*I am grateful for useful comments from John Brehm, and from seminar participants ...... Gerry and King , and the Carolinas Williamson, Butler, Rutledge, and the.
Deliberation and Voting at the Federal Convention of 1787
by John Londregan UCLA Version 0:2 | July 12, 1999
I am grateful for useful comments from John Brehm, and from seminar participants at the 1998 Annual Meetings of the American Political Science Association and the 1999 Midwest Political Science Association Meetings. 1
Abstract
This paper examines the deliberative voting of the Federal Convention of 1787, and contrasts this type of voting with the more commonly observed position taking behavior that characterizes most legislatures. The analysis constructs an empirical model that incorporates proposal valence, which at the federal convention corresponded to proposals' contribution to the \rati ability" of the constitution, and uses information contained in the authorship of proposals to overcome the identi cation problems that plague empirical spatial models of voting. The estimated issue positions of the state delegations reveal a signi cant cleavage on an two issue dimensions; one corresponding to the balance between the states and the central government and the other dealing with the extent of the powers granted to the federal government.
Introduction While most legislatures refer to their public debates as \deliberation", this is rarely an apt description. Instead most public legislative activity is used by legislators to clarify their issue positions to their constituents, thereby solidifying their claim to a particular piece of ideological turf, and reinforcing their image as tireless advocates of the special interests of their reelection constituencies. Real deliberation, in which legislators persuade one another to change their views on policy choices is rare, and usually takes place well away from the public spotlight. The \anaerobic" nature of deliberation, which does not survive the fresh air of publicity, means that it is much more dicult to study. The walls and heavy curtains of the smoke- lled rooms within which such debates transpire block not only the gaze of the contemporary public, but also the retrospective view of political scientists interested in studying the structure of the process. At best one is left to rely on the often self-serving and nostalgic accounts of former participants in such discussions. An unusual and prominent exception to the general invisibility of legislative deliberation is provided by the records of the Federal Convention of 1787. This body met in secret and adopted a set of rules that discouraged agenda manipulation and position taking, and favored deliberation. Moreover, the Convention met at a time of crisis, as the Articles of Confederation were proving an inadequate framework for government. Members of the convention shared a powerful and urgent desire to formulate a politically viable constitution for a national government. Serendipitously, multiple records of these deliberations exist, and they include not only summaries of the speeches given by the various members, but also a record of the 568 roll call votes1
1 The Convention's Journal numbers the votes through 569, but only 568 of these votes actually took place, with one proposal, Number 299, being withdrawn on August 15, 1787 before it could be voted.
1
taken during its three and a half months of meetings. These votes permit us to study deliberative politics through the lens statistical analysis. The estimator used here is adapted to deal with the special problems that haunt ideal point estimation, and which become particularly severe in small legislatures (Londregan, 1999). In addition, I depart from the agnostic norm in dealing with multidimensionality of the issue space. Many analysts estimate spatial locations for legislators using an arbitrary multidimensional issue space, (Heckman and Snyder, 1996), (Poole and Rosenthal, 1996). This is satisfying from one standpoint, in that it \lets the data speak for themselves" about the number of dimensions needed to represent the issue space. Unfortunately for those interested in substance, this leaves analysts to engage in an ex post exercise of assigning ad hoc substantive interpretations to these abstract dimensions. The approach taken here attempts to break away from this arbitrariness by allocating votes to substantive subsets before the estimation takes place. This leaves the analyst in a position to test whether di erent substantive subsets of votes tap the same underlying spatial dimension or not. Of course this procedure is vulnerable to ex ante arbitrariness in the analyst's assignment of votes to substantive categories. The ultimate choice between these two approaches depends on how the analyst weighs the extra work of classifying votes on the basis of their substance against what is probably a higher risk of misclassi cation based on the more commonly used agnostic approach. The resulting estimates reveal a multidimensional issue space, with sharp disagreements about the distribution of power between large states, which preferred seat shares in the legislature to be proportional to population, and the small states that preferred to give each state an equal voice in Congress, regardless of its population. Also intense were disagreements about a set of issues described by Jilson (1981) as \the scope and empowerment of the federal government" which included how commerce should be regulated, and 2
how the national government would treat the peculiar institution of slavery. The parameter estimates shed some light on the question of why deliberations are often referred to with adjectives such as \long" and \drawn out". The parameter estimates reveal that the rate at which \high valence" proposals emerge from deliberation, which is roughly governed by the standard deviation of the common shock to preferences in the model, is low; the expected number of proposals that must be vetted before those with suciently high valence to produce compromise emerge is large. The Federal Convention of 1787 succeeded despite having to prospect for compromise in the very rocky soil of the politics of the Federalist Period primarily because the delegates persevered, continuing to search for the elusive institutional adjustments that would make the proposed Constitution politically viable. Section One of this paper brie y sketches the historical context in which the Federal Convention of 1787 met, and then goes on to outline the design of the rules thwarting both position taking and strategic manipulation of the agenda. Section Two presents a model of deliberation and compromise when political actors share a common goal but nevertheless have di erent interests in how the details are resolved. Section Three presents parameter estimates for a two-dimensional version of this model, and uses these estimates as a window on the Convention's substantive debates about the division of power among the states and the central government, and about a set of issues described by Jilson (1981) as \the scope and empowerment of the federal government" which included how commerce should be regulated, and how the national government would treat the peculiar institution of slavery. A fourth and nal section summarizes the substantive conclusions and discusses their implications about the nature of deliberation.
3
1 Background The records of the Federal Convention that met during the Summer of 1787 in Philadelphia, have been heavily studied by legal scholars in search of \legislative intent", and the participants are frequently referred to with a reverence usually reserved for religious prophets. It is thus very easy to forget that the Constitution recommended by this convention only enjoyed the support of about half of the voting public, and that its supporters only succeeded in winning approval by cleverly manipulating the timing of the state ratifying conventions, and by promising to immediately amend the Constitution with a Bill of Rights (Riker, 1996). The Federal Convention of 1787 was not in a position to return to the states with a constitution carved onto stone tablets and assured of immediate acceptance. Instead it was a gathering of supporters of a strong national government, subsequently referred to as the \Federalists"2 who faced a substantial opposition of \AntiFederalists" advocating a more di use confederation of relatively independent states. In contrast with a representative legislature which contains a cross section of opinion the Federal Convention of 1787 was largely a meeting of partisans with the shared goal of creating a strong national government. Evidently Anti-Federalists, such as Patrick Henry who was selected as part of Virginia's delegation to the Convention but who refused to attend, thought that they could better discredit the Convention's likely recommendation of a strong national government if they were not a part of the deliberations, something that Riker (1996) characterizes as a political blunder. The delegates to the convention thus shared the common goal of producing a strong national government even as they sought to in uence its nature to their own advantage. Because the recommended constitution had to be approved by the states the politicians gathered in Philadelphia needed to be 2 Most of these people subsequently became members of the \Federalist Party" of Washington and Adams. 4
careful; the document they recommended could not contain provisions that would be unacceptable in the states. There was the very real danger that a coalition that pressed its demands too far could win the battle over inclusion of their preferred provisions in the recommended constitution only to lose the war as the provision's unpopularity dragged the proposed constitution to defeat at the state ratifying conventions. Members of each state delegation had considerable political expertise on the likely reaction in their own state to a proposed constitution, and much of the deliberation at the convention involved sharing information about the kind of national government the di erent states could tolerate. This setting created powerful incentives for politicians to take positions for later public consumption, and to attempt to manipulate other delegates beliefs about what their respective states would accept. The enormous potential for this latter type of manipulation was made clear at the outset by the instructions given by their small state to the Delaware delegation, whose credentials contained the stipulation that they were not allowed to alter the status quo language in the fth article of confederation that \In determining Questions in the United States in Congress Assembled each State shall have one Vote."(Farrand, 1966c) p. 575. The small states were reluctant to concede representation proportional to population in Congress, and Delaware, the smallest state to participate in the convention, used its instructions to its delegates to exaggerate its commitment to the \one state, one vote" rule: in the end Delaware backed down from this position, ratifying a compromise that gave them something less. To vitiate the powerful incentives for position taking3 the Convention adopted a rule of secrecy, requiring \That nothing spoken in the House be printed or otherwise published or communicated without leave."(Farrand, 3 See the Appendix for a formal model of the factors favoring position-taking by legislators.
5
1966a) p. 15, a rule that the delegates with few exceptions respected4 . This shroud of secrecy which was preserved for many years after the convention5 essentially removed the incentive to take positions for \public consumption", and allowed members to \climb down" from positions they had taken and defended without incurring \audience costs". Other rules dealt with the incentives to manipulate the agenda. One of these imposed an open rule; \A question which is complicated, shall, at the request of any member, be divided, and put separately upon the propositions, upon which it is compounded."(Farrand, 1966a) p. 8, while another allowed a majority of the Convention to revisit any \matter which had been determined by a majority". These measures made the agenda both hard to predict6 and dicult to manipulate. Another institutional feature of the Convention has had an important impact on political scientists' ability to analyze its deliberations, and that is the small number of voters. Each state's delegation cast a single vote according to a \unit rule". If the majority of the delegation favored a proposition then the state's vote was tallied as \aye", while majority opposition within the delegation led to a recorded vote of \nay". Only on the rare occasions on which a state's delegation were evenly split does the record note that the state's vote was \divided". Thus the votes of individual delegates are not recorded. Moreover, given Rhode Island boycotted the convention, while the attendance dates of the New York delegation, which left early, and the New Hampshire contingent, who arrived late, did not overlap, there were never more than eleven votes cast on any question.
4 New Hampshire delegate Nicolas Gilman violated this requirement of secrecy in letters to his brother John, and to Joseph Gilman(Farrand, 1966c) p. 66., while Joseph Jones, in his correspondence with James Madison reports some minor \leaks"(Farrand, 1966c) p. 80. 5 The Journal of the Convention was not published until 1818, while the notes of Madison, Yates, and several other participants did not reach print until even later on. 6 See the Appendix for a more extended discussion of the importance of unpredictability in limiting the incentives for strategic voting.
6
This raises the well known problem of preference aggregation; the conditions under which the voting position taken by a multimember delegation can be thought of as corresponding to that delegation's median voter are stringent. However, even if one is willing to believe that these conditions hold, at least approximately, a second statistical issue complicates the analysis of the 568 roll call votes taken at the Federal Convention. When the number of voters is small then standard \agnostic" statistical techniques that simultaneously attempt to estimate both the preferred spatial locations of members and the locations of the bills on which they vote are not identi ed, see Londregan (1999). One response to this problem is to incorporate the proposal-making process into the voting model, and this is the approach taken here.
2 A Model of Deliberative Voting This section sets forth a simple variant of the spatial model that is useful in clarifying some of the key concerns of delegates at the Federal Convention. I then go on to incorporate the proposal-making process into the analysis, and to construct a version of the model that is amenable to statistical analysis.
Issue Valence and Issue Positions The model set forth here builds on the distinction drawn by Stokes (1963) between \position" issues, such as minimum wages and gun control, and \valence" issues, such as honesty and integrity, about which voters all agree. Following the development in Chapter One of Londregan (1996) the model here recognizes that most issues combine elements of both position and valence. For example Delaware's position of one vote for every state would have yielded the preferred position for that state's delegation on the continuum of representation schemes with equal representation for each state at one end, 7
and representation proportional to population at the other, however, such a clause in the recommended Constitution would have been very low on the valence issue of rati ability: there is a high probability that voters of the larger states would have rejected a constitution with a unicameral legislature that gave each state equal representation. We can capture preferences over proposals with both position and valence using the following linear-quadratic utility function: 1 (1) 2 This says that a proposed policy with position p and valence v will appeal more to a legislator with a preferred position of x the closer the policy position p is to x, and the higher the policy valence, v . The higher the value for the greater the intensity of caring about the valence aspect of policy relative to the position element. In Figure 1 we can let the horizontal axis represent the degree to which Congressional representation favors small states, with outcomes more favorable to the small states on the left, and outcomes favoring the more populous states on the right. Thus we might expect a state such as North Carolina, where population was expected to grow, to have a preferred outcome at a point like xNC in the Figure 1, with the allocation of seats in the legislature proportional to population, while for New Jersey the preferred representation rule corresponded to the status quo under the articles of confederation, with one vote per state, at a point like xNJ . On June 11, immediately after the Convention had adopted a resolution by James Wilson and Charles Pinckney establishing proportionality to population in the House of Representatives, Connecticut delegates Oliver Ellsworth and Roger Sherman proposed equal representation for the states in the Senate. The result, with equal representation for the states in one house, and U (p; q ; x; ) = , (x , p)2 + v
8
representation proportional to population in the other corresponded to a position like of pCT . Moreover, the proposed compromise had the potential to appeal to the states, making it easier to ratify the proposed constitution. This contribution to rati ability is represented by the valence vCT , see Figure 1. To evaluate the attractiveness of the proposal delegates needed to compare it with the outcome they expected to prevail if the proposal was defeated. In early June a coalition of the largest states, Virginia, Pennsylvania, and Massachusetts, with the southern states of North and South Carolina and Georgia are widely viewed as having formed a dominant block, which could probably have forced through representation proportional to population in both houses, this would result in an outcome with North Carolina's preferred policy position, xNC and with a valence expected to correspond to something like vJune in Figure 1. The proposal at (pCT ; vCT ) lies above the indi erence curve for the New Jersey delegation through the point (xNC ; vJune), but below the hypothetical indi erence curve for the North Carolina delegation through the same point, which would lead New Jersey to favor the proposal, and North Carolina to vote against it, the outcomes that actually transpired. An important feature of the deliberative environment of the Federal Convention of 1787 was the evolution of delegates beliefs about valence. When a grand committee consisting of one member of each state delegation proposed almost the same arrangement7 , with representation proportional to population in the House and equal representation for the states in the Senate, the North Carolina Delegation switched its vote to the armative, and the measure passed. Five weeks of bitter debate on the representation question had convinced delegates that failure to compromise would put passage of the constitution in jeopardy. This meant a downward reassessment of the valence that would result if the large states were to insist on representation 7 Save that the states would each have two Senators, who would vote per capita rather than as a unit. 9
proportional to population, to a level of vJuly . The proposed compromise, at (pCT ; vCT ) is above both the hypothetical indi erence curve for New Jersey passing through the point (xNC ; vJuly ) and the corresponding indi erence curve for North Carolina. This would lead North Carolina to switch its vote, as it did, forming part of a narrow majority in favor of the so-called \great compromise".
Figure 1 The \Great Compromise". v
d
x
(pCT ; vCT )
xNJ
xNC
q
q
vJune
t
(xNC ; vJuly ) The presence of a valence issue, such as a shared interest in writing a viable constitution, creates room for compromise among delegations that would otherwise be gridlocked. In Figure 1 any movement toward the preferred outcome of New Jersey and away from the expected outcome with no compromise of xNC constitutes a move away from the preferred outcome for North Carolina, but the delegations were nevertheless able to compromise on a point like (pCT ; vCT ) because it o ered North Carolina enough extra assurance of a rati able constitution to compensate for having to tolerate a less favorable outcome on Congressional representation. As the delegates deliberated the provisions of the Constitution, their beliefs about the con10
tribution of di erent proposals to the valence issue of rati ability evolved. After watching the Convention bog down in acrimony over the question of equal representation for the states delegates updated their beliefs about the rati ability of a Constitution that did not contain a compromise on that issue. It was these evolving beliefs that created majority support for the great compromise.
Operationalizing the Model To close the model described in the previous section, we need to incorporate the origin of proposals. In other work I discuss the statistical obstacles confronting models that place little structure on the agenda Londregan (1999). To avoid the statistical pitfalls we require a structural model with a nite number of agenda parameters. This results in a likelihood function that is then maximized with respect to the voters' preference parameters and the parameters characterizing the agenda. The model here identi es voters' preference parameters using the structure placed on the agenda. This identi cation comes at a cost: the interpretation of the preference parameters is contingent on the distribution of proposals posited by the structural model. Starting with Equation 1 the condition for legislator i whose preferred outcome is xi to vote in favor of proposal j with position pj and valence vj when the individual's idiosyncratic preference shocks for the proposal and the status quo are v~ij and v~isq is:
, 21 (xi , pj )2 + vj + u~ij > , 21 (xi , psq )2 + vsq + u~isq (2) Letting wij = u~isq , u~ij , while ij = wij , E fwij g and bi = ,E fwij g, gj = pj , psq , and cj is a composite parameter8 we can rewrite this condition
as:
8 cj = (vsq , vj ) + gj
p +p sq j . 2
11
ij < gj xi + bi + cj
(3)
We can divide the parameters of this model into two groups; the parameters that pertain to the voters, (xi; bi), and the proposal parameters (gj ; cj ). As noted above, attempting to estimate a separate set of voter parameters (xi; bi ) for each legislator, and a separate set of proposal parameters (gj ; cj ) for each question brought to a vote is not a practical option. However, proposals do not emerge from a vacuum, they are proposed by legislators in pursuit of an agenda. We can adjust the model in Equation 3 by noting that proposals from the same individual share some important features. Following the development in Londregan (1999) I posit that gj = gp for all solo authored proposals from the same proposer, p, while for proposals with multiple authors gj is a weighted average of its coauthors' gp's. The composite parameter cj contains elements of valence, which are more likely to vary even among proposals from the same individual, and to capture this cj = cp(j ) + v~j where v~j is a random variable and p(j ) is the identity of the identity of the author of proposal j . As with the g parameter, for multiauthored proposals cj is taken to be a weighted average of its coauthors' cp(j) parameters. Substituting into the Inequality 3 we have: ij < gp(j ) xi + bi + cp(j ) + v~j
(4)
Here I operationalize condition 4 by positing that the ij and v~j are normally distributed.
3 Empirical Results The data used in this analysis come from (Farrand, 1966a) and (Farrand, 1966b). Farrand assembles the Journal for the Federal Convention which 12
was maintained by William Jackson, the ocial secretary, as well as day by day accounts written by various participants, most notably James Madison whose notes are often more extensive than those of the ocial Journal. These accounts provide a list of all 568 roll call votes taken at the convention, and record the votes cast by each state's delegation. However, in order to apply the model set forth in Inequality 4 we also need information on the authorship of the questions brought to a vote. In many cases this is not clear from the record, though most such votes are on relatively mundane questions such as motions to adjourn for the day. With the able help of my research assistant, Eduardo Aleman, I have been able to make more con dent attributions for 362 of the votes. Because a handful of proposals had authors who only proposed a few times, making estimation of the corresponding gp and cp parameters impractical, I included only proposals with relatively active authors9, and this leaves me with 354 useable roll calls. Analysts who have studied the Convention Speak of at least two coalition structures during the Convention. The rst of these dealt with the contentious issue of representation, the second with con icting ideas about the scope of government, including the details of regulating commerce, and the status of slavery. Ideally each of the 354 roll calls could be individually studied and classi ed as to its content. As a rst approach to the data I have taken advantage of earlier work by Jilson (1981), who identi es the con ict over the \scope and empowerment of government" with the roll call votes taking place between July 17 and August 28. As a rst approximation I treat the remainder of the Convention as being dominated by the issues surrounding representation. While this issue was laid to rest by the key vote on the so-called \Great Compromise" which took place on July 16, it reemerged during the nal weeks of the Convention as delegates put the Constitution 9 For the states of Pennsylvania, Delaware, Maryland, and North Carolina I estimate a set of composite parameters for the infrequent proposers in each delegation, but even this expedient did not leave enough data to permit estimation for the infrequent proposers from Massachusetts, New York, and Virginia. 13
into its nal form. Admittedly these classi cations are only as robust as the views of the insights of the historians and political scientists who have created them. The risk here is that these classi cations have themselves been \mined" from the data over two centuries of informal analysis.
Parameter Estimates The spatial model of the preceding section is estimated separately using data from the representation and the scope & empowerment deliberations. The model is then reestimated on the entire dataset, thereby imposing the constraint of unidimensionality, which is rejected. To estimate the model several normalizations are needed. First, the leftright ideological scale must be calibrated by arbitrarily xing the preferred outcomes of two delegations, much as one calibrates a temperature scale with two reference points, such as the freezing and boiling points of water. To identify the proposer parameters an additional normalization is needed, it is sucient to set the proposal parameters (gp; cp) for one proposer equal to zero. Then all other proposers can be identi ed relative to this reference individual. Here I have xed the preferred location of the New Jersey delegation at ,1, on the left of the scale, and the Virginia delegation at 1, on the right. This was done because of the close association of each of these delegations with competing drafts of the Constitution, respectively the New Jersey and Virginia Plans, whose consideration dominated the rst two months of the Convention. I have chosen James Madison, the most proli c proposer, to anchor the proposer parameters, and so the (gp; cp) parameters are set to 0 for Madison. Proposals from authors with values of cp above zero enjoyed greater consensus appeal than did those written by Madison, controlling for their issue position, those from authors with negative values of cp enjoyed 14
less consensus appeal. Likewise, those with negative values of gp made proposals to the \left" of Madison, those with gp > 0 proposed to Madison's \right". The bv parameters tell us about the tendency for delegation v to favor proposals from Madison, with positive values of bv indicating a greater tendency to vote for proposals from James Madison. These normalizations are used for both subsets of the data to permit comparison across the two dimensions. Parameter estimates for the representation deliberations appear in Tables 1a and 1b. These estimates paint a fairly clear picture of the primary line of cleavage at the Convention. We observe delegations from the populous states of Pennsylvania and Massachusetts in alliance with the southern states, Virginia, North Carolina, South Caronlina, and Georgia. These states, here represented on the \right" of the spectrum thanks to North Carolina's location having been normalized to equal 1. These delegations on the \right" of the representation issue advocated representation schemes based on population that would bypass the states. On the other side of the divide we see a group of delegations from the mostly smaller Middle Atlantic states of New Jersey, Connecticut, Maryland, New York and Delaware on the \left". These delegations preferred the status quo under the Articles of Confederation, which gave equal representation to each state, regardless of its population.
15
Table 1a: Parameter Estimates; Con ict over Representation Parameter ^ (^x) ^b ^ (^b) g^ Sd ^ (^g) Sd Sd Delegation x^ c^ New Hampshire 0.239 0.296 -0.009 0.305 . . . Massachusetts 0.406 0.218 -0.099 0.264 . . . Elbridge Gerry . . . . -0.660 0.257 -0.076 Rufus King . . . . 0.413 0.365 1.303 Connecticut -0.505 0.196 -0.558 0.294 . . . Oliver Ellsworth . . . . -1.753 0.418 0.557 Roger Sherman . . . . -1.295 0.317 0.656 New York -0.601 0.314 -0.408 0.328 . . . New Jersey -1.000 . -1.052 0.350 . . . Various Delegates . . . . -2.013 0.506 -0.527 Pennsylvania 0.767 0.278 0.299 0.302 . . . Gouverneur Morris . . . . -0.505 0.277 0.422 James Wilson . . . . -0.196 0.277 0.506 Other Delegates . . . . -0.371 0.616 -0.438 Delaware -0.618 0.207 -0.609 0.280 . . . John Dickinson . . . . -1.675 0.425 1.117 Other Delegates . . . . -1.042 0.375 0.407 Maryland -0.780 0.221 -0.587 0.311 . . . Daniel Carroll . . . . -1.633 0.469 -0.046 Luther Martin . . . . -3.487 0.747 -1.375 Other Delegates . . . . -2.739 0.983 2.280 Continued as Table 1b.
16
^ (^c) Sd
. . 0.298 0.420 . 0.454 0.367 . . 0.587 . 0.328 0.310 0.748 . 0.490 0.465 . 0.574 0.771 0.947
Table 1b: Parameter Estimates; Con ict Over Representation (continued) Parameter ^ (^x) ^b ^ (^b) ^ (^g) Sd Sd Sd Delegation x^ g^ c^ Virginia 0.999 0.277 0.406 0.346 . . . James Madison . . . . 0.000 . 0.000 George Mason . . . . -0.649 0.267 0.352 Edmund Randolph . . . . 0.0 58 0.290 1.039 North Carolina 1.000 . 0.523 0.327 . . . Hugh Williamson . . . . -0.143 0.274 0.290 Other Delegates . . . . 0.028 0.471 -0.324 South Carolina 0.473 0.295 0.022 0.291 . . . Pierce Butler . . . . -0.796 0.417 0.579 Charles C Pinckney . . . . -0.514 0.310 0.504 Charles Pinckney . . . . -0.359 0.349 0.238 John Rutledge . . . . -0.525 0.318 0.039 Georgia 0.060 0.217 -0.143 0.271 . . . ^ (^v~) = 0:053, ln(Lik) = ,1339:2156. ^v~ = 0:669, Sd The estimates reveal that these two camps were far from homogeneous. On the left Connecticut, New York, Delaware and Maryland all have similar estimated positions, with New Jersey being somewhat more extreme.10 On the right, Georgia, New Hampshire and Massachusetts are all statistically signi cantly more moderate than North Carolina, while Virginia and Pennsylvania are virtually as extreme as North Carolina, with the small estimated gap for Pennsylvania falling far short of the threshold for statistical signi cance. Many analysts claim that the states of the deep south, North Carolina, South Carolina, and Georgia, allied themselves with populous Vir10 Despite the numerical gap, in pairwise tests for statistical signi cance only Connecticut is statistically signi cantly to the \right" of New Jersey at the = 0:05 signi cance level. 17
^ (^c) Sd . . 0.335 0.327 . 0.321 0.582 . 0.491 0.374 0.442 0.323 .
ginia and Pennsylvania in the expectation that their populations would grow with westward migration. With the exception of New York, none of the states on the left of the spectrum had an active western frontier. While the estimated locations for the various delegations, which correspond to the vertical axis in Figure 2, can tell us about the positions of the majorities of the various state delegations, the proposal parameters can tell us about some of the individuals who made up these delegations. Recall that James Madison, the most proli c proposer at the Convention, has bee chosen as the \reference proposer". Positive estimated values for g, of which there are but three11 indicate delegates who sponsored proposals with a greater bias that Madison's toward representation proportional to population, corresponding to the \right" of the representation spectrum. Negative values indicate a bias towards the status quo of equal representation for the states regardless of population, on the \left" of the spectrum. Close to Madison's position at the far \right" were Virginia Governor Edmund Randolph, the author of record for the \Virginia Plan", which was one of the proposed constitutions deliberated at the Convention, James Wilson of Pennsylvania, another noted \nationalist" among the Convention delegates, Charles Pinckney of South Carolina, and Hugh Williamson of North Carolina. At the other end of the spectrum we nd Luther Martin and Daniel Carroll of Maryland, John Dickinson of Delaware, and Oliver Ellsworth and Roger Sherman of Connecticut. Among the state delegations on the \right" of the representation issue, only two delegates, George Mason of Virginia and Elbridge Gerry of Massachusetts, were statistically signi cantly to the \left" of James Madison.12 . These two \dissidents" from their states' delegations were also among the only three delegates who refused to sign the Constitution eventually proposed by the Convention, and it was Gerry's vote that 11 None of these the common parameter for the infrequently proposing North Carolina delegates, is statistically signi cantly to the right of Madison. 12 This despite the latter's considerable skills at the manipulation of electoral districts that representation proportional to population makes possible! 18
split the Massachusetts delegation on the key July 16 vote on the \Great Compromise", thereby allowing the compromise to prevail by a single vote. The b parameter measures the state delegations predispositions to support proposals from James Madison, the reference proposer. The lowest values for this parameter are encountered in the delegations from New Jersey, Maryland, Delaware, and Connecticut, on the \left", while the highest values correspond to the Virginia, North Carolina, and Pennsylvania delegations on the \right". These are the reactions we would expect given Madison's close identi cation with the \right" of the spectrum, and his estimated value for g to the \right" of virtually all of the remaining delegates. The c parameter estimates the \consensus appeal" of the stream of proposals coming from a particular individual, again relative to the value for James Madison which is normalized to equal 0. The two highest estimated values are for Rufus King, Governor of Massachusetts, and Virginia Governor Edmund Randolph. These were two states in which the Constitution absolutely had to be rati ed, and as it transpired, both rati cation battles were very close. If the interpretation of valence as corresponding to the political viability of the eventual proposed Constitution is correct, then it is hardly a surprise that Convention delegates were receptive to proposals from these experts on the current state of opinion in two \must win" contests.
19
Table 2a: Parameter Estimates; Scope and Empowerment of Government Parameter ^ (^x) ^b ^ (^b) g^ Sd ^ (^g) ^ (^c) Sd Sd Sd Delegation x^ c^ New Hampshire 0.652 0.342 -0.104 0.276 . . . . Massachusetts 0.903 0.369 -0.111 0.289 . . . . Elbridge Gerry . . . . 0.821 0.337 0.274 0.466 Rufus King . . . . 0.699 0.583 -0.896 0.800 Connecticut -1.344 0.457 -0.256 0.338 . . . . Oliver Ellsworth . . . . -0.606 0.263 0.499 0.381 Roger Sherman . . . . -0.958 0.434 -1.109 0.511 New Jersey -1.000 . -0.228 0.321 . . . . Various Delegates . . . . -0.703 0.458 -0.568 0.560 Pennsylvania -0.567 0.417 0.073 0.282 . . . . Gouverneur Morris . . . . -0.257 0.225 0.109 0.312 James Wilson . . . . -0.509 0.249 0.186 0.388 Other Delegates . . . . -0.540 0.366 -0.066 0.541 Delaware -0.343 0.386 -0.074 0.269 . . . . John Dickinson . . . . 0.085 0.303 -0.243 0.451 Other Delegates . . . . -0.962 0.437 -0.027 0.603 Maryland -0.683 0.373 0.166 0.283 . . . . Daniel Carroll . . . . -0.670 0.335 -0.082 0.558 Luther Martin . . . . 0.637 0.351 0.432 0.499 Other Delegates . . . . -0.384 0.485 -0.250 0.694 Continued as Table 2b.
20
Table 2b: Parameter Estimates; Scope and Empowerment of Government (continued) Parameter ^ (^x) ^b Sd ^ (^b) g^ Sd ^ (^g) ^ (^c) Sd Sd Delegation x^ c^ Virginia -0.082 0.383 0.236 0.264 . . . . James Madison . . . . 0.000 . 0.000 . George Mason . . . . 0.208 0.282 0.208 0.444 Edmund Randolph . . . . 0.177 0.242 0.449 0.354 North Carolina 1.000 . 0.016 0.291 . . . . Hugh Williamson . . . . 0.940 0.358 0.135 0.458 Other Delegates . . . . 2.656 0.994 1.681 1.275 South Carolina 1.394 0.443 0.021 0.318 . . . . Pierce Butler . . . . 0.982 0.791 -1.904 1.081 Charles C Pinckney . . . . 0.726 0.380 -0.033 0.577 Charles Pinckney . . . . 0.686 0.280 0.305 0.412 John Rutledge . . . . 0.967 0.463 0.433 0.605 Georgia 0.815 0.405 0.078 0.278 . . . . ^ (^v~) = 0:064, ln(Lik) = ,816:3002. ^v~ = 0:659, Sd Estimates for the deliberations on the issues Jilson (1981) refers to as the \scope and empowerment of government", dealing with commerce, westward expansion, and slavery, appear in Tables 2a and 2b. The rst question to address is whether this issue area is indeed distinct from the representation debate, or whether delegates behaved as though both touched on the same underlying dimension of con ict. If the two issue areas are genuinely homogeneous, then none of the model's parameters should change when we move from one issue to the other. We can test by comparing a constrained model, which forces all parameters to match across the two issue areas, with the unconstrained model, whose estimates appear in Tables 1a, 1b, 2a, and 2b, that allows all parameters to di er 21
across the two areas. The likelihood ratio test allows us to reject this hypothesis at all standard signi cance levels13 . Of course, this failure to match could be caused by di erences in the technical diculty of formulating proposals in the two issue areas even though the underlying distribution of issue preferences was in fact unidimensional across the two issues. To assess this possibility I test the more restrictive hypothesis that all of the x's match. The New York delegation was not present for the \scope and empowerment" deliberation, while New Jersey and North Carolina are used to anchor the two issue scales. This leaves us with 9 state delegations whose locations can be compared across the two issues. If the issue space is genuinely one dimensional, then all nine of the x's from the representation deliberations will equal their counterparts from the \scope and empowerment" deliberation, with di erences in their estimated values coming entirely from estimation error. This hypothesis can be rejected at all standard signi cance levels.14
13 The likelihood ratio test statistic is asymptotically distributed as 2 with 63 degrees of freedom. The observed test statistic of 166:905 therefore corresponds to a p-value of 8:365 10,11 . 14 The Wald statistic for the null hypothesis is distributed as 2 with 9 degrees of freedom. The realized value for the test statistic of 24:383 corresponds to a p-value of 0:0037.
22
Figure 2 Positions on Representation and The Scope and Empowerment of Government.
Representation VA
PA
CT
NJ p p p
p
p
p
p
p
p
MD
p
p
p
p
p
p
p
p
p
p
p
DE
23
p
p
p
p
p
p
p
p
p
p
p
p
p
p
NC p
p
MA
NH GA
p
p
p
p
p
p
SC Scope of Gov't
The estimated preferred outcomes for the state delegations participating in both sets of deliberations are shown in Figure 2. This gure suggests that except for Pennsylvania and Virginia the state delegations are distributed more or less along the 450 line. This suspicion receives further support from the data when we subject it to a formal test. We can accept the null hypothesis that all seven of the remaining states, excluding Pennsylvania and Virginia, have the same preferred outcome in both issue areas at all standard levels of signi cance.15 The \scope and empowerment of government" issue divided the Mid-Atlantic states, Connecticut, New Jersey, Maryland, and Delaware, now joined by Pennsylvania, on the right, from the states of the deep South; South Carolina, North Carolina, and Georgia in alliance with Massachusetts and New Hampshire. In contrast with their extreme position in favor of representation proportional to population, the delegations from W Virginia had middle-of-the-road preferred outcomes on the \scope and empowerment" issue, which made it the key element of winning majorities. The Mid-Atlantic delegates wanted to see the importation of slaves stopped, agricultural exports (mostly from the South) taxed, and easy to amend laws governing \navigation". The southern delegates favored continued importation of slaves, a prohibition on export tari s, and super-majority clauses for \navigation" acts. Part of Virginia's moderation stemmed from it's position as the domestic supplier of slave labor. While the states of the deep south wanted to continue importing enslaved Africans, Virginia preferred to keep the domestic market for slaves closed to foreign competition. Contemporary readers of the records of the Federal Convention nd the cynical protestations of slave owning Virginians against the immorality of the slave trade particularly outrageous. In any event, in the con ict over how to regulate \navigation", as maritime commerce was called at the time, and what to do 15 The Wald statistic for the null hypothesis is distributed as 2 with 7 degrees of freedom. The realized value for the test statistic of 7:622 corresponds to a p-value of 0:367. 24
about slavery, the Virginia delegation took a position between that of the Mid-Atlantic delegates, and those of the deep south. In contrast with the representation issue, on which James Madison was at the extreme right, proposals from Madison on the \scope and extent of government" issue were bracketed to the right by delegates from Massachusetts (Gerry and King), and the Carolinas (Williamson, Butler, Rutledge, and the Pinckney's), while well to Madison's left were delegates from Connecticut (Ellsworth and Sherman), Pennsylvania (G. Morris and Wilson), and Maryland (Caroll and Martin). While the consensus parameters for the \scope and empowerment" deliberations were mostly insigni cant.16 the estimated standard deviation for v~, representing the random shock to the consensus appeal parameter, is virtually the same for the two issue areas17 . To get some idea of how this parameter translates into the probability of compromise, consider a proposal in the ambit of representation from Edmund Randolph, author of the \Virginia Plan" that favored representation proportional to population, with an average value for the consensus shock v~ of 0. Based on the parameter estimates of the model the probability such a proposal would receive the support of the Massachusetts delegation is about 0:83. For a proposal from Connecticut delegate Oliver Ellsworth, closely identi ed with equal representation for each state, the consensus shock would have to assume a value of 1:81 in order to achieve the same probability of approval from the Massachusetts delegation. Given a normal distribution for v~ fewer than one in twenty eight proposals from Mr. Ellsworth would have received such a favorable shock. In concrete terms the shock corresponds to an idea with enough consensus appeal to overcome the disparate issue positions of the 16 Only Roger Sherman of Connecticut and his ideological opposite, Pierce Butler of South Carolina have estimated values for c that even approach statistical signi cance, with both earning negative parameter estimates 17 The Wald statistic for the null hypothesis is distributed as 2 with 1 degrees of freedom. The realized value for the test statistic of 0:0156 corresponds to a p-value of 0:9004. 25
delegates, and the low variance for the shock term indicates that the course of deliberation encountered such positive opportunities for compromise at a very low rate. The similar parameter for the \scope and empowerment" issue indicates that there too consensus producing high valence proposals were rare. The low value for the variance of the consensus shock v~ relative to the level of polarization among the positions of the state delegations on the representation and \scope and empowerment" issues at the Federal Convention suggests why the Convention took so long to produce a Constitution that was, in many respects, very similar the drafts considered by the Convention at its outset. The low variance for v~ meant that many proposals had to be deliberated to uncover those with suciently high valence to overcome the delegations polarized issue positions and so produce the consensus the convention needed. Had higher \valence" proposals that would have made passage of the Constitution more readily assured been easier to nd, the delegates to the Convention might have been spared a long Summer in Philadelphia, and legal scholars might have been deprived of an fertile source of interesting and quotable speeches!
4 Conclusions One well known political historian recently characterized applications of political methodology to the Constitutional Convention as seeking the key to understanding the \great compromise" in \two- or ve-factor solutions with the verimax rotation (ortho)."(Rakove., 1996). In contrast with this heavyhanded characterization, political methodology has been steadily developing models that incorporate more politics into the estimation. Rather than simply trying to \carve matrices at the joint" recent advances in roll call analysis have attempted to directly estimate increasingly realistic versions of the spatial model. Rather than moving us away from the rest of political science, 26
increasing methodological sophistication is bringing methodology closer to politics. This paper has used the roll call voting record to look behind the closed doors of the Federal Convention of 1787, nding a remarkably polarized environment, with important di erences among the state delegations on issue dimensions corresponding to representation, and to the scope and empowerment of the federal government. The most salient di erence between the two issue dimensions was the position of the Pennsylvania and Virginia delegations, which sided with Massachusetts, New Hampshire, and the deep south on the representation issue, while on the \scope and empowerment" issue Virginia took a more independent position, and Pennsylvania joined the remaining Middle Atlantic states, Delaware, Maryland, New Jersey, and Connecticut. The proposal behavior of the various delegates reveals some interesting rifts within the state delegations, notably identifying both George Mason and Elbridge Gerry as Mavericks within their state delegations on the representation issue, and also highlighting the importance of the political acumen of Governors King and Randolph, with their special knowledge of what it would take to get the Constitution through the dicult rati cation conventions in their respective states. Further research would bene t by a more ne-grained coding of the roll call votes taken at the Federal Convention. However this must be balanced against the limitations of the data; only 568 roll calls were taken at the convention, and this will not permit analysts the intellectually satisfying luxury of allowing each roll call vote its own issue dimension. On the purely methodological side, a promising avenue for future research would be to incorporate a random error term into the displacement parameters (the g). Because these parameters interact with the preferred outcomes, represented by the x parameters, this addition would create a challenging form of heteroscedasticity 27
into what is already, in some cases, an eleven-variate probit model. Neverthe-less, hardy methodologists will not long be deterred by this extra complication. Another area of potential fertile methodological work would be to compare the substantive dimensions identi ed here with the more abstract and hard to interpret \dimensions" revealed by agnostic multidimensional scaling.
28
A Position Taking and Agenda Manipulation The formal model developed in Section 2 can be used to illustrate some of the incentives for legislators to engage in position-taking, or strategic voting rather than to vote sincerely for their most preferred outcome. Suppose that we can represent the \true" policy preferences of a legislator using Equation 1. However, like most legislators, the subject of this model cares about more than just policy. The views of constituents, both those whom the legislator currently represents, and those whom in the future he would like to represent, are also important, and these constituents have their own ideas about policy. To capture this consider a very simple model in which the voters want the legislator to take an ideological position of xe on all public votes, with departures penalized on election day. The policy preferred by the voters, xe, may di er from the member's truly preferred policy outcome of x. To keep things simple, this model is intended only to illustrate the issues at work, let's suppose that the representative evaluates the electoral consequences according to a quadratic loss function that depends on the legislator's public vote, 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 ( ; xe ; pP ; pSQ; qP ; qSQ) = , (pP ,xe )2 ,(1, )(pSQ ,xe )2 + (qp+(1, )qSQ)
Letting () denote the probability the bill is enacted, the expected value of the member's overall wellbeing resulting from the bill is: ( )U (pP ; qP ; x)+(1 , ( ))U (pSQ; qSQ ; x)+ wV ( ; xe ; pP ; pSQ; qP ; qSQ) (5) 29
In equation (5) the w term corresponds to the weight the legislator attaches to the electoral payo of taking the position the voters' want him to advocate relative to the direct payo of enacting his own most preferred policies.18 The dependence of () on recognizes that the legislator's vote may a ect the probability the policy is eventually adopted. Some straightforward manipulation reveals that the legislator 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 (6) 2 where xe + xw( (1) , (0)) (7) 1 + w((1) , (0)) Provided 1 + w((1) , (0)) > 0 the legislator behaves \as if" his most preferred outcome was x . Research on retiring members of the US Congress Lott (1992) indicates members' actual preferred points, the x's, are very similar to the positions they advocated throughout their tenure, the x 's. x =
Whether this nding generalizes to other legislatures is, of course, an open question, but it is suggestive. The member's induced ideal point, x in Equation (7) depends on several factors. First there is the e ect the member's vote will have on the probability the measure is passed: ((1) , (0)). Then there is the weight the member places on policy as opposed to reelection, represented by w, and nally the distance between the member's true preferred outcome; x, and the outcome his constituents expect him to pursue; xe. It is conceivable that the impact of the member's vote on the probability the proposal is nally enacted, ((1) , (0)), is negative, that is, it is 18 Since the w term essentially calibrates the weight the legislator places on impressing constituents, or potential constituents, we can restrict attention to non-negative values of w. 30
possible that in some cases by voting against an amendment the legislator actually increases the probability the amendment will be adopted. This might happen in a very low information environment in which a legislator's opponents attempted to free ride on his information, or it might arise with a very structured and predictable agenda. Consider the decision of many anti-spending Republicans to vote strategically in favor of the famous Powell Amendment to the 1956 School Aid Bill, which if implemented would have increased spending, but which many of them knew would lead to subsequent defeat of the bill, see (Riker and Shepsle, 1985). The less predictable the agenda, the smaller the incentives to manipulate it by casting strategic votes. This will occur when there are lots of later opportunities for change, as the reconsideration rules of the Federal Convention of 1787 ensured there would be. This shows up in the formal model as a small value for ((1) , (0)); the legislator's expectations about the ultimate form of the bill after it has passed through the subsequent and unpredictable agenda are barely a ected by how he votes on the current amendment. The incentives for position-taking correspond to the importance the legislator places on the public's reaction, as measured by w. When the votes are not made public they cannot a ect the public's view of the legislator. At the Federal Convention of 1787 the rule of secrecy largely removed votes from the public view. To be sure, the remaining delegates included many of the most prominent politicians of the day, but the rules against revelation impeded the ability of these other politicians' to reveal votes to the public. At the Federal Convention of 1787 the rule permitting the convention to revisit questions that had already been voted attenuated the incentives to vote strategically, while the secrecy rule greatly limited delegates' ability to engage in position-taking19. 19 Perhaps the most resourceful of the delegates on this score was James Madison, whose copious notes on the Convention left him with plenty of room to put his own "spin" on his role in the proceedings, dominating the post humus view of the Convention. 31
References Farrand, Max. 1966a. The Records of the Federal Convention of 1787, Volume I. New Haven: Yale University Press. Farrand, Max. 1966b. The Records of the Federal Convention of 1787, Volume II. New Haven: Yale University Press. Farrand, Max. 1966c. The Records of the Federal Convention of 1787, Volume III. New Haven: Yale University Press. Heckman, James J. and James Snyder. 1996. \Linear Factor Models of the Demand for Attributes with an Application to Estimating the Preferences of Legislators." typescript. Jilson, Calvin C. 1981. \Constitution-Making: Alignment and Realignment in the Federal Convention of 1787." American Political Science Review 75:598{612. Londregan, John B. 1996. Legislative Institutions and Ideology in Chile's Democratic Transition . New York: Cambridge University Press, in press. Londregan, John B. 1999. \Estimating Legislators' Preferred Points." forthcoming in Political Analysis. Lott, John R. 1992. \Political Cheating." Public Choice 52:169{186. Poole, Keith T. and Howard Rosenthal. 1996. Congress: A PoliticalEconomic History of Roll Call Voting. New York: Oxford University Press. Rakove., Jack. 1996. Original Meanings: Politics and Ideas in the Making of the Constitution. New York: Vintage Books. Riker, Arthur T. Denzau William and Kenneth Shepsle. 1985. \Farquharson and Fenno: Sophisticated Voting and Home Style." American Political Science Review 79:1117{33. 32
Riker, William H. 1996. The Strategy of Rhetoric: Campaigning for the American Constitution. New Haven: Yale University Press. Stokes, Donald. 1963. \Spatial Models of Party Competition." American Political Science Review 57:368{77.
33
