際際滷

際際滷Share a Scribd company logo

Election Polling & Forecasting 2004

Dennis Sweitzer
Dennis Sweitzer

A statistics presentation I gave at work using simulation to predict election results from state-by-state polling data.

1 of 21
Download to read offline
But first.. Something of topical interest Psephology The M&M graph
Poll of polls for October 8th, 2004 - Posted October 15th, 2004 Dead Heat Before the second and third debates, Bush's lead has dropped to less than 0.1%... http://tis.goringe.net/pop/pollofpolls.html
Reweighting by Party ID? Shifting Party Affiliation Strong Republican 100% Weak Republican 90% 80% Independent 70% Republican 60% Independent 50% Independent 40% Independent 30% Democrat 20% Weak Democrat 10% 0% Strong Democrat 2 6 0 4 8 2 6 0 4 8 2 6 0 '5 '5 '6 '6 '6 '7 '7 '8 '8 '8 '9 '9 '0 Year
Dewey Beats Truman Chicago Tribune, Nov 3, 1948 Sources of bias: ≒ distribution of telephones favored wealthy Dewey voters rather than poor Truman supporters ≒ the pollsters stopped polling two weeks or more before the election. ≒ Convenience Sampling
Golden Age of Polling? ≒ >95% Households have a telephone Easy to get random sample ≒ Until recently: good response ≒ Now: caller id, answering machines, etc. 30% response rate Cell Phones: No landline 竪Younger, Democratic
Online Polling?
Demographics ≒ "As people do better, they start voting like Republicans--unless they have too much education and vote Democratic, which proves there can be too much of a good thing." --Karl Rove
Electoral College Map
Electoral College Approach KERRY ELECTORAL VOTE PROJECTIONS from State-Level Poll Data 18-Oct-2004 Total number of Electoral states Votes Bush > 55% 15 110 Bush total: Bush 50-55% 12 146 256 Kerry 50-55% 15 187 Kerry total: Kerry > 55% 9 95 282 51 538 polls used beginning ending in estimate: 10/7/2004 10/12/2004
Basis for Simulation State EV K.share SD AL 9 44.3% 0.7% AK 3 37.0% 1.2% AZ 10 46.1% 3.2% AR 6 53.1% 3.1% CA 55 54.6% 2.8% CO 9 46.8% 2.2% CT 7 55.1% 2.7% DE 3 54.5% 2.4% DC 3 89.5% 0.7%
Simulate Elections 1 Random Electoral Votes Rand# SD Mean Normal Kerry Bush AL 0.396 0.73% 44.35% -5.84% 9 AK 0.281 1.20% 37.03% -13.66% 3 AZ 0.314 3.15% 46.08% -5.44% 10 AR 0.244 3.12% 53.15% 0.98% 6 CA 0.105 2.78% 54.63% 1.14% 55 CO 0.441 2.19% 46.76% -3.56% 9 CT 0.113 2.70% 55.15% 1.87% 7 DE 0.44 2.40% 54.50% 4.13% 3 DC 0.474 0.73% 89.49% 39.44% 3 22-Jul 4-Aug 30-Aug 2-Sep 7-Sep Elect Wins 76.0% 73.0% 56.4% 37.0% 39.2% Loses 23.6% 26.1% 42.2% 61.2% 59.7% Ties 0.4% 0.9% 1.4% 1.8% 1.1% Odds 3.22to1 2.80to1 1.34to1 0.60to1 0.66to1
Estimating States (1) ≒ Historical data: Compare state outcomes vs national ≒ Current data: Adjust national poll by historical difference ≒ Combine with recent state polls Example: http://www.mydd.com/story/2004/5/28/115448/006
Estimating States (2) Election data: 1980-2000 (6 elections) state vote counts For pair of states (2550 pairs): Regress each states vote on all other states For each available poll: Predict votes in all other states For each state: Use the median of predicted votes www.pollkatz.homestead.com/files/kerryEVproj.htm
Forecasting Models Predict Vote from: Economic conditions, social climate, incumbency, party fatigue, etc., etc. Limited by: Too little data Result: perfect predictions of past elections Explanatory, not predictive
Bread & Peace Equation 14 Votet = 留 + 硫1 了 j ln Rt j + 硫 2 CUM KIAt j =0 Real Income Growth and the Two-Party Vote Share of the Votes as a function of: Incumbent Party's Presidential Candidate % 65 ≒Income growth 1972 60 1984 1964 ≒War Casualties 1956 Two-party vote share Vietnam 55 2004: Predicts ~53% Bush 1996 1988 1960 50 Korea Explanatory, not predictive! 1976 1968 1992 45 (Whos income?) 1980 1952 40 -1 -0.5 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 % Weighted-average growth of real disposable personal income per capita during the presidential term
Incumbent Advantage or Disadvantage? ≒ Forecasting models: Incumbent advantage ≒ Final Polls vs. Final Votes: Incumbent disadvantage Apparently, undecideds lean toward challenger (at about 2:1)
My Favorite Forecasting Model Electability = 4P - V - S + R + 9G + 95DCI + 95GEN + 95NUC ≒ Elections since 1932 ≒ Predicts all elections since 1932 ≒ Developed using stochastic trials (i.e., guessing until something worked) ≒ Source: Annuals of Improbable Results 2004: Bush, 70; Kerry -20 http://members.bellatlantic.net/~vze3fs8i/air/pres2004.html
Another Favorite Washington Redskins, Last home game prior to election Redskins Win 竪 Incumbent wins Redskins Lose 竪 Incumbent loses ≒ True for entire history of Washington Redskins (15 elections) (1932 & earlier: Boston Braves, no predictive power) ≒ October 31, 2004: vs Green Bay Not anymore 28 to 14 defeat, favor Kerry
University of Iowa's Electronic Market http://128.255.244.60/graphs/graph_Pres04_WTA.cfm
Obligatory Bayesian Methods A Bayesian Truth Serum for Subjective Data. Prelec, Drazen. Science, Vol 306, Issue 5695, 462-466 , 15 October 2004 ≒ Reward based system ≒ Respondents compete ≒ Counterbalances tendency to agree with perceived majority ≒ Best: large samples, rational participants
Bayesian Truth Serum Example: Q1: Do you prefer painting A or B? Q2: Which would others prefer? ≒ Compute Information score+prediction score using sums of logarithms for each respondant and etc etc ≒ Truth Telling is Bayesian Nash equilibrium (I.e., reduced payoffs for anything else) ≒ Does it work? When? For who? For only Bayesians? What about frequentists? Under what conditions? Cost? Etc... ≒ A work in progress

More Related Content

Election Polling & Forecasting 2004

  • 1. But first.. Something of topical interest Psephology The M&M graph
  • 2. Poll of polls for October 8th, 2004 - Posted October 15th, 2004 Dead Heat Before the second and third debates, Bush's lead has dropped to less than 0.1%... http://tis.goringe.net/pop/pollofpolls.html
  • 3. Reweighting by Party ID? Shifting Party Affiliation Strong Republican 100% Weak Republican 90% 80% Independent 70% Republican 60% Independent 50% Independent 40% Independent 30% Democrat 20% Weak Democrat 10% 0% Strong Democrat 2 6 0 4 8 2 6 0 4 8 2 6 0 '5 '5 '6 '6 '6 '7 '7 '8 '8 '8 '9 '9 '0 Year
  • 4. Dewey Beats Truman Chicago Tribune, Nov 3, 1948 Sources of bias: ≒ distribution of telephones favored wealthy Dewey voters rather than poor Truman supporters ≒ the pollsters stopped polling two weeks or more before the election. ≒ Convenience Sampling
  • 5. Golden Age of Polling? ≒ >95% Households have a telephone Easy to get random sample ≒ Until recently: good response ≒ Now: caller id, answering machines, etc. 30% response rate Cell Phones: No landline 竪Younger, Democratic
  • 7. Demographics ≒ "As people do better, they start voting like Republicans--unless they have too much education and vote Democratic, which proves there can be too much of a good thing." --Karl Rove
  • 9. Electoral College Approach KERRY ELECTORAL VOTE PROJECTIONS from State-Level Poll Data 18-Oct-2004 Total number of Electoral states Votes Bush > 55% 15 110 Bush total: Bush 50-55% 12 146 256 Kerry 50-55% 15 187 Kerry total: Kerry > 55% 9 95 282 51 538 polls used beginning ending in estimate: 10/7/2004 10/12/2004
  • 10. Basis for Simulation State EV K.share SD AL 9 44.3% 0.7% AK 3 37.0% 1.2% AZ 10 46.1% 3.2% AR 6 53.1% 3.1% CA 55 54.6% 2.8% CO 9 46.8% 2.2% CT 7 55.1% 2.7% DE 3 54.5% 2.4% DC 3 89.5% 0.7%
  • 11. Simulate Elections 1 Random Electoral Votes Rand# SD Mean Normal Kerry Bush AL 0.396 0.73% 44.35% -5.84% 9 AK 0.281 1.20% 37.03% -13.66% 3 AZ 0.314 3.15% 46.08% -5.44% 10 AR 0.244 3.12% 53.15% 0.98% 6 CA 0.105 2.78% 54.63% 1.14% 55 CO 0.441 2.19% 46.76% -3.56% 9 CT 0.113 2.70% 55.15% 1.87% 7 DE 0.44 2.40% 54.50% 4.13% 3 DC 0.474 0.73% 89.49% 39.44% 3 22-Jul 4-Aug 30-Aug 2-Sep 7-Sep Elect Wins 76.0% 73.0% 56.4% 37.0% 39.2% Loses 23.6% 26.1% 42.2% 61.2% 59.7% Ties 0.4% 0.9% 1.4% 1.8% 1.1% Odds 3.22to1 2.80to1 1.34to1 0.60to1 0.66to1
  • 12. Estimating States (1) ≒ Historical data: Compare state outcomes vs national ≒ Current data: Adjust national poll by historical difference ≒ Combine with recent state polls Example: http://www.mydd.com/story/2004/5/28/115448/006
  • 13. Estimating States (2) Election data: 1980-2000 (6 elections) state vote counts For pair of states (2550 pairs): Regress each states vote on all other states For each available poll: Predict votes in all other states For each state: Use the median of predicted votes www.pollkatz.homestead.com/files/kerryEVproj.htm
  • 14. Forecasting Models Predict Vote from: Economic conditions, social climate, incumbency, party fatigue, etc., etc. Limited by: Too little data Result: perfect predictions of past elections Explanatory, not predictive
  • 15. Bread & Peace Equation 14 Votet = 留 + 硫1 了 j ln Rt j + 硫 2 CUM KIAt j =0 Real Income Growth and the Two-Party Vote Share of the Votes as a function of: Incumbent Party's Presidential Candidate % 65 ≒Income growth 1972 60 1984 1964 ≒War Casualties 1956 Two-party vote share Vietnam 55 2004: Predicts ~53% Bush 1996 1988 1960 50 Korea Explanatory, not predictive! 1976 1968 1992 45 (Whos income?) 1980 1952 40 -1 -0.5 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 % Weighted-average growth of real disposable personal income per capita during the presidential term
  • 16. Incumbent Advantage or Disadvantage? ≒ Forecasting models: Incumbent advantage ≒ Final Polls vs. Final Votes: Incumbent disadvantage Apparently, undecideds lean toward challenger (at about 2:1)
  • 17. My Favorite Forecasting Model Electability = 4P - V - S + R + 9G + 95DCI + 95GEN + 95NUC ≒ Elections since 1932 ≒ Predicts all elections since 1932 ≒ Developed using stochastic trials (i.e., guessing until something worked) ≒ Source: Annuals of Improbable Results 2004: Bush, 70; Kerry -20 http://members.bellatlantic.net/~vze3fs8i/air/pres2004.html
  • 18. Another Favorite Washington Redskins, Last home game prior to election Redskins Win 竪 Incumbent wins Redskins Lose 竪 Incumbent loses ≒ True for entire history of Washington Redskins (15 elections) (1932 & earlier: Boston Braves, no predictive power) ≒ October 31, 2004: vs Green Bay Not anymore 28 to 14 defeat, favor Kerry
  • 19. University of Iowa's Electronic Market http://128.255.244.60/graphs/graph_Pres04_WTA.cfm
  • 20. Obligatory Bayesian Methods A Bayesian Truth Serum for Subjective Data. Prelec, Drazen. Science, Vol 306, Issue 5695, 462-466 , 15 October 2004 ≒ Reward based system ≒ Respondents compete ≒ Counterbalances tendency to agree with perceived majority ≒ Best: large samples, rational participants
  • 21. Bayesian Truth Serum Example: Q1: Do you prefer painting A or B? Q2: Which would others prefer? ≒ Compute Information score+prediction score using sums of logarithms for each respondant and etc etc ≒ Truth Telling is Bayesian Nash equilibrium (I.e., reduced payoffs for anything else) ≒ Does it work? When? For who? For only Bayesians? What about frequentists? Under what conditions? Cost? Etc... ≒ A work in progress