Add Trend to Polls

Calculates a poll aggregation and stores it in the popPolls object, where it is picked up by plot.popPolls().

Usage

popAddTrend(
  data,
  name = NULL,
  type = "kalman",
  args = list(),
  interpolations = list(),
  houseEffects = FALSE
)

Arguments

data

A popPolls object.

name

Name of the trend. Defaults to a name built from type and the applied interpolations.

type

Name of a trend function (see details) or a function.

args

Arguments passed on to the trend function.

interpolations

Named list of interpolations that should be applied to the trend, see details.

houseEffects

Whether the polls should be corrected by the house effects of their firms (see popHouseEffects()) before the trend is calculated. The correction removes the differences between the firms, not the errors they share, so it does not necessarily bring the trend closer to election results; see popAccuracy().

Value

The popPolls object with the trend added to $trends.

Details

Available trend functions are:

kalman

Kalman filter, arguments: sd = 0.003, the daily standard deviation of the true support on the share scale, smoothing = FALSE and missingSampleSize = "firm". With smoothing = TRUE every estimate takes the later polls into account as well (Rauch-Tung-Striebel smoother). The estimates are calculated for the dates with polls only; together with linearInterpolation the smoothed trend reproduces POLITICO's daily kalmanSmooth trend. Polls are weighted by their sample size; polls without one get the median sample size of their firm (or of all polls), or the number given as missingSampleSize. POLITICO uses 400, so missingSampleSize = 400 reproduces its trends exactly.

kalmanKFAS

Kalman filter based on the KFAS package, arguments: sd = 0.003, smoothing = TRUE, missingSampleSize = "firm".

weightedMeanLastDays

Linearly weighted rolling mean, arguments: days = 30, maxObs = Inf.

ident

Plain mean of all polls published on the same day, no arguments.

Instead of a name, type can be a function. It is called with the popPolls object as argument data and the elements of args, and has to return a data.frame with the columns date, party (the codes of data$parties) and value, plus optionally variance, which is used for the uncertainty bands.

Available interpolations are:

lastInterpolation

Carries the last value forward, no arguments.

linearInterpolation

Linear interpolation, no arguments.

bernoulliConvInterpolation

Binomial smoothing over consecutive trend values, arguments: n = 20, k = 6.

Interpolations keep the variance of a trend, which is interpolated (and smoothed) like the values; between two dates with polls it is therefore an approximation.

Examples

if (FALSE) { # \dontrun{
de = popRead('DE-parliament')
de = popAddTrend(de, name = 'Kalman 0.003', type = 'kalman', args = list(sd = 0.003))
de = popAddTrend(de, name = 'Kalman smoothed', type = 'kalman',
                 args = list(smoothing = TRUE),
                 interpolations = list('linearInterpolation' = list()))
plot(de)

# a custom trend: the median of the polls of the last 14 days
rollingMedian = function(data, days = 14) {
    polls = popLong(data)
    dates = seq(min(polls$date), max(polls$date), by = 'day')
    polls[, .(date = dates,
              value = vapply(dates, function(d) median(value[date > d - days & date <= d]),
                             numeric(1))), by = party]
}
de = popAddTrend(de, type = rollingMedian, args = list(days = 21))
} # }