This is a document explaining how to run the bayesian dynamic linear model behind bottenada.se. The Program running Ada is based on R code and is contained in an R-package on github.
Below is an example of installing and using the botten Ada model to do your own predictions of the Swedish election 2014.
As the first step we need to install the package from github to R. Just fire up R and run the following code.
library(devtools)
library(repmis)
library(lubridate)
devtools::install_github(repo = "Ada", username = "MansMeg", subdir = "Package",
dependencies = TRUE)
library(bottenada)We use the polling data from Sweden that is contained in the SwedishPolls github repo. We download the data as polls the following way:
data_url <- "https://github.com/MansMeg/SwedishPolls/raw/master/Data/Polls.csv"
polls <- repmis::source_data(data_url, sep = ",", dec = ".", header = TRUE)Downloading data from: https://github.com/MansMeg/SwedishPolls/raw/master/Data/Polls.csv
SHA-1 hash of the downloaded data file is:
fbfb151fd9484ccd4fb0ef58f91c5459dfad559cWe first choose which parties to do the analysis for as an character vector.
parties <- c("M", "FP", "C", "KD", "S", "V", "MP", "SD")We then create a new dataset to use for the analysis.
predData <- polls[, c(parties, "n", "house")]
predData[, parties] <- predData[, parties]/100Then we add time points to use in the model and then add the house variable.
predData$time <- ymd(polls$collectPeriodTo)
predData$publDate <- ymd(polls$PublDate)
predData$periodFrom <- ymd(polls$collectPeriodFrom)
predData$periodTo <- ymd(polls$collectPeriodTo)
predData$house <- factor(predData$house)We now remove data with missing values (this is actually not neccessary for the timeperiod chosen below, but can be needed for other time periods).
predData <- predData[!(is.na(predData$periodFrom) | is.na(predData$periodTo)),
]
predData <- predData[!(is.na(predData$house) | is.na(predData$n)), ]
predData <- predData[!(is.na(predData$publDate)), ]
predData <- predData[-which(rowSums(predData[, 1:8], na.rm = TRUE) == 1), ]We now need to choose the time period. Below I choose the same period as is used in bottenada.se.
predData2014 <- predData[predData$time > ymd("2012-11-01") & !is.na(predData$time),
]The next step is to set all the parameters used by the model.
parameters2014 <- list()
# Day of election
parameters2014$electionDay <- ymd("2014-09-14")
# Day of prediction NULL if fitting model with all available data, else ex.
# ymd('2014-06-01')
parameters2014$predictionDay <- NULL
# Time period variable names
parameters2014$periodVarNames <- c("periodFrom", "periodTo")
# Publication variable name (used in plot())
parameters2014$publicationDateVarName <- "publDate"
# Timepoint used if not period is used
parameters2014$timeVarName <- "time"
# Variable name for sample size
parameters2014$sampleSizeVarName <- "n"
# Variable name for polling houses
parameters2014$houseVarName <- "house"
# Variable names for period
parameters2014$periodVarNames <- c("periodFrom", "periodTo")
# Variable name for parties
parameters2014$partyVarNames <- c("M", "FP", "C", "KD", "S", "V", "MP", "SD")
# Election results (if fitting model to earlier years were the results are
# known) ex: parameters$electionResults <- c(1791766, 390804, 420524,
# 333696, 437435, 1827497, 334053, 339610) / 5960408
# names(parameters$electionResults) <- c('M', 'C', 'FP', 'KD', 'MP', 'S',
# 'V', 'SD')
parameters2014$electionResults <- NULL
parameters2014$timeInterval <- 1 # Timescale of data (number of days) Not implemented.The second part of the parameters are the MCMC parameters settings. If rejectElectDay is set to TRUE rejection sampling is used so that only samples that sum to less than one at election day is accepted.
parameters2014$MCMCsamples <- 1000 # MCMC Samples
parameters2014$burnin <- 500 # Burnin samples
parameters2014$thin <- 2 # Thinning
parameters2014$seed <- as.numeric(Sys.Date())
parameters2014$silent <- FALSE # Should the analys be done using without printing messages
parameters2014$rejectElectDay <- TRUE # Reject samples that sum > 1 at election dayIf you want to replicate this run exactly you need to set the seed to 16292 and the predictionday to 2014-08-10.
The next part is to set the prior parameters. Most priors (execpt the prediction on election day) are vague priors in this example, but can of course be more informative and they are similar to the ones used at bottenada.se.
The first step is to set the prior for the covariance matrix for the day-to-day change in party opinion. The prior is defined as the parameters
As an example of a vague prior we set
priorData <- polls[ymd(polls$PublDate) > ymd("2013-01-01") & ymd(polls$PublDate) <
ymd("2013-12-31") & !is.na(polls$PublDate), parameters2014$partyVarNames]/100
priors2014 <- list()
priors2014$nub <- 10
priors2014$Sb <- cov(priorData)The next step is to set the priors for the election. This can be done in any type of way. We used our model based mainly on previous election results from Statistics Sweden and incorporating small party effects. We tried different ways to asses the priors based on the previous elections 2010 and 2006 to put the probability mass higher for smaller parties.
You need to define your prior for the election in two steps. First give your best estimates where you believe the election will end. Then you need to specify this as a covariance matrix. In Ada we specify our priors the following way (as of august 9).
priors2014$m0 <- c(0.227, 0.0665, 0.0645, 0.0595, 0.353, 0.08, 0.08, 0.081)
priors2014$C0 <- diag(c(0.001024, 0.000144, 0.000144, 0.000144, 0.001024, 0.000625,
0.000625, 0.000625))
colnames(priors2014$C0) <- rownames(priors2014$C0) <- names(priors2014$m0) <- c("M",
"FP", "C", "KD", "S", "V", "MP", "SD")The house effects also needs priors. Setting the house effects to NA means that we use this polling agency as "anchor" with the 'design effect' set to 1. In bottenada.se we use Statistics Sweden (SCB) as "anchor".
priors2014$tau2matrix <- matrix(data = c(1, 1), ncol = 2, nrow = length(levels(polls$house)),
dimnames = list(levels(polls$house), c("alpha", "beta")), byrow = TRUE)
priors2014$tau2matrix["SCB", ] <- c(NA, NA)Now we are done with setting the priors and can start the analysis.
To run the MCMC sampler we just run the electPredict function in the bottenada package.
prediction2014 <- electPredict(priors = priors2014, parameters = parameters2014,
data = predData2014)It will take some time, for me it took 48 minutes to run the analysis with the parameters set above.
Once the analysis has been run we can check the results. We should check the convergence of the chains to assess that they have converged.
The resulting object contains the chains as separate matrices/arrays in the prediction2014, see ?electPredict for details.
We can also get some summary statistics of the analysis with summary() and get plots with plot().
plot(prediction2014)summary(prediction2014)## ElectPred class object:
## Election (at 2014-09-14) predicted at .
## A total of 150 polls from 9 houses (starting at 2012-10-03).
## MCMC: 1000 samples with a burnin of 500 and a thinning of 2.
## Total runtime: ~ 48 minutes.
##
##
## Expected results:
## Expected results
## M 0.21583
## FP 0.06388
## C 0.05762
## KD 0.04960
## S 0.32987
## V 0.07431
## MP 0.09408
## SD 0.08131
## Other 0.03349
##
##
## Probability Intervals (95%):
## 2.5% 50% 97.5%
## M 0.172098 0.21581 0.25545
## FP 0.046142 0.06398 0.08175
## C 0.042530 0.05708 0.07358
## KD 0.035040 0.04958 0.06482
## S 0.293939 0.32986 0.36749
## V 0.049856 0.07428 0.09939
## MP 0.066746 0.09430 0.12225
## SD 0.049351 0.08163 0.11461
## other 0.005372 0.03258 0.06498
##
##
## Parliment results probability (with limit set to 0.04):
## 2.5% 50% 97.5%
## M 0.17983 0.22396 0.26690
## FP 0.04766 0.06659 0.08417
## C 0.04378 0.05939 0.07689
## KD 0.00000 0.05130 0.06718
## S 0.30288 0.34284 0.38326
## V 0.05156 0.07728 0.10377
## MP 0.06987 0.09791 0.12710
## SD 0.05122 0.08448 0.11970