In DouglasMesquita/RASCO: An R Package to Alleviate Spatial Confounding

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  fig.path = "README-"
)

Acknowlodgment

The authors would like to thank CAPES, CNPq and FAPEMIG for partial financial support.

RASCO: An R package to Alleviate the Spatial Confounding

Overview

RASCO is an R package that allow the practioners to fit restricted spatial models for three class of models:

• Generalized linear mixed models

• Shared component models

• Spatial frailty models

Installing LSB Core

You may need to install the LSB Core

lsb_release is part of a software package called the LSB core, which is not necessarily installed on your system by default. To install it, run the command below that corresponds to your specific system:

Ubuntu, Debian

sudo apt-get update && sudo apt-get install lsb-core

CentOS

sudo yum update && sudo yum install redhat-lsb-core

Fedora

sudo dnf update && sudo dnf install redhat-lsb-core

OpenSUSE

sudo zypper update && sudo zypper install lsb-core

Arch

pacman -Syu lsb-release

Installation

# Install from CRAN (when available)
install.packages("RASCO")
# Or the development version from GitHub
# install.packages("devtools")
devtools::install_github("DouglasMesquita/RASCO")

library(RASCO)
library(spdep)

set.seed(11022020)

Restricted generalized linear mixed models

##-- Spatial structure
data("neigh_RJ")

beta <- c(-0.1, 0.7)
tau <- 1

##-- Data ----
family <- "poisson"
data <- rglmm(beta = beta, tau = tau, family = family,
              confounding = "linear", neigh = neigh_RJ,
              scale = TRUE)

##-- Models ----
sglm_mod <- rsglmm(data = data, formula = Y ~ X1 + X2,
                   family = family,
                   proj = "none", nsamp = 1000)

sglmm_mod <- rsglmm(data = data, formula = Y ~ X1 + X2,
                    area = "reg", model = "besag", neigh = neigh_RJ,
                    family = family,
                    proj = "none", nsamp = 1000)

rglmm_mod <- rsglmm(data = data, formula = Y ~ X1 + X2,
                    area = "reg", model = "r_besag", neigh = neigh_RJ,
                    family = family,
                    proj = "rhz", nsamp = 1000)

sglm_mod$unrestricted$summary_fixed
sglmm_mod$unrestricted$summary_fixed
rglmm_mod$unrestricted$summary_fixed
rglmm_mod$restricted$summary_fixed

sglmm_mod$unrestricted$summary_hyperpar

The SVIF function provides the coefficients variance ratio between two models. We can notice that the restricted model alleviates the variance inflation

SVIF(rglmm_mod$restricted, rglmm_mod$unrestricted)

Restricted shared component models

##-- Spatial structure
data("neigh_RJ")

##-- Parameters
alpha_1 <- 0.5
alpha_2 <- 0.1
beta_1 <- c(-0.5, -0.2)
beta_2 <- c(-0.8, -0.4)
tau_s <- 1
tau_1 <- tau_2 <- 10
delta <- 1.5

##-- Data
data <- rshared(alpha_1 = alpha_1, alpha_2 = alpha_2, beta_1 = beta_1, beta_2 = beta_2, delta = delta,
                tau_1 = tau_1, tau_2 = tau_2, tau_s = tau_s,
                confounding = "linear",
                neigh = neigh_RJ)

##-- Models
scm_mod <- rscm(data = data,
                formula1 = Y1 ~ X11 + X12,
                formula2 = Y2 ~ X21 + X12,
                E1 = E1, E2 = E2,
                family = c("poisson", "poisson"),
                area = "reg", neigh = neigh_RJ,
                priors = list(prior_prec = c(0.5, 0.05), 
                              prior_gamma = c(0, 0.5)),
                proj = "none", nsamp = 1000,
                random_effects = list(shared = TRUE, specific_1 = TRUE, specific_2 = TRUE))

rscm_mod <- rscm(data = data,
                 formula1 = Y1 ~ X11 + X12,
                 formula2 = Y2 ~ X21 + X12,
                 E1 = E1, E2 = E2,
                 family = c("poisson", "poisson"),
                 area = "reg", neigh = neigh_RJ,
                 priors = list(prior_prec = list(tau_s = c(0.5, 0.05), 
                                                 tau_1 = c(0.5, 0.05), 
                                                 tau_2 = c(0.5, 0.05)), 
                               prior_gamma = c(0, 0.5)),
                 proj = "spock", nsamp = 1000,
                 random_effects = list(shared = TRUE, specific_1 = TRUE, specific_2 = TRUE))

##-- Summary
scm_mod$summary_fixed
rscm_mod$summary_fixed

scm_mod$summary_hyperpar
rscm_mod$summary_hyperpar

SVIF(rscm_mod, scm_mod)

Restricted spatial frailty models

#-- Spatial structure
data("neigh_RJ")

##-- Individuals and regions
n_reg <- length(neigh_RJ)
n_id <- sample(x = 3:5, size = n_reg, replace = T)
beta <- c(0.3, -0.3)
tau <- 0.75 # Scale of spatial effect

##-- Data
data <- rsurv(n_id = n_id,
              coefs = beta, cens = 0.5, scale = FALSE,
              cens_type = "right", hazard = "weibull",
              hazard_params = hazard_params <- list(weibull = list(alpha = 1.2, variant = 0)),
              spatial = "ICAR",
              neigh = neigh_RJ, tau = tau, confounding = "linear", proj = "none")

##-- Models
weibull_mod <- rsfm(data = data,
                    formula = surv(time = L, event = status) ~ X1 + X2,
                    model = "none", family = "weibull",
                    proj = "rhz", nsamp = 1000, approach = "inla")

rsfm_mod <- rsfm(data = data, area = "reg",
                 formula = surv(time = L, event = status) ~ X1 + X2,
                 model = "r_besag", neigh = neigh_RJ, family = "weibull",
                 proj = "rhz", nsamp = 1000, approach = "inla")

weibull_mod$unrestricted$summary_fixed
rsfm_mod$unrestricted$summary_fixed
rsfm_mod$restricted$summary_fixed

SVIF(weibull_mod$unrestricted, rsfm_mod$unrestricted)
SVIF(weibull_mod$unrestricted, rsfm_mod$restricted)

DouglasMesquita/RASCO documentation built on Nov. 16, 2022, 9:42 p.m.