An Electoral Vote Forecast Formula: Simulation or Meta-analysis Not Required
Richard Charnin
Oct. 31, 2011
It’s very surprising that election forecasting blogs and academics who use the latest state polls as input to their models don’t apply basic probability, statistics and simulation concepts in forecasting the electoral vote and corresponding win probabilities.
It is interesting to note that neither a simulation or a meta-analysis is required to calculate the expected electoral vote. Of course the individual state projections will depend on the forecasting method used. But the projection method is not the main issue here; it’s how the associated win probabilities are used to calculate the expected EV, win probability and frequency distribution.
2004 Monte Carlo Election Simulation Model (Pre-election and Exit Polls)
This model runs 200 election trials (both pre-election and post election)
https://docs.google.com/spreadsheet/ccc?key=0AjAk1JUWDMyRdDU5VERHay1mZExaT0lMRVhOXzg2aHc#gid=1
Meta-Analysis is an unnecessarily complex method and overkill for calculating the expected Electoral Vote; the EV is calculated by the simple summation formula given below. Princeton Professor Wang projected that Kerry would win 311 electoral votes with a 98% win probability. His model was close to the exit polls. But he was wrong to suggest that his forecast was “wrong” because Bush won undecided voters. Not so. Kerry easily won the late undecided vote.
http://election.princeton.edu/code/matlab/EV_estimator.m
http://synapse.princeton.edu/~sam/pollcalc.html
Wang never considered that the election was stolen. Then again, neither did AAPOR, the media pundits or political scientists. But overwhelming statistical and other documented evidence indicates that the 2004 election was fraudulent.
In every election, unadjusted state and national exit polls have differed from the recorded vote. On average the Democrats do 4-5% better than the recorded vote indicates.
The post-election 1988-2008 True Vote model confirms the exit polls.
https://docs.google.com/spreadsheet/ccc?key=0AjAk1JUWDMyRdGN3WEZNTUFaR0tfOHVXTzA1VGRsdHc#gid=0
Calculating the expected electoral vote is a three-step process:
1. Project the 2-party vote share V(i) for each state(i) as the sum of the final pre-election poll share PS(i) and the undecided voter allocation UVA(i):
V(i)= PS(i) + UVA(i)
2. Calculate the probability P(i) of winning each state given the margin of error (95% confidence):
P(i) = NORMDIST (V(i), 0.5, MoE/1.96, true)
3. Calculate the expected electoral vote EV. It is the product sum of the state win probabilities and the electoral votes:
EV = ∑ P(i) * EV(i), for i = 1,51
The challenger is expected to win the majority (60-90% UVA) of the undecided vote, depending on incumbent job performance. Gallup allocated 90% of undecided voters to Kerry in their final projection, pollsters Zogby and Harris: 75-80%. The National Exit Poll indicated that 65% of undecided voters broke for Kerry. Bush had a 48% approval rating on Election Day.
After calculating the individual state probabilities, we can calculate the EV win probability. The best, most straightforward method is Monte Carlo simulation. This technique is widely used in many different applications when an analytical solution is prohibitive. It is perfectly suited for calculating the EV win probability.
The Election Model uses a 5000 election trial simulation. The win probability is the total number of winning election trials/5000. The average electoral vote is calculated for the 5000 election trials. Of course, the average will only be an approximation to the theoretical value based on the summation formula. But the Law of Large Numbers (LLN) applies: the EV average and median are usually within one or two electoral votes of the theoretical mean. The close match between the Monte Carlo average, median and theoretical (expected) mean electoral vote is proof that 5000 election trials are more than sufficient for the simulation.
The Election Model includes a sensitivity (risk) analysis of five undecided voter (UVA) scenario assumptions. This enables one to view the effects of the UVA factor variable on the expected electoral vote and win probability. Kerry won all scenarios.
Electoral vote forecasting models which do not provide a risk factor sensitivity analysis are incomplete.
Links to the 2004 and 2008 Election Model simulations:
2004 Election Model
http://richardcharnin.com/ElectionModel.htm
The model projected that Kerry would have 337 electoral votes with a 99% win probability and a 51.8% two-party vote share. I allocated 75% of the undecided vote to Kerry. The unadjusted, pristine state exit poll aggregate provided by exit pollsters Edison-Mitofsky 3 months after the election indicated that Kerry had 52.0% and an identical 337 electoral votes. In the post-election True Vote model Kerry had 53.5% – a 10 million vote landslide. But it was not enough to overcome the massive fraud which gave Bush his bogus 3.0 million vote “mandate”.
2004 Election Model Graphs
National Trend
http://www.richardcharnin.com/index_files/ElectionModel_9609_image001.png
Electoral vote and win probability
http://www.richardcharnin.com/index_files/ElectionModel_9609_image002.png
Electoral and popular vote
http://www.richardcharnin.com/index_files/ElectionModel_9609_image003.png
Undecided voter allocation impact on electoral vote and win probability
http://www.richardcharnin.com/index_files/ElectionModel_9609_image004.png
National Poll Trend
http://www.richardcharnin.com/index_files/ElectionModel_9609_image008.png
Monte Carlo Simulation
http://www.richardcharnin.com/index_files/ElectionModel_9609_image011.png
Monte Carlo Electoral Vote Histogram
http://www.richardcharnin.com/index_files/ElectionModel_9609_image012.png
2008 Election Model
http://www.richardcharnin.com/2008ElectionModel.htm
The model exactly matched Obama’s 365 EV. His win probability was 100% as he won all 5000 election trials. His projected 53.1% share was very close to his recorded 52.9%. But it was wrong. The model utilized final pre-election likely voter (LV) polls which low-balled Obama’s True Vote. Registered voter (RV) polls indicated he would have a 57% share. The post-election True Vote Model and the unadjusted state exit poll aggregate indicated that Obama had 58% with 420 EV.
2008 Election Model Graphs
Aggregate state polls and projections (2-party vote shares):
http://www.richardcharnin.com/2008ElectionModel_12777_image001.gif
Undecided vote allocation effects on projected vote share and win probability:
http://www.richardcharnin.com/2008ElectionModel_32191_image001.gif
Obama’s projected electoral vote and win probability:
http://www.richardcharnin.com/2008ElectionModel_29371_image001.gif
Monte Carlo Simulation Electoral Vote Histogram
http://www.richardcharnin.com/2008ElectionModel_30550_image001.gif
This is a one-sheet summary of 2004 and 2008 True Vote calculations with many links to relevant posts and data.
http://richardcharnin.com/TrueVoteCalcSheet.pdf
Susan
October 31, 2011 at 5:11 pm
Just wondering if the hackers/programmers use something like this when writing the algorithms…