Exam Flashcards
Componets of the model and what they do?
The prediction model combines two parts:
1. structural fundamental forecast of national popular vote (NPV) (prior)
2. state-level polling forecast (model describing the distribution of data - Mathematical formulas that characterize the trends and spreads in the data which has different parameter values control knobs)
› model of differences between states (NPV –> SPV for each state) –> prior differences
› When we go from the prior (NPV) to the states vote share we use state relative positions (based on previous elections) to see the difference between each state so that all states don’t have TFC as baseline
› national and state-level polls
› models for sampling and non-sampling error in polls
› model for state and national opinion changes during campaign
–> This is our correlations matrix. Here we use sociodemographics/how similar the states are to share information from state-level polls
Three parameters that the model explores (‘samples’) using MCMC method
› fundamentals ( It’s standard deviation – to reflect uncertainty in the national popular vote (NPV) forecast.)
› potential temporal drift of the polls ( different values for the rate and direction of temporal changes in public opinion during the campaign period.)
› different types of polling bias ( polling errors, such as: Sampling Error: Variability due to the limited sample size of polls. Non-Sampling Error: Systematic biases introduced by polling methodologies (e.g., house effects)
› outcome -> exploration of the posterior distribution
› After many iterations, the model converges on a posterior distribution that reflects the most likely parameter values (even very unlikely ones are in but not as often as more likely scenarios) and provides probabilities for outcomes like the state vote shares and Electoral College results.
› After many iterations => converging towards the ‘actual’ distribution
The theoretical assumptions behind our bayesian model (Structural + polls)
The need for structural fundamentals: Economic voting
- polls can be problematic many months in advance - capture the moment
- Swing voters converge towards ‘true’ opinions closer to election (economic voting of swing voters) based upon the state of the country. structural fundamentals uses the state of the country as a predictor
- Deliver some information that aren’t in polls (Graefe 2018)
To include polling data and make it dynamic: The need for the Michigan Model
- Stupid not to use the available information we have
- fundamentals is a national and lack of data make it unreliable - Lauderdale & Linzer 2015. Grafe 2018 shows that it is the most unpredictable from 2004 to 2016.
- Make a state covariance, so we can share information from polls to similar states. Here we utilize the insights from the Michigan model and assume that similar states (based on similar sociodemographics) vote similar and move similar ways
How does the Bayesian approach handle uncertainty regarding polling?
We include models (that desribes the data) for sampling and non-sampling error in polls. During the simulation
it randomly jumps to different values of this (parameter values) and they are accepted in regards to how probable they are based on the prior and data –> so how well do they fit with the data with the prior in mind.
By running thousands of simulations, the model generates a posterior distribution that accounts for the influence of polling errors on the forecast.
The probability of a candidate’s victory is calculated as the fraction of simulations they win, reflecting both likely and improbable scenarios (e.g., winning the Electoral College but losing the popular vote). This approach ensures comprehensive modeling of election uncertainties.
We have also set a high prior in the beginning when the polls are more noisier because of voters intentions becomes clearer in the end
Newer polls are weighted more than early polls
What kind of election? (AP votecast and Exit polls)
We saw almost a nationwide shift towards trump –> could potentially be a realignment (not uniform) - otherwise balancing
- especially looking at education
He won in all groupings - for example he lost with less among urban citizens than in 2020
significant inroads in Latinos voters - Improved on younger voters too
However education may have happened in 2016 - Trump as an outlier?
What is wrong with the fundamentals model? Lauderdale & Linzer 2015
Three sources of uncertainty
1) Fundamentals models often fail to account for the full uncertainty in their predictions, such as the variability in coefficient estimates - the uncertainty of the point prediction. This leads to overly narrow confidence intervals - which actually should be around 7-10 points –> not very much information in close elections
2) There is no consensus on which variables (e.g., GDP growth, unemployment) are most predictive –> Different specifications can yield significantly different predictions, making results less reliable
- This is because of the limited evidence of elections
3) Electoral College Dynamics: Fundamentals models typically focus on the national popular vote, overlooking the complexities of the Electoral College system, which aggregates state-level outcomes. This can result in incorrect predictions, particularly in close elections. –> for example you can win the popular vote but lose the electoral college
Data limitations: The models rely on a limited number of past elections (often fewer than 20), making it difficult to draw strong conclusions about patterns and relationships. –> limited strength of the evidence.
There is a significant change in political context regarding 1948 to 2024
How much can the data and trends from 1948 be used now - different context. Argument for taking term out for example
Why is simulation used in Bayesian analysis?
To handle cases where analytical solutions are not possible, like when multiple parameters are involved.
What is the main points from the Bitter end?
Calcification –> is due to polarization
o Calcification means less willingness to defect from their party, such as by breaking with their party’s president or even voting for the opposite party.
o There is thus less chance for new and even dramatic events to change people’s choices at the ballot box.
o –> This means smaller fluctuations from year to year in election outcomes.
o However not the same winner because of so close in sheer numbers
o 2016-2020 calcified very small shifts
Not polarization as such –> polarization is a larger gap between the two parties on different areas. Affective polarization
Three elements in the Trump presidency, 2020 election and aftermath upshot was more calcified politicas
o (1) Long-term tectonic shifts have pushed the parties apart while making the views within each party more uniform –> gradually increasing partisan polarization.
o (2) Shorter-term shocks, catalyzed especially by Trump, have sped up polarization on identity issues
o Race, gender, ethnicity etc.
o (3) It is precisely these identity issues that voters in both parties care more about— exacerbating divisions even further and giving politicians every incentive to continue to play to them.
o They voted based on these divided issues
Party polarization explains changes in the candidates’ coalitions. One consequence should be a stronger association between people’s own ideological predispositions and their voting behavior. not so sociodemographic?
o This increased polarization, as Trump’s performance among conservatives and Biden’s performance among liberals made Americans’ ideological identification a stronger predictor of how they voted in 2020
o The ideological polarization in voting behavior was more likely to come from a third source: voters changed their issue positions in ways that aligned with their partisanship.
Partisan perceptions of the economy and its impact on structural fundamentals? (Brady et al. 2022)
Partisans may exhibit motivated reasoning, attributing a good economy to their preferred party and a bad economy to the opposition.
Consensus: Incumbent Partisans are comparatively more optimistic about the economy when their party controls the executive branch than they are otherwise
the gap in economic perceptions approximately doubled between 1999 and 2020, and that partisan economic perceptions no longer seem to converge during economic crises.
Attribution for good/bad economy influenced by partisanship –> Perceptual bias
- also who they give the credit to
Argument for TFC and the objective variables doesn’t play a role
Problems with survey-based opinion (Linzer 2013)
Not every state is polled on every day, leading to gaps in the time series. Data are especially sparse in less-competitive states and early in the campaign.
Second, measured preferences fluctuate greatly from poll to poll, due to sampling variability and other sources of error.
We can eradicate these problems by pooling the polls (Jackman 2005)
Maybe Erickson & Wlezien 2008/2014
Synopsis questions: Why don’t we just use polls or the TFC model?
Not use the TFC only –> but…assumes states move in same way from election to election… –> how about sociodemographic changes in the state and state-specific issues?
Polls –> Erickson & Wlezien
TFC –> uncertain data and becomes more accurate with polls - static (Linzer 2013)
What are the five types of elections? (The american voter revisited)
- Maintaining: 1964 (LBJ)…
- A maintaining election is one in which stable partisan attachments continue to be a major determinant of election re- sults. - Deviating: 1952 (Eisenhower)…
- A deviating election would then be one in which the short-term partisan attitudes lead to the election of the presidential nominee of the minority party, without a fundamental shift in the party identification bal- ance in the nation. - Re-instating: 1960 (JFK), 1980 (Carter)
- Re-alignment: 1930s (FDR), 1980s (Reagan - wwc, gender, white evangelicals)
- The realigning election is one in which the partisanship of people changes. Systematic change occurs when issues motivate social groups to move to one party and when change is reinforced through one’s social groups.
Definition: Significant shift in the sociodemographics group –> can lead to balancing or dominance - Balancing: 2016, 2020
- One in which neither party has a majority in party identification.
- The partisan balance is so close that either party could win.
- This is not an election in which short-term forces work against the majority party, but one in which it is not possible to speak of a majority party.
The real story of the last 50-plus years of American presidential elections is a weakening of the Democratic lead in party identification to the extent that elections are very close and can be swung by any number of short-term matters.
The Linzer State-level model? (Linzer 2013)
Combines fundamentals and state-level polls.
› Uses a sequence of state-level preelection polls to estimate both current voter preferences and forecasts of the election outcome, for every state on every day of the campaign, regardless of whether a survey was conducted on that day. –> share information between states (simple model)
› Forecasts from the model gradually transition from being based upon historical factors early in the campaign to survey data closer to the election Gives more weight to the data in the end as it has more information and seems more credible
› In states where polling is infrequent, the model borrows strength hierarchically across both states and time, to estimate smoothed within-state trends in opinion between consecutive surveys.
› Possible because The temporal patterns in state-level opinion are often similar across states. And you can look at the trend in a correlated state –> Similar states vote similkar –> sociodemographics argument
› Very simple –> we use a correlation matrix
› The model also filters away day-to-day variation in the polls due to sampling error and national campaign effects, which enables daily tracking of voter preferences toward the presidential candidates at the state and national levels.
SO: we get a proportion of voters in state i on day j who say vote D based upon the fundamental model and the polls. The polls are used to show the proportion of voters in state i on day j who say vote D – when there isn’t polls we use state-level effect – LR dynamics of voter preferences in state I (here the prior is included) – and national effects – borrow trends from other states (campaign effects). To capture the vote share on election day we incorporate the fundamentals model (normal prior distribution) – set a weight for the prior which determines how sensitive to polling data. We use a Bayesian reverse random walk from election day and back to the last day with polls which have become the prior distribution
What factors influence the Perry-Gallup likely voter index?
Seven questions that include voter’s past voting behavior, level of interest in the election, and self-reported likelihood of voting. - then assign points based on the answers
Deterministic likely voter model
Why do we pool the Polls? (Jackman 2005)
To create a more precise estimate of the vote intentions. Individual polls are subject to a lot of noise regarding house effects and bias especially.The chief benefit of pooling poll results (after correcting for house effects) is that we are much better positioned to ascertain movements in levels of voter support in response to campaign events
Also the sampling error and therefor uncertainty regarding the point estimate lowers when pooling the polls.
–> the sample sizes are simply too small to reliably detect small fluctuations in support for the parties over the course of an election cam- paign.
Combine the information in the published polls, leveraging them against one another, so as to obtain a clearer picture of what might be going on in the electorate over the campaign
Individual polls are snapshots in time (rolling average – each poll not independent)
Dependent on the poll from the day before
Different ways of pooling:
* simple averages
§ Benefit of sample size
§ The challenge of the average will be very static –> one poll doesn’t change much
§ Because the information means the same
* Bayesian pooling -> prior + new information (new polls)
§ Much more fastmoving
§
§ Gives weight to the polls to capture a better estimate
§ Time in the bayesian model means a lot
* Detect movement in voter support due to campaign events
USES Bayesian pooling -> prior + new information (new polls)
- Gives weight to the polls to capture a better estimate
- Detect movement in voter support due to campaign events
o Prediction
- A recursive process
- Only the estimated state from the previous time step and the current measurement are needed to compute the estimate for the current state
- Continuously try to predict the next point in the time series:
o Prediction is made
o Once the result is measured (i.e., new polling data comes in, or a new missile location is reported), we calculate how far off prediction was
o Update prediction of new latent state, make new prediction for next observation
- Updates it self based on the data
We can’t be sure of the house effects - can account them for last time however the firms will work to account for this –> Linzer does this. Otherwise we can throw it in to our bayesian model with a prior
However lack polling data in some states
Bad things with only using Fundamentals? (Linzer 2013)
They are subject to a large amount of uncertainty, as Most historical forecasts are based on data from just 10 to 15 past elections, and many only generate national- level estimates of candidates’ voteshares.
Moreover, in the event that an early forecast is in error, structural models contain no mechanism for updating predictions once new information becomes available closer to Election Day.
Preelection polls provide contextual information that can be used to correct potential errors in historical forecasts, increasing both their accuracy and their precision.
Polls conducted just before an election generate estimates that are very close to the eventual result, on average
What is the Time for Change Dynamic
It is based on the hypothesis that
voters attach a positive value to periodic alter-
nation in power by the two major parties and
that regardless of the state of the economy and
the popularity of the current president, when a
party has held the White House for two or
more terms, voters will be more likely to feel
that it is time to give the opposing party an
opportunity to govern than when a party has
held the White House for only one term.
What are the two main types of polling, and how do they differ in methodology? (Bailey 2024)
- Probability-based Polling (Random Sampling):
Uses a sampling frame to identify a pool of respondents representative of the population. –> for example land lines or adress based sampling
Relies on interview modes like live calling or interactive voice response (IVR) to collect data. –> can still be biased based on non-response bias - Non-probability-based Polling:
Does not rely on random sampling. Instead, respondents are recruited from convience, pre-existing panels or quota (mathing) sampling methods.
Often used in internet polling, which may introduce bias due to the lack of randomness in the sample selection.
Nonprobability samples face “one fundamental problem: There is no comprehensive sampling frame for the internet, no way to draw a national sample for which virtually everyone has a chance of being selected”. In the random sampling paradigm, the contact list is a random sample which will include people willing and unwilling to respond. For internet panels, the people contacted have already said they will respond to polls.
PROBABILITY IS THE BEST
But internet isn’t doing that bad right now (Silver 2021)
Why Forecast? - benefits of forecasting (Gelman et al. 2020)
- First, the popularity of forecasts reflects revealed demand for such information.
- Second, by collecting and organizing relevant information, a forecast can help people and organizations make better decisions about their political and economic resources.
-If people are going to follow election night results online — and they do, by the millions — they ought to have the context to understand them (Cohn & Katz 2018) - Third, the process of building — and evaluating — forecasts can allow scholars and political observers to better understand voters and their electoral preferences, which can help us understand and interpret the results of elections.
Derek slides:
1. There is demand
2. Can help to organize political and economic resources
4. Can help advance understanding of voters and their electoral preferences
What does Erickson & Wlezien (2008 and 2014) show about economic performance as a predictor?
Why does we use structural fundamentals?
They show the economic voting of swing-voters –> The economic perceptions are important for the vote
Voters converge towards ‘true’ opinions closer to election (economic voting of swing voters)
- here they use economic indicators to guide their vote when they couldn’t care less in April. This is because the campaign makes it salient
In April (200 days before the election), perceived business conditions only moderately predict polling results.
By November (Election Day), there is a much stronger correlation between perceived economic conditions and the vote share for the incumbent party, showing how economic perceptions shape the final vote.
What is a Probabilistic Likely Voter model vs. a deterministic one? And what is Rentch et al. 2020’s insight?
A probabilistic model predicts each respondent’s probability of voting rather than assigning a strict cutoff. This approach is more flexible and accounts for uncertainty, often using logistic regression.
Deterministic has a cutoff
Probabilistic models, offer clear benefits. - Each respondent is assigned an estimated probability that they will vote. This probability is then used as a weight: Responses from those who are more likely to vote are weighted more heavily than responses from those who are unlikely to vote, but all are included in the election prediction.
Probabilistic approaches serve as a compromise between registered-voter and likely-voter methods; the preferences of all respondents are utilized only to an extent proportional to their assessed probability of actually voting.
Rentch et al. 2019 –> Propose a probabilistic likely-voter model that not only uses the typical items (such as those from the Perry-Gallup index), but also adds demographic information about respondents such as age, education, race, income, gender, and strength of partisanship. These demographics are correlated with turnout
Why use a Sociodemographic probabalistic voter model?
- a likely-voter model that is probabilistic uses information from all respondents in the sample rather than discarding those that fail to meet a particular threshold.
- And a likely-voter model that makes use of demographic information for its predictions takes advantage of data that most pollsters collect anyway and which happen to be good predictors of turnout and overreporting.
Key steps/core idea of Bayesian analysis
Key steps:
Identify data – For us this is polling
Define a descriptive model for the data - model for polling bias and state and national.
- This includes choosing the appropriate distribution (e.g., normal distribution) and defining its parameters (e.g., mean and standard deviation).
Assign a prior probability distribution to the parameters of your model. This represents your initial beliefs or expectations before seeing the data.
–> for example Our Fundamentals model
Use Bayesian inference to re-allocate credibility across parameter values.
o Combine your prior beliefs with the observed data to calculate the posterior distribution (updated belief)
o It is essensially a compromise
When new data arrives (like a poll), you evaluate how consistent the data is with your prior belief
Based on how strong your prior is the posterior will be a compromise
Downside to forecasting? (Victor 2021)
- Partisan Polarization Perverts the Fundamentals –> Partisan polarization introduces systematic bias into models since partisans view fundamentals (like economic status or presidential approval) differently depending on their political affiliation. This can reduce the reliability of forecasts.
- Forecast may affect turnout - greater emphasis on forecasts in 2016 cycle increased voters’ certainty about the election outcome and therefore depressed turnout
- false sense of certainty –> overstate certainty or fail to report
- misguides people
Make better visualisations –> for example the needle (The article in ny times)
What is probability-based polling vs. Non-probability based polling? (Bailey 2024)
Probability based sampling – previously the golden standard –> Random drawn
- Sample frame –> How to identify the pool of respondents that are representative. Before: random digit dialing. Some probability-based pollsters now use address-based sampling (ABS) or registration-based sampling (RBS)Can still be biased if only old people have landlines
- Interview mode: Live-calling respondents, Interactive voice response, web based address sampling and FTF (Anes e.g.)
Both non contact and nonresponse
Non-probability-based sampling – Opt-in –> unrepresentative NON-random
- Most nonprobabilistic polls are conducted via the internet or text messages –> Based on opt-in behaviour
- Convience, internet panels and Quota sampling (matching) –> risk of professionals
- No response rate - Unrepresentative
Regarding representation and nonresponse–> they use weights. Both have problem with nonignorable nonresponse
- weighting does not solve – and could potentially exacerbate – nonignorable nonresponse bias.
What is important by the 538’s model?
Correlation matrix, fundamentals and what do they sample using the MCMC
Polling data :
- Allow movement to be correlated between states in addition to between a state and the nation as a whole. lets polls in one state influence similar states – so the trend and Movement in one state in one poll is shared to other states, where the influence is based on similarity
Three factors of similarity: The first is how similarly they have voted in presidential elections since 1948 (we only do 1998-2020), geographic similarity (10 political regions) and demographic similarity between states
The fundamentals model:
- Economic and political fundamentals
11 economic variables and primarily the TFC model with where the model decreases the effect of presidential approval and economic growth in more polarized elections. State factors are also in regarding vote and home-state of candidates.
Three parameters that the model explores (‘samples’) using MCMC method
› fundamentals ( It’s standard deviation – to reflect uncertainty in the national popular vote (NPV) forecast.)
› potential temporal drift of the polls ( different values for the rate and direction of temporal changes in public opinion during the campaign period how much polling averages change)
› different types of polling bias ( polling errors, such as: Sampling Error: Variability due to the limited sample size of polls. Non-Sampling Error: Systematic biases introduced by polling methodologies (e.g., house effects)
› outcome -> exploration of the posterior distribution
› After many iterations, the model converges on a posterior distribution that reflects the most likely parameter values (even very unlikely ones are in but not as often as more likely scenarios) and provides probabilities for outcomes like the state vote shares and Electoral College results.
What is the key concept of Bayes’ Rule?
It updates prior beliefs using observed data to calculate a posterior probability.
MSE and evaluation?
focus on predicted vote share by state
Average of the squared difference between the observed and predicted values for each state
Lower value: better fit
Squared: always positive, sensitive to outliers
How and why do we use “Markov Chain Monte Carlo” (MCMC) simulation?
MCMC allows for the estimation of posterior distributions that are too complex for analytical solutions - we have a lot of parameter values.
Therefore we use MCMC to explore the parameter space efficiently and account for uncertainty in the predictions. The model samples different combinations of parameter values (e.g., national vote share and polling errors) to generate a range of possible election outcomes.
The cleverness of the method is that representative parameter values can be randomly sampled from complicated posterior distributions by using only simple proposal distributions.
All we need to decide whether to accept a proposed parameter value is the mathematical formulas for the likelihood function and prior distribution
Start by sampling Parameter Values
Each parameter represents an important aspect of the forecast, such as:
Polling bias (e.g., house effects or demographic underrepresentation).
Temporal drift (e.g., changes in public opinion over time).
Fundamentals’ variability (e.g., standard deviation in structural predictions).
The Role of MCMC:
The model starts at an initial guess for these parameter values.
It takes a “random walk” through the parameter space, proposing new sets of values at each step.
The likelihood (fit with observed polling data) and prior (historical expectations) determine whether a proposed set of parameter values is accepted or rejected. –> Can see that we don’t have any observations of landslide victories –> impossible according to our fundamentals and the observed data
This gives us a posterior distribution. It reflects:
The most plausible combinations of parameter values after considering prior knowledge and polling data.
Each combination of parameter values corresponds to a potential election outcome.
Parameter Values and Posteriors: The posterior reflects the most likely combinations of parameter values (e.g., polling bias, vote share) given the data.
Multiple Parameter Values Can Yield the Same Outcome –> A high polling bias (overestimating support for one candidate) might balance out with strong fundamentals favoring another, leading to the same predicted election result.
What is the difference between “vote intentions” and “official vote count” as described by Gelman et al. (2020)?
Vote intentions: The total number of votes each candidate would receive if everyone who intended to vote did so and all votes were counted.
- what we meause with polls
Official vote count: The actual count of votes after some people decide not to vote or face barriers to voting.
How did polarization affect the 2020 election?
- The closeness of the election stemmed from the ideological and affective polarization between the two parties that has helped create an increasing calcification of people’s political choices.
○ Calcification
○ Between 2016 and 2020 there were only small shifts in outcomes in various states and counties as well as among individual voters.
Polarization was visible in how the public perceived the candidates:
○ Trump was perceived as more conservative than in 2016, and Biden was perceived as more liberal than Hillary Clinton.
Voting behavior was also more polarized:
○ Trump gained votes among conservatives and Biden gained votes among liberals.
○ Explains Trumps stronger performance among latinos –> Trump’s gains among moderate and conservative Latinos exceeded his losses among liberal Latinos.
–> The ideological polarization in voting behavior was more likely to come from that voters changed their issue positions in ways that aligned with their partisanship.
A final consequence of calcification is that the election year’s impor- tant events did not appear to affect the outcome very much. –> Local voting patterns did not depend much on Biden’s advantage in televised campaign advertising.
There was also little local impact of the rate of COVID infections or deaths in communities
For example Strong partisanship helped stabilize Trump’s approval ratings, driven both by the resolute antipathy of Democrats and the continued support of Republicans
What is the concept of ‘Calcified Politics’ as introduced in ‘The Bitter End’?
Calcified politics refers to a system where political opinions and party alignment become more rigid over time. It results in:
Less likelihood for voters to switch party allegiance.
Smaller year-to-year fluctuations in election outcomes.
Increased partisan parity, where Democrats and Republicans are numerically closer in support.
What are the key components of the “Time for Change” model by Alan Abramowitz?
The “Time for Change” model predicts presidential elections using three fundamental features:
- The growth rate of the economy during the second quarter of the election year –> Predictor of how they see the economy in November
-(economic voting)
2 The incumbent president’s approval rating at mid-year (June approval rating)
(captures performance on other issues)
- The length of time the incumbent president’s party has controlled the White House—the time-for-change factor.
(assumes that voters attach a positive value to periodic alternation in power by the two major parties - will feel more positive of giving the opposing party a chance after two terms) - polarization has probably made this less accurate
The assumption underlying the time-for-change model is that a presidential election is a referendum on the performance of the incumbent president.
Some of the correlation between presidential approval and presidential vote is a byproduct of party identification. However, even after controlling for party identification, there is a strong relationship between approval and vote choice. ‘
How many electors are there today, and how are they allocated?
There are 538 electors, based on:
100 Senators.
435 Representatives.
3 electors for Washington, D.C. (added by the 23rd Amendment).
Elector allocation is adjusted every 10 years following the Census.
How do long-term and short-term factors interact in the Michigan Model?
Long-term factors, such as party identification, provide a stable foundation for voter choice. Short-term factors, like campaign events and candidate qualities, interact with this foundation, but their influence is relatively limited.
For instance, a charismatic candidate might sway some voters, but the majority will vote according to their party loyalty
2016 and 2020 polling errors?
2016 - under-representation of R voters in polls (+3% D)
- lack of lesser educated in samples -> weighting by education in 2018, 2020
late-deciding voters broke for R
2020 - under-representation of R voters in polls (+4.3% D) (COVID!)
- highest state-level error where: more white, more rural, less populated states
- not due to errors in sampling frames due to polling mode (coverage) = it was the right population
DUE to non-response bias of R voters + R who responded more likely to support Biden! –> Nonignorbale responsebias
What is Rentch et al. 2020’s upgrade of LV-models?
Propose a probabilistic likely-voter model that not only uses the typical items (such as those from the Perry-Gallup index), but also adds demographic information about respondents such as age, education, race, income, gender, and strength of partisanship.
Since these variables are consistently related to turnout and may provide valuable context for responses to conventional indicators, they should help produce more precise likely-voter probabilities
What is the Michigan Model and explain the four key factors in the Michigan Model for voting behavior and the main argument ? (The American Voter revisited)
Longerterm factors detertime the shorter-factors. It is a sociological approach –> focusing on socialisation, how voters gets locked in. They get locked in based on their sociodemographics and partisan identity - short term factors matter based on their partisan identification
o Many doesn’t consider voting for other parties
The Michigan Model emphasizes the long-term, stable factors that shape voter behavior. It highlights the importance of socio-demographic characteristics, party identification, and group-based loyalties. This model suggests that voters are often “locked in” by long-term party affiliation, and while short-term events like campaigns and candidate qualities can have an effect, they are less influential
- Long term: Social demographic factors and predispositions as party identification and political ideology.
- Short term: The candidates competing in a campaign and the issues raised in it.
- Sociodemographics: Social characteristics like race, religion, education, and income that shape party identification.
- Party Identification: A stable psychological attachment to a political party, often formed through socialization.
- Party identification = (relatively) stable identity over time at level of individual voter –> occurs through socialization… (parents…) - Issue Stances: Voters’ evaluations of political issues and candidate positions.
- Campaign Effects: The influence of candidate qualities, campaign events, and elite cueing in the short term.
Long-term factors like sociodemographics and party identification determine the baseline, while short-term factors like campaigns play a smaller but significant role.
Why do we need sociodemographics?
We use sociodemographics to move to the state-level predictions
In our model -.-> polling data from one state informs us about opinion in other states
* We do this through a covariance matrix of states, using sociodemographic factors
o Using sociodemographics to see which states look alike and can predict each other –> which ones that goes in the same direction
Why important:
1. states are very different sociodemographically
2. states do not move in same direction!
Based upon an assumption that similar states regarding sociodemographics should move similar directions
- We want to say to our model that these two states doesn’t look alike because of this socio-demographic
How group membership can matter? (Lewis-Beck et al. 2011)
- psychological identification –> personal experiences, social integration
- ‘linked fate’ -> shared experiences
- membership (unions…)
Why do we need state-level models? (Lock & Gelman 2010)
Structural fundamentals (TfC or others (Higgs, 2008))
can estimate outcome through NPV – SPV for each state (hi) * ev of state
- Separate the national forecast (based on fundamentals) from the relative positions of the states (for which past elections and current polls can be combined to make inferences). Preelection polls, not surprisingly, are more reliable as they get closer to the election.
but…assumes states move in same way from election to election… –> how about sociodemographic changes in the state and state-specific issues?
and why not also use polling data (national and state)?
- Preelection polls provide contextual information that can be used to correct potential errors in historical forecasts, increasing both their accuracy and their precision.
Challenge = how to combine when TfC is election day prediction + polls from campaign?
moving to state-level predictions
simplest model : state partisan index : NPV – SPV
in our model -> polling data from one state informs us about opinion in other states
we do this through a covariance matrix of states, using sociodemographic factors
why important:
1. states are very different sociodemographically
2. states do not move in same direction!
How does the Michigan Model differ from proximity models like the Median Voter Theorem? (The American Voter revisited)
The Michigan Model posits that party identification is a much stronger determinant of voting behavior. Main argument: Highlights the importance of long-term factors in shaping ones voting behavior –> Fordi partiidentifikation påvirker ens syn på kandidater og holdninger til enkelte problemstillinger, som afgør ens partivalg.
- The long-term factors can be seen as providing a baseline from which short-term, campaign-specific factors operate.
- Long term: Social demographic factors and predispositions as party identification and political ideology.
- Short term: The candidates competing in a campaign and the issues raised in it.
> < disagrees with proximity models (median voter theory) that ascribe a high level of rationality and effort on the part of the average voter economic voting gives more influence to short term factors
Polarization and Partisan identity (The bitter end)
Party identification = (relatively) stable identity over time at level of individual voter
- stronger and stronger party identification –> more polarized on identity issues and the most important issues is also where people are divided. Elite cue –> people aligned their issue-positions with their partisan identity
Polarization: Polarization refers to the growing ideological distance and division between political parties and their supporters. It leads to stronger alignment with party ideologies and increased hostility toward opposing parties ( affective polarization)
Calcification is the rigidity in political behavior, where voters and politicians rarely deviate from party lines, limiting the impact of even significant political events on voting behavior.
What does it mean that the us politics is calcified? (The bitter end)
Calcification produces hardening and rigidity –> people are more firmly in place and harder to move away from their predispositions.
○ Not the same as polarization:
§ Calcification means less willingness to defect from their party, such as by breaking with their party’s president or even voting for the opposite party. - Polarization means more distance between voters in opposing parties in terms of their values, ideas, and views on policy.
There is thus less chance for new and even dramatic events to change people’s choices at the ballot box. –> the election year’s important events did not appear to affect the outcome very much.
○ Local voting patterns did not depend much on Biden’s advantage in televised campaign advertising.
○ There was also little local impact of the rate of COVID infections or deaths in communities
§ --> This means smaller fluctuations from year to year in election outcomes.
For example –> this explains why Trumps approval rating was the same. The same about the economy
Balanced election in 2020?
Why are win probabilities often misunderstood in election forecasts? (Gelman et al. 2020)
A vote share of 60% implies a decisive win, but a win probability of 60% implies a relatively close race. This confusion stems from the difference in how people interpret probabilities versus vote shares. Reporting win probabilities as exact figures (like 87.4%) can mislead people into overestimating the certainty of the prediction.
What is the difference between subjective and objective evaluations in economic voting?
Objective evaluations are based on measurable economic indicators like GDP, unemployment rates, and inflation, while subjective evaluations are based on how voters “feel” about the economy.
The subjective measurement is important because Attribution for good/bad economy influenced by partisanship –> Perceptual bias
* Brady et al. (2022)–> there is a substantial gap between the proportion of Democrats and the proportion of Republicans that believe the economy is improving. There is also a polarization in this gap from 1999 to 2020
Sum up:
Partisanship influences how voters interpret economic conditions. Partisan voters are more likely to perceive the economy as doing well when their preferred party is in power, even if objective indicators suggest otherwise. This effect is driven by motivated reasoning, where individuals selectively interpret information to reinforce their existing beliefs
Components of the our/economist model
The prediction model combines two parts:
1. structural fundamental forecast of national popular vote (NPV) (prior) –> Static prediction
- Based upon economic voting theory as the voters hold the president accountable for the economy etc. –> People converge towards ‘true’ opinions closer to election (economic voting of swing voters) as predicted by structural fundamentals (Erickson & Wlezien 2008)
- Structural fundamentals (TfC or others (Higgs, 2008))
can estimate outcome through NPV – SPV for each state (hi) * ev of state but…assumes states move in same way from election to election…
and why not also use polling data (national and state)?
= how to combine when TfC is election day prediction + polls from campaign? –> The Michigan model
- state-level polling forecast (i.e. model describing the distribution of data)
2.1 model of differences between states (NPV –> SPV for each state) –>
–> The prior differences of each state in voting behaviour
2.2 national and state-level polls
2.3 sampling and non-sampling error in polls
2.4 model for state and national opinion changes during campaign
Here we utilize the insights from the Michigan model and assume that similar states (based on similar sociodemographics) vote similar and move similar ways
How does the Michigan model view voter behavior in comparison to the proximity model? (The American Voter revisited)
It argues that most voters are “locked in” due to long-term sociological factors, unlike the proximity model which assumes rational evaluation of candidates.
What is ‘economic voting’ and how does it influence elections? (Lewis-Beck et al. 2011)
Economic voting is a proximity model contra Michigan model. It factors in the short-term factors and that the voters are rational.
It refers to the idea that voters base their electoral decisions on the current state of the economy. Voters tend to reward incumbents when the economy is performing well and punish them when it is performing poorly. This concept emphasizes voter accountability, where the president or governing party is held responsible for economic outcomes, even though the role of Congress or external factors might also influence economic conditions
It can be retrospective (past economic performance) or prospective (future expectations). Economic voting is influenced by both personal (egocentric) and national (sociotropic) perspectives.
Source: The American Voter Revisited, Chapter 13
How does the Economist model 2020 look like? (Heidemanns et al. 2020) –> descriptive way of our model for example the models
Their model provides an election day forecast by partially pooling two separate predictions:
* (1) a forecast based on historically relevant economic and political factors such as personal income growth, presidential approval, and incumbency; and
* (2) information from state and national polls during the election season.
Use TFC including an term of economic growth and swing voters as their prior for NPV
How polls are coming in: State-level polling
* Include polls at the national and state level and take each poll to be an estimate of that day’s average support for the Democratic and Republican candidates for president with modeled bias and variance. (national polls are summed out to the 51 states depending on size)
* Model for state and national opinion changes throughout the campaign –> Share information across states contemporaneously and over time.
Treat the development of state-level public opinion as a correlated random walk for which we have prior information from the fundamentals-based prediction at t = T (Election Day).
* The correlation in the random walk imposes our assumption on the estimates that in the absence of data similar states will move in similar ways, i.e., if we have polls for Washington but not for Oregon, then the daily trend for Oregon will look similar to the trend in Washington with added uncertainty.
models for sampling and non-sampling error in polls
* Include adjustment terms for pollsters, polled population (e.g. likely voters and registered voters), and polling mode (e.g. live caller or online). –> Our priors for the parameters are centered at 0 with different standard deviations –> are estimating the deviation given this year’s polling data, thus not using potential information about the quality of the pollsters or the reliability of likely voter adjustments from past data, out of concern that these may not provide reliable indications for the current election.
* Partisan nonresponse adjustment –> If a pollster does not adjust for the partisan composition of their sample, shifts in support can reflect a changing sample composition.
* Add additional uncertainty to each poll where the scale varies based on whether it is a state or a national poll.
* Include state and national level polling error terms, which allows for unmodeled measurement error for each poll beyond the stated margin of error
Uncertainty:
The model estimates a large number of parameters with a relatively small number of polls. To estimate the parameters, we must supply relevant prior information. we started with values that we deemed reasonable a priori such as a 3% polling error for each poll based on historical data, but also evaluated the model output to determine whether the model gave sensible results
Model Vote intensions which we then use to predict the electoral outcome
Rentch et al. 2019’s proposed likely voter model?
They propose augmenting the set of questions typically used to assess likelihood of voting
Propose a probabilistic likely-voter model that not only uses the typical items (such as those from the Perry-Gallup index), but also adds demographic information about respondents such as age, education, race, income, gender, and strength of partisanship.
Bayesian Probabilies - inside head (Kruschke 2014)
Bayesian probability:
* Inferential statistical techniques: assign precise measures to our uncertainty about possibilities
* Uncertainty is measured in terms of probability
Sees probability as a measure of subjective belief about how likely an event is – not long-run frequency. For example, saying there’s a 53% chance of Harris winning reflects your degree of confidence, not a repeated frequency.
- Whenever we ask about how likely an outcome is, we always ask with a set of possible outcomes in mind. This set exhausts all possible outcomes, and the outcomes are all mutually exclusive. This set is called the sample space.
What is a posterior distribution?
The updated belief distribution of a parameter after taking into account the evidence from observed data.
Why do you include political regions?
Political regions – arbitrary however geographic factors provide some residual information that electoral history does not. These residual factors could include things such as the concentration of manufacturing jobs in the Midwest, natural disasters such as Superstorm Sandy affecting the Mid-Atlantic or really anything else that would not be evident in states’ political or demographic similarities.
Things that can have an shared effect on the electorate in these places
What is the sampling frame? (Bailey 2024)
It is the “pool” from which the sample is drawn and ideally should closely match the target population to ensure representativeness.
sampling frame -> how to identify pool of respondents that is representative
If your sample frame is only landlines and only old people have landlines the sample frame is biased
Why do we pool the polls and what should we worry about (Jackman 2005)
There is a lot of noise and little signal in individual polls (and short-term variations) –> they are filled with house effects and bias. Especially in state level polls are the swings often due to noise rather than opinion change
Therefore we can pool the pools so that the
Sample size increase -> uncertainty about estimates decrease as so do the 95 % confidence interval
o Statistical uncertainty decreases with the higher sample size minimizes the sampling error
- Combine the information in the published polls, leveraging them against one another, so as to obtain a clearer picture of what might be going on in the electorate over the campaign
Can’t pool them naively - polls can be biased due house effects. Use previous estimates of these as indicators when pooling the polls.
Statistical model for pooling the polls that:
○ (1) the model pools the polls, over- coming the limits to precision inherent in any one poll;
○ (2) the model smoothes over time, consistent with the notion that although support for either party fluctuates or trends over the course of a campaign, each campaign day need not be considered de novo—yesterday’s level of Coalition support will generally be an excellent predictor of today’s level of support;
○ (3) the model estimates and corrects for the possibility that any single poll is subject to bias or ‘house effects’
Silvers critique of 538’s forecast (and Gelman’s comment to the critique)
They have a very strong prior -> 538 trust less in polls and don’t use them until 75 days prior
–> “treats the fundamentals as a strong prior and Silver treat them as a weak one that you should be pretty eager to discard once you get enough polling. And I think one should be wary of strong priors in data-poor environments (only one election every four years) like election forecasting.” –> The Fivethirtyeight model expresses a lot of uncertainty because it’s allowing for large, Dukakis-versus-Bush-in-1988-style swings in public opinion and The Economist’s model and Nate’s model are making stronger predictions by assuming that swings during the campaign will stay in the narrow range that we’ve seen in recent national elections.
-OUR MODEL –> I turned down the effect of our prior with a small amount
Silver argues that fundamentals are also quite unpredicatable
Say that the reason behind 538’s mistrust in polls is irrational because state-polls and polls from 1948-2020 and polls now are a lot better
As a result of these assumptions about polling movement, 538’s model is extremely uncertain about much of anything at this stage.
- Example of Texas being a likely tipping-state –> because of to much uncertainty.
in Pennsylvania, 538’s 95th percentile forecast covers outcomes ranging from roughly Biden +18 to Trump +17.
Polls in 2020 and what it means for polling later on? (Silver 2021)
Remember on class 2016 vs. 2020
On average in the 2019-20 cycle, all election polls underestimated the performance of the Republican candidate by a whopping 4.8 percentage points.
–> So the big issue in 2020 wasn’t that the polls were that inaccurate — they were only slightly more inaccurate than usual — but that they almost all missed in the same direction.
Bigger change for systematic polling errors in the future:
There’s less chance for errors — overestimating the Democrat in one state, and the Republican in another — to cancel each other out.
More and more Online polling data (see Bailey 2024 for probability vs. non-probability sample)
- However It’s no longer clear that live-caller telephone polls are outperforming other methods, so they’ll no longer receive privileged status in FiveThirtyEight’s pollster ratings and election model
What is the role of the correlation matrix in the model?
It links polling data across states, allowing for information sharing between similar states.
account for trends
Why do we use likely-voter models? - what is it? (Rentch et al. 2024)
Since only a fraction of registered voters actually vote, polls that include all registered voters risk over-representing groups less likely to vote. Likely Voter models correct for this by identifying the subset of voters most likely to participate, thereby improving the accuracy of polling predictions.
Registered voters (RV) out of eligible = ~70%
- however there are still some voters who doesn’t show up
Different models: RV, LV and partisan identification
- if we just used RV as sample –> Overrepresenting D’s because they are more likely to answer - nonresponse bias
–> if we use ‘certain to vote’ -> swing voters who might SWING can be left out and people misreport whether they will vote
- Turnout is skewed based on sociodemographics, education – hispanics is less likely to vote for example
Its population = subset of citizens who can be expected to vote on election day
prediction of who will turnout + their choice today
However Nonignorable nonresponse bias can be an issue?
What are the main insight from Lock & Gelman 2010?
Key contribution:
- Separate the national forecast (based on fundamentals) from the relative positions of the states (for which past elections and current polls can be combined to make inferences). Preelection polls, not surprisingly, are more reliable as they get closer to the election. –> use only one set of preelection polls from 9 months out - Linzer says that updating continously increases precision. Their probability was because 2008 wasn’t close
o They generate both a posterior distributions for the national popular vote and each state’s position relative to this (based on previous elections) they add them together to get posterior distri- butions for the proportion voting Democratic in each state.
* Focus on state relative positions –> how they normally deviate from the national vote share of democrats
o Thats what we want to do with state-polls –> see where they correlate and make them predict each other
- Used in our state-correlation matrix –> we use how state relative positions –> how they normally deviate from the national vote share of democrats
When we go from the prior (NPV) to the states vote share we use state relative positions (based on previous elections) to see the difference between each state
What is the two kinds of Non-response bias and which is problematic? (Bailey 2024)
Ignorable nonresponse (coverage) –> Nonresponse associated with observed characteristics such as age and education
–> Solution: Weighting can correct for this, for example, by adjusting the proportion of young men in the final dataset to match their share in the overall population –> for example likely voter models is a kind of weighing
Nonignorable nonresponse (missingness) When people’s willingness to respond is related to their opinions even after accounting for demographics.
o if they differ with regard to some characteristic related to the quantity we are trying to measure.
* Example: In political polling, for example, nonresponse appears nonignorable for questions about intention to vote - or if Trump supports doesn’t respond.
* Cannot use weights because we don’t know how many people are x in the whole population
Examples of this:
2016 - under-representation of R voters in polls (+3% D)
lack of lesser educated in samples -> weighting by education in 2018, 2020
late-deciding voters broke for R
2020 - under-representation of R voters in polls (+4.3% D) (COVID!)
highest state-level error where: more white, more rural, less populated states
not due to errors in sampling frames due to polling mode (coverage)
non-response bias of R voters + R who responded more likely to support Biden!
The presidential approval of 2020 and prediction of TFC? (The bitter end)
In short, Trump stood out for having an extraordinarily stable approval rating as well as a low one.
TFC:
- The party of the incumbent president tends to win a larger share of the vote when the economy is growing, the incumbent is popular, the incumbent is running for reelection, and the party has held the White House for only one term rather than two or more in a row. All in 2020 Trump case
–> The standard tfc also favored Trump
- However The outcome was mostly in line with people’s subjective economic evaluations, which did forecast a narrow defeat for Trump and was very much in line with what Trump’s approval rating predicted.
What types of sociodemographics are there and what role does they play in the Michigan Model? (The American Voter revisited)
Two types of sociodemographics (Lewis-beck book) that can affect ones party identification and sets a baseline for short term factors
* Group identification - refers to an individual’s psychological attachment or identification with a specific social or political group, such as unions, racial or ethnic groups, or religious affiliations.
* Class voting - objective demographic traits of individuals, such as age, gender, education, income, and geographic location.
Group membership: refers to an individual’s psychological attachment or identification with a specific social or political group, such as unions, racial or ethnic groups, or religious affiliations. –> The group shapes the politics of its members because they psychologically identify with it – the more identification the more it influences
o Identification is based on a sense of shared purpose, identity, or norms within the group collective identity. Psychological identification –> personal experiences, social integration, socialization
2. ‘linked fate’ -> shared experiences –>
o Group membership tends to be stable over time, as these identities are deeply rooted in socialization and historical context.
o Examples: African Americans’ consistent Democratic voting patterns due to shared experiences and group cohesiveness
Social characteristicsare objective demographic traits of individuals, such as age, gender, education, income, and geographic location.
o These characteristics describe the individual without necessarily implying psychological attachment or shared identity.
o Social characteristics influence politics indirectly by shaping an individual’s experiences, environment, opportunities
Social class –> objective (income, type of employment, education) and subjective
* People with social characteristics can differ for at least two reasons.
o First, the same social situation tends to engender some of the same experiences and correspondent political attitudes. For example, college students who rely on student loans may be sensitive to campaign promises of educational aid.
o Second, a changed social placement can lead to different expectations about one’s political role
How our model works as a bayesian forecast? (The economist 2024)
First, before conducting a study, researchers explicitly state what they believe to be true, and how confident they are in that belief. This is called a “prior”.
Next, after acquiring data, they update this prior to reflect the new information—gaining more confidence if it confirms the prior, and generally becoming more uncertain if it refutes the prior (though not if the new numbers are so definitive that leave little room for doubt).
In this framework, the expected distribution of potential vote shares in each state derived above is the prior, and polls that trickle in during the course of the campaign are the new data. The result—a “posterior”, in Bayesian lingo—is our forecast.
Linzers approach - similar to ours? (2013)
Select the 50 posterior draws of π_(iJ ) produced in a single iteration of the sampling algorithm, and tally the total number of electoral votes in states where the Democratic candidate is forecast to receive more than 50% of the two-party vote.
I then add the three electoral votes of the District of Columbia, which is reliably Democratic.
Repeating this calculation across multiple sampling iterations produces a distribution of predicted electoral vote outcomes.
The proportion of these outcomes in which the Democratic candidate receives an absolute majority—270 or more—of the 538 electoral votes is the Democratic candidate’s probability of victory.
Why do polls sometimes fail to predict election outcomes? Bring in 2016 and 2020 as examples (class)
Possible reasons include late-deciding voters, nonresponse bias, and under-sampling of key demographic groups.
What was the Great Compromise?
Negotiations among the 13 states at the constitutional convention –> Two discussed models
o The Virginia plan: number of representatives depend on state population
o New Jersey plan: Unicameral legislature, where each state receives one vote
The Great Compromise established a bicameral legislature:
House of Representatives: Representation based on state population.
Senate: Equal representation, with each state having two senators.
Weighting in polls (Bailey 2024)
The problem weighting sets out to solve is that survey samples often have too many people from some groups and too few people from other groups regarding the desired population. Weighting takes on this problem by giving more weight to the people in the “too few” category and less weight to people in the “too many” category.
- Here we again differ from nonignorable and ignorable nonresponse can weight ignorable easily
- Two widely used algorithms for generating survey weights: Cell-weighting & Raking.
Problems: can have very few in one group (imprecise) and are vulnerable to unrepresentativeness within subgroups
Demographic tendencies in 2024 (Pew Research center 2024)
Race
White voters are much more likely than those in other racial and ethnic groups to associate with the Republican Party. Hispanic and Asian voters tilt more Democratic. Black voters remain overwhelmingly Democratic.
o Have been some shifts toward the GOP in most groups in recent years.
We see a shift in every race group towards trump in 2024 - improves in latino
Education
- Has come a diploma gap – realignmet towards The Republican Party now holds a 6 percentage point advantage over the Democratic Party (51% to 45%) among voters who do not have a bachelor’s degree. - college degree 55 % to Democrats
- Since 2017, the gap in partisanship between college graduates and those without a degree has been wider than at any previous point since 1990’s
- Gender gap between white women with and without
White voters are far more polarized along educational lines than are Hispanic and Black voters.
Also see voters without a college degree differ substantially by income in their party affiliation Social class?
Bidens Winning coalition in the primaries 2020?
Part of the reason Biden’s coalition helped him win is simple math: the demographic groups that tended to support him make up a big chunk of the party.
o older voters are more numerous than the youngest voters.
o Strong Democrats are more numerous than in- dependents.
o Black Americans are numerous in the primary electorates in key states like South Carolina but not in Iowa and New Hampshire.
What is the difference between deterministic and probabilistic LV models? (Rentch et al. 2024)
Deterministic likely-voter models—typically called “threshold” or “cutoff” models—create a likely-voter score for each respondent and then create a decision rule for whether to include or exclude a response when calculating election predictions
o Perry-gallup index
o Loss of information – we know some non-likely voters voted
Probabilistic models: Each respondent is assigned an estimated probability that they will vote. This probability is then used as a weight: Responses from those who are more likely to vote are weighted more heavily than responses from those who are unlikely to vote, but all are included in the election prediction.
What are the three steps of Bayesian inference in election forecasting? - think about our model build
- Prior Prediction: Use national polling and fundamentals (like the economy) to estimate the popular vote. This step avoids overfitting by simplifying through techniques like elastic-net regularization. Models are trained using leave-one-out cross-validation to ensure robustness, with adjustments made for state partisan lean.
- Incorporate New Data: Add recent polling data, reflecting shifts in voter sentiment.
- Update Prior Using Markov Chain Monte Carlo (MCMC): Explore thousands of parameter values to evaluate how well they fit the data and align with prior expectations. This generates a range of possible election outcomes, with more likely scenarios appearing more often in the simulations.
What is the role of “priors” in Bayesian analysis and our model?
Priors are the initial beliefs about parameter values before observing data. They play a critical role in Bayesian inference as they combine with observed data to form the posterior distribution.
Our prior is the fundamentals model
Sociodemographics and group membership (Lewis-Beck et al. 2011)
- Sociodemographic characteristics like race, gender, education, religion, and social roles create distinct groups within a society.
- Group membership often fosters collective identity or shared experiences. For instance, people within a group may perceive shared benefits or challenges (e.g., racial minorities supporting policies that promote equality).
- Members of a group often make decisions based on aggregated rational cost/benefit calculations, meaning they consider how policies impact their group rather than solely focusing on personal outcomes.
Example: Evangelical Christians are typically more supportive of Republican candidates.
Over time, shifts in political alignment can occur due to major historical events or policy changes. For example can Trump’s appearance maybe be seen as something that has created a realignment
Bayes rule
Probability as a measure of subjective belief about how likely an event is – not long-run frequency
- Bayesian data analysis adopts this version of probability and utilizes Bayes’ rule to update prior probabilistic beliefs based on observed data. It’s a way of combining what you already believe (prior) with what the new data tells you (likelihood) to get an updated belief (posterior).
- Likelihood You calculate how well the new data supports different possible values of the parameter.
For example, if polls/evidence favor Harris more than the prior, the likelihood will push the posterior belief upward.
The posterior is a compromise between the likelihood (=Evidence) and the prior
What are the types of elections and what would you say 2024 was?
A maintaining election is one in which stable partisan attachments continue to be a major determinant of election re- sults.
* The party that has a majority in terms of party loyalties wins the presidency, generally through affecting the partisan attitudes has more partisan identification
* 1964 LBJ
A deviating from dominance election would then be one in which the short-term partisan attitudes lead to the election of the presidential nominee of the minority party, without a fundamental shift in the party identification bal- ance in the nation.
* It is important to emphasize that this definition of a deviating election is based on short-term defection from partisanship due to the six partisan attitudes, not a long-term decline in the strength of partisan ties. Long-term partisan decline could be demonstrated only when the majority party loses election after election
Reinstating dominance - the majority party wins again
* JFK 1960 and Carter 1980
The realigning election is one in which the partisanship of people changes.
* The realigning election is one in which the partisanship of people changes.
* Definition: Significant shift in the sociodemographics group –> can lead to balancing or dominance
* Moving a lot of voters around
* New Deal realignment FDR the movement of Catholics and Jews into the Democratic Party marked the New Deal realignment of the 1930s
* Lesser extent with Reagan 1980 –> White working class, white evangelists and gender
* 2024 realignment of working class (education) – Age, racial groups
“Balancing” election (2016 and 2020)
o One in which neither party has a majority in party identification.
o The partisan balance is so close that either party could win.
o This is not an election in which short-term forces work against the majority party, but one in which it is not possible to speak of a majority party.
o (speaks to Sides et al.’s point about calcification)
o Maybe also 2024
The real story of the last 50-plus years of American presidential elections is a weakening of the Democratic lead in party identification to the extent that elections are very close and can be swung by any number of short-term matters.
What are the two main components of Bayesian data analysis?
Two fundamental components to Bayesian data analysis:
The reallocation of credibility across possibilities
The possibilities over which we allocate credibility are parameter values in meaningful mathematical models
Parameter values are the unknown quantities we aim to estimate -> the polling error, standard deviation of Fundamentals model
The prior represents our initial belief about the plausible range of parameter values before seeing the data. - can be informative or non-informative
Each set of parameter values is evaluated by comparing how well it fits the observed data.
Each sampled parameter value represents one possible “state of the world.”
The likelihood tells us how well a specific set of parameter values explains the observed data (e.g., polling results).
Posterior
- The posterior is the updated belief about the parameter values after combining the prior and the likelihood (based on Bayes’ Rule):
- The posterior distribution assigns probabilities to all possible parameter values, reflecting how plausible they are given the data.
Simulation where we randomly sample the posterior
Take a random walk through the parameter space, favoring parameter values which have a high posterior probability based on the prior and polling data
To take next step, there is a proposed jump from current position, with jump sampled randomly from a proposal distribution
Proposed jump accepted or rejected probabilistically, depending on the relative densities of the posterior at the current and proposed positions
Key elements:
Descriptions of the data: Mathematical formulas that characterize the trends and spreads in the data
Formulas contain parameter values which determine the shape of mathematical forms
Control knobs on mathematical devices that simulate data generation; change a parameter value, and you alter a trend in the simulated data
The economist 2024 model?
The first step: Generate a prediction for the national popular vote on election day (prior)
Fundamentals model are quite stable and often foreshadow voters behaviour - economic voting (Erikson & Wlezien). average Error as late-campaign polls
-Avoid overfitting the fundamentals-based model by simplifying (elastic-net regularisation) and training model (leave-one-out cross-validation: chopping up a dataset into lots of pieces, training models on some chunks, and testing their performance on others. In this case, each chunk is one election year)
Their fundamentals model:
- No benefit for term - they assigned a penalty to parties that had already been in power for at least two terms (in keeping with the spirit of the “Time for Change” brand).
- Broad range of economic metrics – change in them
2) State-level polling forecast –> Go from TFC/NPV to State partisan vote share.
first Model differences between states to get SPV
- Same process as above but project each state’s “partisan lean”: How much it more it favours Democrats or Republicans than America as a whole does, and thus how it would be expected to vote in the event of a nationwide tie.
- So, to get the prior for the states vote share: TFC +- the states partisan lean.
Based upon each state vs. national difference in the 2020 (and 2016 - economist) elections
In this framework, the expected distribution of potential vote shares in each state derived above is the prior, and polls that trickle in during the campaign are the new data.
3) Polling - first a model for polling error, then the state-correlation matrix for
Unceartanty regarding polls: Non-sampling error
- (1) polls are subject to the vagaries of voter turnout.
- (2) House effects
The final step in our treatment of polls is pooling the information they provide. Information sharing across states about state and national opinion changes
- (1) an adjustment for overall national trends
- (2) extend this method to state polls. state correlation matrix - Sociodemographics
o Every state poll to influence its estimate of where voter preferences stand in every other state, by varying amounts
Making the forecast: Simulation
MCMC –> Using the posterior, Take a random walk through the parameter space, favoring parameter values which have a high posterior probability –> they are more likely to happen depending on the posterior.
Allows state-polling averages to drift a little. Each step of this “random walk” can either favour Democrats or Republicans but is more likely to be in the direction that the “prior” prediction (The tfc and national polls) would indicate than in the opposite one. As the election draws near, there are fewer days left for this random drift to accumulate, reducing both the range of uncertainty surrounding the current polling average and the influence of the prior on the final forecast.
The ultimate result is a list of 10,001 hypothetical paths that the election could take. The more likely a scenario (based upon the posterior probabilities), the more often it will appear in these simulations—but even highly improbable ones (such as Ms Harris winning the electoral college despite losing the popular vote) will show up every so often. The resulting probabilities of victory are simply the fraction of these simulations that a given candidate wins.
How has Trump’s presidency affected calcification?
Three elements in the Trump presidency, 2020 election and aftermath –> upshot was more calcified politicas
o (1) Long-term tectonic shifts have pushed the parties apart while making the views within each party more uniform –> gradually increasing partisan polarization. The Democratic and Republican parties have become more internally homogeneous and more different from each other in political ideology, certain demographic characteristics, and certain policy issues
important implications for elections.
higher levels of partisanship in presidential approval and voting behavior.
The weakening power of other factors that have traditionally affected evaluations of presidents and voting in presidential elections.
(2) Shorter-term shocks, catalyzed especially by Trump, have sped up polarization on identity issues
* Race, gender, ethnicity etc.
The Democratic and Republican parties have rapidly divided on issues related to identity, especially race, ethnicity, nationality, religion, and gender
○ There has been a sharp increase in the magnitude of the already existing divisions
* These shocks stem directly from the identity, rhetoric, and decisions of political leaders and how the public has reacted to them. –> Main argument is elite cues
- A central part of this story is Trump himself.
An example Immigration –> The same pattern characterized attitudes on other immigration topics and identity-inflected issues: any longer-term partisan gap quickly became much larger.
(3) It is precisely these identity issues that voters in both parties care more about— exacerbating divisions even further and giving politicians every incentive to continue to play to them. –> The political priorities
The basic logic behind the Markov Chain Monte Carlo
The problem: How do we sample a large number of values from an unknown distribution?
We use a statistical technique called Markov Chain Monte Carlo (MCMC)
MCMC takes a “random walk” through the parameter spaces while exploring new values (e.g., NPV, state biases). Steps/parameter values are accepted or rejected based on depending on the relative densities of the posterior at the current and proposed positions –> favoring a set of parameter values which have a high posterior probability. How credible a particular set of parameter values are, is based on the product of the prior (historical or fundamentals-based predictions) and the likelihood (fit with observed polling data - The probability of observing the polling data regarding the new parameter value ).
- Focuses more on parameter values that fit the data well, building a posterior distribution.
One draw of these parameter values therefore represents one election outcome.
This is done 10.000 times –> Together all of these draws define the posterior probability
The ultimate result is a list of 10,001 hypothetical paths that the election could take. The more likely a scenario, the more often it will appear in these simulations—but even highly improbable ones (such as Ms Harris winning the electoral college despite losing the popular vote) will show up every so often based upon the proposal distribution. The resulting probabilities of victory are simply the fraction of these simulations of different parameter values that a given candidate wins.
Elements of a bayesian analysis
Descriptions of the data: Mathematical formulas that characterize the trends and spreads in the data
- These formulas contain parameter values which determine the shape of mathematical forms
- Control knobs on mathematical devices that simulate data generation;
o –> change a parameter value, and you alter a trend in the simulated data
For example: Normal distribution has two parameters:
o Mean: controls location of distribution central tendency (location parameter)
o Standard deviation: controls width or dispersion of distribution (scale parameter)
o Change of them and you would change the distribution of the data
What role does polling mode play in survey accuracy?
Different modes (e.g., phone vs. online) can produce systematic differences in responses due to social desirability bias or coverage issues.
Explanations of the Trump win in 2024
Hispanic swung towards trump
- Gains across non-white racial groups
- Could be because of issue-voting sociodemographics not as strong anymore
- Border counties shifts right-wards Starr County, population 65,920 and 97% Latino, shifted 75 PERCENTAGE POINTS in Trump’s direction.
o Class-vote/issue vote more than they identify as non-american/latinos
Trump was especially strong with less consistent voters. likely voter models vrs actual turnout underrepresenting Trump coalition?
- young male low-propensity voters turned out
- first time voters more republican because of Covid-19
Polling error is the same direction – pretty much all underestimated trump again non-response bias of Trump voters?
- Third time they miss the same direction
- Actually, underestimated him less weight on education can be something but maybe not enough
White educational attainment to Trump realignment in 2016 already haven’t moved since
- A big part of why Trump won Wisconsin in 2024 is that he continued to make gains in the wards with the lowest levels of college education, while Harris only matched Biden’s performance in the most educated wards.
- Low income and low education for trump just 12 years ago, the inverse was true. (Obamas coalition)
Presidential approval set to predict popular vote well
- The structural fundamentals worked out to be quite predictive You can’t argue a bad economy apparently (Dickinson 2014)
- Immigration and economy
What were the key proposals for selecting the president and What solution did the framers devise for selecting the president?
- Selection by the legislature: Risk of corruption and cabals, and concern that the president wouldn’t be independent.
- Direct election by the people: Concerns about voter information in a large country, potential advantage for big states, and Southern states losing power (as enslaved people couldn’t vote).
Solution –> The Electoral College, where the president is indirectly chosen:
States appoint electors equal to their total representatives and senators in Congress.
Electors vote for president and vice president, and results are certified by Congress.
Large states received the most electors due to population.
Small states were guaranteed a disproportionate share of electors through the Senate.
Southern states benefited as 3/5 of the enslaved population was counted for determining electors.
How does a Deterministic Likely Voter model work? (Rentch et al. 2019)
A deterministic model uses a fixed cutoff score to classify likely voters. For example, the Perry-Gallup Index assigns a score to respondents based on their responses to a series of questions (like past voting history, vote intent and interest in the election). Voters with scores above a certain threshold are classified as “likely voters”
Rentch et al. 2019: * Deterministic likely-voter models—typically called “threshold” or “cutoff” models—create a likely-voter score for each respondent and then create a decision rule for whether to include or exclude a response when calculating election predictions
○ BUT cutoff approaches suffer from the loss of information inherent in labeling each respondent as either clearly a voter or a nonvoter.
What is sociotropic vs. egocentric voting?
Pocketbook vs. National
Sociotropic voting occurs when voters base their electoral decisions on the national economic situation rather than their personal financial situation.
Egocentric voting, on the other hand, occurs when voters focus on their own personal finances. Sociotropic voting is generally more influential because media coverage tends to focus on national economic indicators, which affects a broader public perception
* The popular assumption holds they peer into their pocketbooks - If my personal finances are suffering, I put blame on the president, * Overall, the pocketbook voting hypothesis receives faint support. ○ If economics makes an important difference, that difference must come from elsewhere. * “Sociotropic” effects: Sociotropic evaluations assess the collectivity, rather than the individual.
(Lewis beck et al. 2011)
Why do we need sociodemographics in our model?
Polling data from one state informs us about opinion in other states. We do this through a covariance matrix of states, using sociodemographic factors, because they in theory affect voting behavior.
why important:
1. states are very different sociodemographically
2. states do not move in same direction!
Importance of both fundamentals and polling (Graefe 2018)
This study analyzes the relative accuracy of experts, polls, and the so-called ‘fundamentals’ in predicting the popular vote in the four U.S. presidential elections from 2004 to 2016.
Fundamentals:
o The fundamentals consistently overpredicted the Republican vote by substantial margins.
o In other words, in each of the past four elections, the Republicans underperformed relative to the fundamentals, and achieved less votes than what could be expected from historical data.
The results further suggest that experts follow the polls and do not sufficiently harness information incorporated in the fundamentals.
The key here is that the fundamentals-based forecast provided different in- formation than both experts and polls, thus increasing the likelihood that the combined forecast would bracket the true value
Combining expert forecasts and polls with a fundamentals-based reference class forecast reduced the error of experts and polls by 24% and 19%, respectively. The findings demonstrate the benefits of combining forecasts and the effectiveness of taking the outside view for debiasing expert judgment.
How does the Electoral College work today?
Electors chosen by political parties: Typically loyalists or donors; their names rarely appear on ballots.
Winner-take-all system: All states except Maine and Nebraska allocate all electors to the statewide winner.
Statewide elections: All electors are chosen through elections, not state legislatures.
Discretion discouraged: Most electors are bound to vote for their pledged candidate.
What was the Three-Fifths Compromise?
The Three-Fifths Compromise determined that enslaved people would be counted as three-fifths of a citizen for population and taxation purposes.
This increased Southern states’ representation in the House of Representatives as they had a lot of slaves
For example, in 1793, Southern slave states gained 14 extra seats (47 instead of 33) due to this compromise.
What were the two main plans proposed during the negotiations for the U.S. legislature?
- Virginia Plan: Representation based on state population, favoring larger states.
- New Jersey Plan: A unicameral legislature where each state receives one vote, ensuring equal representation for all states.
Dickinson 2014 - fundamentals in 2012
The key point is that the economy, and not a new form of “identity” politics based in part on candidate qualities, was the major election issue despite what was the media’s interpretation. Based on the fundamentals it called for a close Obama win –> 2024 case???
The evidence suggests that Voters, while perhaps not deeply informed regarding candidates and issues, do cast their vote based on their understanding of fundamentals, such as the state of the economy, as viewed through their own partisan predispositions.
What are “house effects” in election polling and what are common sources for it? (Silver 2021 & Gelman 2021)
“House effects” refer to systematic differences in polling results between survey organizations. For example, some polling firms may consistently overestimate support for a specific party due to differences in survey design, weighting, LV’s or data collection methods.
However we can’t counter actual house effects in to our model - we only know house effects after an election. But we can use prior knowledge of past biases and incorporating them as adjustments in models.
And look at pollster ratings
Gelman 2021 argument –> they cancel each other out because we use so many polls
Simulation and Markov Chain
Because we have many parameters, it quickly becomes analytically untrackable we use Simulation to sample the posteriour
We use a statistical technique called Markov Chain Monte Carlo (MCMC), that explores thousands of different values for each parameter in our model, and evaluates both how well they explain the patterns in the data and how plausible they are given the expectations from our prior.
It takes random steps (jumps) to new parameter values based on a “proposal distribution.” Steps are accepted or rejected probabilistically, depending on how much better (or worse) the new position explains the data compared to the current one.
The probability of a step depends on the prior (initial belief) and the likelihood (the new evidence - polling). Together, they define the posterior probability.
So we start at a random set of parameter values (gives one electoral outcome) - jump to another which is accepted or rejected (gives a electoral outcome)
Some of them involve large nationwide, regional, or demographic polling errors benefiting one party or another. Some will show registered-voter polls suffering from a large bias in one direction; others little difference between types of survey populations or polling methods. The more likely a scenario, the more often it will appear in these simulations—but even highly improbable ones (such as Ms Harris winning the electoral college despite losing the popular vote) will show up every so often. The resulting probabilities of victory are simply the fraction of these simulations that a given candidate wins.
