House Effects

Estimates how much each polling firm deviates from the consensus: the mean difference between its polls and a trend, by party. A positive effect means that the firm sees the party stronger than the other firms.

Usage

popHouseEffects(x, trend = NULL, minPolls = 5, sd = 0.003, iterations = 20)

Arguments

x

A popPolls object.

trend

Name of a trend in x$trends the polls are compared to as they are, instead of estimating trend and effects together.

minPolls

Firms with fewer polls of a party are left out.

sd

Daily standard deviation of the Kalman trend, see popAddTrend().

iterations

Maximum number of iterations.

Value

A data.table with the columns firm, party, polls (number of polls), effect (mean difference to the trend) and se (its standard error).

Details

By default the effects are estimated together with the trend they are measured against: a smoothed Kalman trend is calculated from the polls corrected by the current effects, the effects are updated with the remaining deviations, and so on until they no longer change. Otherwise a firm that publishes a large share of the polls would pull the trend towards itself and its effect would be underestimated. The effects are measured relative to the average of the firms, each counting once however many polls it publishes: an error all firms share, as seen at some elections, cannot be told apart from the trend.

The effects are assumed to be constant over time, so for long series it can be worth restricting the polls to the last years first.

Examples

if (FALSE) { # \dontrun{
de = popRead('DE-parliament')
de$polls = de$polls[date >= as.Date('2022-01-01')]
effects = popHouseEffects(de)
# one row per firm, one column per party
data.table::dcast(effects, firm ~ party, value.var = 'effect')

# a trend of polls corrected by the house effects
de = popAddTrend(de, type = 'kalman', houseEffects = TRUE)
} # }