The Evolution of Output in Transition Economies: Explaining the Differences

What are the relative roles of macroeconomic variables, structural policies, and initial conditions in explaining the time path of output in transition and the large observed differences in output performance across transition economies? Using a sample of 26 countries, this paper follows a general-to-specific modeling approach that allows for differential effects of policies and initial conditions on the private and state sectors and for time-dependent effects of initial conditions. While showing some fragility to model specification, the results point to the preeminence of structural reforms over both initial conditions and macroeconomic variables in explaining cross-country differences in performance and the timing of the recovery.

Abstract

What are the relative roles of macroeconomic variables, structural policies, and initial conditions in explaining the time path of output in transition and the large observed differences in output performance across transition economies? Using a sample of 26 countries, this paper follows a general-to-specific modeling approach that allows for differential effects of policies and initial conditions on the private and state sectors and for time-dependent effects of initial conditions. While showing some fragility to model specification, the results point to the preeminence of structural reforms over both initial conditions and macroeconomic variables in explaining cross-country differences in performance and the timing of the recovery.

I. Introduction

Following the universal collapse in measured output at the beginning of transition, the experience of the transition countries of Central and Eastern Europe (CEE) and the countries on the territory of the former Soviet Union has been quite varied. While the output paths of most countries are qualitatively similar — an asymmetric “U” or “V”-shape, with a sharp initial decline giving way to gradual recovery after a sometimes protracted “bottoming out” phase1—countries have differed greatly both in terms of the magnitude of the initial decline and the timing and strength of the recovery. In particular, transition countries in the Baltics, Russia and other countries of the former Soviet Union (BRO) have, on average, experienced sharper declines and slower recoveries than transition countries in Central and Eastern Europe, although there are large differences within these groups as well (Figures 1 and 2, Table 1).2

Figure 1.
Figure 1.

Output Paths in Calendar Time (1989 = 100)

Citation: IMF Working Papers 1999, 073; 10.5089/9781451849448.001.A001

Figure 2.
Figure 2.

Output Paths in Transition Time (Pre-Transition Year = 100)

Citation: IMF Working Papers 1999, 073; 10.5089/9781451849448.001.A001

Notes: Transition year zero is defined as the year in which central planning was decisively abandoned. This is taken to be 1992 for the BRO countries, 1990 for Poland, Hungary and countries on the territory of the former Socialist Federated Republic of Yugoslavia and 1991 for the remaining Eastern European countries.
Table 1.

Transition Economies: Growth of Real GDP, 1990-1997

(in percent per annum)

article image

The objective of this paper is to account for both the common transition experience in the time dimension—i.e. the U-shaped output profile—and the considerable crosscountry differences in output paths in terms of three main groups of explanatory factors: macroeconomic variables, structural reforms and initial conditions (including some other controls, such as wars and internal conflict). There is by now large literature on output and growth in transition,3 which has typically emphasized one or two of these three groups, raising questions about the extent to which its results might be picking up the effects of omitted variables. In contrast, we hope to disentangle the relative contributions of factors that may have contributed to the U-shaped output profile and to cross-country differences in output performance by studying the main potential determinants jointly, in the context of a panel regression that uses data from 26 transition economies.

In the context of modeling output in transition, the attempt to avoid the standard omitted variables problem is complicated by a host of other methodological problems. In dealing with these problems, we extend previous studies in four respects. First, we try to incorporate the “structural change” aspect implicit in the transition process by parametrizing the model in a way that allows policies and initial conditions to have a differential impact on the “old” and the “new” productive sectors, and thus allows the effect of policies and initial conditions on aggregate output to change as the transition process unfolds. Second, we allow for a flexible dynamic structure and in particular do not impose constant effects of initial conditions across time. While these modeling strategies are, in our opinion, necessary to avoid implicitly imposing inappropriate restrictions, they also compound the potential omitted variables problem, as we now have many more potential right side variables (lags and interaction terms of our main variables of interest) to worry about. This motivates a third methodological extension, which is to use a general-to-specific approach to generate relatively parsimonious models after starting out with a fairly extensive set of potential growth determinants. Finally, we also try to address the policy endogeneity problem as far as the macroeconomic policies go, i.e. the fact that either policies themselves or the variables that are often used to proxy for them (such as the fiscal balance) might themselves depend on output or growth. We do this by using IMF program targets as instruments for the macroeconomic right-hand side variables that are most likely to be endogenous, as explained in Section II below.

Three recent papers by de Melo, Denizer, Gelb and Tenev (1997), Wolf (1997) and Havrylyshyn, Izvorski and van Rooden (1998) share our broad objective of studying the effects of relevant policy variables while controlling for the role of initial conditions.4 None, however, addresses the full range of methodological problems that we consider. We explore much more systematically the robustness of results with respect to variations in model specification, while Wolf, de Melo et al. and Havrylyshyn et al. focus on a small set of regressions. In addition, our initial specification is substantially, and in our view appropriately, more flexible, particularly in its treatment of time-varying effects of initial conditions and differential effects of policy on the state and private sectors.5

On the other hand, the first two papers go beyond ours in that they also analyze the effects of initial conditions on policies. While we recognize that this link is interesting, it is not examined in what follows. As such, it could be claimed that our findings on the relative importance of policies and initial conditions understate the overall importance of the latter, as they ignore any “indirect” effect of initial conditions on growth via their influence on policies. However, it is not clear whether a statistical or even behavioral dependence of policies on initial conditions implies that the former should no longer be viewed as government choices.6 We take the view that they can, and are thus interested in initial conditions primarily as controls in a broad regression that includes policy variables, and in terms of their direct effects of output.7

We have two sets of results, depending on the degree to which a structure is imposed on the general-to-specific procedure adopted. First, we subject the policy variables (including lags and interaction terms) to a series of exclusion tests, conditioning on the presence of either a parsimonious set of initial conditions or dummies to control for country specific effects. The objective of this exercise is to give a sense of the range of model specifications that is consistent with the data. The main finding is that the data generally rejects the hypothesis that no structural reform variables and/or no macroeconomic variables belongs in the model, but beyond this it is not very informative—in the absence of additional prior information to guide the model selection process—in telling us what variables should be included. In other words, the same data set could be used to make contradictory claims about the significance or lack of significance of certain policies, an observation that reinforces our skepticism vis à vis ad hoc regression models of growth in transition.

Our second step is to put more structure on the general-to-specific model selection process (priors about the likely importance of variables and some simplification conventions) in ways that we believe will be transparent and acceptable to most readers. This helps us arrive at a small set of “final” specifications, which we discuss and analyze in terms of the main questions asked at the beginning. The main results are as follows. (1) The “U” shape in output is explained by the combination of (i) post-communist initial conditions that, by themselves, generate a contraction in output, and (ii) structural reforms, which are the driving force of the recovery. The effects of structural reforms themselves can typically be decomposed into a mostly positive effect on the private sector and mostly negative effects on the state sector; as the transition proceeds, the shifting relative size of these two sectors implies that the positive effect becomes stronger. (2) Even though structural reforms often affect private and state sectors in opposite ways, the net effect of structural reforms appears to be positive from the beginning, i.e. we find little evidence that reforms significantly exacerbate the output decline initially. (3) The impact of macroeconomic variables, while significant, is much smaller than that of either initial conditions and structural reforms. (4) The role of initial conditions in explaining cross-sectional variation in growth is surprisingly minor; in particular, the difference in performance between the CEE and the BRO countries is mostly explained by differences in structural reforms (even at the beginning of transition), rather than initial conditions.

Section II describes our estimation methodology in some detail. Section III presents and discusses the data, sensitivity analyses and estimation results. Using these results, Section IV interprets the transition experience—both over time and across countries—and answers the main questions that motivate the paper. Section V discusses some extensions, including whether the speed of reform matters, and whether the effects of policies are modified by initial conditions. Section VI concludes.

II. Specification Issues

Modeling the evolution of output in transition as a function of its many possible determinants gives rise to a number of methodological problems. In our view, four main issues need to be addressed. First, with little specific guidance from economic theory, there is a large potential for misspecifying the regression model by omitting relevant variables. This suggests a need to either “test down” from more general to more specific structures, or to explore the robustness of the estimated correlations in a systematic way, as has been done in the empirical literature on long-term growth (Levine and Renelt 1992, Sala-I-Martin 1997), or both. Second, we are faced with potential endogeneity problems, both through the presence of unaccounted country-specific effects and because some of the right-hand side determinants of output—in particular, macroeconomic variables such as the fiscal balance or inflation—could depend on output themselves. The question then arises as to what are the appropriate instruments to test for, and if necessary address, this potential endogeneity. Third, there is an issue regarding the stationarity of output during the transition, and therefore, whether the appropriate left-hand side variable should be output or growth. Finally, the way in which policy variables and initial conditions are to enter the regression model poses some questions. In particular, one would like a specification which allows for the possibility that the same policy change (as measured by, say, a given increase in some liberalization index) might have quite different effects depending on whether it occurs at the beginning of transition or well into transition.8 Similarly, one would not want to impose that initial conditions continue to play the same role throughout the transition process.

A. Basic Setup and Endogeneity Problems

Suppose the structural econometric model of aggregate output or output growth during transition is as follows:

Yt,i=F(Yt1,i,;Pt,i,Pt1,i,;St,i;XI;ZI)+εt,i(1a)
Pt,i=G(Yt,i,Yt1,i,;Pt,i,Pt1,i,;St,i;XI;ZI)+μt,i(1b)
St,i=H(Yt,i,Yt1,i,;Pt,i,Pt1,i,;XI;ZI)+ηt,i(1c)

Yt, i is our main dependent variable (either the level of output or growth), Pt, i denotes a vector of policy variables (including macroeconomic policy variables and structural reform indices), XI denotes observable country-specific effects (including initial conditions), and ZI stands for unobservable country-specific effects. St, i, finally, denotes a state variable such as the extent of structural change since the beginning of transition; it is included to capture the possibility that the effect of policies or initial conditions on growth changes as the transition progresses. As always, t and I index the time period and the country. Our main interest is in estimating the first of these three equations.

As it stands, this system is not identified and the output/growth equation cannot be consistently estimated. In principle, this problem could be addressed in two ways. First, we might be able to find variables which are unrelated to growth, but are correlated with one or more policy or state variables. In system 1, these variables would show up on the right-hand sides of (1b) and/or (1c) but not (1a), and could thus be used to instrument for Pt, i and/or St, i on the right-hand side of (1a). For example, consider a stabilization proxy such as the fiscal deficit. While it can be argued that the deficit is a reasonable measure of a country’s attempt to stabilize (Fischer, Sahay and Végh 1996a), deficits are clearly also susceptible to endogeneity problems as they may depend on current output via tax revenues. One way of resolving this inverse causality problem could be to use (I) deficits targets under IMF-supported programs and (ii) an indicator variable expressing whether the country is “on track” or not in the context of an IMF program as instruments for the actual deficit. These variables should be correlated with the actual deficit but, unlike actual deficits, they can be assumed to be independent of the contemporaneous error term in the output growth equation.9

Second, for some state variables and policy variables such as liberalization indices it might be reasonable to assume that contemporaneous growth does not determine policy decisions and the state of transition. In other words, suppose the true model was

Yi,t=F(Pi,t,Pi,t1,;Si,t;XI;ZI)+εt,i(2a)
Pi,t=G(Yi,t1,;Si,t;XI;ZI)+μt,i(2b)
Si,t=H(Yi,t1,;Pi,t,Pi,t1,;XI;ZI)+ηt,i(2c)

P and S are now weakly exogenous with respect to Y. The output/growth equation is identified and can in principle be consistently estimated-provided that we take care of potential fixed effects problems entering through the correlation of Z (the unobservable country effect) with the remaining right-hand side variables, bearing in mind that strong exogeneity of the right-hand side variables P and S is unlikely to hold.

In Section III below, we follow a combination of these two approaches. In the case of current macroeconomic variables, where weak exogeneity can clearly not be assumed, we use IMF program targets—which are widely available since almost all transition countries had an IMF-supported stabilization program at some point—as instruments. In the case of indices of structural reform, we assume weak but not necessarily strong exogeneity. In addition, we attempt to address the presence of fixed effects in two ways: first, by including a very large set of initial conditions and other country-specific controls—i.e. observable XIs—in the model, which may lend sufficient plausibility to the idea that the remaining set of unaccounted ZIs can be neglected; and alternatively, by explicitly estimating a model including country dummies.

B. Output versus Growth

Should the left-hand side variable be an index of real output or should it be real output growth? Clearly, the question can only be answered in relation to the right-hand side of the equation, particularly for the structural reform indices. Most of the literature has output growth on the left and levels of structural reform indices on the right (Fischer, Sahay and Végh (1996b, 1997), Sachs (1996), Selowsky and Martin (1997), and Wolf (1998)). This makes sense if one thinks that economic reforms leading to a permanent change in, for example, the degree of openness of the economy or the ownership structure of firms will have permanent effects on growth rates as opposed to just output levels. Based on, say, endogenous growth models, this is a natural assumption which has some backing in the empirical growth literature (e.g. Sachs and Warner (1995)). Alternatively, one could argue that in the context of transition economies reforms should be viewed as important primarily for the length and severity of the transitional recession and not for growth afterwards. This would argue for regressing the output level on transition indices, as in Hernández Catá (1997), which implicitly assumes that structural reforms have permanent effects on output levels, but not on how output continues to evolve after the transition. Finally, several papers in the early literature on growth in transition take the opposite view and regress growth on a cumulative liberalization index.10 In a panel context, a positive coefficient in this regression would imply that a reform measure that increases the liberalization index by a given amount in year t will have permanent effects on growth even if the measure is reversed in the following year. This is hard to justify. It thus seems safe to exclude this last approach, but the choice between the former two is difficult based on a priori arguments only.

One way of approaching the problem is to attempt to get some help from the time-series properties of the data themselves. First, one can reasonably assume that the right-hand side policy variables are stationary, as they presumably evolve (or “revert”) toward some international standard defined by market economies. Conditioning on this assumption, the problem of how to define the endogenous variable then boils down to deciding whether output is I(0) or I(1). If output is I(1), changes in stationary policy variables can only have permanent effects on growth, not output. In that case, we would be in the first of the settings discussed in the previous paragraph, and the endogenous variables should be defined in terms of growth. If, on the other hand, output is I(0), we are in the second setting and the endogenous variable should be the level of output.

Table 2 reports t-values from country-by-country Dickey-Fuller (DF) and augmented Dickey-Fuller (ADF) regressions. The first column gives the Dickey-Fuller statistic using the longest available sample for each transition economy (i.e., 1990-1997 for Hungary and Poland, 1992-1997 for the FSU countries and 1991-1997 for the remainder). The next two columns give the Dickey-Fuller and augmented Dickey-Fuller statistics using the maximum sample minus one observation. The final column of the table indicates which of these two regressions was given preference by the standard information criteria. Finally, the last row in the table computes the “t-bar” statistic corresponding to the panel unit root test proposed by Im, Pesaran, and Shin (1997), this statistic is just the average of the individual unit root statistics.

Table 2.

Output Unit Root Tests

article image

Maximum Sample = Full Sample -1.

Note: Critical Values for T=5, N=25:at the 1 percent level: -2.51at the 5 percent level: -2.11at the 10 percent level: -1.96

The table shows that the individual unit root tests reject the unit root null at the five percent level in 9 out of 26 cases using the ADF(1) test and 13 cases based on the DF test performed on the maximum sample. The t-bar statistics suggest a rejection of the unit root null hypothesis at the five or even the one percent levels, depending on the type of test. However, this conclusion is somewhat sensitive to outliers; removing FYR Macedonia from the sample imply that the ADF(1) based test could no longer reject the unit root null at the five percent level for ADF(1), although we would still reject at the ten percent level.

In summary, there is sufficient evidence against the unit root null to justify defining the endogenous variable as the output level. However, the evidence is less than fully conclusive. In addition, the fact that, with only one exception, the literature so far models the output dynamics of transition in terms of growth rather than output argues in favor of also trying growth as endogenous variable to achieve some comparability of results. Finally, as we shall see in the next section, defining the endogenous variable as growth has some important advantages in enabling us to cleanly parametrize the model in a way that distinguishes between the private and public sector effects of policies and initial conditions, and also has some advantages when it comes to presenting the results. As a result, in what follows we go both routes, i.e. define the dependent variable alternatively in levels or growth rates.

C. The Role of Structural Change

Creation versus destruction effects of policies and initial conditions

In addition to systemic transformation—i.e. the emergence of market institutions to guide resource allocation—transition involves structural adjustment, i.e. the decline of certain sectors of the economy and the creation of new ones, with the latter eventually leading to long-term economic growth. Since policies during transition are likely to affect both destruction and creation, the dependence of aggregate growth on policies can be written as follows (ignoring country subscripts, lagged dependent variables and country-specific variables):11

Yt=λtYN(Pt,Pt1,)+(1λt)YO(PtPt1,)(3)

where YN(Pt, Pt−1,…) and Yo(Pt, Pt−1,…) denotes the dependence of current growth in the newly emerging and declining sectors, respectively, on current and past policy, and λt denotes the share of the new sectors in GDE Clearly, λt is time-variant. It will be small at the beginning of transition and increase with time, and its rate of increase will depend on YN and YO.

Assuming that the dependence of YN and YO on policies can be captured by a linear relationship, equation (3) can be written as

Yt=λtα(L)Pt+(1λt)β(L)Pt(4)

where α(L) and β(L) denote lag polynomials. In general, the coefficients of α and β will not the same; in many cases one would even expect them to be of opposite sign (for example, liberalization may hurt the old sectors but help the new ones).

Suppose that one runs a regression of growth on current and past levels of policies, as in Selowsky and Martin (1996) or de Melo, Denizer, Gelb and Tenev (1997). If (4) is the appropriate specification, such a regression amounts to an attempt to estimate a composite “coefficient” λtα(L) + (1 − λt)β(L), a weighted average of potentially offsetting effects with time-varying weights. In particular, if λt increases in time and the α coefficients are positive while the β coefficients are negative, the effect of, say, further liberalization on growth would always be underestimated.

To estimate the effect of policy measures today on growth today and in the future, it is necessary to isolate the time invariant components of (4). Provided one has a measure of λt, this is straightforward. Ignoring country-specific variables, one needs to run a regression of the form:

Yt=α1(L)Pt+α2(L)λtPt+εt(5)

Comparing the coefficients of (5) and (4), it is clear that a1 (L) will identify β(L) while a2(L) will identify α(L) − β(L), so that an estimate of α(L) can be recovered by adding a1(L) and a2(L). Using measures of the private sector share in GDP as proxies for λt, we estimate a generalized version of this equation in Section III.

Time-varying effects of initial conditions

A related question is whether the effects of initial conditions can be assumed to continue with the same intensity over time. In a panel regression context, the existing literature tends to treat initial conditions as observable country-specific fixed effects (in practical terms, the same value of the initial condition is entered into the data set for each year of the sample). For a study on transition, this seems much too strong an assumption: the impact of inherited macroeconomic distortions, for example, would be expected to vanish as the economy is liberalized and stabilized, and would no longer have a notable influence on output thereafter.

To allow the regression model to account for this type of structure one would need to fit a time path for each initial condition allowing a decaying effect over time and encompassing a flat path as a special case. For example, one could fit an exponential function ae(b)xit+c where t stands for transition time, xI is the initial condition for country I, and a, b, and c are parameters to be estimated. The problem is that this requires non-linear estimation of the entire model, which proved infeasible in practice given our large initial number of regressors and the large number of regressions required at the model specification stage (see next section). To retain linear estimation, one is forced to fit the time path of initial conditions as a time-polynomial, such as a cubic function of time. However, this is impractical as a basis for simplifying the time path and testing the hypotheses that interest us. For example, testing that the effect of the initial condition becomes zero and stays zero after some time is impossible on the basis of a cubic function.

As a result, we took the following two-step approach. In a first step, we estimated a cubic time path for each initial condition. In the second step, we approximated the estimated time path via a piecewise linearization (see Appendix for details). This involved estimating the following three-parameter functional form for each initial condition xI:

axiDτ+bxiDτt+cxI(1Dτ)(6)

where Dτ ≡ 1 for tτ and 0 else; t, as before, is transition time, and τ is a parameter we picked on the basis of the third order polynomial fitted first. In other words, for each initial condition, we fitted a piecewise linear function which allowed the data to choose (I) an initial effect (intersection with the y-axis, given by parameter a), (ii) the slope of a linearly increasing or decreasing time path, (iii) the level of a flat effect after t = τ. We then went on to test hypotheses on this functional form to attempt further simplification, e.g. ruling out discontinuities between the linear segments of the function, or testing whether the initial condition becomes zero (c = 0). In the end, this enabled us to characterize most initial conditions using only one or two parameters.

D. Omitted Variables, Robustness, and Path Dependency

Since many of the policy variables and initial conditions are mutually correlated, it would clearly be incorrect to test their significance “one or two at a time”. For example, a regression equation featuring only the policy variable of interest (plus, say, a few non-policy controls, as in DDG (1996), pp. 10-11) will be misspecified unless the omitted policy variables are either uncorrelated with those included or have no effect on output or growth. This cannot safely be assumed.

The alternative is to estimate a general model that includes all major policy variables and initial conditions which might have some bearing on growth. For this model, the absence of misspecification can be assumed, and as a result, valid inferences can be based on it. The obvious drawbacks of this approach is that we might not have sufficient data points to estimate such a model at all, or if we do, that the parameter estimates of interest might lack precision and consequently the tests we want to conduct might have very low power. At worst, we might not be able to detect any significant policy effects at all.

It is important to realize that if the true coefficients on the variables in the general model are in fact non-negligible and mutually correlated, we cannot do any better than estimate the most general model possible. In this case, the information used to estimate the model parameters (including control variables which are themselves of little interest) is well spent—it prevents us from conducting a possibly erroneous inference based on a misspecified model—and the only way to improve our results is to get more data, or maybe more prior information that would enable us to restrict the model in good conscience. If, however, some of the true coefficients are in fact zero or negligible, then the information used on estimating them is wasted. Excluding the negligible variables would have led to more precise estimates of the parameters of interest without misspecifying the model.

To address this trade-off, we apply the following approach, which is a loose application of David Hendry’s “General-to-Specific” methodology.12 First, we estimate the model in the most general form that is feasible under the current data set. This includes a rich set of initial conditions, which are fitted assuming the flexible time structure described above. As an alternative, we also estimate the model using country-specific dummies (fixed effects) instead of initial conditions. At its most general level, this includes 26 country-dummies plus 26 interaction terms with the private sector share for each country dummy.13

Next, we apply a sequence of F-tests to reduce the models to more parsimonious specifications admissible under our data set. This leads to the problem of path-dependency. The order of elimination clearly matters: for example, it is typically the case that most policy variables we are interested in could be “legally” excluded early on from the model as individually or jointly not significant. Thus, it is possible to obtain parsimonious specifications where these variables play no role at all.

In order to obtain stronger results on the significance and quantitative importance of policy variables and initial conditions, one must move beyond an entirely agnostic position in which all variables in the general model are assigned equal prior importance. To make this process as transparent as possible, i.e., give the reader a sense of how the priors influence our results and what the sensitivity of the results would be to the adoption of different sets of priors, we work in two broad steps. Reflecting our interest in the role of policies, we begin by adopting the following minimalist simplification convention: we always simplify first among time constants, then initial conditions (or country-dummy interaction terms, see Appendix), and finally policy variables. The justification for this is to give the set of policy variables a chance of being estimated with reasonable precision before we decide whether and which policies matter.

The remaining path-dependency problem—including dependence of the parsimonious specifications on how we simplify within the set of policy variables—is dealt with from two angles:

  • We begin by showing the reader how broad the range of admissible specification would be if—conditional on the specification(s) achieved after simplifying among time dummies and country-specific variables—we attempted to eliminate the main sets policy variables (i.e. fiscal balance, inflation, and three structural reform indices: price and internal liberalization, external liberalization and private sector conditions). In other words, we show all admissible simplification paths that result from testing the statistical significance of policy variables as groups (i.e. including all lags and interaction terms). This is somewhat akin to sensitivity analyses by Learner (1983) and Levine and Renelt (1992) in that we are trying to map out the sensitivity of the results to a range of “extreme priors”, i.e. priors that contend that certain sets of variables should not matter at all.

  • While the imposition of these extreme priors is useful in the context of a sensitivity analysis, neither we nor (we suspect) most readers would actually want to embrace any of them. Our second angle is thus to adopt a more mainstream set of priors and simplification rules which is sufficiently strong to yield a reasonably small set of parsimonious specifications, i.e. a set which can be characterized by showing the regression results for two or three “final” models. These rules include: (i) simplifying macroeconomic variables before simplifying structural reform variables (thus giving the latter an advantage in terms of estimation precision during the model reduction process);14 (ii) moving sequentially from one variable group (e.g. internal liberalization with its lags and interaction terms) to the next; (iii) exploiting within group simplification possibilities—conditional on never deleting interior lags and testing lags “from the back”—rather than testing for the significance of the entire group first. In some of the models presented below, this latter rule was critical in avoiding the extreme outcomes traced out by the sensitivity analysis, i.e. the elimination of several variable groups altogether. The economic assumption underlying this rule is that all three structural reform indices are potentially important, and should thus be given a chance to survive in the model by eliminating insignificant lags prior to the decision whether to eliminate the index entirely, i.e. with all lags and interaction terms.

All regression results reported in this paper obey these three rules in addition to the “minimalistic” hierarchy among variable groups established earlier. While they may be sensitive to relaxations or reversals of these rules (within bounds implied by the sensitivity analyses performed separately) the results are robust in the sense that variations within the guidelines described below will not affect our results beyond the ranges suggested in the tables. In addition, the coefficients reported are robust in the sense that beginning with any of the specifications reported below, the addition of other variables from our data set will not result in a statistically significant change in the remaining variables.15

III. Estimation

A. Data

The sample period spans the transition period for 10 CEE countries, the three Baltic Republics, 12 CIS countries, and Mongolia. It covers the period 1990 to 1996 for Hungary and Poland, 1991 to 1996 for the remaining CEE countries and Mongolia, and 1992 to 1996 for Baltics and the CIS countries. Our left-hand side variable is either the logarithm of an index of real Gross Domestic Product or the annual growth rate of real GDP As stated in the introduction, we use official GDP numbers (or, in some cases, IMF estimates based on official GDP numbers) which suffer from considerable, well-known, measurement problems, and in particular are widely believed to overstate the initial output decline by inadequately capturing newly emerging activities and by using pre-transition relative prices, which tend to give low weight to new activities.16 However, the only practical alternative—output estimates based on electricity consumption—seems even more problematic for the purposes of a panel regression, quite apart from the fact that these estimates are not available for all countries in our sample.17

Our right-hand side data falls into the following categories:

  • Macroeconomic variables. This includes the fiscal balance as a percentage of GDP (Fbal), the natural log of one plus the CPI inflation rate (expressed as a decimal) (Inf) and a dummy for the exchange rate regime. Inflation is our main stabilization proxy. Controlling for inflation, the fiscal balance could be expected to have an effect on growth either through crowding out or through a short run aggregate demand stimulus. The rationale for including the exchange rate regime, finally, is the notion that the output costs of stabilization might depend on whether monetary or exchange rate targets are used. For all macro variables, IMF data was used.18

  • Structural reform indices. These consist of an index of internal liberalization (LII), which scores price liberalization and the dismantling of trading monopolies in domestic markets; an index of external liberalization (LIE) which measures the removal of trade controls and quotas, moderation of tariff rates and foreign exchange restrictions; and an index of private sector conditions (LIP) which measures progress in privatization and financial sector reforms. These indices were constructed by de Melo, Denizer and Gelb (1996a, b); we updated them for 1996 using information on recent structural reforms from the 1996 EBRD Transition Report.19

  • Initial conditions. We drew on a data set put together by de Melo, Denizer, Gelb and Tenev (1997). This includes data on initial (i.e. pre-transition) levels of per capita income in PPP terms (ypc89) and growth (GrIni); degree of urbanization (Urban), natural resource endowment (NatRR, a dummy variable for natural resource rich countries), initial macroeconomic distortions as measured by estimates of repressed inflation (RepInf) and/or actual fiscal imbalances and inflation just prior to the elimination of planning, initial economic structure including the share of agriculture (AgSh89), trade dependency (Traddep) and a measure of overindustrialization (OverInd), time under communism, and the state of pre-transition reforms, i.e. liberalization steps taken before the final collapse of central planning (LIIni). For precise definitions of these variables, see notes to Table 5.

  • Other controls. This included average growth in the OECD, the terms of trade, and dummies for war or conflict episodes.

  • Private sector share estimates. These were only used for the purposes of creating interactions with other variables (see Section II above). We constructed these estimates by combining information provided in the EBRD’s Transition Reports, the World Bank’s 1996 World Development Report, country data on shares of employment in the non-state sectors compiled by the World Bank, and, in some cases, estimates from IMF economists working on these countries (for details, see Appendix). The notation used for these interaction terms is to precede the variable name with the letter “l” (i.e., “lInf” stands for the private sector share times year on year inflation). We recognize that private sector shares are only a crude approximation to the “new” sector share, in particular, because the private sector may include privatized “old” industries which are not necessarily restructured (see Aghion and Carlin (1996)). However, there is no superior measure available at this point.

Table 3.

Regressions including Initial Conditions: Coefficients on Lagged Endogenous Variables and Policy Variables

article image
Variable Definitions: Output is natural log of an index of real output; Growth is annual average output growth (in percent), Inf is natural log of (1+p), where p denotes average annual inflation expressed as a fraction; Fbal is the fiscal balance (in percent of GDP); LII, LIE and LIP are de Melo, Denizer and Gelb’s (1996) indices of internal liberalization, external liberalization, and private sector entry conditions, respectively (see text). Notation: the prefix “1” denotes an interaction (multiplication) with the estimated private sector share; “D” denotes the first difference operator. The suffix “s” denotes that the series only contains observations corresponding to the transition sample, i.e. pre-transition lags are truncated (replaced by zero entries), see text. Estimation sample: N = 143
Table 4.

Growth Effects of a Permanent 0.1 Increase in Structural Reform Indices at t = 0 (Initial Conditions Models)

article image
Table 5.

Effects of Initial Conditions in Transition Time

article image
Variable Definitions: GrIni denotes average percentage growth 1985-89, ypc89 is income per capita in 1989, measured at PPP exchange rates; dInf-1 and dFbal-1 are Inflation and Fiscal Balance, respectively, in year prior to beginning of transition; RepInf is average repressed inflation 1987-1990, measured as percent change in real wage less the precent change in real GDP; NatRR is a dummy for resource rich countries, it takes the value 1 for Azerbaijan, Kazakhstan, Russia and Turkmenistan and 0 for the remainder; Urban is the percentage of population living in urban areas; AgSh89 is the share of agriculture in GDP in 1989; OverInd is the degree of “overindustrialization” measured as the percentage difference between actual and predicted share of industry in 1989, where the latter is based on a paper by Syrquin and Chenery (1986); LIIni is the value of the de Melo, Denizer and Gelb (1996) index of liberalization for 1989 (multiplied times 100 for easier comparability of units); Traddep denotes total trade in 1989 as a percentage of GDP.

The most obvious absentee from our list of right-hand side variables is a measure of property rights and the quality of the legal framework. Several sources have recently constructed related indices,20 but they are only available for the last few years of our sample period (typically from 1994 or 1995 onwards), and do not exist for all countries.

At the most general level, we included first and second lags of the macroeconomic variables and first, second and third lags of the structural indices in the model, in addition to the contemporaneous variables. However, we also took the view that these lags should not be allowed to extend into the pre-transition period, as the collapse of the central planning system implied a drastic structural break in most countries. Furthermore, pre-transition information is already being independently captured through initial conditions (which include information on pre-transition liberalization). Thus, zeros were substituted for the pre-transition values of lagged output and the policy variables; these truncated lagged variables were then used in conjunction with time dummies for the early transition years for which these truncations were relevant. In the tables, the use of a truncated series is indicated in the tables by the letter “s” (for “sample”) following the lag number.

B. Results

We start with two general regressions involving all the variables described above: one with growth on the left hand side and one with output in levels. Neither is satisfactory: with such a large number of regressors, the data are not sufficient to pin down the coefficient values.21 We thus conduct a first round of model simplifications, in which time dummies and initial conditions are tested and, if possible, eliminated following a general-to-specific approach. As it turns out, this does not lead to a unique outcome for the class of models with growth on the left hand side, even with some prior assumptions on the relative importance of the various initial conditions.22 However, the extent of remaining path dependency can be well summarized by two outcomes. As a result, for the models with growth on the left hand side, the analysis that follows is based on two alternative ways of modeling the effect of initial conditions; there is one specification with output as the dependant variable.

Even with this simplification of the specification of time dummies and initial conditions, there are still too few degrees of freedom to draw confident inferences about the effects of policy and macroeconomic variables. For each of the three basic models, we thus carry out two sorts of analysis, as described at the end of Section II. First, we show the extent to which the main policy variables can be legally eliminated from the model. Second, we present the results from a final set of “parsimonious” specifications at which we arrive after continuing to simplify following the guidelines discussed in Section II

Sensitivity Analysis of Policy Variables

Figures 3 and 4 show the scope for elimination of policy variables when testing for the significance of each policy variable as a group that includes all lagged terms and interaction terms. Figure 3 shows two trees, 3a and 3b, one for each of the two variants with growth on the left hand side. Figure 4 shows the exclusion possibilities with output as the dependent variable.

Figure 3a:
Figure 3a:

Exclusion Possibilities from Model with Growth on LHS (Variant A)

Citation: IMF Working Papers 1999, 073; 10.5089/9781451849448.001.A001

Figure 3b:
Figure 3b:

Exclusion Possibilities from Model with Growth on LHS (Variant B)

Citation: IMF Working Papers 1999, 073; 10.5089/9781451849448.001.A001

Figure 4:
Figure 4:

Exclusion Possibilities from Model with Output on Left Hand Side

Citation: IMF Working Papers 1999, 073; 10.5089/9781451849448.001.A001

Below the long box at the top of each figure, we show the results of all exclusion tests on the five main groups of policy variables: Fiscal Balance (Fbal) and Inflation (Inf) (referred to below as the “macroeconomic variables”) and the three structural reform indices LII, LIE and LIP, each with lags and interaction terms. The model is then simplified by eliminating groups of variables for which exclusion from the model could not be rejected at the 0.1 level. At the next level, exclusion tests are repeated for the remaining groups of variables (however, only the test results for variable groups with p > 0.1 are shown). The process is repeated until all groupwise elimination possibilities are exhausted. The boxes at the end of each tree state the policy variables remaining and in addition give the p-value associated with testing whether the exclusion restrictions imposed along the path leading to the specific model can be jointly rejected or not.

The main results are as follows:

  • The hypotheses that none of the macroeconomic policy variables matters and/or that none of the structural policy variables matter are strongly rejected in both figures 3a and 3b, i.e. in all tests with growth on the left hand side.

  • Within the group of structural variables, LIE is impossible to eliminate at the most general level in Figure 3a while LII (conditioning on the presence of LIE) is easiest to eliminate. Note that stepwise elimination in Figure 3a suggests at first that the elimination of LIE is possible provided that LII remains in the model, but the path that leads to the elimination of LIE turns out to be inadmissible at the five percent level when the restrictions embodied in it are tested as a group (p=0.03). In addition, fiscal balance can never be eliminated.

  • Figure 3b, while agreeing with Figure 3a that some macroeconomic and some structural policy variable matters, shows LIP rather than LIE or LII as the most robust structural policy variable. Regarding the macroeconomic variables, it now turns out that neither fiscal balance nor inflation can be eliminated in any of the final branches of the tree.

  • Finally, it is much easier to eliminate policy variables from the “levels specification” of the model (output on the left-hand side) than from the “growth specification” (Figure 4). The surviving variables are structural policy variables (either LIE or LII), while macroeconomic variables do not survive.

In summary, based on the above, it is only possible to say that some structural variable and (in the context of the regressions with growth rates on the left-hand sides) some macroeconomic variables are important, but not which.

An analogous sensitivity analysis based on country dummy regressions (see Appendix) concurs with the above results. Both some macroeconomic and some structural variables are robust in the models with growth on the left hand side, but weaker results obtain in the model with output on the left-hand side. However, in the latter class it is now the structural variables which can be eliminated, while the macroeconomic variables survive. The previous finding that some structural reform index is important without being able to say which reappears in the context of the models with growth on the left hand side. Given the correlation of the three structural reform indices, the inability of the data to discriminate within this group is perhaps not surprising.

Regression Results

We now move to the regression results which follow from the stronger but still reasonable set of priors discussed at the end of Section II. Most critically, we test macroeconomic before structural variables. We also simplify among lags and interaction terms for a variable (for example LIP) before moving on to test the significance of the next variable. Finally, we test the variables in a sequence that reflects prior beliefs about the relative importance of the variables (we test the least likely variables first). Although there are still several admissible variations of the final model after adopting these rules (depending on whether one begins by simplifying LII, LIE or LIP among the structural variables), the resulting variable sets and coefficient are quite similar and do not lead to qualitatively different conclusions. For each of the three basic specifications that follow from the simplification of initial conditions and time dummies, we thus concentrate on one of these admissible variations in discussing the results (Table 3).

In the case of the growth specification (growth of output on the left-hand side), this leads to two final models (“gA” and “gB”), one for each of the two variants of initial conditions specifications discussed above, while in the class of models with output on the left hand side there is just one final model (“y”). All three models were estimated in two variants: OLS (i.e. ignoring the potential endogeneity problems discussed above) and IV (using Fund program targets as instruments for the contemporaneous macroeconomic right-hand side variables). As it turns out, somewhat surprisingly, the use of IV makes little difference to the results. To keep the presentation manageable, we only show and discuss the OLS results below (Table 3). The IV versions are reproduced in the Appendix.

(i) Policy Variables. Focusing first on the two growth specifications and the macroeconomic variables, note that there are no important contradictions between specifications gA and gB. In particular, increases in inflation have a strong adverse effect on private sector growth (sum of the coefficients on DInf and DlInf)23 and a positive effect on state sector growth. However (and unlike the regression model involving country dummies, see Appendix), we do not see a contemporaneous effect of inflation levels on growth. However, we do find such a contemporaneous effect of the fiscal balance on growth. The signs of the private and public sector effects are somewhat paradoxical (recall that a positive fiscal balance means a surplus), as they suggest that tight fiscal policy, i.e. small deficits or large surpluses, sustain production/growth in the state sectors but negatively impact the private sector. Since this is a contemporaneous effect, one might suspect that it is driven by reverse causality, however, the effect shows even in the instrumented regressions (see Appendix). We are thus left with something of a puzzle: while the negative effect of a contraction on the private sector could consistent with an aggregate demand effect (in particular, since we are already controlling for inflation in this regression), the positive effect on the state sector is hard to interpret. Finally, note that the effects of macroeconomic variables appear very weak in the levels version of the model, consistent with the sensitivity analyses above. In particular, we see no effect of inflation and no contemporaneous effect of fiscal balances. There is, however, a lagged effect of Fbal along the same lines as the contemporaneous growth effect (positive on state sector, negative on private sector).24

Turning to the structural variables, note the effect of internal liberalization in specification gA, namely a contemporaneous positive impact on private sector growth but destructive impact on the state sector. This is in line with standard theory on the creative and destructive effects of structural reforms. Unfortunately, the effect does not seem robust: in model gB we merely find a positive lagged effect of the increase in the internal liberalization index on growth, i.e. without distinction between private and public sector effects. For the case of the external liberalization index (LIE), an interesting pattern emerges. On impact, and perhaps contrary to expectations, the effect of external liberalization on the state sector is positive while in specification gA the private sector effect appears negative (although it is borderline insignificant). As is clear from adding the two contemporaneous coefficients, the positive state sector effect (coefficient on LIE) outweighs the negative private sector effect (sum of coefficients on LIE and lLIE). As time passes, however, the two effects reverse sign. At two lags, the effect on the state sector is negative and on the private sector positive in both specifications. At one lag, both effects are positive and a significant difference between the effects cannot be detected. Finally, for the index reflecting privatization and private sector conditions, we find insignificant contemporaneous effects on growth in Model gA, but significant effects of changes in LIP after one lag, in the direction which theory would predict (destructive on the state sector, creative on the private sector). In contrast, for models gB and “y” we see the same qualitative contemporaneous effect of LIP that we found for LII in the case of model gA, namely a destructive effect of LIP (in levels) on the state sector and a much stronger creative effect on the private sector.

Of these findings, those referring to LIE are clearly the hardest to interpret. One might have expected external opening to be associated with an immediate destructive effect on the state sector and creative effect on the private sector, as we find after two years. However, the fact that the direction of these two effects is reversed is not as implausible as it may seem at first. In particular, at the beginning of transition a country’s exportables, which should benefit from external opening, are probably concentrated in the state sector (energy, manufacturing). While consumer goods, which are especially vulnerable to import competition, were initially produced in both the state and the (small) private sector, they are likely to constitute a larger share of the latter. In addition, state sector production is likely to be initially less vulnerable to import competition than the private sector as it enjoys greater state support (via credit and direct subsidies). After a few years, however, one would expect this support to taper off as budget constraints are increasingly enforced, while the easing of import constraints benefits newly emerging private firms.

Given the fairly complicated dynamics implied by the coefficients on the three policy indices and the partly contradictory effects on the state and private sectors, the discussion so far makes the overall effects of structural reforms over time (say, from the beginning of transition) difficult to gauge. To clarify these effects, Table 4 combines the coefficients on all three structural reform indices from both versions of the growth model of Table 3 to show how a 0.1 increase in LII, LIE and LIP at t = 0 would have affected growth during transition. The top two lines of each panel reflect the separate effects on the state and private sectors. The rest of each panel contains weighted averages of these two lines using a set of actual paths of the private sector share. In addition to the average BRO and non-BRO private sector shares over time, we pick three countries as examples within each of these groups. The idea is to show countries who differed widely in terms of their (estimated) initial share of the private sector and subsequent paths: thus, Albania started out with a very low share (an estimated 0.05) but privatized quickly, Turkmenistan started out with a low share (0.1) that remained low, Bulgaria stated out with a low but somewhat higher share (about 0.17) that grew rather slowly, and the Czech Republic and Estonia started out with shares in the 0.2 to 0.25 range that grew quickly, reaching about 0.7 by the fifth year of transition. Russia represent an intermediate case, it started out at a little around 0.25 in 1992 and is estimated to have reached 0.6 in 1996. The growth effects shown in Table 4 for each country condition on these realized paths of the private sector share; thus, the table shows what would have happened to growth at the margin in reaction to an increase in reforms. Since models gA and gB do not contain lagged endogenous variables (they were found to be insignificant and eliminated in the course of the model simplification process), the paths of Table 4 are essentially impulse responses with respect to a change in structural policies at the beginning of transition policy variable, treating the private sector share as exogenous.

The main robust finding of Table 4 is that structural reforms in aggregate help all countries in the later transition years and helps most of them even in the early transition years, with only one exception. Thus, while reform tends to hurt the state sector, our findings offer little support for the widespread view that structural reforms have an aggregate “destruction effect” at the beginning of transition, which importantly contributes to the initial output decline.25 The main exception is Turkmenistan, which kept its private sector share so low throughout the transition that the opposite effects of structural reform on the private and state sector in model gB continue to imply a negative aggregate effect even in the later years.26

A number of differences between models gA and gB underline the sensitivity of some of our findings to model specification. In model gA, contractionary effects of reforms on the state sector are registered only temporarily, particularly at t = 1, i.e. with a one year lag after the beginning of the reform experiment. This is driven by the negative lagged effect of changes in LIP on the state sector (see variable DLIP-1s in Table 3), and is consistent with results from the regression models using country dummies (see Appendix). In contrast, in model gB the contractionary effect on the state sector is permanent, driven by the negative coefficient on the level of LIP. In the aggregate, this effect is more than offset by a large positive private sector effect of the same variable.

(ii) Initial Conditions. As described in Section II, the initial conditions were parametrized in a way that allowed us to model and test time-varying effects, namely by approximating generally non-linear time paths by piecewise linear functions of time which were subsequently simplified by testing various exclusion and equality restrictions. The disadvantage of this procedure is that it yields estimation coefficients on piecewise linear time paths which require some extra notation and are hard to interpret without actually plotting them. Rather than discussing the coefficients, in Table 5 we show the time paths implied by these coefficients, for each initial condition surviving in the final model models. The coefficients themselves are tabulated in the Appendix.

The thought experiment underlying Table 5 is analogous to that of Table 6: what would be the change in the growth rate (or, in the case of model y, in the output level) in each transition year if before the beginning of transition the initial condition variable had been larger by one unit, everything else equal? “Everything else equal” implies that we ignore any dynamics via lagged dependent variables; thus, the time paths can be interpreted as impulse response functions only in the case of models gA and gB, which have no lagged dependent variable in the model. For the case of model y, however, there are several lagged dependent variables (see Table 5). In this case, the time paths of Table 5 are really no more than a way of representing the coefficients on the initial conditions terms.

Table 6.

Accounting for Growth in Transition (Model gA)

(in percent per year)

article image

In interpreting Table 5, note that the magnitude of the effects across initial conditions can obviously only be compared when the units are the same, i.e. only for the cases where initial conditions are being measured in percent. Observe also that whenever we found a significant difference in the effects of a particular variable on the state and private sectors (i.e. an interaction term with the private sector share that insignificantly different from zero), we show both effects; if only one line is shown (as in the case of Traddep) this means that the hypothesis of identical effects on both sectors could not be rejected.

The overall impression from Table 5 is that the effects of individual initial conditions are fairly similar across the three specifications shown. The robust effects are as follows:

  • higher trade dependency has an adverse aggregate effect on the initial output decline

  • over-industrialization also makes the initial output decline worse; paradoxically, this seems to be driven by an adverse effect on private sector growth, rather than a faster collapse of the state sector

  • higher urbanization is associated with initial faster growth of the private sector, but only at the beginning of transition, and there appears to be a reversion in later years when it affects growth adversely;

  • natural resource exporters (Russia, Azerbaijan, Kazakhstan and Turkmenistan) tend to suffer larger output declines, at least at the beginning

  • a higher share of agriculture is associated with lower private sector growth

  • open macroeconomic imbalances in the year preceding the end of central planning aggravate the output decline; however, average repressed inflation in the 1987-1990 period is generally positively associated with growth.

While many of these effects—which generally appeared robust not just in these three specifications but also in other regressions we do not show—are intuitive and have standard interpretations,27 others are clearly not, such as the positive effect of repressed inflation. In these cases, while one can sometimes rationalize the findings (for example, that repressed inflation went along with pent-up demand for unavailable or shortage goods that had a positive impact on private sector growth) we would not put much stock in the results, particularly in view of the somewhat crude way in which some of the initial conditions are measured.28

The joint effect of the initial conditions does not always conform to the expected time pattern of a strong initial output decline that later disappears. The memorandum item at the bottom of the table shows the overall implications of the initial conditions for the fitted growth path for the average transition economy, i.e. the joint effect weighted with the actual values for each variable. For model gA, the aggregate effect of initial conditions is to generate sharply negative growth in the first year, followed by diminishing but still negative effects on growth over time. This supports the notion of a strongly adverse effect of initial conditions that slowly vanishes over time (with a “half life” of about 5 years). In model gB, in contrast, an initially upward sloping profile, that is gradually receding negative effects of initial conditions on growth, reverses direction in the last two years. While the mechanics of this effect is not very interesting in this particular model,29 it is important to note that the nicely upward sloping path of model g A is not a robust finding. As will become clearer below, this has implications for our overall conclusions when interpreting the time path of growth in transition.

IV. Accounting for the Path of Output in Transition

We now attempt to answer the basic motivating questions of this paper: what factors account for the decline and recovery in output, and what explains the considerable cross-country differences in performance? To answer these questions, we decompose output growth into the contributions of the major groups of explanatory variables. This is feasible in models gA and gB because neither contain any lagged dependent variables; thus, the fitted values can be written as a linear combination of the independent variables alone. For this reason, we concentrate on these two growth specifications in what follows.

We work from two angles. First, we focus on the time dimension of the transitional recession and recovery by decomposing the fitted growth path in transition time, both for the average economy and for two major country blocks, “BRO” and “non-BRO”. This will also show whether our models have greater difficulty fitting the BRO experience than the experience of the Central and Eastern European countries. Second, we focus on the cross-sectional dimension by decomposing fitted growth by country for two major time blocks, the “early years” (transition time 0,1 and 2) and the “later years” (transition time 3 and 4). This gives a sense of what drives the differences in the growth experience across countries and whether the driving factors are different in the early transition period, when most countries experienced varying degrees of output decline, and in the later period, when some countries began to recover while others continued to slide. It also reveals countries that may be considered outliers in the sense that our models do not adequately fit their experience.

A. Accounting for the Transition in the Time Dimension

Tables 6 and 7 show actual and fitted growth for models gA and gB and three groups of countries: the (unweighted) average of all transition economies, the BRO countries and the non-BRO countries including Mongolia. Below the “fitted growth” line, we show the decomposition of fitted values into the major groups of variables (macroeconomic, structural, initial conditions including the regression constant, and the effect of wars). For each time period, the contribution of each group is calculated by summing the product of each right-hand side coefficient and the corresponding data within each group of variables, and then averaging either over all countries (upper panel) or over the BRO/non-BRO groups (lower two panels). The tables thus show how the actual paths of explanatory variables combine with the regression coefficients to explain the transition experience and the main differences between the BRO and non-BRO groups. Figures A3, A4 and A5 in the Appendix plot the actual paths of explanatory variables along with their main growth effects.

Table 7.

Accounting for Growth in Transition (Model gB)

(in percent per year)

article image

Focusing first on the pure time dimension, i.e. on the upper panel that accounts for the average transition experience, we first note that the key conclusions are consistent across both models. The major results are as follows.

  • The “fit” of the average growth path over time is near perfect. This may not be very surprising given the data-driven way in which the regression models were derived and also the fact that we are averaging fitted values over 26 countries. However, it is worth noting that the unexplained residuals are tiny in all of the five time periods without any obvious pattern. This clears the way for the exercise that follows, namely to “account” for growth by decomposing the fitted path.

  • The output decline (transition years 0, 1 and 2) is overwhelmingly attributed to initial conditions and (to a much lesser extent) macroeconomic imbalances. Among the adverse initial conditions, trade dependency and overindustrialization play a prominent role in the initial output decline, accounting for more than three quarters of the impact of initial conditions on the output decline in year 0.

  • the small initial negative impact of macroeconomic variables is due to offsetting effects of inflation and the fiscal balance. Notably, the net initial effect of inflation appears positive; this in turn is attributable to a positive effect on the state sector (see previous section) that more than offsets the adverse affect on the private sector at a time when the state sector is still large.

  • We find no evidence that, controlling for the other factors, structural reforms initially aggravated the output decline. In one of the two models (gB) we do find a substantial negative impact of structural reforms on the state sector (particularly after two years), but this is more than offset by its positive impact on the private sector.

  • The driving force behind the recovery is the impact of structural reforms and—to a lesser extent, and primarily in Model gA—the tapering off of the effect of initial conditions.

Next, we consider what the tables have to say about the differences in performance between the BRO and non-BRO group. As stated in the introduction, two facts require explanation: (i) why was the initial output decline steeper in the BRO; (ii) why did the BRO take longer to recover? Again, the two models gA and gB largely agree on the answers. These are as follows:

  • the larger initial output decline is attributed to some extent to more adverse initial conditions (particularly in Model gB), but to a greater extent to the fact that structural reforms got off to a slower start in the BRO.

  • the poorer growth performance in the BRO in the later transition years is overwhelmingly attributed to less advanced structural reforms.

Overall, this would seem to put most of the blame for the BRO’s poorer performance on policies, rather than initial conditions per se.30

B. A Cross-Sectional View

Consider now Tables 8 and 9, in which the emphasis is on explaining cross-country differences in the transition experience. The “raw material” of these tables is the same as that of Tables 6 and 7, i.e. the product of right-hand side coefficients and data values, summed over groups of related explanatory variables. However, rather than averaging these fitted values over countries as in Tables 6 and 7, we average over time, distinguishing only between two broad “time-phases”—earlier and later transition years. For each phase, we show the fitted values for each individual country. In addition, the last lines of the tables show the cross-sectional correlation, within each time phase, between the fitted values corresponding to a given group of explanatory variables and actual growth.31 This is introduced as a summary measure of the extent to which each of the major groups of right-hand side variables contribute to explaining the observed cross-country differences in growth performance, within each phase.32

Table 8.

Accounting for Growth in Transition: Cross-Sectional Perspective (Model gA)

(in percent per year)

article image
Table 9.

Accounting for Growth in Transition: Cross-Sectional Perspective (Model gB)

(in percent per year)

article image