F5 Pooling the polls in a model Flashcards
What is the best predictor of public opinion?
The polls of yesterday
What is the basic idea of the state covariance matrix?
Swings between states are correlated. If a candidate is doing better in one state, they are expected to do better in another.
The closer to states resemble each other the more we expect swings in one state to predict swings in the other.
What is the advantage of pooling the polls? Draw it
The sample size artificially increases and uncertainty of estimates decreases.
Outliers mean less in an average.
Three distributions: Poll A, poll B and the pooled estimate.
What is sampling error?
Statistical uncertainty as we know it (dependent on sample size). By chance a sample can be different from the population
What is house effects?
A systematic bias towards GOP or dem. Arise from the secret sauce - polling mode, weighting, LV models etc.
Draw the model/timeline for a poll and possible error
Target population: Coverage error
Sampling frame: Sampling error
Sample: Nonresponse bias
Respondents: Adjustment error
What is the difference between pooling with a simple average and a Bayesian approach?
Pooling: Benefit of sample size but the average is very static because you simply add one poll to a gigantic pool of polls.
Bayesian/Kalman filter: Time plays a really important role. The poll from 20 days ago is there but does not weigh as much as the one from yesterday.
What is the Kalman filter?
A model designed to incorporate new but uncertain information (polls) about the progress of a system (support for Kamala).
It’s a recursive process: You only need t - 1 and t to predict t + 1.
What does the Kalman filter assume?
The support for Harris is the same as yesterday, save for random chocks.
t + 1 is a function of t + noise
What is the transition equation? (related to the Kalman Filter)
The polls act like a stock market. We think of the voters as rational and with full information.
Support for a candidate at any given time should reflect all information available (also that might change in the future).
What can cause support in polls to change?
Only new information that wasn’t forecastable/unforeseeable
Who pushed back of the assumption of full information for voters?
Gelman points to that people get information about fundamentals during a campaign (that’s why we incorporate them).
What are Jackman and Gelmans view on information and voters?
The Jackman model assumes people have full information.
Gelman assumes that full information only comes close to election day
What does Gelman see the fundamentals as in terms of the Kalman Filter?
Our understanding of where the missile is going.
What is the measurement equation? (related to the Kalman Filter)
We expect sampling errors to be independent so that a given poll is distributed around the true estimate.
BUT with house effects it’s true estimate + bias.
