F9 Introducing the Economist model Flashcards
What are the two primary components of our model?
A prediction from strutural fundamentals (NPV)
National and state-level polling
What are three parameters that the model explores using MCMC?
1) Fundamentals and its standard deviation
2) Potential temporal drift of the polls
3) Different types of polling bias
How do we arrive at a posterior distribution in the model?
STAN draw samples from a posterior distribution given the different parameters we set out
What are the six types of bias we estimate the uncertainty for in the model? How describes all these biases?
Measurement error national
Measurement error state level
House effects
Population effects
Mode effects
Adjustment for nonresponse
We assume they are centered on 0 and we determine the uncertainty for this (in pct.)
Heidemanns et al. (2020)
What is measurement error nationally?
How much uncertainty is there in polls nationally? Good old fashioned sample error:
The risk that a perfectly random sample of a given size may not reflect the characteristics of the population as a whole.
What is measurement error on a state level?
How much uncertainty is there in polls on state-level? Typically larger than nationally.
Good old fashioned sample error:
The risk that a perfectly random sample of a given size may not reflect the characteristics of the population as a whole.
What do we assume about state-level measurement error?
That it’s correlated across states.
Maybe not so realistic.
What are house effects?
The secret sauce for each polling house – systematic bias towards GOP or dem.
Linzer (2013) and Gelman assumes the even out when pooling polls.
What are populations effects?
Bias that arise from trying to estimate who is going to vote (likely voter models vs. registered voters).
What are mode effects?
Bias from different ways of collecting data (telephone, text etc.). Different people respond more/less to different modes
How does adjustment for partisan nonresponse work?
Certain pollsters weigh on recalled vote. They are trying to account for how voters say they voted in the last election.
The pollster weights the number of Biden ’20 or Trump ’20 voters to match the outcome of the last election.
Normally considered a mistake but is used to try and account for underestimation i 2016 and 2020 (making sure you have enough Trump voters in your sample - Heidemanns et al. approve).
This mean a shifts in support can reflect a changing sample composition.
Why do we have a baseline correlation in the covariance matrix?
To ensure som kind floor for national similarities due to culture/national media. State are after all part of the same country.
What does TFC predict the NPV to be? What is it when to include CSI?
TFC: 50,85%
TCF + CSI: 51,34%
What is a potential problem with adjusting for partisan nonresponse?
Weighting of polls to 2020 results reduces ability to capture re-alignments.
Assume the election in 2020 reassembles 2024.
Derek and Cohn from NYT are skeptical.
