In soichiroy/BridgeChange: Bayesian Change-point models with Bridge prior for analyzing time-series and panel data.

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  fig.path = "man/figures/README-",
  out.width = "100%"
)

BridgeChange

R package BridgeChange constains functions useful to analyze time-series data and panel data with possibly large number of covariates and change-points. This package offers a Bayesian inference on the linear regression mode under high-dimensinal covariates whose effects on the outcome are allowed to be heterogeneous across time. This package will be useful to discover a set of predictive variables under dynamic setting where time-varying effect is expected to exist in many cases.

Table of Contents

1. Overview
2. Installation
3. Time-series Data

Installation

You can install the most recent version of BridgeChange from Gitub using the devtools package.

# install BridgeChange from Github
# you might need to instal "devtools"
devtools::install_github("soichiroy/BridgeChange")

Time-series Data

Model

BridgeChangeReg() can be used to analyse time-series data with possible change-points.

The functions fits the linear model: $$ y_{t} = \mathbf{X}^{\top}{t}\beta{s_{t}} + \epsilon_{s_{t}} $$ where $s_{t} \in {1, \ldots, M}$ is an indicator of states.

# load packages 
require(BridgeChange)

Example

set.seed(1973);

##
## simulate dataset 
##

out <- BridgeChangeSim(ntime=100, predictor = 50, rho=0.2, constant.p = 0.5,
                       positive.jump=FALSE, varying.p = 0.2, break.point = 0.5, dgp.only=TRUE)

# plot simulated coefficients 
plot(1:length(out$true.beta[1,]), out$true.beta[1,], 
     xlab="predictor", ylab="coefficients", ylim=range(out$true.beta), type='n')
points(out$true.beta[1,], col="red", pch="1", cex=1)
points(out$true.beta[2,], col="blue", pch="2", cex=1)

# plot simulated data 
plot(lm(y ~ x, data = out)$resid, type = 'l', lwd = 1.3)
abline(v = 100/2, col = 'red', lty = 3, lwd = 1.5)

##
## fit models 
##

# set mcmc parameters 
mcmc <- burn <- 500; thin <- 1; verbose <- 500

# fit no-break model
set.seed(11173)
out0 <- BridgeChangeReg(y=out$y, X=out$x, scale.data=TRUE, intercept=TRUE,
                        mcmc=mcmc, burn = burn, thin=thin, verbose=verbose,
                        alpha.MH=TRUE, n.break = 0, waic=TRUE)

# fit one-break model 
set.seed(11173)
out1 <- BridgeChangeReg(y=out$y, X=out$x, scale.data=TRUE,intercept=TRUE,
                        mcmc=mcmc, burn = burn, thin=thin, verbose=verbose,
                        alpha.MH=TRUE, n.break = 1, waic=TRUE)

# fit one-break model 
set.seed(11173)
out2 <- BridgeChangeReg(y=out$y, X=out$x, scale.data=TRUE,intercept=TRUE,
                        mcmc=mcmc, burn = burn, thin=thin, verbose=verbose,
                        alpha.MH=TRUE, n.break = 2, waic=TRUE)

##
## Post-estimation 
##

# model selection by waic 
waic <- WaicCompare(list(out0, out1, out2), print = TRUE)


# state transition plot (provided by `MCMCpack`)
par(mfrow = c(1,2), mar = c(4, 2.5, 3.5, 1))
MCMCpack::plotState(out1)
MCMCpack::plotState(out2)

soichiroy/BridgeChange documentation built on Feb. 14, 2022, 11:49 p.m.