- Twenty methods to estimate the Log Gaussian Cox Course of mannequin with level and aggregated case knowledge: the rts2 bundle for R
Authors: Samuel I Watson
Summary: The R bundle rts2 gives knowledge manipulation and mannequin becoming instruments for Log Gaussian Cox Course of (LGCP) fashions. LGCP fashions are a key methodology for illness and different sorts of surveillance, and supply a way of predicting danger throughout an space of curiosity based mostly on spatially-referenced and time-stamped case knowledge. Nonetheless, these fashions could be troublesome to specify and computationally demanding to estimate. For a lot of surveillance eventualities we require leads to close to real-time utilizing routinely accessible knowledge to information and direct coverage responses, or resulting from restricted availability of computational assets. There are restricted software program implementations accessible for this real-time context with dependable predictions and quantification of uncertainty. The rts2 bundle gives a spread of recent Gaussian course of approximations and mannequin becoming strategies to suit the LGCP, together with estimation of covariance parameters, utilizing each Bayesian and stochastic Most Probability strategies. The bundle gives a set of information manipulation instruments. We additionally present a novel implementation to estimate the LGCP when case knowledge are aggregated to an irregular grid reminiscent of census tract areas.
2. On the Laplace Approximation as Mannequin Choice Criterion for Gaussian Processes
Authors: Andreas Besginow, Jan David Hüwel, Thomas Pawellek, Christian Beecks, Markus Lange-Hegermann
Summary: Mannequin choice goals to search out the most effective mannequin when it comes to accuracy, interpretability or simplicity, ideally all of sudden. On this work, we concentrate on evaluating mannequin efficiency of Gaussian course of fashions, i.e. discovering a metric that gives the most effective trade-off between all these standards. Whereas earlier work considers metrics just like the probability, AIC or dynamic nested sampling, they both lack efficiency or have important runtime points, which severely limits applicability. We tackle these challenges by introducing a number of metrics based mostly on the Laplace approximation, the place we overcome a extreme inconsistency occuring throughout naive utility of the Laplace approximation. Experiments present that our metrics are comparable in high quality to the gold normal dynamic nested sampling with out compromising for computational pace. Our mannequin choice standards permit considerably sooner and top quality mannequin collection of Gaussian course of fashions
