EXISTING EVIDENCE ON THE ELECTORAL EFFECT OF HOUSE MEMBERS' VOTING

Almost every existing study that finds an electoral effect of legislative voting focuses on a small set of elections. These include the studies by Schoenberger (1969) for 1964, Erikson (1971) for Republicans in the 1950s–1960s, Bernstein (1989) for 1978 and 1980, Johannes and McAdams (1981) for 1978, Erikson and Wright (1993) for 1990, and Brady et al. (1996) and Jacobson (1996) for 1994.³

This research provides a critical basis for considering the importance of legislative voting in affecting electoral outcomes but, as a whole, does not establish that members' actual probability of seat loss is affected by their roll-call decisions, particularly for so-called safe members. In general, results on electoral margins are not equivalent to results on seat loss because, among other reasons, members behave differently if they are minimizing their probability of defeat rather than seeking the largest number of votes.⁵

Even the contention that legislative voting systematically affects electoral margins is put into question by work that analyzes elections across several decades. For example, political economics research on members' ideological “shirking” from 1960 through 1990 has produced inconsistent findings. In studies that estimate shirking as the absolute difference between interest-group voting scores and the predicted values from these scores regressed on constituency characteristics, some evidence has suggested that shirking is correlated with electoral defeat (e.g., Wright 1993), while other evidence has suggested that the factors are uncorrelated (e.g., Goff and Grier 1993). One possible reason for the discrepancy is that these studies equate shirking with noise that could result from omitted variables; to the extent that the constituency characteristics included in a particular analysis do not fully capture district preferences, shirking is overestimated and the effects of it underestimated.⁶

Other research that examines elections over time, but that does not characterize itself as studying shirking, also produces inconclusive results as to whether legislative voting systematically affects elections. Gaines and Nokken (1999) analyze all midterm elections between 1958 and 1994 and find that in almost every year, an incumbent's roll-call support for the president did not significantly affect his probability of reelection. Brady, Canes-Wrone and Cogan (2000) find more evidence of an impact, showing that for the 1954–1994 elections, winning incumbents tended to have more moderate voting records than losing incumbents. This difference, however, is significant for only about one-third of the elections.The most compelling evidence of an electoral effect over time comes from Ansolabehere, Snyder and Stewart (2001) and Erikson and Wright (2000). Ansolabehere, Snyder and Stewart pool the elections of 1952–1974 and 1976–1996 to compare candidates' voting records in races in which a sitting incumbent was defeated and find that a candidate's vote share was lower the more she supported the extreme of her party. The key limitation of the analysis is that it does not examine the roll-call voting of members who are never unseated, and therefore, as the authors acknowledge, it remains possible that most House races are not affected by legislative voting. Erikson and Wright do examine all incumbents for the elections of 1976–1996 and similarly find that voting with the ideological extreme of one's party decreases one's vote share. Their analysis does not, however, control for standard electoral determinants such as challenger quality and campaign spending, and thus it remains possible that the effect of legislative voting would disappear once such controls were added to the analysis.⁷

In sum, the literature is far from conclusive that legislative voting systematically affects House electoral outcomes. It offers some evidence that roll-call voting, particularly voting with the extreme of one's party, affects electoral margins, but this relationship has not been established across time and members. Moreover, even the existing support does not show that legislative voting affects the probability of reelection, particularly for safe members. What is needed, therefore, is a study that examines the relationship between vote shares and roll-call decisions across elections for all districts and that explicitly analyzes the effect of the decisions on the probability of winning for safe as well as marginal members.

LEGISLATIVE VOTING AND ELECTORAL MARGINS

Building on previous research, we estimate the electoral impact of legislative voting by analyzing the effect of voting with the extreme of one's party, a phenomenon we refer to as roll-call ideological extremity. We hypothesize that controlling for district ideology, Democrats lose electoral support by voting more liberally and Republicans by voting more conservatively. Thus for two Democratic (Republican) incumbents from districts with identical voter preferences, the Democrat (Republican) with the more liberal (conservative) voting record should have a lower electoral vote share, holding all else equal. We refer to this prediction as the Roll-Call Ideological Extremity Hypothesis.

While much of the previously discussed work on electoral margins details the theoretical framework underlying the Roll-Call Ideological Extremity Hypothesis (in particular, see Erikson 1971), we briefly review the logic here. Arguably the most critical assumption is that the Republican and Democratic candidates of each district diverge ideologically, with the Democrat being more liberal than the Republican. Evidence of such divergence is provided by Ansolabehere, Snyder and Stewart (2001), and other studies justify this phenomenon by describing the long-run career benefits a member may obtain from voting with his or her party (e.g., Aldrich 1995; Cox and McCubbins 1993; Rohde 1991; Snyder and Groseclose 2000). Thus even on roll-calls for which a member's electoral incentive is to vote against her party, she can have long-term career incentives to vote with it.

Applying Downsian (1957) logic to this stylized fact of ideological divergence between the Republican and the Democratic candidates in a district, ideological moderation should increase an incumbent's vote share. Given Downs' assumptions that voters can order the candidates on an ideological spectrum, that voters have positions on the spectrum, and that electoral choices depend upon candidates' positions, either candidate can increase her vote share by moderating towards the other. Thus if voters' assessment of the incumbent's ideology is not independent of her legislative voting, her electoral margin is higher the less she votes with the ideological extreme of her party.⁸

Notably, while this framework assumes that voters select candidates on the basis of ideology, the Roll-Call Ideological Extremity Hypothesis does not require voters to be fully informed about all details of an incumbent's legislative record. The literature identifies several mechanisms by which roll-call voting should affect constituents' assessments of an incumbent's ideology absent their complete knowledge of his or her record.⁹

Model, Data, and Measurement for Analysis of Electoral Margins

To test the Roll-Call Ideological Extremity Hypothesis, we employ an econometric model that is similar to those adopted in previous studies of the relationship between legislative voting and electoral margins. In particular, we regress each incumbent's vote share on a measure of roll-call ideological extremity, controlling for a range of factors.¹³

where the variables in brackets are naturally included only in tests that pool across elections. Because the Roll-Call Ideological Extremity Hypothesis does not imply asymmetric effects across parties, we analyze Democratic and Republican incumbents jointly, with the parties in a given year separated by the factor In Party. The substantive results still hold, however, if the effect of roll-call ideological extremity or the entire model is estimated separately for each party.¹⁴

The measurement of each variable in the model is as follows.

Incumbent's Vote Share. As is common in studies of electoral margins (e.g., Jacobson 1996), the dependent variable equals the incumbent's percentage of the two-party vote.¹⁶

Roll-Call Ideological Extremity. We base our measure of legislative voting on Americans for Democratic Action (ADA) scores, which reflect the proportion of liberal positions taken by a member in a given year on votes selected by the Americans for Democratic Action. These votes generally encompass the key policy issues of a session. For example, the scores include the major tax and budget votes of the 1980s and 1990s, the abortion and school prayer votes of the 1970s and 1980s, and the Civil Rights votes of the 1960s. The ADA ratings are widely known among members and often publicized in their districts. Therefore, to the extent that legislative voting affects members' electoral fate, the scores should have this effect. The specific measure that we adopt provides a consistent ranking of ideological extremity across the parties and is based on a member's ADA score in the year prior to the election if the member is Democratic and 100 minus this score if the member is Republican. Thus for a Democrat, a higher value indicates a more liberal member, while for a Republican, it reflects a more conservative one.¹⁷

Presidential Vote. Following Erikson and Wright (1989, 1993, 1997) and Brady et al. (1996), among others, we use the presidential vote to control for district ideology. Specifically, Presidential Vote equals the average proportion of the two-party vote received by the presidential candidate of the incumbent's party in the two most recent elections in his or her district, with each combination of presidential elections normalized around its mean. Thus for Democratic incumbents, the variable is based on the vote for the Democratic presidential candidate, and for Republican incumbents, on the vote for the Republican presidential candidate. We average across the two most recent elections and normalize each of these averages, to mitigate the impact of factors that may be specific to individual presidential races.

Challenger Quality. Research suggests that an incumbent's vote share is lower when she faces a challenger who has previously won an elected position (e.g. Jacobson and Kernell 1983). We therefore control for challenger quality with a variable that equals 1 if the challenger has held elective office and 0 otherwise.

ln(Challenger Spending) − ln(Incumbent Spending). A great deal of work has examined the electoral effects of candidates' spending, and while the findings vary, the literature as a whole suggests that a candidate's spending increases his vote share. As in some previous research (e.g., Erikson and Wright 1993), we control for this effect with a variable reflecting the difference between challenger and incumbent spending, taking the natural log of each. Our data are from Federal Elections Commission (FEC) Reports. Because these reports were not edited by the FEC until 1980, we include the control for spending only in the 1980 through 1996 elections.¹⁹

Freshman. This dummy variable equals 1 if the House term was the incumbent's first and 0 otherwise. Research suggests that there exists an incumbency advantage due to, among other factors, members' capacity to bring home pork and perform constituency service (Mayhew 1974a). Because freshmen generally have less political clout than more senior members, for example, less influential committee assignments (Munger 1988), freshmen should have a lower incumbency advantage and thus a lower average vote share.

ΔPersonal Income (coded by In Party), Presidential Popularity (coded by In Party), Midterm Loss (coded by In Party), and In Party. We include several factors that vary over time²⁰

We include each of these factors as a single term that is coded by affiliation with the in party rather than without this coding but with interaction terms according to affiliation with the in party because the latter approach would create a high degree of multicollinearity.²²

Results on Electoral Margins

We first analyze Eq. (1) for each election of 1956 through 1996.²⁴

Table 1 presents the central results. In particular, the table describes the parameter estimates on the key variable Roll-Call Ideological Extremity, reporting White's robust standard errors because the Cook and Weisberg (1983) test rejects the null of homoskedasticity in almost every election. We have chosen to make the parameter estimates on the control variables and the regression diagnostics available upon request to conserve space, particularly because these findings largely replicate those presented subsequently (and more succinctly) for the pooled-election analyses.

As Table 1 shows, the results of the individual-election analysis strongly support the Roll-Call Ideological Extremity Hypothesis. In all of the elections, an incumbent receives a significantly lower electoral margin the more he votes with the extreme of his party holding constant the range of control variables listed in the table. Thus comparing two Republican members who have identical district ideologies, who are running against the same quality of challenger, who are not freshmen, and who face the same distribution of spending between themselves and their challengers, the member with the more conservative voting record obtains a lower vote share. Similarly, the results suggest that between two otherwise equivalent Democratic incumbents, the one with the more liberal voting record will receive the lower vote share. In each year, this effect is statistically significant at the conventional level of p = 0.05 (two-tailed).²⁶

To interpret the size of this effect, we consider a 25-point shift in a member's ADA score toward the extreme of his party, a shift that approximates one standard deviation of the data, or 3 to 4 high-profile votes.²⁷

Given the consistent significance of the relationship between roll-call decisions and vote margins, we proceed to estimate the average impact of this relationship. Pooling the data, we control not only for factors that vary only within an election but also ones that vary across time, as specified in Eq. (1). Because the control for candidate spending exists only for 1980–1996, we conduct two tests. The first examines these later years with the control, and the second all years without it.

Table 2 presents the results. As with the individual-election analysis, the Cook and Weisberg (1983) test suggests evidence of heteroskedasticity, and White's robust standard errors are therefore reported.²⁹

Moreover, this effect has a magnitude comparable to that of other commonly recognized electoral determinants. To interpret the average impact of ideologically extreme voting, we again focus on a 25-point shift in a member's ADA score toward the extreme of his party, which, as mentioned previously, approximates a standard deviation. The coefficients for each sample indicate that such a shift would decrease a member's vote share by approximately 2 percentage points. In comparison, the 1980–1996 results suggest that an incumbent's vote share is lower by 2.5 percentage points when she faces a high-quality challenger, by 1 percentage point if the ratio between the challenger's and the incumbent's spending increases by 25 percentage points, and by 1 percentage point if she is a freshman. Even the coefficient on the midterm loss phenomenon has a comparable impact; off-year elections are estimated to lower the vote share of members of the president's party by 3 percentage points. The 1956–1996 coefficients are similar, with a high-quality challenger lowering a member's vote share by approximately 5 percentage points, a freshman receiving a lower vote share of 2 percentage points, and off-year elections lowering the vote shares of members of the president's party by 2 percentage points.

As these results suggest, the effects of the control variables are generally consistent with our expectations. The coefficients on challenger quality, presidential vote, candidates' spending, freshmen, personal income, and the midterm loss phenomenon all have the predicted signs with statistical significance (p < 0.05, two-tailed). The only control variable that has an unexpected effect is presidential popularity, and even though the factor was included merely to prevent overestimating any impact of legislative voting, we still investigated why the variable did not have the same effect as in the study by Tufte (1975). Employing Tufte's specification, which regresses seat change across the parties (rather than individual members' vote shares) on presidential popularity and personal income, we found presidential popularity to have the predicted impact, suggesting our finding derives from differences across the specifications. Specifically, the Tufte effect does not hold once the control for spending is added or if the dependent variable represents vote share rather than seat loss; our subsequent analysis on the probability of reelection demonstrates this variation in the impact.

Before moving on to this analysis, however, we first summarize the key findings in Tables 1 and 2. The results are important because previous work had left open the possibility that legislative voting affects members' electoral margins only for particular elections. By establishing a consistent effect in every year, we have foreclosed this possibility. Moreover, we have shown that the magnitude of the effect is not insubstantial but rather comparable to that of other widely recognized electoral determinants such as campaign spending, challenger quality, and the midterm loss phenomenon.

LEGISLATIVE VOTING AND THE PROBABILITY OF REELECTION BY DISTRICT SAFETY

The previous section still leaves open whether most members should be concerned that legislative voting might affect their probability of defeat. On the one hand, a majority of incumbents win by the so-called safe margin of at least 60% of the two-party vote, and the findings indicate that a shift from perfectly moderate to perfectly extreme voting alters a member's vote share by only 4 percentage points. On the other hand, the findings indicate that safety itself is likely to be dependent upon members' prior legislative records, and this relationship would suggest that even safe members might need to fear the electoral ramifications of legislative voting. For example, the results on electoral margins would lead us to expect that between two otherwise similar incumbents, the one with the more moderate record should be likelier to win by a safe margin, yet if that member were subsequently to shift toward her party's extreme, then her probability of winning reelection could significantly decrease. In fact, to the extent that a member is safe due to the moderation of her prior voting, a change toward more ideologically extreme voting could have as large of an effect as for a marginal member, whose lack of safety may derive from having voted ideologically extremely.

To assess whether members should indeed fear that roll-call voting might affect their prospects for reelection, we conduct a final test that estimates the relationship between legislative voting and the probability of reelection directly. In the test, we account for the possibility that safe members may not need to be concerned with the electoral ramifications of legislative voting even if marginal members do. Moreover, we incorporate that safety itself may depend upon members' prior voting records.

Results on the Probability of Reelection

We estimate the two-equation system by maximum likelihood and, following our previous tests, employ robust standard errors.³³

We used the package TSP 4.5 to maximize the likelihood function, which Kimhi (1999) defines.

Although the equations are estimated jointly, we describe the results separately for presentation purposes, beginning with the first-stage equation that estimates the safety of the incumbent. Table 3 states these results. It is immediately apparent from the table that safety is a function of the factors we previously found to determine incremental changes in vote shares. In other words, when a member is described as safe, this descriptor refers not to an exogenously imposed invulnerability but, instead, to an outcome based on political conditions, some of which are in the member's control.

Most importantly for our purposes, the results indicate that a member is significantly more likely to be designated safe the more moderate is his legislative voting holding other factors equal. Interpreting the probit coefficients at the means of the independent variables, we focus as before on a 25-point shift in an incumbent's ADA score toward the extreme of his party, a shift that approximates a standard deviation.³⁴

The standard deviation in members' ADA scores is 24.96 points. This standard deviation is close but not identical to that of the data in Table 2 since the current analysis excludes freshmen running in election t.

Specifically, such a change decreases a member's probability of holding a safe seat by 6% for each sample of the data. So-called safety is therefore not independent of legislative voting, suggesting that safe members may well need to fear the electoral ramifications of roll-call decisions.

The control variables generally have the same effect on safety as they did on marginal changes in vote share. The most notable difference is in the 1980–1996 sample: the coefficients on Midterm Loss and Freshman are no longer significant at p = 0.05 (two-tailed) in this sample. These effects are still in the expected direction, however, and that on freshman remains marginally significant (p = 0.06, one-tailed).

Turning to the main results, those on the probability of reelection, Table 4 presents them by the estimated regime of seat safety. In particular, the table describes the parameter estimates of each electoral determinant for a safe member versus a marginal one. The findings show that even safe members face a significantly lower probability of reelection when they increase the extent to which they vote with the extreme of their party. For both safe and marginal members and in each sample of the data, an incumbent's likelihood of retaining her seat decreases the more ideologically extreme is her voting.

These results support the long-standing findings that members view legislative voting as an important component of the electoral connection and consider constituency cues in roll-call decisions. Perhaps most notably, the results show why such legislative behavior is consistent with the fact that most incumbents win reelection by safe margins. Even safe legislators' roll-call voting affects their risk of electoral defeat. Thus members are correct in assuming that legislative votes have an impact on the probability of reelection.

Comparing the coefficients on Roll-Call Ideological Extremity across regimes, the one for marginal members is higher than that for safe members in the test with the control for campaign spending, while the reverse occurs in the other test. Neither difference, however, is statistically significant (p > 0.4 in each case). This result is consistent with the first-stage findings that safety is a function of members' prior voting records. As we argued at the outset, to the extent that safety depends upon roll-call decisions, they may have a similar impact on the probability of winning for safe and marginal members.

In terms of the magnitude of the coefficients, interpretation is less straightforward than for even a typical probit analysis because the means of the variables differ across the regimes of safety. To describe the magnitudes similarly by regime, we interpret the probit coefficients at specified parameter values, beginning with the following ones, which were chosen for their moderate values: 0.75 for roll-call ideological extremity (an ADA of 75 for a Democrat, 25 for a Republican); a difference of 0 between the district vote for the presidential candidate of the incumbent's party and the average vote for that candidate across the districts; 1 for challenger quality, midterm loss, and in party; 50% for presidential popularity; and 0 for the other variables. At these values, a member who shifts his ADA score 25 points toward the extreme of his party decreases his probability of reelection by 11% in the marginal regime of the 1980–1996 test, 9% in the safe regime of that test, and 4% in each regime of the 1956–1996 test. Altering the values such that the incumbent does not face a high-quality challenger, the magnitudes stay relatively constant, with such a shift in an incumbent's roll-call voting decreasing his probability of reelection by 9% in the marginal regime of the 1980–1996 test, 8% in the safe regime of this test, 3% in the marginal regime of the 1956–1996 test, and 2% in the remaining regime. Likewise, reducing the initial value of roll-call ideological extremity does not substantially alter the results. For example, at 0.50 (an ADA of 50 for each member) a 25-point shift in a member's ADA toward his party's extreme decreases his likelihood of reelection by 8% in each regime of the 1980–1996 test and by 3 to 4% in each regime of the 1956–1996 test.

As with the results on vote shares, the impact of roll-call voting is comparable to that for other electoral determinants. For example, at the initial parameter values of the previous paragraph, the 1980–1996 test suggests that an incumbent's probability of winning would decline by 5 to 7% if the ratio of challenger to incumbent spending increased by 25%. Similarly, each test indicates that a high-quality challenger decreases an incumbent's probability of reelection by 6 to 15%. The impact of legislative voting on the probability of reelection is thus not only statistically significant for safe as well as marginal members, but also similar to the effect of factors that are commonly recognized to be important determinants of electoral outcomes.

In general, the results on the control variables comport with our expectations. The effects of campaign spending and district ideology are consistently in the expected direction and statistically significant. Those on challenger quality also have the correct sign in each regime and sample and are significant with the exception of the marginal regime of the 1980–1996 test. In addition, the coefficients for the remaining variables that are not included as a main effect typically have the predicted sign, and they are significant only with the expected sign. Specifically, the effect of the economy is significant in each regime of the 1956–1996 test, the effect of the midterm loss is significant in the marginal regime of each test, and the effect of presidential popularity is significant in the marginal regime of the 1956–1996 test.

Overall, the results in Tables 3 and 4 are important because they show that a typical member should rationally be concerned with the electoral impact of legislative voting. Controlling for a wide range of other electoral determinants, roll-call voting has a significant effect on the probability of reelection. Notably, this effect holds for not only marginal but also safe members. These results are consistent with our finding that safety itself is a function of members' voting. That is, safe members are “safe” partially as a consequence of their roll-call decisions. Consequently, when such members change their voting pattern toward their party's extreme, they face a significantly higher probability of defeat, just as other members do.

CONCLUSION

While surveys of members have consistently found that they are concerned with the potential electoral ramifications of legislative voting, previous work had not established that this concern is necessarily rational. Surveys of the mass public have suggested that the typical voter is fairly ignorant of her representative's policy decisions, and studies of elections themselves have provided mixed results as to whether legislative voting affects electoral margins. Moreover, even the positive evidence has left open the possibility that roll-call decisions may not affect the likelihood of defeat for the majority of incumbents. The literature has thus failed to provide evidence that a typical member should believe that he is accountable to voters with regard to his roll-call decisions.

Our study presents three types of evidence to this end. First, we show that, holding district ideology constant, in every election between 1956 and 1996 an incumbent's vote share decreased the more he voted with the extreme of his party. Second, pooling across the years, we establish that the average impact of this effect is comparable to that of commonly recognized electoral determinants such as challenger quality. Third, by directly examining the probability of reelection, we demonstrate that the probability decreases significantly as an incumbent's voting support for his party increases, with this effect holding not only for marginal but also safe incumbents. The vulnerability of the safe members occurs in part because moderate voting increases the probability of holding a safe seat or, in other words, because safety itself derives from a member's roll-call positions.

Notably, while we find evidence of House members' electoral accountability for their policy actions, we certainly do not conclude that voters are highly knowledgeable about all such actions. Earlier we noted several rationales for the seemingly distinct states of the world. For example, “uninformed” voters may take cues from informed elites or, alternatively, may become informed by challengers when an incumbent votes out of step with her district. Accordingly, future work might examine how elites' and challengers' behaviors influence the degree to which legislative voting affects electoral outcomes.

Our analysis also provokes a number of other issues for future research. One question is the extent to which the effect of legislative voting varies according to the salience of the policy issue. Research on legislative behavior finds that members are more likely to enact legislation reflecting public opinion when the policy issue is salient (Canes-Wrone 2001; Hutchings 1998; Kollman 1998). In combination with our findings, this research suggests that members' legislative votes regarding relatively salient issues may have larger electoral effects than votes on less salient matters.

A second extension would be to examine whether changes in the number of marginal seats derive in part from variation in party discipline. Standard explanations for the “vanishing marginals” of the 1960s–1980s emphasize member activities other than legislative voting, such as constituency service (Fiorina 1989). Meanwhile, research has found substantial variation in the influence of parties over members' voting throughout these decades (e.g., Snyder and Groseclose 2000). The evidence presented in this paper suggests that these phenomena of party discipline and the vanishing marginals could be related.