In hunzikp/dimstar: Discrete Multivariate Spatio-temporal Autoregressive Models

knitr::opts_chunk$set(echo = TRUE)

dimstar is an R package implementing a novel MCEM algorithm for estimating latent-Guassian spatio-temporal autoregressive models for potentially multivariate binary and count outcomes.

Installation

Install dimstar straight from github:

require(devtools)
devtools::install_github('hunzikp/dimstar')

NOTE: This is a development version; documentation is pending. The package assumes that Python (2.7 / 3.6) and scipy are installed on your machine.

Usage

In the following, we replicate the county-level Spatial Probit analysis of the 1996 presidential elections reported in Section 5.1 of Liesenfeld et al. (2016). The outcome is a dummy indicating whether a county was won by Democratic candidate Bill Clinton.

## Dependencies
library(spdep)
library(Matrix)
library(dimstar)

## Prepare W
coords <- as.matrix(elections.df[,c("lon" ,"lat")])
el_knn <- knearneigh(coords, k = 6, longlat = TRUE)
el_nb <- knn2nb(knn = el_knn, sym = TRUE)
W <- nb2mat(el_nb, style = "W")
W <- Matrix(W, sparse = TRUE)

## Prepare X
X <- model.matrix(clinton ~ log_urban + prop_smcollege + prop_associate + prop_college + prop_gradprof, data = elections.df)

## Instantiate model object 
X.ls <- list(X)
y.ls <- list(elections.df$clinton)

model <- DISTAR$new(X.ls = X.ls, y.ls = y.ls, W_t = W, 
                         N = nrow(X), G = 1, TT = 1,
                         count = FALSE, 
                         spatial_dep = TRUE, 
                         temporal_dep = FALSE, 
                         outcome_dep = FALSE)

## Train
set.seed(0)
ch_vec <- model$train(maxiter = 15, M = 50, abs_tol = 1e-4, verbose = TRUE)

## Standard errors
model$compute_vcov(M = 100)

## Report
model$coeftest()

Note that the estimates are near-identical to the ones reported by Liesenfeld et al. (2016).

hunzikp/dimstar documentation built on May 17, 2019, 6:36 a.m.