For reasons of cost, logistics, or the ability to compare vote choice across multiple congressional districts, survey researchers often make use of generic ballot questions when polling races for the U.S. House of Representatives. However, it isn’t clear that generic ballot questions give results that are fully comparable to questions that use the names of candidates. Given the prevalence of generic ballot questions in analyses of House elections, understanding the source and extent of this bias is vital. We explore the source and extent of this bias through a series of survey experiments carried out during 2010 Congressional elections and a historical analysis of bias in House elections.
Generic ballot experiments
To test the extent of bias resulting from the use of generic ballot questions, respondents in a statewide survey of Delaware carried out by Fairleigh Dickinson University’s PublicMind poll were randomly assigned to receive either generic congressional question (“I know it’s early, but in the upcoming race for the House of Representatives, do you plan to support the Republican or the Democratic candidate?”) or the same question with the names of the candidates. Delaware makes an ideal test case, as there’s only one House seat: a race that pitted a novice Republican, Glen Urquhart against a well-known former Lieutenant Governor, John Carney, Jr. Urquhart had won his primary race by <1 point; Carney had run unopposed. Carney also outspent Urquhart by a margin of $2.1 million to $1.4 million.
When respondents were given the generic ballot question, Carney led by 5 points, a margin of 45 to 40. When the candidates were named, however, Carney’s lead expanded to 15 points, 51 to 36. In this case, a survey that relied on the generic ballot question would have dramatically underestimated Carney’s lead.
Perhaps this is an extreme example. To check, we ran the same experiment in a statewide poll of New Jersey. In 2010, New Jersey Republicans, bolstered by Tea Party candidates, ran a higher than usual number of qualified candidates against Democrats who had only won their seats in the 2008 wave election. In the generic ballot question, Democrats outpolled Republicans by 5 points, 42 to 37. The generic ballot question, Democrats outpolled Republicans by 5 points, 42 to 37. When the candidates, including the relatively strong Republican challengers, were named, however, the Republicans outpolled the Democrats by 4 points, 41 to 37 (the 9-point swing is significant; t=3.1; p<0.01). While the direction of the effect goes toward Republicans in one state and Democrats in the other, the overall story is clear. Highly qualified candidates tend to outperform the generic ballot, while less qualified candidates tend to underperform the generic ballot.
Historical analysis
If it is the case that the degree and direction of bias in the generic ballot question is a function of candidate selection, the pattern should be evident in national polls as well. Moreover, we should be able to predict bias in the generic ballot question as a function of the same factors that lead to relatively strong or relatively weak candidates. Past research indicates that qualified challengers are highly risk averse, looking for signs of weakness in the candidate, or favorable national trends before running (Berkman and Eisenstein 1999; Bianco 1984; Groseclose and Krehbiel 1994; Jacobson and Kernell 1983). While individual level characteristics of an incumbent or race are of no use in predicting national level generic ballot bias, elements of the national political environment should be.
For this analysis, we compared the results of the last available generic ballot question (in most cases, about a month before the election) with the actual election results for all of the House elections since 1954 (a total of 29 elections) – all elections for which polling data is available. Election results are from federal election records, adjusted to reflect two-party vote shares. While the exact wording of the question differs (for instance, in a period in the 1950s, Gallup asked the question using a box in which respondents placed their answers), there does not appear to be any systematic differences by time period.
We model the bias based on the factors that a prospective candidate would take into account when deciding whether or not to run for the House in a given year: the president’s approval rating (ranging from 32 to 76, with a mean of 56), how well the president’s party did in previous elections (ranging from 41 to 57 percent of the vote, with a mean of 49), and the interaction of the two. We ran separate analyses based on the values of these variables 3, 6, 9, 12, 15, and 18 months before the election.
Together, these results were used to calculate the size and direction of the bias, with the results of the generic ballot corrected to reflect the two party vote, and the direction coded such that positive values mean that the president’s party did better in the actual election than in the generic ballot (Figure 1).
Because our dependent variable is at least theoretically unbounded (the difference between actual and generic ballot vote ranges from 0 to 9 points, mean of 2.24, standard deviation of 2.55), and can take on positive or negative values, simple OLS regression seemed appropriate for the analysis.
From the results of the analysis, it seems clear that the bias in the generic ballot question is strongly predicted by the performance of the president’s party in the prior election, and the approval of the president 12 to 18 months before the actual election, with the strongest results coming at a lag of 15 months (full results in Table 1). This lag corresponds to August of the year before the election, about the time that candidates would have to decide whether or not to run.
Positive values of the dependent variable represent bias in the favor of the president’s party – the extent to which the generic ballot understates how well the president’s party will do in the House elections. The results indicate that the potential for bias is greatest when the president’s approval is high. High presidential approval 15 months before the election, combined with a poor performance of the president’s party in the previous House election, leads the generic ballot question to significantly overstate how well the president’s party will do in the election. In 1960, for instance, President Eisenhower had a 67 percent approval rating 15 months before the election, and Republicans had only earned 44 percent of the vote in the 1958 election. Generic polls had predicted that Republicans would get 40 percent of the two-party vote; they earned 45 percent. The popularity of the Republican president encouraged qualified Republicans to run for office, and the poor showing of Republicans in the previous election meant there may have been more qualified Republicans not already in Congress.
In most elections, of course, the bias is rather smaller, averaging 2.6 points over the elections studied, and the results indicate that when the presidential approval is at 40 percent, bias is expected to be almost zero, regardless of the party’s performance in the previous election. Similarly, when the party earned close to 50 percent in the prior House election, there’s expected to be almost no bias, regardless of what the president’s approval is before the election (predicted bias for various values are in Table 2).
Conclusions
These results give us insight both into what causes the bias in generic ballot questions and how to correct for it. The bias seems to result from the fact that, in some years, one party runs candidates who are better than generic members of the party in the minds of voters. When they do so, their party outperforms the generic ballot.
Knowing this, we can estimate how much bias is expected to be present in the generic ballot questions in any given election, based on political conditions 15 months prior to the election. For the 2014 election, the model leads us to expect a relatively small bias of 0.62 points in favor of the Republican Party. Any estimate of seat distribution in the House based on national generic ballot questions would do well to take this into account.