Pitfalls in predicting voter turnout, 64% turnout rate projected for 2012

Readers who followed our pre-election series last spring may recall MPNblog.com projected that 2010 voter turnout in Missouri would likely set a midterm record. We now know that didn't happen, and by a long-shot.

Only 47 percent of registered voters cast ballots in Missouri's U.S. Senate race, or about 43 percent of the state's voting-age population. The only worse election for turnout in the last decade was the 1998 midterm which was dominated with morality attacks against former President Bill Clinton, Sen. Kit Bond's (R) re-election, and resulted in no major power shifts in Congress.

So what was wrong with our model published last March? Why were our predictions so far off from reality?  It turns out there are several reasons for the discrepancy.

Problem 1: Assumptions
Our model made assumptions eight months before the election. Coefficients used for model forecasts are based on history and are only as good as the data that feeds them. Projections of future data to feed into the model are really just best guesses. Back in March 2010, we couldn't have fully predicted this election would be a structural realignment for Congress that it turned out to be. In fact, few would have guessed last March that Democratic defeats would have been so numerous. We made assumptions about the unemployment rate and registered voters, both of which were generous assumptions.

Specifically, we assumed the statewide unemployment rate would be 9 percent in Nov. 2010, it turned out to be 9.6 percent. We assumed the voting age population of the state would be 4.51 million, off by about 25,000.  More importantly though, we assumed turnout would not be driven by other exogenous factors, which results in the next problem: the 2010 was clearly an atypical election.

Problem 2: Exogenous Factors
Aside from the results, we know 2010's results were atypical for Missouri because in the last 30 years it's the only election to fall outside the 95 percent confidence interval established in our turnout model. In other words, using the actual results from 2010 our model last year projected turnout at 2.5 million (+/- 230,000). The only election to come close to this level of error was in 1994, which still fell within the confidence interval. In 1994, the model was off by 195,000 votes, but in that case projecting lower turnout than really occurred.

Problem 3: Imputation of Voting Age Population
One of the questions I've received in multiple emails about our model is why we used voting age population instead of registered voters. The reality is that both measures are valuable, yet historically voter registrations were misaligned with actual voting patterns as rolls became outdated due to deceased or migrant voters. Further, voting age population provides a smoother baseline for which to compare trended data, thus allowing at a theoretical level a better comparison point for overall voting trends.

To highlight this point, see the chart below which captures total voting age population and total registered voters. Note that the voting age population is substantially more smooth. The bars signify the differential between the two lines (voting population and registered voters). Of particular note is the incredibly small differential in 2004 when the total registrations were just 80,000 below the voting age population; this year has been the subject of considerable discussion regarding irregularities in the voter registration database.

However, the issue here rests with the fact that population between Census counts is provided as only an estimate.  Our model takes the Census estimate which is a best guess of state population over 18 years of age and adjusts it to reflect an estimate at the time of the election.  In other words, an imputation built on an estimate built on an estimate.

Revised Model

These three problems, though, are realities of modeling and forecasting.  When we update the model to include real estimates from 2010 then re-project turnout one major, yet interesting, factor comes in to play:   the significance of several regression terms falls off substantially.  In particular, the unemployment rate and the dummy variable for a Senate race are no longer significant.  The most significant driver of the model results becomes whether there was a Presidential race in the election cycle.  In other words, most of the statistically significant drivers of variability in the projection are minimized and result in a voter turnout rate 5 percent higher than that realized in the 2010 election, mostly a result of the problems noted above.

So what does this revised model show for 2012? Assuming a 9 percent unemployment rate and a voting age population in line with the Census Bureau's current 2012 estimates, the model currently projects 2.9 million voters in the 2012 election cycle and a turnout rate of 64 percent. The rate is much higher than 2010, but then again, it will be a presidential election cycle.

How accurate will this projection be? We'll let you know after the election, after all, forecasts are just best estimates.  We can be certain of one thing, it will be an interesting election cycle for Missouri with an incumbent President and Senator both vying for reelection.

Related Posts by Category: