Philipp Hunziker July 13, 2017
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.
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.
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)## Loading required package: sp
## Loading required package: Matrix
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)## [1] "Setting up log-determinants..."
## [1] "Done."
## Train
set.seed(0)
ch_vec <- model$train(maxiter = 15, M = 50, abs_tol = 1e-4, verbose = TRUE)## [1] "Iteration 1; Max. parameter change = 6.172"
## [1] "Iteration 2; Max. parameter change = 0.773"
## [1] "Iteration 3; Max. parameter change = 0.738"
## [1] "Iteration 4; Max. parameter change = 0.233"
## [1] "Iteration 5; Max. parameter change = 0.147"
## [1] "Iteration 6; Max. parameter change = 0.611"
## [1] "Iteration 7; Max. parameter change = 0.628"
## [1] "Iteration 8; Max. parameter change = 0.199"
## [1] "Iteration 9; Max. parameter change = 0.966"
## [1] "Iteration 10; Max. parameter change = 0.289"
## [1] "Iteration 11; Max. parameter change = 0.083"
## [1] "Iteration 12; Max. parameter change = 0.339"
## [1] "Iteration 13; Max. parameter change = 0.137"
## [1] "Iteration 14; Max. parameter change = 0.641"
## [1] "Iteration 15; Max. parameter change = 0.045"
## Standard errors
model$compute_vcov(M = 100)
## Report
model$coeftest()##
## z test of coefficients:
##
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) 0.605057 0.120679 5.0138 5.338e-07 ***
## log_urban 4.275470 5.346903 0.7996 0.423933
## prop_smcollege -2.838460 0.442548 -6.4139 1.418e-10 ***
## prop_associate 0.795818 0.845979 0.9407 0.346855
## prop_college -1.721855 0.808532 -2.1296 0.033204 *
## prop_gradprof 4.474180 1.427493 3.1343 0.001723 **
## rho_1 0.643665 0.023906 26.9249 < 2.2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Note that the estimates are near-identical to the ones reported by Liesenfeld et al. (2016).