Core Functionality of the 'spatstat' Family

Functionality for data analysis and modelling of spatial data, mainly spatial point patterns, in the 'spatstat' family of packages. (Excludes analysis of spatial data on a linear network, which is covered by the separate package 'spatstat.linnet'.) Exploratory methods include quadrat counts, K-functions and their simulation envelopes, nearest neighbour distance and empty space statistics, Fry plots, pair correlation function, kernel smoothed intensity, relative risk estimation with cross-validated bandwidth selection, mark correlation functions, segregation indices, mark dependence diagnostics, and kernel estimates of covariate effects. Formal hypothesis tests of random pattern (chi-squared, Kolmogorov-Smirnov, Monte Carlo, Diggle-Cressie-Loosmore-Ford, Dao-Genton, two-stage Monte Carlo) and tests for covariate effects (Cox-Berman-Waller-Lawson, Kolmogorov-Smirnov, ANOVA) are also supported. Parametric models can be fitted to point pattern data using the functions ppm(), kppm(), slrm(), dppm() similar to glm(). Types of models include Poisson, Gibbs and Cox point processes, Neyman-Scott cluster processes, and determinantal point processes. Models may involve dependence on covariates, inter-point interaction, cluster formation and dependence on marks. Models are fitted by maximum likelihood, logistic regression, minimum contrast, and composite likelihood methods. A model can be fitted to a list of point patterns (replicated point pattern data) using the function mppm(). The model can include random effects and fixed effects depending on the experimental design, in addition to all the features listed above. Fitted point process models can be simulated, automatically. Formal hypothesis tests of a fitted model are supported (likelihood ratio test, analysis of deviance, Monte Carlo tests) along with basic tools for model selection (stepwise(), AIC()) and variable selection (sdr). Tools for validating the fitted model include simulation envelopes, residuals, residual plots and Q-Q plots, leverage and influence diagnostics, partial residuals, and added variable plots.

Functions for Optimal Non-Bipartite Matching

Perform non-bipartite matching and matched randomization. A "bipartite" matching utilizes two separate groups, e.g. smokers being matched to nonsmokers or cases being matched to controls. A "non-bipartite" matching creates mates from one big group, e.g. 100 hospitals being randomized for a two-arm cluster randomized trial or 5000 children who have been exposed to various levels of secondhand smoke and are being paired to form a greater exposure vs. lesser exposure comparison. At the core of a non-bipartite matching is a N x N distance matrix for N potential mates. The distance between two units expresses a measure of similarity or quality as mates (the lower the better). The 'gendistance()' and 'distancematrix()' functions assist in creating this. The 'nonbimatch()' function creates the matching that minimizes the total sum of distances between mates; hence, it is referred to as an "optimal" matching. The 'assign.grp()' function aids in performing a matched randomization. Note bipartite matching can be performed using the prevent option in 'gendistance()'.

Wrapper Algorithm for All Relevant Feature Selection

An all relevant feature selection wrapper algorithm. It finds relevant features by comparing original attributes' importance with importance achievable at random, estimated using their permuted copies (shadows).

Multinomial Logit Models with Random Parameters

An implementation of maximum simulated likelihood method for the estimation of multinomial logit models with random coefficients as presented by Sarrias and Daziano (2017) . Specifically, it allows estimating models with continuous heterogeneity such as the mixed multinomial logit and the generalized multinomial logit. It also allows estimating models with discrete heterogeneity such as the latent class and the mixed-mixed multinomial logit model.

Markov Chain Monte Carlo

Simulates continuous distributions of random vectors using Markov chain Monte Carlo (MCMC). Users specify the distribution by an R function that evaluates the log unnormalized density. Algorithms are random walk Metropolis algorithm (function metrop), simulated tempering (function temper), and morphometric random walk Metropolis (Johnson and Geyer, 2012, , function morph.metrop), which achieves geometric ergodicity by change of variable.

Cholesky Decomposition of the Wishart Distribution

Sampling from the Cholesky factorization of a Wishart random variable, sampling from the inverse Wishart distribution, sampling from the Cholesky factorization of an inverse Wishart random variable, sampling from the pseudo Wishart distribution, sampling from the generalized inverse Wishart distribution, computing densities for the Wishart and inverse Wishart distributions, and computing the multivariate gamma and digamma functions. Provides a header file so the C functions can be called directly from other programs.

Random Network Model Estimation, Selection and Parameter Tuning

Model fitting, model selection and parameter tuning procedures for a class of random network models. Many useful network modeling, estimation, and processing methods are included. The work to build and improve this package is partially supported by the NSF grants DMS-2015298 and DMS-2015134.

Distances and Routes on Geographical Grids

Provides classes and functions to calculate various distance measures and routes in heterogeneous geographic spaces represented as grids. The package implements measures to model dispersal histories first presented by van Etten and Hijmans (2010) . Least-cost distances as well as more complex distances based on (constrained) random walks can be calculated. The distances implemented in the package are used in geographical genetics, accessibility indicators, and may also have applications in other fields of geospatial analysis.

Raster Randomization for Null Hypothesis Testing

Randomization of presence/absence species distribution raster data with or without including spatial structure for calculating standardized effect sizes and testing null hypothesis. The randomization algorithms are based on classical algorithms for matrices (Gotelli 2000, ) implemented for raster data.

Normal aka Gaussian 1-d Mixture Models

Onedimensional Normal (i.e. Gaussian) Mixture Models (S3) Classes, for, e.g., density estimation or clustering algorithms research and teaching; providing the widely used Marron-Wand densities. Efficient random number generation and graphics. Fitting to data by efficient ML (Maximum Likelihood) or traditional EM estimation.