Description
Combines repeated surveys from several vendors or panels into one daily latent trend, treating each reading as noisy and possibly biased, and estimates every source's bias alongside the trend.
When several sources measure the same quantity at different times, such as competing pollsters or several survey vendors and panels, their readings disagree because of sampling error, timing and systematic house effects. This model treats each reading as the latent level on its field date plus a constant bias for its source plus sampling error with variance p(1 - p)/n. The latent level follows a daily random walk (or local linear trend) whose innovation variance is estimated. Because levels and house effects are not separately identified from the readings alone, the series is anchored with known benchmarks, such as actual election results at each end, or with a sum-to-zero constraint on biases, which identifies only relative biases. Fitting by Gibbs sampling with forward filtering and backward sampling gives a smoothed daily estimate with credible bands. In 239 Australian polls, one face-to-face pollster overstated Labor by about 2.7 points.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Unmigrated
- Stage
- Post-launch, Strategy
- Data
- Quantitative
- Effort
- Weeks
Steps
- Assemble every reading with its date (the midpoint of fieldwork), source and sample size; treat different modes from one source as separate sources.
- Set each reading's sampling variance from its proportion and sample size, inflating for design effects where known.
- Specify a daily latent level following a random walk (or local linear trend) and a constant bias term per source.
- Anchor the scale with known benchmarks (for example billing data, census figures, election results) or impose a sum-to-zero constraint on source biases.
- Place priors on biases and the innovation standard deviation at plausible magnitudes, fit by MCMC and check mixing, since biases correlate strongly with the level.
- Report the daily level with credible bands and each source's bias with its interval.
Sources
- Bayesian Analysis for the Social Sciences (Jackman, 2009)
Also known as
- Poll aggregation with house effects
- Local-level dynamic linear model for survey series
- Bayesian tracker pooling
Categories
- Analytics & Behaviour
- Statistics & Modelling