Central bank independence and inflation.

Acemoglu, Daron ; Johnson, Simon ; Querubin, Pablo 等

One of the most studied types of policy reform is the introduction of CBI. Theoretical work in the early 1980s argued that when unanticipated changes in monetary policy can reduce the rate of unemployment, the government will be unable to commit to low inflation, and consequently inflation will be suboptimally high. (36) Kenneth Rogoff's proposed solution to this problem, namely, to delegate monetary policy to a conservative central banker, established a theoretical rationale for creating independent central banks. (37)

A large number of empirical studies over the past fifteen years have examined the impact of CBI on inflation, economic growth, and a variety of other variables. (38) Some early studies used a measure of de jure CBI (we describe the construction of the various measures of CBI below) and exploited cross-sectional variation within OECD countries. (39) Alberto Alesina and Lawrence Summers report a near perfect negative correlation between inflation and de jure CBI. (40)

This de jure index of CBI was further developed by Alex Cukierman, Steven Webb, and Bilin Neyapti and extended to seventy-two independent countries over the period 1950-89. (41) Using their index in pooled time-series and cross-sectional regressions, that study and another by Cukierman show that the negative correlation between de jure CBI and inflation does not hold for a cross section of developing countries, although they confirm the earlier negative correlation for developed countries. (42) They also show that their index of de facto CBI (based on the turnover of central bank governors) is negatively correlated with inflation in developing countries, but not significantly correlated with inflation in developed countries. (43)

Nevertheless, other studies find very different results. Using updated data, Christopher Crowe and Ellen Meade do not find the same correlations. (44) Marta Campillo and Jeffrey Miron argue, as does Thomas Oatley, that the correlation between de jure CBI and inflation is not robust to the inclusion of various covariates, such as measures of openness or the government deficit. (45) Oatley, Gabriel Mangano, James Forder, and King Banaian, Richard Burdekin, and Thomas Willett also document that the results depend on the subjectively coded details of CBI measures and are not generally robust. (46)

Philip Keefer and David Stasavage, (47) in work related to this paper, argue that CBI will be effective only if it cannot be reversed, and that this will happen only if there are political checks and balances. In their empirical work they interact a measure of checks and balances from Thorsten Beck and coauthors with CBI and find that introducing CBI increases inflation unless checks and balances are sufficiently strong. (48) Their work, like much of the rest of this literature, exploits only cross-sectional variation. This strategy makes omitted-variable bias potentially quite severe, since the countries that have introduced CBI typically have different macroeconomic equilibria than the rest.

In contrast to almost all of this literature, we focus on within-country variation. Although not a panacea against omitted-variable bias, fixed-effects panel data regressions provide more convincing and more relevant conditional correlations, focusing on whether inflation declines following the introduction of CBI. Using such regressions, we will show that the introduction of CBI appears to be associated with declines in inflation in countries with intermediate political constraints. The benefits of CBI in more developed economies appear to be more limited. (49)

The Motivating Theory

In this section we use a simple model to clarify our approach to the political economy of reform, and we derive hypotheses concerning the circumstances under which CBI should have a significant impact on inflation. Our purpose is not to contribute to the theoretical literature on the political economy of reform, but rather to highlight why specific institutional reforms might have different effects depending on the constraints facing politicians. For this purpose we choose the simplest model possible to communicate the major forces rather than strive for generality. Our model is a simplified version of--and a slight variant on--Gene Grossman and Elhanan Helpman's model of lobbying. (50) In this model a single organized lobby tries to convince a politician to choose distortionary policies, but underlying constraints and reform of specific institutions relating to this policy place limits on the lobby's influence on the politician and on policy choices. We first use this model to highlight the interactions between policy reform and constraints on politicians. We then use a simple extension of the model to show how successful policy reform in one sphere can lead to a deterioration in other dimensions of policy (the seesaw effect).

Our framework is chosen both for its simplicity and to emphasize the commonalities between CBI and other types of policy reform; it does not do justice to some aspects of the macroeconomic equilibrium leading to high inflation. For example, the time-inconsistency problems emphasized by Robert Barro and David Gordon and by Rogoff are absent. (51) We provide some justification for why we think these time-inconsistency problems are not first-order in the context of high inflation in developing countries. The war-of-attrition aspect of the conflict over policy reform, emphasized by Alesina and Allan Drazen, is also absent in our setup. (52) Although one could develop the same general insights using a war-of-attrition model, we prefer to use our simpler model to highlight the basic political economy factors affecting the effectiveness of policy reforms.

Model and Main Result

The economy consists of three actors: the citizenry, a politician, and an organized interest group or lobby. There is a single policy variable [pi] [greater than or equal to] 0. To make the transition to the empirical work easier, this policy can be thought of as inflation policy, although nothing in this section depends on this interpretation. In addition to policy [pi], the variable [rho] [member of] {0, [[rho].sub.R]}, with [[rho].sub.R] > 0, denotes whether or not there has been policy reform ([rho] = 0 if there has not) and parameterizes its intensity (see below). The large group of citizens has preferences given by

(1) u([pi]) = -[eta][pi],

where [eta] is a strictly positive constant. These preferences imply that the political bliss point of the citizens (that is, their most preferred policy) is [pi] = 0 (since [pi] [greater than or equal to] 0 by assumption), and any increase in [pi] away from zero reduces citizens' welfare. Thus [pi] should be thought of throughout as a distortionary policy. We could also make u directly depend on whether there has been policy reform, that is, on [rho]. This has no effect on the major results we would like to emphasize. In addition, u is made linear in [pi] only to simplify the exposition.

The second actor, the politician, has a utility function given by

(2) v([pi],[rho],t)= [lambda]u([pi]) + (1 - [lambda])t - [rho][pi].

Here t [greater than or equal to] 0 denotes a transfer from the lobby, which might consist of explicit bribes or campaign contributions. The variable [lambda] [member of] [0, 1] captures how much weight the politician's utility function places on the welfare of the citizens. We think of [lambda] as a measure of general institutional constraints on the politician (such as those measured, in our empirical work, by constraints on the executive or control of corruption). (53) When [lambda] = 1, the politician must act as a perfect agent of the citizens, perhaps because any deviation from the policies preferred by the citizens will be punished by quick replacement. In contrast, when [lambda] is close to zero, there are few constraints on the politician's behavior, perhaps because he or she is not accountable to the citizens or because politicians are difficult to replace using elections or other means. In this case the politician can pursue with impunity policies that increase the transfers he or she receives.

The other important feature of the preferences in equation 2 is the dependence on [rho][pi]. This captures the idea that policy reform makes distortionary policies more costly for the politician. For example, the politician may find it more difficult to provide credit to favored firms or groups, or to enact inflationary policies aimed at winning support. All else equal, this will discourage the use of distortionary policies by the politician. Making the use of such policies more costly for the politician is not the only way to model the effects of policy reform. (54) An alternative would be to model policy reform as introducing a hard constraint, for example imposing [pi] [less than or equal to] [bar.[pi]] for some upper bound on policy [bar.[pi]]. This is not a useful modeling strategy for understanding policy reform in societies with weak institutions, however, because such hard constraints would leave no room for pursuing distortionary policies after the reform, whereas our focus is on whether reform will prevent politicians from choosing distortionary policies.

The third actor is an organized lobby, which benefits from [pi]. Suppose that the utility of the lobby is given by

(3) w([pi],t) = [alpha][pi] - [beta]/2 [[pi].sup.2] - t,

where [alpha] and [beta] are strictly positive constants. The quadratic form is again assumed for convenience, as is the specification that these preferences do not directly depend on [rho]. These assumptions also have no effect on the qualitative results. These preferences immediately imply that the lobby's political bliss point is

[[pi].sup.*] [equivalent to] [alpha]/[beta] > 0,

and so the lobby will try, using the only instrument available to it for this purpose, the transfer t, to shift policy toward higher levels of [pi] than preferred by the citizens. Examples of policies for which citizens and lobbies have conflicting preferences include industrial policy, tariffs, and agricultural subsidies. Inflation is another potential example, since it is often used as a means of generating funds (for example, through the inflation tax) for redistribution to politically powerful groups, such as public sector employees or companies receiving procurements or industrial subsidies, at the expense of the citizens at large; inflation may also result from the use of government credits for favored firms.

Although there are three actors in all, the citizens are passive, and the main interactions are between the lobby and the politician. We model this game as follows:

--The parameter [lambda] and the reform variable [rho] are given.

--The lobby makes an offer ([??], [??]) to the politician. As in Grossman and Helpman's model, this implies that if the politician accepts the transfer [??], he or she has to implement policy [??]. This is presumably supported by a continuation game with repeated interactions, but as in much of the literature, to simplify the analysis we do not model these.

--The politician chooses policy [??]. If [??] = [??], the politician also receives transfer [??]. Otherwise, the politician receives t = 0.

This is a simple game, and we characterize its subgame perfect equilibrium, as is usually done, by backward induction. In the last stage of the game, the politician will choose whichever policy maximizes his or her utility. Clearly, this will be either [??] = [??], so that the politician receives the transfer [??], obtaining a utility [lambda]u([??]) + (1 - [lambda])[??] - [rho][??], or [??] = 0, in which case the politician receives zero transfers and obtains utility [lambda]u(0).

Therefore we can summarize the best response of the politician as follows: (55)

(4) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII].

If the lobby wishes to have its own policy preferences implemented, it must satisfy this incentive compatibility constraint (or respect the fact that the politician will play a best response in the last stage). This implies that when the lobby wishes to see implemented a policy close to its own preferences, it must choose ([??], [??]) as a solution to the following program:

(5) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] subject to equation 4.

Let the solution to this maximization problem give the lobby utility [??]. Since ([??] = 0, [??] = 0) is a feasible strategy, we must have [??] [greater than or equal to] 0. Moreover, it is also evident that this problem will lead to a solution with [??] > 0, when [??] > 0, and [??] > 0 will be chosen if and only if the solution to expression 5 also involves [??] > 0.

We next characterize the solution to this problem. The incentive compatibility constraint of the politician (equation 4) requires that if [??] > 0, then

[??] = [[lambda][eta] + [rho]/1 - [lambda]] [??].

Substituting this into the objective function of the lobby (equation 3), we have the problem faced by the lobby expressed as

(6) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII],

whenever it wants to implement policy [??] > 0. Inspection of this maximization problem establishes our first result:

Result 1. If [lambda] [greater than or equal to] [bar.[lambda]] [equivalent to] [alpha]/([alpha] + [eta]), then the lobby prefers not to have an influence on policy regardless of whether [rho] = 0 or [rho] = [[rho].sub.R].

This result follows immediately, since when [lambda] [greater than or equal to] [??], the utility-maximizing policy for the lobby is to choose [??] = 0. Since [lambda] corresponds to a measure of the quality of constraints on the politician, this result suggests that when these constraints are sufficiently strong, the political system will generate a policy choice that is not distortionary, regardless of whether there has been reform or not. In terms of our empirical work below, this result suggests that in societies that place significant constraints on politicians, reform should have relatively small effects. This can be understood by considering the example of inflation: although CBI might limit inflation in well-governed societies, one would not expect a very large decline in inflation to result from implementing CBI, since these societies would not have chosen highly distortionary policies to start with.

More generally, the solution to the lobby's maximization problem in expression 6, and thus the (subgame perfect) equilibrium level of policy, is given by

(7) [??] = max {1/[beta]([alpha] - [lambda][eta] + [rho]/1 - [lambda]); 0}.

This equation shows that the sensitivity of equilibrium policy [??] to [rho] (policy reform) will be lower, the lower is [lambda]. Intuitively, when there are no checks on the politician in power, the politician will do whatever maximizes his or her utility, and this will involve maximizing the transfers the politician receives. Consequently, transfers can outweigh the costs that policy reform imposes on the politician's use of distortionary policies. In terms of the inflation example, the politician in power can exert pressure or use other means to force the central bank to increase the money supply and inflation, even if choosing high inflation might have become more difficult or costly. (56) Interpreted differently, equation 7 suggests that in societies with low [lambda], de jure reform may not translate into de facto reform, because despite the greater cost of [pi] to the politician, the political equilibrium will induce him or her to choose policies not so different from those before the reform.

The following result now readily follows from equation 7:

Result 2. Suppose that [lambda] < [bar.[lambda]]. Then a reform that increases [rho] from 0 to [[rho].sub.R] will reduce [??]. Moreover, for [lambda] < [alpha] - [[rho].sub.R]/[eta] + [alpha], the greater the decline following policy reform.

This result therefore implies that when constraints are not so strong as to have avoided the use of distortionary policies in the first place, policy reform might be effective. How effective it will be is a function of the constraints on the politician. The greater is [lambda], the more transfers are necessary for the politician to adopt the distortionary policy after reform, and thus the lower will be the equilibrium distortionary policies following reform.

Putting these two results together, we conclude:

Result 3. Policy reform will have the largest effect on distortionary policies in societies with intermediate levels of constraints on politicians, and it will have no or only limited effects in societies with the strongest and the weakest constraints.

In the empirical work that follows, we investigate whether the effects of CBI reform on inflation are consistent with the predictions in Result 3.

The Seesaw Effect

We now use the model from the previous subsection to illustrate the seesaw effect, whereby successful policy reform might lead to a deterioration in other dimensions of policy. To do this, we augment the previous model with another policy dimension, denoted by [theta] [greater than or equal to] 0, and modify citizen preferences to

u([pi], [theta]) = -[eta][pi] - [eta]'[theta],

where [eta]' is also a strictly positive constant. This implies that [theta] is another distortionary policy, and thus the political bliss point of the citizens now corresponds to [pi] = [theta] = 0. In this instance policy reform is narrowly targeted at [pi] and thus only makes policy [pi] more costly for the politician. Some reforms, which involve the introduction of greater accountability for politicians, would not fit this pattern. CBI reform is a natural candidate in this context, since it is primarily focused on monetary policy and inflation.

The preferences of the lobby are modified to

w([pi], t) = [alpha][pi] + [alpha]'[theta] - [beta]/2[([pi] + [theta]).sup.2] - t,

with again [alpha]' > 0. The preferences of the politician are unchanged. We again look for a subgame perfect equilibrium.

The politician will choose [??] and [??] greater than zero if

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII].

By reasoning identical to that in the previous subsection, the optimal policy-transfer combination for the lobby is then given by the solution to the following maximization problem:

(8) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII].

To simplify the discussion, let us impose the following assumption:

(9) [eta]' > [eta] and [alpha]' < [alpha]

This assumption implies that policy [theta] is more costly for the citizens and less beneficial for the lobby than policy [pi]. In view of this, the following result is immediate:

Result 4. Suppose expression 9 holds and there has been no policy reform; that is, [rho] = 0. Then [??] = 0.

Intuitively, it is more economical for the lobby to receive policy favors through [pi], which is both more beneficial for the lobby and less costly for the citizens. Consequently, policy [theta] will never be used in equilibrium (either [pi] > 0 and [theta] = 0, or [pi] = [theta] = 0). Result 4 is a consequence of the simplifying assumptions made in this subsection; in particular, it depends on the assumption that the two policies, [pi] and [theta], are perfect substitutes. Without this assumption, both policies might be used simultaneously before policy reform. Nevertheless, our main result, Result 5 below, would continue to apply even when these policies are not perfect substitutes.

Next suppose that policy reform is enacted, so that [rho] = [[rho.sub.R] > 0. Our main result in this subsection is that following such policy reform, it may become beneficial to use the alternative distortionary policy [theta]. The following result summarizes the conditions under which this will happen:

Result 5. Suppose expression 9 holds and consider policy reform increasing [rho] from 0 to [[rho].sub.R]. If

(10) [alpha] - [lambda][eta] + [rho]/1 - [lambda] < [alpha]' - [lambda][eta]'/1 - [lambda],

and if

[lambda] < [??] [equivalent to] [alpha]'/[alpha]' + [eta]',

then the equilibrium following policy reform involves [??] = 0 and

[??]' = 1/[beta]([alpha]' - [lambda][eta]'/1 - [lambda]) > 0.

This result follows readily from the maximization problem in expression 8, combined with expressions 9 and 10. Note that those two expressions are consistent with each other provided that [[rho].sub.R] is sufficiently large--meaning that reform is effective in making policy [pi] costly. Result 4 implies that before policy reform, the equilibrium involves [??] = 0. Moreover, given expressions 9 and 10, before policy reform the equilibrium involves [??] > 0. Consequently, policy reform in this case creates a typical seesaw pattern: the policy that is regulated with the reform is used less intensively (moving from [??] > 0 to [??]' = 0), but at the same time the political process generates rents for the still-powerful lobby by using an alternative policy instrument more intensively (moving from [??] = 0 to [??]' > 0). This result highlights that when the political-economic interactions leading to distortionary policies remain unchanged, the imposition of a (specific or narrowly targeted) policy reform might change only the form of redistribution toward politically powerful groups, rather than eliminate policy distortions entirely.

An interesting implication of the specific configuration of parameters given by expressions 9 and 10 should also be noted: in the case discussed in Result 5, policy reform makes both citizens and the lobby worse off. Instead of redistribution being accomplished through the less costly instrument, it now uses the more costly instrument. This is not a general result: policy reform might improve the allocation of resources despite the presence of seesaw-like effects. Nevertheless, this result points out the potential pitfalls of specific policy reforms in societies where constraints on politicians are absent and political-economic interactions lead to dysfunctional policy choices.

Discussion

The model we have developed here is in the spirit of standard approaches to redistributive politics. In particular, as noted above, it does not include any element of time inconsistency. We do not believe that time-inconsistency issues play a major role in understanding high-inflation or hyperinflationary episodes in less developed economies. Most instances of high inflation are instead directly related to the inability of governments to fund their (often politically motivated) expenditure through taxation and borrowing. Such policies are much more likely to emerge when there is severe distributional conflict that the political system is unable to resolve, or when politicians have only limited instruments to use to distribute patronage. The increasing inflation in Zimbabwe, depicted in figure 1, is an example: President Mugabe turned to hyperinflation when he had no further resources left to redistribute as patronage to the military and to his key supporters.

Finally, although we have followed Grossman and Helpman in setting up the model with lobbies making offers to politicians, (57) the model could be reformulated to allow politicians to make offers to lobbies or interest groups without affecting any of the results. This reformulation might be more in line with the political experiences of countries in sub-Saharan Africa or Latin America, where interest groups are as likely to be captured by politicians as vice versa. (58) For instance, in his seminal study of economic policy in Ghana, Tony Killick notes that President Kwame Nkrumah succeeded in capturing the lobbies, making them dependent on him instead of himself on them. (59) Argentina under Juan Peron illustrates the same pattern. (60) First as minister of labor and then as president in the 1940s, Peron played an active role in creating the interest groups that then constituted his support base. Certain groups in society will have (political) resources that politicians need, and politicians will offer them redistribution in exchange for this support. The model above could be reformulated along these lines, so that policy reform would represent an increase in the cost to politicians of supplying policy favors to interest groups, possibly because they now have to use patronage to control those who run the independent central bank as well. Our major results, in particular Results 3 and 5, would continue to apply in this modified setup.

Data Sources and Construction

Throughout the paper we focus on the post-Bretton Woods period 1972-2005. From 1946 until 1971, when the United States suspended the convertibility of dollars to gold, countries under the Bretton Woods system agreed to coordinate monetary policy in order to fix their currencies with respect to gold. This naturally limited the discretion that both governments and central banks enjoyed in managing their monetary policies. This implies that the post-Bretton Woods era is a natural period for us to focus on in analyzing the relationship between CBI and inflation.

Several approaches have been used to measure CBI, and some controversy surrounds the advantages and disadvantages of each. (61) The key question centers around whether one should use de jure measures, which capture how much independence the central bank has under the law, or de facto measures, which correspond to how much independence the central bank has in practice. Another distinction that has been made is between goal independence, in which the central bank is able to determine its own objectives, and operational independence, in which it does not set its own goals but is free to achieve any goal set for it without interference.

Vittorio Grilli, Donato Masciandaro, and Guido Tabellini developed an index based on work by Hans Aufricht that emphasizes political independence, as measured by the procedures regarding the appointment of the central bank board, the relationship between the bank and the government in the formulation of monetary policy, and the formal responsibilities of the bank. (62) They also measured economic independence, which focuses on whether or not the central bank must finance government debt. The current state of the art of measurement of de jure CBI stems from Cukierman and his collaborators, (63) who constructed an index of de jure CBI for seventy-two countries over the period 1950-89, coding the variable for each of four ten-year periods, with the index remaining constant within a given period. Their index is a weighted average of sixteen different central bank characteristics. (64)

Each of the above indices has various drawbacks. Mangano argues that both the choice of criteria in the different indices and the interpretation of laws involve significant subjectivity. (65) It appears that these indices do not capture de jure CBI only but are, at least in part, informative about how monetary policy is being conducted in practice. Our focus on policy reform makes it important that we focus on de jure CBI. A related problem is that most existing work reports values for CBI indices computed at a specific point in time. However, within-country variation is essential for our empirical strategy based on panel fixed-effects regressions. (66)

To overcome these problems, we measure de jure CBI by a dummy variable that takes a value of one in every year after a major reform to the constitution or the central bank law leading to increased independence, and zero elsewhere. The advantage of this measure is that it does not incorporate information on the de facto conduct of monetary policy. A drawback is that this measure assumes that CBI increases by the same magnitude in every country following a reform, which we know is not true, because reforms in different countries have introduced different levels of independence. Nevertheless, it captures in a simple way the effect of central bank reform and changes in de jure CBI.

Our main sample consists of fifty-two countries for which information on changes in central bank legislation was reported by Luis Jacome and Francisco Vasquez, and by Simone Polillo and Mauro Guillen. (67) We exclude all former socialist countries, because data for these countries before CBI are limited, and many enacted CBI at the same time as many other reforms associated with the transition from planned to market economies. We also exclude Africa, since central bank reforms in these countries are hard to identify and interpret, and in most cases their reforms have fallen short of creating truly independent central banks. (68)

Our sample does, however, include almost every OECD and Latin American country and a sample of thirteen Asian countries. (69) To construct our CBI dummy for Latin American countries, we used the dates of major central bank reforms provided by Jacome and Vasquez. (70) For the remaining countries the CBI dummy takes a value equal to one starting in the year in which the Cukierman index constructed by Polillo and Guillen increases. (71) Moreover, for every country we used information from Marco Arnone and coauthors, (72) who report the years of major central bank reforms, to verify that a major reform did take place in the year in which the index constructed by Polillo and Guillen increases. Whenever there was disagreement between the two sources, we consulted additional sources (such as the central bank's website) to identify the year in which the most substantial reform toward CBI (if any) took place. Similarly, for countries for which the index constructed by Polillo and Guillen did not increase between 1989 and 2000, we used Arnone and coauthors and other sources to explore whether a central bank reform took place after 2000. A list of the countries included in our sample and the details of the coding, including the laws that have amended central bank charters, as well as additional sources used, can be found in the online appendix. Most central bank reforms in these countries took place during the 1990s.

We also examined the impact of CBI on the forty of these fifty-two countries whose degree of CBI changed over our sample period. Focusing on this sample enables us to obtain identification from differences in the timing of CBI; the countries in this sample might also be more homogeneous, since all have undergone the same policy reform.

Although we believe our CBI dummy provides a transparent way of investigating the relationship between CBI and inflation, we also study the robustness of our results by using the Cukierman index. Since these data are not in the form of an annual time series, and earlier data exhibit almost no variation, we take the value of the index in 1989 and assume that this value holds for all pre-CBI periods. We then use the 2003 value from Crowe and Meade for all post-CBI periods. (73)

Our main measure of political institutions is constraints on the executive, from the Polity IV dataset, which codes the extent of constitutional limits on the exercise of arbitrary power by the executive. (74) The Polity dataset reports a qualitative score between 1 and 7 for every independent country. We computed the average of the constraints on the executive variable for the period 1972-2004 and then classified each country in our sample according to whether it has weak, medium, or strong constraints on the executive. (75) Countries within one standard deviation of the sample mean for the constraints measure were assigned to the medium-constraints category, and the rest were assigned to either the strong-constraints or the weak-constraints group. Figure 3 shows the distribution of average constraints on the executive for our sample as well as the cutoffs for the different categories. Appendix table A1 shows the category in which each country was classified. The weak- and medium-constraints countries are found mainly in Asia and Latin America; most OECD countries are in the strong-constraints category.

To verify that our results are not driven by the specific cutoffs used in creating the weak-, medium-, and strong-constraints groups, we also look at the interaction of a quadratic function of the average constraints on the executive with the CBI dummy. We show that the results obtained using this approach are similar to those estimated using the dummies for weak, medium, and strong constraints.

In addition, we check the robustness of our results using different measures of institutions. In particular, we use the rule-of-law and the control-of-corruption indices constructed by Daniel Kaufmann, Aart Kraay, and Massimo Mastruzzi for the period 1996-2005. (76) These indices are reported on a scale from -2.5 to 2.5 and are based on subjective perceptions reported by experts from the public sector, the private sector, and nongovernmental organizations. (77) The classification of countries on these variables, following the same one-standard-deviation rule described above, is reported in appendix table A1. Broadly speaking, the classifications under the three institutional variables are similar. The most notable exception is Singapore, which is classified as having weak constraints on the executive, but strong institutions in terms of both the rule-of-law and control-of-corruption indices.

[FIGURE 3 OMITTED]

Our main dependent variable, inflation, was obtained from the International Financial Statistics (IFS) of the International Monetary Fund and is defined as the annual percentage change in the consumer price index. (78) Figures B1, B2, and B3 in appendix B plot the inflation data over time as well as the date of CBI for all the countries in our sample. Figure B1 focuses on the eight countries categorized as having weak constraints on the executive, figure B2 on the twenty-five medium-constraints countries, and figure B3 on the nineteen countries with strong constraints. To isolate changes in inflation in individual countries, in all three figures the average of normalized world inflation is subtracted from each country's normalized inflation rate; thus a value of zero indicates that the country's inflation rate is at the world average. (79)

Figure B1 presents a mixed picture with respect to the association between inflation and CBI in countries with the weakest constraints. In China and Guyana a decline in inflation is seen following the introduction of CBI, whereas in Guatemala and Nepal inflation drifts upward over time, and the introduction of CBI in 2002 in both countries appears to have had no effect on this process. Figure B2 shows several cases of significant disinflation concurrent with the introduction of CBI. This happens in Colombia and Argentina, as discussed in the introduction, but it also seems to occur in Turkey and Uruguay. In a number of countries, such as Bolivia, Chile, and Greece, inflation starts declining before CBI. Finally, figure B3 shows no case where CBI is followed by a large fall in inflation.

Comparison of figures B1 and B2 shows that in contrast to our theoretical model, inflation is generally higher in medium-constraints countries than in weak-constraints countries. We do not have an explanation for this pattern. It may reflect the fact that in countries with weak constraints, politicians have access to other instruments for achieving their political objectives and do not need to rely primarily on monetary policy, which may be a relatively blunt instrument for achieving clientelistic objectives or self-enrichment.

Our data on government expenditure as a percentage of GDP were obtained from the International Monetary Fund's Worm Economic Outlook. (80) In our robustness checks we also use additional macroeconomic variables to control for other time-varying determinants of inflation. We use the logarithm of GDP per capita (in constant prices, using the chain series from the Penn World Tables 6.2). Also, to control for the extent to which fixed exchange rate regimes might have provided an anchor for inflation and limited the discretionary use of monetary policy, we use an index of exchange rate flexibility constructed by Carmen Reinhart and Rogoff. (81) This index takes values between one and six, with higher values corresponding to more-flexible exchange rate regimes. The index is reported on an annual basis and covers the period 1972-2001.

Central Bank Independence and Inflation

In this section we present our main empirical results concerning the relationship between CBI and inflation. Throughout, the left-hand-side variable is defined as

(11) [y.sub.c,t] = [inflation.sub.c,t]/1 + [inflation.sub.c,t],

where [inflation.sub.c,t] denotes the inflation rate (for example, 0.1 for 10 percent inflation) for country c in year t. This transformation is useful, since otherwise hyperinflationary episodes entirely dominate the analysis. An alternative would have been to use the logarithm of inflation, but this is not possible when there are cases of zero inflation. Moreover, using logarithms would shift the outlier problem to cases in which inflation is very low. With a slight abuse of terminology, we refer to [y.sub.c,t] as inflation rather than as normalized inflation.

It should be noted throughout that changes in CBI are neither randomly assigned nor exogenous, so the empirical work we present here is intended to uncover robust correlations between central banking reform and contemporaneous or subsequent declines in inflation. As noted in the introduction, CBI is often adopted as part of a package of anti-inflation policies, and this can be seen in figures B1 and B2 in appendix B, which reveal a number of instances in which inflation starts declining a few years before CBI. This is indicative of a pattern in which other anti-inflation measures are adopted simultaneously with or just before the introduction of CBI.

The Main Effect of Central Bank Independence on Inflation

Our general estimating equation for the results reported in table 1 is

(12) [y.sub.c,t] = [k.summation over (j=1)] [[zeta].sub.j] [y.sub.c,t-j] + [[phi].sub.0] [x.sub.c,t] + [k.summation over (j=1)] [[phi].sub.j] [x.sub.c,t-j] + [[delta].sub.c] + [[omega].sub.t] + [[epsilon].sub.c,t].

Here the first summation includes lags of inflation and helps us control for potential serial correlation in this variable. The variable of interest is the dummy for CBI, described in the previous section and denoted by [x.sub.c,t]. Since CBI may have delayed effects, we sometimes also control for lags of CBI, as shown by the second summation. For simplicity, the same number of lags is used for inflation as for CBI. In addition, all regressions include [[delta].sub.c], which stands for a full set of country dummies, and [[omega].sub.t], which is a full set of year dummies. The error term [[epsilon].sub.c,t] captures all omitted influences. Since CBI is not randomly assigned, this disturbance term may be correlated with some of the right-hand-side variables. It is therefore important to interpret the estimates we report as conditional correlations rather than causal effects, [[epsilon].sub.c,t] may also have residual serial correlation, and to control for this we report throughout standard errors that are fully robust for arbitrary serial correlation at the country level (that is, clustered at the country level). (82)

Columns 1-1 through 1-3 of table 1 examine the whole sample of countries. Columns 1-4 through 1-6 repeat the same regressions but focus on the sample of countries that changed their CBI during the sample period, and thus rely only on differences in the timing of CBI to identify its effect on inflation. Column 1-1 includes only the contemporaneous CBI dummy, [x.sub.c,t]. The estimate of [[phi].sub.0] is -0.036, with a (robust) standard error of 0.034. Thus, in this first specification, the effect of CBI on inflation is negative, but this effect is not statistically significant at the 10 percent level, and the coefficient is rather small. This estimate implies that CBI is associated with a reduction in the annual inflation rate by 9 percentage points at the sample mean inflation rate of 58 percent. (83)

The specification in column 1-1 does not control for persistence in inflation, however. Since inflation is a highly serially correlated variable, doing so can have major effects on the estimates. Column 1-2 therefore includes five lags of inflation but still includes only the contemporaneous CBI dummy. When these lags are included, one can distinguish between the short- and the long-run effects of CBI. The short-run effect is still given by

[[phi].sub.0], whereas the long-run effect is [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII.]

Column 1-2 reports both of these effects, together with the p-value for the statistical significance of the long-run effect. It shows that neither the short-run nor the long-run effect is significant. For example, the long-run effect is about double the impact of CBI on inflation suggested in column 1-1, but it falls short of conventional levels of significance (p = 0.13). The lags of inflation themselves are highly significant. The results in the first two columns therefore show that despite the preponderance of theory and some evidence that CBI is important in practice, in this broad sample of countries the effect of CBI on inflation is modest and not statistically significant.

Column 1-3 also includes the five lags of the CBI dummy. In this case, however, the long-run effect is given by [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]. In this specification the short-run effect is -0.031 (suggesting that CBI decreases inflation by 8 percentage points) and is again insignificant. The long-run effect is estimated to be -0.087 (suggesting that CBI decreases inflation by 22 percentage points) but is also insignificant.

Column 1-4 reestimates the basic model reported in column 1-1 using the smaller subsample of countries that underwent a change in CBI. The estimated effect is now -0.063, with a standard error of 0.030, and is thus now significant, as are the results in columns 1-5 and 1-6, which repeat the regressions in columns 1-2 and 1-3, respectively, using the smaller subsample. Both the short- and the long-run effects are negative, and the estimated effects are larger. The basic finding of table 1, then, is of some weak evidence that CBI does reduce inflation, especially when one looks only at the sample of countries that changed their CBI over the period.

Central Bank Independence, Political Constraints, and Inflation

Why does CBI not have a quantitatively more significant effect on inflation? The answer we will propose is based on our motivating theory, namely, that CBI will have a strong effect on inflation only in societies with intermediate levels of constraints on politicians. Let [S.sub.c], [M.sub.c], and [W.sub.c] be dummy variables denoting a strong-, a medium-, or a weak-constraints country, respectively. Including interactions with these dummies, the general estimating equation becomes

(13) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII].

This equation implies that [[phi].sup.S.sub.0] measures the short-run effect of CBI on countries with strong constraints on the executive, whereas with the full set of lags included,

[long-run effect.sup.S] = [k.summation over (j=0)][[phi].sup.S.sub.j]/1 - [k.summation over (j=1)][[zeta].sub.j]

corresponds to the long-run effect for these countries. The short-run and the long-run effects for medium- and weak-constraints countries are defined analogously. Equation 13 also includes a vector of country- and time-varying covariates [Z.sub.c,t] (with a vector of coefficients denoted by v). These covariates will be included in the robustness checks reported in table 7.

Table 2 reports the results of estimating equation 13 on our base sample, with constraints measured in terms of the average constraints on the executive from the Polity IV data as described above. The first three columns use data for the whole sample. Column 2-1 is again the most parsimonious specification, omitting any lags of inflation or of the CBI dummy. The estimates suggest that CBI has no effect on inflation in weak- and strong-constraints countries but has a marginally significant effect of -0.071 (standard error of 0.044) in the sample of medium-constraints countries. This effect is twice the size of the estimate in column 1-1, showing that almost all of the effect in table 1 is due to the impact of CBI on inflation in countries with intermediate constraints. Moreover, the p-values at the bottom of the table indicate that we can reject the null hypothesis that the impact of CBI in the medium-constraints group is equal to that in the weak- or the strong-constraints groups.

Column 2-2 shows similar results when the lags of inflation are included, and again the effect of CBI on inflation is statistically significant at about the 7 percent level. However, in this case, even though the long-run effect in the medium-constraints group is almost twice as large as for the weak-constraints group, we can no longer reject the null hypothesis that they are equal. (But we continue to reject the null hypothesis that the effect is the same in strong- as in medium-constraints countries.)

Column 2-3 reports the specification that includes the lags of CBI as well as the lags of inflation. Now the long-run effect in the sample of countries with intermediate constraints is larger, at -0.125, and is significant at about the 6 percent level. The effects in countries with strong or weak constraints continue to be insignificant. In this specification we can reject at the 10 percent level the null hypothesis that the effect of CBI is the same in the medium- and weak-constraints countries (and we continue to reject, at the 1 percent level, the null hypothesis that it is the same in the strong- and medium-constraints countries).

Columns 2-4 through 2-6 estimate the same models for the sample of countries that experienced a change in CBI. The effects of CBI in the medium-constraints countries are now quantitatively larger and more significant. The short-run effect in column 2-4, -0.097, is significant and implies that CBI is associated with a decline in inflation of 24 percentage points for this group of countries, (84) We continue to reject the null hypotheses that this effect is the same as that for strong- or weak-constraints countries. In columns 2-5 and 2-6, once we include lags of the CBI dummy and inflation, the short-run effect becomes smaller and less significant, but the long-run effects are statistically and economically significant. For example, the effect in column 2-6, -0.196, implies that CBI reduces inflation by 49 percentage points. As in the first three columns, there is no evidence that CBI influences inflation in countries with strong constraints. However, the results reported in these columns show some evidence that CBI might have a negative long-run effect on inflation in weak-constraints countries, although this effect is smaller than that for medium-constraints countries.

We conclude from this evidence that the conditional correlations between the introduction of CBI and changes in inflation are broadly consistent with the ideas suggested in the theoretical discussion above. In particular, CBI appears to be followed by a decline in inflation in countries with intermediate constraints on politicians, but there is no evidence of a negative effect of CBI on inflation in strong-constraints countries. In weak-constraints countries there is typically no effect of CBI on inflation, although some specifications show negative but (usually) less significant effects. Although we can reject the null hypothesis that the effect of CBI on inflation differs between strong- and medium-constraints countries, we can reject the same hypothesis in only some of the specifications for medium- and weak-constraints countries.

Robustness

Tables 3 through 7 investigate the robustness of the results in tables 1 and 2. Table 3 replicates columns 2-1, 2-3, 2-4, and 2-6 of table 2 but replaces the Polity IV executive constraints measure with the rule-of-law and control-of-corruption indices constructed by Kaufmann, Kraay, and Mastruzzi. (85) The results in columns 3-1 through 3-4, based on the full sample, are similar to those in columns 1-1 through 1-3 of table 1 and do not show statistically significant effects of CBI. The results in columns 3-5 through 3-8, which focus on the sample of countries that experienced a change in CBI, are more encouraging. They are broadly similar to those in columns 2-4 through 2-6 of table 2. Columns 3-5 and 3-7 show evidence of a statistically significant short-run effect of CBI on inflation only in countries with institutions of medium strength, whether using the rule-of-law index (column 3-5) or the control-of-corruption index (column 3-7). These coefficients are similar in magnitude to those reported in column 2-4 of table 2. Columns 3-6 and 3-8, on the other hand, show evidence of a negative and statistically significant long-run effect of CBI on inflation in both the medium- and the weak-institutions countries. These results again show that the effects in the medium- and strong-institutions countries are statistically different, but we cannot reject the null hypothesis that the effects in the weak- and the medium-institutions countries are equal. Overall, the results using the alternative measures of institutions are generally consistent with those reported in table 2, although with these measures the differences between the weak and the medium categories are less pronounced.

An econometric problem in the estimation of equation 13 is that it includes both fixed effects and lagged dependent variables. In the presence of the latter, the key regressor [x.sub.c,t] (and its lags) will be mechanically correlated with [[epsilon].sub.c,s] for s < t, so that the standard fixed-effects estimator, used in table 2, is inconsistent. (86) However, it can be shown that the fixed-effects estimator becomes consistent as the number of time periods in the sample increases (that is, as T [right arrow] [infinity]). Even though our sample covers a relatively large number of time periods, this source of inconsistency might still be important depending on how close to a unit root the inflation dynamics are. The most common way of ensuring consistency in this case is to use the generalized method of moments (GMM) estimator developed by Manuel Arellano and Stephen Bond. (87) This involves differencing equation 13 to eliminate fixed effects, so that we have

(14) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII],

and then using lags of [y.sub.c,t] and [x.sub.c,t] to instrument for the differenced terms (here [DELTA][y.sub.c,t] = [y.sub.c,t] - [y.sub.c,t-1] and so on). These lags will be valid instruments to solve the endogeneity problem resulting from the mechanical correlation between [x.sub.c,t] and the error term if there is no additional serial correlation in [[epsilon].sub.c,t] (so that there is no second-order serial correlation in [DELTA][[epsilon].sub.c,t]), which can be tested. Table 4 reports the results of estimating equation 14 on the same samples and specifications as in columns 2-2, 2-3, 2-5, and 2-6 of table 2. (Here and throughout the rest of the paper, [S.sub.c], [M.sub.c], and [W.sub.c] are again based on the Polity IV executive constraints variable.) The results are similar to those of table 2. In particular, the interactions between the medium-constraints dummy and CBI are still negative, and the estimates are almost all statistically more significant than in table 2. There is no evidence of a negative effect of CBI on inflation in strong-constraints countries. The results in table 4 also show some evidence of a negative effect of CBI in weak-constraints countries once we focus on the sample of countries that changed their CBI, in columns 4-3 and 4-4. Nevertheless, these results should be interpreted with caution: although the tests for serial correlation cannot reject the null hypothesis of no second-order correlation in the differenced residuals, the Sargan overidentification tests reject the null hypothesis in all columns, which implies that different lags estimate different magnitudes for the coefficients of CBI and lagged inflation (which is often a problem when using this GMM procedure with large T).

Table 5 investigates an alternative to the use of the three dummy variables [S.sub.c], [M.sub.c], and [W.sub.c] for strong-, medium-, and weak-constraints countries to capture the nonlinear effects of this variable, namely, the use of a quadratic term in the average of constraints on the executive interacted with the CBI dummy. The coefficients reported evaluate the quadratic at levels 1, 4, and 7 of constraints on the executive. The results are broadly similar to those in table 2. The short- and the long-run effects of CBI on inflation are always negative and significant for countries with medium constraints. There is no evidence of a significant effect of CBI in countries with strong or weak constraints. However, we can reject the hypothesis that the effects are equal only when comparing the strong- and the medium-constraints groups. Overall, these results suggest that the specific way of parameterizing the non-linearity in the exact cutoffs between weak and medium, and medium and strong, constraints does not matter greatly to our results.

Table 6 reports the results of a simple falsification exercise. Each regression this time includes the interaction of each of the executive constraints dummies with both the contemporaneous CBI and its five-year lead. To the extent that inflation was already declining in some countries before central bank reform, this lead variable will capture the decline. When this is the case, we cannot have much confidence that CBI is in fact a major factor in the decline in inflation. The table shows that the contemporaneous effect of CBI in countries with intermediate constraints is still significant. The coefficient on the five-year lead is never negative for this group and is typically very small. Hence the negative association between CBI and inflation in this group of countries is not driven mainly by a secular decline in inflation. For strong-constraints countries, the lead of CBI is likewise never significant and is typically very small. For weak-constraints countries there is some evidence of an upward trend in inflation preceding CBI. This suggests that in some of these cases CBI may have been introduced in the midst of episodes of particularly high inflation.

Finally, table 7 reports results from a range of other robustness checks. The top panel focuses on the full sample and reports results of specifications corresponding to columns 2-1 and 2-3 in table 2. In the bottom panel the sample is restricted to countries that changed their CBI during the sample period, and specifications corresponding to columns 2-4 and 2-6 of table 2 are reported. Columns 7-1 and 7-2 in both panels report the robustness of the results to including potential time-varying determinants of inflation. In particular, we include five lags of the logarithm of GDP per capita and five lags of the exchange rate flexibility index constructed by Reinhart and Rogoff, described above. (88) Although the negative effect of CBI on inflation is not significant in the full sample, the results in the bottom panel show a pattern similar to that depicted in table 2, with a negative and statistically significant long-run effect of CBI only in the group with intermediate constraints on the executive. The results in columns 7-3 and 7-4 include a quadratic time trend interacted with regional dummies to control for differing trends in inflation across world regions. These additional controls have little effect on the estimates. In column 7-3 in the top panel, the estimate of the impact of CBI in medium-constraints countries is -0.065, with a standard error of 0.044, which is similar to that reported in column 2-1 of table 2. The long-run effect in column 7-4, however, is smaller and less significant than that in column 2-2. Nevertheless, the results in the bottom panel for columns 7-3 and 7-4 show that in the sample of countries that changed CBI, once we control for differences in regional trends, the results are again similar to those in table 2.

In columns 7-5 and 7-6 we exclude the Western European countries from the sample entirely. This robustness check is motivated by the concern that policy dynamics and the effect of CBI on inflation might be different in Western Europe than in the rest of the world. The results, however, are again similar to the baseline estimates.

Columns 7-7 and 7-8 include an additional dummy for the introduction of the European Central Bank, or for the country in question joining the European monetary union after the European Central Bank was established. Again there is little change from the results in table 2.

Finally, columns 7-9 and 7- 10 show the robustness of our results to using the CBI index constructed by Cukierman rather than the CBI dummy. Despite the concerns discussed above related to the coding and interpretation of this index, it still constitutes a useful check on our measure of CBI and on our results. The results using this index are again similar to those reported in table 2: only the interaction between the medium-constraints dummy and the Cukierman CBI index is negative and statistically significant. (89) Interestingly, in columns 7-5 through 7-10 for both panels we can now reject at the 10 percent level the null hypothesis that the effects in the intermediate-constraints group are equal to those in the weak-constraints group as well as in the strong-constraints group.

We also examined the robustness of the results reported in table 2 to the type of CBI that different countries implemented. To do this we ran the same regressions as in table 2 but included a dummy for goal independence of the CBI. This dummy takes a value of one for countries where CBI reform defined inflation as the unique objective of monetary policy, and zero elsewhere. This is a useful robustness check against the possibility that CBI may have very different effects where the central banker is truly conservative, with price stability the unique objective (as with the German Bundesbank), than where CBI has multiple objectives (as with the U.S. Federal Reserve). The results (not reported here) using the goal independence dummy are analogous to those in table 2 and suggest a similar pattern, with a negative effect of goal independence on inflation found only for countries with intermediate constraints on the executive. We also briefly investigated whether changes in the dynamics and persistence of inflation in the mid-1980s, documented for example by Haroon Mumtaz and Paolo Surico and by James Stock and Mark Watson, (90) might be responsible for our results. As a crude control for this possibility, we interacted the lags of inflation with a dummy for the period 1972-90 and a dummy for 1991-2005. The results (again not reported here) exhibited a pattern similar to that in table 2, with a negative effect of CBI found only in countries with intermediate constraints on the executive, although the statistical significance of some of the estimates became weaker.

The Seesaw Effect

Here we investigate the implications of CBI reform for other dimensions of policy. In particular, we look at whether CBI is followed by an increase in government expenditure as a percentage of GDP in countries with intermediate political constraints. Our basic econometric model is again equation 13, but now the dependent variable measures fiscal policy. Table 8 reports the different effects of CBI in strong-, medium-, and weak-constraints countries on government expenditure as a percentage of GDP. No evidence is found that introducing CBI in countries with strong or weak constraints has any impact on government expenditure relative to GDP in either the short or the long run. However, in medium-constraints countries, where CBI was found above to be associated with declines in inflation, CBI also appears to be associated with increases in government expenditure. Both the short- and the long-run effects are positive for these countries in all specifications, and in columns 8-2, 8-3, and 8-5 these effects are statistically significant. In column 2 the estimated [[phi].sup.M.sub.0] is 0.011, with a standard error of 0.004, which suggests an immediate increase in government expenditure equivalent to just above 1 percent of GDP. The long-run effect of 0.050 is also significant and corresponds to a 5-percent-of-GDP increase in government expenditure in the long run in medium-constraints countries that have introduced CBI.

Table 9 shows that the broad pattern is similar when we use the Arellano-Bond GMM estimator as in table 4. In this case the evidence consistent with the seesaw effect is somewhat stronger, although now we also find some positive long-run effect of CBI on government expenditure in weak-constraints countries (see, for example, column 9-2). Finally, table 10 reports results of a falsification exercise similar to that in table 6. Here we again include a five-year lead of the CBI dummy to check whether the estimates in table 8 might be capturing a secular upward trend in government expenditure in countries with intermediate constraints. The results show no evidence of such a pattern. The interaction between the medium-constraints dummy and the five-year lead of the CBI dummy is never significantly different from zero.

We also investigated whether CBI is associated with a change in the government budget balance (again measured as a fraction of GDP). These results (not reported here) reveal no significant effect. One reason might be that governments in medium-constraints countries, which do show a significant increase in spending, may have also raised revenue. Alternatively, government balances may have improved automatically because these governments held nominal debt and benefited from the decline in interest rates associated with reduced inflation. Counteracting this, however, they are also likely to have experienced a reduction in the inflation tax.

We conclude that there is some suggestive evidence consistent with the seesaw pattern, in that in some of the cases where CBI reduces inflation, there may also be an increase in government expenditure as a share of GDP. Nevertheless, this result is not always robust, and further investigation is necessary before reaching a firm conclusion.

Conclusions

We have argued that in order to understand when policy reform will be effective, one must understand the political context that leads to distortionary policies in the first place. Although such a claim would seem natural to those steeped in the study of political economy, it appears not to have influenced either the design of policy reform in the past or the recent wave of revisionist research on the failures of reform. Rather, scholars have taken some recent instances of failure to imply that the reforms themselves were misconceived because of second-best problems. We believe that a much more fruitful approach is to study the political economy constraints that interact with both the implementation and the effects of reform.

To develop this perspective, we constructed a simple political economy model of policy distortions and showed that when such distortions arise as a result of the political equilibrium, strong forces exerted by those actors initially benefiting from the distortions may negate the effects of the reform. The model illustrates in a simple way that both the extent of distortionary policy and the effects of reform will generally depend on political institutions, for example on various measures of the constraints on political officeholders. The first major insight of the model is that policy reform may not be effective when constraints are so weak that reform can be undermined, or when constraints are sufficiently strong that policy is unlikely to have been highly distortionary to start with. Rather, it is at intermediate values of constraints that policy reform might be most effective. The second important lesson from the model is that with multiple policy instruments, reform may lead to a seesaw effect, whereby effective or partially effective reform in one dimension leads to more-intensive use of other distortionary instruments.

The paper also provides empirical evidence consistent with these patterns in the context of CBI. Our evidence suggests that CBI is associated with significant declines in inflation in countries with medium constraints, but with no decline, or smaller declines, in countries with strong or weak constraints. There is also some evidence that, consistent with the seesaw effect, countries where CBI is associated with reduced inflation also experience an increase in government expenditure. This might partly reflect the use of an alternative policy instrument to achieve goals that high inflation was previously being used for.

Do these results imply that policy reform can never work? We do not think this is the correct conclusion, for two reasons. First, our evidence is consistent with the view that targeted reforms, such as anti-inflation programs and CBI, can be effective. Second, the evidence is difficult to reconcile with a naive approach that assumes that politicians adopt reforms for well-meaning reasons or because they have decided that their old policies were inappropriate. In this light, if reforms fail to be fully effective, it must be at least partly because of the constraints imposed by existing political economy factors. In particular, when reforms are promoted by international organizations and imposed on reluctant leaders, they are less likely to be effective or, at the very least, need to be designed much more carefully to ensure their effectiveness. Otherwise our model and the empirical results suggest that de jure reform may not translate into de facto reform, and when it does, other dimensions of policy may deteriorate.

The general message of our paper is that the analysis of policy reform should start with an understanding of the political economy constraints that lead to distortionary policies in the first place. This type of political economy analysis might ultimately be useful both in predicting when reforms are likely to be effective and in understanding how reforms can be better designed to reduce the risk of backlash. Although we are currently too far from a satisfactory theoretical or empirical understanding to be able to design policy reforms that are robust to such challenges, this paper and others in this vein should help make the case that an analysis of the political economic roots of distortionary policies must be part of any debate on the effectiveness and design of reforms.

Table A1. Classification of Countries by Strength
of Executive Constraints and Institutions

 Coding bused Coding Coding based
 on executive based on on control of
Country constraints (a) rule of law (b) corruption (b)

Argentina Medium Medium Medium
Australia Strong Strong Strong
Austria Strong Strong Strong
Belgium Strong Medium Medium
Bolivia Medium Medium Weak
Brazil Medium Medium Medium
Canada Strong Strong Strong
Chile Medium Medium Medium
China Weak Medium Medium
Colombia Medium Weak Medium
Costa Rica Strong Medium Medium
Denmark Strong Strong Strong
Dominican Republic Medium Weak Medium
Ecuador Medium Weak Weak
El Salvador Medium Weak Medium
Finland Strong Strong Strong
France Medium Medium Medium
Germany Strong Strong Strong
Greece Medium Medium Medium
Guatemala Weak Weak Weak
Guyana Weak Medium Medium
Honduras Medium Weak Weak
India Medium Medium Medium
Indonesia Weak Weak Weak
Ireland Strong Strong Medium
Israel Strong Medium Medium
Italy Strong Medium Medium
Japan Strong Medium Medium
Korea Medium Medium Medium
Malaysia Medium Medium Medium
Mexico Medium Medium Medium
Mongolia Medium Medium Medium
Nepal Weak Medium Medium
Netherlands Strong Strong Strong
New Zealand Strong Strong Strong
Nicaragua Medium Weak Weak
Norway Strong Strong Strong
Pakistan Medium Weak Weak
Paraguay Weak Weak Weak
Peru Medium Weak Medium
Philippines Medium Medium Medium
Portugal Medium Medium Medium
Qatar Weak Medium Medium
Singapore Weak Strong Strong
Spain Medium Medium Medium
Sweden Strong Strong Strong
Switzerland Strong Strong Strong
Turkey Medium Medium Medium
United Kingdom Strong Strong Strong
United States Strong Strong Medium
Uruguay Medium Medium Medium
Venezuela Medium Weak Weak

Sources: Authors' determinations based on sources indicated.

(a.) From Polity IV dataset.

(b.) From Kaufmann, Kraay, and Mastruzzi (2007) data.

APPENDIX A

[FIGURE B1 OMITTED]

[FIGURE B2 OMITTED]

[FIGURE B3 OMITTED]