程序代写代做代考 game graph ada Hive html C DO VOTERS AFFECT OR ELECT POLICIES? EVIDENCE FROM THE U. S. HOUSE*

DO VOTERS AFFECT OR ELECT POLICIES? EVIDENCE FROM THE U. S. HOUSE*
DAVID S. LEE ENRICO MORETTI MATTHEW J. BUTLER
There are two fundamentally different views of the role of elections in policy formation. In one view, voters can affect candidates’ policy choices: competition for votes induces politicians to move toward the center. In this view, elections have the effect of bringing about some degree of policy compromise. In the alternative view, voters merely elect policies: politicians cannot make credible promises to moderate their policies, and elections are merely a means to decide which one of two opposing policy views will be implemented. We assess which of these con- trasting perspectives is more empirically relevant for the U. S. House. Focusing on elections decided by a narrow margin allows us to generate quasi-experimental estimates of the impact of a “randomized” change in electoral strength on subse- quent representatives’ roll-call voting records. We find that voters merely elect policies: the degree of electoral strength has no effect on a legislator’s voting behavior. For example, a large exogenous increase in electoral strength for the Democratic party in a district does not result in shifting both parties’ nominees to the left. Politicians’ inability to credibly commit to a compromise appears to dominate any competition-induced convergence in policy.
I. INTRODUCTION
How do voters influence government policies? An economist’s answer is that they do so by compelling politicians to adopt “middle ground” platforms. Competition for votes can force even the most partisan Republicans and Democrats to moderate their policy choices. In the extreme case, competition may be so strong that it leads to “full policy convergence”: opposing parties are forced to adopt identical policies [Downs 1957].1 More realisti-
* An earlier version of the paper, “Credibility and Policy Convergence: Evi- dence from U. S. Roll Call Voting Records,” is online as NBER Working Paper No. 9315, October 2002. We thank David Card, John DiNardo, and Melvin Hinich for helpful discussions, and David Autor, Hongbin Cai, Anne Case, Dhammika Dhar- mapala, Justin Wolfers, and participants in workshops at the University of North Carolina—Chapel Hill, University of Texas—Austin, University of Chicago De- partment of Economics and Graduate School of Business, Princeton University, University of California—Los Angeles, Stanford University Graduate School of Business, University of California—Davis, University of California—Riverside, and University of California—Irvine, and two anonymous referees for comments and suggestions. We also thank James Snyder and Michael Ting for providing data for an earlier draft. This material is based upon work supported by the National Science Foundation under Grant No. 0214351.
1. Empirical studies indicate that Republican and Democratic legislators vote very differently, even when they share the same constituency. For example,
© 2004 by the President and Fellows of Harvard College and the Massachusetts Institute of Technology.
The Quarterly Journal of Economics, August 2004
807

808 QUARTERLY JOURNAL OF ECONOMICS
cally, competition leads to “partial policy convergence”: candi- dates do pursue more moderate policies, even if they are not forced to adopt identical platforms [Wittman 1983; Calvert 1985]. This less rigid and arguably more realistic understanding of Downs’ insight has become central to how economists think about political competition. Indeed, the so-called “Downsian paradigm” has remained the backbone of many models in political economy.
There is, however, a growing recognition of a serious short- coming of this paradigm. In a recent survey of the literature, Besley and Case [2003] emphasize that the assumptions about politicians’ commitment and motivation in the Downsian para- digm “are unreasonable and outcomes are highly unrobust to deviations from them.” Downsian convergence depends on the assumption that elected politicians always implement the policies that they promised as candidates. But Alesina [1988] shows that when partisan politicians cannot credibly promise to implement more moderate policies, the result can be full policy divergence: the winning candidate, after obtaining office, simply pursues his most-preferred policy. In this case, voters fail to compel candi- dates to reach any kind of policy compromise.
What emerges, then, are two fundamentally different views of the role of elections in a representative democracy. On the one hand, when electoral promises are credible—as in a Downsian partial convergence—candidates seek middle ground policies, and general elections bring about some degree of policy “compro- mise.” On the other hand, when promises to enact moderate policies are not credible—as in full policy divergence—general elections are merely a means to decide which candidate’s pre- ferred policy will be implemented. Which of these two competing views is empirically more relevant? This paper assesses the rela- tive importance of the two contrasting perspectives in explaining how Representatives vote in the U. S. House.
As is apparent from Alesina’s [1988] analysis of the role of credibility, the two broad views have sharply different predictions for how a politician’s electoral strength influences her policy choices. When politicians have incentives to moderate their plat- forms—as in partial policy convergence—the relative electoral
Poole and Rosenthal [1984] show that senators from the same state but from different political parties have different voting records. This is inconsistent with Downs’ original model, in which candidates adopt identical positions—“complete policy convergence.” See also Snyder and Groseclose [2000] and Levitt [1996].

DO VOTERS AFFECT OR ELECT POLICIES? 809
strength of the two parties matters. More specifically, when elec- toral support is high, a candidate can afford to vote in a relatively more partisan way if he is elected; a weaker candidate would be forced to choose a more moderate policy. An increase in electoral strength for the Democratic party in a district, for example, would cause both parties’ nominees to shift to the “left.” On the other hand, when voters do not believe promises of policy compro- mises—as in full policy divergence—the relative electoral strength of the two candidates is irrelevant, as politicians simply pursue their own personal policy views. That is, an increase in electoral strength for the Democratic party in a district leaves legislators’ actions unchanged.
Therefore, an assessment of the relative importance of the two views requires estimating the effect of a candidate’s electoral strength on subsequent roll-call voting records. To do so, we consider electoral races where a Democrat holds the seat—and hence an electoral advantage—and measure the roll-call voting records of the winners of these elections. We measure the extent to which they are more liberal than the voting records of winners of elections where the Republican had held the seat; i.e., where the Democrat was relatively weaker.2 Of course, which of the two parties holds a district seat—and hence the electoral advan- tage—is clearly endogenously determined, influenced by the po- litical leanings of the voters, the quality of candidates, resources available to the campaigns, and other unmeasured characteris- tics of the district and the candidates. A naive comparison that does not account for these differences between Democratic and Republican districts is likely to yield biased estimates. What is needed is an exogenous variation in who holds the seat—and hence greater electoral strength—in order to measure how poli- ticians’ actions respond to the odds of winning an election.
To isolate such exogenous variation, we exploit a quasi ex- periment embedded in the Congressional electoral system that generates essentially “random assignment” of which party holds a seat—and therefore which party holds the electoral advantage. In particular, we focus our analysis on the set of electoral races in
2. Our empirical strategy obviously accounts for the fact that Democrats are more liberal than Republican. That is, the roll-call behavior of a winner of an electoral race where a Democrat held the seat will tend to be more liberal simply because—due to the advantage of incumbency—the winner will more likely be a Democrat. This fact in itself would cause a difference in voting records, even if Representatives ignored electoral pressures and simply voted their own ideologi- cal position. It is easy to account for this factor—as we will show in Section II.

810 QUARTERLY JOURNAL OF ECONOMICS
which the incumbent party had barely won the previous election (say by 0.01 percent of the vote). The key identifying assumption is that districts where the Democrats barely won are compara- ble—in all other ways—to districts where the Republicans barely won. We present empirical evidence that strongly supports this assumption: Democratic and Republican districts are in general very different, but among close elections, they are similar in every characteristic that we examine, including various demographic characteristics of the population, racial composition, size of the district, income levels, and geographical location. Our quasi ex- periment, then, addresses the endogeneity problem by isolating arguably independent and exogenous variation in candidates’ electoral strength across Congressional districts.
Using this regression-discontinuity design and voting record data from the U. S. House (1946 –1995), we find that the degree of electoral strength has no effect on a legislator’s voting behavior.3 Candidates with weak electoral support do not adopt more mod- erate positions than do stronger candidates, holding other factors constant. For example, a large exogenous increase in electoral strength for the Democratic party in a district does not result in shifting both parties’ nominees to the left. This suggests that voters seem not to affect politicians’ choices during general elec- tions; instead, they appear to merely elect policies through choos- ing a legislator. That is, they do not influence policy through their Representatives’ choices as much as they are implicitly presented with policy choices by different candidates.4
Our findings are consistent with the inability of opposing candidates to credibly commit to a policy compromise. It appears that the central prediction of the Downsian paradigm—that in- dividual politicians’ policy choices are constrained by voters’ sen- timents—has little empirical support, at least in the context of U. S. House general elections. Our findings provide some empiri- cal justification for the notion that candidates confront a credi-
3. The voting score data include (1) the actual roll-call data from ICPSR Study 4 “United States Congressional Roll Call Voting Records,” (2) Groseclose, Levitt, and Snyder’s [1999] Inflation-Adjusted ADA Scores [Groseclose 2002], (3) McCarty, Poole, and Rosenthal’s [1997] DW-NOMINATE scores [Poole 2002a], (4) Poole and Daniels’ [1985] Interest Group Ratings [Poole 2002b], and (5) Poole’s [1999] rank order data [Poole 2002c].
4. This leaves open the question of how candidates are selected. There are several models where candidates are endogenous. (See, for example, Persson and Tabellini [2000] for an introduction to this literature.) In this paper we take the candidates’ ideologies as exogenous. We return to this point below.

DO VOTERS AFFECT OR ELECT POLICIES? 811
bility problem. This notion has been explicitly adopted in recent theoretical analyses [Besley and Coate 1997, 1998].
It is important to recognize that our findings say little about whether members of the U. S. House generally represent their “constituencies.” Instead, our analysis focuses on the role of gen- eral elections in inducing candidates with different policy stances to move toward the center. Although we find a small effect of the pressures of a general election on candidates, this does not imply that election outcomes do not “represent” the desires of the elec- torate. First, and most obviously, voters still do choose between the two available policy platforms. Second, “representativeness” does not necessarily occur only through general elections. Pre- election channels (primary elections, for example) may also be important in inducing representativeness. Indeed, within each district, the Republican and Democratic nominees may, respec- tively, represent the “median” Republican and “median” Demo- cratic voter.
The paper is organized as follows. Sections II and III provide background and motivation for our analysis, and describe our empirical strategy, first informally, and then within a formal conceptual framework. Section IV describes the context and the data, and Section V presents our empirical results. We relate our findings to the existing literature in Section VI. Section VII concludes.
II. BACKGROUND AND CONCEPTUAL FRAMEWORK
II.A. Role of Credibility in Political Competition
Voters can influence policy in two distinct ways. Competing political candidates have incentives to adopt positions that reflect the preferences of the electorate because doing so raises the chances they will win the election. That is, voters can affect the policy choices of politicians. Alternatively, voters always impact policy outcomes by selecting a leader among several candidates, who each may have already decided on a particular policy based on other reasons. In this way, voters may simply elect policies. Whether voters affect or elect policies depends on whether or not candidates are able to make credible promises to implement moderate policies.
A large class of models of political competition assumes that they can. The most well-known example is the simple “median

812 QUARTERLY JOURNAL OF ECONOMICS
voter” model of political competition [Downs 1957]. Two can- didates, who care only about winning office, compete for votes by taking a stance in a single dimensional policy space. Voters cast their vote based on these positions, and the equilibrium result is that the politicians carry out identical policies—the one most preferred by the “median voter.” In this extreme example, voters have a powerful effect on politicians’ choices, to the point where it is irrelevant which of the two candidates is ultimately elected.
A similar outcome results when opposing candidates care not only about winning the election, but also about the implemented policies themselves. Opposing parties may not choose identical positions, but in general electoral competition will compel them to choose policies more moderate than their most preferred choices [Wittman 1983; Calvert 1985]. The basic insight that voters affect candidates’ positions by inducing spatial competition is robust to various generalizations of the simple model utilized by Downs [Osborne 1995].
But it is much less robust to the assumption that candidates can commit to policy pronouncements, as emphasized in Besley and Case [2003]. When politicians have ideological preferences over policy outcomes, credibility becomes an issue. Specifically, Alesina [1988] points out that Downs’ equilibrium may fall apart if parties care about policies and there is no way to make binding precommitments to announced policies. After winning the elec- tion, what incentive does a legislator have to keep a promise of a more moderate policy? In a one-shot game, the only time-consistent equilibrium is that candidates carry out their ex post most- preferred policy. Electoral pressures do not at all compel opposing candidates to moderate their positions. Voters’ only role in affect- ing policy outcomes is to elect a politician, whose policy position is unaffected by electoral pressures.
In a repeated election framework, both policy convergence and divergence are possible, as politicians can establish cred- ibility through building reputations. If voters and opposing parties believe that there are sufficiently high costs to deviat- ing from moderate promises, it is possible to achieve some degree of policy convergence [Alesina 1988]. Voters affect poli- cies because of candidates’ incentives to maintain a reputation. But if both parties and voters do not expect any compromise, the fully divergent outcome occurs in every election. Candi-

DO VOTERS AFFECT OR ELECT POLICIES? 813
dates do not deviate from their ex post most-preferred policy, and voters only elect policies.5
The goal of this paper is to examine which phenomenon is more empirically relevant for describing roll-call voting patterns of U. S. House Representatives. Does the expectation of how voters will cast their ballot affect how legislators vote, or do voters simply elect a legislator among candidates with fixed policy posi- tions? The answer to this question has important implications for understanding and modeling policy formation in a representative democracy.
If voters primarily affect politicians’ decisions, then “centripe- tal” political forces generated by the broader voting population would largely outweigh any “centrifugal” forces that pull candi- dates’ positions apart (e.g., party discipline, special interest groups). It would also imply that candidates are able to convince voters that they will compromise on policy, through the building of reputations or other mechanisms. The Downsian paradigm would then seem to be a reasonable, first-order description of policy formation as it relates to U. S. House elections.
On the other hand, if voters primarily elect policies, then “centrifugal” forces largely would dominate any Downsian con- vergence. It would then become more important, for example, to understand how a nominee, and the policies that she supports, is chosen by the party: primary elections could be more influential than general elections for policy formation. It would also provide an empirical basis for assuming that candidates face a serious credibility problem in their policy pronouncements. There is a growing recognition of the inadequacy of the Downsian paradigm on this point [Besley and Case 2003].
Existing studies have established that, controlling for con- stituency characteristics, Democratic representatives possess more liberal voting records than Republican members of Con- gress.6 This constitutes strong evidence against the extreme case of complete policy convergence (e.g., the median voter theorem), but is too stringent a test of the more general notion of Downsian electoral competition. Therefore, to measure the relative impor-
5. It is also true that even if discount rates are sufficiently low, the fully divergent outcome still remains a subgame perfect equilibrium of the repeated election game.
6. The full convergence hypothesis has been tested, and rejected by many authors. For example, Poole and Rosenthal [1984] show that senators from the same state but from different political parties have different voting records. For a discussion of empirical regularities in the literature, see Snyder and Ting [2001a].

814 QUARTERLY JOURNAL OF ECONOMICS
tance of competition-induced convergence, it is necessary to em- pirically distinguish between partial convergence, where voters affect politicians’ policy choices—despite the undeniable party effect—and complete divergence, where voters merely elect poli- cies. This is the goal of our study.
II.B. Identification Strategy
We now describe the main difficulties of addressing this question, and how we confront them with our identification strat- egy. Here we will intentionally be less formal, in order to provide the intuition of our approach. A more rigorous exposition of our conceptual and econometric framework is presented in the next section. Throughout the discussion we assume a two-party politi- cal system.
The most straightforward way to determine whether voters primarily affect or elect policy choices is to simply compare can- didates’ most-preferred policies (hereinafter “bliss points”) and the policies they would actually choose. If the voting records were more moderate than their bliss points, this would indicate that the expected voting behavior of the electorate factored into the candidates’ decisions. If there were no difference between their choices and their bliss points, this would imply that voters merely influence the relative odds of which of the two candidates’ policies is “elected.” Unfortunately, such a comparison is impossible, since there are no reliable measures of candidates’ bliss points.
In this paper we utilize a simple empirical test of whether voters primarily affect or elect policy choices, based on how Rep- resentatives’ roll-call voting behavior is affected by exogenous changes in their electoral strength. The test is based on the predictions of Alesina’s [1988] model of electoral competition. In the next section we formally develop the idea, but the intuition is very simple. If candidates are constrained by their constituents’ preferences, we should observe that exogenous changes in their electoral strength have an impact on how they intend to vote if elected to Congress. On the other hand, if promises to adopt moderate policies are noncredible, then the electoral strength of a candidate should be irrelevant to how (s)he intends to vote.
Throughout the paper we use the following notation for the timing of elections. t and t 􏰁 1 represent separate electoral cycles. For example, when t 􏰏 1992, it includes the 1992 cam- paign, the November 1992 election, and the 1993–1994 Congres- sional session. Similarly, t 􏰁 1 would include the 1994 campaign,

DO VOTERS AFFECT OR ELECT POLICIES? 815
the November 1994 election, and the 1995–1996 Congressional session.
Our strategy is based on the following thought experiment. Imagine that we could decide the outcome of Congressional elec- toral races in, say, 1992 with the flip of a coin (but we allow all subsequent elections to be determined in the usual way). This initial randomization guarantees that the group of districts where the Democrat won would be, in all other respects, similar to the newly Republican districts. For example, the two groups of districts would be similar in the ideological positions of the voters and candidates, the demographic characteristics, the resources that were available to the candidates, and so forth.
Because incumbents are known to possess an electoral ad- vantage, the outcome of the 1992 race would impact what hap- pens in the 1994 election. Democrats are likely to be in a rela- tively stronger electoral position where they are incumbents, and similarly for Republicans. The key point is that the random assignment of who wins in 1992 essentially generates random assignment in which party’s nominee has greater electoral strength for the 1994 election. We could use this change in elec- toral strength to test the hypothesis of complete divergence against the alternative of partial convergence.
Specifically, we could examine the 1995–1996 voting “scores” of the winners of the 1994 elections where the Democrats had held the seat during the 1994 campaign, and compare them with the scores of winners of elections where a Republican held the seat. This difference would represent a valid causal effect of who holds the seat during the 1994 electoral races on 1995–1996 voting records. We call this the “overall effect,” and it is the sum of two components.
The first component would reflect that the 1995–1996 voting scores of the winners where a Democrat held the seat during the 1994 electoral race will tend to be more liberal simply because— due to the electoral advantage of holding the seat—the winner will more likely be a Democrat. And as we know, Democrats have more liberal voting scores. This first component reflects how voters elect policies: how they impact policy by simply altering the relative odds of which party’s nominee is chosen. As we show more formally in the next section, this component can be directly estimated by answering the questions how much more likely is the winner to be a Democrat if the seat is already held by a

816 QUARTERLY JOURNAL OF ECONOMICS
Democrat and what is the expected difference between how Re- publicans and Democrats vote, other things constant.
The remaining, second component would reflect how candi- dates might respond to an exogenous increase or decrease in the probability of winning the election in 1994. If legislators are pressured to keep their election promises, then a Democrat who is challenging an incumbent Republican in 1994 would be expected to have less liberal voting records in 1995–1996 (if elected) com- pared with an incumbent Democrat. After all, the challenger would be in a much weaker electoral position than the incumbent. This second component reflects how expected voting behavior affects the policy choices of candidates. It is computed by sub- tracting the first component from the overall effect.
The relative magnitudes of the two components indicate which equilibrium—full divergence or partial convergence—is relatively more important. If the “elect” component is dominant, it suggests full policy divergence: politicians simply vote their own policy views, unaffected by electoral pressures. If the “affect” component is important, it suggests partial policy convergence: policy choices are constrained by electoral pressure imparted by voters.
What allows us to perform this decomposition into the two components? The initial “random assignment” of who wins the 1992 election does. Without the random assignment, it would be difficult to distinguish between any of these effects and differ- ences due to spurious reasons. After all, in the real world, the party that holds a district seat—and the electoral advantage—is clearly endogenously determined, influenced by the ideologies of the voters and candidates, and other unmeasured characteristics of the districts. A naive comparison that does not account for all these unobservable differences between Democratic and Repub- lican districts is likely to yield biased estimates.
For example, Democratic legislators will have more liberal voting scores than Republicans (for simplicity, consider the period of the 1990s). But Democrats are also more likely to be elected in places like Massachusetts and than in places like Alabama. So it is not clear how much of this voting gap reflects the typical difference between Republican and Democratic nominees and how much of the gap reflects the typical difference between Rep- resentatives from Massachusetts and Alabama.
How do we generate the initial “flip of the coin” decision of who wins the 1992 election? We use a quasi experiment that is

DO VOTERS AFFECT OR ELECT POLICIES? 817
embedded in the Congressional electoral system. Specifically, our empirical strategy focuses on elections that were decided by a very narrow margin in 1992, as revealed by the final vote tally. For example, we begin by examining elections that were decided by less than a 2 percent vote share. We argue that among these elections, it is virtually random which of the two parties won the seat [Lee 2003]. For the sake of exposition, we defer to a later section the discussion of why we believe this to be true, and the description of the empirical evidence that strongly supports this assumption. We have used 1992 and 1994 in this explanation of our empirical strategy. In practice, in our empirical analysis we use data for the period 1946–1995.
III. THEORETICAL AND ECONOMETRIC FRAMEWORK
In this section we 1) formally define what it means to ask the question of whether voters primarily affect or elect policies, and 2) explain how our empirical strategy is able to distinguish be- tween these two phenomena.
III.A. Model
We utilize the repeated election framework of Alesina [1988], adopting that study’s modeling conventions and notation. Con- sider two parties, D (Democrats) and R (Republicans), in a par- ticular Congressional district. The policy space is unidimensional, where party D’s and R’s per-period policy preferences are repre- sented by quadratic loss functions, u(l ) 􏰏 􏰂(1/ 2)(l 􏰂 c)2 and v(l) 􏰏 􏰂(1/2)l2, respectively, where l is the policy variable and c(􏰐0), and 0, are their respective bliss points. As in Alesina, the analysis makes no distinction between the “party” and an indi- vidual nominee, so that the “electoral strength” of the party in a district is equivalent to the “electoral strength” of the party’s nominee in that district, during the election. Also, candidates’/ parties’ bliss points are assumed to be exogenously determined.7
The timing of elections is as follows. Before election t, voters
7. This framework has little to say on the question of how candidates are selected. Alternative frameworks are possible and may generate different predic- tions. For example, the models proposed by Bernhardt and Ingberman [1985] and Banks and Kiewiet [1989] are quite different in spirit from the model used here. In those models, the challenger is at disadvantage because she cannot adopt the incumbent’s position and is therefore forced to take a more extreme position. In equilibrium the low probability of defeating incumbent members of Congress deters potentially strong rivals from challenging them [Banks and Kiewiet 1989].

818 QUARTERLY JOURNAL OF ECONOMICS
form expectations of the parties’ policies, denoted xe and ye. At this point, the outcome of the election is uncertain to all agents in the model, with the probability of party D winning being P, which is “common knowledge.” P(xe,ye) is a function of xe and ye, and by assumption, when xe 􏰐 ye, then 􏰑P/􏰑xe, 􏰑P/􏰑ye 􏰒 0; that is, more votes can be gained by moderating the policy position. If party D wins the election, x is implemented, and if party R wins, y is implemented. A rational expectations equilibrium is assumed throughout; x 􏰏 xe, and y 􏰏 ye. The game then repeats for period t 􏰁 1. Note that period t includes both the election and the subsequent Congressional session, and similarly for t 􏰁 1. For example, if t 􏰏 1992, t refers to the November 1992 election and the roll-call votes RCt in the 1993–1994 Congressional session; t 􏰁 1 refers to the November 1994 election and the roll-call votes RCt􏰁1 in the 1995–1996 session.
Alesina [1988] shows that the efficient frontier is given by x* 􏰏 y* 􏰏 􏰓c, where 􏰓 􏰠 (0,1). Because of the concavity prefer- ences, both parties prefer a moderate policy with certainty to a fair bet. Three Nash equilibria are possible.
(a) Complete Convergence: x* 􏰏 y* 􏰏 􏰓*c.
In this equilibrium, opposing parties agree to a moderate policy, by Nash bargaining on the efficient frontier. The “Folk Theorem” equilibrium is one where both parties “announce” the same, moderate policy, and the voters expect the moderate out- come, but as soon as a party deviates from the announced posi- tion, reputation is lost, and the game reverts to the uncooperative outcome, y* 􏰏 0, x* 􏰏 c. As long as discount rates are sufficiently low, promises to adopt policy compromises are credible.
For our purposes, the key result is that dx*/dP* 􏰏 dy*/dP* 􏰏 (d􏰓*/dP*)c 􏰐 0, where P* represents the underlying “popular- ity” of party D: the probability that party D would win at fixed policy positions, xe 􏰏 c and ye 􏰏 0.8 An increase in P* represents an exogenous increase in the popularity of party D, which would boost party D’s “bargaining power” so that the equilibrium moves closer to her bliss point. This exogenous increase comes about from a “helicopter drop” of Democrats in the district, or campaign resources, or the advantage that comes from being the incumbent in the district. In this equilibrium, policy choices are implicitly
8. 􏰓 is used to characterize the entire efficient frontier. 􏰓*, on the other hand, denotes the Nash bargaining equilibrium.

DO VOTERS AFFECT OR ELECT POLICIES? 819
constrained by voters. Thus, when dx*/dP*, dy*/dP* 􏰐 0, we say that voters affect candidates’ policy choices.
Indeed, in this equilibrium—similar to Downs’ original “me- dian voter” model—voters exclusively affect policy choices, and do not elect policies at all: it is irrelevant for policy which party is actually elected.
(b) Partial Convergence: 0 􏰡 y* 􏰡 x* 􏰡 c.
Is the result that voters affect policies—dx*/dP*, d y*/dP* 􏰐 0—robust to minor deviations from the complete convergence equilibrium? We show that it is. This agrees with our intuition that voters can induce policy compromise, even if they cannot force them to adopt identical positions. It also agrees with our intuition that a rejection of complete convergence says little about the relative degree to which voters affect or elect policies. Reject- ing complete convergence simply implies that y* 􏰒 x*, but noth- ing about whether 0 􏰒 y* or x* 􏰒 c.
It is possible to extend Alesina’s model to allow for parties to care about winning the seat, per se, in addition to caring about the policy outcome.9 The result is that in general, 0 􏰡 y* 􏰡 x* 􏰡 c, because there are values where x 􏰏 y is not Pareto efficient. Both parties can be made better off by one party moving closer to its bliss point, because there is an explicit benefit to obtaining office. A detailed proof is available on request.
The important point, for our purposes, is that the compara- tive static dx*/dP*, dy*/dP* 􏰐 0 is robust to this logical exten- sion to the model. With an exogenously higher P*, party D has a better “bargaining position” and therefore can compel the parties to agree on a position closer to party D’s bliss point.
(c) Complete Divergence: x* 􏰏 c, y* 􏰏 0.
In this equilibrium, voters expect nothing else than the par-
ties to carry out their bliss points if elected, and the parties do just this. This can arise if promises to implement policy compro- mises are not credible. In this case, an increase in P* now does nothing to the equilibrium: dx*/dP* 􏰏 dy*/dP* 􏰏 0. This is a “corner solution,” whereby an exogenous shock to P* has no effect on candidates positions. Here, voters merely elect politicians’ fixed policies.
9. Our extension should not be confused with that of Alesina and Spear [1988], in which parties agree to split the benefits of office. In our extension, they cannot split the benefits of office. This case should also not be confused with the partial convergent equilibria that can arise if discount rates are too low to support fully convergent equilibria. Alesina [1988] proves existence of these equilibria.

820 QUARTERLY JOURNAL OF ECONOMICS
Among the above three equilibria, the full convergence equi- librium is not very realistic, and has already been empirically rejected by several authors. But a rejection of full convergence says little about whether politicians’ behaviors are better char- acterized by partial convergence (voters can affect policy out- comes) or complete policy divergence (voters only elect policies). Distinguishing between these two equilibria is our goal. For this purpose, the key result of the theoretical framework is that dif- ferentiating between partial and complete divergence is equiva- lent to assessing whether dx*/dP*, dy*/dP* 􏰐 0 or dx*/dP*, dy*/dP* 􏰏 0.
We assume that voters are forward-looking and have rational expectations. This implies that voting records RCt􏰁1—roll-call votes after the election—are on average equal to voters’ expecta- tions. It is important to note that this is not the same as assuming that candidates can make binding precommitments. Politicians always have the option of not carrying out their preelection policy pronouncements. But in Alesina’s repeated game equilibrium, candidates do carry out their “announced” policies because of the need to maintain a reputation.10
III.B. Estimating Framework
The above framework directly leads to our empirical strat- egy. Note first that the roll-call voting record RCt of the repre- sentative in the district following the election t can be written as
(1) RCt 􏰏 􏰔1 􏰂 Dt􏰕yt 􏰁 Dtxt,
where Dt is the indicator variable for whether the Democrat won election t. A similar equation applies for RCt􏰁1. Simply put, only the winning candidate’s intended policy is ultimately observable. In Appendix 1 we provide conditions under which the above expression can be transformed into
(2) RCt 􏰏 constant 􏰁 􏰖0P*t 􏰁 􏰖1Dt 􏰁 εt
(3) RCt􏰁1 􏰏 constant 􏰁 􏰖0P*t􏰁1 􏰁 􏰖1Dt􏰁1 􏰁 εt􏰁1,
where P* is the measure of the electoral strength of party D—the probability of a party D victory at fixed platforms c and 0 —and ε reflects heterogeneity in bliss points across districts. This equa-
10. Of course, the equilibrium depends on candidates not discounting the future too much.

DO VOTERS AFFECT OR ELECT POLICIES? 821
tion simply parameterizes the derivatives dx*/dP*, dy*/dP* as 􏰖0. It also allows an independent effect of party, 􏰖1, which is reasonable given the existing evidence that party affiliation is an important determinant of roll-call voting records. In this equa- tion, partial convergence (voters affect policy choices) implies that 􏰖0 􏰐 0. Full divergence (voters only elect policies) implies that 􏰖0 􏰏 0.
In general, we cannot observe P*, so equation (2) cannot be directly estimated by OLS. But suppose that one could randomize Dt. Then Dt would be independent of εt and P*t. Also, if bliss points are exogenous—and hence are not influenced by who won the previous election—then Dt will have no impact on εt􏰁1. It follows that
DR E􏰗RCt􏰁1􏰢Dt 􏰏 1􏰘 􏰂 E􏰗RCt􏰁1􏰢Dt 􏰏 0􏰘 􏰏 􏰖0􏰗P*t􏰁1 􏰂 P*t􏰁1􏰘
􏰁􏰖􏰗PD 􏰂PR 􏰘􏰏􏰙 1 t􏰁1 t􏰁1
E􏰗RCt􏰢Dt 􏰏 1􏰘 􏰂 E􏰗RCt􏰢Dt 􏰏 0􏰘 􏰏 􏰖1
E􏰗D 􏰢D 􏰏1􏰘􏰂E􏰗D 􏰢D 􏰏0􏰘􏰏PD 􏰂PR ,
where D and R superscripts denote which party held the seat—
and hence held the electoral advantage. For example, PD de- t􏰁1
notes the equilibrium probability of a Democrat victory in t 􏰁 1 given that a Democrat held the seat during the campaign of t 􏰁
R
1; P*t􏰁1 represents the “electoral strength” of the Democrat dur-
(4)
(5)
(6)
t􏰁1t t􏰁1t t􏰁1t􏰁1
ing the campaign of t 􏰁 1, given that a Republican held the seat. DR
Note that while we cannot estimate P*t􏰁1 and P*t􏰁1, we can estimate the PD and PR from the data.11
t􏰁1 t􏰁1
These three equations form the basis of our empirical analy-
sis. Equation (4) shows that the total effect 􏰙 of a Democratic victory in t on voting records RCt􏰁1 is the sum of two components,
DR DR
􏰖1[Pt􏰁1 􏰂 Pt􏰁1], and the remainder, 􏰖0[P*t􏰁1 􏰂 P*t􏰁1]. The first
term is the “elect” component. The second term is the “affect” component. The equation shows that the overall effect 􏰙 can be estimated by the simple difference in voting scores RCt􏰁1 be- tween districts won by Democrats and Republicans in t.
The next two equations show how to estimate the “elect”
11. It is important to distinguish between P* and P. P* is a measure of the underlying “popularity” of a party, the probability that party D will win if parties D and R are expected to choose c and 0, respectively. A change in P* represents an exogenous change in popularity. On the other hand, P is the probability that party D will win, at whatever policies the parties are expected to choose.

822 QUARTERLY JOURNAL OF ECONOMICS
component, which is the product of 􏰖 and [PD 􏰂 PR ]. 􏰖 is
1 t􏰁1 t􏰁1 1 estimated by the difference in voting records RC .12 PD 􏰂 PR
t t􏰁1 t􏰁1 is estimated by the difference in the fraction of districts won by
Democrats in t 􏰁 1.
The “affect” component, 􏰖0[P*t􏰁1 􏰂 P*t􏰁1], can be estimated
DR
by 􏰙 􏰂 􏰖 [PD 􏰂 PR ]. If voters merely “elect” policies (com-
1 t􏰁1 t􏰁1
plete divergence), we should observe little change in the candi-
dates’ intended policies following an exogenous increase in the
DR
probability of victory; that is, 􏰖0[P*t􏰁1 􏰂 P*t􏰁1] should be small.
If voters not only choose politicians, but also affect their policy
choices (partial convergence), candidates should move toward
their bliss points in response to an exogenously higher probability
DR
of winning; that is, 􏰖0[P*t􏰁1 􏰂 P*t􏰁1] should be relatively large.
This simple decomposition allows us to make quantitative state-
ments about the relative importance of the “affect” and “elect”
phenomena. We can compute what fraction of the total effect 􏰙 is
explained by the “elect component” 􏰖 [PD 􏰂 PR ], and what 1 t􏰁1 t􏰁1
DR fraction by the “affect component” 􏰖0[P*t􏰁1 􏰂 P*t􏰁1].
Note that the initial “random assignment” of Dt is crucial here. Without this, the estimated differences above would in general be biased for the quantities 􏰙, and 􏰖1.13 As an example, without this random assignment, the simple difference in how Republicans and Democrats vote after election t would reflect both 􏰖1, and that candidates are likely to have more liberal “bliss points” where Democrats hold the seat.
We argue that the examination of suitably “close” elections in period t isolates “as good as random assignment” in Dt. As would be expected from a valid regression-discontinuity design, among
12. As will be evident below, in principle, one could obtain an alternative estimate of 􏰖1, by examining the difference in records RCt􏰁1 among close elections in time t 􏰁 1. In practice, however, this makes little difference because we are pooling data from many years (i.e., the difference between estimating 􏰖1 from data 1946–1994 and estimating it using data from 1948–1994).
13. More formally, without random assignment of Dt, the three expressions would become
E􏰗RCt􏰁1􏰢Dt 􏰏 1􏰘 􏰂 E􏰗RCt􏰁1􏰢Dt 􏰏 0􏰘 􏰏 E􏰗RCt􏰢Dt 􏰏 1􏰘 􏰂 E􏰗RCt􏰢Dt 􏰏 0􏰘 􏰏
E􏰗D 􏰢D􏰏1􏰘􏰂E􏰗D 􏰢D􏰏0􏰘 􏰏 t􏰁1 t t􏰁1 t
􏰙 􏰁 E􏰗εt􏰁1􏰢Dt 􏰏 1􏰘 􏰂 E􏰗εt􏰁1􏰢Dt 􏰏 0􏰘 􏰖0􏰔E􏰗P*t􏰢Dt 􏰏 1􏰘 􏰂 E􏰗P*t􏰢Dt 􏰏 0􏰘􏰕 􏰁􏰖1 􏰁 E􏰗εt􏰢Dt 􏰏 1􏰘 􏰂 E􏰗εt􏰢Dt 􏰏 0􏰘
PD 􏰂PR . t􏰁1 t􏰁1
It is clear, from the expressions above, that without random assignment, the parameter estimates 􏰙 and 􏰖1 would be biased.

DO VOTERS AFFECT OR ELECT POLICIES? 823
elections decided by a very narrow margin, as long as there is some unpredictable component of the ultimate vote tally, who wins the election will be mostly determined by pure chance (e.g., unpredictable components of voter turnout on election day). This is shown more formally in Appendix 1.
By now it should be clear why it is necessary to examine the impact of who wins in t on RCt􏰁1, roll-call votes in period t 􏰁 1. The impact of who wins in t on RCt, roll-call votes in t, only yields 􏰖1. By estimating 􏰖1 alone, it is only possible to test complete convergence 􏰖1 􏰏 0. This extreme hypothesis has already been tested by several studies.14 But 􏰖1 alone is not sufficient to say anything about the size of the “elect” phenomenon relative to the “affect” phenomenon. It is not sufficient for testing full divergence against partial convergence, and hence it is not sufficient for evaluating the Downsian perspective versus the alternative view that politicians face difficulties in credibly committing to policy compromises.
IV. ROLL-CALL VOTING RECORDS IN THE U. S. HOUSE
IV.A. Context
There are several reasons why the U. S. House of Represen- tatives provides an ideal setting in which to empirically assess whether voters primarily affect or elect policies. First, the U. S. federal legislative body is virtually a two-party system, and policy convergence is frequently modeled in a two-party context. When there are more than two candidates, the basic insight of Downs’ [1957] approach to policy convergence arguments becomes more complicated (see Osborne [1995]).
Second, it is well-known that Democrats and Republicans have different (and often directly opposing) policy positions. It is meaningful to ask whether electoral competition compels oppos- ing parties’ nominees to moderate their positions in the face of strong incentives to vote along party lines. If the U. S. House were a relatively nonpartisan environment (with “bliss points” rela- tively close together), the distinction between voters affecting or electing policies would be less important, and a test to distinguish between them less useful.
14. See, for example, Poole and Rosenthal [1984], Levitt [1996], and Snyder and Groseclose [2000].

824 QUARTERLY JOURNAL OF ECONOMICS
Third, the U. S. House is arguably the most likely setting in which to observe policy convergence, if establishing reputations is important. U. S. House elections are held every two years, and there are no term limits (as opposed to gubernatorial and presi- dential elections), meaning that political careers can consist of several terms in office. Furthermore, political tenure in the House is often a stepping-stone to participating in electoral races for higher offices. For these reasons, it is plausible that candidates for the U. S. House have high discount factors, which would allow reputation to support convergent equilibria.
Finally, our empirical analysis focuses on Representatives’ voting records. These votes are directly observable, and are part of the public record. In principle, voters can compare a legislator’s record with their platforms and promises as candidates (and opponents can advertise any deviations during election cam- paigns). Convergent equilibria of the kind described in Alesina [1988] require that policy positions are perfectly observable by voters and that it can be determined whether politicians deviate from policy pronouncements.
IV.B. Data Description
We now discuss the choice of the dependent variable.15 There are several alternative ways to measure Representatives’ voting on legislation. A widely used measure is a voting score provided by the liberal political organization, Americans for Democratic Action (ADA). For each Congress, the ADA chooses about twenty high-profile roll-call votes, and creates an index that varies be- tween 0 and 100 for each Representative of the House. Higher scores correspond to a more “liberal” voting record. Throughout the paper our preferred voting record index is the ADA score. Later, we show that our results are robust to many alternative interest groups scores and other voting record indices.
We utilize data on ADA scores for all Representatives in the U. S. House from 1946–1995, linked to election returns data during that period.16 There is considerable variation in ADA scores within each party. For example, the distribution of ADA scores for Democrat and Republican Representatives in the three
15. All the data and the programs used in this paper are available at http:// www.econ.ucla.edu/moretti/papers.html
16. To make the comparison across Congresses possible, we follow the litera- ture and use “adjusted” ADA scores throughout the paper. This adjustment to the nominal ADA score, was devised by Groseclose, Levitt, and Snyder [1999].

DO VOTERS AFFECT OR ELECT POLICIES? 825
most recent Congresses shows significant overlapping between the parties. It is not uncommon for Democrat representatives to vote more conservatively than Republican candidates, and vice versa.
One advantage of using ADA scores is that it is a widely used index in the literature. However, one limitation is that it includes only twenty votes per Congress, and the choice of what issues to include and what weight to assign to each issue is necessarily arbitrary. To assess how robust our results are to alternative measures of “liberalness” of roll-call votes, we have reestimated all our models using three alternative sets of voting record measures.
First, we use the DW-NOMINATE scores constructed by McCarty, Poole, and Rosenthal [1997]. Poole and Rosenthal [1985] developed the NOMINATE procedure to estimate a low- dimensional measure of political ideology in a complex multidi- mensional political world. NOMINATE is an attempt to estimate the underlying ideology that drives observed roll-call behavior by assigning legislators the ideological points that maximize the number of correctly predicted roll-call votes. The NOMINATE data have the advantage of including all roll-call votes, not an arbitrary subset of votes. It also ignores the Representative’s political party and the legislative issue in question, so it is argu- ably more exogenous than the ADA scores.17
Second, for each member and each Congress, we construct our own measure of loyalty to the party leadership using the individual vote tallies on every issue voted on in the House. For
17. Poole and Rosenthal [1997] note that a single dimension would be un- likely to capture the division between Northern and Southern Democrats during the Civil Rights Era. Therefore, the NOMINATE procedure estimates a two- dimensional measure of ideology where the first dimension captures party loyalty and can be thought of as a liberal to conservative scale, and the second dimension captures the issues of race that divided the Democrats until the mid-1970s. To remain consistent with our discussion of a single ideological dimension, we re- strict our analysis to the first dimension during the period where the second dimension had little predictive power. Specifically, we restrict our DW-NOMI- NATE analysis to 1975 and beyond. However, we have reestimated our models including DW-NOMINATE data for the entire 1946–1995 period and obtained very similar results. For completeness, we have also reestimated our ADA models including only data for the 1975–1995 period and obtained very similar results. We use the DW-NOMINATE scores as opposed to the Poole and Rosenthal’s [1991] earlier D-NOMINATE scores because the DW data cover up through the 106th Congress while the D-NOMINATE data ends with the 99th Congress. McCarty, Poole, and Rosenthal [1997] note that the D-NOMINATE and DW-NOMINATE scores are highly correlated where both scores are available. See Poole and Rosenthal [1997] for a description of the NOMINATE procedure. Poole’s [1999] rank order data yield similar results.

826 QUARTERLY JOURNAL OF ECONOMICS
this measure, we calculate the percent of a representative’s votes that agree with the Democrat party leader.18
Third, we use ratings from interest groups other than the ADA. We include both liberal and conservative ratings from groups such as the American Civil Liberties Union, the League of Women Voters, the League of Conservation Voters, the American Federation of Government Employees, the American Federation of State, County, Municipal Employees, the American Federation of Teachers, the AFL-CIO Building and Construction, the United Auto Workers, the Conservative Coalition, the U. S. Chamber of Commerce, the American Conservative Union, the Christian Vot- ers Victory Fund, the Christian Voice, Lower Federal Spending, and Taxation with Representation. Not all the ratings are avail- able in all years, so sample sizes vary when using these alterna- tive ratings.
As we show below, our results are remarkably stable across alternative measures of roll-call votes. This finding lends some credibility to the conclusion that our estimates are not driven by the unique characteristics of one particular measure. See the Data Appendix for a detailed discussion of our samples and data sources.
V. EMPIRICAL RESULTS
In this section we present our empirical results. Subsection
V.A presents our main results with a simple graphical analysis
that illustrates that changes in electoral strength appears to
affect future voting records entirely because it alters the relative
odds of which party’s nominee will be elected to the House. That
is, candidates do not seem to change their intended policies in
response to large exogenous shocks to electoral strength. This is
followed by more formal estimates of the key parameters of in-
DR DR
terest: 􏰙, 􏰖1[Pt􏰁1 􏰂 Pt􏰁1], and 􏰖0[P*t􏰁1 􏰂 P*t􏰁1]. Subsection V.B
provides evidence supporting our main identifying assumption— that among elections that turn out to be close, who wins is “as good as randomly assigned.” In subsection V.C, we show that our results do not change substantially when we utilize a number of alternative voting record indices. Finally, in subsection V.D, we examine the sensitivity of our results to a functional form as-
18. The results are nearly identical if one uses the party whip instead of the party leader.

DO VOTERS AFFECT OR ELECT POLICIES? 827 sumption utilized in our base model. The analysis within this
more general framework confirms our findings.
V.A. Main Empirical Results
Graphical Analysis. As discussed above, the way to distin-
guish between full divergence and partial divergence is to analyze
the effects on roll-call votes of an exogenous change in the prob-
ability of winning the election. The total effect of such an exoge-
nous change on roll-call behavior (􏰙) can be split into two compo-
nents: the elect component (􏰖 [PD 􏰂 PR ]) and the affect
1 t􏰁1 t􏰁1
component (􏰖0[P*t􏰁1 􏰂 P*t􏰁1]). If voters merely “elect” policies
DR
(complete divergence), we should observe little change in the candidates’ intended policies following an exogenous increase in
DR
the probability of victory: 􏰖0[P*t􏰁1 􏰂 P*t􏰁1] should be small. If
voters not only choose politicians, but also affect their policy
choices (partial convergence), candidates should move toward
their bliss points in response to an exogenously higher probability
DR
of winning: 􏰖0[P*t􏰁1 􏰂 P*t􏰁1] should be relatively important.
We begin with a simple graphical analysis of ADA scores. Figure I plots ADA scores at time t 􏰁 1 against the Democrat vote share at time t. As an example, we are relating the ADA scores for the representative from, say, the first California district during the 1995–1996 Congressional session to the Democratic vote share observed in the 1992 election in that district. In practice, we use all pairs of adjacent years from 1946 to 1995, except for the pairs where we cannot link districts due to redistricting (pairs with years ending with “0” and “2”).
Throughout the paper the unit of observation is the district in a given year. But to give an overall picture of the data, each point in Figure I is an average of the ADA score in period t 􏰁 1 within 0.01-wide intervals of the vote share at time t. The vertical line marks 50 percent of the two-party vote share. Districts to the right of the vertical line are districts won by Democrats in elec- tion t, districts to the left are districts won by Republicans in election t. The continuous line is the predicted ADAt􏰁1 score from a regression that includes a fourth-order polynomial in vote share and a dummy for observations above the 50 percent threshold, and an interaction of the dummy and the polynomial. The dotted lines represent pointwise 95 percent confidence intervals of this approximation.
A striking feature of the figure is that ADA scores appear to

828 QUARTERLY JOURNAL OF ECONOMICS
FIGURE I
Total Effect of Initial Win on Future ADA Scores: 􏰙
This figure plots ADA scores after the election at time t 􏰁 1 against the Democrat vote share, time t. Each circle is the average ADA score within 0.01 intervals of the Democrat vote share. Solid lines are fitted values from fourth- order polynomial regressions on either side of the discontinuity. Dotted lines are pointwise 95 percent confidence intervals. The discontinuity gap estimates
􏰙􏰏􏰖 􏰔P*D 􏰂P*R 􏰕􏰁􏰖 􏰔P*D 􏰂P*R 􏰕. 0 t􏰁1 t􏰁1 1 t􏰁1 t􏰁1
“Affect” “Elect”
be a continuous and smooth function of vote shares everywhere, except at the threshold that determines party membership. There is a large discontinuous jump in ADA scores at the 50 percent threshold. Compare districts where the Democrat candidate barely lost in period t (for example, vote share is 49.5 percent), with districts where the Democrat candidate barely won (for example, vote share is 50.5 percent). If the regression disconti- nuity design is valid, the two groups of districts should appear ex ante similar in every respect—on average. The difference will be that in one group, the Democrats will be the incumbent for the next election (t 􏰁 1), and in the other it will be the Republicans. Districts where the Democrats are the incumbent party for elec- tion t 􏰁 1 elect representatives who have much higher ADA scores, compared with districts where the Republican candidate

DO VOTERS AFFECT OR ELECT POLICIES? 829
barely won and became the incumbent—on average. The size of
the jump appears to be fairly large, at around twenty ADA points.
What does this discontinuity mean? Formally, the gap is a
credible estimate of the parameter 􏰙 in equation (4). Intuitively,
it is unsurprising to observe some discontinuity. We know that
party affiliation is an important determinant of roll-call behavior.
We also know that if a Democrat (Republican) is elected in period
t, a Democrat (Republican) is more likely to be elected in period
t 􏰁 1 in the same district, due to the incumbency advantage. The
party effect, together with the electoral advantage of incumbency,
suggests that we should expect to find a gap in Figure I. It is not
surprising to observe that, for example, the 1995–1996 voting
records are more liberal in the districts that were won by Demo-
crats in 1992. The 1995–1996 representative, after all, is more
likely be a Democrat. In Sections II and III we called this par-
ticularmechanismthe“electcomponent,”anddenotedit􏰖 [PD 􏰂
PR ]. t􏰁1
There is a second component that contributes to 􏰙. If candi-
dates are constrained by expected voters’ behavior, then a Demo-
crat who is challenging an incumbent Republican (left side of the
graph) would be expected to moderate his intended policies more,
compared with an incumbent Democrat (right side of the graph).
After all, the incumbent would be in a much stronger electoral
position compared with the challenger. This is the other reason
why voting scores should be more liberal where the Democrat is
the incumbent, and hence why there should be a gap in Figure I.
In Sections II and III we labeled this phenomenon “voters affect-
DR ing policies,” and denoted it 􏰖0[P*t􏰁1 􏰂 P*t􏰁1].
The discontinuity 􏰙 illustrated in Figure I is equal to DRDR
􏰖0[P*t􏰁1 􏰂 P*t􏰁1] 􏰁 􏰖1[Pt􏰁1 􏰂 Pt􏰁1]. While it is not surprising DR
to find that 􏰙 􏰐 0, the real question is whether 􏰖0[P*t􏰁1 􏰂 P*t􏰁1] or 􏰖 [PD 􏰂 PR ] dominates. Our empirical strategy is simple.
1 t􏰁1 t􏰁1
We can directly estimate 􏰖 [PD 􏰂 PR ], by separately estimat-
1 t􏰁1 t􏰁1
ing 􏰖1, the expected difference in voting between the two parties,
as well as [PD 􏰂 PR ], the electoral advantage to incumbency. t􏰁1 t􏰁1
We can subtract this from the total effect 􏰙 to determine the DR
magnitude of the “affect component” 􏰖0[P*t􏰁1 􏰂 P*t􏰁1].
In Figure II we illustrate the two elements that make up the
elect component, 􏰖 and [PD 􏰂 PR ]. The top panel in Figure 1 t􏰁1 t􏰁1
II plots ADA scores at time t against the Democrat vote share at time t. As in Figure I, average ADA scores appear to be a con- tinuous and smooth function of vote shares everywhere, except at
1 t􏰁1

830
QUARTERLY JOURNAL OF ECONOMICS
FIGURE IIa
Effect of Party Affiliation: 􏰖1
FIGURE IIb
Effect of Initial Win on Winning Next Election: (PD 􏰂 PR )
t􏰁1 t􏰁1
Top panel plots ADA scores after the election at time t against the Democrat vote share, time t. Bottom panel plots probability of Democrat victory at t 􏰁 1
against Democrat vote share, time t. See caption of Figure III for more details.

DO VOTERS AFFECT OR ELECT POLICIES? 831
the threshold that determines party membership. The jump is a credible estimate of 􏰖1 in equation (5).
Compare a district where the Democrat candidate barely lost at time t (for example, vote share of 49.5 percent), with a district where the Democrat candidate barely won at time t (for example, vote share of 50.5 percent). Again, if the regression discontinuity design is valid and has generated random assignment of who wins in t, then the average voting records of Democrats who are barely elected will credibly represent, on average, how Democrats would have voted in the districts that were in actuality, barely won by Republicans (and vice versa). The observed difference in voting scores represents a credible estimate of the average policy differences between the two parties across districts—the direct influence of party affiliation on voting scores. The difference at the 50 percent threshold appears quite large, with a gap of about 45 points.
Finally, the bottom panel of Figure II plots estimates of the probability that the Democrat will win election t 􏰁 1 for a given Democratic vote share at t. As in the previous cases, the figure shows a smooth function of vote shares everywhere, except at the threshold that determines which party won t. The size of the jump estimates [PD 􏰂 PR ] in equation (6).19
t􏰁1 t􏰁1
The discontinuity around the 50 percent threshold indicates
that, for example, districts which barely elected a Democrat in t are more likely to elect a Democrat in t 􏰁 1, consistent with a causal incumbency advantage. This is consistent with these dis- tricts experiencing exogenous increases in the probability of elect- ing a Democrat (Republican) in 1994.
The total effect 􏰙, given by the discontinuity in Figure I,
appears to be about twenty ADA points. Figure IIa shows that the
estimate of 􏰖1 is about 45 points, and Figure IIb shows that
[PD 􏰂 PR ] is around 0.5. The “elect component” 􏰖 [PD 􏰂 t􏰁1 t􏰁1 1t􏰁1
PR ] is thus approximately 45 􏰚 0.5 􏰏 22.5. The small difference t􏰁1
(20 –22.5) implies that the entire effect of an exogenous change in electoral strength on future ADA scores is not operating through how candidates’ policy choices respond to changes in the proba- bility of winning. Instead, the effect is operating through simply changing the relative odds that a party will retain control over the seat. That is, this graphical analysis indicates that voters
19. This regression-discontinuity estimate of the incumbency advantage is documented in Lee [2001, 2003].

832
QUARTERLY JOURNAL OF ECONOMICS
TABLE I
RESULTS BASED ON ADA SCORES—CLOSE ELECTIONS SAMPLE
Variable
ADAt
(1.3)
DEMt􏰁1 (col. (2)*(col. (3)) (col. (1)) 􏰂 (col. (4))
(5)
􏰂1.64 (2.0)
Total effect
􏰙 ADAt􏰁1
(1)
21.2 (1.9)
Elect component Affect component
􏰖 (PD 􏰂PR )􏰖[(PD 􏰂PR )] 􏰖[P*D 􏰂P*R ] 1 t􏰁1 t􏰁1 1 t􏰁1 t􏰁1 0 t􏰁1 t􏰁1
(2)
47.6 0.48
Estimated gap
(3) (4) (0.02)
22.84 (2.2)
Standard errors are in parentheses. The unit of observation is a district-congressional session. The sample includes only observations where the Democrat vote share at time t is strictly between 48 percent and 52 percent. The estimated gap is the difference in the average of the relevant variable for observations for which the Democrat vote share at time t is strictly between 50 percent and 52 percent and observations for which the Democrat vote share at time t is strictly between 48 percent and 50 percent. Time t and t 􏰁 1 refer to congressional sessions. ADAt is the adjusted ADA voting score. Higher ADA scores correspond to more liberal roll-call voting records. Sample size is 915.
primarily elect policies (full divergence) rather than affect poli- cies (partial convergence).
Here we quantify our estimates more precisely. In the analy- sis that follows, we restrict our attention to “close elections”— where the Democrat vote share in time t is strictly between 48 and 52 percent. As Figures I and II show, the difference between barely elected Democrat and Republican districts among these elections will provide a reasonable approximation to the discon- tinuity gaps. There are 915 observations, where each observation is a district-year.20
Table I, column (1), reports the estimated total effect 􏰙, the size of the jump in Figure I. Specifically, column (1) shows the difference in the average ADAt􏰁1 for districts for which the Democrat vote share at time t is strictly between 50 percent and 52 percent and districts for which the Democrat vote share at time t is strictly between 48 percent and 50 percent. The esti- mated difference is 21.2.
In column (2) we estimate the coefficient 􏰖1, which is equal to the size of the jump in Figure IIa. The estimate is the difference in the average ADAt for districts for which the Democrat vote
20. In 68 percent of cases, the representative in period t 􏰁 1 is the same as the representative in period t. The distribution of close elections is fairly uniform across the years. In a typical year there are about 40 close elections. The year with the smallest number is 1988, with twelve close elections. The year with the largest number is 1966, with 92 close elections.

DO VOTERS AFFECT OR ELECT POLICIES? 833
share at time t is strictly between 50 percent and 52 percent and districts for which the Democrat vote share at time t is strictly between 48 percent and 50 percent. The estimated difference is 47.6.
In column (3) we estimate the quantity [PD 􏰂 PR ], which t􏰁1 t􏰁1
is equal to the size of the discontinuity documented in Figure IIb. The estimated jump is 0.48. This indicates that if the Democrat (Republican) candidate wins a close election in a given district in, say, 1992, the Democrat (Republican) candidate in the same district has a 0.48 higher probability of winning in 1994. This is indeed consistent with the notion that the party that already holds a seat holds a substantial electoral advantage.
In column (4) we multiply the estimates in columns (2) and (3) to obtain an estimate of the elect component, 􏰖 [PD 􏰂 PR ].
1 t􏰁1 t􏰁1 The product is 22.84, which is not statistically different from the
estimate of 􏰙 in column (1). Because estimates of 􏰙 and 􏰖 [PD 􏰂 1 t􏰁1
PR ] are quite similar, we conclude that the “affect component” t􏰁1
is quite small. In column (5) we subtract the estimate in column
DR
(4) from the estimate in column (1) to yield 􏰖0[P*t􏰁1 􏰂 P*t􏰁1]. The
difference is virtually zero.
What is the relative importance of the “elect component”
DR DR 􏰖1[Pt􏰁1 􏰂 Pt􏰁1], and the “affect component” 􏰖0[P*t􏰁1 􏰂 P*t􏰁1]?
Our results indicate that 􏰖 [PD 􏰂 PR ] overwhelmingly domi-
1 t􏰁1 t􏰁1
nates 􏰖0[P*t􏰁1 􏰂 P*t􏰁1]. Indeed, it entirely explains the overall
DR
effect 􏰙. Below, we show that this is true not only on average, but also for every decade taken separately.
V.B. Tests for Quasi-Random Assignment
It is important to note that our empirical test crucially relies on the assumption of random assignment of the winner in close elections in t. Specifically, the key identifying assumption in our analysis is that as one compares closer and closer elections, all predetermined characteristics of Republican and Democratic dis- tricts (including the district-specific bliss points) become more and more similar. If this assumption does not hold, our estimates are likely to be biased.
Intuitively, this assumption seems to make sense. While Republican and Democrat districts are likely to be very different in general, the difference should decline as we examine elections whereby “pure luck” is a more important determinant of who wins—in other words, elections that turn out to be won by a tiny

834 QUARTERLY JOURNAL OF ECONOMICS
margin. In the Appendix we provide a formal discussion of this assumption. Here, we provide two pieces of empirical evidence to support this assumption.
First, if examining close elections truly provides random assignment, characteristics determined before time t should be the same on both sides of the 50 percent threshold—on aver- age.21 We find that as we compare closer and closer elections, Republican and Democrat districts do have similar observable characteristics. Consider, for example, geographical location. There are sizable geographical differences in the entire sam- ple. Averaging over the entire time period, Democrats are significantly more likely to be elected in the South than in the North and the West. However, as we start restricting the sample to closer and closer elections, the geographical differ- ences decrease. For elections that are within only two percent- age points from the threshold, the differences are not statisti- cally significant.
This is shown graphically in Figures III and IV, which plot average district characteristics against Democratic vote share. Other than geographical location, we consider the following pre- determined characteristics: real income, percentage with high school degree, percentage black, percentage eligible to vote, and size of the voting population. Generally, the figures indicate that the difference at the 50 percent threshold is small and statisti- cally insignificant.
Table II illustrates the same point by quantifying the differ- ence between Democrat and Republican districts for a larger set of characteristics. In particular, we examine all the characteris- tics shown in Figures III and IV, as well as the fraction of open seats, percent urban, percent manufacturing employment, and percent eligible to vote.22 Column (1) includes the entire sample. Columns (2) to (5) include only districts with Democrat vote share between 25 percent and 75 percent, 40 percent and 60 percent, 45 percent and 55 percent, and 48 percent and 52 percent, respec- tively. The model in column (6) is equivalent to Figures III and IV, since it includes a fourth-order polynomial in Democrat vote
21. See Lee [2003] for the conditions under which RD designs can generate variation in the treatment that is as good as randomized.
22. Data on districts’ characteristics in each election year are from the last available Census of Population. Because the census takes place every ten years, standard errors allow for clustering at the district-decade level.

DO VOTERS AFFECT OR ELECT POLICIES? 835
FIGURE III
Similarity of Constituents’ Characteristics in Bare Democrat and Republican Districts–Part 1
Panels refer to (from top left to bottom right) the following district character- istics: real income, percentage with high-school degree, percentage black, percent- age eligible to vote. Circles represent the average characteristic within intervals of 0.01 in Democrat vote share. The continuous line represents the predicted values from a fourth-order polynomial in vote share fitted separately for points above and below the 50 percent threshold. The dotted line represents the 95 percent confidence interval.
share. The coefficient reported in column (6) is the predicted difference at 50 percent. The table confirms that, for many ob- servable characteristics, there is no significant difference in a close neighborhood of 50 percent. One important exception is the percentage black, for which the magnitude of the discontinuity is statistically significant.23
As a consequence, estimates of the coefficients in Table I from regressions that include these covariates would be expected to produce similar results—as in a randomized experiment—since
23. This is due to few outliers in the outer part of the vote share range. When the polynomial is estimated including only districts with vote share between 25 percent and 75 percent, the coefficients becomes insignificant. The gap for percent urban and open seats, while not statistically significant at the 5 percent level, is significant at the 10 percent level.

836 QUARTERLY JOURNAL OF ECONOMICS
FIGURE IV
Similarity of Constituents’ Characteristics in Bare Democrat and Republican Districts—Part 2
Panels refer to (from top left to bottom right) the following district character- istics: voting population, North, South, West. Circles represent the average char- acteristic within intervals of 0.01 in Democrat vote share. The continuous line represents the predicted values from a fourth-order polynomial in vote share fitted separately for points above and below the 50 percent threshold. The dotted line represents the 95 percent confidence interval.
all predetermined characteristics appear to be orthogonal to Dt. We have reestimated all the models in Table I conditioning on all of the district characteristics in Table II, and found estimates that are virtually identical to the ones in Table I.
As a similar empirical test of our identifying assumption, in Figure V we plot the ADA scores from the Congressional sessions that preceded the determination of the Democratic two-party vote share in election t. Since these past scores have already been determined by the time of the election, it is yet another predeter- mined characteristic (just like demographic composition, income levels, etc.). If the RD design is valid, then we should observe no discontinuity in these lagged ADA scores—just as we would ex- pect, in a randomized experiment, to see no systematic differ- ences in any variables determined prior to the experiment. The

DO VOTERS AFFECT OR ELECT POLICIES? 837
TABLE II
DIFFERENCE IN DISTRICT CHARACTERISTICS BETWEEN DEMOCRAT AND REPUBLICAN DISTRICTS, BY DISTANCE FROM 50 PERCENT
North
South
West
Log income
Percentage high-school grad.
Percentage urban
Percentage black
All (1)
􏰂0.211 (0.018) 0.250 (0.015) 􏰂0.031 (0.013) 􏰂0.086 (0.013) 􏰂0.035 (0.003) 0.070 (0.011) 0.082 (0.005) 􏰂0.002
􏰁/􏰂 25 (2)
􏰂0.156 (0.019) 0.145 (0.014) 􏰂0.012 (0.015) 􏰂0.036 (0.012) 􏰂0.024 (0.003) 0.065 (0.011) 0.042 (0.004) 0.000 (0.001) 3019.2 (3723.0) 0.010 (0.002) 0.065 (0.011)
10229
􏰁/􏰂 10 (3)
􏰂0.096 (0.021) 0.093 (0.016) 􏰂0.036 (0.020) 0.014 (0.014) 􏰂0.008 (0.004) 0.053 (0.012) 0.013 (0.004) 0.004 (0.002) 4961.5 (4562.4) 0.007 (0.003) 0.053 (0.012)
4174
􏰁/􏰂 5 (4)
􏰂0.054 (0.024) 0.053 (0.019) 􏰂0.003 (0.017) 0.026 (0.017) 􏰂0.001 (0.004) 0.053 (0.014) 0.003 (0.005) 0.004 (0.002) 3211.4 (5524.2) 0.006 (0.004) 0.053 (0.014)
2072
􏰁/􏰂 2 (5)
􏰂0.059 (0.036) 0.009 (0.028) 0.001 (0.020) 0.030 (0.026) 0.001 (0.007) 0.056 (0.023) 􏰂0.003 (0.009) 0.005 (0.004) 8640.4 (8427.9) 􏰂0.003 (0.006) 0.056 (0.023)
910
Polynomial (6)
􏰂0.041 (0.045) 0.015 (0.036) 0.001 (0.036) 0.052 (0.033) 0.008 (0.008) 0.053 (0.028) 􏰂0.053 (0.013) 0.003 (0.005) 2007.5 (10483.0) 􏰂0.003 (0.007) 0.053 (0.028)
13413
Manufacturing
employment (0.001)
Total population
Percentage eligible to vote
Open seats
Number of observations
􏰂1817.9 (3517.3) 0.005 (0.002) 0.070 (0.011)
13413
Standard errors are in parentheses. The unit of observation is a district-congressional session. Columns (1) to (5) report the difference in average district characteristics between Democrat and Republican districts. Column (1) includes the entire sample. Columns (2) to (5) include only districts with Democrat vote share between 25 percent and 75 percent, 40 percent and 60 percent, 45 percent and 55 percent, and 48 percent and 52 percent, respectively. The model in column (6) includes a fourth-order polynomial in Democrat vote share that enters separately for vote share above and below 50 percent. The coefficient reported in column (6) is the predicted difference at 50 percent. All standard errors account for district-decade clustering.
lack of discontinuity in the figure lends further credibility to our identifying assumption.24
Overall, the evidence strongly supports a valid regression discontinuity design. And as a consequence, it appears that among close elections, who wins appears virtually randomly as- signed, which is the identifying assumption of our empirical strategy.
24. The estimated gap is 3.5 (5.6).

838 QUARTERLY JOURNAL OF ECONOMICS
FIGURE V
Specification Test: Similarity of Historical Voting Patterns between Bare Democrat and Republican Districts
The panel plots one time lagged ADA scores against the Democrat vote share. Time t and t 􏰂 1 refer to congressional sessions. Each point is the average lagged ADA score within intervals of 0.01 in Democrat vote share. The continuous line is from a fourth-order polynomial in vote share fitted separately for points above and below the 50 percent threshold. The dotted line is the 95 percent confidence interval.
V.C. Sensitivity to Alternative Measures of Voting Records
Our results so far are based on a particular voting index, the ADA score. In this section we investigate whether our results generalize to other voting scores. We find that the findings do not change when we use alternative interest groups scores, or other summary measures of representatives’ voting records.
Table III is analogous to Table I, but instead of using ADA
scores, it is based on two alternative measures of roll-call voting.
The top panel is based on McCarty, Poole, and Rosenthal’s DW-
NOMINATE scores. The bottom panel is based on the percent of
individual roll-call votes cast that are in agreement with the
Democrat party leader. All the qualitative results obtained using
ADA scores (Table I) hold up using these measures. When we use
the DW-NOMINATE scores, 􏰙 is 􏰂0.36, remarkably close to the
corresponding estimate of 􏰖 [PD 􏰂 PR ] in column (4), which 1 t􏰁1 t􏰁1
is 􏰂0.34. The estimates are negative here because, unlike ADA scores, higher Nominate scores correspond to a more conservative voting record. When we use the measure “percent voting with the Democrat leader,” 􏰙 is 0.13, almost indistinguishable from the

DO VOTERS AFFECT OR ELECT POLICIES? 839
TABLE III
RESULTS BASED ON NOMINATE SCORES AND ON PERCENT VOTED LIKE DEMOCRAT LEADERSHIP—CLOSE ELECTIONS SAMPLE
Variable
(Pt􏰁1 􏰂 Pt􏰁1) DEMt􏰁1
(3)
0.62 (0.04)
􏰖1[(Pt􏰁1 􏰂 Pt􏰁1)] (col. (2)*(col. (3)) (4)
􏰂0.34 (0.04)
0.13 (0.02)
􏰖0[P*t􏰁1 􏰂 P*t􏰁1] (col. (1)) 􏰂 (col. (4)) (5)
􏰂0.02 (0.04)
0.00 (0.02)
Total effect
􏰙 􏰖1
Zt􏰁1 Zt (1) (2)
Elect component Affect component
DRDRDR
(a) Results based on Nominate scores Estimated gap 􏰂0.36 􏰂0.58 (0.03) (0.02)
(b) Results based on percent voted like Democrat leadership Estimated gap 0.13 0.29 0.46 (0.01) (0.006) (0.02)
Standard errors are in parentheses. The unit of observation is a district-congressional session. The sample includes only observations where the Democrat vote share at time t is strictly between 48 percent and 52 percent. The estimated gap is the difference in the relevant variable for observations for which the Democrat vote share at time t is strictly between 50 percent and 52 percent and observations for which the Democrat vote share at time t is strictly between 48 percent and 50 percent. Time t and t 􏰁 1 refer to congressional sessions. The top panel uses the DW-NOMINATE score constructed by McCarty, Poole, and Rosenthal. Higher Nominate scores correspond to more conservative roll-call voting records. The bottom panel uses the percent of a representative’s votes that agree with the Democrat party leader. Sample size is 276 in top panel and 1010 in bottom panel.
estimate 􏰖 [PD 􏰂 PR ] in column (4), which is 0.13. We show 1 t􏰁1 t􏰁1
the graphical analysis for the estimate of 􏰖1 in Figure VI.
Our empirical findings are also not sensitive to the use of ratings from various liberal and conservative interest groups. Liberal interest groups include the American Civil Liberties Union, the League of Women Voters, the League of Conservation Voters, the American Federation of Government Employees, the American Federation of State, County, and Municipal Employees, the American Federation of Teachers, the AFL-CIO Building and Construction, and the United Auto Workers. Conservative groups include the Conservative Coalition, the U. S. Chamber of Com- merce, the American Conservative Union, and the Christian Voice. All the ratings range from 0 to 100. For liberal groups, low ratings correspond to conservative roll-call votes, and high rat- ings correspond to liberal roll-call votes. For conservative groups
the opposite is true.
These alternative ratings yield results that are qualitatively
similar to our findings in Table I and III. Instead of presenting these results in a table format as we did in Table I and III, we present the main results in graphical form. We summarize our

840 QUARTERLY JOURNAL OF ECONOMICS
FIGURE VI
Nominate Scores, by Democrat Vote Share; and Percent Voted with Democrat Leader, by Democrat Vote Share
The top panel plots DW-Nominate scores at time t against the Democrat vote share at time t. Circles represent the average Nominate score within intervals of 0.01 in Democrat vote share. The bottom panel plots the fraction of a Representative’s votes that agree with the Democrat party leader at time t against the Democrat vote share at time t. Circles represent the percent voted with Democrat leader within intervals of 0.01 in Democrat vote share. The continuous line is from a fourth-order polynomial in vote share fitted separately for points above and below the 50 percent threshold. The dotted line is the 95 percent confidence interval.

DO VOTERS AFFECT OR ELECT POLICIES? 841
FIGURE VII
Total Effect [􏰙] Versus “Elect Policy” Component [􏰖 (PD 􏰂 PR )], for
1 t􏰁1 t􏰁1 Alternative Interest Group Ratings
The figure plots estimates of the Total Effect [􏰙] on the y-axis against estimates of the “Elect Policy” component [􏰖 (PD 􏰂 PR )] on the x-axis for eleven interest
1 t􏰁1 t􏰁1
groups ratings. The line is the 45° line. The closer the total effect is to the “elect component” effect for an interest group, the closer that group symbol is to the 45° line. All ratings are between 0 and 100. Higher ratings of liberal interest groups correspond to more liberal roll-call voting records. Higher ratings of conservative
interest groups correspond to more conservative roll-call voting records.
ACLU is America Civil Liberties Union; LWV is League of Women Voters; LCV is League of Conservation Voters; AFGE is American Federal Government Em- ployees; AFSCME is American Federation State, County, Municipal Employees; AFT is American Federation of Teachers; BCTD is AFL-CIO Building and Con- struction; UAW is United Auto Workers; CC is Conservative Coalition; CCUS is U. S. Chamber of Commerce; ACU is American Conservative Union.
results in Figure VII, where we plot our estimate of 􏰙 against our estimate of 􏰖 [PD 􏰂 PR ] for each of these alternative interest
1 t􏰁1 t􏰁1
group ratings. The diagonal is the 45° degree line. Most estimates
are on the line or close to the line, indicating again that across a variety of different interest groups scores, the results are highly consistent with the full policy divergence hypothesis.25
Our qualitative findings seem insensitive to the choice of voting score. Representatives’ policy positions, on a wide array of
25. Figures VIII to XI show that the relationship between ratings, and the democrat vote share shares the same general features as the relationships for ADA, DW-NOMINATE, and “Percent vote with Democrat Leader.” In all cases, we find a large discontinuity.

842 QUARTERLY JOURNAL OF ECONOMICS
FIGURE VIII
Liberal Interest Groups Ratings, by Democrat Vote Share—Part 1
The top panel on the left refers to ratings from United Auto Workers. The top panel on the right refers to ratings from American Federation of Teachers. The bottom panel on the left refers to ratings from American Federation of Govern- ment Employees. The bottom panel on the right refers to ratings from League of Conservative Voters. Circles represent average ratings within intervals of 0.01 in Democrat vote share. The continuous line is from a fourth-order polynomial in vote share fitted separately for points above and below the 50 percent threshold. The dotted line is the 95 percent confidence interval.
issues, do not seem to respond to exogenous changes in electoral strength. Voters appear to elect, rather than affect, candidates’ platforms.
V.D. Heterogeneity
We now turn to the important issue of heterogeneity. Candi- dates’ bliss points can be very different across districts or over time. For example, in any given year, a Democrat from Alabama is likely to have a bliss point that is quite different from a Democrat from Massachusetts. Our main results in subsection V.A are based on a model where candidates’ positions can vary across districts and years. But one implicit functional form as- sumption that we adopted was that the difference in policy posi- tions between Democrat and Republican candidates is constant across districts and over time. This assumption is violated if, for

DO VOTERS AFFECT OR ELECT POLICIES? 843
FIGURE IX
Liberal Interest Groups Ratings, by Democrat Vote Share—Part 2
The top panel on the left refers to ratings from League of Women Voters. The top panel on the right refers to ratings from American Civil Liberties Union. The bottom panel on the left refers to ratings from Taxation with Representation. Circles represent average ratings within intervals of 0.01 in Democrat vote share. The continuous line is from a fourth-order polynomial in vote share fitted sepa- rately for points above and below the 50 percent threshold. The dotted line is the 95 percent confidence interval.
example, the gap in intended policies between Democrats and Republicans from Alabama is different from the gap between Democrats and Republicans from Massachusetts.
We show, however, that our findings are robust to a more general framework that allows for virtually unrestricted hetero- geneity in the gap between opposing candidates’ policies across time and districts. The details can be found in Appendix 3.
Intuitively, we know that the total effect 􏰙 is partially driven by the impact of who wins election t on the composition of Demo- crats and Republicans in office after election t 􏰁 1. In the Ap- pendix we show that it is possible, ex post, to identify the “mar- ginal” districts that switched from Republican to Democrat in t 􏰁 1, because the Democrat won in t. By deleting these districts from the sample, we can examine the impact of an increase in electoral strength on policy positions, without the confounding effects of the compositional change (i.e., the “elect component”).

844 QUARTERLY JOURNAL OF ECONOMICS
FIGURE X
Conservative Interest Groups Ratings, by Democrat Vote Share—Part 1
The top panel on the left refers to ratings from the Lower Federal Spending Support Score. The top panel on the right refers to ratings from the Conservative Coalition. The bottom panel on the left refers to ratings from the Chamber of Commerce. The bottom panel on the right refers to ratings from the Christian Voters Victory Fund. Circles represent average ratings within intervals of 0.01 in Democrat vote share. The continuous line is from a fourth-order polynomial in vote share fitted separately for points above and below the 50 percent threshold. The dotted line is the 95 percent confidence interval.
Our findings using this more general framework indicate
DR
small estimates of 􏰖0[P*t􏰁1 􏰂 P*t􏰁1]. This implies that our finding
that voters primarily elect policies is not an artifact of a func- tional form assumption of our basic specification.
We conclude this section considering a different type of het- erogeneity: heterogeneity over time. In Table IV we replicate the estimates in Table I, presenting separate estimates by decade.26 Column (2) shows that the discontinuity estimated by pooling all the years (Table II, column (2)) masks some variation in the discontinuity gap across states and years. This is not surprising, as the political science literature, for example, has noted that in the South, Democrats and Republicans are ideologically closer
26. Because there are few observations in 1940 and 1990, we merge 1940 with 1950, and 1980 with 1990.

DO VOTERS AFFECT OR ELECT POLICIES? 845
FIGURE XI
Conservative Interest Groups Ratings, by Democrat Vote Share—Part 2
The top panel on the left refers to ratings from the American Conservative Union. The top panel on the right refers to ratings from the Christian Voice. The bottom panel on the left refers to ratings from the National Taxpayers’ Union. Circles represent average ratings within intervals of 0.01 in Democrat vote share. The continuous line is from a fourth-order polynomial in vote share fitted sepa- rately for points above and below the 50 percent threshold. The dotted line is the 95 percent confidence interval.
than they are in the North. The estimated discontinuity is rela- tively smaller during the 1970s, and relatively larger during the 1990s. When we stratify by region and decade (not shown), the discontinuity is relatively smaller in the South in the 1950s and 1970s. Consistent with previous evidence, column (3) shows that the incumbency advantage has increased over time. The key findings of the table are in column (5). Notably, the relative importance of the “elect component” and the “affect component” remains fairly stable over time. Entries in column (5) indicate that our findings in Table I hold in each decade from 1940 to 1990.27
27. We have also estimated separate models for presidential election years and midterm election years. Our results are similar to the ones in Table I. For example, estimates similar to column (5) for presidential election years and midterm election years are, respectively, 􏰂0.1 (2.8) and 􏰂2.5 (3.0).

846
QUARTERLY JOURNAL OF ECONOMICS
TABLE IV
RESULTS BASED ON ADA SCORES, BY DECADE—CLOSE ELECTIONS SAMPLE
Variable ADAt􏰁1 1946–1958 14.2
(3.2)
1960–1968 23.5 (3.5)
1970–1978 11.5 (4.7)
1980–1996 46.8 (3.7)
ADAt DEMt􏰁1 41.7 0.41
(2.3) (0.05)
49.5 0.51 (2.7) (0.05)
46.6 0.40 (3.1) (0.06)
56.6 0.76 (2.8) (0.05)
(col. (2)*(col. (3))
17.0 (4.8)
25.2 (4.9)
18.6 (5.1)
43.0 (4.9)
(col. (1)) 􏰂 (col. (4))
􏰂2.8 (4.0)
􏰂1.7 (4.1)
􏰂7.1 (5.1)
3.8 (4.5)
(1) Total effect 􏰙
(2) (3)
􏰖 (PD 􏰂PR
1 t􏰁1 t􏰁1 1 t􏰁1 t􏰁1 0 t􏰁1 t􏰁1
(4) (5)
Elect component Affect component
) 􏰖[(PD 􏰂PR )] 􏰖[P*D 􏰂P*R ]
Standard errors are in parentheses. The unit of observation is a district-congressional session. The sample includes only observations where the Democrat vote share at time t is strictly between 48 percent and 52 percent. The estimated gap is the difference in the average of the relevant variable for observations for which the Democrat vote share at time t is strictly between 50 percent and 52 percent and observations for which the Democrat vote share at time t is strictly between 48 percent and 50 percent. Time t and t 􏰁 1 refer to congressional sessions. ADAt is the adjusted ADA voting score. Higher ADA scores correspond to more liberal roll-call voting records. Sample sizes are 322 in 1946 –1958; 245 in 1960 –1968; 183 in 1970 –1978; 164 in 1980–1996.
VI. RELATION TO PREVIOUS EMPIRICAL LITERATURE
A number of empirical studies have directly or indirectly examined the policy convergence issue.28 Typically, the studies examine whether party affiliation matters for the observed voting records of the legislator. Most studies find evidence of this, which is strictly inconsistent with the complete policy convergence re- sult. For example, Poole and Rosenthal [1984] show that senators from the same state belonging to different parties have signifi- cantly different voting records.
28. An example of early empirical work in this area is Miller and Stokes [1963]. The literature is too large to be summarized here. Other examples include, but are not limited to, Snyder and Ting [2001a], Fiorina [1999], Poole and Rosenthal [2001], Snyder and Ting [2001b], Lott and Davis [1992], Canes-Wrone, Brady, and Cogan [2002], Krehbiel [2000], Bender [1991], McArthur and Marks [1988], and McCarty, Poole, and Rosenthal [2000].

DO VOTERS AFFECT OR ELECT POLICIES? 847
More recently, Snyder and Groseclose [2000] estimate the effect of party affiliation on roll-call votes using an identification strategy based on the assumption that on votes with 65 percent or more legislators on one side, legislators are not subject to party influence. Their main concern is that the correlation between party affiliation and voting may be spurious, possibly reflecting legislators’ or constituency preferences. They conclude that in the majority of cases, party affiliation is a significant determinant of roll calls.29
We have argued here, however, that evidence against the (arguably unrealistic) outcome of complete convergence says little about the relative importance of the broader Downsian notion that electoral competition pressures politicians to moderate their positions. An appropriate assessment of this phenomenon rests on being able to empirically differentiate between partial conver- gence and complete policy divergence.
Other studies have noted a correlation between the elector- ate’s and legislator’s ideologies across constituencies and over time [Levitt 1996; Snyder and Ting 2001a]. This observation does not constitute evidence that candidates are being forced to devi- ate from their bliss points (i.e., partial convergence). Since bliss points are likely to vary across districts within the United States, this correlation could simply reflect that candidates’ ideologies may be correlated with those of his fellow citizens in the district. These correlations suggest that the U. S. Congress may achieve a modest degree of “representation” across constituencies, in voting on legislation. Our findings suggest that whatever degree of rep- resentation that may arise, it appears not to occur through Down- sian competition during a general election.
Our findings are generally consistent with existing evidence that candidates exhibit ideological rigidity. Many studies have shown that representatives do not change their positions during their careers, even when their constituency changes. For exam- ple, redistricting does not seem to make much of a difference.30
VII. CONCLUSION
The arguably central notion of the Downsian paradigm is that the pressure of electoral competition constrains candidates’
29. See Krehbiel [2000] and McCarty, Poole, and Rosenthal [2000] for a different interpretation of the Snyder and Groseclose results.
30. See Poole and Rosenthal [1997] for an overview.

848 QUARTERLY JOURNAL OF ECONOMICS
policies. This paper finds that voters do not affect policies in this way, but rather primarily elect policies. The empirical analysis finds that there is little evidence that members of the U. S. House alter their positions (in either direction) in response to a large exogenous change in the probability of winning the election. This result is unlikely to be spurious or a mere statistical artifact, because it is derived using a quasi experiment embedded in the Congressional electoral system that isolates arguably exogenous variation in the probability of winning the election.
The finding that voters primarily elect policies does not nec- essarily constitute evidence regarding the overall “representa- tiveness” of members of the U. S. Congress. For example, one can define “representativeness” as the degree to which a politician represents the preferences of his/her core supporters. Our analy- sis would say nothing about this degree of representativeness. In loose terms, we have no way of knowing whether the Democrat’s (Republican’s) preferred policy exactly coincides with the most- preferred policy choice of the “median” Democratic (Republican) voter. It is possible that electoral competition in the U. S. primary system can work to force each party’s nominee to represent the median voter within each party. Instead, our analysis takes can- didates’ preferred policies—representative or not—as given, and asks whether electoral competition in general elections compels politicians to deviate from those bliss points.
A leading explanation for our findings is that the difficulty of establishing credible commitments to moderate policies is an important real-world phenomenon, and dominates any Downsian convergence effect. This complete divergence hypothesis yields a very strong prediction—that exogenous changes in the relative popularity of the candidates has no impact on the candidates’ positions. Our quasi-experimental evidence is consistent with this bold prediction. Thus, at a minimum, there appears to be empirical support for the assumption that politicians are unable to overcome this credibility problem—an assumption that has been adopted in recent theoretical analyses of representative democracy (e.g., Besley and Coate [1997, 1998]).
Indeed, if one accepts this interpretation, it suggests the need for a more nuanced view than the Downsian perspective of how the American political system translates voters’ preferences into actual policy. When voters merely “elect” instead of “affect” poli- cies, voters do not influence policy choices as much as they are presented with choices. Investigating how citizens can affect

DO VOTERS AFFECT OR ELECT POLICIES? 849 which candidates, and hence which policies appear “on the ballot”
seems to be a useful direction for future empirical research. APPENDIX 1
1. Derivation of Equation (2)
By definition, we have RCt 􏰏 y*(k,􏰓*,c) 􏰁 (x*(k,􏰓*,c) 􏰂
y*(k,􏰓*,c)) Dt. Both x* and y* depend on k, the Nash bargaining
point 􏰓* 􏰠 (0,1) and the bliss point c. Symmetry of the P function
yields a symmetric efficient frontier (around the line y* 􏰏 c 􏰂 DR
x*). Consider that P*t can take on only two values, P*t and P*t DR
(with P*t 􏰐 P*t ), corresponding to which party was the incum- bent party in election t. Assuming symmetry of the incumbency
DR
advantage, this implies P*t 􏰏 1 􏰂 P*t . Given the above sym-
metries, it follows that the Nash bargains are such that
DR
􏰓*(P*t ) 􏰏 1 􏰂 􏰓*(P*t ). Symmetry then implies that x*(k,
DDR 􏰓*(P*t ), c) 􏰂 y*(k, 􏰓*(P*t ), c) will be equal to x*(k, 􏰓*(P*t ),
R
c) 􏰂 y*(k, 􏰓*(P*t ), c), and this gap we can denote, 􏰖1. Defining
DRDR
􏰖0 􏰛 [y*(k, 􏰓*(P*t ), c) 􏰂 y*(k, 􏰓*(P*t ), c)]/[P*t 􏰂 P*t ], we
obtain
RCt 􏰏 constant 􏰁 􏰖0P*t 􏰁 􏰖1Dt 􏰁 εt,
where we have added a translational shift ε of the entire setup, to reflect heterogeneity across districts around the “average” district (the bliss points would be (εt, c 􏰁 εt)).
2. Random Assignment from Close Elections
Consider the vote share production function for all districts: (7) VSt 􏰏 vs􏰔 x*􏰔P*t,εt􏰕, y*􏰔P*t,εt􏰕, εt, et􏰕
ˆ*
(8) 􏰏 vs􏰔Pt,εt􏰕 􏰁 eˆt,
where εt and P*t vary across districts, and εt represents transla- tional shifts in the parties’ bˆliss points. et is an unpredictable component of the vote share. vst is the forecast of the vote share, which depends on information at the time of the election, and eˆt is the forecast error.
By Bayes Rule, we have
(9) 􏰜􏰔P*t, εt􏰢VSt 􏰏 v0􏰕 􏰏 f􏰔v0􏰢P*t,εt􏰕 􏰣 g􏰔P*t,εt􏰕 , h􏰔v0􏰕

850 QUARTERLY JOURNAL OF ECONOMICS
where 􏰜( 􏰣 􏰢 􏰣 ) is the joint density of P*t, εt conditional on VSt. f( 􏰣 􏰢 􏰣 ) is the density of VSt conditional on P*t and εt, g􏰤 is the marginal joint density of P*t, εt and h􏰤 is the marginal density of VSt.
ˆ
If eˆ t has continuous density conditional on vs, then for a fixed
P*t, εt, f(v0􏰢P*t,εt) is continuous in v0. h(v0) is also continuous in v0, which implies that 􏰜(P*t,εt􏰢VSt 􏰏 v0) is continuous in v0. This implies that the distribution of P*t, εt is very nearly the same betweenVSt 􏰏1⁄2 􏰁􏰝andVSt 􏰏1⁄2 􏰂􏰝,for􏰝small.31
APPENDIX 2: DATA
The election data used in our analysis come from ICPSR Study 6311, “Elections to the United States House of Represen- tatives, 1898–1992,” compiled by Gary King [1995]. ICPSR 6311 reports the state, district, incumbency status, and the total votes received by the Republican and Democrat candidates. ICPSR 6311 has the advantage that King distinguishes between “nor- mal” elections that can be characterized as a Republican running against a Democrat, with no important independent candidates, and all other elections labeled as “exceptions.” We drop the 304 “exceptions” between 1946 and 1994 leaving 10,138 observa- tions—roughly 422 observations per congressional session from 1946 to 1992. The 1994 election data were all obtained from the U. S. House of Representatives’ Office of the Clerk’s Web site.
The Americans for Democratic Action interest group ratings were taken from Timothy J. Groseclose’s Web site [2002]. As mentioned in the text, we use “adjusted” ADA scores to be able to compare ratings across sessions of congress. See Groseclose, Lev- itt, and Snyder [1999] for a description of the procedure used to calculate the “adjusted” ADA scores. We replicated the analysis in the paper with the “nominal” ADA scores and found no qualita- tive difference in our findings.
Congressional District Data for 1962–1970 were taken from ICPSR Study 10, “United States Congressional District Data Books.” Congressional District Data for 1972–1980 were taken from ICPSR Study 11, “United States Congressional District Data Book for the Ninety-Third Congress.” Congressional District
31. See DiNardo and Lee [2002] for a similar proof when P*92 and ε have discrete support, and Lee [2003] for conditions under which an RD design can generate random assignment of treatment.

DO VOTERS AFFECT OR ELECT POLICIES? 851
Data for 1982–1990 were taken from ICPSR Study 8710, “Census of Population and Housing, 1980 [United States]: STF3D Con- gressional District-level Extract.” Congressional District Data for 1992 were taken from the 1990 Congressional District level ex- tract. All districts during a given decade were assigned the most recent census value. For example, 1974 congressional districts were assigned the 1970 census values.
Keith T. Poole’s [2002] Web site provided much of the data used in this paper. The non-ADA interest group ratings came from Poole and Daniels’ [1985] interest group rating data which contain 59 interest group ratings scores, each covering some or all of the years between 1959 and 1981. McCarty, Poole, and Rosenthal’s DW-NOMINATE scores and Poole’s rank ordering data were also found at Poole’s Web site. For a description of the NOMINATE procedure, see Poole and Rosenthal [1997], and for a specific description of the DW-NOMINATE scores see McCarty, Poole, and Rosenthal [1997] and for a description of the rank order procedure see Poole [1999].
James Snyder was kind enough to provide us with the Presi- dential Election returns by Congressional District from 1952 to 1996 for a previous draft of this paper.
APPENDIX 3: ADDITIONAL ESTIMATES UNDER UNRESTRICTED HETEROGENEITY
To see how it is possible to introduce heterogeneity in the difference between parties’ positions across districts and legisla- tures, we focus, as before, on districts that experience close elec- tions at time t and assume that the winner in close elections is random.
Imagine that after election t 􏰁 1, the econometrician can divide this set of districts into three groups. The first group (which we call the top group) includes districts that were going to be Democrat in time t 􏰁 1, irrespective of the outcome at time t. This group includes the districts that were (exogenously) as- signed a Republican representative at time t but become Demo- crat at time t 􏰁 1. But it also includes districts that were (exogenously) assigned a Democratic representative at time t, remained Democrat at time t 􏰁 1, and would have been Demo- crat at time t 􏰁 1 even if they had been assigned a Republican winner at time t. In the terminology of Angrist, Imbens, and Rubin [1996], these districts are “always takers,” because irre-

852 QUARTERLY JOURNAL OF ECONOMICS
spective of the quasi-random assignment of DEMt, they are Demo- crat at time t 􏰁 1.
It is crucial to remember that even though the econometrician can identify the districts in this particular group, we maintain the assumption that all agents in the model cannot. If any party knew with certainty that it was going to win in t 􏰁 1, irrespective of the outcome of election t, there would be little incentive to moderate positions. We emphasize that we are not introducing a new as- sumption at this point. The assumption that electoral outcomes are not completely certain (particularly among close elections in t) provides the basis for Alesina’s theoretical framework, and the analysis throughout this paper.
In a parallel fashion, we can identify a “bottom group,” which includes districts that, irrespective of the outcome at time t, become Republican at time t 􏰁 1. This group includes the dis- tricts that were (exogenously) assigned a Democrat representa- tive at time t but are Republican at time t 􏰁 1. It also includes districts that were (exogenously) assigned a Republican represen- tative at time t, remained Republican after time t 􏰁 1, and would have been Republican at time t 􏰁 1 even if they had been assigned to be Democrat at time t. These districts could be called “never takers” [Angrist, Imbens, and Rubin 1996], because irre- spective of the “random assignment” of DEMt, they are Republi- can at time t 􏰁 1. Again, we can only identify these districts ex post, from the point of view of the econometrician.
The remaining districts make up the third and final group, which we call the middle group. In the terminology of Angrist, Imbens, and Rubin [1996] these districts are the “compliers.”
We can think of the parameter 􏰙 that we estimated in our main result section as a weighted average of each group specific 􏰙:
(10) 􏰙 􏰏 Pr1􏰙1 􏰁 Pr2􏰙2 􏰁 Pr3􏰙3,
where 􏰙1 is the 􏰙 for the top group, 􏰙2 is for 􏰙 for the middle group, and 􏰙3 is the 􏰙 for the bottom group; and Pr is the relative size of each group.
If we can identify the three groups, we can provide a test of complete divergence against partial convergence based on the size of the group specific 􏰙s. In particular, we have sharp predic- tions on the size of 􏰙1 and 􏰙3 depending on whether complete divergence or partial convergence is true.

DO VOTERS AFFECT OR ELECT POLICIES? 853
Under complete divergence, we expect the ADA scores of representatives from both the top and bottom group to be unaf- fected by who wins the election at time t. This is because, despite the underlying change in electoral strength, candidates are un- able to credibly promise anything but their extreme bliss points. So it should not matter whether they had a strong or moderate chance of winning the election in t 􏰁 1.
By contrast, under partial convergence, a Democrat victory in t should cause ADA scores to rise for the top and bottom groups. A Democrat victory in t causes an increase in electoral strength for the Democrats for election t 􏰁 1. With greater “bargaining power” the compromise would move toward the Demo- crats’ bliss points.
Essentially, by dividing the districts into these three groups,
we have eliminated the “elect component” when focusing only on
the top and bottom groups. Among the top and bottom groups, the
change in electoral strength will not change the composition of
Democrats and Republicans in office after t 􏰁 1, because we have
deleted the marginal districts (the middle group). Thus, both 􏰙1 D
and 􏰙3 are direct estimates of the “affect component” 􏰖0(P*t􏰁1 􏰂 R
P *t 􏰁 1 ) .
How do we identify these three groups in practice? In most
contexts, it would be impossible to identify these groups of dis- tricts. However, in our context, it is possible because we actually observe the Democratic vote share, which is the underlying index that perfectly determines who wins in t 􏰁 1.
To see how this is done, consider elections t and t 􏰁 1. The “vote production function” is VSt􏰁1 􏰏 vs(x*t􏰁1(P*t􏰁1),y*t􏰁1(P*t􏰁1),et􏰁1), where VSt􏰁1 is the vote share for the Democrats. e is an unpre- dictable component of the vote share, which must be present if electoral outcomes are uncertain. Taking a linear approximation to this function yields (normalizing all coefficients to 1) VSt􏰁1 􏰏
P*t􏰁1 􏰁 et􏰁1.
Consider the districts where the Democrat defeated the in-
cumbent Republican party in t 􏰁 1. These are the districts in R
which VSt􏰁1 􏰏 P*t􏰁1 􏰁 et􏰁1 􏰐 1⁄2 . Next, consider the Democratic incumbents who succeeded in t 􏰁 1 who have vote share VSt􏰁1 􏰏
D
P*t􏰁1 􏰁 et􏰁1 􏰐 1⁄2 . To make this group comparable, we must
select a subset of these districts. That is, we must choose VSt􏰁1 􏰐
1􏰥􏰥
DR
⁄2 􏰁 􏰞, where 􏰞 􏰏 (P*t􏰁1 􏰂 P*t􏰁1). Doing this allows us to focus

854 QUARTERLY JOURNAL OF ECONOMICS R
on those districts such that P*t􏰁1 􏰁 et􏰁1 􏰐 1⁄2 . Intuitively, some of the districts where the Democratic incumbents won in t 􏰁 1 would not have won if the district had been won by a Republican in t. This selection procedure eliminates these “marginal” districts.
Since Dt is essentially randomly assigned (among close elec- tions), we have that E[RCt􏰁1􏰢D0 􏰏 1, VSt􏰁1 􏰐 1⁄2 􏰁 􏰞􏰥] 􏰏
DR
E[x*t􏰁1(P*t􏰁1)􏰢P*t􏰁1 􏰁 et􏰁1 􏰐 1⁄2] and E[RCt􏰁1􏰢D0 􏰏 0, VSt􏰁1 􏰐
RR
1⁄2] 􏰏 E[x*t􏰁1(P*t􏰁1)􏰢P*t􏰁1 􏰁 et􏰁1 􏰐 1⁄2]. Therefore, the difference
DR
in these quantities is E[ x*t􏰁1(P*t􏰁1) 􏰂 x*t􏰁1(P*t􏰁1)􏰢Dt 􏰏 0, Dt􏰁1 􏰏
1]. This is the response of Democrats’ policy choice to differing
DR probabilities of winning t 􏰁 1, P*t􏰁1 and P*t􏰁1.
How do we practically choose this higher cutoff 􏰞􏰥? We observe R
the proportion of the Dt 􏰏 0 group of districts that have P*t􏰁1 􏰁 et􏰁1 􏰐 1⁄2 . This proportion is 􏰟 􏰛 Pr[Dt􏰁1 􏰏 1􏰢Dt 􏰏 0]. Now consider the Dt 􏰏 1 group of districts. If Dt is randomly assigned, then we must choose 􏰞􏰥 so that Pr[VSt􏰁1 􏰐 1⁄2 􏰁 􏰞􏰥􏰢Dt 􏰏 1] 􏰏 􏰟. This can be done since VSt􏰁1 is perfectly observable.
Empirical Results with Heterogeneity. We now describe how to empirically implement the theoretical framework described above. Our empirical strategy is similar to the one used to obtain our main results in subsection V.A. As with our basic results, we include in our analysis only districts that experienced close elec- tions at time t. In these districts we assume again that the winner party at time t is virtually random. The main difference with our previous empirical strategy is that we now allow for a more unrestricted form of heterogeneity across districts and divide our sample of close elections in three groups.
We begin by showing in Panel A in Table V that in the sample of close elections, among districts that were won by Democrat in period t, 72.6 percent remain Democrat in t 􏰁 1, and 27.4 percent are won by Republicans. Among districts that were won by Re- publican in period t, 75.9 percent remain Republican in t 􏰁 1, and 24.1 percent are won by Democrat.
If challengers, who are electorally weaker, are responding to electoral pressures by moderating their positions, we should see that a successful Democrat challenger’s ADA score is less liberal than its comparison group, and that a successful Republican challengers’ ADA score is more liberal than its comparison group.
Take, for example, the top left corner of panel A. It represents

DO VOTERS AFFECT OR ELECT POLICIES? 855
TABLE V
ADDITIONAL ESTIMATES USING ADA SCORES—CLOSE ELECTIONS SAMPLE
Panel A Percent DEMt􏰁1 􏰏 1
Percent DEMt􏰁1 􏰏 0 Total
24.1 75.9 100
(2) (3) (4)
72.6 27.4 100
Difference in average ADAt􏰁1 (col. (2))
􏰂(col.(1)) Obs. (3) (4)
􏰂1.7 224 (3.0)
47.4 441 (1.8)
􏰂4.6 250 (2.3)
Districts, where DEMt 􏰏0 (1)
Districts, where DEMt 􏰏1
Panel B
Districts in top group Districts in
middle group Districts in
bottom group
Average ADAt􏰁1 in districts, where DEMt 􏰏 0 (1)
67.6 (3.0) 18.6 (1.7) 21.1 (2.2)
Average ADAt􏰁1 in districts, where DEMt 􏰏 1 (2)
65.9 (2.9) 66.1 (1.8) 16.5 (2.1)
Standard errors are in parentheses. The unit of observation is a district-congressional session. The sample includes only observations where the Democrat vote share at time t is strictly between 48 percent and 52 percent. Column (1) refers to observations for which the Democrat vote share at time t is strictly between 48 percent and 50 percent. Column (2) refers to observations for which the Democrat vote share at time t is strictly between 50 percent and 52 percent. Time t and t 􏰁 1 refer to congressional sessions. Higher ADA scores correspond to more liberal roll-call voting records.
Democrats who were able to win the district even though it was held by a Republican. We would like to compare the voting score of those districts with the voting score of otherwise comparable districts where the Democrat party is the incumbent party, and the candidate from the Democrat party was elected at time t.
The empirical challenge is to identify the correct counterfac- tual. The methodology described above provides a solution and leads us to split the districts into three groups, as shown in Panel B. We empirically identify the three groups as follows. As before, we consider only the 915 observations that experience close elec- tions at time t. We include in the top group of districts where a Republican (barely) won at time t but a Democrat regained the district at time t 􏰁 1 as well as the districts with the largest Democrat vote shares at time t 􏰁 1, among the districts where a Democrat won at time t. Specifically, we included the top 24.1

856 QUARTERLY JOURNAL OF ECONOMICS
percent districts with the largest Democrat vote share among the districts where a Democrat won at time t. There are 224 such districts.
Similarly, we include in the bottom group districts where a Democrat (barely) won at time t but a Republican regained the district at time t 􏰁 1, as well as districts with the largest Republican vote shares at time t 􏰁 1, among the districts where a Republican won at time t. Specifically, we included the 27.4 percent districts with the largest Republican vote share, since among all the districts where the Republican lost at time t, the Republican candidate won at time t 􏰁 1 in 27.4 percent of the cases. There are 250 such districts. Finally, we include in the middle group the remaining districts. There are 441 such districts.
Because of the way we constructed the three groups, none of the top group districts in column (1) has a representative in period t 􏰁 1 who is the same representative in period t. On the contrary, 91 percent of the top group districts in column (2) has a representative in period t 􏰁 1 who is the same representative in period t. The corresponding figures for the bottom group are, respectively, 0 and 96 percent. The corresponding figures for the middle group are 90 percent and 89 percent.
Panel B shows the average voting scores, for each of the three groups, by the winning party at time t. In column (1) we report the average ADA at time t 􏰁 1 for districts that were Republican at time t for each of the three groups. In column (2) we report the average ADA score at time t 􏰁 1 for districts that were Democrat at time t. In column (3) we report the difference.
Consider the top and the bottom group. Consistent with the hypothesis of complete divergence, there is virtually no difference in the voting scores of the two sets of districts. Consider the middle group. As expected, there is a large difference in the voting scores of the two sets of districts.
Overall, the results in Table V lead us to conclude that our earlier results are robust to a more general framework that allows for unrestricted heterogeneity in the differences in pol- icy positions across districts and over time. Finally, we have estimated similar models using alternative measures of roll- call behavior (DW-Nominate scores and percentage of vote consistent with Democrat leadership). In results available

DO VOTERS AFFECT OR ELECT POLICIES? 857 from the authors, we find similar results: virtually no response
of voting records to changes in electoral strength.
UNIVERSITY OF CALIFORNIA, BERKELEY, AND NATIONAL BUREAU OF ECONOMIC RESEARCH
UNIVERSITY OF CALIFORNIA, BERKELEY, AND NATIONAL BUREAU OF ECONOMIC RESEARCH
UNIVERSITY OF CALIFORNIA, BERKELEY
REFERENCES
Alesina, Alberto, “Credibility and Policy Convergence in a Two-Party System with Rational Voters,” American Economic Review, LXXVIII (1988), 796–805. Alesina, Alberto, and Stephen E. Spear, “An Overlapping Generations Model of
Electoral Competition,” Journal of Public Economics, XXXVII (1988), 359–379. Angrist, Joshua, Guido Imbens, and Donald Rubin, “Identification of Causal Effects Using Instrumental Variables,” Journal of the American Statistical
Association, XCI (1996), 444–445.
Banks, Jeffrey S., and D. Roderick Kiewiet, “Explaining Patterns of Candidate
Competition in Congressional Elections,” American Journal of Political Sci-
ence, XXXIII (1989), 997–1015.
Bernhardt, Daniel, and Daniel Ingberman, “Candidate Reputation and the In-
cumbency Effect,” Journal of Public Economics, XXVII (1985), 47–67. Bender, Bruce A., “The Influence of Ideology on Congressional Voting,” Economic
Inquiry, XXIX (1991), 416–428.
Besley, Timothy, and Stephen Coate, “An Economic Model of Representative
Democracy,” Quarterly Journal of Economics, CXII (1997), 85–114.
Besley, Timothy, and Stephen Coate, “Sources of Inefficiency in a Representative Democracy: A Dynamic Analysis,” American Economic Review, LXXXVIII
(1998), 139–156.
Besley, Timothy, and Anne Case, “Political Institutions and Policy Outcomes,”
Journal of Economic Literature, XLI (2003), 7–73.
Calvert, Randall L., “Robustness of the Multidimensional Voting Model: Candi-
date Motivations, Uncertainty, and Convergence,” American Journal of Po-
litical Science, XXIX (1985), 69–95.
Canes-Worne, Brandace, David W. Brady, and John F. Cogan, “Out of Step, Out
of Office: Electoral Accountability and House Members’ Voting,” American
Political Science Review, XCVI (2002), 127–140.
DiNardo, John, and David S. Lee, “The Impact of Unionization on Establishment
Closure: A Regression Discontinuity Analysis of Representation Elections,”
NBER Working Paper No. 8993, June 2002.
Downs, Anthony, An Economic Theory of Democracy (New York, NY: Harper and
Row, 1957).
Fiorina, Morris P., “Whatever Happened to the Median Voter?” paper prepared for
the MIT conference on Parties and Congress, Cambridge, MA, October 2,
1999.
Groseclose, Timothy J., ADA SCORES, 1947–1999 [Computer File]. http://
faculty-gsb.stanford.edu/groseclose/archive.htm 12 February 2002. Inter-university Consortium for Political and Social Research, and Carroll Mc- Kibbin, ROSTER OF UNITED STATES CONGRESSIONAL OFFICEHOLD- ERS AND BIOGRAPHICAL CHARACTERISTICS OF MEMBERS OF THE UNITED STATES CONGRESS, 1789–1996: MERGED DATA [Computer file], 10th ICPSR ed., Ann Arbor, MI, Inter-university Consortium for Politi-
cal and Social Research [producer and distributor], 1997.
Inter-university Consortium for Political and Social Research, UNITED STATES CONGRESSIONAL ROLL CALL VOTING RECORDS, 1789–1990 [Parts 1–202] [Computer file], ICPSR ed., Ann Arbor, MI, Inter-university Consor-
tium for Political and Social Research [producer and distributor], 2001.

858 QUARTERLY JOURNAL OF ECONOMICS
Inter-university Consortium for Political and Social Research, and Congressional Quarterly, Inc., UNITED STATES CONGRESSIONAL ROLL CALL VOTING RECORDS, 1789 –1994 [Parts 203–206] [Computer file], ICPSR ed., Ann Arbor, MI, Inter-university Consortium for Political and Social Research/Washington, DC, Congressional Quarterly, Inc. [producers], 1997, Ann Arbor, MI: Inter- university Consortium for Political and Social Research [distributor], 1997.
Inter-university Consortium for Political and Social Research, and Congressional Quarterly, Inc., UNITED STATES CONGRESSIONAL ROLL CALL VOTING RECORDS, 1789 –1996 [Parts 207–208] [Computer file], ICPSR ed., Ann Arbor, MI, Inter-university Consortium for Political and Social Research/Washington, DC, Congressional Quarterly, Inc. [producers], 1998, Ann Arbor, MI: Inter- university Consortium for Political and Social Research [distributor], 1998.
King, Gary, ELECTIONS TO THE UNITED STATES HOUSE OF REPRESEN- TATIVES, 1898 –1992 [Computer file]. ICPSR version, Cambridge, MA, Gary King [producer], 1994, Ann Arbor, MI, Inter-university Consortium for Po- litical and Social Research [distributor], 1995.
Krehbiel, Keith, “The Coefficient of Party Influence,” Stanford Graduate School of Business Research Paper No. 1646, 2000.
Lee, David S., “The Electoral Advantage to Incumbency and Voters’ Valuation of Politicians’ Experience: A Regression Discontinuity Analysis of Close Elec- tions,” University of California, Berkeley Center for Labor Economics, Work- ing Paper No. 31, April 2001.
——, “Randomized Experiments from Non-Random Selection in U. S. House Elections,” University of California, Berkeley manuscript, September 2003.
Levitt, Steven D., “How Do Senators Vote? Disentangling the Role of Voter Preferences, Party Affiliation, and Senator Ideology,” American Economic Review, LXXXVI (1996), 425–441.
Levitt, Steven D., Timothy Groseclose, and James M. Snyder, Jr., “Comparing Interest Group Scores across Time and Chambers: Adjusted ADA Scores for the U. S. Congress,” American Political Science Review, XCIII (1999), 33–50.
Lott, John, and Michael Davis, “A Critical Review and Extension of the Political Shirking Literature,” Public Choice, LXXIV (1992), 461–484.
McArthur, John R., and Stephen V. Marks, “Constituent Interests versus Legis- lator Ideology: The Role of Political Opportunity Cost,” Economic Inquiry, XXVI (1988), 461–470.
McCarty, Nolan M., Keith T. Poole, and Howard Rosenthal, Income Redistribution and the Realignment of American Politics (Washington, DC: The AEI Press (1997).
McCarty, Nolan, Keith T. Poole, and Howard Rosenthal, “The Hunt for Party Discipline in Congress,” American Political Science Review, XCV (2000), 673– 687.
Miller, Warren E., and Donald E. Stokes, “Constituency Influence in Congress,” American Political Science Review, LVII (1963), 45–56.
Osborne, Martin J., “Spatial Models of Political Competition under Plurality Rule: A Survey of Some Explanations of the Number of Candidates and the Position they Take,” Canadian Journal of Economics, XXVIII (1995), 261–301.
Persson, Torsten, and Guido Tabellini, Political Economics (Cambridge, MA: MIT Press, 2000).
Poole, Keith T., “Changing Minds? Not in Congress!” GSIA WP No. 1997-22, in Changing Minds, Michael Munger and Bryan Roberts, eds. (New York: Cam- bridge University Press, 1999).
——, DW-NOMINATE SCORES FOR HOUSE, 1st–106th SESSION OF CON- GRESS [Computer File]. http://voteview.uh.edu/default_nomdata.htm, 2002. ——, INTEREST GROUP RATINGS, 1959–1981 [Computer File]. http://
voteview.uh.edu/dwnominate.htm, 2002.
——, RANK ORDER FOR THE HOUSE, 1947–1996 [Computer File]. http://
voteview.uh.edu/default_nomdata.htm, 2002.
Poole, Keith T., and R. Steven Daniels, “Ideology, Party and Voting in the U. S.
Congress 1959–1980,” American Political Science Review, LXXIX (1985),
373–399.
Poole, Keith T., and Howard Rosenthal, “The Polarization of American Politics,”
Journal of Politics, XLVI (1984), 102–131.

DO VOTERS AFFECT OR ELECT POLICIES? 859
Poole, Keith T., and Howard Rosenthal, “A Spatial Model for Roll Call Analysis,” American Journal of Political Science, XXIX (1985), 331–347.
Poole, Keith T., and Howard Rosenthal, “Patterns of Congressional Voting,” American Journal of Political Science, XXXV (1991), 331–347.
Poole, Keith T., and Howard Rosenthal, Congress: A Political-Economic History of Roll Call Voting (Oxford, UK: Oxford University Press, 1997).
Poole, Keith T., and Howard Rosenthal, “D-Nominate after 10 Years,” Legislative Studies Quarterly, XXVI (2001), 5–30.
Snyder, James, and Timothy Groseclose, “Estimating Party Influence in Congres- sional Roll-Call Voting,” American Journal of Political Science, XLIV (2000), 187–205.
James Snyder, and Michael Ting, “Party Labels, Roll-Call Votes, and Elections,” mimeo, 2001a.
James Snyder, and Michael Ting, “An Information Rational for Political Parties,” mimeo, 2001b.
U. S. Department of Commerce, Bureau of the Census, CENSUS OF POPULA- TION AND HOUSING, 1980 [UNITED STATES]: SUMMARY TAPE FILE 3D CONGRESSIONAL DISTRICT-LEVEL EXTRACT [Computer file]. Mil- waukee, WI: John McAdams, Marquette University [producer], 198?. Ann Arbor, MI: Inter-university Consortium for Political and Social Research [distributor], 1987.
——, CENSUS OF POPULATION AND HOUSING, 1990 [UNITED STATES]: SUMMARY TAPE FILE 3D CONGRESSIONAL DISTRICTS OF THE UNITED STATES, 103RD CONGRESS [Computer File]. Washington, DC, 1993.
——, UNITED STATES CONGRESSIONAL DISTRICT DATA BOOKS, 1961– 1965 [Computer file]. ICPSR ed., Ann Arbor, MI: Inter-university Consortium for Political and Social Research [producer and distributor], 196?.
——, UNITED STATES CONGRESSIONAL DISTRICT DATA BOOK FOR THE NINETY-THIRD CONGRESS, 1973 [Computer file]. ICPSR ed., Ann Arbor, MI: Inter-university Consortium for Political and Social Research [producer and distributor], 1999.
Wittman, Donald, “Candidate Motivation: A Synthesis of Alternative Theories,” American Political Science Review, LXXVII (1983), 142–157.