Hierarchical Accountability in Government

Transcription

1 Hierarchical Accountability in Government Razvan Vlaicu y Alexander Whalley z May 2014 Abstract This paper studies a setting where a relatively uninformed voter holds a policymaker accountable through an informed intermediary. In equilibrium the voter uses the intermediary to insulate the policymaker from pandering incentives if the voter s policy expertise is low or the policymaker s congruence is high. The voter can thus exploit the bene ts of bureaucratic expertise without forfeiting electoral responsiveness. We examine the model s predictions using U.S. city-level data and nd that hierarchically-accountable managers reduce popular city employment, and adjust it more exibly, than electorally-accountable mayors. The estimated policy e ects vary with informational and political determinants of policymaker incentives and are robust to instrumentation by precipitation shocks that in uenced early 20th century manager government adoptions for reasons obsolete today. JEL Classi cation: D72, D73, H70. Keywords: pandering, accountability, incentives, city manager. We would like to thank Jim Davies, Allan Drazen, Fernando Ferreira, Price Fishback, Lisa George, Gabriele Gratton, Alex Hirsch, Shawn Kantor, Ethan Kaplan, Ken Shotts, Gergely Ujhelyi, Richard Van Weelden, Emanuel Vespa, Joel Waldfogel, John Wallis, and seminar participants at IIES, New York University, Stanford University, Stockholm University, University of Chicago, University of Houston/Rice University, University of Illinois at Urbana, University of Pittsburgh, Canadian Public Economics Conference, Econometric Society NAWM, and Princeton Political Economy Conference. Erin Moody, Krystal Tapia, Michael Wall, and Marko Zivanovic provided excellent research assistance. y University of Maryland, Department of Economics, and Department of Government and Politics, 3114 Tydings Hall, College Park, MD 20742, USA. vlaicu@econ.umd.edu. z University of California - Merced, School of Social Sciences, Humanities & Arts, 5200 North Lake Road, Merced, CA 95343, USA; and NBER, 1050 Massachusetts Ave., Cambridge, MA 02138, USA. awhalley@ucmerced.edu. Electronic copy available at:

2 1 Introduction The accountability of government o cials to the general public is a fundamental principle of democratic governance. Yet, this principle s most direct manifestation electoral accountability while a useful safeguard may nevertheless introduce distortions in the policymaking process. Having to periodically face a public whose policy expertise is limited gives electorally-accountable o cials incentives to disregard their private expertise and adopt popular policies not necessarily in the public interest by engaging in pandering (Canes-Wrone, Herron, and Shotts 2001) or electoral manipulation (Rogo 1990). Delegating policymaking to unaccountable technocrats allows expertise to be used without fear of electoral retribution but runs the risk that technocrats may pursue private goals that can disconnect policymaking from the public interest (Maskin and Tirole 2004). An important question then is how to resolve this tension between expertise and responsiveness, that is, how to allow public o cials expertise to be expressed in policymaking while also preserving accountability to the general public. In this paper we focus on a setting where the policymaker is accountable to the voter through an intermediary. Policymakers such as prime ministers, city managers, and school district superintendents cannot be directly removed by voters. However, they remain accountable to the voter in the sense that they can be replaced at will by the intermediary, e.g., legislature, city council, school board, and the intermediary in turn is electorally accountable to the voter. Following contract theory terminology we refer to this type of principal-intermediary-agent relationship as hierarchical accountability. 1 At rst sight, holding a policymaker accountable through an intermediary seems to disconnect policymaker choices from voter preferences. If the voter is fully informed he can, nevertheless, exploit the intermediary s reelection motivation to control policymaker moral hazard just as e ectively as the voter could directly (Persson, Roland, and Tabellini 1997). 2 In an incomplete information environment, however, the voter also faces an adverse selection problem which weakens electoral accountability by producing pandering incentives. Even if the voter s best interest is to limit policymaker pandering a commitment problem renders him unable to do so. Ex ante an uninformed voter may prefer to insulate the better-informed policymaker. Ex post, however, he is better o replacing an unpopular policymaker because 1 "If the manager is not responsive to the governing body, it has the authority to terminate the manager at any time." (ICMA 2007, p. 2). About half of U.S. cities are run by city managers. Other forms of indirect accountability grant the policymaker a xed term, i.e., top regulators, central bank governors. 2 One caveat is possible policymaker-intermediary (executive-legislature) collusion. In that case, the executive would have more discretion to engage in rent-seeking. 2 Electronic copy available at:

3 unpopular policies signal the policymaker has di erent preferences. 3 When accountability to an imperfectly-informed voter distorts the policymaker s incentives in this way, delegating policymaker accountability to an informed intermediary seems justi able. The case for delegation is less clear-cut, however, if the intermediary s preferences may also di er from the voter s. In that case the voter needs to provide incentives for the intermediary to act in the voter s interest. Under what conditions can an informed intermediary help the voter insulate the policymaker from pandering incentives? To understand behavior in this setting we rst develop a hierarchical agency model (voterintermediary-policymaker model) where the voter is uncertain about the optimal policy but leans toward a "popular" policy. The intermediary is a political expert, i.e., informed about the policymaker s type, and the policymaker is a policy expert, i.e., informed about the optimal policy. The model s key insight is that the voter can credibly use the intermediary to insulate the policymaker from popular pressure when pandering is relatively detrimental to the voter. That happens when the voter s policy uncertainty is high or the policymaker s preferences are likely to be aligned with the voter s. In these cases the voter is ex ante better o retaining an unpopular policymaker and the intermediary allows him to commit to do so ex post, since retaining an unpopular policymaker signals the intermediary s preferences are aligned with the voter s. Hierarchical accountability thus o ers the voter the exibility to allow the exercise of policymaking expertise without forfeiting electoral responsiveness. Voters can switch between erring on the side of expertise and erring on the side of responsiveness when their informational and political environment changes. 4 In our empirical application we study policymaking by U.S. city managers, hierarchicallyaccountable o cials with the same major policy responsibilities as electorally-accountable mayors, i.e., writing the budget and hiring personnel. While manager government provides a natural measure of hierarchical accountability how to measure pandering behavior is less clear. We propose that pandering behavior can be identi ed empirically in policy issues that satisfy three conditions: (i) is a primary (not secondary) policy issue over which the policymaker has jurisdictional control, (ii) the optimal policy is state-contingent, and (iii) the public has a clear stance on this issue. We argue that police o cer employment satis es these requirements as (i) crime has consistently ranked among the top two local policy issues 3 Besley and Smart (2007) and Smart and Sturm (2013) notice the commitment issue in the context of scal rules and term limits, respectively. 4 In contract theory and corporate nance hierarchical agency models study optimal incentives through wage contracts (e.g., Strausz 1997, Park 2000). Our main result echoes the nding in this literature that under asymmetric information an intermediary allows the principal to commit to a broader range of incentive structures. 3

4 in Gallup surveys of local attitudes since 1959 (see Gallup 2000) and city executives have juristictional control over public safety, a substantial local budget item, (ii) the probability of crime uctuates with economic and social conditions, changing the optimal policy response, and (iii) due to the salience of public safety, demand for police o cers by relatively uninformed voters can be expected to be high even when the probability of crime is low; in contrast, civilian police employment, e.g., administrators, dispatchers, or other city employee categories, should not elicit such a clear popular preference. Across a number of speci cations we nd that on average managers employ 8-14% fewer police o cers per capita than mayors but a comparable number of police civilians and non-police employees per capita. A central challenge to estimating institutional e ects is that institutions may be endogenous to policymaking, for instance through unobserved voter preferences (Aghion, Alesina, and Trebbi 2004). To address potential endogeneity in accountability form we propose an instrument for manager government. The instrument is based on the observation that before the 1936 Flood Control Act transferred ood prevention from local governments to the Army Corps of Engineers cities often responded to ood-related infrastructure crises by adopting manager government because it facilitated the ascension of engineers into the top executive o ce. We document that pre-1936 precipitation shocks in uenced early switches to manager government for reasons obsolete today and show that the police employment patterns noted above also appear in this IV setting. 5 We further explore the theory model s incentive mechanisms and nd that policy volatility is higher in manager governments as insulated managers can act on their expertise to adjust the policy to the stochastic state. We also propose measures of voter beliefs and policymaker congruence. We nd that the manager-mayor o cer employment di erential decreases in cities a ected by the crack epidemic and increases in cities with increasingly competitive elections. It is also more pronounced in election years when the theory model predicts incentives should be sharper. We note that these patterns cannot be fully explained by alternative mechanisms, such as pure patronage motivations or policymaker type selection. 6 Our paper relates to the literature on the career concerns of expert policymakers. Maskin and Tirole (2004) show how relying on unaccountable technocrats, e.g., unelected bureaucrats, tenured judges, eliminates pandering incentives but makes removing noncongruent policymakers harder; thus there is a tradeo between expertise and responsiveness in pol- 5 To our knowledge this is the rst instrument for manager government in the literature, if we exclude Baqir s (2002) use of lags of city institutions as instruments for current city institutions. 6 The existing evidence for electoral cycles focuses on scal policy, is mostly at the national or state level, and is generally weaker for developed countries (Drazen 2000). 4

5 icymaking. Similarly, media providers improve the monitoring of public o cials, but face incentives to be yes men themselves (Ashworth and Shotts 2010); term limits improve political selection, but also allow noncongruent o cials to implement their preferred policies (Smart and Sturm 2013); candidate competition reduces pandering, but encourages anti-pandering (Kartik, Squintani, and Tinn 2013). Our theory shows, by contrast, that hierarchical accountability provides the voter with the exibility to choose between insulating congruent policymakers or providing pandering incentives to noncongruent policymakers. The voter can thus enjoy the bene ts of bureaucratic expertise without forfeiting electoral responsiveness. Our contribution to the empirical literature is two-fold. First, we provide an empirical measure of pandering behavior grounded in the logic of the theoretical pandering literature and use it to quantify how pandering incentives respond to informational and political factors. Second, we contribute to the literature on U.S. city managers which has largely focused on di erences in public spending from city mayors. Coate and Knight (2011) survey this literature, note that results have been mixed, and provide new evidence that managers outspend mayors. 7 A smaller literature has looked at other policy outcomes: Managers are more likely to privatize city services (Levin and Tadelis 2010), more fexibly adjust tax revenues (Vlaicu and Whalley 2011), and reduce full-time city employment (Enikopolov 2012). In contrast, our paper documents di erences in policymaking incentives for popular and neutral policies. While the previous literature has treated city government form as exogenous, we o er an instrumental variable strategy that addresses the important issue of government form endogeneity. The rest of the paper is organized as follows. Section 2 presents the model. Section 3 provides historical background and introduces the data. Section 4 presents the empirical strategy and results. Section 5 concludes. 2 Theory To gain insight into how hierarchical accountability operates we build on Maskin and Tirole (2004) and Smart and Sturm (2013) by introducing an asymmetrically informed intermediary in the voter-policymaker relationship. For easy linking with our empirical application to police employment this section uses terminology speci c to that particular policy domain. The 7 Coate and Knight s (2011) citizen-candidate model attributes di erences in spending not to incentives, but to voters electing di erent councilmember types. Other papers (Baqir 2002, MacDonald 2008) found no spending di erential between managers and mayors. 5

6 model, however, applies in any setting featuring a hierarchically-accountable policymaker with asymmetric expertise. Model. At time t a policy choice x t needs to be made from the set of policy alternatives f0; 1g: For instance, the policy issue can be police employment, in which case x t = 1 can stand for "high police" and x t = 0 for "low police." The policy s e ect depends on the state of the world s t 2 f0; 1g prevailing in that period. In the illustration state s t = 1 can represent "high probability of crime" and state s t = 0 can represent "low probability of crime." The state is i.i.d. across periods. Let = P fs t = 1g and assume > There are three kinds of players: voter, intermediary, and policymaker. The voter gets a unit of payo when the policy matches the state, zero otherwise: v(x t ; s t ) = 1 fx t = s t g. Notice that based on the crime probability prior the voter prefers policy x t = 1. Using Maskin and Tirole s (2004) terminology, we say that "high police" is the popular policy. 9 The voter delegates policymaking to a policymaker (P ) and delegates policymaker accountability to an intermediary (I). These two agents are both policy-motivated and o cemotivated. Their policy motivation depends on their type, congruent or dissonant: P t ; I t 2 fc; Dg. A congruent agent has the same preferences as the voter: u(x t ; s t jc) = 1 fx t = s t g. A dissonant agent has preferences opposite to the voter s: u(x t ; s t jd) = 1 fx t = 1 s t g. 10 The timing of the in nite-horizon game is as follows. (I) Every period the policymaker observes the state s t 2 f0; 1g and chooses between the unpopular/popular policy: x t 2 f0; 1g: (II) Every period the intermediary observes policy and policymaker type (x t ; P t ) and decides whether to replace/retain the policymaker: y t 2 f0; 1g: (III) Every other period the voter, having observed two periods of choices [(x t 1 ; y t 1 ); (x t ; y t )] ; but not types P t ; I t or the current state s t ; decides to replace/reelect the intermediary: z t 2 f0; 1g: An agent exiting at t is succeeded by an agent (challenger) whose type is a new draw from the type distribution. An exiting agent cannot run for o ce again Gallup s annual Crime Survey asks the question "Is there more crime in your area than there was a year ago, or less?" Since the survey started in 1973 the percentage of respondents who say "More" exceeded those that say "Less" except for three years 1998, 2000, 2001 (see Gallup 2010). Assuming the state is i.i.d. across periods means that past policy choices do not a ect the voter s current crime probability prior. This seems to be a good approximation if police have only a short-term e ect on committed crime (actual crime), while the probability of crime (latent crime) is driven by more fundamental forces like economic inequality, economic growth, and race relations. 9 Stucky (2005) reviews the literature on how the public s crime probability prior (also known as "fear of crime" in the criminology literature) a ects popular preference for police. 10 The analysis would be similar if we assumed u(x t ; s t jd) = 1 fx t = 0g. 11 A key feature of hierarchical accountability is that the policymaker can be replaced between elections. A city manager, for example, serves at the city council s pleasure. His "job tenure is only secure until the next council meeting" (Stillman 1977, p. 664). The model captures this feature by assuming that the 6

7 For expositional simplicity we assume that reelection incentives a ect only election-period behavior (cf. Rogo 1990, Shi and Svensson 2006). Let ( j t) t=1;2;::: be sequences of i.i.d. preference shocks, for j = P; I where j t 2 fc; Dg and P j t = C = j is the congruence probability. Assume j > 1 2 :12 An agent s type is either last period s or the current period s preference shock: j t 2 j t 1; j t and P j t = j t 1 = : Assume 0 < < 1: These assumptions have two implications: First, an agent s type is correlated over time between two adjacent periods, but uncorrelated between two non-adjacent periods. In particular, an agent s type at t + 1 (non-election period) is correlated with his type at t (election period), but uncorrelated with his type at t 1 (non-election period). 13 Second, although initially policymaker and intermediary types are independent, correlation between their types can nevertheless develop through the intermediary s equilibrium retention decisions. The policymaker cares about policies he himself chooses. The intermediary cares about the choices of the policymaker he is keeping in o ce. Formally, lifetime utility for a policymaker at the beginning of an electoral term, i.e., at t 1; is u(x t 1 ; s t 1 j P t 1) + P 1 i=t i t+1 u(x i ; s i j P i ) Q i =t y 1; for an intermediary it is [u(x t 1 ; s t 1 j I t 1)+u(x t ; s t j I t )]+ P 1 i=t 2i 2t+2 [u(x 2i t+1 ; s 2i t+1 j I 2i t+1) + u(x 2i t+2 ; s 2i t+2 j I 2i t+2)] Q i =t z 2 t: Here is a time discount factor, with 0 < < 1: As is standard our policymaker is a policy expert who knows whether the popular policy is optimal (the state s t ). The intermediary, on the other hand, is a political expert, in the sense of Gailmard and Jenkins (2009), who knows the policymaker s preferences (his type P t ). This assumption seems reasonable in settings where the intermediary has frequent contact with the policymaker. 14 The voter, on the other hand, knows neither the popular policy s optimality (moral hazard) nor the policymaker s preferences (adverse selection). Thus, in this model hierarchical accountability features two forms of the classic delegation tradeo : by delegating policymaking to a policy expert the voter may bene t from more informed policy choices but may su er from the policymaker s dissonance; by delegating policymaker accountability to a political expert the voter may bene t from more informed retention decisions but may su er from the intermediary s dissonance. We denote voter posterior beliefs about agent types by j t = P j t = CjX t ; Y t ; where j = P; I and (X t ; Y t ) = (x ; y ) t =1 is the public history at time t. We refer to posterior intermediary can replace the policymaker in the middle of an electoral term. 12 The parameter j can be interpreted as the fraction of voters favoring state-matching policies. Then, an entering agent can be thought of as a random draw from the electorate. 13 The type persistence probability is Pf j t = j t 1 g = Pfj t = j t 1 ; j t 1 = j t 1g = (1 ) ; for any t: 14 That is probably the case for a city council since the policymaker they choose, the city manager, has to attend all city council meetings and sometimes committee meetings. 7

8 beliefs about agent type as the agent s "reputation." Equilibrium. The game described above is a principal-intermediary-agent model with incomplete information. We solve for its Perfect Bayesian Equilibrium, namely: (i) policymaker, intermediary, and voter strategies (x t ; y t ; z t ) that are sequentially rational; and (ii) voter beliefs P t ; I t that are consistent with agent strategies. We add the requirement that voter strategies be robust to a small fraction of non-strategic agents implying that in a pooling equilibrium both on- and o -equilibrium path beliefs are derived from sincere, i.e., preference-based, behavior. 15 We restrict attention to pure strategies. Throughout this section we assume that the discount factor is large enough that both agent types seek to remain in o ce; the equilibrium will thus characterize agents extrinsic incentives. 16 Since the voter can only replace the policymaker through the intermediary, his problem is to reelect that intermediary that seems more in tune with the voter s preferences. Lemma 1 in Appendix A.1 shows that following a two-period electoral term [t 1; t], after the voter has observed choices [(x t 1 ; y t 1 ); (x t ; y t )], the incumbent intermediary s public reputation I t increases whenever ^ I t (x t ; y t ) P I t = Cj (x t ; y t ) ; I t = I t I : In words, the voter evaluates congruence based on election-period behavior (x t ; y t ) under the assumption that the intermediary s type is based on current preferences. Since the voter s reelection decision at t depends only on (x t ; y t ), the intermediary at t 1 is free to retain only policymakers of his own type; this makes the intermediary s type in non-election-periods matter to the voter. The following proposition states the implications of this voter calculation for policymaker incentives. For contrast we also include the case of no intermediary, i.e., electoral accountability, studied by Maskin and Tirole (2004). Proposition 1 For popular policy issues ( > 1=2) hierarchical accountability either insulates the policymaker pre-election, if P > ; or creates policymaker pandering incentives pre-election, if P < ; electoral accountability always creates pre-election pandering incentives for the policymaker. Agents follow their preferences post-election under both accountability forms; in particular, an intermediary retains only a policymaker of his own type. Proof. See Appendix Section A.1. Under electoral accountability the voter uses the policy signal x t policymaker s type. to learn about the Because > 1=2 a popular policy is more likely optimal thus the choice of a congruent policymaker. Therefore a popular policy is a signal of congruence: ^ P t (x t = 1) = P P +(1 P )(1 ) > P > ^ P t (x t = 0) = P (1 ). The voter s best P (1 )+(1 P ) 15 See Maskin and Tirole (2004) for application of this re nement in a related setting. 16 In Appendix A.3 we also solve the case where agents have no career concerns and instead follow their preferences; that equilibrium captures behavior driven by selection. 8

9 response then is to reelect only after the popular policy, giving the policymaker incentives to pander, i.e., choose the popular policy regardless of its optimality. Hierarchical accountability provides the voter an additional signal: the retention signal contained in the intermediary s retention decision y t : Because the intermediary prefers to retain only policymakers of his own type, retention is a signal of intermediary congruence if the policymaker is more likely congruent. Thus, because for a popular policymaker that is the case: ^ P t (x t = 1) > P > 1=2; retaining a popular policymaker signals intermediary congruence: ^ I t (x t = 1; y t = 1) = I P > I P +(1 I )(1 P )(1 ) I ; and replacing a popular policymaker signals intermediary dissonance: ^ I t (x t = 1; y t = 0) = I (1 P )(1 ) < I (1 P )(1 )+(1 I ) P I : For an unpopular policymaker ^ P t (x t = 0) > 1 i 2 P >. That is, he may still be believed by voters to be more likely congruent if: (a) the strength of the policy signal is low, or (b) his initial reputation P was high. In these cases retaining an unpopular policymaker signals intermediary congruence: ^ I t (x t = 0; y t = 1) = I P (1 ) I P (1 )+(1 I )(1 P ) (> I i P > ), and replacing him signals intermediary dissonance: ^ I t (x t = 0; y t = 0) = I (1 P ) I (1 P )+(1 I ) P (1 ) (> I i P < ). In other words, when the retention signal dominates the policy signal the voter gives the intermediary incentives to retain the policymaker independent of his policy choices, in e ect insulating the policymaker. Conversely, if the policy signal dominates the retention signal, then retaining an unpopular policymaker signals a dissonant intermediary. In this case the voter gives the intermediary incentives to retain only popular policymakers, which in turn gives the policymaker the incentive to pander. 17 Note that ex ante voter welfare under insulation is P whereas under pandering it is : The equilibrium therefore implements the policymaking behavior that maximizes voter ex ante welfare: insulation if P > ; and pandering if P <. In that sense the hierarchical structure o ers the voter the exibility to optimally choose the types of policymaker incentives that best serve the voter s interest. Neutral Issues. One may argue that another way in which hierarchically-accountable policymakers di er from electorally-accountable policymakers is that they have weaker motivations to deliver political patronage to various constituencies. While patronage could potentially a ect popular policies, such as public safety or low taxes, it also typically in- uences neutral policies such as in-kind transfers or city jobs that may not elicit a clear popular preference. This can be captured by setting = 1=2; which makes the voter ex ante 17 These two types of intermediary behaviors are consistent with case studies of U.S. city councils that distinguish between council strategies that insulate the manager, termed "blind faith," versus council strategies that transmit popular preferences, termed "political." See Stillman (1977). 9

10 indi erent between policies x t = 1 and x t = 0: Proposition 2 For neutral policy issues ( = 1=2) policymaking outcomes do not vary with the accountability form in election periods or non-election periods. Proof. See Appendix Section A.1. Intuitively, a balanced state prior silences the policy signal. Now the voter can credibly insulate an electorally-accountable policymaker since the voter cannot link a certain policy to a particular type. Under hierarchical accountability the equilibrium is also insulating because the retention signal is always stronger than the policy signal, P > = 1=2. These two e ects imply that pandering incentives are absent on neutral issues for both accountability forms. Discussion. Propositions 1 and 2 are robust to a setup where the voter factors into his reelection decision both election-period and non-election-period policy (cf. Martinez 2009); see Appendix Section A.2 for a solution to the model under this relaxed assumption. 18 The propositions show how the incentives of hierarchically-accountable policymakers may depend on their informational and political environment; observationally, this creates a negative election-period policy di erential in popular policies between hierarchical and electoral accountability when P >. The propositions also imply that in this case policy outcomes are more volatile under hierarchical accountability because it gives the policymaker more freedom to adjust the policy to the stochastic state. The model can also be solved under the assumption that the accountability form s primary role is not to provide incentives but to select certain types of policymakers; see Appendix Section A.3 for details. The key insight of the selection model is that policymaker selection through an informed intermediary is more precise, since the intermediary is better informed, but may back re if intermediary selection by the voter is weak. On average, a pure selection model predicts less popular policies by hierarchically-accountable policymakers because they are more likely dissonant, i.e., biased against the popular policy. This prediction matches that coming from the incentives model. However, the timing of policy choices is di erent. Under pure selection the popular policy di erential should be observed in nonelection years since the e ects of selection are realized post-election. Under the incentives model the popular policy di erential should occur in election years because policymakers change their behavior pre-election to a ect election outcomes. Thus, the election-year popular policy di erential is negative under an incentives model but positive under a selection 18 We nd that election-period pandering is driven by an intertemporal tradeo : choosing a popular policy against preferences at the beginning of the term results in a sure loss, whereas postponing it for the end of the term results in a potential loss, i.e., because the election period s state may be such that the policymaker prefers the popular policy. The same logic applies to the intermediary. 10

11 model. The di erence in predictions between the incentives and the selection models allows us to distinguish between the two models empirically. 3 Empirical Application Our theoretical results imply the following policy patterns: (H1) A negative mean policy di erential should exist on popular policy issues between hierarchically-accountable and electorally-accountable policymakers. No such di erential should exist on neutral policy issues. (H2) A positive policy volatility di erential should exist on popular policy issues between hierarchically-accountable and electorally-accountable policymakers. No such di erential should exist on neutral policy issues. (H3) The negative mean di erential on popular issues should decrease in the voter s crime prior and increase in policymaker congruence. (H4) The negative mean di erential on popular issues should be larger in election years. Our empirical application examines popular (police o cer employment) and neutral (police civilian employment, non-police employment) policy issues in U.S. cities where the executive is either hierarchically-accountable (manager) or electorally-accountable (mayor). Historical Background. Early U.S. city executives were appointed by state governors or city councils. After major cities like Boston and St. Louis started to popularly elect their mayors in 1822, electoral accountability became nearly universal. By the end of the 19th century city politics had became dominated by "machine" politicians drawing their support from workers and immigrants, and often using illegal means to stay in o ce. This crisis in city government accountability sparked an urban reform movement that coincided with broader social reforms taking place during the Progressive Era (roughly ). Initially reformers targeted cities electoral systems: from district-based partisan to at-large non-partisan elections. The goal was to dilute the power of minorities and parties on which the machines thrived. The Progressives, who wanted not only cleaner government but also more e cient government, were later inspired by the "e ciency movement" s success with the new corporate form in business enterprises and sought to apply the model to city governance. 19 Manager government, rst experimented with in the small city of Staunton, Virginia in 1908, attained 19 In the corporate model the CEO is accountable to shareholders through a board of directors. 11

12 broad recognition when the National Municipal League made it their recommended government form in the 1915 edition of the Model City Charter. During the 1910s and 1920s most major cities, including New York City, were debating switching to a manager govenment. Despite strong opposition from incumbent mayors and party bosses manager government advanced steadily. 20 Knoke (1982) attributes successful switches to manager to strong business interests, weak unions, high population mobility, small immigrant population, and small city size. As these factors are likely to persist and independently a ect policy today, empirically identifying the e ect of manager government seems challenging. Manager government, however, also lled a growing need for technical expertise at the top of city government, a need felt more acutely in times of crisis. To mitigate the e ects of natural disasters the new technologically-intensive public infrastructure, such as levees and bridges, had to be e ectively managed. The city of Dayton, Ohio provides an illustration. In March 1913 after days of heavy rainfall the Great Miami River over owed the city s levees causing a ood that destroyed over 20,000 homes. In the immediate aftermath local leaders sought to rebuild the ood control system with a large public works campaign. Expediency dictated the adoption of manager government so that an engineer could be appointed to lead the reconstruction e ort. 21 Subsequent crises such as the Great Mississippi Flood of 1927 and the Northeast Flood of 1936 resulted in substantial losses across multiple local jurisdictions and helped swing the balance toward federal takeover of ood control from local authorities. In 1936 Congress passed the Flood Control Act (FCA) that moved responsibility for ood prevention and management to the Army Corps of Engineers. The hundreds of miles of levees and 375 major reservoirs constructed by the Army Corps of Engineers after 1936 signi cantly weakened the link between heavy precipitation and the incidence of oods. 22 That adopting a manager government was no longer a local response to ood hazards after 1936 suggests an instrumental variable identi cation strategy: using precipitation shocks during the local ood control era ( ) to isolate exogenous variation in manager government. Data. Our sample consists of all U.S. cities with year-1900 Census population over 17, A number of 87 cities adopted manager government between , another 153 between , and 84 more between (Judd and Swanstrom 2010). 21 After Dayton s rst choice for the position George Washington Goethals, the engineer overseeing the Panama Canal construction declined, the engineer Henry M. Waite became the city s rst manager. 22 Appendix Table A1 shows that the relationship between extreme precipitation and ood incidence is markedly stronger in the local ood control period ( ). Once we control for pre-1936 precipitation shocks, 20th century precipitation shocks have little relationship with ood incidence. 12

13 residents. This sample selection criterion is not a ected by government form choice, since manager reforms do not occur until the small city of Staunton, Virginia (pop. 7,289 in 1900) starts experimenting with it in After dropping Washington DC, since it has a federally-appointed city government until 1973, a number of 248 present-day cities satisfy this criterion. The sample period for our panel is Our government form data comes from ICMA s Municipal Year Book. We point out two features of government form variation in our sample. First, the maps in Figures 1 and 2 display little geographic clustering in manager government. Second, Figure 3 shows that most manager charter adoptions in our sample occur before In fact, the majority of adoptions occurred before the federal takeover of ood control in 1936, and only ten occurred after We measure local policymaking behavior using police employment. This policy area has the attractive feature that it allows us to distinguish popular from neutral policy issues by disaggregating police employment into o cer and civilian employees, according to the distinction made in the FBI s Uniform Crime Reports. 24 We construct our instrument using weather reports from the U.S. Historical Climatology Network s Daily Temperature, Precipitation, and Snow Data. This dataset contains daily readings for temperature extremes, rainfall, and snowfall collected from weather stations throughout the U.S. We match a city to the closest weather station based on geographic coordinates. 25 density-adjusted snowfall at the station level. 26 Our main precipitation measure is the yearly sum of rainfall and We de ne an extreme precipitation event as a year when precipitation exceeds the 99th percentile of the national 20th century yearly precipitation distribution. Our cross-sectional measure of precipitation shocks for a given city is the frequency of extreme precipitation events in a given period, referred to below as precipitation shocks. For instance, precipitation shocks in the local ood control period ( ) are referred to as LFC precipitation shocks. In addition to these key variables, we analyzed an extensive set of geographic, demographic, economic, and crime variables. The Data Appendix (see Section A.4) provides the 23 Two intervening annexations and one merger slightly alter the sample: Pittsburgh, PA annexed Allegheny, PA in 1907; Omaha, NE annexed South Omaha, NE in 1915; and West Hoboken, NJ merged with Union Hill, NJ to form Union City, NJ in Census of Governments city employment data has the advantage that it distinguishes between full-time and part-time city employees, but distinguishes between police o cers and civilians only starting in The U.S. has 126 weather stations reporting in The median distance to the closest weather station for our sample cities is 47 miles. As the opening of new stations could be related to changes in local weather or local economy we keep every city matched to the same station over time. 26 We adjust snowfall for water density by dividing it by ten, as suggested by the U.S. Department of Agriculture. See 13

14 complete list of variables, with details about their measurement and sources. Table 1 reports descriptive statistics for our major variables by government form. Panel A shows that manager cities employ on average about 21% fewer o cers per capita than mayor cities, and virtually the same number of civilians per capita. Interestingly, Panel B suggests that before the advent of manager government there were no major di erences in either type of police employment. Other statistically signi cant di erences are: manager cities have on average higher reported crime, higher fraction of cleared crime (Panel A), higher incidence of Progressive institutions and Republican mayors (most manager cities maintain an honorary mayor position), lower non-police employment per capita (Panel C), and higher fraction college graduates (Panel D). Figure 4 provides a rst look at how manager government a ected police employment historically. Manager governments employ fewer police o cers from 1960 onward, but not in In Panel B we divide cities based on whether they experienced LFC precipitation shocks and nd that cities hit by these shocks have lower o cer employment after 1960, with the di erence attenuating over time. Manager and Popular Policy. We rst estimate a simple model of the form: log(employment i;t ) = 1 Manager i;t + 2 x i;t + t + i;t (1) where Employment i;t is government employees per capita for city i in year t, Manager i;t is a dummy variable indicating manager form of city government, x i;t is a vector of controls, t s are year xed e ects, and i;t is the error term. The coe cient 1 measures the conditional di erence in mean employment between manager and mayor cities. According to hypothesis (H1) we expect 1 < 0 for police o cer employment and 1 = 0 for neutral employment categories. This model allows us to account for measurable city geographic, demographic, economic, and political characteristics, and to control for national trends a ecting local police employment. As government form changes infrequently during and the observations are unlikely to be independent within a city we cluster the standard errors at the city level. 27 Table 2 presents estimates of the o cer employment di erential. Accounting for year and division xed e ects (the U.S. has nine Census divisions 28 ) column (1) shows that manager governments employ 11:9% fewer o cers per capita. Adding demographic 27 We explore alternative standard error de nitions in Table 4 Panel B. 28 The nine U.S. Census divisions are: New England, North Atlantic, South Atlantic, East North Central, East South Central, West North Central, West South Central, Mountain, and Paci c. 14

15 controls to account for voter preferences in column (2) shows that the point estimate remains negative and highly statistically signi cant. 29 Manager government may be related to other institutional and political factors that could a ect policy. The descriptive statistics in Table 1 re ect the historical fact that manager reform often came on the heels of other urban reforms: non-partisan ballots, at-large elections, and civil service rules. The literature has sometimes packaged these other reforms together with manager reform, distinguishing only between "traditional" and "reformed" cities (Stucky 2005). Manager cities may also di er in media penetration and police unionization, which may a ect voter crime probability prior () and policymaker congruence ( P ). Adding these institutional and political controls in column (3) does little to alter the estimate. Finally, column (4) shows that manager remains negatively correlated with o cer employment when controlling for city government spending. In the theory model above hierarchical accountability performs an informational role in the sense that it more often allows the policymaker to act on his private expertise. It is then useful to compare the magnitudes in Table 2 with impacts of information on government policy estimated in prior work. For instance, Stromberg s (2004) estimated elasticity of federal unemployment relief spending with respect to radio penetration (see his Table II column IV) would imply a 20.1% upper bound, in absolute value, on the o cer di erential, corresponding to a change from zero radio penetration to full penetration. All the estimates of the o cer di erential in Table 2 are within this range. Endogeneity. Even if the speci cations above control for all relevant confounds, estimates may still be biased by reverse causality. Despite the infrequency of actual changes in accountability form (e.g., 0.8% of 1,420 U.S. cities had changed their government form between according to Table 1 in Baqir 2002) city charters are endogenous by virtue of being subject to revision by popular referendum. Measurement error is another potential source of bias in estimates leading to attenuation. As manager cities typically maintain an honorary mayor, city clerks in these cities sometimes mistakenly report a mayor form of government on ICMA survey forms (see Coate and Knight 2011). To address concerns with bias in the estimates we develop an IV approach. Our strategy is based on public infrastructure crises triggering early 20th century switches to manager government to facilitate the ascension of engineers into the top executive o ce. Floods caused by extreme precipitation before the 1936 Flood Control Act were one such crisis; see Historical Background above. We thus instrument for manager government using 29 Voter or special interest preferences can a ect institutional choice. Morelli and Van Weelden (2013) argue that pandering can depend on how divisive an issue is for the electorate and Aghion, Alesina, and Trebbi (2004) discuss how the distribution of voter preferences may a ect the choice of institutions. 15

16 the frequency of precipitation shocks in the local ood control era ( ), in short, LFC precipitation shocks. Formally, the rst stage model is: Manager i;t = 1 LF C_P recipitation_shocks i + 2 w i;t + t + i;t (2) where LF C_P recipitation_shocks i is the frequency of precipitation shocks in the local ood control era for city i, w i;t is a vector of controls, t s are year xed e ects, and i;t is the error term. The controls include Century_P recipitation_shocks i, the frequency of city precipitation shocks during the 20th century, and Median_P recipitation i, median annual city precipitation during the 20th century. These variables help make the exclusion restriction credible as previous research found that climate a ects economic growth (Dell, Jones, and Olken 2012), crime (Jacob, Lefgren, and Moretti 2007), con ict (Miguel, Satyanath, and Sergenti 2004), and the origins of trust (Durante 2010). 30 Our identi cation strategy requires that, conditional on typical local precipitation patterns: (i) cities hit by LFC precipitation shocks have the same average unobserved characteristics as spared cities, and (ii) LFC precipitation shocks a ect police employment decades later only through their e ect on government form. In the Appendix we provide supportive evidence for these identifying assumptions. 31 Table 3 presents IV estimates of the o cer employment di erential controlling for typical local precipitation patterns, together with counterparts. Columns (1),(2) control for year xed e ects and typical precipitation patterns. Columns (3),(4) add water proximity controls, e.g., swamps, rivers, to account for the fact that cities close to water may implement private solutions to ood risk. Columns (5),(6) add division xed e ects. The rst stage shows a strong relationship between LFC precipitation shocks and present-day manager government: the F-statistic exceeds the critical value of 10 below which nite-sample weakinstrument bias could be a concern (Bound, Jaeger, and Baker 1995). In the second stage we nd signi cantly lower o cer employment in manager cities. The larger IV point estimates 30 In a related IV strategy using country-level data Bruckner and Ciccone (2011) exploit the fact that in non-democratic societies the cost of popular opposition to an authoritarian regime is lower during times of economic distress, making negative rainfall shocks correlated with democracy. 31 In support of instrument validity and the exclusion restriction we nd that: (i) LFC precipitation shocks are not correlated with early city characteristics (Appendix Table A2 Panel A); (ii) LFC precipitation shocks are not related to other present-day city institutions (Appendix Table A2 Panel B), and (iii) in contrast to LFC precipitation shocks, federal ood control (FFC) precipitation shocks ( ) are not related to police o cer employment today (Appendix Table A3). Trends in city managers educational backgrounds provide additional support for our identi cation strategy. More than 95 percent of city managers were engineers in 1918, 77 percent in 1934, and only 18 percent in 1977 (Stillman 1977). In other words, LFC precipitation shocks are related to manager government for reasons obsolete today. 16

17 relative to the baseline results suggest that measurement error in government form might be present in our sample (Coate and Knight 2011); another factor could be unobserved voter preferences, e.g., more conservative cities are both more likely to adopt manager government and have a tougher stance on crime. Overall the IV results uphold the substantive conclusions derived from the estimates. 32 If the costs of changing government form are heterogeneous, the identi ed e ects will be local to a subset of cities and potentially di erent from the population-wide treatment e ect. For example, Acemoglu, Robinson, and Torvik (2013) argue that voters dismantle exogenously imposed checks and balances when special interests are strong. One way to explore this issue is to examine the sensitivity of the IV estimates to changes in instrument construction. If the e ects of government form are heterogeneous and highly speci c to cities induced to change government form by our proposed instrument, alternative instrument de nitions would likely change the magnitude of the estimated e ects. Appendix Table A4 shows that the IV results are robust to alternative instrument de nitions, strengthening their external validity. 33 Using a di erent sample, a di erent period, and exploiting switches to manager government Coate and Knight (2011) nd that managers spend more than mayors. We nd that manager governments employ fewer police o cers. To reconcile these seemingly incongruous results we examine total city government spending and police department spending data in Appendix Table A5. Columns (1),(2) replicate Coate and Knight s (2011) nding in our sample; accounting for endogeneity in column (3) weakens the manager-mayor total spending di erence although the point estimate remains positive. Columns (4),(5) imply little di erence in police spending per capita while column (6) shows smaller police spending by managers, with the estimate not statistically signi cant. Robustness. Table 4 reports sensitivity checks for our preferred IV speci cation and two related speci cations. Panel A checks the sensitivity of our results to minor sample alterations: excluding extremely large/small cities, excluding the pre-1972 period, which was characterized by racial tensions, and the post-1994 period, when the federal government intervened more forcefully in local police employment through COPS grants, and dropping dependent variable outliers and cities far away from weather stations. Overall, the and 32 Baqir (2002) and Whalley (2013) nd little evidence that related forms of institutional endogeneity contaminate results in the context of U.S. local governments. 33 An additional external validity question is whether the identi ed e ects apply at other levels of government using hierarchical accountability, e.g., countries and school districts; this is an interesting question for future research. 17