error.INLA uses analytical approximation and numerical integration rather than Monte Carlo simulation.Source: §1 · pp. 2–3
- The authors use a regular square grid rather than Lindgren et al. (2011)’s triangulation approach, retaining flexibility in covariance choice through Rue and Held (2005)’s method.The article also outlines sections on Gaussian-field approximation, the Cox process, MCMC, INLA, and conclusions.Source: §1 · pp. 3–4
2 Spatially Continuous Gaussian processes and their Approx-
The section constructs sparse GMRF approximations to spatially continuous Gaussian fields by fitting local grid neighborhoods, then evaluates their accuracy and computation. In the simulations, 2-neighborhoods substantially improve on 1-neighborhoods, though optimization can be less robust.
2.1 Theory
- The dense covariance and its inverse are computationally difficult to manipulate, but toroidal grid extension makes the covariance block-circulant and enables DFT-based computations.The section describes a 128 × 128 extended grid whose covariance matrix has dimension 16,384 × 16,384.Source: §2.1 · pp. 5, 6, 8
- A sparse GMRF is fitted by parametrizing its precision matrix so its implied covariance is close to the dense covariance of the continuous field.Neighborhoods specify local dependencies; small neighborhoods yield sparse precision matrices that support fast matrix operations.Source: §2.1 · pp. 5, 8, 9
2.2 Simulation Study
- The field approximation was assessed by simulating Gaussian realizations and estimating integrated mean square error across repeated realizations.The simulations fixed σ at 1, varied φ from 0.025 to 0.2, and used two extended-grid resolutions to reduce dependence on grid-size artifacts.Source: §2.2 · pp. 11–12
- 2-neighborhood approximations significantly improve field accuracy over 1-neighborhoods, while 3-neighborhoods add no apparent accuracy and are more robust to optimization problems.BFGS appeared to converge to local sub-optimal values in two 2-neighborhood cases; Nelder–Mead improved those results.Source: §2.2 · p. 12, Table 1 · p. 13
- For sensible φ relative to the observation window, the study found very good sparse approximations and optimization times around a minute for 256 × 256 grids, but optimization was not always robust.The 2-neighborhood was judged sufficiently accurate for the later study because its φ values were small relative to the observation window.Source: §2.2 · pp. 12–13
3 The Spatial Log-Gaussian Cox Process
The spatial log-Gaussian Cox process models events in a window using an intensity with a known spatial component and a latent Gaussian field. The paper uses a grid-discretized field for prediction, treating intensity as constant within cells.
3 The Spatial Log-Gaussian Cox Process
- The process models event locations in a spatial window with a random intensity, and event counts in subsets are conditionally Poisson given that intensity.Source: §3 · p. 13
- The intensity separates a known spatial component and expected event count from a continuous Gaussian field that captures residual variation and spatial clustering.When the known components are specified, inference about the residual field informs extraneous spatial clustering.Source: §3 · pp. 13–14
- The model uses a discretized field on a regular square grid, treating intensity as constant within each cell and choosing fine spacing so multiple counts per cell are rare and approximation error is negligible.The finite field has an approximate covariance matrix computed at grid-cell centroids.Source: §3 · p. 14
4 Markov Chain Monte Carlo
MCMC prediction for the log-Gaussian Cox process samples the latent field conditional on observations, using its full covariance matrix. The section focuses on MALA, a Langevin-type Metropolis–Hastings method, and discusses tuning and computation for this approach.
4 Markov Chain Monte Carlo
- Predictive inference about Y requires samples from the conditional distribution of the latent field Y given observations X, using the full covariance matrix.The full covariance matrix is used in the MCMC approach to inference for this model.Source: §4 · p. 15
- MCMC draws dependent samples that approach the target density in the limit, but poor mixing and uncertain convergence can require prohibitively long runs.Reliable inference from relatively short runs depends on apparent convergence to stationarity and low autocorrelation.Source: §4 · p. 15
- For the log-Gaussian Cox process, the proposal kernel is the Metropolis-Adjusted Langevin algorithm, a more sophisticated Metropolis–Hastings variant.A Langevin-type proposal is natural when the transition-density gradient can be written explicitly.Source: §4 · p. 16
- Møller et al.’s linear transformation of Y_ext makes Monte Carlo simulation from π(Y|X) more efficient.The transformation is written using Γ = Σ^−1/2.Source: §4 · p. 16
- The target acceptance probability is often set to 0.574, approximately optimal for a Gaussian target as dimension tends to infinity.An automatic choice of h that achieves this target while preserving ergodicity is implemented in the lgcp R package.Source: §4 · p. 17
- The target log density is specified for the model, and its gradient can be computed using the DFT.S denotes observation-window grid cells, CA their area, and X(s) the event count in cell s.Source: §4 · p. 17
5 The Integrated Nested Laplace Approximation
The GMRF/INLA approach first approximates the full covariance matrix of Y, then uses that approximation for Bayesian inference with INLA. INLA approximates posterior marginals for the latent field and its hyperparameters.
5 The Integrated Nested Laplace Approximation
- The GMRF/INLA approach uses two approximations: it approximates Y’s full covariance matrix, then performs Bayesian inference with INLA using that covariance approximation.The covariance approximation yields substantial computational benefits.Source: §5 · pp. 17–18
- INLA provides approximate posterior marginals for the latent field given observed data and for the latent-field parameters.The paper denotes the latent-field hyperparameters by ϑ, distinguishing them from parameters θ and φ mentioned earlier.Source: §5 · p. 18
- INLA replaces exact relations for posterior marginals of the latent field and hyperparameters with approximations.The approximate density is denoted by ˜π; the text also describes Gaussian approximations used in this procedure.Source: §5 · pp. 18–19
- A quadrature scheme over ϑ is used to evaluate marginal posteriors, with optimized routines implemented in the R INLA package.For the simulation study, the authors used INLA’s generic0 model, fixed redundant τ = 1, and omitted an offset because the mean µ was known.Source: §5 · p. 19
6 Simulation Results
Across the simulated scenarios, MALA generally gave more accurate probability predictions than the GMRF-plus-INLA methods, while INLA ran faster in its approximate settings. A follow-up study without the GMRF approximation reduced the visible S-shaped bias, but MALA still outperformed both INLA variants.
6.1 The Simulated Scenarios
- The 18 scenarios combined two fixed spatial components with exponential covariance parameters φ = 0.02, 0.04, 0.6 and σ = 0.5, 1, 2.The authors expected larger σ to make prediction harder for both methods and larger φ to challenge the GMRF approximation.Source: §6.1 · pp. 19–20
6.2 Algorithm Parameters
- GAIN1 and GAIN2 used simplified Laplace inference with neighbourhood sizes 1 and 2; GAIN3 used neighbourhood size 2 with the full Laplace approximation.MALA ran for 100,000 iterations with a 10,000-iteration burn-in, retaining every 90th sample.Source: §6.2 · pp. 20–21
6.3 Results
- Across the 18 scenarios, GAIN3 took 39 minutes, MALA 20 minutes, GAIN2 around 4 minutes, and GAIN1 around 1.5 minutes.The timings used an INLA version available on 8 February 2012; the authors were informed that later Laplace-strategy work could reduce computation time by 30–50%.Source: §6.3 · p. 21, Table 2 · p. 21
- GAIN1 and GAIN2 outperformed MCMC in only a few scenarios; GAIN2’s relative MSE2 ranged from 1.045 to 82.41, with a median of 3.38.Increasing neighbourhood size from 1 to 2 did not reduce median relative MSE; GAIN3’s median relative MSE versus MALA was 1, with values of 0.76 for fixed λ1(s) and 3.72 for fixed λ2(s).Source: §6.3 · p. 24, Table 4 · p. 24
- The INLA probability estimates showed an S-shaped bias around y = x, whereas MALA estimates were approximately symmetrically distributed around it.The plots also indicate that GAIN2 and GAIN3 would give poor probability estimates in two scenarios.Source: §6.3 · p. 25, Figure 3 · p. 25
7 Discussion
The discussion concludes that MCMC can provide more accurate predictive probabilities than INLA-based methods, depending on settings, and argues that INLA should be treated with caution. It also identifies limits in the study’s spatial scope, MCMC methods, and software comparison.
7 Discussion
- The article provides a tour of techniques for spatial prediction in log-Gaussian Cox processes and independently evaluates GMRF approximation alongside a critical comparison of Bayesian inference methods.Source: §7 · p. 28
- The evaluation covers spatial rather than spatio-temporal prediction, and pilot runs found INLA slow for spatio-temporal prediction on a regular grid.The authors chose MALA for ease of implementation and stability, while noting that more sophisticated methods have better theoretical mixing properties.Source: §7 · p. 28
- MCMC can yield more accurate predictive probabilities than INLA-based methods, depending on settings, while producing probabilities comparatively quickly and with less bias.The authors therefore advise appropriate caution when using GMRF approximation followed by INLA for this model class.Source: §7 · pp. 28–29
- The study did not assess the full capabilities of lgcp and INLA, including INLA’s broader latent-Gaussian model support and hyperparameter-estimation framework.The authors note that hyperparameter sampling is challenging for MCMC.Source: §7 · p. 29