constituent errors in assessing their senators

21
Public Choice 83: 251-27t, 1995. © 1995 Kluwer Academic Publishers. Printed in the Netherlands. Constituent errors in assessing their Senatars* THOMAS A. HUSTED Department of Economics, The American University, Washington, DC 20016-8029 LAWRENCE W. KENNY Department of Economics, University of Florida, Gainesville, FL 326l 1-2017 REBECCA B. MORTON Department of Political Science, University of Iowa, Iowa City, IA 52242-1409 Accepted 3 June 1993 Abstract. We attempt to explain why some constituents are well informed and others are poorly informed about the positions taken by their U.S. Senators. The acquisition of political information is modeled in a Bayesian framework. A constituent with virtually no information about a candi- date assigns him an average position on a liberal/conservative spectrum. As more political infor- mation is acquired with more political involvement, the constituent shifts her prior toward the poli- tician's actual position and thus has a smaller error in assessing positions taken by her representative. In the Bayesian framework, voters make larger errors in evaluating the records of party mavericks than of typical party members. The model is tested using data from the 1982 American National Election Study. Data on the respondent's perceived ideology of their Senators and their socioeconomic characteristics are corn- bined with information on the Senators' actual ideology, length of time in the Senate, political party, and candidacy for re-election in 1982. The empirical analysis provides support for our predictions. 1. Introduction 1 Many models of electoral processes use a principal agent approach. Candidates compete for votes through their choice of policies and the candidate whose po- titicies are preferred by more voters is elected to office. If the legislator wants to be reelected, his subsequent behavior must not disappoint his principals who voted him into office, The success of the electoral threat hinges crucially on an electorate well in- * We are grateful to an anonymous referee, John Aldrich, David Denslow, Bernie Grofman, Jinook Jeong, Laura Langbein, Mark Rush, and seminar participants at the University of Florida, SUNY-Stony Brook, 1990 Public Choice Meetings, and the 1990 American Political Science As- sociation Meetings for their helpful comments.

Upload: nyu

Post on 13-Jan-2023

2 views

Category:

Documents


0 download

TRANSCRIPT

Public Choice 83: 251-27t, 1995. © 1995 Kluwer Academic Publishers. Printed in the Netherlands.

Constituent errors in assessing their Senatars*

T H O M A S A. H U S T E D

Department of Economics, The American University, Washington, DC 20016-8029

L A W R E N C E W. K E N N Y

Department of Economics, University of Florida, Gainesville, FL 326l 1-2017

R E B E C C A B. M O R T O N

Department of Political Science, University of Iowa, Iowa City, IA 52242-1409

Accepted 3 June 1993

Abstract. We attempt to explain why some constituents are well informed and others are poorly informed about the positions taken by their U.S. Senators. The acquisition of political information is modeled in a Bayesian framework. A constituent with virtually no information about a candi- date assigns him an average position on a liberal/conservative spectrum. As more political infor- mation is acquired with more political involvement, the constituent shifts her prior toward the poli- tician's actual position and thus has a smaller error in assessing positions taken by her representative. In the Bayesian framework, voters make larger errors in evaluating the records of party mavericks than of typical party members.

The model is tested using data from the 1982 American National Election Study. Data on the respondent's perceived ideology of their Senators and their socioeconomic characteristics are corn- bined with information on the Senators' actual ideology, length of time in the Senate, political party, and candidacy for re-election in 1982. The empirical analysis provides support for our predictions.

1. Introduct ion 1

M a n y models o f electoral processes use a pr incipal agent approach . Candida tes

compete for votes th rough their choice o f policies and the candida te whose po-

titicies are prefer red by m o r e voters is elected to off ice. I f the legis lator wants

to be reelected, his subsequent behav io r must not d isappoint his pr incipals who

vo ted h im into off ice,

The success o f the electoral threa t hinges crucial ly on an electorate well in-

* We are grateful to an anonymous referee, John Aldrich, David Denslow, Bernie Grofman, Jinook Jeong, Laura Langbein, Mark Rush, and seminar participants at the University of Florida, SUNY-Stony Brook, 1990 Public Choice Meetings, and the 1990 American Political Science As- sociation Meetings for their helpful comments.

252

formed about the legislator's behavior in office. There is, however, some skep- ticism regarding the efficiency of the electoral mechanism. This is because the individual voter has little incentive to vote, much less to study the political op- tions, since she is unlikely to be a swing voter. We attemt to shed some light on the viability of the electoral process by examining how accurately consti- tuents place their Senators on a liberal/conservative scale.

There is a large literature focusing on the extent to which the elected representative's policy and voting decisions are motivated by personal ideolo- gy, rather than by constituent interests. 2 These studies have generally found that the legislator's ideology plays an important role in his votes on social issues such as abortion and school prayer and on weakly monitor legislation such as general spending bills. 3 Our examination of why some constituents have more information than others can be used to provide a better understanding of the determinants of ideological voting. The legislator's ideology is expected to play a smaller role among representatives with well informed constituencies.

Section 2 examines the factors that determine the constituent's ability to evaluate accurately the politician's ideology. We test this model in the third sec- tion using data from the 1982 American National Election Study. Section 4 concludes.

2. A model of constituent errors in evaluation

Previous analyses have revealed that many constituents are poorly informed about the positions taken by their elected representatives and that there is con- siderable variation in the size of the errors that constituents make in their evalu- ations. 4 Converse (1970) suggested that the errors are largely due to a lack of political sophistication. He contended that there are two types of respondents: elites with "attitudes", who, because of superior knowledge, are better able to evaluate candidates, and a larger group of voters with "nonattitudes", who respond to questions concerning candidate positions randomly. In contrast, Achen (1975) and Erikson (1979) suggested that the apparent error may be a result of the vagueness of the questionnaire. Consequently, they argued that political sophistication is not highly correlated with the error, because it mostly reflects the inherent measurement problems of the survey.

Palfrey and Poole (1987) constructed an index of voter information using es- timates of the true positions of candidates and parties constructed through the Aldrich-McKelvey (1977) scaling procedure. The Aldrich-McKelvey procedure corrects for the measurement problems in the survey questions and allows for a construction of estimates of the true positions of a candidate and the ideal points of the respondents from the survey responses. The index constructed by Palfrey and Poole (1987: 53) is "an average measure of correlation between the

253

(estimated) true location of candidates and parties and a voter 's (reported) per- ception of these locat ions." They found that voter information was positively correlated with "ideological extremism;" that is, voters with more extreme po- sitions on the issues were more likely to have accurate information on the true positions of the candidates and parties. Moreover, informed voters are much more likely to vote and their voting decision is more "predictable" according to a standard spatial model given their personal positions.

In a related study, Powell (1989) found that contributors to political cam- paigns are more accurate in specifying the " t r u e " positions of candidates than other constituents who are less involved. She used interest group measures of the candidates' positions as estimates of their true locations on the issues. Moreover, Powell notes that even the error of non-contributors can be partially explained by factors that affect political information collection such as consti- tuent education levels, age, length of residence, and candidate seniority. Powell 's work suggests that most observed guessing contains some knowledge of the true candidate position and does not merely reflect nonattitudes. In this paper, we explore the acquisition of this knowledge using a Bayesian approach.

It is well known that the incentive for voting and other constituent involve- ment is extremely small and that the benefit f rom obtaining better information about candidates is correspondingly meager (see, for example, Downs, 1957; and Becker, 1983, 1985). This reasoning is consistent with evidence that the majority of voters are not well informed. Although it is a stylized fact that con- stituents have very little information about candidates, there are few formal theoretical models that describe constituents' acquisition of information. 5

The standard approach that we adopt and extend in this paper is to assume a Bayesian learning procedure through which voters update their evaluation of candidates with new information. Calvert (1980) demonstrated that such a process has a well defined solution for the constituent when there is a positive cost to information, acquisition and voter utility from candidate positions is bounded.

In our model, a constituent who has virtually no information about a candi- date assigns him a position on a liberal-conservative scale that is based initially on easily observable political characteristics. For example, the constituent could assume that the candidate is as liberal as the average Senator, the average Senator in his party, or the average Senator in his party and region (e.g., a Southern Democrat). We assume that the constituent estimates the true posi- tion of, say, the Senator 's political party (GENERAL). As the constituent gathers more and more information specific to the Senator, she shifts her prior away from GENERAL toward the Senator's actual ideology. The constituents' priors are assumed normally distributed with a mean of GENERAL and a finite variance equal t o ffi 2.

It is probably true that any Senator-specific information that constituents

254

receive has quite a bit of noise in it. For example, candidates are often accused of misleading constituents in political advertisements either by refusing to take a stand - so-called "ambiguous behavior" - or by outright deception. 6 Con- sequently, each constituent also observes "no i sy" signals of the Senator 's actu- al rating. We assume that the signals can be either " t ru th fu l " o r " false." Each truthful signal is assumed normally distributed with mean ACTUAL and a finite variance air 2 and each false signal is assumed normally distributed with mean GENERAL and a finite variance air2. Therefore, with Bayesian up- dating in a noisy information environment, the constituent's estimate of the Senator 's ideology is a weighted average of 1) her estimate of the average poli- tician's position and 2) the average position indicated by the information the constituent has received about the candidate or

PREDICT = or(ACTUAL) + (1 -a ) (GENERAL) (1)

where ot = (Nt/ai2)/(Nt/ait 2 + Nf/aif2 + 1/ai 2) and 0 _< o~ < 1; N t is the num- ber of truthful signals about the politician's position that have been received by the voter; and Nf is the number of false signals about the politician's posi- tion that have been received by the voter. As the voter receives more of the truthful (false) signals, the weight placed on the actual or true position, o~, rises (falls). 7

The constituent's error in her prediction is

ERROR = A C T U A L - PREDICT (2)

Substituting (1) into (2) yields

ERROR = A C T U A L - [otACTUAL + ( 1 - o 0 G E N E R A L ] ,

o r

ERROR = (1 - ~)[ACTUAL - GENERAL]. (3)

The absolute value of this error is

IERRORI = (1 - ot) l ACTUAL - GENERAL I. (4)

Thus, the constituent will make smaller errors if she is better informed or if the Senator 's voting record is closer to what she would have predicted without knowing very much about the politician. It follows that there will be smaller errors associated with the voting records of the standard party members than with the positions taken by party mavericks, all other information about the Senators equal.

255

Recent experimental research suggests that an adaptive model of voter infor- mation acquisition such as the one just outlined is appropriate. Collier, Or- deshook, and Williams (1989) found that constituents place strong emphasis on retrospective information in forming evaluations, particularly when candi- date positions are stable over time.

As shown in equation (1), a is inversely related to the variance of the original predictors and is a function of the characteristics that determine the demand for political information. These characteristics can he subdivided into voter and politician specific characteristics, or

= f ( v , c ) , (5)

where V and C are vectors of voter and Senator characteristics, respectively. Substituting (5) into (4) yields the relationship in equation (6) that is to be em- pirically estimated:

IERRORI = [ACTUAL - GENERALI - f(V,C) ] A C T U A L - GENERAL I (6)

The constituent's demand for political information is derived from her desired level of political participation. That is, it makes little sense for the con- stituent to acquire information about politicians if she does not expect to vote or to contribute money to one of the candidates. In fact, it seems likely that voter participation and political information gathering are jointly determined events and our focus in this paper is on those characteristics that would deter- mine either voter participation or political information gathering. We are, however, estimating a reduced form model of political information gathering. Consequently, we do not include any direct measure of political participation (e.g., voting or campaign contributions) as an exogenous variable, since this is endogenous in our model specification. Nevertheless, we can draw on the large literature of the determinants of political participation to find the rele- vant voter characteristics (V) to include in this reduced form model of political information gathering.

Political participation is inversely related to its cost. The cost of voting is measured by the respondent 's income (a measure of the time cost of voting), age (a measure of capability and experience), population density (a measure of closeness to polling locations as well as access to political information through larger media markets), length of residence in the area (a measure of knowledge of local affairs), and whether the respondent belongs to a union or is married (miscellaneous indicators of shared knowledge). Education is also an impor- tant determinant of political participation. It is well documented that educa- tion raises skills and, in particular, raises allocative abilities (see Schultz, 1975).

256

Constituents with greater allocative skills are better able to distinguish between competent and inept candidates and to select those candidates who will best represent their interests. Accordingly, those with more education face a lower cost of acquiring good political information and become better informed. It is also well documented that Blacks have attended poorer quality school sys- tems, particularly before schools were integrated. For this reason, Blacks are expected to be less informed than comparable whites, even if the attained educational levels are the same.

Candidates can influence the benefits from participating by offering plat- forms that discriminate among groups through the progressive tax system and through government spending levels. Filer, Kenny, and Morton (1993) report- ed that voter turnout depends on the group's position on the income distribu- tion. Similarly, homeowners have a greater stake in voting because public poli- cies can produce capital gains or losses for them.

The demand for specific information about the candidates depends not only whether the constituent votes but also on whether the constituent ever switches parties when voting. Suppose that voters support the candidate whose platform is closest to their preferred point, which is monotonically related to income. Republican platforms are almost always more preferred by the very rich, while the poor generally support Democrats, because of their historical emphasis on income redistribution. Middle income voters sometimes switch parties, de- pending on the individual party platforms. For example, a number of these voters favored Johnson over Goldwater and, in a later election, Nixon over McGovern. As a result, middle income voters, who are potential switchers, are more likely to acquire additional political information to assist them in their party and candidate selection.

As postulated above, the constituent's acquired information also depends on the politician's characteristics (C). Constituents might be expected to acquire more information about a representative up for re-election than about those not facing re-election. However, as discussed above, candidates may try to mis- represent or obfuscate their political positions or voting records, if they have not been good agents of their constituents. Shepsle (1972), Alesina and Cukier- man (1990), Chappell (1989), Harrington (1992a, 1992b), and Reed (1989) have examined motives underlying this behavior. Thus, if a candidate up for reelec- tion sends out more "false" signals than "truthful" ones, then o~ will fall when a candidate is up for re-election causing voter errors to increase. Consequently, it is uncertain what effect a re-election campaign has on the constituent's ability to acquire candidate information.

Likewise, the effect of the representative's length of tenure in office on the constituent's demand for political information is also complex. Certainly, voters have received more information over the years about individuals who have been in office a longer time. But, more senior legislators have greater

257

clout in Congress; accordingly, the specific positions they take on many issues are less important to their constituents, since these legislators bring more Fed- eral largesse to their districts. 8

3. Empirical analysis

3.1. Data

The empirical test of the model focuses on the constituent's evaluation of her two U.S. Senators. The individual data used for the test come from the 1982 American National Election Study (hereafter, ANES), a nationally representa- tive pre- and post-election survey of 1418 adults. Among the several socioeco- nomic and political questions, each respondent is asked to rate in a series of questions her elected representatives on a liberal to conservative scale. This ser- ies begins with the following:

We hear a lot of talk these days about liberals and conservatives. Here is a seven-point scale on which the political views that people hold are arranged from extremely liberal (1) to extremely conservative (7). Where would you place yourself on this scale, or haven't you thought much about this?

Of the original group of respondents, 471 individuals answered that they "ha- ven't thought much" about rating themselves or their elected representatives on a liberal-conservative scale and therefore are excused from the remaining questions in the series. 9 The remaining individuals are asked in two subse- quent questions to rate their U.S. Senator(s) and U.S. Senate candidates (if the incumbent was not up for re-election) on this same seven-point scale - (1) ex- tremely liberal to (7) extremely conservative. In order to correspond with a liberal voting rating system, these answers were converted so that the ANES scale ranges from (1) extremely conservative to (7) extremely liberal. The respondent 's modified answers to these questions are used to measure the vari- able PREDICT, which was defined in equation (1). 1°

Of the 897 individual responses to the initial screening question, 790 adults f rom 36 states evaluated at least one of their U.S. Senators on the liberal/con- servative scale. 11 From these 790 individuals, 149 are deleted because of miss- ing socioeconomic data, leaving 641 respondents rating at least one Senator. The fact that respondents were not asked to evaluate Senate incumbents not seeking re-election (if their term ended in 1982) resulted in 152 of these respon- dents rating only one incumbent Senator. In total, there are 1130 (641 + 489) evaluations (PREDICT) on 71 Senators from the ANES post-election respon- dents. Along with PREDICT, the respondent's age, education level, income,

258

race, sex, union membership, marital status, housing situation, and own ideol- ogy are obtained from the ANES. The respondent-specific data are combined with information on the 71 individual Senators, including their length of time in the senate, political party, and whether they were up for re-election in 1982.

As in Powell (1989), we use the Americans for Democratic Action (ADA) rating to compute the various estimates of the Senator's true position, ACTU- AL. It is one of several voting measures that are generated by special interest groups. These ratings reflect the percentage of votes on specific issues cast by the politician that agreed with the policy positions of the particular interest group. In addition to the liberal ADA rating, a conservative rating is compiled by the Americans for Constitutional Action (ACA) and additional ratings are developed by labor unions, environmental groups, and energy groups, among others. These ratings are highly correlated, and Kau and Rubin (1979), Peltz- man (1984) and others have obtained similar results using several different rat- ing measures. 12 This evidence is consistent with the empirical findings of Feld and Grofman (1981), Poole (1981), and Poole and Rosenthal (1989) that only one dimension is generally necessary to characterize legislators' positions or preferences. According to Poole and Rosenthal (1989: 22), "the House is cur- rently virtually unidimensional." Therefore, since each of these ratings places legislators on the same dimension, it does not matter which measure is used.

The ADA ratings potentially range from 0 (most conservative) to 100 (most liberal). Unfortunately, one weakness of this rating is that it counts an absence as a vote for conservatism. Since we are examining voter information, it seems more appropriate to treat absenses as neither a vote for conservatism nor a vote for liberalism and we have modified the ADA ratings accordingly. For exam- ple, in 1980 Edward Kennedy missed 12 of the 18 votes used by the ADA to the rating. Without adjustments, he would be given a conservative rating of 33~ Since Kennedy's other votes were in line with the ADA, we would adjust his 1980 ADA score to a 100. After similar adjustments are made for all Senators, we then calculate each Senator's average 1981 and 1982 ADA scores. These average ADA ratings range from 0 for conservative Senators Helms and East to 100 for liberal Senators Dodd, Levin, and Tsongas.

One final modification is required to make each Senator's ADA rating com- parable with the ANES respondent's PREDICT. Because the ADA ratings potentially range from 0 to 100, they must be scaled to correspond to the 1-7 rating that was taken from the ANES. One potential complication with the straightforward scaling of ADA ratings is that Senators holding extreme posi- tions are generally unlikely to get elected or reelected. Consequently, Senators with extremely liberal or conservative senate voting records may not be per- ceived by their constituents to be very liberal or conservative. As evidence of this, the mean values of constituents' ANES rating for each Senator range be- tween 1.89 and 4.80.13 Powell (1989) also found that contributors to congres-

259

sional campaigns who say they know the incumbent's position on issues very well tended to rate their representatives between 2.0 and 6. t on the same 7 point scale. 14 Accordingly, for the bulk of our empirical analysis we scale the actual ADA ratings to take on a narrower range of values between 2 and 6 [TADA (2-6)], rather than 1-7.15 In one case, we do transform the ADA ranking to take on values on the 1-7 scale [TADA (1-7)]. 16

One potential drawback to our analysis is that the scales used by the respon- dents in the ANES might not be comparable with the adjusted ADA scales. In order to check whether these scales are equivalent, statistical tests for the equivalence of the means and variances of the Senator's transformed ADA rankings [both TADA (1-7) and TADA (2-6] and the predicted ideological ranking from the observations in the ANES [PREDICT] are performed. With TADA (1-7) the value of the test statistic in a test for the equality of means is 0.263 and is 1.13 with TADA (2-6). Since the critical value for the t- distribution at the 5% level of significance is 1.96, the null hypothesis that the means of the Senator's transformed ADA ratings and PREDICT are equal cannot be rejected in either situation. The calculated statistic for the test of the equality of the variances is 1.68 using the 1-7 transformation and 1.34 with TADA (2-6). With a critical point for the F distribution equal to 1.35 at the 5% level of significance, the null hypothesis that the variances are equal is on the borderline for TADA (1-7) and cannot be rejected with TADA (2-6). The results from these statistical tests are supportive that the average respondent in the ANES is using the same scale as the actual ADA rankings. In addition, the positive Pearson correlation coefficients between the two TADA measures and PREDICT (0.313 for TADA (1-7) and 0.309 for TADA (2-6)) indicate a positive relationship between the actual ADA ratings and the constituent's prediction.

However, it might still be the case that a respondent with an extreme ideolo- gy is using a different scale to evaluate her Senators than is used by the ADA. For example, "extreme liberals" may rate all other positions to their political right (leaning liberal, moderate, and conservative) as "conservative". If this occurred, voters with more extreme preferences would have larger evaluation errors. Alternative, Palfrey and Poole (1987) find that more extreme voters are more likely to be better informed. The absolute value of the difference between the respondent's own ideology rating and the "middle-of-the-road" rating of 4 (ABOWNID) is included in the empirical analysis to control for these poten- tial effects.

The definitions and summary statistics for all individual and Senator specific variables are listed in Table 1.

260

Table 1. S u m m a r y s ta t is t ics

Var iab le Descr ip t ion M e a n Var iance

Individual characteristics

I N C O M E F a m i l y i ncome 25729 15695

A G E Age in years 43.481 16.020

E D U C E d u c a t i o n in years 13.534 2.520

T E N U R E Years at in te rv iew 20.449 18.676

ci ty or t o w n

R E N T Rents = 1; o the r = - 1 - 0.409 0.913

B L A C K Black = 1; o the r = - 1 - 0.869 0.495

F E M A L E Fema le = 1; m a l e = - 1 - 0,067 0.998

M A R R I E D M a r r i e d = 1; o the r = - 1 0.335 0.943

U N I O N F a m i l y m e m b e r in u n i o n = 1; - 0 . 5 4 3 0.840

o ther = - 1

L O W D E N 1 i f cen t ra l c i ty > 2 mi l l ion 3.627 1.458

2 i f cent ra l c i ty < 2 mi l l ion

3 i f subu rb o f t

4 i f subu rb o f 2

5 i f ad jacen t to me t ro a rea

6 i f ou t ly ing a rea

A B O W N I D A b s o l u t e va lue o f the 1.052 0.909

d i f fe rence be tween the

r e sponden t ' s own ideo logy

and 4 (Modera te , Midd le

o f the R o a d )

P R E D I C T C o n s t i t u e n t ' s conserva t ive 3.584 2.193

ra t ing for he r two U.S.

Sena tors f r o m 1

(ext remely conservat ive)

to 7 (ex t remely l ibera l )

Senator characteristics T A D A ( 1 - 7) Average 81 - 82 A D A ra t ing

( 1 - 7) 3.645 3.690

T A D A ( 2 - 6) Ave rage 8t - 82 A D A ra t ing

(2 - 6) 3.763 1.640

D E M D e m o c r a t = 1;

Repub l i can = - 1 0.115 0.994

S O U T H D E M Sena tor is Sou the rn and - 0 . 6 9 4 0.720

Democra t i c = 1; o ther = - 1

N O N S O U T H D E M Sena to r is n o n - S o u t h e r n and - 0.191 0.982

Democra t i c = 1; o ther = - 1

T E R M Years in senate 8.658 7.100

R E E L E C T Sena to r r an for re-e lect ion - 0.375 0.927

in 1982= 1; o t h e r = - 1

G O V R A C E G o v e r n o r ' s race in 1982 = 1 0.540 0.842

other = - 1

A B D I F A b s o l u t e va lue o f 1.378 1.136

( T A D A - P R E D I C T )

261

3.2. Empirical model

As described in the theoretical section of the paper, one focus of our paper is to determine the measure that best describes the initial position assigned to a Senator by a constituent with virtually no information about her Senator (GENERAL). Three alternative measures of GENERAL are considered:

1. The average ADA score for 1981 and 1982 for all Senators. 2. The average ADA score for 1981 and 1982 for all Senators by political party

(Democrat/Republican). 3. The average ADA score for 1981 and 1982 for all Senators by political party

and region (South/non-South).

The first measure implies that the constituent assigns an average ADA score to each of her Senators and does not even use the Senator's political party to rate him on a liberal-conservative scale. The second measure is based on the assumption that uninformed constituents assign to each of their Senators the average ideological position taken by all Senators in his party. Finally, the third measure assumes the most sophisticated behavior of Bayesian voters. For it to be appropriate, uninformed constituents must start off by assuming their Sena- tor to be as liberal as the average Senator in his party and region. Following equation (4) in the theoretical section, the variables DIFFALL, DIFFPAR, and DIFFREG are the absolute values of the difference between the Senator's transformed average 1981, 1982 ADA rating [TADA], measuring ACTUAL, and each of the three proposed measures of GENERAL.

As described above, the respondent in the ANES must place her Senators on a discrete seven point scale. Our empirical analysis concentrates on the absolute value of the difference between this discrete rating and the Senator's actual continuous mean 81-82 ADA ratings [ E TADA-PREDICT I l, which is the con- stituent's "evaluation error". However, to be consistent with the 1-7 integer choices provided to the respondents for evaluating their Senators in the ANES, this evaluation error is rounded off to the closest integer. The resulting depen- dent variable (ABDIF) takes on six integer values (0-5) with TADA (2-6) and seven values (0-6) with TADA (1-7), with a 0 indicating a perfect evaluation in both cases.

Table 2 shows the simple frequency relationship between ABDIF and DIFFALL. Twenty-two percent of the evaluations involve no error, 41 percent are off by 1 (out of 7), 21 percent are off by 2, 10 percent by 3, and 5 percent are off by 4. The column percentages and marginals indicate that larger values of DIFFALL are associated with larger constituent evaluation errors, so it seems larger errors tend to be recorded, as predicted, for party mavericks than for party regulars.

262

Table 2. Table of DIFFALL by ABDIF (2 - 6) (column frequencies are in parentheses)

ABDIF

DIFFALL 0 1 2 3 4 5 Total

0.0 38 57 30 7 0 0 132 (15) (12) (13) (6) (0) (0) (12)

0.5 37 85 23 9 1 0 155 (15) (18) (10) (8) (2) (0) (14)

1.0 49 117 58 18 4 0 246 (20) (25) (25) (15) (7) (0) (22)

1.5 75 125 52 38 11 4 305 (30) (27) (22) (32) (20) (36) (27)

2.0 48 84 69 45 39 7 292 (19) (18) (30) (38) (71) (64) (26)

Total 247 468 232 117 55 11 (22) (41) (21) (10) (5) (1)

Note: GENERAL equals mean 1981-82 transformed ADA rating for all Senators. DIFFALL has been rounded.

In our empirical analysis, we consider two related models o f consti tuent in-

fo rmat ion . First, we explain the const i tuent ' s initial response to whether she

has thought much about political ideology. A binary dependent variable (RESPONSE) is defined as 1 if the consti tuent rates her own ideology on the

1 - 7 scale and 0 if she has " n o t thought much about [ ideology]" . Since this

original quest ion asks about the individual 's evaluat ion o f her own ideology,

only individual-specific characteristics are assumed to determine this decision.

A simple probit model is used to estimate the model parameters . Co lumn (1)

in Table 3 lists the parameter estimates for this model , where R E S P O N S E is

the dependent variable. We next examine the determinants o f A B D I F ( 2 - 6 ) and A B D I F ( 1 - 7 ) ,

which directly measures how much in fo rmat ion the respondent has about her

Senator ' s ideology. The creat ion o f this variable has already been described and, because it takes on ordered and discrete values, an ordered-response

model or an ordered probi t model is the appropr ia te est imation technique for these regressions, iv The observed values o f A B D I F are assumed to take one o f

these discrete values if the underlying unobserved value o f A B D I F (ABDIF*) falls within cells defined by threshold parameters that are u n k n o w n and to be

estimated. For example, A B D I F -- 0 if A B D I F * _< 0, A B D I F -- 1 if 0 <

A B D I F * _ ~1, and so on. The estimated coefficients and threshold parameters (/zi) f r o m five ordered

probi t regressions are reported in columns 2 - 6 o f Table 3. The first three regressions incorpora te one o f the three priors described above to fo rm D I F F A L L , D I F F P A R , and D I F F R E G . In order to estimate equat ion (6) f r o m

263

Table 3. Parameter estimates for response model (RESPONSE) and error o f evaluation regressions

(ABDIF)

(absolute values o f t-statistics in parentheses)

Ind . /Dep .

(1) (2) (3) (4) (5) (6) RESPONSE ABDIF ABDIF ABDIF ABDIF ABDIF

( 2 - 6 ) ( 2 - 6 ) ( 2 - 6 ) ( 2 - 6 ) ( 1 - 7 )

C O N S T A N T - 1.550 0.427 0,802 0,797 1.401 0.284

0 .75 ) (4.85) (12.98) (12.65) (5.04) (3.13) DIFFALL 1 - 0.791 - - - 0.930

(3.74) (6.53) DIFFPAR 2 - - 0.652 - - -

(2.38) DIFFREG 3 - - - 0.928 - -

(2.17)

Individual characteristics'

EDUC 0.120 - 0.020 - 0,029 - 0,045 - 0.031 - 0.021

(7.20) (1.97) (2.03) (2.07) (2.11) (3.11)

INCOME 0.235 - 0 , 1 1 8 - 0 . 1 4 4 - 0 . 0 3 2 - 0 . 1 2 9 - 0 . 0 9 8 (10 -4 ) (2.11) (2.06) (1.69) (0.26) (1.51) (2.61)

INCSQ - 0 . 2 1 9 0.145 0,155 0.010 0.178 0.123

(10 -.9 ) (1.20) (1.60) (1.20) (0.05) (1.34) (2.14)

F E M A L E - 0.209 0.042 0.011 0.005 0.035 0.013

(4.91) (1.83) (0.33) (0.10) (1.05) (0.82) BLACK - 0.047 0.044 - 0.012 - 0.003 0.112 0.066

(0.69) (1.04) (0.16) (0.02) (1.75) (2.22) U N I O N - 0 . 0 9 3 0.028 0.010 0.008 - 0 . 0 0 9 0.018

(1.75) (0.96) (0.24) (0.13) (0.21) (0.93)

M A R R I E D 0.074 0.041 0.028 0.018 0.054 0.017

(1.48) (1.59) (0.68) (0.30) (1.40) (0.94)

A G E 0.000l - 0 . 0 0 3 - 0 . 0 0 5 - 0 . 0 0 9 - 0 . 0 0 5 - 0 . 0 0 3

(0.05) (1.75) (1.80) (2.40) (2.15) (2.28) T E N U R E - 0.002 0.003 0.002 0.(KM 0.004 - 0.0001

(0.93) (t .83) (0.98) (1.39) (1.90) (0.08)

RENT 0.117 0,048 0.031 0.048 0,062 0.008

(2.15) (1.70) (0.75) (0.79) (1.45) (0.42) L O W D E N - 0.053 0.028 - 0.03 - 0.040 0.021 0.013

(I.91) (1.67) (1.22) (t.22) (0.86) (1.12) A B O W N I D - - 0.012 0.065 0.084 0.058 - 0.023

(0.5I) (1.73) (1.53) (1.65) (1.47) R E E L E C T - 0.131 0.102 0 . i10 0.111 0.047

(5.03) (2.92) (2.18) (3.05) (2.70) NON-

S O U T H D E M

S O U T H D E M

T E R M

0.072 0.281 0.333 0.223 0.074 (2.60) (7,19) (6.17) (6.04) (3.87)

0.024 - 0,776 - 0.026 - 0.098 - 0,027 (0.31) (1,35) (0.30) (1.78) (0.56)

0.009 0.019 0.022 0.010 -0 .0001

(2.62) (3.64) (2.81) (2.22) (0.04)

264

Table 3. Cont.

Ind . /Dep .

(1) (2) (3) (4) (5) (6)

RESPONSE ABDIF ABDIF ABDIF ABDIF ABDIF

( 2 - 6 ) ( 2 - 6 ) ( 2 - 6 ) ( 2 - 6 ) ( 1 - 7 )

G O V R A C E - 0.057 0.085 0.157 0.094 0.051

(1.92) (1.49) (1.98) (2.30) (2.52)

t~ 1 - 1.182 1.160 1.148 1.17t 1.022

(25.76) (25.72) (25.65) (25.71) (22.13)

~t 2 - 1.894 1.851 1.826 1.866 1.733

(32.44) (32.36) (32.66) (32.85) (30.35)

P,3 - - 2.552 2.482 2.440 2.494 2.351 (31.82) (32.29) (32.91) (32.46) (33.63)

/x 4 - 3.423 3.310 3.245 3.321 2.937

(24.42) (26.12) (26.16) (25.09) (32.93)

~5 . . . . . 3 . 8 0 4

(25.26) OBS 1124 1130 1130 1130 1130 1130

X 2 179.52 146.83 97.35 68.16 110.33 273.31

1DIFFALL = I T A D A (1 - 6) - G E N E R A L I (mean = 1.22)

GENERAL = average A D A

= 3.71

2DIFFPAR = I T A D A ( 2 - 6) - G E N E R A L I (mean = 0.81)

G E N E R A L = party specific average

= 4.71 for Democrats

= 2.86 for Republicans 3DIFFREG = 1TADA (2 - 6) - G E N E R A L I (mean = 0.55)

G E N E R A L = par ty/ region specific average

= 5.09 for Northern Democrats

= 3.03 for Northern Republicans

= 3.65 for Southern Democrats

= 2.28 for Southern Republicans

Note. All the individual characteristics, Senator characteristics, and state variables in regressions

(2)-(4) have been multiplied by DIbTALL, DIFFPAR, and DIFFREG, respectively. The variables

in (6) have been multiplied by DIFFALL with average A D A = 3.57.

the theoretical section, each of the voter characteristics variables in V and each of the Senator-specific characteristics variables in C must be multiplied by the included DIFFALL, DIFFPAR, or DIFFREG. The parameter estimates reported in columns (2)-(4) in Table 3 are from the three ordered probit models using TADA (2-6). The coefficients in column (5) are from a non- Bayesian model that assumes no prior is used by the constituent. In this model, none of the individual or Senator characteristics is multiplied by DIFFALL, DIFFPAR, or DIFFREG. Finally, the last column, (6), lists the parameter esti- mates from a fifth ordered probit regression that uses TADA (1-7), rather than TADA (2-6), and DIFFALL as the prior. 18

265

The results from the RESPONSE and ABDIF models can be jointly used to test the predictions of the underlying theoretical model. The overall empirical results provide strong support for the model's Bayesian framework. First, the Bayesian model predicts that there will be larger evaluation errors (i.e., larger values of ABDIF) associated with atypical Senators. That is, according to equation (6) in the theoretical section, an increase in the absolute value of the difference between ACTUAL and GENERAL should result in larger errors. This prediction is supported by the significant and positive coefficients on DIFFALL, DIFFPAR, and DIFFREG in the results reported in columns (2)-(4) and (6) of Table 3.

Second, by comparing the reported chi-square statistics, we get the best fit overall in the model where uninformed constituents are not sophisticated enough even to take advantage of party differences in ideology. That is, unin- formed constituents assume that their Senators have the ideology of the aver- age Senator (DIFFALL). This is also true in comparisons between the reported results in column (6) and unreported regressions using ABDIF (1-7). For ex- ample, the chi-square statistic for the ordered probit with ABDIF (1-7) without the prior is 190.95.

Finally, since we expect that those constituents who are most likely to have thought about political ideology are more likely to have smaller prediction er- rors, the signs on the coefficients in the simple probit RESPONSE model in column (1) are expected to be opposite those reported in the ordered probit AB- DIF models in columns (2)-(6). This is generally the case, particularly for those variables most closely related to political information acquisition.

Consistent with the expectation that an increase in skills results in more knowledge and smaller errors, we find that the constituent's level of education (EDUC) has a positive impact on her probability of being aware of ideology. Similarly, there is a consistently negative relationship estimated in the ABDIF regressions between EDUC and constituent assessment errors. Adams and Kenny (1986) found that states with higher educational attainment were less likely to place a restriction on gubernatorial tenure. Our results generally sup- port their reasoning that those who are more educated are better able to distin- guish the quality of positions taken by their elected representatives.

The relationship between income and constituent information is complex. Filer, Kenny, and Morton (1993) showed that the stakes in the election tend to be greatest for those at the upper and lower tails of the income distribution. On the other hand, the cost of voting also tends to rise as family income in- creases. On net, there is a positive relationship between family income and voter turnout in cross-sectional data. This implies that the derived demand for information about candidates also rises with income, leading to smaller en'ors. The demand for information about the individual Senator also depends on the constituent's party loyalty, which appears to be weakest for those close to the median-income voter.

266

The parameter estimates support this reasoning. The generally significant coefficients for both family income (INCOME) and its square (INCSQ) imply a quadratic relationship between political information and income. As income rises, respondents are more likely to be aware of ideology and have smaller evaluation errors. This is consistent with the turnout results. The effects of in- come level on ABDIFF diminish as income continues to increase. Families making $53,700 (in 1982 dollars) are most likely to be aware of ideology, and (using the results from the first ordered probit regression with prior DIFFALL) families making $40,800 have the smallest errors in assessing their Senators' ideology. Income has a positive impact on knowledge for most of our sample, since only about 20 percent have family incomes greater than $40,800 and only 7 percent make more than $53,700. The fact that there is a lower turning point for errors in evaluating individual Senators than for a general awareness of ideology is consistent with the acquisition of information by potential switch- ers being a middle income phenomenon. Subsequent increases in income are associated with a deterioration of this knowledge.

The respondent's gender (FEMALE) and race (BLACK) are important de- terminants of political information. Women are less likely to have thought about political ideology and, in our best fitting model reported in column (2), are more likely to have larger errors in evaluating their Senators. Although BLACK is insignificant in the RESPONSE probit and in the ABDIF ordered probit models based on the 2-6 variable, the significant coefficient for BLACK based on the 1-7 transformation implies that Blacks make larger mis- takes in assessing their Senators. Filer, Kenny, and Morton (1991) reported that Blacks areqess likely to vote, in part because of inferior schooling, so the finding of larger evaluation mistakes by Blacks is expected.

According to the definition of LOWDEN found in Table 1, population den- sity falls as this variable becomes larger in value. Consequently, it becomes harder to get to the polls, which discourages people from participating in polit- ics, and political information may not be as available as in a more densely populated area. The negative coefficient reported for LOWDEN in the RESPONSE model and the positive coefficient reported in the first ABDIF model (in column (2)) are consistent with this reasoning.

Of our remaining constituent characteristics, the respondent's age and hous- ing situation are the other significant contributors to his response and evalua- tion error. AGE is, as expected, negativdy related to evaluation error in all of the ABDIF models. TENURE and RENT are positively related in the model with DIFFALL in column (2). Renters have less at stake at election time, since they face no potential capital gains or losses and have been found to be less like- ly to vote than owners. The results in Table 3 are not fully consistent with this evidence. While renters have larger evaluation errors, they are more likely to have thought about ideology. The insignficance of UNION in the ABDIF

267

models does not necessarily indicate that it is an unimportant determinant of political information. Surprisingly, those who have family members in a union are less likely to have thought about political ideology.

The coefficients estimated for ABOWNID are generally insignificant and, therefore, it appears that there are no systematic differences in the abilities of the respondents to evaluate the Senators based on the respondent's ideology.

The results in Table 3 also illustrate the effects of regional differences, elec- tions, and the Senator's characteristics on the respondent's assessment error. The positive coefficients estimated for TERM indicate that Senators who have been in office for a longer time have larger assessment errors. This may be be- cause Senators with more seniority are less likely to lose and also bring home more Federal largesse. The positive and significant coefficients on REELECT in all of the ABDIF models suggest the ambiguous or deceptive behavior on the part of Senators standing for re-election in 1982. In fact, the positive coeffi- cients on both TERM and REELECT indicate that Senators up for re-election and in office for longer periods are likely to provide more" false" than "truth- ful" signals about their actual voting behavior.

The positive and significant coefficients reported for GOVRACE in the or- dered probit models are consistent with the presence of a Governor's race in the state in 1982 drawing additional attention away from the state's Senators. As a result, constituents from those states would make larger errors in evaluat- ing their Senators.

Finally, the insignificant coefficients on SOUTHDEM and the positive and significant coefficients reported for NONSOUTHDEM reflect that, relative to constituents of Republican Senators, respondents from nonsouthern regions with Democratic Senators (NONSOUTHDEM) make larger errors in evaluat- ing the ideology of their Senators.

The parameter estimates using ABDIF (1-7) reported in the last column (6) of Table 3 lend additional support to our Bayesian framework. In this regres- sion model, DIFFALL, constructed with TADA (1-7), is positively related to ABDIF. As with the other models, there is a U-shaped relationship between the respondent's income and ABDIF, and, as expected, EDUC and AGE are nega- tively related to ABDIF, while BLACK is positively related. GOVRACE, NONSOUTHDEM and REELECT remain positively related to the respon- dent's evaluation error.

4. Conclusions

We found that a Bayesian framework helps to explain the acquisition of politi- cal information by constituents. The prior that seems best to describe unin- formed voters assigns all Senators the average Senator's ideology. In that case,

268

the vo te r does n o t t ake in to accoun t p a r t y or reg iona l d i f ferences in Senate vot -

ing pa t t e rns . W e pred ic t and f ind tha t there a re la rger e r rors a ssoc ia ted wi th

p a r t y maver icks t h a n with p a r t y regulars . Al loca t ive skills, as measu red b y

educa t i ona l a t t a i n m e n t and age, p l ay an i m p o r t a n t role in reduc ing er rors in

assessing the pos i t ions t aken by Senators . The re la t ionsh ip wi th i ncome is con-

sistent b o t h wi th the fac t t ha t t u r n o u t rises wi th i ncome and wi th the observa-

t ion tha t cand ida te - spec i f i c i n f o r m a t i o n is m o r e va luab le to po ten t i a l p a r t y

switchers , who t end to be m i d d l e - i n c o m e class. These resul ts s u p p o r t the w o r k

o f Powel l (1989) b y d e m o n s t r a t i n g tha t the er rors m a d e by voters a re depen-

den t u p o n i n f o r m a t i o n acquis i t ion skills and tha t the extent o f " g u e s s i n g " is

no t s imply a func t ion o f pol i t ica l sophis t i ca t ion .

This analys is suggests t ha t cons t i tuen ts a re b o t h re la t ive ly u n i n f o r m e d a b o u t

c a n d i d a t e pos i t ions a n d d i f fe r in the i r inab i l i ty to eva lua te them. As discussed

in the I n t r o d u c t i o n , this lack o f i n f o r m a t i o n m a y a l low cand ida te s to engage

in ideo log ica l vo t ing or even a m b i g u o u s behav io r , since cand ida tes m a y f ind

it easier to vote accord ing to their ideo logy on issues tha t concern g roups o f

vo te rs who a re least capab le o f de t e rmin ing elected of f ic ia l s ' t rue pos i t ions .

M o r e o v e r , the d i f fe rences in vo te r abi l i t ies to eva lua te c a n d i d a t e b e h a v i o r can

inf luence the d i rec t ion o f such ideo log ica l vot ing . Such cons t i tuen t e r rors ,

then , can have i m p o r t a n t effects u p o n the eff ic iency o f the represen ta t ive pol i t -

ical system.

Notes

1. In order to clarify the discussion, Senators and constituents were assigned different sexes. The flip of a coin determined that Senators would be male and constituents female.

2. For examples of this large literature, see Kau and Rubin (1979), Kau, Keenan, and Rubin (1982), Mitchell (I979), Kalt (1981), Kalt and Zupan (1984, 1990), Peltzman (1984), Nelson and Silberberg (1987), and Dougan and Munger (1989).

3. For an examination of the effects of the electorate's information problems on the efficiency of legislative representation see Morton (1991). Using a rational expectations framework, McKelvey and Ordeshook (1985, 1987) have shown that under certain assumptions about the distribution of preferences, voter knowledge of these preferences, and the anticipated behavior of other voters, plurality maximizing candidates will still converge to an efficient (non- shirking) electoral equilibrium, However, the requirements on voter knowledge of the elec- torate's preference structure and the level of awareness of other constituents' behavior in- creases with the complexity of issues addressed by candidates (McKelvey and Ordeshook, 1987) and decreases the strength of the conclusion that efficient representation can be achieved in the face of an uninformed electorate.

4. Some researchers view the incomplete collection of information as rational. Voters are less likely to obtain information on politicians if the candidates' strategies are stable or if their vote will not decide the election (Collier, Ordeshook, and Williams, 1989). However, these candi- date specific characteristics cannot explain why some constituents are better informed than other constituents about the same candidates in the same race.

5. For a review of this literature, see Calvert (1986: 14-28).

269

6. Some recent models of the amount and kind of information that politidans reveal conclude that under various conditions it is rational for candidates not to reveal complete and/or truth- ful information. Calvert (1986) reviews this literature. Shepsle (1972) first explored equilibria involving candidates with ambiguous platforms.

7. We implicitly assume that the cost of information is such that no constituent chooses to be fully informed. Kmenta (1986) derives a similar expression for prediction based on information from an outside source.

8. Bennett and Mayberry (1979) documented the effect of seniority on the net benefits received by constituents from the Federal government.

9.50 other responses were coded as "missing". 10. As noted above, Aldrich and McKelvey (1977) pointed out that respondents could be assigning

different scales to the survey's seven point scale. There are several problems with implementing the procedure they developed to deal with this situation, however. First, their assumption that the individual's perceptions of the candidate's position are centered around the candidate's true position contrasts with our Bayesian assumption that the perceived position is a weighted average of the true and average positions, and is thus biased. Second, the t982 ANES data we use do not include enough evaluations from each respondent of a common set of candidates (e.g., Presidential candidates) to estimate the respondent-specific "regressions" that are need- ed to implement Aldrich and McKelvey's procedure.

t l . This initial 115 difference is due to "missing data". 12. The correlation between ADA and ACA ratings is 0.94 (Kalt and Zupan, 1984) and between

the ADA and the AFL-CtO's Committee on Political Education ratings in 0.90 (Peltzman, 1984).

13. The fact that the mean of PREDICT has a narrower range than TADA (2-6) is consistent with our hypothesis that some uninformed constituents ascribe average positions to their represen- tatives.

14. More precisely, she regressed the informed contributor's estimation of how conservative the representative is on the ACA score. The values given above come from evaluating the regres- sion at ACA equal to 0 and 100.

15. TADA (2-6) = 0.04 (ADA) + 2. 16. TADA (1-7) = 0.06 (ADA) + 1.

17. See McKelvey and Zovonia (1975) for a description of the ordered probit technique. 18. In this case the mean DIFFALL = 3.57.

References

Achen, C.H. (1975). Mass political attitudes and the survey response. American Political Science Review 69: 1218-1231.

Adams, J.D. and Kenny, L.W. (1986). Optimal tenure of elected public officials. Journal o f Law and Economics 29 (October): 303-328.

Aldrich, J.H. and McKelvey, R.D. (1977). A method of scaling with applications to the 1968 and 1972 presidential elections. American Political Science Review 71: 11 I - 130.

Alesina, A. and Cukierman, A. (1990). The politics of ambiguity. The Quarterly Journal o f Eco- nomics 105 (4): 829-850.

Becker, G.S. (1983). A theory of competition among pressure groups for political influence. Quarterly Journal o f Economics 98 (August): 371-400.

Becker, G.S. (1985). Public policies, pressure groups, and dead weight cost. Journal o f Public Eco- nomics 28 (December): 329-347.

270

Bennett, J.T. and Mayberry, E.R. (1979). Federal tax burdens and grant benefits to states: The impact of imperfect representation. Public Choice 34 (3/4): 255-269.

Calvert, R.L. (1980). The role of imperfect information in electoral politics. Unpublished Ph.D. dissertation, California Institute of Technology.

Calvert, R.L. (1986). Models of imperfect information in politics. London: Harwood Academic Publishers.

Chappell, H.W. (1989). Campaign advertising and political ambiguity. Working paper, University of South Carolina.

Collier, K., Ordeshook, P.C. and Williams, K. (1989). The rationally uninformed electorate: Some experimental evidence. Public Choice 60: 3-29.

Converse, P.E. (1970). Attitudes and non-attitudes: Continuation of a dialogue. In E.R. Tufte (Ed.), The quantitative analysis ofsocialproblems, 168-189. Reading, MA: Addison-Wesley.

Dougan, W.R. and Munger, M.C. (1989). The rationality of ideology. Journal o f Law and Eco- nomics 32: 119-142.

Downs, A. (1957). An economic theory of democracy. New York: Harper and Row. Erikson, R.S. (1979). The SRC panel data and mass political attitudes. British Journal o f Political

Science 9:89-114. Feld, S.L. and Grofman, B. (1981). Ideological consistency as a collective phenomenon. American

Political Science Review 82: 773-788. Filer, J.E., Kenny, L.W. and Morton, R. (1991). Voting laws, educational policies, and minority

turnout. Journal o f Law and Economics 34, 2 pt. 1 (October): 371-393. Filer, J.E., Kenny, L.W. and Morton, R. (1993). Redistribution, income, and voting. American

Journal o f Political Science 37 (February): 63-87. Harrington, J.E. (1992a). The revelation of information through the electoral process: An ex-

ploratory analysis. Economics and Politics 4: 255-276. Harrington, J.E. (1992b). Modelling the role of information in elections. Mathematical and Com-

puter Modelling 16: 133-t45. Kalt, J.P. (1981). The economics and politics ofoilprice regulation. Cambridge, MA: MIT Press. Kalt, J.P. and Zupan, M.A. (1984). Capture and ideology in the economic theory of politics.

American Economic Review 74 (June): 279-300. Kalt, J.P. and Zupan, M.A. (1990). The apparent ideological behavior of legislators: Testing for

principal-agent slack in political institutions. Journal o f Law and Economics 33 (April): 103-131.

Kau, J.B. and Rubin, P.H. (1979). Self-interest, ideology, and logrolling in congressional voting. Journal o f Law and Economics 22 (October): 365-384.

Kau, J.B., Keenan, D. and Rubin, P.H. (1982). A general equilibrium model of congressional vot- ing. Quarterly Journal o f Economics 97, 2, (May): 271-293.

Kmenta, J. (1986). Elements o f econometrics. New York: Macmillan. McKelvey, R.D. and Ordeshook, P.C. (1985). Elections with limited information: A fulfilled ex-

pectations model using contemporaneous poll and endorsement data as information sources. Journal o f Economic Theory 36: 55-85.

McKelvey, R.D. and Ordeshook, P.C. (1987). Elections with limited information: A multidimen- sional model. Mathematical Social Sciences 14: 77-99.

McKelvey, R.D. and Zavonia, W. (1975). A statistical model for the analysis of ordinal level de- pendent variables. Journal o f Mathematical Sociology 4 (1): 103-120.

Mitchell, E.J. (1979). The basis of congressional energy policy. Texas Law Review 57: 591-613. Morton, R.B. (1991). An analysis of legislative inefficiency and ideological behavior. Public

Choice 69:211-222. Nelson, D. and Silberberg, E. (1987). Ideology and legislator shirking. Economic Inquiry 25 (Janu-

ary): 15-25.

271

Palfrey, T. and Poole, K. (1987). The relationship between information, ideology, and voting be- havior. American Journal o f Political Science 31:511-530.

Peltzman, S. (1984). Constituent interest and congressional voting. Journal o f L a w and Econom- ics 27 (April): 181-210.

Poote, K.T. (1981). Dimensions of interest group evaluation of the U.S. Senate, 1969-1978. American Journal o f Political Science 25: 49-67.

Poole, K.T. and Rosenthal, H. (1989). Patterns of congressional voting. Carnegie Mellon GSIA Working Paper # 88-89-07.

Powell, L.W. (1989). Analyzing misinformation: Perceptions of congressional candidates' ideolo- gies. American Journal o f Political Science 33 (February): 272-293.

Reed, W.R. (I989). Information in political markets: A little knowledge can be a dangerous thing. Working paper, Texas A&M University.

Schultz, T.W. (1975). The value of the ability to deal with disequilibria. Journal of Economic Literature 13 (September): 827-846.

Shepsle, K.A. (1972). The strategy of ambiguity: Uncertainty and electoral competition. American Political Science Review 66: 555-568.