Measuring Party Cohesion on Roll Call Votes with an Application to the Labor Committee of the Chilean Senate
by John Londregan UCLA Version 1.0 | August 21, 1997
Introduction
While organized political parties play an important role in almost every legislature in the world, the channels by which parties operate are less clear. It may be that political parties are composed of like-minded legislators who would vote alike in any event because they share a similar political outlook, call this the a nity hypothesis. This similarity of outlook may arise spontaneously, or it may be the result of concerted action across many electoral districts to promote candidates with a common view. Another model, which one could refer to as the discipline hypothesis", is that, shared outlook or no, legislators belonging to the same political party compromise their basic ideological positions on votes that are important to their parties. Of course it could be that parties combine features of both of these pure types, with members banding together in part because of a shared outlook, one that makes it easier to coordinate, and less costly to compromise among themselves when party orders are given. While often interesting and entertaining, interview evidence from legislator's and their assistants on this point is hard to interpret. The incentives to dissemble are not trivial. Except among the Communist parties, which make a virtue of discipline, few politicians will eagerly admit in public to having compromised their principals on a particular issue. And among those for whom party discipline is a virtue, or at least a convenient excuse, the
which this research was undertaken. For their useful comments I would like to thank Larry Bartels, Henry Brady, Gary King, Howard Rosenthal, and seminar participants at the 1996 Political Methodology Meetings in Ann Arbor.
I am grateful to the Hoover Institution at Stanford University for providing a congenial environment in
1
credibility problem simply changes sign", as politicians take cover for their actions behind a mask of party discipline. The requirement that politicians take public positions on issues is designed to overcome the incentives to dissimulate about their stands, and their motives for taking them. In part this is the role of the press; by asking questions they set a hypothetical agenda and let political candidates and o ceholders respond. However the same problems that beset interviews with academics also cloud the content of press statements. Successful politicians are masters of the rhetorical gymnastics and social acrobatics used to de ect and evade di cult questions. Another institution designed to force politicians into the open is the roll call vote in which each legislator must register his or her vote of yes", no" or abstain" which becomes part of the public record. This restricted vocabulary limits the space for semantic maneuver, and so can do much to clarify a politician's position. Moreover, even when the vote is not brought to immediate public attention it may return in the next election to haunt the legislator who cast it, as political opponents and news reporters have powerful incentives to identify and publicize controversial or unpopular positions a legislator may have taken while the legislator himself will be quick to point to popular public votes. Here I develop an approach that permits us to recover information about the importance of the a nity" and discipline" hypothesis of party cohesion. Doing this requires us to rst construct a model of voting on legislative roll calls. If one maintains the hypothesis that such votes are sincere, then methods are available which can be adapted to the purpose of measuring party cohesion. The details of these methods will occupy our attention in the rst section of the paper. Attention then turns to the question of sincerity: is the hypothesis of sincere voting a reasonable one? The answers to this question depend heavily on the institutional context. To illustrate the methods used here I will draw on data from the Labor Committee of the Chilean Senate, and in section two the discussion turns to the institutional arrangements in Chilean Senate Committees that discourage members from casting votes designed to manipulate the information available to others. Given non-strategic voting, and a one dimensional issue space, the third section shows how roll call data can be used to construct one test of the a nity hypothesis, and two tests for behavior linked to party discipline. The fourth section presents empirical analysis of voting data from the Labor committee of the Chilean Senate. A short concluding section follows. This material is adapted from Chapters 1, 4, and 5 of a longer manuscript on the role of legislative institutions in Chile's transition toward democracy where interested readers can nd more detail about the institutional context.
1 2
1 A Self-Scoring Ideology Measure
Political presure groups take an intense interest in members voting records, and in the US these groups often compute summaries of the voting records of members of Congress,
1 2
Kingdon 1981 at 60. Arnold 1991 10-11, 49, 68-71. Londregan 1997 .
2
marking o points for incorrect votes" much as teachers might grade a true-false exam, and calculating a summary score reporting the percentage of correct" votes cast by the legislator. This approach can be useful in lobbying; as with students, members of Congress pay more attention when they are told something will be on the test" in the sense that a particular roll call vote will be included in the AFL CIO's COPE score or in the score for the Chamber of Commerce of the US. Moreover, umbrella groups, such as the Americans for Democratic Action ADA which unites many on the left, often face con icting agendabased presures over which votes to include in their rating. Should the ADA include a few environmental votes when calculating its rating, or instead place greater emphasis on votes of interest to social liberals" on abortion and AIDS funding? The ratings received by Congress members depend not only on how they chose to vote, but also on the outcome of the later struggle between people like Barbara Boxer and Barney Frank over the de nition of a perfect" liberal voting score. The very usefulness of interest group ratings in lobbying members of Congress and as weapons in the struggle to assert agenda control within the ratings groups themselves makes them awed measures of preferences, as they a ect the behavior they measure" while the political agenda whose perspective they re ect is itself often unstable. The analogue between interest group ratings of voting records and exam grades is a close one. While ratings groups seek to measure a legislator's latent ideology by tallying correct" and incorrect" votes, educators often seek to measure a student's latent mastery of a subject, or sometimes the student's latent ability, scoring correct" and incorrect" answers. The literature on the educational testing problem is extensive, and raises the interesting prospect of a self-scoring test, in which students' answers are used to gauge the latent di culty of the questions, and even to identify the correct answers, while the questions are then used to measure the student's latent abilities. This same approach can be brought to bear in the analysis of voting records. The use of statistical analysis to measure ability and other latent traits has a long history . This model emerges as the response to the question When is the widespread use of the fraction of correct answers on a test justi ed?". Put into the language of statistics this question becomes When is the fraction of correct answers a su cient statistic for the latent characteristic of ability or subject mastery the test seeks to measure?". The answer to this question is surprisingly straightforward . If we add a few sensible technical conditions then the fraction of correct answers is a su cient statistic for latent ability if and only if the probability that the subject v gives a correct answer to question p, vp is:
3 4 5
See for example early work by Thurstone 1925 , but a useful point of departure is the testing model developed by Rasch 1961 . 4 The development here is based on Fischer 1995 . 5 These are i the random variable vp in equation 1 below can take on any value on the real line, albeit with very low probability, ii no guessing", so that at very high levels of mastery the probability the subject answers correctly converges to 1, while at abysmal levels it converges to 0, iii random disturbances to the subject's response, caused by factors such as momentary distractions during the test, are uncorrelated across questions.
3
3
v , p = e v , p vp 1+e Many readers will recognize this as the logit probability. It is useful to rewrite this condition in terms of the discriminant function. Subject v will give a correct answer to item p if the following condition is met:
v , p ,vp 1 Here vp is a random error term that obeys the so-called extreme value" distribution that gives rise to logit probabilities . In this context vp corresponds to idiosyncratic factors, such as momentary distractions during the test. The parameter p measures the di culty of the question. Higher values of p correspond to lower values for the lefthandside of 1 and so to a lower probability of encountering a value of vp that will satisfy the inequality, and lead to a correct answer. The parameter is an arbitrary positive constant, and by convention it is always set equal to 1 in applications of the model. Given the value of , the parameter v corresponds to the ability of subject v to provide correct answers. All else equal, a higher value of v corresponds to a higher value for the lefthandside of 1 , and so to a higher probability of encountering a value of vp that will satisfy the inequality. This means that increases in the value of the di culty measure p decrease the probability of a correct answer, increases in the value of the ability measure v increase the probability of a correct answer. The Rasch model takes for granted that we know which answers are correct, and it is formulated in terms of the number of correct answers a subject gives. However, we could easily reparameterize it in terms of the number of times the subject answered true" on a test with binary options, e.g. true" and false". In this case instead of using the standard normalization of = 1 for all questions, would vary among questions, with p = 1 for questions for which the correct answer was true" and p = ,1 for questions whose correct answer was false". We could then replace the di culty parameter with a new parameter p equal to p for questions whose correct answer was true", but equal to , p for questions whose correct answer was false". For this setting the probability that examinee v answers true" to question p is given by:
6
2 An extension of the basic Rasch model takes this approach a step further. Instead of imposing the value of p a priori the two parameter logit model" treats this as an additional parameter to be estimated , so that we have the formulation in inequality 2 but without any restriction on the values that p can take on, instead we estimate not only the examinee parameter v , but also the question parameters p and p. In other words, we do not need
7
p v , p
,vp
6 7
Maddala 1986 . Birnbaum 1968 .
4
to know a priori which are the correct answers to the questions! Our estimates of p will tell us which answers are correct . If we knew the examinees' ability parameters, the v , then estimating the p and p parameters would be straightforward. The probability that examinee v with high ability answers true" to question p will exceed the probability of an answer of true" from a examinee v0 with lower ability if the item discrimination parameter" p is positive, while the high ability examinee will be less likely to answer true" if the item discrimination parameter is negative. Thus we can estimate the correct" answer by using our knowledge of the v to estimate p. If p 0 the correct answer is true", if p 0 the correct answer is false". Better still, in cases of ambiguity the estimator will alert us by returning a value of p near 0. Similarly we can estimate the di culty parameter for the question using p . correspond to a more stringent threshold for an answer of true" and so High values of p higher values of p reduce the probability that all subjects given an answer of true". This corresponds to increased di culty when the correct answer is true" and to reduced di culty when the appropriate answer is false". Of course educators are seldom if ever in the position of having absolute knowledge of examinees' abilities. However, it is in principal possible to simultaneously estimate the examinees' ability parameters and the question parameters. In practice professional testing organizations such as the Education Testing Service which administers the widely used Scholastic Aptitude Test SAT develop banks of questions which they update by introducing new questions a few at a time as part of a battery of test questions, the remainder of which have known characteristics. In principal the known questions serve to measure subjects ability parameters, and then the subjects' responses are used to gauge the di culty and item discrimination parameters for the new questions. This means that in a very important sense such questions are self scoring" the item discrimination parameters are not imposed by the graders, they are rst estimated using exam-takers' responses. The construction of self scoring tests has much to recommend it, especially as a means of at least partially escaping from preexisting biases on the part of the examiners. In particular, while the process is unlikely to identify answers the examiners had gotten wrong, it can often ag ambiguous questions e.g. those with low values for p, the item discrimination parameter . However, the simultaneous estimate of the examinee and question parameters by maximum likelihood can lead to biased estimates. This is because of what is called the parameter proliferation problem" which stems from the attempt to simultaneously estimate examinee parameters, v , whose estimated values depend on the question parameters, p and p , and question parameters, whose estimated values are contingent on the ability parameters. As the researcher adds data, for example by giving the exam to more subjects, she also adds more parameters to estimate, e.g. the ability parameters of the newly added test takers. This leads to biases as we either hold the number of subjects xed and allow the length of the test battery to grow without bound, or as we hold the number of test questions xed and arbitrarily increase the number of test takers. Under somewhat fanciful conditions;
8
In fact they will order the responses and the respondents, leaving the analyst to determine which end of the axis corresponds to high mastery, or high ability, usually a trivial task.
8
5
the number of examinees goes to in nity, and the ratio of the number of questions to the number of examinees also goes to in nity, maximum likelihood estimates remain consistent . The idea behind this is that we want to have so many questions that we can, in some sense, independently estimate the ability of each examinee using a separate, and arbitrarily large, set of questions. An alternative to searching for conditions under which the estimator is not biased is to gauge and correct for the biases. This has been done in the context of the so-called three parameter model" which is a further re nement of the two parameter logit model that allows for guessing by examinees. In that context the biases resulting with 96 questions and several million examinees, the numbers encountered on the Verbal SAT, have been found to be fairly small, and bias correction formulas adapted for that application have been developed . However, the biases are inversely proportional to the number of questions. When the number of questions becomes small the biases can be very large. Turning to voting in legislatures, it is useful to analyze voting within the contest of a version of the widely used spatial model. This version of the model recognizes that while the map of the political battle eld is drawn in terms of ideology, Stokes noted that not all political issues were of the con ictive variety the standard spatial model handles best. He divided issues into two types; position issues", that resemble the school bond example of the preceding paragraph, and valence" issues, such as honesty and integrity, about which voters all agree: all would agree that scandal is undesirable, virtually all would agree that a hard-working politician is preferable to one who is not. Nor do valence issues pertain only to characteristics of candidates. In fact many legislative votes are unanimous ones including very important decisions such as increases in the debt ceiling or declarations of war. While the valence aspect of many issues gets relatively little attention in game-theoretic analyses valence is very much a part of politics as it is practiced on a daily basis: many issues are simultaneously position issues and valence issues. Consider the example of a bill to increase taxes to pay for a war of national defense. While politicians of all persuasions, and their constituents, can readily agree on the importance of marshaling resources to avoid military defeat and foreign occupation, those on the right might be expected to favor sales taxes and the draft, which fall fairly evenly on all income groups, while those on the left might be expected to favor luxury taxes, and condemnation of large estates to be converted into air elds and training grounds, tax schemes likely to equalize the distribution of income. While there will be sharp disagreements over the details of how such a tax bill should be written, a shared interest in survival will create a powerful incentive for all to compromise. While few issues have the impact and urgency of a wartime tax increase, most combine elements of position" and valence" issues, which we may view as idealized types. Most issues, no matter how consensual, can be translated into policy in di erent ways in order to appeal to di erent groups, most issues, no matter how divisive, leave politicians with some
9 10 11
9 10 11
Haberman 1977 . Lord 1983 . Stokes 1963 .
6
areas of agreement. A simple analytic model that combines the valence and position aspects of issues represents policies in two dimensions. Along the rst, or position dimension, preferences are represented as in the standard spatial model; all else equal, the closer a policy is to an individual's preferred outcome along this rst dimension the better the individual likes that policy. Along the second dimension all individuals agree about what makes for better policy, for example all would nd a given level of environmental protection more attractive if it could be achieved at a lower economic cost. A simple model that captures these basic features represents an individual's preferences using the following linear-quadratic" utility function:
U p; v; x = , x , p + v 3 Legislator v, with preferred policy outcome xv , will prefer proposal p, corresponding to position pp and valence vp, to the status quo, with position psq and valence vsq , if the following condition is met:
2
U pp; vp; xv U psq ; vsq ; xv 4 This says that the proposal yields greater utility than the status quo from the perspective of someone with a preferred policy outcome of xv , accounting for both the valence and the position characteristics of the proposal and the status quo. Substituting from equation 3 this becomes: , xv , pp + vp , xv , psq + vsq To allow for the possibility that the legislator has an idiosyncratic reaction to the proposal, perhaps because it lies in an area of the individual's own personal experience, or because there is a particular constituent interest in the legislation, or because there is particular lobbying presure, for example from another member of the legislature o ering a favorable vote at a later date on an issue of greater interest to legislator v, we can add an idiosyncratic error term, vp, so that the condition for a favorable vote becomes:
2 2
, xv , pp + vp + vp , xv , psq + vsq
2 2
A little algebra establishes that this condition is equivalent to:
12
Expanding the square terms on both sides we have: ,x2 + 2xv pp , p2 + vp + vp ,x2 + 2xv psq , p2 + vsq v p v sq 2 " terms on both sides cancel. Next collect the all terms save The -x vp on the lefthandside: pp 2xv , pp , psq 2xv , psq + vp , vsq ,vp That is: 2 pp , psq xv , pp + psq + vp , vsq ,vp 2
12
7
gp xv , mp + vap ,vp Where the variables are de ned as follows: gp = 2 pp , psq This is twice the gap" between proposal p and the status quo. mp = pp psq The midpoint" between proposal p and the status quo. vap = vp , vsq The valence advantage" for proposal p relative to the status quo.
+ 2
5
If we go on to specify a probability distribution for vp, the idiosyncratic error term for voter v on proposal p equation 5 says that the probability that we encounter a value of vp high enough to lead voter v to favor proposal p will be higher the larger the value of vap, that is, the greater the valence advantage for the proposal relative to the status quo. If the legislator's most preferred outcome, xv is to the right of the midpoint of the proposal and the status quo, with xv , mp 0 then the probability the legislator will vote for the proposal is an increasing function of gp, meaning that the legislator is more likely to vote for proposals that move policy rightward from the status quo. Likewise if the legislator's preferred outcome xv is to the left of the proposal midpoint, so that xv , mp 0 then probability the legislator votes for the proposal declines as gp increases, meaning the legislator is less likely to vote for proposals moving policy rightward. This interaction between the legislator's preferred outcome and the proposal is a key feature of the spatial model. Collecting terms from the inequality 5 we see that voter v will vote for proposal p if:
gpxv + cp ,vp 6 Where cp = vap , gpmp measures the e ect of proposal characteristics that a ect all voters in the same way. While it would be desirable to be able to separately measure vap and mp we cannot do so on the basis of voting data alone. Without outside information we cannot tell whether a proposal to move policy to the right by gp gained almost no votes because the midpoint mp was to the right of the entire legislature, so that the policy sought to make an extreme element of the status quo even more extreme, or whether instead the proposal was voted down because it had a substantial valence disadvantage, as measured by a large negative value of vap. What is unambiguous is that increases in cp make every legislator more likely to vote for proposal p. It thus represents the characteristics of the proposal about which there is consensus. It can be thought of as a measure of the consensus appeal of the proposal. In contrast to vap and mp whose separate e ects we cannot disentangle, we can hope to estimate xv , gp, and the consensus appeal parameter, cp. While the spatial model and the measurement models from educational testing have very di erent foundations, and are designed for very di erent applications, notice that they share a very similar structure. If the vp error term obeyed the extreme value distribution the
8
voting model of equation 6 would correspond exactly to the two parameter logit model of 2 with gp = p, xv = v and cp = , p . This means that measuring ability using a truefalse" test and measuring political conservatism as represented in the spatial model using votes of yeah" and nay" involve solving the same type of statistical estimation problem. One insight from the educational literature is immediately portable to the question of measuring ideology. The use by interest groups of the percentage of correct votes to measure legislators' ideology is only the appropriate measure when the assumptions of the Rasch model are satis ed. While in this case there is little uncertainty about the interest group's de nition of a correct" vote, we also need the votes to all have the same item discrimination parameter . In the context of the spatial model in 6 this means that we must have gp = for all p. That is, the distance between the status quo and the proposal must be the same for all votes used to construct the interest group's rating scale, a condition that is very unlikely to be met in practice. As with the literature on testing the empirical analysis of roll call voting has evolved in sophistication over time . As with the educational testing literature early research used factor analysis and Guttman scales . As with the educational testing literature a more modern, and more carefully model-based approach emerged, in applications to multivalued questionnaire responses from public opinion surveys , and to roll call voting in the US Congress . The in uential work of Pool and Rosenthal uses a formulation of utility that di ers slightly from that developed in equation 3 combined with various error structures, one leading to a logit-style model and another to a version of the probit model Others have worked with exactly the formulation of 6 and the assumption of normal errors, leading to what might be called the two parameter probit model" . As with the literature on education testing, the literature on measuring legislators' preferences raises a cautionary note. Poole and Rosenthal conduct extensive Monte Carlo studies of their estimator, and nd that when the number of legislators falls much below 100 substantial biases in legislators' estimated preferred outcomes, the xv , emerge. Unlike Lord, they do not work out bias correction formulas, noting that in the application of their estimator to legislators the size of the US Congress, their primary interest, the biases are small. The biases in the ideal point estimators from the Poole and Rosenthal model, and in the ability and question-speci c parameters from the three parameter logit model stem from the attempt to simultaneously estimate two sets of interdependent parameters, and our inability to add data without at the same time adding additional parameters to at least one of the
13 14 15 16 17 18 19 20 21
For an anachronistically titled survey of the early literature see Poole 1988 . Chester 1948 . Gage and Shimberg 1949 , MacRae 1954 . 16 Cahoon, Hinich and Ordeshook 1978 . 17 Poole Rosenthal 1985 , 1996 . 18 Poole and Rosenthal 1991 . 19 Poole and Rosenthal 1996 . 20 See Ladha 1991 , also Platt, Poole, and Rosenthal 1992 . 21 Poole and Rosenthal 1991 .
13 14 15
9
two sets. In the legislative context this means that we try to estimate parameters speci c to each item brought to a vote; gp and cp, and to estimate an ideology parameter for each legislator; xv . This is a particularly serious problem in the Chilean context, where Senate committees have only 5 members, all of whom occasionally miss votes. The biases entailed by this small number of voters for maximum likelihood estimates of the model in 6 are severe. Moreover, the option of conditional maximum likelihood estimation" CML developed by Rasch, which permits consistent estimation of test subjects ability parameters even when the number of questions per examinee is small cannot be applied to the two-parameter logit model. As previously noted, to take advantage of that technique, one would rst need to be willing to assume that the gp were the same for all proposals, so that the distance between the status quo and the proposal was the same for all proposals not what we expect to see in a legislative context. The estimation quandary is, in one sense, a byproduct of our insistence on remaining agnostic about the content of the proposals, and about the determinants of legislators' preferences. Thus we ignore information about the party a liation of legislators, their district characteristics, and professional backgrounds, as we ignore the characteristics of the proposals on which they vote, including information about their content, their sponsors, and the presure groups that support and oppose them. Incorporating this sort of information into our model is in any event intrinsicly worthwhile, and returns us to important questions about the linkages between ideology and policy choices. This approach can also solve the parameter proliferation problem that leads to biased estimates. Returning to the voting model in 6 denote the expected value of the idiosyncratic preference shock for legislator v by bv = E fvpg. Subtracting E f,vp g from both sides of 6 , and letting vp = ,vp , E f,vpg we have:
22
gpxv + cp + bv vp 7 Where the error term vp has an expected value of 0. As it is formulated in 7 the model treats the legislator and proposal characteristics symmetrically. In the language of the psychometric literature, including bp and cv in the model has the e ect of double-centering" the data. The next step is to model the structure of the legislator and proposal parameters. Let us consider rst the proposal parameters. Let the ideological displacement of proposal p be given by: gp = ~ 0wgp g~ 8 Here wgp is a column vector of explanatory variables. For example it might consist of a ~ set of proposer indicator variables. In the context of the Chilean Senate's Labor Committee the rst element of wgp might be dummy" variable equal to 1 if Senator Ruiz de Giorgio ~ was among the cosponsors of proposal p, and equal to 0 otherwise. If we were so inclined
This technique has been heavily borrowed by economists to deal with what they refer to as xed e ects" logit models, see Chamberlain 1984 .
22
10
we could also code proposals by their content, so that the second element of wgp might be ~ an indicator variable for whether the proposal enjoyed the support of the unions, equal to 1 if it did so, and to 0 if it did not. We might expect such sponsorship to be associated with bills moving labor policy leftwards". Likewise, if such data were available for multiple authored bills, we might want to include a measure of how much of the bill was written by each of its sponsors, and indeed, how much was written by lobbyists. The other element on the righthandside of 8 , ~ , is a vector of parameters measuring the impact of each of our g explanatory variables, and it is the same vector for all proposals. These parameters, common to all proposals, must be estimated. This means, for example, that if having been written by Socialists moves one proposal half a unit to the left of the position of a proposal with otherwise identical measured characteristics, for the sake of the example, one also written by lawyers and backed by a labor union, but not coauthored by socialists, then socialist authorship will have the same a ect on all proposals, moving each a half a unit to the left of proposals with the same remaining observed characteristics. Just as we can model proposals' displacements, so too we can build a similar model of the consensus appeal bp parameters:
cp = ~0wcp + p c~ 9 As in equation 8 wcp is a column vector of explanatory variables that a ect cp. These ~ may include all of the measures discussed in the context of gp and in addition might include various measures of expertise, such as measures of the proposal's authors' professional backgrounds before coming to the legislature. The parameter vector ~ measures the impact of c the measured characteristics on proposals' consensus appeal, and it's value is the same for all proposals. As with the ~ parameters, the value of ~ must be estimated. g c Equation 9 di ers from equation 8 in that it includes the random variable p. This random variable captures two sources of variation in proposals' consensus appeal. First there is the well known tendency for people to have both good" and bad" ideas. Even the most expert policymakers sometimes make technically awed proposals, and even those with little expertise will occasionally make proposals everyone recognizes as improvements. Thus we expect a distribution of valence advantages, vap, among proposals with the same author, or group of authors. The second source of variation is variation in the proposal midpoint. This will come about naturally as legislators and the Executive seek to identify amendable elements of the status quo. Rather than sitting down at their word processors, entering the ideological coordinates of the policy they would most like to see, and then watching as the text of a proposal with the desired characteristics rolls o the printer, legislators seeking to propose changes to policy must seek to match ideas for reform with elements of the legislation they are considering. Consider an example from the Chilean Senate. Socialist Senators Calder
n, Gazmuri, o Nu~ez, and Vodanovic, joined by Christian Democrat Arturo Frei, and by Laura Soto of n the Party for Democracy proposed an amendment to the Labor Code reform bill with the object of guaranteeing minimum rest periods for workers on factory ships. Among other provisions it required that workers be guaranteed a minimal rest period of at least eight
11
uninterrupted hours per day. This proposal struck the Senate Labor Committee as reasonable, and it was unanimously approved. Shortly thereafter the committee received an urgent joint communiqu
from the worker's association, the Federaci
n Nacional de Tripulantes de e o Naves Especiales de Chile, and the association representing the owners of factory ships, the Asociaci
n de Buques Fabricas Arrastreros e Industrias Conexas. This missive informed o the committee that the standard work schedule at sea involved two six hour shifts of work and two six hour rest periods every 24 hours, and that to change this to accommodate the eight hour rest requirement would utterly disrupt the ships, a prospect that dismayed the shipowners and met with erce opposition among the workers, whose earnings were calculated in part on the basis of the quantity of sh processed . Article 23-A of the Labor Code dealing with workers aboard shing vessels had looked to the amendment's authors like a natural element of the status quo for a proposed liberalization", in the form of limits on working hours. Nevertheless, and because of exactly the sort of uncertainty about the linkage between proposals and policy emphasized by Gilligan and Krehbiel , the proposal did not achieve its objective of moving policy leftward" with no loss in the status quo valence. Of course, not all proposals by the authors of the amendment to require eight hour rest periods aboard shing vessels were as dismal as the one just described. It is this variability of the consensus appeal of proposals from the same author or group of authors that is captured by including the random p term in equation 9 . In the context of the model, the eight hour rest" proposal corresponded to a very low value of p, a fact belatedly recognized by the committee. Just as for the proposal parameters, we can model the voter characteristics, xv and cv :
23 24
xv = ~ 0 wxv x~ bv = ~0wbv b~
10 11
As with the proposal characteristics, wxv and wbv are are vectors of variables that measure ~ ~ observable characteristics of legislators that are likely to in uence their voting behavior, such as their party a liation, their election margin, perhaps their career background or the region of the country they represent. Alternatively, these characteristics may simply represent vectors of legislator-speci c indicator variables, so that we wind up with a more agnostic measure of legislators' ideal points. Likewise, ~ and ~ are parameters to be estimated. x b Substituting these expressions for the proposal and legislator characteristics we obtain: ~ 0wgp ~ 0 wxv + ~0wcp + p + ~0 wbv vp g~ x~ c~ b~ To operationalize this model we need to specify probability distributions for p and vp . The normal and logit formulations of such disturbance terms are both widely used. However, the logit formulation assumes a fairly restrictive condition known as independence
23 24
Sesiones del Senado, 30 marzo, 1993 p. 4625. Gilligan and Krehbiel 1987 .
12
of irrelevant alternatives". As there is no reason to assume that this condition applies here, and because some of the later analysis will explore the correlations between pairs of the vp's, I work with a normal distribution for the vp's. The p errors need not also be normal, the routine used to estimate the model integrates over p numerically, and so can just as easily accommodate non-normal distributions. However, as a rst approach, the p are taken to be normal as well. The mean of the p will simply be absorbed" by a uniform shift in the ~ c parameters, so that without loss of generality we can assume the p have mean zero. However, there are no a priori grounds for restricting the variance of p relative to the variance of the vp 's. A straightforward way to handle this is to replace p, normally distributed with mean 0 and variance , with p, where p is normally distributed with mean 0 and variance 1. Substituting this into our model, we see that equation 7 becomes:
2
~ 0wgp ~ 0wxv + ~0wbv + ~0 wcp + p vp g~ x~ b~ c~ 12 In addition to the usual probit normalization that the vp have mean zero and variance one, four other normalizations are required to identify the model. Two of these amount to an arbitrary selection of origin and scale for legislators' preferred ideologies. We could measure most preferred outcomes on a scale from 0 to 100, or on a scale of ,1 to 1, or on any scale we liked. In fact we could reverse measures, putting the political left" on the right of the scale, e.g. by giving them higher numbers, as is done in the US by organizations like Americans for Democratic Action, whose ADA score places political liberals at 100 on the right of the scale and political conservatives at 0, on the left of the scale. In this setting, a change of the units used to measure ~ can simply be o set by changes in ~ and ~. To x g c avoid ambiguities, it is su cient to normalize the locations of two legislators. Thus for each committee a legislator generally thought to be on the political right, such as institutional Senator Olga Feli
, is normalized to have a location of 1, while a Senator thought to be on u the left, such as Christian Democratic Senator Maximo Pacheco, is normalized to have a location of ,1. The remaining normalizations have the e ect of choosing a reference proposer", or when explanatory variables are used in place of dummy variables for the proposers' identity, a reference proposal type". For this proposal type the corresponding element of ~ and the c corresponding element of ~ are set equal to zero. Substituting these normalizations into 12 g we see that the condition for supporting proposals of the reference type is:
~0 wbv + b~
p
vp
This means that the probability that voter v supports a proposal of the reference type depends on the level of ~0 wbv , so that the ~ parameters measure the tendency to vote for b~ b proposals of the reference type. As with the selection of two legislators to normalize the ideological preference parameters, the choice of reference proposer is arbitrary. Changes are simply o set by changes in the estimated value of the remaining parameters. 13
Given our identifying normalizations, the parameters of the model set forth in 12 can be estimated consistently using maximum likelihood estimation . The secret to the success of this approach is that by xing the number of parameters, that is, the lengths of the vectors ~ , ~, ~ , and ~, we no longer add parameters to estimate whenever we add data, and so we x b g c avoid the parameter proliferation problem. As a practical matter, the number of legislators voting in a given Committee is small, and so we may still choose to estimate preference parameters xv and cv for every legislator who participated in a committee's votes. This choice is readily encompassed in the framework of 12 by letting wxv and wbv correspond to ~ ~ vectors of legislator-speci c dummy variables. However, as we add votes the key is not to try to estimate a separate parameter for each proposal. While the Constitution Committee voted on over 800 proposed amendments during the sample period, no attempt is made to separately measure correspondingly many proposal parameters. Instead proposals are identi ed with their authors, or in some cases, with the party a liations of their authors using author-speci c, or party-speci c indicator variables. While it would be very useful to have detailed information about which of a proposed amendment's cosponsors was the real author, and what the relative contributions of the cosponsors really was, this information is simply not available. While one might hope to learn something about it from the legislators themselves; those who have retired might be willing to speak, there is little hope of obtaining comprehensive and reliable data of that nature. Instead multiple authored amendments are taken to be weighted averages of the characteristics of their authors, with each of N authors' characteristics receiving a weight of 1=N . Thus if legislator i was among 5 cosponsors of amendment p the elements of wgp and of wcp corresponding to legislator i would equal 1 5. ~ ~ At this point it is straightforward to write down the contribution of proposal p to the ~ log of the likelihood function. It is a function of the parameter vector :
25
8 Z ~ = ln 1 Y
Y es v n ~ 0wgp ~ 0wxv + ~0 wbv + ~0wcp + lp g~ x~ b~ c~ : ,1 v V p
p
o
p d
9 = p;
13
Where:
~ x ~7 b7 ~ 7 14 ~7 g5 ~ c While V p is the set of legislators who voted on proposal p and Y es v = 1 if legislator v voted in favor of the proposal, and equals ,1 if instead she voted against it. As a practical matter it is convenient to estimate characteristics of proposers whose number is bounded above by the membership of the Senate plus the President of the Republic. Likewise, it is convenient to work with the set of legislators voting on a committee. But this approach entails di culties for voters and proposers who seldom participate.
25
2 6 6 =6 6 4
3
See Londregan 1996 for details.
14
Consider the case of Chilean Institutional Senator Fern
ndez who participated in but one a proposal considered by the Senate Labor Committee. This proposal was rejected, and so Labor Committee proposals by Senator Fern
ndez have an observed failure rate of 100. If a one were to include a proposer-speci c parameter for Senator Fern
ndez in the speci cation a maximum likelihood estimation would attempt to set the c parameter corresponding to that Senator equal to ,1, attributing all of the blame" for the proposal's rejection to the participation a Senator making proposals with in nitely poor consensus appeal. Of course, this would be a classical case of over tting" the data, with the sparse participation of Senator Fern
ndez in the work of the Labor Committee essentially creating a stray parameter a corresponding to a single datapoint. For this reason it is impractical to attempt to estimate individual-speci c parameters for proposers and voters who participate infrequently, and so these are grouped with other members of their parties. Individual-speci c preference and proposal parameters are only calculated for frequent participants.
2 Sincere Voting in Chilean Senate Committees
The preceding section set forth an approach to recovering preferences when we are con dent that legislators vote sincerely for the policy option closest to their most preferred outcome. But will they vote sincerely? Given the enormous amount of information contained in a sincere vote, legislators are often reluctant to cast roll call votes. Likewise, we must be mindful that the high information content of votes that are believed to be sincere means such votes may in uence the beliefs of others, and so can undermine the incentives to vote sincerely. Because public voting can convey so much information legislatures jealously restrict the agenda that is brought to a roll call vote, as when the nal passage in the US House of Representatives of the Tax Reform Act of 1986, a major tax reform measure was decided by voice vote The Chilean Congress is no less di dent about roll call votes than are most other legislatures. Many important bills are decided by a count of hands in which the record shows the vote totals for and against the measure, but does not report how each individual member voted, it is also common for sessions to be interrupted while the president of the chamber holds brief private meetings with the leaders of the caucuses to resolve the disposition of amendments, the continuation of debate, and even the passage of a bill. In these meetings the leaders' votes are weighted by the size of their caucuses, and the votes are not publicly reported. Although taking clear public positions can be dangerous for politicians, sometimes circumstances overcome their basic reticence. This is the case for standing committees of the Chilean Senate. The Senate must overcome its informational disadvantages relative to a well informed executive and a conference committee which includes policy experts from the Chamber of Deputies. In the preceding Chapter we saw that the Senate has developed insti26 27
26 27
Birnbaum and Murray 1987 . votaci
n economica. o
15
tutions to overcome this di cult situation, by carefully vetting the members it sends to the conference committee, and by forcing the expert members of its own standing committees to take detailed public positions on the bills they consider. Key to this forced position-taking is the requirement that Senate committees take public roll call votes on each amendment considered during the second reading of a bill . How one interprets these votes depends on whether one believes the public spotlight will be su cient to induce sincere" voting on the part of committee members, or whether instead they will vote strategically". To begin to address this question consider the technology" for writing amendments. A typical bill, such as the labor law reforms considered in 1993, will contain a host of margins on which policy can be adjusted. Just as the concentration of a medication in a patient's bloodstream can be increased by administering doses through injection, or by pill, or elixir, or perhaps using an inhaler, so too the pro-labor" content of a bill can be increased or reduced to the desired level by adjusting policy on any of multiple margins: by extending coverage of the maximum work week, by expanding the liability of contractors for workplace injuries, by increasing the number of mandatory vacation days, and in many other ways. A legislator seeking to move the bill moderately in a pro-labor direction, but not as far as the unions might like, could thus achieve this by shortening the work week to 48 hours, but leaving the status quo liability of employers intact, or by increasing employers liability for injuries sustained by employees of subcontractors, but leaving the work week untouched, or there could be requirements about the provision of workday meals, or onsite medical care, or more formalized grievance procedures. In order to move the bill moderately in the pro-labor direction the law could be modi ed in just one or two of the ways suggested above. However, substantial modi cations made simultaneously on all margins would lead to a much larger pro-labor shift in policy. The multiplicity of margins in the ambit of labor law reform is shared by other issue areas as well; telecommunications policy can be made less regulated via partial privatization and little regulation, or complete privatization and extensive regulation. The level of deregulation can be adjusted by changes in the rate base companies can use in calculating their fees, or in the form of increasing or removing explicit price caps, or by compelling local phone service to be extended to remote and unpro table rural customers. Antiterrorism policy can be made less severe by removing some crimes from the jurisdiction of the military courts, or by reducing the budget for enforcement, or through an adjustment of the penalties for those convicted. Across the policy spectrum leftward" or rightward" shifts in policy positions can be achieved through adjustments on a variety of margins. The connection between these multiple margins and the valence of policy about which voters care is often complex, requiring considerable study and expertise to understand. What will be the valence e ects of a 6 month increase in the mandatory sentence for possession of marijuana, in the form of prison crowding, congestion of the court calendars, deterrence of drug abuse? Will the valence of this proposal be greater than a measure requiring special
28 29
28 29
Reglamento del Senado, art
culo 40. i Diarios de Sesiones del Senado March 30, 1993, pp. 4617-80.
16
courts to hear drug cases? What about increased spending on enforcement? It is about questions such as these that informed committee members on their way to the conference committee have special expertise. However, once legislators reach the conference committee they will choose only some arrows from their quiver of potential amendments. A conference committee that decides to move labor policy moderately in favor of labor will, for example, reduce the workweek but not for farmworkers, nor perhaps will it increase the number of mandatory holidays. The precise choice of instruments used to achieve the desired policy shift is something to which the committee is not committeed ex ante. Moreover, given some bargaining and heterogeneity within the conference committee we may expect that even future conference committee members will not be able to accurately forecast the set of instruments upon which the committee will settle. Now consider the voting decision of a pro-labor Christian Democratic Senator in the Senate Labor Committee at the second reading of the labor reform bill. Will this Senator vote the pro-labor position on a change in the length of the workweek? Suppose for the sake of argument that by strategically taking the opposite position the Senator could deceive pro-business legislators into favoring the amendment, if it was subsequently adopted by the conference committee. The cost of taking this position is that it will be used in the next election, the press will label the member as somewhat less pro-labor. This would mean that workers in the member's district would be more likely to stay home, or to vote for a more reliably pro-labor candidate on the same party list, the legislator will incur an erosion of his reputation. The policy bene ts of the deception depend on whether the amendment is actually adopted by the conference committee. As the members of the committee are informed the impact of the member's vote on whether the committee adopts the amendment may be quite small, though the member's misleading committee vote will serve to deceive uninformed members of the chambers, should the conference committee adopt it. Thus the member weights certain costs against uncertain bene ts. The greater the uncertainty about which margins the conference committee will adjust, the lower the Senator's expected payo from casting a deceptive vote. The payo to deception is further diminished by the prospect that even if the committee adopts the policy about which the member cast the deceptive vote, that deception will have been for naught unless it actually changes the eventual legislative outcome. If the pro-labor amendment would have been enacted anyway, then the member would have incurred the costs of the deceptive vote unnecessarily, likewise if the amendment is defeated despite the member's e orts. Unlike sincere votes, the member gets no electoral credit for having furthered the position of labor, instead just the opposite will happen as the Senator's reputation for taking a pro-labor position erodes. Even if the amendment isn't adopted; the cries that The Senator voted against reducing the work week!" would cost a Christian Democratic Senator votes, while if the proposal prevails the Senator's excuse that he cast a deceptive vote to aid in passage of the pro-labor position, thereby tricking his colleagues, is less likely to be believed by the voters than it is by his Senate colleagues, who will respond by according his statements and votes less credibility in the future.
30
30
McGraw 1991 .
17
When the issue area is at all complex the conference committee has many margins on which it can adjust policy, and which of these it will use is hard to forecast at the time of the second reading of the bill, even for future conference committee members. This limits committee member's incentive's to cast misinformative votes at the second reading. At the same time the public nature of the committee roll call votes forces members to take positions, and they will pay a political price if they choose incorrectly. Thus there are powerful incentives at work to induce sincere voting in Senate committees at the second reading of the bill. The cross-cutting pressures on US Congressmen to vote either strategically or sincerely can have an important e ect on their voting behavior. In their careful analysis of this question, Denzau, Riker and Shepsle 1985 considered these pressures when a strategic vote could directly a ect the subsequent agenda. They examine the Powell Amendment in detail. Under the Rule by which that amendment was considered, the vote on the amendment was to be followed immediately and without further opportunity to consider additional amendments by a vote on nal passage of the bill. In contrast to that situation, Committees of the Chilean Senate at the second reading have very little direct agenda control, amendments passed by the committee can be rejected by the oor, amendments rejected by the committee can be renewed. To better understand the factors at work in the context of the Chilean Senate the simple spatial model introduced in the preceding section provides useful guidance. Suppose that the committee member's constituents expect him to advocate a policy position of x. Let a public vote cast by the member be represented by the variable f0; 1g, where = 1 represents a vote in favor of the proposal under consideration, while = 0 is a vote against the proposal, and hence implicitly in favor of the status quo. If the member votes on a proposal of pP when the status quo alternative is pSQ the member's electoral payo is:
V ; x; pP ; pSQ; qP ; qSQ = , pP , x , 1 , pSQ , x + qp + 1 , qSQ 15
2 2
Suppose the legislator has true" policy preferences represented by:
U p; q; z = , p , z +
2
16
where z is the member's true" preferred policy outcome. Letting denote the probability the bill is enacted, the expected value of the member's overall wellbeing resulting from the bill is:
V ; x; pP ; pSQ; qP ; qSQ + w U pP ; qP ; z + 1 , U pSQ; qSQ; z 17 In equation 17 the w term corresponds to the weight the legislator attaches to his actual policy preferences, as opposed to those he is expected to advocate . The dependence of
31
Since this term calibrates the weight the legislator places on his or her own agenda, we can restrict our attention to non-negative values of w.
31
18
on recognizes that the Senator's vote at the second reading of the bill in the committee may a ect the probability the policy is eventually adopted. Some straightforward manipulation reveals that the Senator will vote in favor of the proposal if the following condition is met: 2 pP , pSQ x , pP , pSQ + qP , qSQ 1 + w 1 , 0 0 18 2 where 1 , 0 x = x1+ zw 1 , 0 +w Provided 1 + w 1 , 0 0 the Senator behaves as if" his most preferred outcome was x . The value of z represents the legislator's true" policy position. Research on retiring members of the US Congress indicates members' actual preferred points, the z's, are very similar to the positions they advocated throughout their tenure, the x 's. Whether this nding would generalize to other legislatures is, of course, an open question, but it is suggestive. The w values are inversely proportional to the importance of maintaining an ideological posture for electoral reasons. The greater the weight a member places on reelection the smaller will be the member's value for w, and the closer x will approach x, the position that favors the member's electoral chances. The value of 1 , 0 is the impact of the member's vote in the committee on the probability the amendment is actually part of the legislation that is nally enacted. It is conceivable that this impact is negative, that is, that by voting against an amendment in the committee the Senator actually increases the probability the amendment will be adopted. However, it seems very unlikely that this e ect will be large enough to make the 1 + w 1 , 0 term negative . If the other members of the legislature know the value of x then they can learn from the legislator's public committee vote. For this to work it is not necessary that x, the preferred policy of the constituents, be close to z, the preferred policy of the member, nor is it essential that w, the weight the member places on his own policy preferences, be small. It is not even necessary that the e ect of the member's vote on the probability the amendment is eventually enacted is constant, as long as other Senators know what it is. Even if this last condition is not satis ed, if w is small, or z is close to x or the absolute value of 1 , 0 is bounded above by a small number, then the value of x will be nearly constant across amendments and we may expect other Senators to learn what it is. As a practical matter we may expect a committee member's impact on the nal passage of the amendment, 1 , 0 , to be small. Both the conference committee and the president are informed, and so they will recognize the deceptive vote for what was. The conference committee bill, and presidential observations", must be voted under a closed rule", that is, the amendment will be bundled together with the rest of the bill. Moreover, the other committee members votes already provide additional information to the uninformed members
32 33
Lott 1996 . A su cient condition is that the legislator places more weight on the reelection motive than he does on his personal ideological agenda, so that w 1.
32 33
19
of the parent chamber; to change the outcome through a deceptive vote a committee member needs to change the information possessed by the pivotal voter in one of the parent chambers by enough to change their vote on the entire bill, or his vote must change the conference committee's decision of whether to o er the amendment. Once the other legislators have an accurate estimate of the e ective preferred outcomes, the x , of committee members, they can use the committee votes to make inferences about the impact of a proposed amendment on the valence of policy, that is, about qP , qSQ. This information then becomes very useful when evaluating the conference committee's bill, and the contents of the amendments o ered as part of the president's constructive veto if there is one. The objective for members is to try to anticipate all of the policy recommendations they might encounter later, and propose them to the committee during the second reading of the bill. Then if the conference committee bill is loaded with amendments similar to those that split the Senate committee along ideological lines, with for example the pro-labor Senators favoring the amendment, while the pro-employer Senators took contrary positions, the uninformed members of the Senate will know that the bill probably moves policy a considerable distance toward labor's preferred outcome and away from the status quo. If instead there are relatively few amendments in the conference committee bill that split the Senate committee on ideological lines, with many amendments that received unanimous support in the Senate committee, then uninformed Senators on the oor can con dently conclude that the bill is mostly a consensus measure that raises valence while leaving the ideological content of the status quo largely unchanged.
3 Party Cohesion
This framework permits a careful assessment of the degree to which members of the same political party share the same preferences. The most straightforward approach to doing this is to compare the estimated xv 's and bv 's for members of the same party. On most committee votes a party is represented by at most a single Senator. The ideal point estimates for a particular committee are based mostly on nonoverlapping votes, and we seldom observe Senators from the same party voting on the same proposed amendments. As the discussion of the preceding section indicated, comparisons of estimated ideological preferences between committee jurisdictions are problematic, because di erences between the elements of ~ cold be attributable either to di erences in the proposal writing ability and b ideological disposition of the reference proposers for the two committees or to di erences in the ideological preferences of the voters whose b's are being compared. In contrast, for comparisons within the same committee jurisdiction, the reference proposer is the same. In this case di erences among the b's must be attributable either to di erences in the substantive expertise required to master di erent parts of the committee's agenda, or to di erences in the mean of the idiosyncratic preference shocks of the voters. Thus, how we test the hypothesis that two voters bring the same preferences to the committee depends on whether we believe that the technical aspects of the subject matter of the committee's jurisdiction draw on one set of skills. if we think that they do, we should check both whether legislator's estimated 20
preferred outcomes, the x, match, and whether they draw from the same distribution of idiosyncratic disturbances to preferences, so that their b's match as well. However, if we expect that there is heterogeneity within the committee's jurisdiction, so that, for example, legislation dealing with corporate law and bills related to human rights call for di erent kinds of expertise, then we should only check for matches in the x's, as di erences among the b's might result from di erences in the ability to formulate legislation across di erent subsets of the committee's agenda. While multiple votes from members of the same party are rare, they do occur, and they provide us with additional important information about Senators' preferences. We seldom observe Senators from the same party voting di erently when they vote on the same measures. Some of this similarity is surely attributable to the tendency for Senators from the same party to have similar preferred outcomes. Yet party pressures may also play a role. On key votes the party leaders may demand loyalty, pushing Senators to shift from their most preferred outcomes with promises of rewards and threats of punishments. While most legislative scholars readily acknowledge that some degree of party discipline occurs, it is also part of the common view that, at least for Congressional parties in the US, this discipline is applied selectively . Moreover neither the press nor academic observers are regularly invited to attend when party leaders twist the arms of the membership. Thus, we are not in a position to directly measure which votes, if any, were in fact party whip votes. Even if we are prepared to accept that all votes labeled party whip votes in fact are, we will miss any that involved more discrete pressure. Moreover, there is always the concern that the party leadership will attempt to exaggerate its ability to twist arms by claiming some votes as whip votes" when members were all inclined to vote for them anyway. In the pre-coup Chilean setting observers noted considerable heterogeneity in the level of discipline among parties, with discipline tending to be tighter on the left than on the right . During the Aylwin years that are the focus of this study the proposal behavior of the Socialists is extremely disciplined, with a majority of proposals from Socialists Senators being cosponsored by the party's entire Senate delegation. Patterns of coauthorship exhibit much looser cohesion among members of the other parties. However, the Socialists' Senate delegation is sparse, with only four members distributed across the various committees, and we do not observe pairs of Socialists in committees voting on the same proposals. In fact, when we remove legislators' votes on proposals they themselves have written or coauthored we never observe more than two members of any party voting on the same proposal in any committee. For pairs of proposers from the same party voting on the same proposal we can augment the voting equation 12 to allow the idiosyncratic preference shocks, the vp to be correlated, thereby allowing for party whip behavior. While we may not be able to identify which votes the party attempts to in uence, we will nevertheless observe a closer a nity in the voting records of members subject to the same party whip than would be expected from the similarity in their preferred policies alone. This approach is similar to the statistical analysis
34 35
34 35
See for example the discussion of party leadership votes" in Cox and McCubbins 1993 . Agor 1971 .
21
of carcinogenic environmental hazards. Statistical analysis can often tell us whether people living near a nuclear power plant contract cancer at a higher rate than the general population, though if the cancer rate is higher for power plant neighbors statistical analysis cannot tell us which of the excess cancers resulted from proximity to the power plant, and which would have occurred anyway. Likewise, by allowing for correlation between the preference shocks, the vp , for members of the same party voting on the same proposal, we are still unable to identify which votes are in fact whip votes, though we can detect whether some whip voting is taking place. Another possibility is that the correlation between the vp for members of the same party is in fact negative. This might emerge if members of the same party coordinated their committee attendance to avoid sending largely redundant signals with their identical votes, showing up only when at least on of the party's committee members wanted to highlight a proposal about which the party members disagreed. The likelihood function 13 is readily extended to deal with this case by including the correlation between votes cast by members of the same party. Thus, letting !vp represent the nonrandom in uences on the vote of member v on proposal P : !vp = ~ 0wgp ~ 0wxv + ~0wbv + ~0 wcp g~ x~ b~ c~ the contribution to the likelihood function of proposal p becomes:
~z lp ; ~ =
where:
8 Z1 Y ln : Y es v f!vp + ,1 v V1 p
p g
Y
Y es v r f!vp +
2
p
g ; 2Tan,1 z
r = p d p
r W p
b Y es v r f!vp +
1
pg ;
o
19
b a; b; is the bivariate normal cumulative density function with discriminants a, and b, and correlation . V p is the set of all voters who are the only member of their party to vote on proposal p. W p is the set of all parties, r, that have two members voting on proposal p vi r members of party r, i f1; 2g, voting on proposal p. zr a parameter that gauges the correlation between the vp for members of party r with = 2Tan, zr =.
1 1
While this new formulation introduces some additional notational complexity the only real change here is to allow for correlation between the vp for members of the same party. Because the correlation coe cient is theoretically bounded on the interval ,1; 1 it is convenient 22
to rst estimate the parameter zr for each party with members casting simultaneous votes and then to recover the corresponding value for . This so called arctangent transform" is convenient because it simpli es the mechanics of maximizing the likelihood function . A nal technical question which must be addressed before proceeding with the analysis of the committees has to do with votes on one's own proposals. It is only under odd circumstances, that Senators ever vote against their own proposals, and very little is lost in treating the probability a Senator votes in favor of his own proposal as being equal to 1, so that its contribution to the log of the likelihood function is ln 1 = 0. This is tantamount to dropping such votes from the dataset, before estimating the model, and this is the approach taken here. A somewhat di erent approach is possible for votes on proposals from members of one's own party. In this case we can include an element of wbv that is equal to 1 if legislator v is ~ voting on a proposal from a member of his own party. In this case the corresponding element of ~ measures another aspect of party loyalty, that is, it measures how much more likely a b member is to vote for a proposal if it comes from a member of his own party, controlling for the member's ideology and the proposal's ideological and valence characteristics. This complements the party whip parameter, zr , that measures the correlation between the preference disturbances of members of the same party voting on the same proposal. A positive significant value for this parameter would indicate that members given preference to proposals from fellow party members even after controlling for their ideological content. Including this parameter has the added advantage that it purges our ideological estimates of party loyalty e ects. However, to avoid of high collinearity with the other elements of wbv a single party ~ loyalty parameter common to all the legislators voting on a given committee is estimated. As it is formulated, the log-likelihood in equation 19 exploits the fact that we never observe more than two members of the same party voting on the same bill in the dataset. To allow for party whipping" when there are more than two voters from each party the loglikelihood would have to be modi ed, allowing for higher dimensional joint normal densities. Under ordinary circumstances this would be a severe problem, as numerical integration routines tend to break down as the number of dimensions being integrated increases. However, in this case all members of the same party face the same party whip", so that we can think of the vp for members of a particular party, r, as all partaking of a shared common term,
rp and of uncorrelated idiosyncratic terms, vp:
36
vp = v p +
rp We would then need only to integrate numerically over the
rp for each proposal, as well as the p term already in the likelihood. Such a bivariate numerical integral would be manageable.
A further complication arises in that in some cases, when the number of votes cast by members of the same delegation is very small, there are no disagreements between members of the same party. When this occurs the maximum likelihood estimator wants to make the corresponding zr arbitrarily large, coming as close as possible to = 1. In the following chapters I deal with these cases by restricting the value of to the interval ,0:999; 0:999 , which corresponds to restricting the value of zr to the interval ,636:62; 636:62 .
36
23
4 An Application to the Labor Committee of the Chilean Senate
As discussed in section 2 the rules of the Chilean Senate require that every amendment considered during the second reading of a bill must be subject to a roll call vote. Given the incentives built in to the that committee system for sincere voting, these roll call votes provide a valuable window on the policy preferences of the legislators who take part. Having set up a model for measuring party cohesion, it is instructive to apply it to one of these committees. The Labor Committee cast ample votes to permit estimation, while it dealt with a su ciently narrow subject matter to give one some con dence in the assumption of a unidimensional issue space. The Senate Labor and Social Provision Committee has jurisdiction over laws a ecting the labor market, including Chile's innovative pension system. However, during the period under study no major bills a ecting the pension system were considered. The Committee's work focussed instead on more the more conventional labor relations issues that are the bread and butter of ideological politics throughout the world. The government and the opposition lined up largely as one would expect, with every member of the Concertaci
n advocating o policy positions to the left of those espoused by every member of the opposition. However, there was considerable heterogeneity among the positions taken by Senators on the right, with Institutional Senator William Thayer taking a position close to the Concertaci
n, even o as another Institutional, Olga Feli
staked out a position far to the right of even the most u conservative of the elected Senators serving on the Committee. A substantive analysis of the proposals made by di erent Senators, and of the positions they took during the Committee's debates con rms the message of the parameter estimates: Senator Thayer did indeed take a position to the left" of the remainder of the opposition, on questions from working hours to wage oors and grievance procedures. Why would former President Pinochet have appointed an advocate of such positions? One possibility is that he made a mistake in selecting Thayer as one of the Institutionals. Another, and perhaps even less likely possibility is that his true policy preferences on labor issues resemble those of the Christian Democrats, though his 16 and a half year record as dictator provides considerable evidence against this interpretation. Another possibility, the interpretation preferred by this author, is that Pinochet knew what he was doing when he appointed William Thayer. In this interpretation Pinochet's primary goal for the Institutionals was to protect the interests of the military; no human rights prosecutions, no budget cuts for the armed forces. By including a pro-labor appointee among the Institutionals, Thayer was one of only two appointed directly by Pinochet, he reduced pressure from labor to abolish the Institutionals, thereby protecting their ability to perform their primary function of safeguarding the interests of the armed forces. Supporting evidence for this is given by Thayer's well known sympathy for Labor Unions. Not only had he served in the cabinet of the Frei administration during the sixties, but he is co-author of one a well known text on labor law, whose tone is su ciently pro-union that members of the Concertaci
n sometimes bring it to Senate debates, quoting Thayer's own o 24
words in Senate debates. In the words of Christian Democratic Senator Ricardo Hormaz
bal, a Thayer is a prisoner of his own history" . Moreover, Thayer recounts that at one point during the period of military rule Pinochet invited him to be Secretary of Labor. Thayer turned him down, saying people are going to expect me to establish freedom to unionize, and it isn't the hour." . Pinochet would have had to have been very forgetful indeed not to anticipate that Thayer would take a moderate ideological position on labor issues. As noted earlier, four normalizations are required to identify the model. These amount to the choice of a reference proposer" for whom the proposal parameters g and c are normalized to equal 0, and two reference voters used to anchor the scale on which preferred outcomes are measured. Here the National Renovation party is chosen as the reference proposer, and so sponsors with g 0 propose to the left" of National Renovation, those with g 0 propose farther to the right" than does National Renovation. In addition, two voter locations must be pinned down. Many ideological ratings groups normalize their best friends at a location of 100 and their worst enemies at a location of 0. Here I adopt a normalization with more neutral connotations, placing Socialist Rolando Calder
n on the left, with a location of ,1, o and Institutional Senator Olga Feli
on the right with a location of 1. u In addition to these normalizations, one other was necessary. The only time that the Institutional Senators voted together they voted the same way. Thus the maximum likelihood algorithm tries to impute a correlation of 1 for the random error terms associated with these two Senators. Such perfect coincidence seems far fetched, especially in light of the great di erences in policy preferences exhibited by these two Senators. Accordingly I constrain the correlation between their shocks to equal 0:999, just less than perfect correlation. This corresponds to a value of z, the transformed correlation coe cient, of 636:62 Parameter estimates appear in table 1. The rst column, headed by x, reports estimates of Senators' most preferred policy outcomes. As noted above, the preferred policy outcome of Socialist Rolando Calder
n was normalized to equal ,1, while the most preferred policy o outcome of Institutional Senator Olga Feli
was normalized to equal 1. The remaining u Senators estimates are all taken with reference to these two. On the left we encounter Senator Ruiz De Giorgio, a trade unionist, with an estimated preferred policy outcome of ,0:997, virtually identical to the preferred outcome for the Senator Calder
n. Also on the o left, and virtually indistinguishable from Senators Calder
n and Ruiz is Christian Democrat o Humberto Palza, with an estimated preferred outcome of ,0:982. Additional evidence about the ideological preferences of these three Senators is provided by their estimated values of b. Recall from section 1 that a legislator's value of b calibrates the probability the legislator votes in favor of an amendment from the reference proposer, which in this case corresponds to the members of the National Renovation Party. This means that in the context of Table 1 higher values for b imply a higher probability of voting for
37 38 39
Interview with the author, May 1997, Valpara
so Chile. i Interview with the author, May 1997, Santiago Chile. 39 Constraining this parameter to equal zero causes a trivial decline in the log of the likelihood function, it is kept near 1 because this is more consistent with observed behavior. Renormalizing zInst = 0 does not a ect the remaining results.
37 38
25
Participant Partido Democrata Cristiano
Table 1: Parameter Estimates for the Labor Committee Parameter
. . -8.939 -8.973 -8.975 . . 4.297 4.188 204.446 Ricardo Hormaz
bal S. a -0.895 1.302 . . . a 0.567 0.084 . . . Humberto Palza C. -0.982 0.975 . . . 0.075 0.595 . . . Jos
Ruiz De Giorgio e -0.997 0.672 . . . 0.094 0.618 . . . Union Democrata Independiente . . 2.205 -4.225 . . . 3.225 2.974 . Institutional Senators . . 3.458 -3.376 636.62 . . 4.020 3.463 . Olga Feli
S. u 1.000 18.401 . . . . 8.808 . . . William Thayer A. -0.797 2.530 -8.685 -8.073 . 0.116 0.662 4.453 4.353 . Renovaci
n Nacional o -0.595 3.869 0.000 0.000 . 0.184 0.806 . . . Miguel Otero L. -0.012 8.956 . . . 0.434 2.308 . . . P.S. P.P.D. . . -12.173 -9.608 . . . 4.832 4.822 . Rolando Calder
n A. o -1.000 0.222 . . . . 0.501 . . . Executive Proposals . . -9.692 -5.687 . . . 4.557 4.473 . Own Party Proposals . 0.334 . . . . 0.463 . . . a Estimated standard errors are show in parentheses. ^ ^ = 4:866, Sd ^ = 0:562, log lik = ,280:622, n = 244 amendments.
x
b
g
c
z
26
proposals from the National Renovation Party. The estimated value for bCalderon of 0:222 is lower than for any of the other senators on the Committee, and indicates a barely better than 50 chance that this Socialist senator would vote for a proposal written by a member of the National Renovation Party, such as Senator Per
z. However, Senators Ruiz and Palza, e with estimated values of bRuiz = 0:672 and bPalza = 0:975 have the next lowest probabilities of voting for proposals from the National Renovation Party. Moreover, the values of b are estimated with a fairly high degree of precision for all three of these senators. We can test the null hypothesis that all three senators share a common preferred outcome, x, and a common mean for their idiosyncratic preference disturbances b, using a statistic with 4 degrees of freedom. The resulting value of the test statistic of 2.9656 corresponds to a p-value of 0.5636, leading to acceptance at all standard signi cance levels. Senator Hormaz
bal, a Christian Democrat and the remaining member of the Cona certaci
n to have voted on the Labor Committee also has an estimated preferred outcome o near ,1, though it's estimated value of ,0:895 is over a standard deviation to the right of that for Senator Calder
n. In addition the estimated value of bHormazabal of 1:302 is higher o than that for Senator Calder
n, indicating that this Christian Democratic Senator is more o likely to vote for proposals sponsored by the National Renovation Party than is Senator Calder
n. Testing the null hypothesis that in fact Senators Calder
n and Hormaz
bal share o o a the same preferences, with matching values of x and b leads to a test statistic of 3:8318, corresponding to a p-value of 0.1472, not su cient to warrant rejection of the null hypothesis though it does suggest that Senator Hormaz
bal is slightly to the right of his fellow member a of the Concertaci
n Rolando Calder
n. o o Although the Christian Democratic Senators frequently voted together on the same amendments, much of the similarity in their voting records can be attributed to their having similar preferences. The estimated z parameter of ,8:975 for the Christian Democrats corresponds to a negative correlation of ,0:929, indicating the antithesis of party discipline. However, in light of the very large estimated standard error of 204:446 this surprising estimate di ers insigni cantly from 0. The negative value appears to be an artifact of the imprecision of the parameter estimate, and so does not provide much evidence on the question of party discipline. Notice that, just as the party whip" parameter, z was insigni cantly di erent from 0 for the Christian Democrats, so too the estimated shift to the b parameter when a Senator votes on proposals from members of his own party is precisely estimated, with a standard error of only 0:463, less than for any of the estimated b values, and insigni cantly di erent from 0: the estimated value of 0:334 is less than three quarters of a standard deviation from 0. This is further evidence in favor of party cohesion stemming from a coincidence of preferences rather than from party discipline. The picture of voting by Concertaci
n Senators that emerges from this analysis is one of o a group of Senators who largely agree about policy, and cast similar votes for this reason. The data do not permit us to reject the hypothesis that the high observed degree of cohesion among the Christian Democrats on the Labor Committee is entirely attributable to similarity of of outlook. Even the di erences between the most rightward" of the Christian Democrats,
2 2 2
27
Ricardo Hormaz
bal, and the most leftward' member of the Concertaci
n on the commita o tee, Socialist Senator Rolando Calder
n fall short of statistical signi cance. Moreover, the o estimated magnitude of these di erences is small when contrasted with the di erences among members of the opposition, and between the opposition and the Concertaci
n, and it is to o these di erences that we now turn. In contrast with the very homogeneous preferences of the Concertaci
n Senators on the o Labor Committee there is considerable heterogeneity among their counterparts in the opposition. Most striking is the gap between Senator Thayer, with an estimated preferred outcome of ,0:797, over 15 standard deviations to the left of his fellow Institutional Senator Olga Feli
, whose preferred outcome is normalized to equal 1. These two Institutional u Senators bracket the remaining opposition Senators, both in terms of their estimated preferred outcomes, and their b parameters; bThayer = 2:530 indicates that Senator Thayer is less likely than any of the other opposition Senators to vote for a proposal from the National Renovation Party, while the estimated value of bFeliu = 18:401 indicates that Senator Feli
u is even more likely to vote for proposed amendments sponsored by members of the National Renovation than the party's own Senators! A test of the null hypothesis that these two Senators have the same values for x, and b leads to a test statistic of 414.6028, literally leading to o the charts rejection . These parameter estimates indicate that, far from acting as a homogeneous block on the far right, the Institutional Senators exhibit remarkable di erences in their policy preferences, at least in the ambit of labor issues. This nding is all the more remarkable in light of Senator Thayer being one of the two Institutionals appointed directly by then President Pinochet , while Feli
was appointed by the Supreme Court. In one sense it is not surprising that u William Thayer, a former Christian Democrat would take a position closer to the members of his own party than Senator Feli
. Perhaps Pinochet simply miscalculated when he appointed u Thayer to the Senate, expecting him to take a position more like that of Senator Feli
. u By the time Thayer began voting in the Senate Pinochet was no longer in a position to remove him. Yet his great age, he was 71 when he took o ce in March 1990, suggests that Senator Thayer, is probably not seeking reappointment by a civilian president when his current eight year term as Institutional Senator expires, so he is presumably voting his conscience. An intriguing question is whether Pinochet anticipated this when he made the appointment. Certainly there was little evidence during Pinochet's sixteen and a half years of rule to indicate that he was anything but suspicious of trade unions. Yet if he saw the Institutional Senators as primarily guardians of the corporate interests of armed forces, in avoiding human rights prosecutions and cuts in the military budget, perhaps he sought to reduce pressure for Constitutional reform by giving labor unions less reason to want to eliminate the institutionals as a means of obtaining more pro-labor legislation. Because most of the National Renovation Senators who served on the Labor Committee did so only brie y they must be grouped to permit individual estimation. This means that a common ideal point is estimated for the National Renovation Senators, less Senator Otero. 40 The corresponding p-value is 9:335 10,91 .
40 2 2 41
41
The other direct appointee was Sergio Fern
ndez F.. a
28
While Senator P
rez cast slightly more votes than Senator Otero, I estimate Senator Otero's e preferred outcome separately from the other members of his party to permit a subsequent comparison with his estimated ideal point on the Constitution Committee, where he also served. As with the Institutional Senators the members of the National Renovation Party do not take identical positions on Labor policy, with Senator Otero's preferred policy position of ,0:012 considerably to the right of the remaining members of his party, who are estimated to take a position of ,0:595. Likewise, the estimated value of b for Senator Otero is substantially higher than for the remaining members of his party on the Labor Committee. A test of the null hypothesis that Senator Otero and the remaining Senators from the National Renovation Party have the same preference parameters x and b leads to a statistic of 8.0725, corresponding to a p-value of 0.0177, indicating rejection at = 0:10 and = 0:05, though not at the more stringent = 0:01 signi cance level. These estimates indicate the odds are more than 55 to 1 against the estimated di erences between the preference parameters of Senator Otero and those of the remaining members of his party being a mere result of estimation error. Similar tests reveal signi cant di erences between Senators Otero and Thayer , and between Senator Otero and Senator Feli
. u While the estimated position of Senator Thayer, to the left of all of the other opposition Senators on the Labor Committee is surprising it is nevertheless the case that his position as the leftmost of the opposition Senators is still to the right of the rightmost of the estimated preferred policies of the Concertaci
n Senators, that of Senator Hormaz
bal. Likewise, o a while Senator Thayer's estimated value for b indicates that he is less likely than any other member of the opposition to vote for a an amendment from the reference proposer, the National Renovation Party, he is nonetheless more likely to do so than any member of the Concertaci
n. In fact, whether we rank the Senators on the Committee using their values for o x or b we come up with the same ordering, with Senators whose estimated preferred policies, the x's, are to the right also having higher estimated values for b. The estimated preference parameters for Senator Hormaz
bal and the National Renovaa tion Senators excluding Senator Otero is statistically signi cant, leading to a statistic of 10.074, corresponding to a p-value of 0.0064, and indicating rejection of the null hypothesis that Senator Hormaz
bal and the National Renovation Senators share the same values for x a and b. In contrast to this unambiguously signi cant gap, the di erences between the intermediate preference parameters of Senator Thayer and those of the more moderate Senators of the National Renovation Party is just at the threshold of signi cance , while a similar comparison with the estimated preferences of the rightmost member of the Concertaci
n o delegation to the Labor Committee, Senator Hormaz
bal, falls just a shade below statistical a signi cance . Similar tests lead to borderline evidence of a di erence between the preferences
2 2 42 43 2 2 44 45
The 2 statistic takes on a value of 10.058, corresponding to a p-value of 0.0064 2 The 2 statistic takes on a value of 14.324, corresponding to a p-value of 0.0008 2 44 A test of the null hypothesis that Senator Thayer and the National Renovation Senators less Senator Otero share the same x and b parameters generates a 2 statistic of 4.779, corresponding to a p-value of 2 0.0917, indicating rejection at = 0:10 but not at either = 0:05 or = 0:01 45 A test of the null hypothesis that Senators Hormaz
bal and Thayer share the same x and b parameters a generates a 2 statistic of 4.274, corresponding to a p-value of 0.1180, indicating acceptance even at = 0:10. 2
42 43
29
of Senator Thayer and each of the remaining Christian Democrats on the Committee , and more decisive evidence of a di erence between Senator Thayer and Socialist Senator Rolando Calder
n . These results are consistent with the hypothesis that Senator Thayer occupies o an intermediate position, between the left of the National Renovation Party and the right of the Christian Democrats. The proposal parameters provide additional evidence about the di erences between Senator Thayer and the remaining members of the opposition. Recall that the value of g for the National Renovation Party has been normalized to equal 0. The estimated values of g for the Institutional Senators less Senator Thayer and the UDI Senators are statistically insigni cantly to the right of the normalized value for the National Renovation Senators. Because of the relatively small numbers of proposals coming from each individual it was not practical to estimate di erent values of the proposal parameters for each Senator, hence the party-wide parameter estimates for National Renovation and UDI. Because of the surprising evidence about Senator Thayer's preference parameters an individual set of proposal parameters has been estimated for this Senator. The proposal parameters for the Christian Democrats and for President Aylwin, also a Christian Democrat, are very similar, and signi cantly to the left of the normalized value for the National Renovation Senators, with t-ratios of -2.080 and -2.127 respectively. The Socialist PPD Senators who frequently cosponsor amendments tend to propose somewhat leftward of the other members of the Concertaci
n, the estimated gap, with the PS PPD o 3.217 to the left of the Christian Democrats is statistically signi cant at = 0:05 but not at = 0:01 . Senator Thayer again emerges well to the left of the remainder of the opposition. The t-ratio for the di erence between estimated value of g for Senator Thayer and the value of gRN = 0 for the National Renovation Senators is -1.950, corresponding to a p-value of 0.0511, so that we can reject the null hypothesis that he and the National Renovation Senators make proposals with the same ideological content at = 0:10 but not at = 0:05. A test of the hypothesis that Senator Thayer and the remaining Institutionals make ideologically identical proposals leads to similar results . In contrast there is virtually no di erence between the estimated value of g for Senator Thayer of ,8:685 and the corresponding value of ,8:938 for the Christian Democrats , though his estimated value for g is signi cantly rightward of the Socialists . The net proposal quality parameters, which amalgamate proposers' valence and ideo46 47 48 49 50 51
The 2 statistic corresponding to the null hypothesis that Senators Thayer and Palza share the same 2 values for x and b of 5.164 corresponds to a p-value of 0.0756, the 2 statistic for the hypothesis of identical 2 preferences for Senators Ruiz and Thayer is 5.491, with a p-value of 0.0642. 47 The 2 statistic is 10.058, corresponding to a p-value of 0.0065, leads to rejection even at = 0:01. 2 48 The 2 statistic of 5.9604 corresponds to a p-value of 0.0146. 1 49 The 2 statistic of 3.541 corresponds to a p-value of 0.0598, signi cant at = 0:10 but not = 0:05. 1 50 The p-value for the null hypothesis that Thayer and the Christian Democrats have the same value for g is 0.7553 51 The p-value for the null hypothesis is that Thayer and the Christian Democrats have the same value for g is 0.7553
46
30
logical parameters, as discussed in section 1 of the preceding Chapter, are mostly similar, though the Christian Democrats and PPD have values that are just at the threshold of being statistically signi cantly lower than those of the National Renovation Party, with t-ratios of -2.142 and -1.993 respectively. Because Senator Thayer, and the Christian Democratic Senators have virtually the same estimated values for g and similar values for x, the higher estimated value of c for Thayer suggests that he was better able to formulate high valence proposals than were the Christian Democratic Senators, though the estimated di erence is nowhere near the threshold of statistical signi cance. While no estimate of President Aylwin's preferred policy outcome x is possible, as he does not vote in the Congress, it seems reasonable to expect that it is very similar to the ideal points for the remaining members of his party. Given that the estimated value of g for President Aylwin and the Christian Democrats on the Committee are almost identical, we might expect that his higher estimated value for c means the President tends to draw higher valence proposals. The high and statistically signi cant estimate for of 4.866 indicates that variations in the consensus value of proposals swamp idiosyncratic preference shocks, whose variance has been normalized to equal 1. However, this high value for , while it is su cient to produce a large number of unanimous votes, does not render the ideological content of proposals irrelevant. To put it in context, Socialist Senator Rolando Calder
n will be approximately o indi erent between a proposal from another Socialist Senator with an average consensus shock and a proposal from a National Renovation Senator with a consensus value drawn from the top sixth of the distribution. While consensus quality matters, it takes a lot of it to overcome Senators' ideological leanings.
5 Conclusions
The approach developed here provides a means of identifying two distinct forms of party cohesion; a nity" and discipline". While these are not mutually inconsistent forms of behavior, they are di erent. The technique is applied to the Labor Committee of the Chilean Senate. Chilean Senate committees create powerful incentives for Senators to vote nonstrategically. This facilitates the analysis of roll call votes. The analysis reveals the a nity hypothesis, which postulates that Senators share common objectives, applies to the parties on the left, and somewhat among the elected Senators on the right, while the Institutional Senators exhibit marked di erences in their ideological a nities. This may be due to an intentional strategy on the part of former president Pinochet to limit pressure for the abolition of the Institutional Senators. None of the parties exhibit signi cant levels of either party whipping" or deference to proposals of party members. The methods used here can be generalized to other legislative settings with larger numbers
31
References
Agor, Weston H. 1971 The Chilean Senate Austin: University of Texas Press . Arnold, R. Douglas, The Logic of Congress New Haven: Yale, 1991 . Birnbaum, A. 1968 Some Latent Trait Models and Their Use in Inferring an Examinee's Ability." in Statistical Theories of Mental Test Scores. F. M. Lord and M.R. Novick eds. Reading: Addison-Wesley pp. 395 , 479. Birnbaum, Je rey H. and Alan S. Murray Showdown at Gucci Gulch. New York: Random House, 1987 pp. 174-5. Cahoon, Lawrence S., Melvin Hinich, and Peter C. Ordeshook 1978 . A Statistical Multidimensional Scaling Method Based on the Spatial Theory of Voting." in Graphical Representation of Multivariate Data. P. C. Wang ed.. New York: Academic Press . Chamberlain, Gary 1984 Panel Data" in Handbook of Econometrics Vol. 2, Zvi Grilliches and M. Intriligator eds. pp. 1248-1318. Chester, W. Harris 1948 A Factor analysis of Selected Senate Roll Calls, 80th Congress," Educational and Psychological Measurement Vol. 8, pp. 582-91. Cox, Gary W. and Matthew D. McCubbins 1993 Legislative Leviathan: Party Government in the House Berkeley: University of California Press . Denzau, Arthur T., William Riker, and Kenneth Shepsle. 1985 Farquharson and Fenno: Sophisticated Voting and Home Style." American Political Science Review Vol. 79 pp 111733. Fischer, Gerhard H. 1995 Derivations of the Rasch Model" in Rasch Models: Foundations, Recent Developments and Applications Gerhard H. Fischer and Ivo W. Molenar eds. New York: Springer Verlag pp. 15-38. Gage, N.L. and Ben Shimberg 1949 Measuring Senatorial Progressivism," Journal of Abnormal Social Psychology Vol. 44 pp. 112-17. Gilligan, Thomas W. and Keith Krehbiel 1987 Collective Decisionmaking and Standing Committees: An Informational Rationale for Restrictive Amendment Procedures." Journal of Law, Economics and Organization 3:287-335. Haberman, S. J. 1977 Maximum Likelihood Estimates in Exponential Response Models." The Annals of Statistics 5:815-41. Kingdon, John W. Congressmen's Voting Decisions. New York: Harper and Row, 1981 . Ladha, Krishna K. 1991 A Spatial Model of Legislative Voting with Perceptual Error." Public Choice 68:151-74. Londregan, John B. 1996 Estimating Preferred Points in Small Legislatures: Why We 32
Can't Remain Agnostic." typescript. Londregan, John B. 1997 Faustian Bargains: Legislative Institutions and Chile's Democratic Transition, typescript. Lord, Frederic M. 1983 Unbiased Estimates of Ability Parameters, of Their Variance, and of Their Parallel Forms Reliability." Psychometrika 48:477-82. Lott, John R. 1996 Political Cheating." Public Choice Vol. 52 pp. 169-86. MacRae, Duncan 1954 The Role of the State Legislator in Massachusetts," American Sociological Review Vol 19, pp. 185-94. Maddala, G. S. 1986 Limited Dependent and Qualitative Variables in Econometrics New York: Cambridge University Press . McGraw, Kathleen M. 1991 Managing Blame: An Experimental Test of the E ects of Political Accounts." American Political Science Review Vol. 85 No. 4, pp. 1133-1157. Platt, Glenn, Keith T. Poole, and Howard Rosenthal. 1992 Directional and Euclidean Theories of Voting Behavior: A Legislative Comparison." Legislative Studies Quarterly 17:56172. Poole, Keith T. 1988 Recent Developments in Analytical Models of Roll-Call Voting in the US Congress." Legislative Studies Quarterly. vol. 13, pp. 117-33. Poole, Keith T. and Howard Rosenthal. 1985. A Spatial Model for Legislative Roll Call Analysis." American Journal of Political Science 29:357-84. Poole, Keith T. and Howard Rosenthal. 1991. Patterns of Congressional Voting." American Journal of Political Science 35:228-78. Poole, Keith T. and Howard Rosenthal 1996 Congress: A Political-Economic History of Roll Call Voting. New York: Oxford University Press . Rasch, Georg 1961 On General Laws and the Meaning of Measurement in Psychology." in Proceedings of the Fourth Berkeley Symposium on Mathematical Statistics and Probability 4:321-33 Berkeley: University of California Press . Reglamento del Senado de Chile. 1993 . Sesiones del Senado de Chile. Various issues, 1990-1994. Stokes, Donald. 1963 Spatial Models of Party Competition." American Political Science Review. 57:368-77. Thurstone, L.L. 1925 A method of scaling psychological and educational tests." Journal of Educational Psychology Vol. 16, pp. 433-51.
33