From Education to Institutions!

Transcription

1 From Education to Institutions! Alvar Kangur April 30, 2009 Abstract In this paper I take the question of the role of human capital versus institutions in economic growth to a dynamic panel data. Contrary to many recent contributions on the field it is shown that, although there are some merits to it, on theoretical grounds System-GMM cannot be preferred to Difference-GMM. Further it is argued that the policy-regime regressions widely utilized in the literature do not correspond to the dynamic properties of institutions that become important once time-variation is used for identification purposes. Further, it is important to account for the educational distribution of contingencies to capture properly their incentives. It is shown that for crude regime choices incentives stemming from wider educational base are important, while higher education is relevant in explaining dynamic institutional innovations. There seems to be no evidence whatsoever of reverse causality from institutions to education. JEL classification: O11, O40, P16 Keywords: Education, institutions, dynamic panel University of Oxford, Department of Economics and Lady Margaret Hall, DPhil candidate. Alvar.Kangur@economics.ox.ac.uk 1

2 Contents Contents 2 1 Introduction 3 2 Some aspects of institutions a la North (1981, 1990) Static aspects of institutions Dynamic nature of Institutions Reinterpreting Institutional Indices 9 Do available indices measure institutional constraints?. 9 Is Polity IV following constitutional changes? Is Polity IV capturing changes in operating rules? Do these measures capture anything credible? Does history matter? Is Polity IV stock or flow? Another look at link between education and institutions 13 5 Policy Regime Regressions A Look at Existing Evidence Potential Issues Reinterpreting the Education-Institutions Relationship Levels-to-Levels estimation Flows-to-Flows estimation From Institutions To Education 23 8 Conclusion 24 Bibliography 25 2

3 1 Introduction For at least a half of century it has been a conventional wisdom that education leads to democracy and better institutions that eventually leads to higher growth. This hypothesis has most prominently been assigned to Lipset (1960) according to whom education as well as number of other social conditions leads to a deeper culture of democracy being developed. Similar hypothesis though on different grounds has been described by a father of institutional economics, Douglass C. North (1981). For example (North 1981, p. 208) writes: The cumulation in a stock of knowledge [embodied in technological developments and skills of people] has imposed an evolutionary order upon the secular change of political and economic institutions... In recent decade, however, this view has been challenged. After seminal work by Hall and Jones (1999), Acemoglu et al. (2001, 2002) in a series of articles have argued for institutional prevalence. In a cross-section of countries they have used the quasi experiment of colonization patterns to identify strong and significant effects of institutions on income levels. However, these results have not been left uncriticized. Glaeser et al. (2004) have demonstrated that education seems to be much stronger predictor of growth and income levels than available indices of constitutional constraints. They also run a race between average levels of institutions and education in a cross-section of countries and find that, if anything, only education comes out as important determinant of income levels. In Kangur (2008), a companion article to this I have argued that settler mortality put forward by Acemoglu et al. (2001) cannot be considered as plausible instrument for institutions and, if anything, the colonial settlement patterns are more associated with changes in human capital endowments, a result more in line with Lipset (1960) as well as with endowment view of Engerman and Sokoloff (1997, 2002). However, many writers have pointed out that in a simple cross-sectional applications, high multicollinearity in instrumented levels regressions, arising from the fact that instruments for one endogenous variable also predict other instrumented endogenous determinant and vice versa, could lead to violation of the rank condition, and thus the deep parameters of interest are not properly identifiable. For this reason one must resort for other identification strategies. Glaeser et al. (2004) have exploited the time series variation to provide some evidence on the timing of education and institutions. More specifically, they set up the following dynamic equation: X 1it X 1i,t 1 = λx 2i,t 1 + µ i + ν it (1) where X 1 and X 2 is either an average years of schooling or political institutions, the latter being measured as an average of available index values over the respective time interval. µ i denotes fixed country effects, ν it is an error term containing variables that are omitted from the regression, and time t is measured in 5-year intervals. They claim that in a regression where change in 3

4 institutions is regressed on initial schooling, the latter comes out as very significant, whereas in a reverted regression initial levels of institutions do not have any power to predict changes in schooling. In a subsequent paper Acemoglu et al. (2005), however, note that without time-effects the effect of education is identified through the general increase in education and political institutions that in turn says nothing about causal effects. In their empirical application they add time effects to regression (1) and show that education looses its power to predict changes in institutions; in fact, although insignificant, the coefficient on initial level of average schooling always comes out as negative. Thus the authors argue for no causal links between education and changes in institutional quality. Acemoglu et al. (2005) estimation strategy has been subject to criticism as well. From methodological point of view both Bobba and Coviello (2007) and Castelló-Climent (2006) have argued that because of the presence of high persistence in the time-series the Difference GMM used by Acemoglu et al. (2005) may be subject to a large downward bias, a result demonstrated by Blundell and Bond (1998) using Monte-Carlo simulations. Re-estimating the model with System GMM they recover positive effect of education on democracy. Castelló-Climent (2006) also argue that what matters for democracy is not so much the average measure of education but some measure of distribution of education to proxy for the education attained by the median voter. In this article rather than looking for a transition from autocratic to democratic rules of the game I investigate whether education causes institutions as a dynamic process or vice versa. I make four contributions. First, I review critically dynamic properties of institutional measures provided by Polity IV that most prominently feature large discrete changes, high volatility, meanreversion and no path-dependence. 1 I argue that these are the properties of flow variables, whereas the underlying concept of institutions, as evident from the writings and the above quote by North (1981), is inherently a stock. This distinction is important if one resorts to time-series in addition to cross-sectional variation in order to identify the line of causality. Second, I review the existing panel-data literature where the effect of average total education on institutions is identified only through the moment conditions obtained from the levels equation of system-gmm estimator. At the same time there are many issues related to the use of levels equations that make the estimates less reliable. I provide evidence that once a measure more closely linked to the notion of institutional constraints such as Executive Constraints from Polity IV is used, penetration of primary education as more closer proxy of wider educational base, a precondition of the emergence of democracy-oriented constituency, comes out as more robust predictor in both difference and levels equations. Third, I re-estimate the education-institutions link in the spirit of treating the stock of institutions as a latent variable, and observed ordinary measures as rough proxies to policy flows that, when accumulated, give a proxy for the 1 Although I focus on Polity IV the same features are present in most of the available institutional indices. 4

5 institutional stock. In the context of regression (1), if one regresses the change in institutions on the initial level of education this means that not X 1it, but X 1it should be measured by available institutional index values. I present evidence that in this setting only the incentives associated with higher level of education can have any predictive power on institutions. Fourth, when the estimated equations are reverted, there is no evidence of reverse causality running from institutions to education. If anything, the effect of institutions can be negative. The rest is organized as follows. In section 2 I review some of the properties of economic institutions with particular emphasis on its dynamic nature pertinent to a stock variable as opposed to more static choice of regimes. Section 3 then takes another look at popular measures of institutions used in the literature and argues that these are measures of flows, and not institutions subject to a secular change. Thus, in empirical applications, the underlying measure of institutions should be treated as a latent variable, whereas what is observed is a rough measure of changes in the underlying institutions. Section 4 comments on some of the caveats in the link between education and institutions. Section 5 re-estimates the existing policy-regime regressions with particular attention on problems with identification through level equation and incentives stemming from different levels of education. Section 6 in the spirit of this paper treats the institutional stock as a latent variable and provides further evidence on causality running from education to institutions. Section 7 takes a look at issues of reverse causality and finally section 8 concludes. 5

6 2 Some aspects of institutions a la North (1981, 1990) In this section I will review some crucial aspects of institutions as they were defined most prominently by North (1981, 1990). I do not attempt to add anything innovative to it, rather my motivations here is purely empirical. Having emphasized some of the main features of institutions I will then take a look at the existing institutional indices that have been used in empirical literature to see if these actually comply with those underlined here. 2.1 Static aspects of institutions As a norm, any discussion of institutions begins with two interlinked and most obvious what I would call static characteristics: ability to formulate constraints and to ensure their credible execution. The first characteristic refers to why we have institutions in the first place, while the second dimension is needed in order for them to be binding in the society Dynamic nature of Institutions Focusing only on static definitions largely neglects another important feature of economic institutions that is in the heart of this writing - the dynamic nature of institutions. The reasons for institutional change can be plentiful, keeping in the spirit of North (1981) it is conflict of ideologies and, most of all, changes in relative prices and opportunity costs of constituents that lead to institutional innovation. More relevant to the current context, however, is the process of institutional change itself. Among the main components of institutions - constitutional rules, operating rules, and normative behavioral codes - constitutional rules are the most costly to modify and thus usually the most static. Nevertheless, even when constitutional rules are radically altered, institutions themselves do not go through an abrupt discontinuous change: In consequence, there develops an ongoing tension between informal constraints and the new formal rules, as many are inconsistent with each other. The informal constraints had gradually evolved as extensions of previous formal rules.... Although wholesale change in the formal rules may take place, at the same time there will be many informal constraints that have great survival tenacity because they still resolve basic exchange problems among the 2 The following are few classic quotes: Institutions are a set of rules, compliance procedures, and moral and ethical behavioral norms designed to constrain the behavior of individuals in the interests of maximizing the wealth or utility of principals. (North 1981, p ).... it is essential to specify the individual behavioral characteristics that lead to the constraints that make up institutions....individuals in the absence of any constraints maximize at any and all margins; it is, then, constraints that make possible human organization by limiting certain types of behavior. (North 1981, p ). 6

7 participants, be they social, political, or economic. (North 1990, p. 91) Institutions typically change incrementally rather than in discontinuous fashion. How and why they change incrementally and why even discontinuous changes (such as revolution and conquest) are never completely discontinuous are a result of the imbeddedness of informal constraints in societies. (North 1990, p. 6) This underlies the stability of institutions further emphasised by the property of path-dependence: due to many interlinkages and persistence of informal constraints a discontinuous change in formal rules does not translate into an equivalent discontinuous change in institutions. Rather it will show up as a gradual change over time as the restructuring of constraints takes place. The concept of path dependence inherent in the evolutionary changes of institutional constraints is most vividly illustrated by the evolution of common law 3 : The evolution of the common law, a form of institutional change, is helpful in understanding overall institutional change. Common law is precedent based - it provides continuity and essential predictability that are critical to reducing uncertainty among contracting parties. Past decisions become embedded in the structure of law, which changes marginally as new cases arise involving new, or at least in terms of past cases unforeseen, issues; when decided these become, in turn, a part of the legal framework. (North 1990, p ) Path dependence means that history matters. We cannot understand today s choices (and define them in the modeling of economic performance) without tracing the incremental evolution of institutions. (North 1990, p ) The importance of these concepts for empirical investigations of the impact of institutions cannot be underestimated for a number of reasons. First, the persistence or path-dependence is in fact the single most crucial feature that has been used to justify the exploitation of historical events as providing instruments for institutions in Acemoglu et al. (2001, 2002) as well as many other writings. Second, using measures close to constitutional rules as a first-order proxy for institutions in empirical applications that exploit time-variation to identify certain causal effects might not be entirely appropriate. This is reminiscent of the nature of constitutional rules as being the most static component of institutions. Third, note that constitutional rules exist to define a broad framework within which policy decisions are made and, as a result, operating property rights are 3 Similar concepts of path-dependence and incremental institutional innovation) are reinforced also in the writings of Greif (2006) (see pp ). 7

8 created. Thus measures that simply describe outcomes of changes in legislature (which most of the real world polity indices do) do not capture the underlying state of institutions, but rather the nature of political decisions that henceforward will contribute into the formation of property rights and operating rules. These last two points will be taken under closer scrutiny in the next sections. In sum, if one would look for one measure to capture the outlined features of institutions over time, all aforementioned reasonings indicate that a good measure of institutions should behave as a stock variable. Institutions reflect and build upon the entire (feasible) historical experience. Each period the rulers add to the institutional stock either by changing constitutional or more likely operational rules as well as by adjusting the framework of adjudication and enforcement. As consensus ideology develops, also new rules of the game will gradually be introduced, whereas some of the old ones will cease to bind. It is the combination of the constitutional rules with the associated moral and ethical codes of behavior that underlies the stability of institutions and makes them slow to change. The combination produces ingrown patterns of behavior which, like the capital stock, tend to be changed only incrementally. (North 1981, p. 205). The view of institutions by North (1981) is also reflected in the reasoning provided by Rodrik et al. (2002) who think of quality of institutions as a predetermined stock variable representing not only the political choices of current but also those of past political rulers. Thus current policies affecting the institutional quality are naturally a flow variable. One can then write down an equation of motion for institutional quality I as: 4 I = α i p i δi, (2) where α i denotes the impact of policy p i on institutional quality. 4 Rodrik et al. (2002) themselves are not concerned by the dynamic properties of institutions, they use their reasoning as a criticism of Easterly and Levine (2003) who regress income levels on institutional quality and policies at the same time, whereas according to equation (2) policies are already included in I and thus simply do not belong to the regression of income on institutions. 8

9 3 Reinterpreting Institutional Indices In this section I investigate whether some of the institutional indices used in the literature actually share these properties of institutions as defined previously. Much of the discussion here relies on Glaeser et al. (2004), who conduct a careful analysis of three of the most popular sources for these measures used to make causal empirical claims related to effects of institutions. Do available indices measure institutional constraints? Careful inspection by Glaeser et al. (2004) reveals that most popular indices used in the contemporary research as measures of institutions are already by construction not constraints: they capture ex-post outcomes, reflecting the choices of executives, and not the constraints under which these choices are made. The only indicator that possibly stands out in capturing the decision-making environment itself, is Executive Constraints from Polity IV dataset. Indeed, according to Marshall and Jaggers (2005, p ) the executive constraints, or decision rules, are characteristic of the extent to which the head of the unit... must take into account the preferences of others when making decisions... they provide basic criteria under which decisions are considered to have been taken. Given these qualifications in what follows I concentrate on Polity IV database only, though many of the following considerations apply also more generally. Is Polity IV following constitutional changes? I demonstrate some of the features of Polity IV database on Figures 1 and 2, where the former graphs 6 nations whose volatility is at the higher end of the scale, and the latter exhibits three examples of opposite extremes. From the discussion in the previous section operating rules evolve incrementally like the common law, whereas the graphs on Figure 1 show rather abrupt discrete changes, most likely characteristic to changes in constitutional constraints. Experience from Greece shows that after a period of a military regime during , democratic republican constitution was enforced in 1975, leading to a big jump in both indices of democracy and executive constraints. In 1985 new president came to a power and in 1986, at the time of an important constitutional amendment, both indices jump further to their maximum values. Similarly, in South-Korea periods of five republics (from second to sixth) that all started with constitutional changes are clearly evident. However, in most cases this is not what Polity IV indices follow; rather the changes coincide almost perfectly with electoral outcomes or other events, such as military coups, that bring about a major shift in ruling legislature. In Argentina executive constraints follow exactly the changes of democratic presidential and military governments. The current version of Argentine constitution is due to president Carlos Menem in 1994 with no corresponding changes in respective indices at all. Rather Polity IV reflects presidential elections in 1989 and In Chile after a coup of 1973 a military government was established. The new constitution, however, was only adopted in 1980 in an undemocratic ref- 9

10 erendum that elected Pinochet as president. New democratic president and congress were elected in 1988 and 1989 respectively; and since 1989 the constitution was amended about 9 times. Non of the constitutional changes, however, are reflected in the dynamics of Polity IV indices. Two periods of military governments after 1958 and 1977 coups are evident for Pakistan. Polity IV shoots up when Benazir Bhutto was elected as the prime minister, though the constitution of 1973 was reinstated only in After a number of constitutions, charters, and political turbulence, the 1991 military government of Thailand was overthrown in 1992 when Polity IV skyrockets to its maximum levels. However, the landmark of democratic constitutional reform is Thus close inspection reveals that Polity IV provides a rapidly moving assessment of electoral outcomes over time, not a measure of actual political constraints on government, and certainly not a measure of anything permanent or durable (Glaeser et al. 2004, p. 11). It s primary focus is clearly on capturing the operational nature of policy-making by ruling legislature. Nevertheless, it makes a great attempt in classifying these policies on a scale from unlimited authority to executive parity. Is Polity IV capturing changes in operating rules? Operating rules almost by definition evolve like common law, changing marginally and building upon the past inheritance. Closer look on executive constraints reveals a combination of extreme outcomes. As Figure 1 has suggested, countries with frequent changes in legislature or the nature of political system experience corresponding extreme volatility. The experience for most of the OECD countries in Figure 2 falls to another extreme: there is almost no time-variation in most of the long-time stable democracies who get constantly the maximum score of 7. In fact there is no time-variation in any of the countries where electoral outcomes have not changed the political environment: stable dictatorships like Cuba, Qatar and Saudi-Arabia get constantly the lowest score of 1, whereas communist regimes of USSR and China tend to get scores of 3. However, it is hard to believe that even in countries like US or UK operating rules, or institutions more generally, are the same now as half a century ago. This period has witnessed the most rapid technological change in human history, that, consistent with North (1981, 1990), must have been accompanied by institutional innovation. Do these measures capture anything credible? Constitutional rules are costly to modify and involve certain adjustment costs, property rights must be complemented with frameworks of adjudication and enforcement etc. Even when political regimes will be changed abruptly, the newly modified institutions still have to mix with other rules and norms to create a new equilibrium. Following the discussion in the previous section, for an institutional change to be credible this process takes time. However, as seen from Figure 1, Polity IV measures can be rather volatile with many abrupt changes. Glaeser et al. (2004) point out that, on average, Polity IV is almost twice as volatile than education. 10

11 Figure 2 provides few extreme examples. In Spain the new democratic constitutions was approved in 1978, 3 years after the death of general Franco. In Baltic States, after the collapse of Soviet Union, new democratic governments were created in Both events were enough to move these countries from the bottom of the Polity IV scale to the absolute top. Does this mean, that the institutions of Spain in 1978 and of Baltic States in 1991 were credible? Can we really say, that with these regime shifts institutions of Spain and Baltic states were in a blink of an eye made comparable to institutions of these 21 OECD countries, who have received maximum scores for the last forty years and longer? Clearly, what is compared here is not capturing the process of formation of institutions, but the accountability of policy decisions at a given point in time. Does history matter? Glaeser et al. (2004) have found that political institutions in Polity IV are strongly mean reverting and clearly less persistent than education. I demonstrate it in Table 1, where in a 5-year panel I regress Polity IV measures of institutions as well as average years of schooling on their 5-year lags, 5-year lagged difference as well as on time effects. As expected, education exhibits a unit root property, whereas institutional measures follow a mean reverting process. Further, as previously demonstrated, Polity IV follows most recent electoral outcomes, and thus not the entire history of institutional experience. Is Polity IV stock or flow? What has been neglected in the empirical applications is that all these properties - discrete changes, high volatility, meanreversion, no history- or path-dependence, description of policies at given point in time - are properties of a flow variables, not underlying stocks capturing the whole history of the underlying concept that institutional economists are used to talk about. 5 All in all, the documented dynamic patterns clearly do not correspond to an incremental innovations underlying the secular change of institutions a la North (1981). Consistent with the evidence, it is my interpretation that Polity IV, as well as other similar indices, track the nature of executive policy decisions, thus defining the current environment only. The key implication of this observation is that these indices are rather proxies for policy flows α i p i in the general equation of motion (2) that in any given point in time add to the overall institutional stock. 6 5 This point, for example, has been clearly made by Park (2001) in the context of economic freedom and IPR indices: Another source of confusion arises from not recognizing that indexes of economic freedom and patent rights are flows, not stocks. They reflect the value for a particular year or period, and not the entire history of their respective institutions or experiences. 6 Obviously these index scores are not a perfect measures of flows either. They do capture the general nature of environment, but do not say anything about the pace of a decision making 11

12 In the cross-sectional literature where institutions are measured with a longterm average of an index of a choice the distinction between flows and stocks is not essential since the average over flows can be taken as close proxy to a stock. However, in panel data applications that exploit time-series information in addition to cross-sectional variation this distinction is crucial. In their production function type regression (1) of a change in institutions on education, Glaeser et al. (2004) as well as Acemoglu et al. (2005, 2006) and other followers of this literature measure X 1it with an institutional index value at period t. Since this index value measures the overall quality of policy flows during that period, they are effectively regressing a second difference in institutions, and not a change in institutions, on initial level of schooling. If one wants to regress a change in institutions on initial education, then according to fundamental accumulation equation (2) it would be conceptually appropriate to treat X 1it as policies adding to the stock of institutions and to measure those with respective observed index values. In this context the underlying concept of institutions is effectively a latent variable. Figures 3 and 4 illustrate the differences that arise from such misuse of political index measures for regression (1). Figure 3 plots a change in Polity IV democracy index over against initial average schooling in the top graph, whereas lower graph plots an average over 1990 and 1995 democracy index scores as a flow measure for that time period. Figure 4 features similar plots for executive constraints. 7 The top graph in both figures shows clearly the reason for results documented by Acemoglu et al. (2005): weak insignificant relationship between change in an index of political institutions and initial years of schooling. However, the bottom graph gives a much more optimistic view of existence of strong positive correlation between initial human capital and a change in institutions, whichever whey the causality should run. process. 7 Both Figures 3 and 4 are constructed from Glaeser et al. (2004) database. 12

13 4 Another look at link between education and institutions Most of the empirical applications of the Lipset (1960) hypothesis of educational prevalence routinely use total average years of schooling from Barro and Lee (2000) dataset as a measure of education. This is because traditional models of political participation as Glaeser et al. (2007) mainly assume that it is the general base of education that is important for political institutions and democracy since, as opposed to dictatorships, democracy offers weak incentives to wide base of supporters. Castelló-Climent (2006) has suggested that it is not so much the total stock of education as measured by average years of education, but the more equal distribution of education that triggers the movement towards more democratic societies. Measuring institutions similarly to Acemoglu et al. (2005) and exploiting a system GMM estimator popularized by Blundell and Bond (1998) they argue that this hypothesis is robust and once the distribution of education is controlled for, average years of education loses its significance. Castelló-Climent (2006) base their argument on Lipset s oft-cited book. 8 Lipset clearly indicates towards causal relationship between education and democracy, though only in general terms, acknowledging potential caveats: Dewey [1916] has suggested that the character of an educational system will influence its effect on democracy, and this may shed some light on the sources of instability in Germany. The purpose of German education, according to Dewey, writing in 1916, was one of disciplinary training rather than of personal development. The main aim was to produce absorption of the aims and meaning of the existing institutions, and thoroughgoing subordination to them. This point raises issues which cannot be entered into here, but indicates the complex character of the relationship between democracy and closely related factors, such as education. (Lipset 1960, p. 57) The causal link from education to institutions, though with a reference to technological developments and changing relative prices, finds support also from the writings by North (1981, p. 208): The cumulation in a stock of knowledge [embodied in technological developments and skills of people] has imposed an evolutionary order upon the secular change of political and economic institutions... The complexity of this relationship is largely a matter of empirical testing. Here it is useful to differentiate between growth and level effects of institutions. In the context of North (1981), the incentives for growth effect, or 8...and interestingly on the exact same citation as Acemoglu et al. (2005). 13

14 incremental institutional innovations underlying the secular change in institutions, are shaped by the opportunity cost of constituents. These opportunity costs are much more important for highly educated skilled workers, since they form the part of society who are more prone for research and innovation intensive activities, have potentially more entrepreneurial options, are more likely to be elected to the legislature, and thus would benefit more from better and enforced property rights. Contrary, large pool of more moderately educated workforce might hinder the growth, leaving the nation into the trap of extractive property rights that are oriented towards innovation coping technologies. 9 These considerations suggest that, in empirical applications and depending on objective, it is important to distinguish at least between different levels of education. Such differentiation to a first order captures distributional features of education as well as allows to differentiate between incentives of constituencies. If changes in technology are a major part of institutional innovation, this, as in the example of common law, would cause incremental additions to the stock of operating rules. Thus, if one intends to account for a secular change or the dynamic nature of institutions, higher educational attainment would be more appropriate measure to use. Also it is mostly lower levels of education that to a larger extent can be subject to disciplinary training through which the existing institutions can replicate themselves. In this respect, higher education seems to be more important precondition for the dynamic secular change in democratic institutions to take place, as it should be more personal development oriented and less influenced by current ideologies. The current literature has been more concerned with level effects of institutions - discrete changes from authoritarian towards more democratic environment. These changes occur mostly through domestic uprisings as well as referendums, for which large support of similarly thinking constituents is needed. This reflects the conventional assumption that democracy as opposed to autocracy needs a wide base of supporters, as its incentives are not concentrated to a small elite. In the literature the effect of general educational base on the choice of regime / environment (a scale from autocracy to democracy) has been routinely proxied by total average years of schooling, whereas here the penetration of primary education as a proxy for wider constituency, or distributional measures of Castelló-Climent (2006) seem to be more appropriate. This effect might not be without ambiguities, since wide (primary) educational base can generate large working class that in the history has been one of the preconditions for a spread of socialist/communist ideologies, institutions of which are related closer to dictatorships than democracies. However, this argument loses its strength once one is willing to accept that communist regime with a large working class can be seen as a middle step from complete autocracy to democracy. Rajan and Zingales (2006, p. 8) have put it rather 9 The case in point would be China and India, who are known not to respect intellectual property rights to an extent that more advanced OECD countries do. On another hand, Japan after the second World War was not oriented towards coping technologies, but buying the necessary patent rights. 14

15 nicely: Once we accept that institutions, especially bad ones, may not be very persistent without the underlying power structures holding them in place, it becomes easier to understand why we have seen such extraordinary change in countries that were under the yoke of communism... We would argue that one of the virtues of communism is a very strong emphasis on education, and this creates the broad constituencies that can press for market reforms once the stranglehold of the nomenklatura is broken. Ironically, instead of capitalism containing the seeds of its own destruction, the seeds for flourishing capitalism have been nurtured in the soil of communism. Capitalism may well be the final stage of communism! Finally, when it comes to empirical estimation a word of caution is in order. If at any level of education the aim of the system is the absorption of the aims and meaning of the existing institutions then one should take reverse causality from institutions to education seriously. 15

16 5 Policy Regime Regressions 5.1 A Look at Existing Evidence Table 2 sets the stage by following the developments in the literature, and in parallel discussing the various issues involved. The relationship of interest is the dynamic regression of institutions on lagged level of education of the form: y it = αy i,t 1 + βx i,t 1 + η i + υ it (3) Education is measured by total average years of schooling taken from Barro and Lee (2000). I distinguish between two widely used measures of institutions; in Panel A I follow Acemoglu et al. (2005, 2006) who measured institutions with augmented Freedom House index of Political Rights, in Panel B institutions are measured by Polity IV index of Executive Constraints as advocated by Glaeser et al. (2004). In what follows and keeping in line with existing empirical investigations these index values directly enter the regressions for y it and y i,t 1. Column 1 reports significant positive coefficient for lagged total average years of schooling from pooled OLS estimator. Glaeser et al. (2004) instead reported fixed effects estimates without time controls that yielded qualitatively similar results. Acemoglu et al. (2005), however, argued that estimates without netting out time-effects are misleading since they capture general increase in education and institutions that is not of a causal nature. Columns 2 and 3 present their main specifications that they show to be robust to a number of controls. Within estimator with time-effects gets rid of any signs that higher education might lead to better institutions. Exploiting difference-gmm estimator due to Arellano and Bond (1991) to account for endogeneity in lagged dependent variable makes qualitatively no difference in both panels. With this the human capital prevalence hypothesis had received a serious setback. Traditionally researches have treated (lagged) education as strictly exogenous, an assumption that according to Bobba and Coviello (2007) as well as reasoning in section 4 might be too strong. There are good reasons to believe that past, as well as current, institutional innovations can affect educational choices. In columns 4 and 5 I treat education as predetermined or endogenous respectively, an assumptions that imposes greater restrictions on instrument matrix. However, qualitatively this brings no changes to Acemoglu et al. (2005) conclusions. A number of authors, most notably Caselli and ColemanII (2006) and Bobba and Coviello (2007) have argued that, given the persistence in democracy and education, difference-gmm estimator does not provide adequate identification. If true, then past levels become uninformative of future changes and thus, as demonstrated by Blundell and Bond (1998), the instrumentation in heart of difference-gmm estimator becomes weak and subject to serious finitesample bias. Columns 6-8 follow this suit using system-gmm as an estimation method under different assumptions regarding the available instruments for education. As can be seen, coefficient on education becomes positive and is very precisely measured. 16

17 Finally, before turning to various issues associated with the estimation, in the light of discussion in section 4, in Tables 3 and 4 I rerun the same regressions with average years of higher and primary education respectively. It is interesting to note the differences between the two indices used. Varying the components of education does not make any difference for Freedom House political rights; all results come through as before. However, the regressions with executive constraints prove to be a sharp contrast. Here especially primary education proves to be a rather robust predictor of policy regimes across various estimators. This is in concordance with the earlier hypothesis that it is the wider spread of basic education that is needed for formation of a larger working class, a large constituency that can push through the regime changes. Higher education proves in some cases to be only marginally insignificant. 5.2 Potential Issues There seems to be a number of of issues involved that especially concern the exploitation of a system-gmm estimator. I will consider them in turn. Recall that so called system-gmm refers to the following set of equations: y it = αy i,t 1 + βx i,t 1 + υ it (4) y it = αy i,t 1 + βx i,t 1 + υ it (5) Here the difference equation (4), that is simply the first-difference transformation of equation (3), provides the moment conditions for difference-gmm estimator. To get system-gmm estimator, these are then combined with additional (T-2) linear moment conditions implied by the levels equation (5). As was already mentioned above these additional moment conditions can provide huge efficiency gains when the underlying variables for which internal instruments are needed are close to unit root processes. First, many researchers routinely assume that measures of institutions are persistent, whereas as Glaeser et al. (2004) as well Table 1 and the discussion above has shown, they are rather extremely mean-reverting. This is confirmed by regressions in Table 2: the feasible range for lagged dependent variable as given by within and pooled OLS estimators is roughly in between , far away from a unit root, with all difference-gmm estimates lying within that range. 10 Furthermore, as Monte Carlo evidence in Blundell and Bond (1998) shows, bias in difference-gmm becomes important for values of α in equation (4) around 0.8 and above, whereas the respective estimates in Table 2 as well as in Bobba and Coviello (2007) are well below this bound. In this respect, the 10 Bobba and Coviello (2007) motivate system-gmm on the grounds that difference-gmm produces coefficient estimates for an autoregressive component smaller than the lower bound given by within estimator. They use Acemoglu et al. (2005) database that should give results comparable to those in Table 2 panel A. However, in their specifications both education as well as time-effects are excluded from the instrument matrix that is not entirely proper (firststage) instrumentation strategy. With education and/or time-dummies added to the instrument matrix I am not able to confirm their claim across all possible variations of the size of instrument matrix. 17

18 time-series properties of institutional indices per se do not provide us with any evidence of biasness of difference-gmm estimates. In principle, if assumption of strict exogeneity is dropped, persistence in educational attainment could lead to weak instrumentation problems and thus biased estimates of β in regression (4). For both institutions and education these are overidentifying restrictions that can be tested. Looking at columns 4-5 and 7-8 in Table 2, Difference-in-Sargan statistics never rejects the validity of instruments for education in equations in differences. Oddly, this suggests that even persistence in education is not reason to resort to a system GMM. However, depending on the use of an estimator as well as the treatment of education the number of instruments here can rise close to 100. High number of instruments will make these tests of overidentifying restrictions weak as well as cause overfitting and bias towards the within estimator. To check for these concerns Tables 5 and 6 present the specifications exactly comparable to those in Table 2, but reducing the number of lagged observations available as instruments. In Table 5, the instrument set is restricted to 2 most recent available lags only, whereas in Table 6 the instrument matrix is collapsed instead. The coefficient estimates do not change considerably, suggesting that overfitting is not serious concern. However, with marked differences across institutional measures, the problems with instrument exogeneity became evident. Under Freedom House index overidentification tests are on the margin of rejecting the exogeneity of instruments for education, though at the same time evidence in Table 6 clearly rejects the validity of instruments for levels equation (5) as well. Under Executive Constraints, on another hand, validity of educational instruments in difference equations is still not rejected, whereas the p-values for system GMM levels equations in Table 6, although still above 10%, have dropped dramatically. Such results seem to raise more questions than they answer, though in a nutshell they provide relatively little evidence to favor system over difference GMM on theoretical grounds. Second issue comes back to the observation that what these institutional indices are actually measuring is the flow (or the nature) of policies in equation (2). Thus the correct specification of regression (3) would recognize that the institutional stock is essentially a latent variable and thus the levels equation is not observed. Correct specification of a difference equation would then regress observed institutional indices - that represent flows and hence changes or first-differences in institutions - on a lagged first-difference in education, instead of a lagged level of education. Thus both difference as well as levels equations run in the literature and documented in Tables 2-6 are misspecified. Nevertheless, the levels equation specified in these tables, regressing observed indices on lagged level of education, can still be informative. Ignoring the lagged dependent term, what this regression would actually capture is whether the past levels of education can cause changes in institutional quality - the very question Glaeser et al. (2004) were set to investigate in the first place! Given that system-gmm in Table 2 identifies a positive coefficient on education only trough the level equation and the strong positive relationship shown in the bottom panels of Figures 3 and 4, this represents a good news to those in favor of human capital prevalence hypothesis. 18

19 Careful separation between stocks and flows helps to resolve a third more technical issue. The time-series properties indicate that educational attainment data is clearly non-stationary whereas institutional measures appear to be stationary. Hence, the regression of observable institutional measures on levels of education would essentially look for a relationship between I(0) and I(1) variables that, although meaningful in the current context, might show spurious time-series correlation. Recognizing explicitly that this relationship is essentially of flow-stock is important in suggesting appropriate time-series transformations such that a meaningful (non-spurious) relationship could be estimated. I return to this in later sections... Finally, in order for additional moment conditions provided by the levels equation (5) in system-gmm to be valid, an additional assumption on initial conditions must be valid: E( y i2 η i ) = 0. (6) In a dynamic AR(1) model as given by equation (3) η i determines the level of y it, or institutions, towards what each country is converging to. This assumption essentially requires the initial deviations from that convergent level to be uncorrelated with η i (or each country s convergent level). Technically this rules out the dynamic panel bias, the correlation between the lagged dependent variable and fixed effects in the levels equation (5), where fixed effects cannot be removed by the respective transformation otherwise. In the ARspecification this happens when the model has generated y it long enough prior to the sample period so that the influence of the true start-up of the process has become negligible. However, if the starting point of a sample coincides with the true start up of the process, the restriction (6) as well as the system-gmm estimator will be invalidated. The immediate question is if this assumption is valid for most of the colonized world where the process of institutional development was altered. For example, take Nigeria who in 1960 gained its independence from UK and formed an independent republic. Both Polity IV and Freedom House assign very high scores to Nigeria, reflecting the current direction of policies. However, from the point of view of the process for institutions this represents either a starting point of a new (or perhaps a transition back to a true underlying) model. In any case, the initial 1960 deviations from a convergent level of institutions cannot be assumed to be random. 19

20 6 Reinterpreting the Education-Institutions Relationship 6.1 Levels-to-Levels estimation In this section I treat institutions as a latent variable and, consistent with the discussion in section 3, I interpret existing indices - such as those provided by Polity IV and Freedom House - as flow variables, proxing for policy decisions in (2) and through the assessment of policy defining the political regimes. To provide a rough proxy for institutions I accumulate Polity IV flow measures into a stock according to this equation of motion with 3 percent depreciation. The initial year for accumulation is taken to be 1960 since from there onwards Polity IV coverage increases considerably. Considering the shortness of timedimension (database runs in 5-year intervals) the sample used in regressions covers the maximum period possible from 1960 to Before turning to the estimation results few remarks are in order. First, not all countries included in the base sample are independent in For those few pre-independence observations have been estimated mainly by backward extrapolation. Second, the results are robust to both democracy and aggregate polity indices from Polity IV database as well as to altering the depreciation rate. In case of the former the main results seem to come through even more strongly. In case of the latter the results are strongest with pure stock and weaken, though not considerably, when depreciation rate is raised to 10 percent. Table 7 sets the stage by indicating some of the basic time-series properties of accumulated stock measures for 3 Polity IV indices; in Table 8 I re-estimate the same equations as in Table 2 with an accumulated proxy for an institutional stock as a dependent variable. The overall results are much more encouraging, though the use of this stock measure introduces many estimation issues, rendering the results mostly invalid. Columns 1 and 2 provide basic pooled OLS and fixed-effects estimates, defining as the range for an autoregressive term, whereas Table 7 suggests close to unit root behavior. This is clearly what Blundell and Bond (1998) define as a problematic range for traditional estimators based on first-differencing and their instrumentation with respective lagged levels. In addition, Table 7 suggests that these stock measures of institutions have deeper autoregressive structures that further complicates the instrumentation. Columns 3-5 in Table 8 illustrate the point in the context of difference-gmm estimator. The Arellano and Bond (1991) AR-test in all cases detects 2nd order serial correlation in residuals in levels, making only the 4th and deeper lags of institutional stock variable available to instrument for endogenous lagged dependent variable. Even then the difference-in-sargan test detects problems with the exogeneity of these instruments. The 2nd order serial correlation also complicates the instrumentation for education in columns 4 and 5. For example, treating education as predetermined with respect to institutions allows only the past shocks to be correlated with current education, making the first 11 I interpret the regressions with flows as probability of transition between environments, and regressions with stocks as the true causal relationship between education and institutions. 20