Thursday, June 7, 2012

Reviewing Early General Election Polls

As part of Elephant Watcher's shift from the Republican primary to the general election, we have added a page devoted to national match-up polls between Barack Obama and Mitt Romney. The Polls page will be updated frequently, adding national polls as they are released, and averaging the most recent of them. Though polling data during the Republican primary was occasionally sparse, there will always be plenty of polls released on the 2012 presidential election. The trick is in reviewing all of this data in the most intelligent manner possible. In today's post, we will explain how Elephant Watcher will present the polling data during this election season, as well as the reasons behind it.

The Polls page will only include national polls. This is in contrast to many political analysts, who spend their time on examining the polls from swing states. Since the 2000 election, an increasing amount of attention has been paid to state polls, rather than national--because the winner of a presidential election is determined by the electoral college. Earlier, we offered an in-depth explanation of why it's much better to follow national polls rather than state polls. To summarize, state poll analysts are forced to rely on less accurate polling, fewer polls, out-of-date polls, less-reputable pollsters, and must cobble together electoral map projections from multiple, interdependent state outcomes. National polls bypass each of these problems. Unless the election is decided by less than 1% nationally, the winner of the popular vote is virtually guaranteed to win the electoral college anyway.

Having decided to track national rather than state polls, there are still a few things to consider. Some polls are listed as polls of "likely voters," while others are "registered voters" or "adults." Polls of likely voters involve what are called "likely voter screens," which adjust the polls based on which participants are most likely to vote in November. To do this, pollsters ask participants whether they intend to vote, but also ask other questions to determine whether they are likely to turn out (e.g. past voting history, voter enthusiasm, etc.). Polls of registered voters do not.

The difference matters. In presidential elections, Republican voters consistently turn out to vote at a higher rate than their Democratic counterparts. This means that registered voter polls will consistently overestimate the strength of the Democratic candidate. This is not an expected or potential bias--it is a known bias. By the time the election is imminent, nearly all pollsters will switch to polling likely voters rather than registered voters in order to eliminate the bias and increase accuracy.

Early on in the election season, however, many pollsters will not yet employ the likely voter screens. Why not? Because the closer you get to the election, the easier it is to determine whether someone is likely to vote. Often, pollsters will begin the season by polling registered voters, and at some point down the line, they make the switch. Other pollsters will use a likely voter model from the beginning.

If we were to review all of the polling data at face value, likely voter polls would be mixed in with registered voter polls. There would be a consistent gap between the accurate likely voter polls and the inaccurate registered voter polls, which overestimate the Democrat's chances. Later, as the pollsters switch to polling likely voters, the Republican candidate would suddenly increase in the polls, showing momentum where perhaps there is none. Obviously, this is not an ideal situation.

To correct for these problems, Elephant Watcher will adjust the registered voter polls, marked by an asterisk (*), by three points in favor of the Republican candidate. To avoid confusion, the results will be listed the same; only the R+ or D+ number in front of the poll will be changed. (For example, a registered voter poll listed "Obama 45, Romney 46" will be counted as R+4 instead of R+1.) Polls of likely voters will be unaffected. The adjustment is three points, based on the average difference between registered and likely voter polls--diminished somewhat because it assumes a slight Republican bias on the part of Rasmussen polls.

Note that this won't necessarily produce good polling results for Mitt Romney: If Barack Obama is doing well in the polls, he will be ahead of Romney in both likely voter and registered voter polls. Note also that this is a temporary measure, since nearly all pollsters will be switching to likely voter polls in the coming months.

Since it's always better to use multiple polls rather than looking at one poll, the Polls page will include a poll average. This will be a simple average of the ten most recent polls. One final adjustment is that "daily tracking polls," which release polls every day (e.g. Gallup and Rasmussen) will only be added only once every three days. This is done for two reasons. First, the poll average would otherwise be overwhelmed by the large number of polls being released by Gallup and Rasmussen. Second, daily tracking polls include overlapping data. For example, a tracking poll might include polling data from Monday through Wednesday; the next day, from Tuesday through Thursday--reusing the same Tuesday and Wednesday data. Entering daily tracking polls only once every three days solves both problems.