Dynamic estimation of ideal points for the US Congress

Abstract

Theories of candidate positioning suggest that candidates will respond dynamically to their electoral environment. Because of the difficulty of obtaining “bridge votes”, most existing approaches for estimating the ideal points of members of Congress generate static ideal points or ideal points that move linearly over time. We propose an approach for dynamic ideal point estimation using Project Vote Smart’s National Political Awareness Test to construct bridge votes. We use our dynamic estimates to measure aggregate change, to measure individual-level change, and to study the institutional and structural factors that explain the changing positions of House candidates and members of Congress. We demonstrate that while the Republican Party has been selecting increasingly extreme candidates, Democratic incumbents have become more extreme while in office. We also find that the congruence between elected members of Congress and their constituents is mostly explained by the selection as opposed to the responsiveness of the candidate. Nonetheless, we find evidence of dynamic responsiveness of incumbents in specific circumstances. We find that competitiveness, midterm elections, and sharing the president’s party affiliation are associated with greater responsiveness. Conversely, retirement is not associated with a change in responsiveness. We find no evidence of responsiveness of challengers. Finally, we find that close elections draw challengers who are more in line with the district’s ideology.

Appendix

In this section, we provide additional details about the National Political Awareness Test (NPAT). We focus on the bridge observations that allow us to estimate dynamic preferences for US House candidates. We employ two types of bridge. First, bridge votes allow us to connect the NPAT responses from different years. Table 5 reports the number of bridges between each year of the NPAT and for other years. At most times, many bridge items connect a particular year to other years (with 1992 being the exception). In 1992, seven bridge votes emerge between that year’s NPAT and the other NPATs. With the exception of 1992, the number of bridges has ranged from around 20 to more than 100. Moreover, even when only about 20 bridge votes could be identified, the policy issues were quite varied. For example, in 2016, the bridge votes included items on abortion, taxes, campaign contributions, the death penalty, federal government involvement in education, global warming, the Keystone XL pipeline, healthcare, immigration, foreign involvement by the US military, social security, gay marriage, and terrorism.

Table 5 Bridges between biannual NPATs

The second type of bridging is between the NPAT and roll call voting in the US House. The bridge voters we use are for individual members who voted in a particular session of the US House and responded to the NPAT (or had their responses coded by Project Vote Smart) in the election following that session. The number of bridge voters is reported in Table 6. In 1992, 30 bridge voters were identified. Since then, the number of bridge voters always has exceeded 50 and, in most cases, has exceeded 100. Starting with the 2010 election, Project Vote Smart (PVS) began coding the positions of non-respondents. That information is, however, not archived. We have a large number of bridge voters in 2016 because we collected the data from the PVS website after the 2016 election and PVS coded responses for essentially all House candidates. Between 2010 and 2014, we have fewer responses because, although PVS coded the responses for essentially all candidates, for candidates who ran in later years, the NPAT coding would have be overwritten and lost (contrarily, candidate responses to the NPAT survey are available for prior years).

Table 6 Bridges between the NPAT and the House of representatives

In the paper, we report results based on a one-dimensional model. In Table 7, we report the fit of one- to three-dimensional models applied to the entire data matrix and for the NPAT alone. We consider two measures of fit—the percentage of items correctly predicted and the geometric mean probability. Using a one-dimensional model, we can correctly predict 89.4% of responses in the data matrix. Adding a second and third dimension helps only marginally—adding a second dimension allows us to correctly predict an additional 1.2% of items. The percentage correctly predicted is somewhat lower for NPAT items alone—we can predict 80.0% of the items correctly. For the NPAT alone, a larger share of the item responses are from challengers and third-party candidates; the responses of challengers and third-party candidates are somewhat less predictable. In addition, the improvement in fit from a two-dimensional model was somewhat larger for Libertarian candidates.

Table 7 Dimensionality of candidate preferences