1

Accepted for publication: Acta Oeconomica (2015)

INSTITUTIONAL EFFECTS ON ECONOMIC PERFORMANCE IN TRANSITION: A DYNAMIC PANEL ANALYSIS

Adnan Efendic* University of Sarajevo School of Economics and Business

Geoff Pugh Staffordshire University Business School

INSTITUTIONS AND ECONOMIC PERFORMANCE

This article uses dynamic panel analysis to investigate the relationship between institutional improvement and economic performance in transition countries. The contribution of this paper is two-fold. First, we find that per capita GDP is determined by the entire history of institutional reform under transition and that, conditional on this history, per capita GDP adjusts to recent institutional changes. Moreover, we find that the time-horizon over which we measure institutional change matters, with five-year changes showing the clearest effects on current levels of per capita GDP. Secondly, we address the pronounced methodological heterogeneity of this literature. To compensate for incomplete theoretical guidance from the institutional literature, we draw upon an institutional meta-regression analysis to inform our model specification. Then we check the robustness of our estimates in a variety of ways: in particular, by a simple variant of extreme bounds analysis, in which model diagnostics are of central importance.

Key words: Institutions; economic performance; transition countries; dynamic panel analysis; extreme bounds analysis

JEL classification indices: E11; P30; P39

Acknowledgments

The authors acknowledge a particular debt to Nick Adnett. We also thank two anonymous referees for additional improvements. Any shortcomings remaining are the responsibility of the authors.

* Corresponding address: Trg Oslobodjenja A. I. 1, 71000 Sarajevo, Bosnia and Herzegovina. E-

mail: adnan.efendic@efsa.unsa.ba; A.Efendic@staffs.ac.uk

2

1. INTRODUCTION

Many authors agree that transition is largely a process of institutional change (Cornia Popov 2001; North 2005; Redek Susjan 2005; Eicher Schreiber 2010). Accordingly, institutional economics may be particularly relevant in explaining economic differences among transition countries (TCs). However, the study of institutions in transition is still characterized by empirical gaps that remain to be investigated. Most transition studies report that better institutions were supportive in achieving better economic performance.2 However, while results tend to be qualitatively similar, model specifications and empirical strategies in this literature are diverse. In turn, this reflects a lack of theoretical guidance from the institutional literature. Accordingly, the point of departure for this study is the pronounced methodological heterogeneity of this literature. Our response is to use the results from a meta-regression analysis (Efendic et al. 2011) to inform our literature review (Section 2) from which we derive theoretical and econometric reasons for our model specification (Section 3). From this platform, further features that distinguish the methodology of this study are that: (1) we build on the recent introduction of dynamic panel models in the literature by implementing and discussing the full range of diagnostic tests and checks necessary to ensure the validity of such models (Section 4); (2) we address the potential endogeneity of institutional effects on economic performance; and (3) we undertake extensive robustness checking of our reported results (Section 5), in particular an extreme bounds analysis, which we have adapted for application to dynamic panel models. With respect to the substantive issue, we report new findings on the timing of institutional effects on economic performance (Section 4 and the concluding Section 6).

2. LITERATURE REVIEW

Conventional literature review establishes a general consensus that institutions matter for achieving better economic performance in transition. Although the qualitative findings seem homogenous, a meta-regression analysis (Efendic et al. 2011) applied on institutional studies identified five main sources of heterogeneity in this literature: the dependent variable; measurement of the variable of interest, that is the proxy variable for institutions; model specification; the estimators applied; and the approach to addressing the potential endogeneity of institutions. Since these heterogeneities affect estimates of the effect of institutions on economic performance reported in the literature, our review will focus on these differences.

2 For example, in: Sachs (1996); Havrylyshyn Van Rooden (2003); Fidrmuc (2003); Assane

Grammy (2003); Lane Rohner (2004); Chousa et al. (2005); Redek Susjan (2005); Falcetti et al. (2006); Beck Laeven (2006); Di Tommaso et al. (2007); Marelli Signorelli (2007); Gylfason Hochreiter (2009); Paakkonen (2009); Eicher Schreiber (2010).

3

2.1. The dependent variable: output growth or levels?

The core theoretical institutional literature (North 1990) explains the effect of institutions on economic performance; hence, not specifying precisely the definition and measurement of the explained variable. Lack of theoretical precision has permitted substantial heterogeneity in empirical studies, which variously focus on economic growth or level of economic output variables. In their meta-regression analysis (MRA) of the empirical literature studying institutional effects on economic performance, Efendic et al. (2011) identify 20 studies using output growth and 21 studies using output level. This MRA reports more robust findings of positive and statistically significant institutional effects on output levels than on output growth. Yet, in the transition sub-sample, research has been mainly focussed on output growth, leaving institutional influences on output-levels relatively unexplored.

In addition to the finding that investigating output levels is more likely to reveal institutional effects, and that such investigations have been relatively neglected in the transition literature, there are also substantial theoretical reasons for focussing on output levels as the dependent variable. Some economists, including Knowles Weatherston (2006), Basu (2008) and Easterly (2009), argue that the level of output should be the focus for institutional research. The main rationale for this modelling strategy is that national differences in per capita output levels reflect entire histories of time-varying growth performance. Accordingly, analysing the determinants of differing per capita levels helps to avoid non robust and, hence, spurious explanations that arise from potentially unrepresentative samples of impermanent growth processes (reflecting, according to Easterly 2009, p. 30, .. the possibility that if you get a result associating high growth with a particular country in one period, it is likely to vanish in the following period.). Moreover, focusing on per capita values enables us to take the relative country size into account (Busse Hefeker 2007). Accordingly, our dependent variable will be defined in terms of the per capita output level - the logarithm of GDP per capita ( gdppcln ).

2.2. The independent variable of interest: measuring institutional performance in transition

Institutions are a complex phenomenon and empirical research cannot capture all of this complexity; hence, simplified institutional indicators and proxies need to be used in applied research (Williamson 2000). A huge disparity in using institutional proxies in empirical research, without any consensus on the direction of unification, suggests that a single variable representing institutions is not available (Keefer Knack 1997; Raiser et al. 2001; Shirley 2008). Consequently, the second methodological challenge for empirical research on institutions is to find an adequate quantitative proxy for the performance/quality/efficiency of institutions in transition. Looking at previous transition research, most researchers rely on the European Bank for Reconstruction and Development (EBRD) structural and institutional change indicators as their proxy for institutions (for example, Sachs 1996; Raiser et al.

4

2001; Havrylyshyn Van Rooden 2003; Beck Laeven 2006; Falcetti et al. 2006; Di Tommaso et al. 2007; Marelli Signorelli 2007; Eicher Schreiber 2010). Other authors use different indicators; for example: Redek Susjan (2005), and Paakkonen (2009) employ the Heritage Foundation Index of Economic Freedom as institutional measures; Sonin (2000) uses risk ratings indicators; Chousa et al. (2005) base their institutional variable on the shadow economy; Lane Rohner (2004) and Beck Laeven (2006) use the World Bank Worldwide Governance indicators; Estrin et al. (2009) utilize the Corruption Perception Index obtained from Transparency International, while some authors use specifically designed survey data (e.g. Efendic et al. 2014). Most transition papers are based on aggregated institutional indicators. The first general critique is that the institutional variable in this case is a broad indicator usually composed of sub-indices, which measure different institutional features that might be the product of institutions rather than institutions themselves (Shirley 2008). Conversely, De Haan et al. (2006, p. 182) see this aggregation as an advantage of institutional indices; the authors conclude that those indices are both reliable and useful. Indeed, regressions that include such aggregated institutional variables suggest some important empirical regularities and usually explain a large fraction of economic performance (Shirley 2003). However, Knowles Weatherston (2006) remind us that such institutional indices are mainly used as a proxy for formal institutions, while measuring informal institutions is very difficult. A corollary is that informal institutions are usually not reflected in these indices, and so are relatively neglected in the empirical literature (Efendic et al. 2011a). Another potential shortcoming of these institutional measures is the assumption that the institutional framework among different countries has the same structure and size in relation to the economy. Shirely (2008) argues that much less effort has been directed towards measuring institutions in specific countries. Furthermore, as Havrylyshyn Van Rooden (2003) underline, such indices are based on the judgment of outside experts, which may be subjective and contain perceptions bias. Glaeser et al. (2004) argue that potential subjectivity biased measures raise doubt over causality that goes from variables representing institutions to economic growth, since the institutional indices mainly improve with the level of economic growth (performance). This is a simple but a rather convincing criticism. However, Glaeser at al. (2004) focus on political institutions, while economic along with political institutions, including the interrelationships between them, are crucial for economic prosperity and better performance (Sobel and Coyne 2011; Bjornskov et al. 2014). Moreover, Sobel and Coyne (2011) investigate in particular the issue of stationarity and cointegration of different institutional measures and find that indices of formal political and economic institutions are non-stationary, implying that institutional reforms are indeed permanent. Their finding also implies that, in non-transition countries, central parts of the (non-political) institutional framework are very persistent over time, suggesting that subjective bias is unlikely to be a main concern.3

Conversely, De Haan et al. (2006) argue that even though subjectivity may play a role in constructing institutional indices, such subjectivity is acceptable, since

3 For the last sentence we thank an anonymous referee.

5

discretionary decisions must be made. In line with this argument, Leukert (2005) sees institutions more as a qualitative issue rather than a quantitative one; accordingly, personal opinions about institutions are relevant and welcomed. All in all, any strategy in measuring institutions in transition will have its advantages and disadvantages. However, transition papers are rather consistent in using (EBRD) aggregated institutional indices to proxy institutional performance.

2.3. Model specifications in applied transition research

Analysis of the evolution of economic performance in transition is a very complex task, especially because economic theory provides neither clear guidance nor consensus as to how the transition process should be analysed (Havrylyshyn et al. 1998). In such circumstances, empirical modelling should take into account all possible determinants and transition specifics, which per se raise a number of methodological problems. There is a wide range of empirical specifications utilized to model institutional effects. In some studies, institution(s) is/are the only explanatory variable(s) (although often augmented by the lagged dependent and/or lagged values of the institutional variable); for example, Mauro (1995) and Sachs (1996). Although there is no clear guideline about the specification that should be used in institutional research, this simple bivariate specification is less acceptable than a fully-specified model (Gwartney et al. 2004). Ostrom (2005) suggests that to understand and analyse the processes of structural change of any particular situation, we should include one or more of the underlying sets of variables, which Durham (2004, p. 486) calls mainstream economic controls. Adding one or more standard growth-determining factors to an institutional bivariate specification leads us to some form of the extended production function specification, which integrates growth factors, institutions, and often some other variables. Such specifications, in different forms, can be found in: Keefer Knack (1997); Assane Grummy (2003); Glaeser et al. (2004); Redek Susjan (2005); Eicher et al. (2006); Gwartney et al. (2006); Aixala Fabro (2008) and Paakkonen (2009). Finally, we may identify also many other specifications that include institutions as explanatory variables together with control variables that are not standard production factors. The seminal paper written by Rodrik et al. (2004) may be a good representative (also exploited by Sachs 2003; Easterly Levine 2003; Alcala Ciccone 2004; Jacob Osang 2007) in which authors use institutions, trade integration, and geographical location as explanatory variables of economic development. In transition research, there is no consensus concerning the variables to be included in these regression models (Jutting 2003). However, in studies of economic performance in transition, extended production function specifications are applied by only a minority of researchers (Falcetti et al. 2006; Redek Susjan 2005; Paakkonen 2009) although all these transition studies investigate the effect of institutions on economic growth. Yet, it is quite the opposite for non transition research, which include a good number of extended production function specifications and output-level focused studies.

6

2.4. Estimators used in transition research

Regarding the methodology employed to estimate institutional models, existing empirical research on transition is often based on OLS cross-section analysis, although some research has been based on static panels, while Falcetti et al. (2006), and Paakkonen (2009) and Eicher Schreiber (2010), for example, apply a dynamic model. We argue below that dynamic panel models are a methodological advance in comparison to the cross-section and static panel models applied; accordingly, we discuss these papers. Eicher Schreiber (2010) in their dynamic panel regress GDP per capita growth on institutions for a period of 11 years. The institutional variable is constructed from the EBRD indicators. The authors find significant evidence that institutions influence economic growth per capita in transition. Moreover, by analyzing the dynamic contribution of institutions on growth, Eicher Schreiber (2010) find that sustained institutional change is crucial for economic performance in transition. However, in this research the standard model diagnostics are not reported, thereby raising doubts concerning instrument validity, while their bivariate specification may give rise to omitted variables bias. A more developed specification is applied by Paakkonen (2009) and Falcetti et al. (2006) in which the authors, in addition to the (once) lagged dependent variable and an institutional proxy, use other explanatory variables such as investment and government consumption, and include the interaction of the institutional proxy with some of these variables. These specifications might be considered as more fully specified models. Paakkonen (2009) reports a positive effect of increasing economic freedom on economic growth over the period 1998-2005. Falcetti et al. (2006) employ the same proxy for institutions as Eicher Schreiber (2010) and find that institutions are an important determinant of economic growth in transition.

However, none of these studies report the full range of model diagnostics, as recommended by Arrelano Bond (1991); Bond (2002); Roodman (2009a; 2009b); and Sarafidis et al. (2009), and they leave some important aspects of dynamic panel modelling unexplored. Moreover, the potential effect of time-related shocks in transition is not investigated; the authors do not include all TCs in the sample; and the authors do not investigate the timing of short-run institutional effects. All these shortcomings will be addressed in our modelling procedure (below).

2.5. Addressing the potential endogeneity of institutional effects on economic performance in transition

The problem of the potential endogeneity of institutions is one of the most difficult in empirical institutional work (Nye 2008; Ahlerup et al. 2009). Although institutional economists generally recognize institutions as an endogenous factor in economics some empirical studies do not consider the potential endogeneity problem

7

(in the transition context, this applies to Sachs 1996; Havrylyshyn Van Rooden 2003; Lane Rohner 2004; Redek Susjan 2005; Chousa et al. 2005; Ahlerup et al. 2009). Yet Efendic et al. (2011) find that the conclusions of such studies should be treated with great caution, because of their potential overestimation of the institutional effect on economic performance.

The most widely recognized strategies for addressing the potential endogeneity of institutions are those that derive instruments from historical perspectives (Acemoglu et al. 2001), the geographical environment (Rodrik et al. 2004) or linguistic characteristics (Hall Jones 1999). Yet these instruments developed for global samples typically cannot be applied to sub-samples of countries (Eicher Leukert 2009), in particular to TCs. More promisingly, Falcetti et al. (2006), Paakkonen (2009) and Eicher Schreiber (2010) use internally generated instruments in the context of dynamic panel modelling. Our modelling strategy builds upon this approach.

3. EMPIRICAL MODELLING OF INSTITUTIONAL EFFECTS IN TRANSITION

From our literature review, we conclude that best practice in the investigation of institutional effects in transition is to model output levels rather than growth, to proxy institutional effects using some established index, to estimate a fully-specified dynamic model and to address the potential endogeneity of institutions. In this section, we explain how our approach responds to these requirements.

3.1. Institutional proxy

In establishing our proxy variable for institutional quality, we follow the mainstream transition literature and focus on a broad aggregated indicator of institutional change in transition, which is constructed from the EBRD indices of structural and institutional reforms. In general, this index ranks institutions in transition relative to the standards of the industrialized market economies. Justification for this approach is that transition is in essence a process of transformation from centrally planned towards market oriented economies. Raiser et al. (2001, p. 6) see the EBRD institutional indicators as the best available data on institutional change in transition economies. Arguably, the EBRD institutional indicators trump all other institutional indices that we have identified in the literature for at least two reasons. Firstly, they are transition specific, hence, specially designed for TCs. Institutional reforms in transition includes redefining the role of the state, market and business sectors, which this index is designed to capture. Secondly, observations are available annually from the beginning of transition enabling the longest time-span and the largest number of observations. This (unweighted) aggregated institutional EBRD index is scored from 1.0 (minimum) to 4.3 (maximum); we normalize it to a range from 0 to 1. In this transformation, 0.0 indicates completely state-planned economic institutions while

8

1.0 represents market based standards of economic institutions in developed (OECD) economies (Eicher Schreiber 2010, p. 170). In our initial checking procedure we found that almost all components of the EBRD index are highly correlated with each other, which most researchers also report (Ali 2003; Bevan et al. 2004; Lane Rohner 2004; Di Tommaso et al. 2007; Marelli Signorelli 2007; Eicher Schreiber 2010; Bjornskov et al. 2010). Sobel and Coyne (2011) report that in a transition sample of countries even minor reforms to one institution can be reinforcing and result in subsequent reforms to other institutions, suggesting that institutional changes and reforms are simultaneously and permanently maintained.

Since these sub-indicators may capture similar information coming from different aspects of institutional change, these high correlations are not surprising (Di Tommaso et al. 2007, p. 873). However, the choice of indicators that averaged and aggregated to one institutional proxy raises the question of how to combine them in empirical research on institutional change as an underlying process rather than focussing on just one sub-dimension (Raiser et al. 2001, p. 4). Moreover, multicollinearity might be a serious issue in such analysis. Fortunately, Raiser et al. (2001) have exploited the Multiple Indicator Multiple Cause methodology to control for potential measurement errors in this multidimensional variable, as well as for the problem of aggregation of different components of the EBRD index. The authors find that averaging institutional sub-indices into one composite index is an appropriate measure of institutional change in transition. As a check, we calculated Cronbachs Alpha for the eight components of the EBRD index. With complete observations for each index, Cronbachs Alpha = 0.96, which suggests that the aggregation of the EBRD indices is appropriate for the data used in the regressions reported below. Accordingly, we follow standard practice in institutional literature and so investigate the impact of this composite institutional (proxy) index on economic performance.

An additional issue in defining the institutional variable of interest, which has not been properly addressed in the literature, is the timing of the institutional influence on economic performance. The majority of researchers model contemporaneous or short-run institutional effects (Sachs 1996; Redek Susjan 2005) although some authors argue that the institutional influence might work with lags (Gwartney et al. 2004; Le 2008; Raiser et al. 2001). Accordingly, the timing of institutional effects in transition should be further investigated. To this end, we do not use the current and/or lagged values of the proxy for institutions, which is the measure typically applied in previous studies. Sobel and Coyne (2011) report that institutional proxies, often being nonstationary, should not be used in levels as independent variables explaining some (nonstationary) measures of economic prosperity. Indeed, they should be investigated rather in the form of changes in the series, which is the approach that we apply. Accordingly, we use the change in institutional improvement over a five-year period ( itinst5 ), which we characterise as an influence over the medium-term. This approach to estimating the influence of change in institutions (Sobel and Coyne 2011) over a longer period is recommended by Gwartney et al. (2004) and applied, for example, by Raiser et al. (2001) and Le

9

(2008). We follow this practice to allow for institutional influences on economic performance to take place when sustained over time. In the next section, we advance additional econometric reasons for this practice.

3.2. Base model specification

We specify our base model to estimate the effect of changes in institutional quality on economic performance within a dynamic rather than a static framework. Moers (1999) suggests that the dynamic effect of institutional change may be large in TCs even in the short-run, which should be investigated. Moreover, Sachs (2003) argues, with reference to the work of Barro Sala-i-Martin (1997), that determination of per capita income should be specified in a dynamic model, not in the oversimplified static model. Similarly, Eicher Schreiber (2010) conclude that by exploring the time dimension in a dynamic panel, one can analyze how continuous institutional changes influence economic performance in transition. Consequently, in this dynamic model we allow current economic performance to be influenced by past economic performance, which is a well-known feature of economic processes. In a dynamic panel specification, lagged GDP per capita is an endogenous variable by definition. Hence, we control the endogeneity of this variable in its lagged form as a regressor by using internal instruments; namely, lagged levels and lagged differences. To the economic reasons for specifying the institutional variable of interest as change in institutional improvement over a five-year period, we add two econometric reasons. Firstly, models in which institutional measures are from the current or lagged periods may give rise to spurious regression (Andrzej Cizkowicz 2003; Falcetti et al. 2006); conversely, our specification as a stationary variable - should avoid this problem.4 Secondly, the institutional proxy is constructed in such a way as to eliminate reverse causation and from this perspective may be treated as an exogenous variable. It is not likely that current economic performance may explain past institutional changes; moreover, using a longer period in measuring institutional performance is a good way of attenuating endogeneity (Aron 2000). However, in a dynamic panel specification, endogeneity potentially also arises from correlation of institutional quality with unobservable time invariant influences on economic performance captured by the country-specific error terms (vi). Because not all such unobserved variables can be identified (this depends on the state of theoretical understanding) or, even if identified, measured (this is subject to data limitations), we cannot with certainty control for all such potentially correlated variables. In this case, endogeneity may arise from omitted variables. Accordingly, although we have designed our institutional proxy to be free from simultaneity

4 The Im, Pesaran and Shin panel unit root test confirmed that our dependent variable (lngdppc) is a

trended variable, integrated of order one (the null that this variable contains a unit root in all panels cannot be rejected; p=1.000) and that our independent variable of interest (inst5) is stationary (the null being rejected; p=0.000). This precludes spurious correlation via common statistical generating mechanisms.

10

between economic performance and institutional quality as a source of endogeneity, we do not assume that itinst5 is exogenous.

Initial conditions in individual TCs were different. For both economic and econometric reasons we control for the potential impact of different starting positions on later economic performance. Controlling for initial conditions in regressions with either economic growth or the output level as the dependent variable is established practice (Fidrmuc 2003; Havrylyshyn Van Rooden 2003; Beck Laeven 2006). We argue that controlling for initial conditions is important for four main reasons. a) The economic theory of conditional convergence stresses the importance of

controlling initial conditions in such economic models. b) Per capita GDP on the eve of transition may be an aggregative indicator of

relative capability for transformation and performance during transition (in which case, we anticipate a positive sign).

c) Specification with initial per capita income is the solution to a potential inconsistency between our dependent variable and our independent variable of interest: whereas specifying economic performance in terms of levels of per capita GDP captures the result of the entire history of time varying growth performance, our institutional variable of interest by definition measures changes confined to the transition period. However, initial (1989) per capita income controls for the influence of the entire process governing economic performance up to shortly before our sample period. Hence, effects of other independent variables are estimated net of pre-transition influences. This ensures that our dependent variable is subject to further influences predominantly in the period during which our independent variable of interest is able to exert influence.

d) Specification with initial per capita income helps to address potential endogeneity associated with omitted variables. Institutional quality may be correlated with time invariant unobservable influences on economic performance captured by the country-specific error terms (vi). Consequently, specifying our model with initial per capita income is to displace an important but otherwise unobserved time invariant influence from the error term (the vi) into the observed systematic part of the model (Roodman 2009a). Hence, there is less likely to be correlation between our independent variable of interest and unobserved time invariant influences on our dependent variable. Conversely, omitting initial conditions may introduce unmodelled persistence into our model. If not included in the observable part of the model, the influence of initial conditions will be displaced into the group-specific error term where it may be a source of both autocorrelation and endogeneity, either of which might invalidate estimation of a dynamic model by the General Method of Moments (Bond 2002; Roodman 2009a, 2009b).5

Finally, for similar reasons as apply to our institutional variable, initial conditions proxied by per capita GDP from 1989 cannot be subject to endogeneity arising from

5 Evidence for the former was found when excluding our proxy for initial conditions from the model

reported in Table 1; namely, the m2 test yielded a rejection of the null of no second-order serial correlation in the differenced residuals. In contrast, the m2 test reported in Table 1 supports non-rejection of the null.

11

simultaneity effects. By definition, our initial conditions predate the transition period. In addition, correlation between initial conditions and unobserved time invariant influences in the country-specific error terms (vi) is unlikely to be sufficiently substantial to give rise to endogeneity bias. Given that the purpose of transition was to bring about a profound structural break in economic, social and cultural development, we assume the pre-transition time invariant components of the country-specific error terms (vi), with which initial conditions may have been correlated, to be sufficiently different from the post-transition components of the vi for correlation between these and the initial conditions not to be a problem in practice. For example, many of the countries giving rise to pre-transition country-specific effects no longer exist. For this reason, we treat initial conditions as an exogenous variable. We follow some authors and include variables commonly used to control for stabilization policies in transition, which may influence economic performance (Havrylyshyn Van Rooden 2003; Chousa et al. 2005; Redek Susjan 2005; Falcetti et al. 2006; Paakkonen 2009): namely, inflation; the budget deficit; domestic investment; and foreign direct investment. In our specification, we follow existing practice and treat these control variables as exogenous. Over the last twenty years TCs have been going through similar reforms, although with different sequences and speeds. Hence, it is possible that those countries suffered some universal time-related shocks. Moreover, some TCs experienced economic, financial, and political integration or disintegration (particularly the ex-Yugoslavian transition economies), which implies possible time-related interdependencies between countries (De Hoyos Sarafidis 2006). Hence, we include in our specification time-dummy variables in order to control for potential common time-related shocks.

Hence, our model specification has the following form:

itiit

ittitititit

vu

uYXinstgdppcgdppc

+=

+++++=

.5lnln 1

Specification (1)

where i=1, . . . , 29 indexes the TCs and t=1992, . . . , 2007 indexes the 16 years in the sample.6 The dependent variable in Specification 1 is the natural logarithm of GDP per capita denoted as itgdppcln . 1ln itgdppc is the dependent variable with a

6 An important issue in any transition research is the observed time period, hence, sample size. Some

authors, for example Falcetti et al. (2006), estimate their models starting from an earlier period of transition (1989) but not with the full sample (3 SEE countries are omitted). The supporting argument is that some countries (Central European) made significant progress in transition reforms during the initial period (1989-1992). However, the data for this initial period are not reliable and not available for all transition economies. Accordingly, we follow the advice of other authors (Beck Laeven 2006) and rely on data from a more stable period of transition (1992-2007), which entails the advantage that our model can be estimated for the full sample of transition countries (29) with correspondingly more observations (325).

12

one-year lag while estimates its effect on the current value of the dependent variable. is the regression intercept; itinst5 , the variable of interest is the difference in the institutional index over a five-year period where measures the effect of institutions on the dependent variable. itX is a 1k vector of k control variables identified as important co-determinants of economic performance in transition, which includes: domestic investment proxied by the gross capital formation as a percentage of GDP ( itinvest ); foreign direct investment (FDI) inflow measured as a percentage of GDP ( itfdiper ); budget balance measured as a percentage of GDP (

itbudget ); the inflation rate proxied by the annual rate of change of the consumer price index ( itcpi ); and, finally, initial conditions proxied by GDP per capita (Purchasing Power Parity, income per capita in 1989 US dollars in logarithmic form:

iinitialln ). is a k1 vector of parameters to be estimated. tY is a vector of time dummies to be estimated (t= 1993 . . . 2007). Finally, uit is a composed error term, made up of two components: vi the group-level effects, which control for all unobserved influences on countries economic performance that can be assumed constant (or, at least, slowly moving) over the sample period; and it the observation-specific error term. The strategy of specifying independent variables as percentage changes or as ratios to GDP ensures their stationarity and thus precludes spurious regression (Redek and Susjan 2005). Further explanation of the variables, including data sources, is available in Appendix 1.

In addition to the econometric advantages discussed below, we argue that a dynamic specification is particularly well suited to analyse the impact on per capita GDP of institutional reform under transition. Our dynamic model estimates the short-run effect on economic performance of the most recent medium-term changes in institutional quality conditional on the effects of the entire history of institutional reform under transition. We demonstrate this feature by simplifying our specification, while preserving essentials: in equation 2, is the level of per capita GDP of country i in year t, is the change in institutional quality in the previous five years, and is the usual error term. Starting with our simplified specification in equation 2, we repeatedly substitute for the lagged dependent variable.

Substitute for in (2):

= + + (2)

= + + (3)

Substitute (3) into (2)

= + + + + (4)

Substitute for in (4):

13

= + + (5)

Substitute (5) into (4)

= + + + + + +

Gather terms

= + + + + + + (6)

and so on.

By repeated substitution, we demonstrate that dynamic specifications, through the lagged dependent variable, contain the entire history of the independent variables. In equation 6, we find that current GDP per capita is influenced not only by the most recent institutional changes () but also by the cumulated effects from institutional changes one period back () and two periods back (), although these persistence effects attenuate the more remote the period (shown by the increasing exponent on ). Further substitutions demonstrate that our dynamic specification includes the whole history of institutional reform that influences the current level of per capita GDP. By taking this history into account, we are able to identify the additional short-run effects on per capita GDP of recent - medium-term -institutional changes. In turn, these are informative about the process of adjustment of per capita GDP to institutional change. By taking account of the effects of all past institutional changes (reforms) together with estimating the effect of current adjustment to the most recent change, our dynamic model enables the level of per capita GDP to be explained by changes in institutional quality, which is consistent with our earlier discussion motivating the choice of level of per capita GDP as our measure of economic performance. To anticipate the estimates reported below, because the effects of current adjustment do not induce further rounds of effects through time (shown by the non-significance of the long-run coefficient on institutional change), the effects of institutional change on per capita GDP are fully accounted for by the history of institutional change and current adjustment.

Finally, following good practice guidelines suggested by a number of authors, notably Roodman (2009a; 2009b), in Appendix 2 we explain our preference for a dynamic panel model estimated by the System General Method of Moments (SGMM).

14

4. EMPIRICAL FINDINGS

4.1. Model diagnostics

Specification 1 is estimated by SGMM and implemented by xtabond2, a user-written programme for STATA 10 and later versions (Roodman 2009a). The estimated model is for the period 1992-2007 and covers the set of 29 TCs according to the EBRD definition (Kosovo and Turkey are not part of this sample). The results are reported in Table 1.

Table 1 Base model - SGMM dynamic panel two-step robust estimate

The dependent variable is the natural logarithm of GDP per capita in current US$ (Lngdppc)

PREFERRED MODEL: inst5 exogenous

ALTERNATIVE MODEL: inst5 endogenous

variables (SHORT EXPLANATION OF VARIABLE)

COEFFICIENTS COEFFICIENTS

constant (INTERCEPT TERM)

-0.220 (-0.59)

-0.488 (-0.89)

lngdppcL1. (LAGGED DEPENDENT VARIABLE, 1st LAG)

0.913 *** (10.88)

0.880 *** (9.69)

inst5 (INSTITUTIONS, 5 YEAR DIFFERENCE)

0.403 ** (2.28)

0.438 ** (2.04)

cpi (INFLATION, ANNUAL AVERAGE IN %)

-0.001 (-0.78)

-0.0001 (-0.59)

budget (BUDGET DEFICIT, % GDP)

0.001 (0.14)

-0.001 (-0.14)

fdiper (FDI INFLOW, % GDP)

-0.003 (-1.57)

-0.003 (-1.61)

invest (DOMESTIC INVESTMENT, % GDP)

0.003 (1.37)

0.003 (1.12)

lninitial (INITIAL CONDITIONS, GDP PPP 1989)

0.129 (1.10)

0.193 (1.34)

Year dummies for 1996 to 2007: _Iyear_1996 -.253 **

(-2.61) -0.347 ** (-2.31)

_Iyear_1997 -.351 *** (-3.84)

-0.380 *** (-3.56)

_Iyear_1998 -.331 *** (-3.22)

-0.341 ** (-2.66)

_Iyear_1999 -.420 *** (-3.90)

-0.435 *** (-3.78)

_Iyear_2000 -.290 ** (-2.69)

-0.349 *** (-2.98)

_Iyear_2001 -.225 ** (-2.31)

-0.268 *** (-2.78)

_Iyear_2002 -.196 ** (-2.15)

-0.216 ** (-2.28)

_Iyear_2003 -.115 (-1.45)

-0.131 (-1.53)

_Iyear_2004 -.087 (-1.45)

-0.109 * (-1.76)

_Iyear_2005 -.086 ** (-2.13)

-0.111 ** (-2.66)

_Iyear_2006 -.067 ** (-2.51)

-0.089 *** (-3.22)

Notes: *; **; *** denotes test statistic significance at the 10%, 5% and 1% levels respectively. T-statistics (in parentheses) computed from cluster-robust SEs. Source: Authors calculations using STATA 10.

15

Model diagnostics

Number of observations 325 325 Number of groups (countries) 29 29 Number of instruments 41 56 F- test of joint significance: Ho: The estimated coefficients on the independent variables are jointly equal to zero

F (18, 28) = 2,310 Prob > F = 0.000

F (18, 28) = 2,775 Prob > F = 0.000

Arellano-Bond test for AR(1) in first differences: H0: There is no first-order serial correlation in residuals

z = -2.67 Pr > z = 0.008

z = -2.48 Pr > z = 0.013

Arellano-Bond test for AR(2) in first differences: Ho: There is no second-order serial correlation in residuals

z = -1.78 Pr > z = 0.075

z = -1.72 Pr > z = 0.085

Hansen J-test of overidentifying restrictions (a check that the overall model specification is valid): H0: all overidentifying restrictions (all overidentified instruments) are valid (exogenous)

chi2 (22) = 14.44 Prob > chi2 = 0.885

chi2 (37) = 14.57 Prob > chi2 = 1.000

Difference-in-Hansen tests of exogeneity of GMM instrument subsets:

Hansen test excluding the differenced instruments on the levels equation a test of the validity of the instruments on the differenced equation: H0: instruments on the differenced equation are exogenous (valid)

chi2 (10) = 12.32 Prob > chi2 = 0.265

chi2 (26) = 10.80 Prob > chi2 = 0.996

Hansen test excluding SGMM instruments (the differenced instruments on the levels equation); in effect, a test of system versus difference GMM: H0: GMM differenced-instruments on the levels equation are exogenous

chi2 (12) = 2.12 Prob > chi2 = 0.999

chi2 (11) = 3.77 Prob > chi2 = 0.976

Hansen test excluding the instruments on the lagged dependent variable: H0: all other instruments inst5 and the exogenous variables are exogenous (valid)

Not applicable

Chi2 (36) = 16.51 Prob > chi2 = 0.998

Difference-in-Hansen tests of exogeneity of standard IV instrument subsets:

Test of the joint validity of all GMM instruments: H0: GMM instruments without IV instruments are exogenous

Chi2 (4) = 4.21 Prob > chi2 = 0.378

Chi2 (20) = 14.54 Prob > chi2 = 0.802

H0: Standard IV instruments are exogenous and they increase the Hansen J-test

chi2 (18) = 10.22 Prob > chi2 = 0.924

chi2 (17) = 0.03 Prob > chi2 = 1.000

Source: Authors calculations using STATA 12.

The validity of the obtained results in SGMM depends on the model diagnostics. Compared to Ordinary Least Squares (OLS), SGMM does not assume normality and it allows for heteroskedasticity in the data. Dynamic panel models are known for endemic heteroskedasticity of the data, which can be addressed (Baltagi 2008). Accordingly, we report two-step estimates that yield theoretically robust results (Roodman 2009a). Moreover, we apply the two-step estimator to obtain the robust Sargan test, in other words, the (robust) Hansen J-test, which is not available in one-step estimation. A small panel sample may produce downward bias of the estimated asymptotic standard errors in the two-step procedure (Baltagi 2008, p. 154). As a

16

remedy we report corrected results by implementing the Windmeijer correction (Windmeijer 2005).

The SGMM approach assumes that the applied instruments in the model are exogenous. Consequently, an important procedure in testing the statistical properties of this model is testing the validity of instruments, which requires testing for the presence of first- and, in particular, second-order autocorrelation in the error term. Moreover, SGMM requires the steady state assumption throughout the analyzed period (Roodman 2009a), which also needs to be investigated. The results of the relevant statistical tests and checks are as follows: a) According to Arrelano Bond (1991), the GMM estimator requires that there is

first-order serial correlation (m1 test) but that there is no second-order serial correlation (m2 test) in the differenced residuals. As we see from Table 1, these tests support the validity of the model specification.

b) The Hansen J-statistic tests the null hypothesis of valid overidentifying restrictions, or in other words, validity of instruments (Baum 2006). According to Baum (2006, p. 201), the Hansen J- test is the most commonly used diagnostic in GMM estimation for assessment of the suitability of the model. The Hansen test of overidentifying restrictions does not reject the null at any conventional level of significance (p-value=0.885); hence, it is an indication that the model has valid instrumentation.

c) The Hansen J-test evaluates the entire set of overidentfying restrictions/instruments. It is also important to test the validity of subsets of instruments (levels, differenced, and standard IV instruments). For this purpose, one can use a difference-in-Sargan/Hansen test, also known as the C-test (Baum 2006). The null hypothesis of the C-test is that the specified variables are valid instruments. As we see from Table 1, we cannot reject the null hypothesis of the exogeneity of any GMM-instruments used, or of the validity of the standard IV instruments.

d) Sarafidis et al. (2009) utilize a combination of the m2 and difference-in-Hansen tests for testing cross-section dependence. This approach examines whether any error cross section dependence remains after including time dummies in the model (p.149). The null hypothesis of this test is that the cross section dependence is homogenous across pairs of cross section units. In the reported model diagnostics, the m2 statistic is satisfactory with respect to the null, while the difference between the Hansen statistics for the full set of instruments available and for each of the various subsets of instruments is not sufficiently large to reject the null of homogenous cross-section dependence (De Hoyos Sarafidis 2006, p. 484). Conversely, if we run the same regression without time dummies the model diagnostics are much worse (particularly noteworthy is the deterioration of the m2 test), suggesting the presence of unmodelled cross-section dependence (Sarafidis et al. 2009). Hence, inclusion of time-dummies in our specification improves the model diagnostics by removing universal time-related shocks from the error term.

e) The check for the steady state assumption suggested by Roodman (2009a) can be also used to investigate the validity of instruments in SGMM. This

17

assumption requires a kind of steady-state in the sense that deviations from long-term values are not systematically related to the group-specific effects in the error term (vi). This assumption requires that the estimated coefficient on the lagged dependent variable in the model should indicate convergence by having a value less than (absolute) unity (Roodman 2009a, p. 114), otherwise SGMM is invalid. The estimated coefficient on the lagged dependent variable is 0.9, which is consistent with the steady-state assumption. The second condition that Roodman (2009a) suggests is that the convergence process must not be correlated with the fixed effects (the vi), which has been addressed by controlling for initial conditions in the model.

f) Bond (2002) suggests additional investigation of the dynamic panel estimates validity by checking whether the estimated coefficient on the lagged dependent variable lies between the values obtained from OLS and FE estimators, which is confirmed in our model (the following values are obtained: OLS=0.98 > GMM=0.91 > FE=0.60).

g) Roodman (2009b) strongly suggests that one should report the number of instruments used, since dynamic panel models can generate an enormous number of potentially weak instruments that can cause biased estimates. There are no clear rules concerning how many instruments is too many (Roodman 2009b), but some rules of thumb and tell-tale signs may be used. First of all, the number of instruments should not exceed the number of observations, which is the case here (41 instruments < 325 observations). Second, a tell-tale sign is a perfect Hansen J-statistic with the p-value=1.00. At the same time, the p-value should have a higher value than the conventional 0.05 or 0.10 levels; at least 0.25 is suggested by Roodman (2009b, p. 142). In our model, the Hansen J-test reports a p-value=0.88, which satisfies both rules. We estimated a number of other regressions by increasing or decreasing the number of instruments, using Roodmans (2009b) collapse command for decreasing the number of instruments, but any other restrictions worsen the model diagnostics.

h) The F-test of joint significance reports that we may reject the null hypothesis that the estimated coefficients on the independent variables are jointly equal to zero (p=0.000).

Considering together the various diagnostic tests and checks that have been conducted, there is sufficient evidence to satisfy the key assumptions of SGMM estimation and to conclude that this model is an appropriate statistical generating mechanism.

4.2. Addressing endogeneity

As we explain above, in a dynamic panel specification endogeneity potentially arises from omitted and typically unobservable time-invariant variables captured in the country-specific component of the model error term (vi). Taking a differenced value of our institutional variable of interest may help to reduce such correlation and so minimize the effect of endogeneity. However, to investigate the potential endogeneity of our institutional variable of interest, we instrument inst5 in the same

18

manner as the endogenous lagged dependent variable. We continue to instrument the lagged dependent variable minimally; however, we find that a larger than minimum, but fewer than maximum, number of available instruments are necessary to estimate the effect of inst5 with acceptable precision. Considering the model diagnostics, the m1 and m2 tests are consistent with instrument validity (respectively, p=0.013 and p=0.085). However, because we have a relatively small sample at our disposal, the Hansen test is being used to assess more overidentifying restrictions (37, compared to 22 when inst5 is assumed exogenous) without any increase in information, and thus has reduced power to reject the null of instrument validity. Accordingly, the overall Hansen test with p=1.00 may indicate a problem of too many instruments (Roodman 2009b). To investigate this possibility, Roodman (2009b) recommends using the difference-in-Hansen tests, in order to gain statistical power by focusing on the instruments of greatest concern. First, we consider the test on the differenced instruments for the levels equation, which also constitutes a test of the validity of system versus difference GMM estimation as well as of the steady-state assumption of system GMM; there is a clear non rejection of the validity of this group of instruments (p=0.976). In addition, the Hansen test excluding this group a test of the validity of all the other instruments fails to reject the null of instrument validity while reducing the p-value to a little below the tell-tale value of 1.00 (p=0.996). Secondly, a similar result is obtained from the Hansen test excluding the instruments on the lagged dependent variable; in other words, a non rejection of the validity of all other instruments (p=0.998). Finally, the joint test of the validity of the instruments on both the lagged dependent variable (endogenous by definition) and on inst5 (potentially endogenous) yields a non rejection with p=0.802. These difference-in-Hansen tests yield p-values in the range suggested by Roodman (2009b) (0.25p

19

persistence effect (0.913) together with its high level of statistical significance suggest that the current level of per capita GDP reflects the entire history of the process by which it is determined, which includes all previous institutional developments (Greene 2008, p. 469).8 Conversely, the estimated models reported in Table 1 suggest that the impact of current developments on current per capita GDP is limited. However, recent institutional improvement is an exception; this does have an effect on current per capita GDP.

Our variable of interest (inst5) is statistically significant and exerts an economically substantial influence on economic performance. We estimate a dynamic panel model in a Log-Lin form. Hence, a ten percent improvement in institutions over the period of five years is associated, on average, with a 4.03 percent increase in the current GDP per capita level. This implies that as institutions improve so increasingly large absolute improvements are needed to yield a given increase in GDP per capita. For example, very poor institutions with an index of 0.1 require an absolute improvement of only 0.01 to give a percentage improvement of 10 per cent, while excellent institutions with an index of 0.9 require an absolute improvement of 0.09 i.e. almost to perfection to give a similar percentage improvement. In other words, absolute improvements in institutional quality are subject to diminishing returns. Indeed, intuitively, this must be the case; for as institutional quality approaches the ceiling of one, so the potential for institutional improvement to raise economic performance is reduced. It is instructive to compare the estimated short-run impact effect reported in Table 1 and the long-run cumulated impact reported in Table 2.9 The former is statistically significant whereas the latter is not. This suggests that recent institutional improvement sustained over the medium term (five years) adds to current per capita GDP but that this is a current impact effect only, hence does not cumulate thereafter into a larger long-term effect.10 After five years, the effects of institutional improvement are discernable in our sample in a higher level of per capita GDP and thus a higher platform for all future economic activity. In that sense, the benefits of institutional improvement are long-run (indeed, permanent). However, the combination of a statistically significant short-run (impact) coefficient and an insignificant long-run coefficient suggests a once-and-for-all economic performance effect from institutional improvement over the medium term. In sum, institutional change previously sustained over the medium term changes the current level of GDP per capita and thus sets a new starting level for the future evolution of GDP per capita.

8 Greene is definitive on this point (2008, p. 469): Adding dynamics to a model creates a major

change in the interpretation of the equation. Without the lagged variable, the independent variables represent the full set of information that produce observed outcome yit. With the lagged variable, we now have in the equation the entire history of the right-hand-side variables, so that any measured influence is conditional on this history; in this case, any impact of (the independent variables) xit represents the effect of new information. 9 Papke Wooldridge (2004) provide an explanation of how to obtain both the coefficient and the

standard error for the long-run effect in a dynamic panel data model. 10

Mathematically, the long-run coefficient is the value after an infinite number of periods into the future.

20

The coefficient on the lagged dependent variable captures the entire historical process, including institutional development, culminating in the current level of per capita GDP. The coefficient on the institutional change variable measures the additional impact of recent medium-term changes in institutional quality on current GDP. Accordingly, our model encompasses both the effects of institutional reform on per capita GDP during the history of transition and the additional effects of the most recent changes. In comparison to some other transition panel models (Redek Susjan 2005; Falcetti et al. 2006; Eicher Schreiber 2010), in our model the institutional variable in the current or previous period does not appear as significant, suggesting that if institutional change does influence economic performance then it does so only when sustained over a longer but medium-term - period. In other words, the time-horizon over which institutions act in transition does matter. Hence, an improvement of institutions in transition would not come as a stimulus to economic performance overnight. Similar findings are presented by Gwartney et al. (2004, p. 231) in their non transition research, according to whom a time period of 5 to 10 years is necessary for the effects of an improvement in the quality of a countrys institutions to be registered fully. The time-dummy variables used to capture universal time related shocks in transition over the observed period are mainly significant. We do not attempt to explain the reasons for such results, since this is not a primary interest. However, mainly significant time dummies do suggest the presence of time-related cross-country shocks although these may not be specific to transition or, using econometric vocabulary, cross-sectional dependence in the residuals. Since the other variables in the model are not our primary interest and are not estimated with conventionally acceptable levels of precision we will just briefly comment on them. FDI inflow as a percentage of GDP in per capita terms (fdiper) has a negative effect on GDP per capita in the current year, while domestic investment (invest) proxied by gross fixed capital formation appears as a positive influence on GDP. However, if we allow FDI or domestic investment to influence economic performance with two lags, they appear as significant and positive influences on economic performance. In these models, the institutional variable remains statistically significant with almost the same magnitude, although the model diagnostics substantially worsen (most likely reflecting degrees of freedom lost by lagging these variables). Since, these variables are not of primary interest, we report only the base specification with better model diagnostics. Finally, better initial conditions (lninitial) in 1989 have a positive sign suggesting an advantage for those TCs with higher GDP per capita in 1989. Our findings on the non significance of budget balance, inflation and inward FDI are similar to those of Redek Susjan (2005); the finding on FDI inflow is also consistent with Carkovic Levine (2005).

Table 2 Long-run effect of changes in institutions on economic performance Variable Long-run coefficient SE t-statistic P>|t| inst5 4.64 5.80 0.80 0.431 Source: Authors calculations in Stata 10

21

5. SENSITIVITY ANALYSIS OF THE PREFERRED MODEL

5.1. General sensitivity checking

We conducted a range of robustness checks. These checks investigate the sensitivity of our results to: different time-horizons over which recent institutional changes influence economic performance; the inclusion of dummy variables for EU integration and different groups of TCs; the use of external instruments for institutional influence; and different endogeneity assumptions, including a systematic robustness analysis of the chosen specification. Since we are limited with respect to space, we report the main findings only. If we use a four-year difference of institutional quality as the explanatory variable the results are quite consistent, but the model diagnostics are weaker. If we further decrease the difference to three years, the variable of interest becomes insignificant while the model diagnostics weaken further. Similarly, increasing the difference to six or seven years likewise resulted in unacceptable model diagnostics, while the institutional proxy proved to be insignificant. All in all, the most valid results are obtained in the preferred model, suggesting that the time-horizon over which institutional performance is measured greatly affects conclusions concerning the determinants of economic performance in transition. Countries status with respect to the process of EU integration may be also an important explanatory variable in explaining economic performance and institutional effects in transition (Chousa et al. 2005; Di Tommaso et al. 2007). After including an (exogenous) EU dummy variable (the base category is non-EU TCs) the model diagnostics were a little worse than those of the base model, while the estimated effects of other variables remain much the same regarding sign, magnitude, and significance. The EU dummy has a positive sign but was not significant at conventional levels of significance. Because those TCs that entered the EU tend to have the best economic and institutional performance, this dummy variable may potentially be endogenous. Accordingly, we treat the EU dummy as a predetermined variable, using the standard SGMM instruments. In addition, following Di Tommaso et al. (2007), we instrument the EU dummy using the geographical distance from Brussels as an external instrument. However, the effect of EU membership still proves insignificant, while model diagnostics weaken further. If we control in our model specification for different clusters of countries (EU, SEE, and CIS transition economies), the model diagnostics worsen compared to those of the base-line model, while none of these dummy variables proves to be statistically significant. Hence, we do not identify differences in the model between different clusters of TCs. As part of our investigation, we estimated the baseline model augmented by interactions between institutions and domestic/foreign investment. We did not find any significant interaction effects, while the model diagnostics in these cases worsened.

22

We estimated a number of other regressions with the institutional variable from the current period as well as with lags, in each case instrumenting them with additional external instruments to be found in the literature: years under communism (commun); war (war); distance (distance); EU membership (eu); and fractionalisation by religion (chrprob). However, in all cases the model diagnostics were inappropriate, while the institutional variable did not appear as significant. Accordingly, we could not identify precisely the current or lagged influence of the institutional variable on economic performance, which again confirms the key findings on the importance of how we measure the timing of institutional effects.

5.2. Extreme bounds analysis

Institutional theory is not explicit about what variables constitute the essential core of an appropriate empirical specification. Facing this challenge, authors often use the specific-to-general approach in specifying their institutional models (Klomp De Haan 2009) or investigate more specifications with sets (vectors) of different variables (Blume Voigt 2011). In our case, specifying a model with all potentially important variables identified in the transition research is incompatible with estimating a SGMM model on the size of sample available. Yet estimation on different groups of variables yields widely varying results; as reported above, Efendic et al. (2011) find that the link between institutions and economic performance was conditional on, amongst other sources of heterogeneity, empirical specifications. Accordingly, our preferred specification must be further investigated in order to check its robustness.

In order to assess our specification we conducted a variant of Extreme Bounds Analysis (EBA), which we adapted to take account of model diagnostics (more on EBA can be found in: Leamer 1983, 1985; Leamer Leonard 1983). EBA is an econometric methodology to assess whether minor changes in the list of independent variables alter the main conclusions of the model (Leamer 1983; 1985). It is also a test of whether some doubtful omitted variables truly do not belong to the model; if so, then the base specification should produce better estimates (Leamer Leonard 1983).

The key step in the EBA procedure is to estimate the base specification extended for all possible combinations of up to three variables (henceforth, EBA models). For each estimated EBA model, one investigates whether the coefficient on the variable of interest,

mj , where m indexes the coefficient estimated on the variable of interest (inst5) in j regressions, remains statistically significant and of the theoretically predicted sign. The extreme bounds refer to the highest and the lowest values of

mj

obtained from its standard error mj in the EBA model, according to the following formulas: Lower bound = mjmj 2 ; and Upper bound = mjmj 2 + . Researchers(s) report the lower and upper bounds of

mj and assess whether the

23

coefficient of interest is likely to be zero (EBA test is not passed) or not (EBA test passed) (McAleer et al. 1985).

One shortcoming of EBA is lack of clear guidelines about the diagnostics that should be investigated for the EBA models. Different authors have considered some model diagnostics issues, but without achieving consensus. Accordingly, in this EBA we use an important additional criterion. Namely, we refer to the model diagnostics, since SGMM works only under arguably special circumstances (Roodman 2009b, p. 156). Hence, the final judgment on EBA models will be based on the standard EBA tests suggested by Leamer (1983; 1985) but augmented by the model diagnostic tests, establishing in that way the more systematic approach to EBA evaluation suggested by McAleer et al. (1985, p. 306).

There is no clear rule as to which variables should be considered in EBA. Rather, variable selection depends on theory but also on researchers judgments of the potentially important variables (Leamer Leonard 1983; McAleer et al. 1985) as well as on data availability (Sala-i-Martin 1997). Initially, as relevant variables we consider those that are already exploited in the empirical institutional research on TCs. Using this criterion, we come to a set of more than 20 variables that were used in different studies focused on either output growth or output levels. However, some of the variables provide similar information to the existing variables in the base specification, and so are not interesting for our EBA. Considering all the potentially interesting transition variables, our final list includes nine variables (with very short descriptions): commun (number of years under communism); war (number of years during the last two decades in which a particular country endured military conflicts); industry (percentage share of industry in GDP); distance (distance of the capital city from Brussels); trade (total trade share as a percentage of GDP); eu (dummy variables for TCs that are EU members); chrprob (probability that a randomly chosen citizen is Christian); popgr (average annual population growth in percentages); and current (current account deficit as a percentage of GDP). Importantly, none of these examined variables is highly correlated with either our variable of interest (inst5) or with the other variables. This suggests that multicollinearity should not be a significant problem when using these variables together in the same regression.

Using those nine variables in different combinations we estimated 129 EBA regressions. Table 3 reports results from the twenty-four EBA regressions that yield adequate model diagnostics.

24

Table 3 Summary findings from the Extreme Bounds Analysis (EBA) Sets of up to three variables used in the EBA models

Coeff. (inst5)

mj

Lower bound

Upper bound

EBA test (+/-)

Significance test

Sign test (+/-)

Robust/ fragile (+/-)

Trade 0.36 0.06 0.66 + + + + Chrprob 0.47 0.13 0.81 + + + + Popgr 0.42 0.04 0.80 + + + + trade chrprob 0.44 0.08 0.80 + + + + trade popgr 0.43 0.07 0.79 + + + + trade current 0.41 0.05 0.77 + + + + eu popgr 0.51 0.13 0.89 + + + + chrprob popgr 0.47 0.11 0.83 + + + + popgr current 0.49 0.29 0.69 + + + + commun war trade 0.51 0.07 0.95 + + + + commun distance eu 0.58 0.12 1.04 + + + + war industry trade 0.61 0.05 1.17 + + + + war distance trade 0.49 0.11 0.85 + + + + war distance popgr 0.33 - 0.85 1.51 - - + - war trade current 0.59 0.10 0.99 + + + + war popgr current 0.61 0.23 0.99 + + + + industry trade eu 0.45 0.05 0.85 + + + + industry eu current 0.46 0.10 0.82 + + + + distance eu popgr 0.53 0.11 0.95 + + + + trade chrprob popgr 0.42 0.02 0.82 + + + + trade chrprob current 0.53 0.13 0.93 + + + + trade chrprob popgr 0.42 0.02 0.82 + + + + eu chrprob popgr 0.49 0.09 0.89 + + + + eu chrprob current 0.53 0.15 0.91 + + + + eu popgr current 0.45 0.11 0.79 + + + +

Source: Authors calculations using EBA regression results from STATA 10.

Notes: Coeff. (Inst5) mj - coefficient on the variable of interest (inst5) estimated in EBA model Lower bound lower bounds in the EBA model Upper bound upper bounds in the EBA model EBA test whether the EBA model satisfies (+) or not (-) the EBA decision rules for the bounds. Significance test whether inst5 in EBA model is significant (+) or not (-) Sign test - whether inst5 in EBA model is of expected positive (+) or negative (-) sign Robust/fragile whether EBA model has satisfied all tests robust (+) or failed at least one - fragile (-)

Our variable of interest always had the anticipated sign; hence, it is robust in its positive influence on economic performance. Regarding significance and EBA tests, we identify only one EBA model (that includes war, distance and popgr as additional independent variables) which is fragile. Conversely, in 23 of the 24 estimated EBA regressions with acceptable model diagnostics our variable of interest was robust to changes in specification. Accordingly, by applying Sala-i-Martins (1997) suggestion to look at the entire distribution of the EBA results, we conclude that the institutional variable was robust to changes in the base specification when estimated in a statistically well specified model.

The most demanding robustness check will be to estimate the model with an extended data set that includes the period of the global financial crisis and its

25

aftermath. However, as yet, some of the key data are not available for 2008-2012. Moreover, the accuracy of the data for this period is still questionable and subject to revision. This suggests that it may be still too early to expect a reliable robustness check by applying our model to post-crisis data

6. CONCLUSIONS

The relationship between institutions and economic performance in transition has attracted significant attention among applied economists in recent years. Most findings suggest that improving institutions in TCs does influence economic performance significantly and positively. Our study confirms the economic importance of institutions and adds some new findings to be considered by applied economists and policy makers in transition. Firstly, we find that per capita GDP is determined by the entire history of institutional reform under transition and that, conditional on this history, per capita GDP adjusts in addition to recent or medium-term institutional changes. Moreover, we find that the time-horizon over which we measure institutional performance matters. We could identify neither a statistically significant contemporaneous influence of improving institutions on economic output nor a significant effect arising from institutional changes measured over periods longer than five years. Instead, we were able to identify positive and significant effects most robustly when these arose from five-year differences in our measure of institutional quality. This relationship is quite strong: a ten percent increase in the quality of institutions over the previous five years increases GDP per capita in transition countries by four per cent, on average. We also find that absolute improvements in institutional quality are subject to diminishing returns. As institutional quality approaches the ceiling of index equal to one, so the potential for institutional improvement to raise economic performance is reduced. Our model takes into account the long term. Our estimated institutional effects are conditional on the entire history of institutional improvement (and, indeed, of all of the independent variables). However, we find that the response of economic performance to institutional improvement is a medium-term effect. Institutional change previously sustained over the medium term changes the current level of GDP per capita and thus sets a new starting level for the future evolution of GDP per capita. Moreover, this institutional effect is realised once-and-for-all, because it does not cumulate into a larger effect over time. In sum, our model estimates the current effect of recent medium-term institutional improvement on economic performance over and above the effects of the previous history of institutional improvement. Secondly, the findings suggest that TCs as a whole suffered from universal time related shocks captured by period (year) dummy variables. An implication of this finding is that models omitting period dummies are misspecified. A related finding is that our empirical results do not differ significantly for EU and non-EU TCs, or between different clusters (SEE, CIS, EU). In the course of this study, we attempt to set out some features of good practice in the study of institutional effects on economic performance. We advance three

26

practices to increase the validity of reported results: (1) in the absence of precise guidance from institutional theory on model specification, we consult meta-analysis of the literature for guidance on the implications of competing specification choices; (2) we endorse dynamic panel modelling as the preferred approach to identifying institutional effects on economic performance, and of a SGMM approach to estimation, while insisting more than do previous studies on the importance of model diagnostic tests and checks; and (3) a further corollary of lack of guidance from theory on model specification is the requirement to conduct systematic checking of the robustness of reported results by comparison with results from competing specifications, to which end we propose a variant of extreme bounds analysis designed to check the robustness of dynamic panel estimates. Finally, from the perspective of political decision-makers, the preferred results are probably not very encouraging, because the effects of institutional improvement appear to work over longer time horizons than the typical electoral cycle. This may point to an inconsistency between policy-makers short-run priorities and sound policies for the medium and/or long run. This adds to the better understanding of lagging institutional reforms and improvements in some TCs, and so may help to inform potential strategies for maintaining and/or renewing impetus to institutional reform.

27

REFERENCES

Acemoglu, D. Johnson, S. Robinson, J. A. (2001): The colonial origins of comparative development: an empirical investigation. The American Economic Review, 91: 1369-1401. Ahlerup, P. Olsson, O. Yanagizawa, D. (2009): Social capital vs institutions in the growth process. European Journal of Political Economy, 25: 1-14. Aixala, J. Fabro, G. (2008): Does the impact of institutional quality on economic growth depend on initial income level?. Economic Affairs, 28: 45-49. Alcala, F. Ciccone, A. (2004): Trade and productivity. The Quarterly Journal of Economics, 119: 613-646. Ali, A. M. (2003): Institutional differences as sources of growth differences. Atlantic Economic Journal, 31: 348-362. Andrzej, R. Cikowicz, P. (2003): A comment on the relationship between policies and growth in transition countries. Economics of Transition, 11: 743748. Aron, J. (2000): Growth and institutions: a review of the evidence. The World Bank Research Observer, 15: 99-135. Arrelano, M. Bond, S. (1991): Some tests of specification in panel data: Monte Carlo evidence and an application to employment equations. Review of Economics and Statistics, 58: 277-297. Arrelano, M. Bover, O. (1995): Another look at the instrumental variables estimation of error components models. Journal of Econometrics, 68: 29-51. Assane, D. Grammy, A. (2003): Institutional framework and economic development: international evidence. Applied Economics, 35: 1811-1817. Baltagi, B. H. (2008): Econometric analysis of panel data. Chichester: John Wiley & Sons Ltd.. Barro, R. J. Sala-i-Martin, X. (1997): Technological diffusion, convergence, and growth. Journal of Economic Growth, 2: 1-26. Basu S. R. (2008): A new way to link development to institutions, policies and geography. UNCTAD Policy Issues in International Trade and Commodities Study Series, No. 37. Baum, F. C. (2006): An introduction to modern econometrics using Stata. Texas: Stata Press. Beck, T. Laeven, L. (2006): Institution building and growth in transition economies. Journal of Economic Growth, 11: 157-186. Bevan, A. Estrin, S. Meyer, K. (2004): Foreign investment location and institutional development in transition economies. International Business Review, 13: 43-64. Bjornskov, C. Dreher, A. Fischer, A. V. (2010): Formal institutions and subjective well-being: revisiting the cross-country evidence. European Journal of Political Economy, 26: 419-430. Blume, L. Voigt, S. (2011): Does organizational design of supreme audit institutions matter? A cross-country assessment. European Journal of Political Economy, 27: 215-229.

28

Blundell, R. Bond, S. (1998): Initial conditions and moment restrictions in dynamic panel data models. Journal of Econometrics, 87: 11-143. Bond, S. (2002): Dynamic panel models: a guide to micro data methods and practice. Portugees Economic Journal, 1: 141162. Busse, M. Hefeker, C. (2007): Political risk, institutions and foreign direct investment. European Journal of Political Economy, 23: 397-415. Carkovic, M. V. Levine, R. (2005): Does foreign direct investment accelerate economic growth?. In: Moran, T. H. Graham, E. M. Blomstrom, M. (eds): Does foreign direct investment promote development?. Washington: Institute for International Economics, 195-220. Chousa, J. P. Khan, H. A. Melikyan, D. Tamazian, A. (2005): Assessing institutional efficiency, growth and integration. Emerging Markets Review, 6: 69-84. Cornia, G. A. Popov, V. (eds) (2001): Transition and institutions the experience of gradual and late reformers. Oxford: Oxford University Press. De Haan, J. Lundstrom, S. Sturm, J.E. (2006): Market-oriented institutions and policies and economic growth: a critical survey. Journal of Economics Surveys, 20: 157-191. De Hoyos, R. E. Sarafidis, Y. (2006): Testing for cross-sectional dependence in panel-data models. The Stata Journal, 6: 482-496. Di Tommaso, M. L. Raiser, M. Weeks, M. (2007): Home grown or imported? Initial conditions, external anchors and the determinants of institutional reform in the transition economies. The Economic Journal, 117: 858-881. Durham, J. B. (2004): Economic growth and institutions: some sensitivity analyses, 1961-2000. International Organization, 58: 485-529. Easterly, W. (2009): The anarchy of success. The New York Review of Books, 56: 2830. Easterly, W. Levine, R. (2003): Tropics, germs, and crops: how endowments influence economic development. Journal of Monetary Economics, 50: 41-47. European Bank for Reconstruction and Development (EBRD) (2008): Selected economic indicators. Downloaded from: www.ebrd.com in July 2008. Efendic, A. Pugh, G. Adnett, N. (2011): Institutions and economic performance: a meta-regression analysis. European Journal of Political Economy, 27: 586-599. Efendic, A. Pugh, G. Adnett, N. (2011a): Confidence in formal institutions and reliance on informal institutions in Bosnia and Herzegovina: an empirical investigation using survey data. Economics of Transition, 19: 521-540. Efendic, A. Mickiewicz, T. Rebmann, A. (2014): Growth aspirations and social capital: young firms in a post-conflict environment. International Small Business Journal, DOI: 10.1177/0266242613516987. Eicher, T. Leukert, A. (2009): Institutions and economic performance: endogeneity and parameter heterogeneity. Journal of Money, Credit and Banking, 41: 197-219. Eicher, T. Schreiber, T. (2010): Structural policies and growth: time series evidence from natural experiments. Journal of Development Economics, 91: 169-179. Eicher, T. S. Garcia-Penalosa, C. Teksoz, U. (2006): How do institutions lead some countries to produce so much more output per worker than others?. In: Eicher,

29

S. T. Garcia-Penalosa, C. (eds): Institutions, development, and economic growth. Cambridge, Massachusetts: The MIT Press, 65-80. Estrin, S. Korosteleva, J. Mickiewicz, T. (2009): Better means more: property rights and high-growth aspiration entrepreneurship. IZA DP, No. 4396. Falcetti, E. Lysenko, T. Sanfey, P. (2006): Reforms and growth in transition: re-examining the evidence. Journal of Comparative Economics, 34: 421-445. Fidrmuc, J. (2003): Economic reform, democracy and growth during post-communist transition. European Journal of Political Economy, 19: 583-604. Freedom House (2008) Nations in transition 2007: democratization from central Europe to Eurasia. Budapest: Freedom House Europe Kht. Glaeser, E. L. La Porta, R. Lopez-de-Silanes, F. Shleifer, A. (2004): Do institutions cause growth?. Journal of Economic Growth, 9: 271-303. Greene, W. H. (2008): Econometric analysis. New Jersey: Pearson International River. Gwartney, J. D. Holcombe, R. G. Lawson, R. A. (2004): Economic freedom, institutional quality, and cross-country differences in income and growth. Cato Journal, 24: 205-233. Gwartney, J. D. Holcombe, R. G. Lawson, R. A. (2006): Institutions and the impact of investment on growth. KYKLOS, 59: 255-273. Gylfason, T. Hochreiter, E. (2009): Growing apart? A tale of republics: Estonia and Georgia. European Journal of Political Economy, 25: 355-370. Hall, R. E. Jones, C. I. (1999): Why do some countries produce so much more output per worker than others?. Quarterly Journal of Economics, 114: 83-116. Havrylyshyn, O. Izvorski, I. Van Rooden, R. (1998): Recovery and growth in transition economies 1990-1997: a stylized regression analysis. IMF WP, No. 98/141. Havrylyshyn, O. Van Rooden, R. (2003): Institutions matter in transition, but so do policies. Comparative Economic Studies, 45: 2-24. International Monetary Fund (IMF) (2000): World economic outlook focus on transition economies. Washington, IMF. Jacob, J. Osang, T. (2007): Institutions, geography and trade: a panel study. Southern Methodist University Departmental Working Paper, No. 0706. Jutting, J. (2003): Institutions and development: a critical review. OECD Working Paper, No. 210. Keefer, P. Knack, S. (1997): Why dont poor countries catch up? A cross-national test of an institutional explanation. Economic Inquiry, 35: 590-602. Klomp, J. De Haan, J. (2009): Political institutions and economic volatility. European Journal of Political Economy, 25: 311-326. Knowles, S. Weatherston, C. (2006): Informal institutions and cross-country income differences. University of Otago Economics Discussion Papers, No. 0604. Lane, J. E. Rohner, D. (2004): Institution building and spillovers. Swiss Political Science Review, 10: 77-90. Le, T. (2008): Trade, remittances, institutions, and economic growth. MRG Discussion Paper Series, No. 2308.

30

Leamer, E. E. Leonard, H. (1983): Reporting the fragility of regression estimates. The Review of Economics and Statistics, 65: 306-317. Leamer, E. E. (1983): Lets take the con out of econometrics. The American Economic Review, 73: 31-43. Leamer, E. E. (1985): Sensitivity analysis would help. The American Economic Review, 75: 308-313. Leukert, A. (2005): The dynamics of institutional change: formal and informal institutions and economic performance. EEA working paper, No. 1486. Marelli, E. Signorelli, M. (2007): Institutional change, regional features and economic performance in EUs transition countries. Quaderni del Dipartimento di Economia, Finanza e Statistica, No. 37/2007. Mauro, P. (1995): Corruption and growth. The Quarterly Journal of Economics, 110: 681-712. McAleer, M. Pagan, A. A. Volker, P. A. (1985): What will take the con out of econometrics?. The American Economic Review, 75: 293-307. Moers, L. (1999): How important are institutions for growth in transition countries?. Tinbergen Institute Discussion Papers, No. 99-004/2. North, D. C. (2005): Understanding the process of economic change. Princeton: Princeton University Press. North, D. C. (1990): Institutions, institutional change and economic performance. Cambridge: Cambridge University Press. Nye, J. (2008): Institutions and the institutional environment. In: Brousseau, E. Glachant, J.M. (eds): New institutional economics a guidebook. Cambridge: Cambridge University Press, 67-80. Ostrom, E. (2005): Understanding institutional diversity. Princeton University Press, Princeton and Oxford. Paakkonen, J. (2009): Economic freedom as a driver for growth in transition. BOFIT Discussion Papers, No. 1/2009. Papke, L. E. Wooldridge, J. M. (2004): A computational trick for delta-method standard errors. Economics Letters, 86: 413-417. Raiser, M. Di Tommaso, M. L. Weeks, M. (2001): The measurement and determinants of institutional change: evidence from transition economies. EBRD Working Paper, No. 60. Redek, T. Susjan, A. (2005): The impact of institutions on economic growth: the case of transition economies. Journal of Economic Issues, 39: 995-1027. Rodrik, D. Subramanian, A. Trebbi, F. (2004): Institutions rule: the primacy of institutions over integration and geography in economic development. Journal of Economic Growth, 9: 131-165. Roodman, D. (2009a.): How to do xtabond2: An introduction to difference and system GMM in Stata. The Stata Journal, 9: 86136. Roodman, D. (2009b): A note on the theme of too many instruments. Oxford Bulletin of Economics and Statistics, 71: 135-158. Sachs, J. D. (1996): The transition at mid decade. American Economic Review, 86: 128-133.

31

Sachs, J. D. (2003): Institutions dont rule: direct effects of geography on per capita income. NBER Working Paper, No. 9490. Sala-i-Martin, X. X. (1997): I just run two million regressions. AEA Papers and Proceedings, 87: 178-183. Sarafidis, V. Yamagata, T. Robertson, R. (2009): A test of cross-section dependence for a linear dynamic panel model with regressors. Journal of Econometrics, 148: 149-161. Savezni zavod za statistiku (1991): Statistiki godinjak Jugoslavije 1991. Beograd: Savezni zavod za statistiku. Shirley, M. M. (2003): What does institutional economics tell us about development? Paper presented at the ISNIE meetings on September 28, 2003, Hungary, Budapest. Shirley, M. M. (2008): Institutions and development. C