Forecasting the 2012 U.S. Presidential Election: Should we Have Known Obama Would Win All Along? Robert S. Erikson Columbia University Keynote Address IDC Conference on The Presidential Election of 2012: Campaign and Results Jan. 6, 2013
How do we predict elections? We consult Nate Silver (538 blog, New York Times)
Or we consult similar number crunching, such as these predictions by Drew Linzer, Emory University
Or we consult the Election Markets: Here, Intrade Prices on Obama Victory
Or we even might consult political scientists and their multivariate models
Predicting the Next Presidential Election: When do we make our forecast? Right after the previous election, 4 years earlier Two years before the next presidential election, following the midterm election Early in the election year Just before the party conventions Just after the party conventions After the debates Election eve. Election night
Predicting the Next Presidential Election: How do we make our forecast? Polls Polls dominate as a predictor as th election approaches. Not so good earlier. The economy Both objective and subjective measures. Leading economic indicators offer an advanced look at what will happen, which voters will react to. What people say will happen Aggregated, poll respondent predictions are good election forecasters (Wolfers). The wisdom of crowds. Betting markets (gambling) Prices in election markets or by oddsmakers in theory incorporate more than polls. Although this is a hot idea among scholars, their value has been debated by Erikson and Wlezien.
Predicting US Presidential Elections This talk draws on several sources: My recent book, The Timeline of Presidential Elections, How Campaigns Do (and Do Not) Matter, coauthored with Christopher Wlezien Articles by Erikson and Wlezien on election markets and polling accuracy. Erikson and Wlezien forecasting models, such as those in PS (October 2012) and post-mortems in PS (January 2013) I also rely on the work of others, with a focus on the polls in the final days of the 2012 campaign. The framework is time, to see how and how well we can predict presidential election outcomes at various times from the previous election to Election Day.
Predicting the Next Presidential Election: Right after the previous election, 4 years earlier After each election, pundits predict that the winning party will dominate for years. Typically the losing party is a victim of its own ideological extremism, scapegoating its own leaders. In truth, the answer depends. After each turnover of the presidency from one party to the other, the new presidential party wins. 9 of the last ten times. The exception: Carter loses to Reagan, 1980 But after two terms in office, the dominant party loses. 4 of the last five times. The exception: Bush beats Dukakis, 1988
-10-5 0 5 10 Midterm vote does not predict the next presidential election 1964 1948 2012 2004 1996 2000 1968 1992 2008 1960 1976 1988 1952 1980 1956 1984 1972-5 0 5 10 % Dem. Vote for House of Representatives, 2 years earlier
Predicting the Presidential Election: During the election year. This is the subject of The Timeline.
The Timeline of Presidential Campaigns Throughout each presidential election year, trialheat polls are conducted to reveal the division of voters choice between the two major parties presidential candidates. Polling begins long before the party conventions certify their parties nominees, in January if not earlier. How much do these polls tell us about the eventual outcome of the November election? What other factors besides the polls matter?
The Timeline of Presidential Campaigns Data base is all national presidential trial-heat polls involving the major-party candidates over 15 presidential elections, 1952-2008 (January-November) In today s discussion, (1) Illustrate what we have learned with some graphs [scatterplots] from elections 1952-2008 (2) Apply to 2012
The Timeline of Presidential Elections: The Argument in A Nutshell Early in the election year, trial-heat polls are only weak indicators of the outcome. Election trends evolve very slowly from week to week. Vote intentions change least during the official Autumn campaign, by which time most peoples minds are made up. Trends are most volatile (1) early in the election year, (2) following the party conventions, and (3) in the final week. Over the election year, the outcome is increasingly predictable from the polls, with the winner largely evident immediately following the conventions. The leader in the polls following the conventions almost universally wins the election Trends move in the direction of the fundamentals such as the economy
Predicting the Presidential Election: Early in the election year Early in the election year, neither the trial-heat presidential polls nor the president s approval predict the outcome. The economy at the time does somewhat better. The best predictor is the cumulative record of the leading economic indicators. Reason: Leading indicators tell us what the economy will look like 6 weeks ahead, when voters pay more attention.
25 50 75 25 50 75 25 50 75 Polls to Election Day, by days until the election -300-200 -150-120 -90-60 -30-10 -1 25 50 75 25 50 75 25 50 75 Dem. % in Polls Dem % actual 2-party vote Dem. % in Polls Graphs by Days before Election
Predicting the Presidential Election: Early in the election year Election results are almost totally unpredictable from the trial-heat polls early in the election year. Before April, voters have not always been introduced to the eventual candidates. The greatest learning takes place between January and April of the election year
0.25 Adjusted R Squared.5.75 1 Prediction adjusted R 2 by days to election, the final 300 days: from low to high predictability -300-200 -100 0 Days Until Election Adjusted R Squared Lowess Fit
-20-10 0 10 20 30 From January to Election Day 1964 1992 2008 1996 2004 1956 2000 1988 1960 1984 1980-20 -10 0 10 20 30 % Democratic, January Polls
-20-10 0 10 20 30 From April to Election Day 1964 1952 1956 1992 1968 1972 2000 1984 2008 1976 1960 2004 1988 1980 1996-20 -10 0 10 20 30 %Democratic, April Polls
-20-10 -20-10 0 %Democratic, Election Day 0 10 20 30 10 20 30 Evolution January to April 1964 1964 1992 2004 1956 2000 1988 2008 1996 1960 1984 1980 1952 1956 1992 1968 2000 1984 1972 2008 1976 1960 2004 1988 1980 1996-20 -10 0 10 20 30 % Democratic, January Polls -20-10 0 10 20 30 %Democratic, April Polls
Party Conventions Matter! Conventions matter! The largest shifts occur from before to after the two party conventions. Before the conventions, the vote becomes predictable from the polls. But the winner remains in doubt. In 4 of the 5 elections between 1966 and 2004, the preconvention leader lost the popular vote. Exception: 1996 After the conventions, the general election outcome becomes clear. Almost always the candidate who is ahead after the conventions wins. Exceptions: in 1980, the post-election polls tied. In 2000, Gore led post-conventions and won the popular vote. The least volatility in the polls occurs during the official campaign in September and October,. By then, most voters have made up their mind, and the cake is usually baked.
20 30 From Pre-Convention Polls to Election Day 10 1964 2008 1992 0 2000 1960 1968 2004 1980 1952-10 1956 1996 1976 1988 1984-20 1972-20 -10 0 10 Dem. Vote, Pre-Conveniton Polls 20 30
20 From Post-Convention Polls to Election Day 10 1964 0 2008 1968 1996 1992 19602000 1976 2004 1988 1980 1952 1956-10 1984-20 1972-20 -10 0 10 Dem. Vote, Post-Conventions Polls 20
.9.92.94.96.98 Stability of vote choice over the campaign (NES 1952-2008) -60-40 -20 0 Date of interview, Relative to Election Day Lowess Fit Observed Stability
.85.9.95 1 Stability of choice over the campaign, by information level -60-40 -20 0 Date of interview, Relative to Election Day 5 Highest Info 4 3 2 1 Lowest Info
-20-10 0 % Democratic 10 20-20 -10 0 10 20 Weekly Vote intentions by lagged weekly vote intentions, in the final four weeks of the campaign -4-3 % Democratic minus 50%, Previous Week% Democratic minus 50%, Previous Week -2-1 -20-10 0 10 20-20 -10 0 10 20 % Democratic minus 50%, Previous Week% Democratic minus 50%, Previous Week Graphs by Week
The Romney campaign is in trouble By Ezra Klein, September 17, 2012 But I didn t realize quite how dire Romney s situation was until I began reading The Timeline of Presidential Elections: How Campaigns Do and Don t Matter, a new book from political scientists Robert Erikson and Christopher Wlezien. What Erikson and Wlezien did is rather remarkable: They collected pretty much every publicly available poll conducted during the last 200 days of the past 15 presidential elections and then ran test after test on the data to see what we could say about the trajectory of presidential elections. Their results make Romney s situation look very dire. For instance: The least-stable period of the campaign isn t early in the year or in the fall. It s the summer. That s because the conventions have a real and lasting effect on a campaign. The party that gains pre- to post-convention on average improves by 5.2 percentage points as measured from our pre- and post-convention benchmarks, write Erikson and Wlezien. On average, the party that gains from before to after the conventions maintains its gain in the final week s polls. In other words, its poll numbers do not fade but instead stay constant post-conventions to the final week.
Predicting the Presidential Election: after the debates Individual debates have their short term impact. Consider Obama s poor performance after the first 2012 debate and his subsequent (small but visible) slump in the polls. But over the long run from before the first debate to after the final debate the amount of change in the polls is little more than occurs in the autumn when no debate occurs.
Net changes during past debate seasons
The Romney campaign is in trouble (cont.) By Ezra Klein, September 17, 2012 This year, it was the Democrats who made the biggest gains from before to after the conventions. Obama is leading by 3 percent in the Real Clear Politics average of polls, about double his lead before the Republican convention. If that doesn t fade by the end of the week or so that is, if it proves to be a real lead rather than a post-convention bounce then there s simply no example in the past 15 elections of a candidate coming back from a post-convention deficit to win the popular vote. This is about the point where I m supposed to write: That said, the race remains close, and the debates are coming soon. It s still anyone s game. But the most surprising of Erikson and Wlezien s results, and the most dispiriting for the Romney campaign, is that unlike the conventions, the debates don t tend to matter. There s a fairly strong degree of continuity from before to after the debates, they write. That s true even when the trailing candidate is judged to have won the debates. Voters seem to have little difficulty proclaiming one candidate the winner of a debate and then voting for the opponent, Erikson and Wlezien say.
Predicting the Presidential Election: on Election Eve During the final week there is a greater shift from polls to to the vote than from week to week in the trial-heat polls. Polls generally get the winner right, even if they tend to slightly exaggerate the margin. Considering that there is voter movement from the final polls to Election Day, pollsters do a good job.
10 1964 2008 2008 1996 1992 0 1976 2000 1960 1968 2004 1988 1980 1952-10 1956 1984 1972-20 % Democratic, Election Day Vote 20 From the final week to Election Day: Vote for the Democratic Candidate -20-10 0 10 20 % Democratic minus 50%, Final Week's Polls
-5 0 5 10 15 20 From the final week to Election Day: Vote for the President s Party 1972 1964 1956 1984 1996 1988 2004 2000 1968 1960 1976 1992 1952 2008 1980-5 0 5 10 15 20 % for President's Party minus 50%, Final Week's Polls
Predicting the Vote from other Sources (Not Trial Heat Polls)
The Economy and the Vote over the Campaign Timeline In polls early in the election year, economic conditions have virtually no relationship to voter preferences at the time. Poll respondents do not consult the state of the economy when they are asked in April how they will vote in November. The vote becomes more predictable from the economy as the campaign progresses. Even the economy measured early in the election year has some ability to predict the vote.
-20-10 -20-10 % for Pres. Party in Polls (minus 50%) 0 % for Pres. Party (minus 50%) 0 10 20 30 10 20 30 Income growth and the vote, April to November Mid-April 1964 Election Day 1956 1964 1972 1972 1992 1996 1980 1984 1956 1996 1988 1984 1960 2004 2008 19761988 2000 1960 2008 2004 2000 1976 1968 1992 1980 1952 1968 1952 0.5 1 1.5 Cumulatieve P/C Income Growth 0.5 1 1.5 Cumulative P/C Income Growth
-20-10 -20-10 % for Pres. Party in Polls (minus 50%) 0 % for Pres. Party (minus 50%) 0 10 20 30 10 20 30 Perceived economy and vote intentions, April, and vote, November Mid-April 1964 Election Day 1956 1996 1972 1992 1980 1984 19721964 1984 1956 1988 1996 2008 2004 1988 1960 1976 2000 2008 1992 1980 2004 1960 2000 1976 1968 1968 0 50 100 150 Economic Perceptions 0 50 100 150 Economic Perceptions
-5 0 5 10 15 Predicting 2012 from Per Capita Income Growth Through Quarter 14 1972 1964 1984 1956 1996 1988 1960 2004 1976 2000 1968 2012 1992 2008 1980 1952 0.5 1 1.5 Cumulative Per Capita Income Growth
-5 0 5 10 15 Predicting 2012 from Perception of Business Conditions, Quarter 14 1972 1964 1956 1984 1960 2012 2004 1996 1988 2000 1968 1976 2008 1992 1980 0 50 100 150 Business Conditions
45 50 55 60 65 Predicting the Vote from Q13 Leading Economic Indicators 1972 1964 1984 1956 1996 1988 1976 2012 1960 2000 1968 2004 2008 1980 1992 1952-2 -1 0 1 2 3 Cumulative Leading Indicators Growth Through Quarter 13
Predicting the 2012 vote: Hibbs Guns and Butter Model
The Economy and the Vote in 2012 How to measure th economy? Objective (Per Capita Income or GDP, Leading Econ. Indicators) Preferably a weighted average over multiple quarters Subjective(Economy Better or Worse in Polls) In 2012, measures differed in their prediction Income growth Romney Cumulative Leading Indicators->Obama Subjective Economy (Polls) ->Obama Economy is not everything. At most it explains half the variance The economy increasingly predicts vote intentions as the campaign evolves. Why does the economy matter? What is the Voters Psychology? Voters simply reward or punish based on performance Economy->Vote Economy ->Partisan and Ideological Evaluations->Vote
Why does the economy matter? The economy can affect intervening variables (party identification, voter and candidate ideological stances) that in turn affect the vote. The Macro Polity argument: The national vote is influenced by net party identification plus ideology (Stimson s mood) and party ideology (measured by scoring party platforms, moderate platforms win votes) Macro Polity predictions (actually post-dictions predict the vote well AND seem to be influenced by the economy. As if economy party ID and ideological proximity
Predicting the Vote: Some Observations Early in the election year, preferences in trial-heat polls do not reflect the economy, and have little bearing on the ultimate election outcome. But At the start of the presidential year, the economy may be a more important predictor of the outcome than the trial-heat polls This may be why Romney relied on the economy to drive the result, ignoring Obama s small lead in the polls Over the campaign, both the economy and the polls improve as predictors of the vote, with the polls becoming more dominant over time Obama s persistent small lead may be why Romney reset his campaign. The economy was not everything. Ideology matters too, favoring centrist candidates This made the choice of Ryan as the vice presidential candidate more puzzling. But each party thought it held the advantage on issues. The election helped to decide who is right.
Another indicator: Presidential Approval
Election Markets (Betting Markets) Especially among economists, it has become established wisdom that election markets are good election predictors, possibly even better than the polls. (Erikson and Wlezien) There is little evidene that this is true. They reflect the polls but add no further useful information Markets typically overestimate the chances of the underdog. This is interesting because it implies that people overestimate the volatility of campaigns Candidate X (e.g., Romney) might be behind in the polls, but who knows, anything can happen
Intrade Prices on Obama Victory
0.25.5.75 1 Predicting the Market Price from the Vote, 1952-2008 1992 2008 1996 1960 2004 1976 2000 1984 1956 1980 1988 35 40 45 50 55 60 65 Percent Democratic, President
Predicting the Market Price from the Vote, 1880-1932
0.25.5.75 1 1992 2008 1996 1960 2004 1976 2000 1984 1956 1980 1988 35 40 45 50 55 60 65 Percent Democratic, President
Elections used to be predictable even before the polls Before the polls, smart money could detect the outcomes and drove the price. These were efficient markets Now with polls, prices simply follow the polls So far, the current election markets are inefficient, thin markets. If the polls err, so do the markets. Classic case of 1948: Polls were wrong, and everyone (even the oddsmakers) predicted Dewey would defeat Truan.
Predicting the Presidential Election: The challenge to pollsters, who must get it right Consider the problems today s pollsters face The challenge of predicting who will vote (the likely voters) and who will not The declining response rate (now less than 10%!) The growing reliance on cell phones, which cannot be polled easily Some evidence that survey respondents are more likely to cooperate when their candidate is ahead Even after the final, last-minute, poll, the electorate may shift without warning. Pollsters always adjust their samples by reweighting (post-stratifying) to make the sample representative
Are the pre-election polls very good? Pre-election polls are very good. National polls and state polls have been particularly good in 2012 and other recent elections. Why are the polls so accurate? Best answer: Pollsters are so good because if they did not practice their art properly, they would be out of business. The need to get it right serves as the motivator.
Simon Jackman s State Forecasts based Largely on State Polls
Predicting the Vote on Election Day Exit Polls Exit polls interview voters as they leave the voting booth. Unlike pre-election polls, these election-day polls are notorioiusly unreliable. Some reasons Interviewers are one-day amateurs, not professionals Many votes are cast before Election Day Still, when properly reweighted post-election for demographics, etc., they provide useful information regarding who voted for whom and why. They are just not good at predicting the outcome.
Conclusions Presidential election outcomes are difficult to predict far in advance. Early in the election year, the economy is a better indicator than the polls. As the election nears, rely on the polls. The economy is not everything (see 2012) The polls work well despite increasing problems. Will there be a time when conventional polling fails?