Issue Accountability in U.S. House Elections

For Republican incumbents the same logic applies to produce different predictions. Rather than liberal voters potentially being less responsive, under Republican incumbents it is conservative voters. And voter-MC disagreement may matter more when Republican incumbents take conservative positions than when they cast liberal roll call votes. This implies an additional set of comparisons. For voters with liberal preferences, the effect of an MC taking the conservative position instead of the liberal position will be greater for Republican incumbents than Democratic incumbents. For voters with conservative preferences, the effect of an MC taking the liberal position instead of the conservative position will be greater for Democratic incumbents. As explained below (and shown in Table 2), these propositions are not testable with the data analyzed in this paper because on the roll call votes considered there is only variation in the positions taken by MCs for one party or the other.

For Republican incumbents, p_inc(V_con, MC_con) − p_inc(V_con, MC_lib) = 0, and p_inc(V_lib, MC_lib) − p_inc(V_con, MC_lib) = 0.

Later I consider the possibility that liberal voters who disagree with Democratic incumbents and conservative voters who disagree with Republican incumbents may respond through abstention rather than voting for the other party’s candidate.

It is possible that Democratic opposition to these measures could have been based on beliefs that the measures were insufficiently liberal, in which case equating opposition to the measures with a conservative preferences would obviously be problematic. But, when the relationship between opposition and district-level presidential vote is analyzed, it is apparent that opposition to the measures is more likely in Democratic districts where the district-level presidential vote was more Republican, strongly suggesting that Democratic incumbents who voted against the measures did so because they believed the proposed policies were too liberal, not insufficiently so.

All of roll call votes considered in the empirical analysis were classified as “key votes” by Congressional Quarterly or the National Journal or both.

Treating respondents in the fifth category as missing does not alter the patterns of statistical results or substantive findings.

The analysis is based on major party, self-reported voters. Incumbents seeking reelection in uncontested elections are not included in the analysis. If the analysis is limited to the respondents for whom the CCES was able to validate as having voted, the patterns of results are the same.

Party identification is coded −1 (strong identifier of challenger’s party), −.5 (weak or leaning identifier of challenger’s party), +.5 (weak or leaning identifier of incumbent’s party), or +1 (strong identifier of incumbent’s party). Pure independents and other non-identifiers are coded 0. Ideology is also coded on a −1/+1 scale and is the average of two questions, one asking respondents to place themselves on a 7 point ideology scale and one asking them a 5 point version. General policy preferences are measured with 13 policy questions, all coded to range from −1 (most liberal response) to +1 (most conservative response) and then averaged (alpha reliability coefficient = .91). To facilitate interpretation, when analyzing voting for Democratic incumbents, ideological identification and general policy preferences are scaled so higher values indicate more liberal preferences. When analyzing voting for Republican incumbents, they are scaled so higher values indicate more conservative preferences. To preserve cases and avoid listwise deletion of missing data, respondents with missing data on the ideology and policy questions (before averaging) were assigned to the median values of those variables.

In preliminary analyses, several district-level covariates were included (presidential vote and challenger quality), but there were no substantively or statistically significant relationships between them and vote choice. As a result, in the model results reported in Tables 3 and 4 in the Appendix only individual-level control variables are included. In addition, re-estimating the models without the sampling weights produces the same patterns of results.

Specifically, the estimated probabilities are the average estimated effects for the survey respondents. Sometimes effects are estimated for a hypothetical individual with average (or median or modal) values for the independent variables. In the present case, though, this would produce exaggerated effects because that sort of voter (one whose partisanship and ideology is very near the middle of the scale) typically has a probability of voting Democratic (Republican) near .50, which is exactly the place on the logit curve where logit changes produce the largest probability changes.

Based on the parameter estimates reported in Table 3 in the Appendix, the estimated effect of voter-MC agreement versus disagreement among conservative voters is 1.60 (8). The estimated effect of voter-MC agreement versus disagreement among liberal voters is − .55 (7). Thus the estimated difference in the two effects is 2.15.

Among Democratic incumbents who voted for the ACA, the difference in the logit parameter estimates for voters who agreed and disagreed with the incumbent is 1.08 (9). The estimated difference among incumbents who voted against the ACA − .03 (10), producing a difference in the estimated effects of 1.11.

Among Democratic MCs, those who cast a liberal roll call vote on one measure were more likely to cast liberal votes on the others measures. Among voters, those who expressed a liberal preference on one issue were more likely hold liberal preferences on the other issues. Multicollinearity among voter preferences is less of a concern due to the large number of voters in the sample (N = 17,988), but there are considerably fewer Democratic incumbents (N = 232). To investigate this issue further, I re-estimated the voting model four times, each time including only one of the issues and excluding the other three. None of the sets of model estimates provides consistent support for the voter-MC preference congruence model by finding support for Eqs. (1) through (4). And the ACA issue remains the one where support for the voter-MC conditional preference congruence is strongest. The estimated difference in differences for Eqs. (5) and (6) are largest in magnitude for the ACA and the only ones for which both reach conventional levels of statistical significance (p < .05).

The estimated effect for Republican incumbents is somewhat higher, approaching 50 points.

Political knowledge is measured in terciles based on a scale with ten items like party control of the House and Senate, placing president Obama and the Democratic party on the liberal side of an ideological scale, and placing the Republican party on the conservative side of the scale.

Further, in “Congressional races, a centerpiece of the Republican strategy has been to use Democratic ‘yes’ votes on the law to tar incumbents as advocates of expansive government and lockstep followers of their party’s leadership” (Sack 2010b).

For 11% (78 of the 726), p ≤ .05.