Why the Link between Volatility and Growth is both Positive and Negative¤ Jean Imbs London Business School and CEPR April 2002 Preliminary Abstract This paper uses international disaggregated data to investigate the link between volatility and growth. We …nd the relationship to depend on the level of aggregation. At the sectoral level, volatility is associated with high growth. We use non-parametric methods to show that within activity volatility is typically associated with net entry and higher (total factor) productivity. This is consistent with the evidence that most of the process of creative destruction occurs within narrowly de…ned sectors, as surveyed in Caballero and Hammour (2000) or Davis and Haltiwanger (1998). Once sectoral data is aggregated at the country level however, the link becomes negative. We show how aggregation creates this discrepancy. Keywords: Sectors, Growth, Volatility. JEL Classi…cation: E32, O40 ¤ London Business School, Dept. of Economics, Regent’s Park, London NW1 4SA, [email protected]. 1 1 Introduction The relationship between macroeconomic volatility and growth is an old and important issue, both from a theoretical and an empirical standpoint. The debate relates directly to two central questions in macroeconomics: what are the determinants of economic growth, and if volatility is one of them, how does this a¤ect our understanding of business cycles and their welfare e¤ect. Theoretically, the correlation could go both ways: positive on the one hand, as volatility could be a manifestation of the adoption of a new (general purpose) technology, that entails costly restructuring.1 ;2 On the other hand, the relationship could be negative as volatility can result in wasted human capital or deter investment and the implementation of newer technologies.3 Without a theoretical consensus, a considerable body of literature has concerned itself with establishing empirically the sign of the relationship. This paper contributes to this literature using a disaggregated approach. There are several reasons why a disaggregated approach is promising to address the question at hand. Firstly it multiplies the cross-sectional dimension of the data, quite usefully when estimating the determinants of output growth, an endeavour famously sensitive to the conditioning set. In particular, the higher dimensionality of the data relative to cross-country studies, and the fact that the variation of interest is country-sector speci…c opens the possibility to account for all country- and sector-speci…c deter1 The ‡uctuations resulting from the disruptive introduction of a new technology are commonly lab eled “Schumpeterian waves”, or “creative destruction”. A large theoretical literature, spanning macroeconomics and industrial organization, proposes to model creative destruction. See for instance Helpman and Trajtenberg (1998) or Jovanovic and Rob (1990) for an early contribution. 2 A closely related strand of research concerns itself with the long-run e¤ects of business cycles. In particular recessions could have positive long-run e¤ects on growth, through “opportunity costs” arguments whereby productivity-enhancing activities are best left for recessionary p eriods, or “cleansing” e¤ects whereby recessions eliminate less productive units, thus increasing average productivity. This is similar to the present exercise, but the corresponding empirical tests typically involve time-series procedures to identify shocks of interest and track their e¤ect on productivity over time. Unless appropriate asymmetries between recesionnary and expansionary shocks are built in these models, volatility will however remain unrelated with growth. See Aghion and Howitt (1996) for a survey. 3 This will be the case if for instance investment is irreversible, as in Pindyck (1991), or if growth occurs through learning-by-doing, as in Martin and Rogers (1997). Ramey and Ramey (1991) present a model where volatility constrains the choice of technology, and thus results in lower growth. 1 minants of growth, both in a pure cross-section and using panel techniques. We …nd the link between (sectoral) growth and volatility to be non negative at least, and actually signi…cantly positive in most speci…cations. Secondly, the question at hand can be addressed both in the aggregate and at the sectoral level using the same data, an interesting opportunity to investigate the importance of aggregation when computing growth and volatility. Our disaggregated data includes information on the number of establishments at the sectoral level, thus making it possible to decompose growth into an intensive (average …rm growth) and an extensive (average …rm entry) margin.4 This decomposition turns out to be strikingly relevant, as we …nd volatility to be associated signi…cantly with extensive growth only. Furthermore, we use non-parametric methods to show that the link between growth and volatility increases with the rate of (sectoral) productivity growth. Both facts lend support to a Schumpeterian view of economic ‡uctuations where innovative entering …rms induce both high growth and high volatility, as the economy adjusts to the advent of a new technological paradigm. They are also consistent with the empirical fact that a large share of (industry-level) productivity growth originates in the reallocation of factors within narrowly de…ned sectors.5 This surprising result stands in stark contrast with recent (aggregate) studies by Ramey and Ramey (1995) or Martin and Rogers (2000) [RR and MR, respectively], who document a negative relation. We explain this contradiction with an aggregation argument. In particular, we show that the same dataset can be used to generate both a (well-known aggregate) negative and a (surprising disaggregate) signi…cantly positive relationship between growth and volatility. We show that the aggregate negative link stems from composition e¤ects, and document how the range of activities that compose aggregate GDP appears to have a crucial impact on the nature of the aggregate relationship. Thus, this paper reconciles two sets of apparently contradictory theoretical propositions, whose manifestations are simply obscured by aggregation. The negative link between aggregate growth and volatility masks a positive 4 The data come from UNIDO, and cover manufacturing sectors. An establishment is de…ned as “a unit which engages, under a single ownership or control, in one, or predominantly one, kind of activity at a single location”. The measure may not exactly coincide with the notion of …rm, but there is to our knowledge no other data for such a large cross-section. We will indi¤erently use the two appelations in the paper. 5 See for instance Caballero and Hammour (2000). 2 one at the purely disaggregated level, and is to cross-country heterogeneity in the sectoral composition of aggregate GDP. The rest of the paper is structured as follows. In section 2, we present our purely disaggregated results and, using a variety of recent techniques, establish that sectoral growth and volatility are positively related in most cases. In section 3, we show the sectoral results are consistent with a Schumpeterian view of innovation and volatility, commonly labeled “creative destruction”. Section 4 turns to the aggregate evidence and shows how aggregation a¤ects the link between growth and volatility. There, we reconcile our (sectoral) evidence with the literature, and cast some light on the relationship between aggregate growth and volatility. Section 5 presents some sensitivity analysis. Section 6 concludes. 2 2.1 Sectoral Volatility and Growth Data We use data on sectoral value added, employment, factor content and number of establishments in manufacturing activities only, as published by the United Nations Industrial Development Organization (UNIDO). Although observations go as far back as 1963, the data is particularly incomplete in the early part of the sample.In order to limit measurement error, we focus on countries where at least 18 years of observations are available between 1970 and 1992, which selects a maximum of 36 countries.6 We also consider a sub-set of the UNIDO data, composed of OECD economies only, which further reduces the sample to 15 countries. The purpose of a reduced dataset is to control for the putative importance of transitional dynamics and focus on economies at a comparable stage of development. The OECD sample excludes developing countries where industrialization -and the a¤erent structural change- has played an important role in economic growth. There is a maximum of 28 sectors. We de‡ated value added by Producer Price 6 We constrained the numb er of sectors in a given country to remain constant over time, as the aggregate volatility stemming from whole sectors appearing or disappearing over time could be an artefact of measurement error, and eliminated sectors whose de…nition in the ISIC classi…cation system varied in our sample. We also took care of outliers, whose inclusion only reinforced our results. The “System of National Accounts” was changed in 1993, which is why sectoral information of a comparable nature over time and across countries typically becomes incomplete after 1992. 3 Index series taken from the International Monetary Fund’s (IMF’s) International Financial Statistics.7 Data on aggregate (capital and output) growth rates come from the Penn-World Tables. Table 1 presents some summary statistics. Relative sectoral output growth and volatility are much larger on average in the extended sample, as expected given the wider country coverage. The unconditional correlation between growth and volatility is mildly positive in the UNIDO data, although never signi…cantly di¤erent from zero. Figures 1 and 2 show the corresponding unconditional scatterplots. 2.2 Cross-Sectional Evidence In this section, we focus on the pure cross-section in the variables of interest. In particular, we follow an intuition similar to Rajan and Zingales (1998), and hold constant all (time-invariant) country- and sector-speci…c characteristics through …xed e¤ects, and focus instead on individual country-sector variation. Thus we alleviate parts of the issue of conditionality, quite important in standard growth regressions. A substantial literature has indeed concerned itself with the appropriate conditioning set in growth regressions, culminating with Levine and Renelt (1992) who propose a list of four robust conditioning variables. 8 Sectoral data make it possible to control for all time invariant country-speci…c considerations, and thus do away with the potential issue of sensitivity that is prevalent in this area, even in a pure cross-section. Similarly, sector-speci…c e¤ects control for any tendency of one sector to display systematically high growth rates, possibly because of technological breakthrough, and focus on systematic deviations from an international average. Thus, all estimations are run controlling for countryand sector-speci…c e¤ects. Speci…cally, consider µ ¶ yi;j;T yi;j;0 yi;j;t yi;j;t¡1 ln ¡ ln = ¯ 0 + ¯ 1 VT ln ¡ ln + ¯ 2 Xi;j + "i;j Yj;T Yj;0 Yj;t Yj;t¡1 (1) where i and j index industry P and country, respectively, yi;j;t is sectoral value added at time t, Yj;t = i yi;j;t , Xi;j is a vector of control variables and 7 Or alternatively an index of industrial production when the PPI was not available. This follows Rajan and Zingales (1998). Countries and sectors are listed in the App endix. 8 The Levine-Renelt variables are initial income p er capita, average population growth, initial human capital as measured by schooling years and average investment rate. 4 VT (:) is the (time) variance operator, computed over period T.9 The coef…cient of interest is ¯ 1. There are numerous reasons why the residual "i;j is liable to contain both industry- and country-speci…c e¤ects, say °i and ´j , respectively. To list only two, suppose for instance political instability translates in both high aggregate volatility and low growth, as in Alesina et al (1992): the residual is then negatively correlated with the regressor, through ´ j . Suppose instead an industry speci…c technological breakthrough is associated with both high sectoral volatility and growth: the residual is now positively correlated with the regressor, through °i . Accounting for these …xed-e¤ects is a serious issue in most growth regressions, especially so when the issue is to identify the signi…cance of one particular variable, such as volatility in our case. Here however, (1) can be estimated simply in deviations from means as in µ ¶ µ ¶ yi;j;T yi;j;0 yi;j;t ± ln ¡ ln = ¡° 0 + ¯ 1 ±VT ¢ ln + ¯ 2 ±Xi;j + ±"i;j Yj;T Yj;0 Yj;t (2) P P where ±Zi;j = Zi;j ¡ 1I i Zi;j ¡ J1 j Zi;j. The country or sector …xed e¤ects P P are now controlled for, as ° 0 = ¡¯ 0 + 1I i °i + 1J j ´ j .10 Equation (2) leaves open the question of what country-industry speci…c variables ought to be included in the set of controls, denoted by Xi;j . We now turn to this question. Numerous authors have concerned themselves with the dynamics of sectoral specialization, using insights from growth and international trade theories.11 A very stylized model may be helpful here. Consider a two-country two-sector two-factor world. Sector one is capital-intensive, sector two labour-intensive, and country A has higher aggregate capital-labour ratio than country B. Suppose …nally that initial sectoral specialization patterns correspond to the balance of aggregate endowments, i.e. that sector one is 9 Basu and Fernald (1997) argue that gross output is a better measure of sectoral activity than value added when investigating external e¤ects of aggregate activity onto sectoral production. We checked that the results obtain as well when using gross output. 10 This is a generalization of the estimation procedure used in Rajan and Zingales (1998), where only the dependent variable is demeaned, and the focus is on the signi…cance of an interaction term b etween (sector-speci…c) need for and (country-speci…c) availability of external …nance. Since they use the UNIDO data as well, Rajan and Zingales also have to present results for manufacturing sectors only. 11 For theoretical approaches, see Ventura (1997) or Cuñat (2000). For an empirical analysis, see Imbs and Wacziarg (2002). 5 initially larger in country A and sector two is larger in country B. With aggregate diminishing returns to capital, country B accumulates capital faster, and factor price equalization favours growth in sector one there. Similarly, factor price equalization and neo-classical convergence suggest sector two will grow relatively faster in country A.12 Thus, from an empirical point of view, the determinants of relative sectoral growth are two-fold. Firstly, a variable capturing country-speci…c capital-labour growth interacted with sector-speci…c capital content. This is the approach adopted in a recent paper by Bernard and Jensen (2001), which shows sectoral factor content is important in explaining higher than average sectoral output growth in a cross-section of US regions. We estimate (2) with such an interaction term between sectoral capital content and the aggregate growth of the capitallabour ratio. However, the measurement of this variable raises a number of issues: although UNIDO provides information on the sectoral wage bill and sectoral value added (both nominal), the resulting labour shares tend to be quite noisy. Secondly, an assumption of constant returns to scale must be maintained if the capital share is to be inferred from the wage bill. Although the evidence suggests sectoral production functions by and large display constant returns to scale, there seems to be ample cross-sectoral variation.13 Thus, an interaction term à la Bernard and Jensen must be taken as an approximation in the present context. Fortunately, the previous sketch of a model suggests an obvious alternative: in both countries, the fastest growing sector is also the smallest initially. This happens of course because of diminishing returns to capital, and because of our assumption that initial sectoral shares match initial relative aggregate endowments. However, it suggests the inclusion of a measure of the initial relative size of a sector in the estimation of (2), which has the substantial advantage of being immediately available. There is at least one additional reason to include such an “initial condition” term in (2), which is more of an econometric nature. Transition dynamics in the usual neoclassical sense are potentially important here, as they tend to result in both high and monotonically-decreasing growth. This may result in an upward bias when estimating the relationship between growth and volatility.14 We therefore include the initial sectoral share in value added in (2), and note 12 For details, see Ventura (1997). This model guided the choice of using relative growth rates in (1) and (2). 13 See for instance Burnside, Eichenbaum and Rebelo (1996) 14 There are reasons to believe that this bias is not as imp ortant in the sectoral data as it is in the aggregate. Imbs and Wacziarg (2002) have shown that the notion of a “steady state economic structure”, to which growing economies would converge is not supported 6 that a negative sign could be interpreted equally as a “convergence” term or as dictated by comparative advantage. 2.2.1 Results Table 2 presents estimations of (2), for both UNIDO samples and various speci…cations of Xi;j . The main result that should be taken from the table is the signi…cantly positive sign of ¯ 1. The coe¢cient on the variance of value added growth is positive and signi…cant in 75% of the cases, and always in the OECD sub-sample. It is also never signi…cantly negative.15 Volatility is also important economically. In the OECD, the smallest estimate of ¯ 1 across speci…cations is 3.832, which implies one standard deviation of the volatility measure (measured across country-sectors) translates into around a third of percentage point of average yearly sectoral output growth.16 Several other comments are in order. Firstly, our results are in partial agreement with Bernard and Jensen (2001), as we …nd some evidence that capital intensive sectors grow faster in economies with high rates of capital accumulation. This appears to be particularly prevalent in our extended sample, inclusive of developing economies. There are two important di¤erences between this study and Bernard and Jensen’s: …rstly our sample is international whereas theirs is inter-regional in the US, secondly our estimations are ran with both country and sector …xed e¤ects.17 When an estimation more similar to Bernard and Jensen’s equations (7) and (8) is performed, for instance our equation (1) with no variance term and no country …xed e¤ects, the “comparative advantage” interaction term does enter with a very signi…cant positive sign even in the OECD sample. Thus, the results in table 2 do not necessarily cast doubt on a role for international trade based theories of sectoral specialization; rather, they correspond to an estimation strategy whose purpose is elsewhere. in the data. Countries are shown to …rst diversify, thus allocating resources across sectors increasingly equally, but start re-specializing once they reach a relatively high level of income p er capita. 15 From the point of view of the positive bias that could arise from transitional dynamics, it is reassuring that estimates of ¯ 1 should be most positive in the sample where this putative bias is a priori least prevalent, i.e. in the OECD. 16 To be precise, an average yearly growth higher by 0.36% in the OECD sample. 17 Bernard and Jensen focus on the U.S economy, and show sectoral growth is lowest in least capital- and skill-intensive industries. 7 Secondly, table 2 provides evidence in favour of a signi…cant “convergence” term, as measured by initial sectoral value added, either in absolute or in relative term. In both samples, initially “smaller” sectors tend to display higher subsequent growth rates. As expected, the signi…cance of a convergence term is strongest in the extended sample, where the volatility term ceases to be signi…cant once initial conditions are held constant. This con…rms the importance of a bias due to “transitional dynamics” in explaining why growth relates positively to volatility. In the reduced sample however, volatility continues to be signi…cantly positive even controlling for initial conditions, which calls for an alternative explanation than the aforementioned bias. There is evidence that the relationship between growth and volatility is biased upwards by transitional dynamics, but this evidence is quite limited between rich countries.18 The next section generalizes the estimation procedure in making use of the panel dimension of the data. 2.3 Dynamic Panel Evidence Recent years have seen important developments in the use of panel techniques to estimate growth equations.19 In this section, we implement some of them, thus making use of the panel dimension of our data. A generalized speci…cation of (1) is µ ¶ yi;j;T yi;j;t yi;j;T ¡1 P P ln = ¯ 1 VT ¢ ln + (¯ 2 + 1) ln P y y i i;j;T i i;j;t i yi;j;T ¡1 +¯ 3 COMP:ADV:T + ®i;j + (±T ) + "i;j;T (3) with t 2 [T ¡1; T]. The conditioning set now includes initial relative sectoral value added and an interaction term between initial sectoral capital content 18 We also investigate the possibility that the residuals in (2) have country- or sectorspeci…c components with implications on the structure of the covariance matrix, and thus on the consistency of the estimates. We do this by allowing for clustered residuals at the country and the sector level, i.e. relaxing the assumption of independence within clusters (country or sector), but not between them. The technique is standard in the labour literature. The results with standard errors robust to clustered residuals are available upon request. In short, while ¯ 1 becomes less signi…cant, as all other coe¢cients, it never falls below the 13% signi…cance level in the extended sample, and below 2% in the reduced OECD one. 19 See Caselli, Esquivel and Montfort (1996) for a seminal contribution, and Forb es (2000) for a recent one. 8 and aggregate capital growth over the subsequent sub-period [T ¡1; T ].20 ±T denotes a period-speci…c indicator variable, included whenever meaningful. The main di¤erence with equation (1) is that the …xed e¤ects ®i;j are now generalized to be country-sector speci…c. An obvious procedure to obtain ³ ´ yi;j;T yi;j;t P P estimates for ¯ 1 involves evaluating ln and V ¢ ln over T i yi; j;T i yi;j;t sub-periods, and estimating (3) in …rst di¤erences. However, a well-known problem in doing so is that, while it controls for the presence of a timeinvariant component in the residuals, it leaves open the possibility that the lagged dependent variable be correlated with the residuals. In other words, estimates of ¯ 2 could be biased, with possible consequences on estimates of ¯ 1. Anderson and Hsiao (1981) suggest to instrument the …rst-di¤erenced y ¡2 lagged dependent variable with its initial level, as indeed ln P i;j;T is i yi;j;T ¡2 not correlated with "i;j;T ¡ "i;j;T ¡1. Arellano and Bond (1991) generalize this procedure by proposing to include further lags of the dependent variables as instruments, a useful possibility when the instrument set turns out to be insu¢cient, as is often the case in the present context. Crucially, both instrumenting techniques require no serial correlation in the residuals for the instruments to be consistent. We next present the results of three estimations procedures for (3), …rstly using simple within group OLS, secondly using the Anderson-Hsiao instrumental variable technique, and …nally implementing the Arellano-Bond GMM method. 2.3.1 Results Table 3 presents the results of our three dynamic panel speci…cations, with or without the interaction variable capturing comparative advantage. The generality of (3) is appealing, but measurement error is likely to obscure substantially the resulting estimation. Firstly, variances are now computed on fewer observations; secondly, the resulting measurement error is likely to be exacerbated by …rst-di¤erencing. As data quality tends to vary with aggregate income level, measurement error is likely to be strongest in our extended sample. In other words, within-groups estimation of (3) is likely to give estimates of ¯ 1 and ¯ 2 that are seriously biased downward. From this point of view, the results in table 3 are remarkably robust. The top panel presents the results of the estimation of (3) in …rst-di¤erences, where ln Pyi;j;T i yi; j;T ³ ´ yi;j;t P and VT ¢ ln yi;j;t are computed over [1970,1981] and [1982,1992], rei 20 Thus initial values are measured in T ¡ 1. 9 spectively, and initial variables are measured in 1970 and 1982.21 In both samples, the convergence e¤ect is estimated to be strong, with very negative estimates of ¯ 2. As expected, transitional dynamics as proxied by ¯ 2 are both more signi…cant and larger in magnitude (by roughly a third) in the extended sample. Secondly, the comparative advantage variable is non-signi…cant, but its inclusion results in the e¤ects of volatility to become insigni…cant in the extended sample, but not in the OECD. We conclude that the presence of country-sector speci…c …xed e¤ects is not driving the tendency for the correlation between volatility and growth to be non-negative. However, estimates of ¯ 2 are su¤ering from a possible bias, due to the presence of lagged dependent variables in (3). The middle panel in table 3 uses the lagged level of relative value added to instrument for its lagged growth, following Anderson and Hsiao (1981). The results indeed point to an important bias, as ¯ 2 now becomes non-signi…cantly di¤erent from zero in the OECD, whereas it is still negative, but less signi…cantly so, in the extended sample. As for volatility, its e¤ects become virtually zero in most cases. It is however not impossible that this could largely result from the insu¢cient …t of the chosen instruments, as the very low values for …rst-stage R2 suggest in the penultimate row of the middle panel. Extending the set of available instruments is exactly the purpose of the Arellano-Bond GMM estimator, which uses all available lags of the dependent variables as instruments (both in levels and in di¤erences). We report the results of their estimation procedure in the lower panel of Table 3. Several comments are in order: …rstly, ¯ 1 is now positive and signi…cant in all cases. Secondly, the convergence term becomes insigni…cant in the extended sample, and actually signi…cantly positive in the OECD. Thirdly, the coe¢cient on the comparative advantage variable is signi…cantly negative in both samples, a somewhat surprising result. We note however that the signi…cance of the other variables, and in particular the positive estimates of ¯ 1 do not depend on the inclusion of the comparative advantage variable. Finally, Sargan’s tests fail to reject the hypothesis that overidentifying restrictions are satis…ed in both samples, at any standard level of signi…cance. As already mentioned, second-order serial correlation in the residuals of equation (3) is liable to cast doubt on the consistency of the Arellano-Bond 21 The number of observations is lower than in table 2 b ecause of countries missing observations in 1982. 10 estimation. Standard (unreported) t-tests suggest this is a serious problem in the extended sample, where the hypothesis of no serial correlation is rejected in both speci…cations.22 In the reduced sample, however, it is substantially harder to reject the hypothesis of no second-order serial correlation in the residuals. 23 We conclude the results from the OECD sample of countries are probably to be taken most seriously, although we notice that the coe¢cient on volatility is positive in both cases. The maintained assumption in the previous estimations has been that volatility is strictly exogenous to the dependent variable, output growth. While we come back to the issue of endogeneity, and the bias in ¯ 1 that may result from it, in section 5, the ‡exibility of the Arellano-Bond procedure o¤ers a …rst step towards addressing the concern. The estimation remains ³ ´ yi;j;t P consistent even when allowing for the possibility that VT ¢ ln be i yi;j ;t correlated with "i;j;S for all T > S. This makes it possible to consider current volatility as a manifestation of past high growth. We perform once again the Arellano-Bond estimation, eschewing the assumption of strict exogeneity, and compare its performance to the one reported in Table 3. The resuts, unreported for clarity, are quite clear. In all cases, allowing for volatility to be predetermined results in rejection of the overidentifying restrictions at standard con…dence levels. 24 Furthermore, serial correlation becomes a serious problem in the OECD sub-samples as well. The evidence thus favours a speci…cation that assumes perfect exogeneity of output volatility, relative to one that allows volatility to be predetermined. 3 Why the Disaggregated Link is Positive A common justi…cation for a putative positive link between growth and volatility is the presence of so-called “Schumpeterian waves”, whereby the adoption of a new technology entails high volatility as the economy adjusts to its introduction. A direct implication is that the link between sectoral 22 The test-statistics are N(0; 1) = 3:88 and 2:62 depending whether the comparative advantage variable is included. This rejects the null at 1% con…dence level. See Arellano and Bond (1991) for details on the serial correlation tests. 23 The test-statistics are now N(0; 1) = 1:20 and 1:52, resp ectively. This rejects the null at only 23% and 13% con…dence levels. 24 In the order of the sp eci…cation in the low panel of table 3, the p-values of the Sargan tests become 0:093, 0:112, 0:002 and 0:023, respectively. 11 volatility and growth should vary systematically with the sectoral rate of (total factor) productivity growth. In particular, the higher productivity growth, the higher the positive correlation between growth and volatility. In this section we investigate the empirical support for this prediction. 3.1 Sample-Splits We do this in two ways. Firstly, we split our cross-sections according to sectoral measures of productivity growth. For lack of reliable data on capital accumulation at the sectoral level in the UNIDO data, we use labour productivity growth. For each sample, we present estimation results for both (2) and (3), i.e. …rstly for a purely cross-sectional estimation with country-sector …xed e¤ects, and secondly using the Anderson-Hsiao and the Arellano-Bond estimation methods. Each sample is split according to the median level of labour productivity growth across country-sectors. Table 4 presents the results for both samples in the UNIDO data. One main result transpires from all estimation methods: the coe¢cient on volatility ¯ 1 is systematically higher in the high productivity growth sub-sample. Thus, volatility and growth are related positively mostly in country-sectors with high labour productivity growth. Secondly, the “convergence” term is signi…cantly negative in all cases in the full sample, and weaker in the reduced OECD sample. Thirdly, the comparative advantage term enter weakly in most cases, and sometimes with the wrong sign. The rest of this section discusses these results in more details. The cross-sectional estimations yield clear results: in the full sample, ¯ 1 actually changes signs, from signi…cantly negative to signi…cantly positive as one moves from low to high productivity growth sectors. In the reduced sample, the coe¢cient increases in magnitude, from being virtually zero to signi…cantly positive (as well as a point estimate almost twice larger). The Anderson-Hsiao estimations yield similar results, if somewhat weaker. In the extended sample, ¯ 1 goes from zero to signi…cantly positive (and a point estimate twice larger as well), whereas in the reduced sample it is nonsigni…cantly di¤erent from zero in both sets of sectors, although a negative (positive) point estimate obtains in low (high) productivity sectors. Once again, the instrument set used by the Anderson-Hsiao estimator is weak, particularly so in low productivity sectors. Nevertheless, the expanded instrument set used by the Arellano-Bond estimator leads to the same conclusion: from virtually zero in low productivity sectors, ¯ 1 becomes signi…cantly 12 positive in high-productivity sectors, or at the very least the point estimate doubles in the OECD sample. Arellano-Bond estimations are well-speci…ed, as the overidentifying restrictions are never rejected.25 Serial correlation continues to be a concern for the full sample estimates, although only in high productivity sectors, but once again we fail to reject the hypothesis of no serial correlation in the reduced sample.26 3.2 Non-Parametric Estimations A more general approach involves estimating the link between growth and volatility over rolling sub-samples of country-sectors, ordered by increasing rates of productivity growth. This avoids imposing an arbitrary split on the data, and makes it possible to investigate the shape of the relationship between growth and volatility with as few restrictions as possible on the functional form. In particular, for each observation pair (y n; xn) on sectoral growth and volatility, we run a regression akin to (2) or (3) using a small amount of data around xn and collect the corresponding estimate of ¯ 1. We then repeat the operation for the next pair of observations on growth and volatility, ordered by increasing rate of labour productivity growth. Thus, we investigate in a much more general manner than in the previous section whether estimates of the link between growth and volatility display any systematic tendency to increase with the sectoral rate of productivity growth. The procedure derives closely from robust locally weighted scatterplot smoothing (lowess ), and requires an arbitrary choice of two parameters: the size, or “bandwidth”, of each sub-sample, and a step parameter governing the extent of overlap between two successive sub-samples. There are of course many ways of choosing these, but we checked that the general shape of the obtained curve did not depend on any particular parameter choice. In general, kernel estimations further require the choice of a weighing scheme that assigns (often decreasing) weights onto observations farther away from the central point of each sample. The present procedure is akin to a “rectangular” or “‡at” kernel procedure, where equal weights are assigned to observations within the bandwidth, whereas observations outside are disregarded altogether. 27 Another advantage of the approach relative to 25 The p-values associated with the Â2 tests reported in the table are 0.306, 0.134, 0.385 and 0.661, in the same order as in table 4. 26 In the order of Table 4, the t-statistics are N(0; 1) = 1:41, 2:66, 1:38 and 0:92. 27 For more details, see Imbs and Wacziarg (2002). 13 polynomial or semi-parametric estimations is that it is locally robust: the shape of the relationship linking growth and volatility in, say, low productivity sectors is not a¤ected by observations in high productivity sectors. By the same token, it highlights the possible role for outliers. Figures 4 through 9 report the non-parametric estimates of ¯ 1 over sub-samples ordered by increasing average labour productivity growth over [1970;1992]. For each sample, we report results for three estimations: the pure cross-sectional approach in equation (2), the within-OLS estimation of (3), and the Arellano-Bond estimation of (3).28 The bold sections of the curves correspond to estimates of ¯ 1 signi…cant at the 5% con…dence level, and the X-axis reports average labour productivity growth between 1970 and 1992 in the median sector of the considered sub-sample. The non-parametric evidence con…rms the results based on sample splits. In the extended sample, estimates of ¯ 1 increase with labour productivity growth in the three cases. As expected, measurement error is more prevalent in dynamic estimations, that rely on measures of growth and volatility based on shorter time periods; this is manifested by sudden jumps in the estimates of ¯ 1 in …gures 5 and 6. In spite of noise, however, the trend is strikingly positive in all three cases, and does not depend on the choice of a bandwidth.29 Results in the reduced sample are slightly di¤erent, and appear to be obscured by the presence of a few outliers in low productivity growth sectors. The overall shape of the curve is ‡at in …gure 7 and 8, corresponding to the cross-sectional and within estimators. This explains why parametric estimates of ¯ 1 were more signi…cantly positive in the OECD sample: crosssectoral heterogeneity appears to be less of a problem there. Nevertheless, the Arellano-Bond estimator con…rms the presence of an upward trend in ¯ 1, partly masked by the presence of a few outliers in low productivity growth sectors, appearing to display both high growth and high volatility.30 Insofar as high rates of labour productivity growth re‡ect technological progress, 28 The weakness of instruments in the Anderson-Hsiao estimations resulted in essentially ‡at non-parametric estimations. 29 Figure 4 is based on a sample of 714 observations, and the bandwidth is 300. Figure 5 is based on a sample of 440 observations, with a bandwidth of 200. Figure 6 is based on a sample of 714 observations, and a bandwidth of 300. In all three cases, the step parameter is two observations. The general upwards shape prevails for alternative (reasonable) parameters. 30 The apparent culprit is Petroleum Re…neries (ISIC 353). When this activity is suppressed from the sample (which entails 6 lost observations), the initial section where ¯1 is signi…cantly positive disappears, while the subsequent upward trend remains prevalent. It is reassuring that the outlier should b e a sector whose activity is closely related to the price of a commodity, and ‡uctuations thereof. 14 the evidence in this section suggests the positive link between growth and volatility originates in “Schumpeterian waves”, or “creative destruction”, whereby the introduction of better technologies is accompanied by sectoral volatility. In the next section, we investigate whether it is also accompanied by net …rm entry. 3.3 The Mechanics of Creative Destruction In their survey on “creative destruction”, Caballero and Hammour (2000) argue factors shifting from low to high productivity activities account for a large share of productivity growth; furthermore, most of this factor reallocation occurs within narrowly de…ned sectors, contrary to a ascribing this role to structural change, or factor reallocation across sectors. Our results are perfectly consistent with this evidence, as they underline the association of within-sector volatility with high (productivity) growth. Several studies have attempted to further decompose these productivity e¤ects into net entry of new …rms endowed with a better technology, and learning on the part of incumbents. Foster, Haltiwanger and Krizan (2001), for instance, …nd a roughly equal contribution of both in the US manufacturing data. On the other hand, the industrial organization literature provides theoretical arguments that newer technologies actually tend to be implemented by new entrants.31 The question whether creative destruction is accompanied by a wave of new entrants can be addressed in a natural way in the present context, with two empirical shortcuts. As already mentioned, the UNIDO data provides information on the number of establishments per sector. The relationship between establishments and …rms is a loose one, as several establishments could very well be part of one …rm. Thus, growth in the number of establishments captures net entry imperfectly at best. Furthermore, the number of establishments is available from 1981 only, thus reducing the time series over which average growth rates are computed to a little more than ten years on average, less than half the data otherwise available. 32 With those two caveats however, it is possible to decompose output growth into an intensive (output per …rm) and an extensive (number of …rms) margins, and examine whether volatility is related with the former or the latter. Under the premise that the positive link between (sectoral) volatility and growth 31 See for instance Jovanovic and Nyarko (1996) or Asplund and Nocke (2001). This is also the reason why no dynamic estimations are possible for this growth decomposition. 32 15 re‡ects creative destruction, documented in the previous section, this may clarify the mechanism whereby technologies are adopted. Table 5 reports the results of this growth decomposition in both UNIDO samples when estimating (2), i.e. controlling for country and sector- e¤ects. They are unequivocal: the link between extensive growth and volatility is signi…cantly positive in both cases, as if entering …rms were the innovative units, and the source of sectoral volatility. The link between intensive growth and volatility, on the other hand, is non-signi…cant at best. Interestingly, the positive role for the comparative advantage variable, evident in table 2 for the full sample, seems to correspond to extensive growth as well, as if net entry were stimulated in those sectors intensive in a factor accumulating relatively fast in the aggregate. Taken together, the evidence in this section suggests the link between growth and volatility is an important symptom of technology adoption at the sectoral level, and con…rms that, at least in manufacturing, new technologies are introduced by new entrants, or at least by productive units newly created by incumbent …rms. 4 Why the Aggregate Link is Negative Our result stands in stark contrast with previous similar exercises. RR and MR used a number of aggregate cross-sections to establish what appears to be an extremely robust feature of cross-country data: aggregate growth (conditionally) correlates negatively with aggregate volatility. Both papers refer extensively to the empirical growth literature to choose the conditioning set. The choice of these controls has been the focus of an enormous literature, culminating with Levine and Renelt (1992), who suggested the determinants of aggregate growth are (i) initial income, (ii) average investment share of GDP, (iii) average population growth and (iv) initial human capital. When conditioning GDP growth on this set of variables, both papers …nd high variance of GDP to be associated with low aggregate growth on average, although the unconditional correlation is actually positive for a sub-sample of OECD countries. The two papers di¤er somewhat in their methodology, but not fundamentally in their results. RR implement a maximum likelihood estimation and focus on the variance of innovations to GDP, using the time series dimension of the data, while MR focus purely on a cross-section and 16 deviate somewhat from the Levine-Renelt conditioning variables.33 Both papers show the negative impact of volatility is strongest for a sub-set of OECD countries.34 In this section, we reconcile our results with theirs, by examining the role of aggregation in the estimation. 4.1 Aggregation First consider table 6. There, we report results that are meant to reproduce the aggregate evidence. In particular, we estimate à ! X X X ln yi;j;T = ¯ 1 VT ¢ ln yi;j;t + (¯ 2 + 1) ln yi;j;T ¡1 + ®j + (±T ) + "j;T i i i (4) Thus, we investigate the relationship between aggregate growth and volatility. Admittedly, the aggregates used here di¤er somewhat from standard GDP measures, as they are sums over manufacturing activities only; the exercise remains however interesting, at least to document the e¤ects of aggregation if not for direct comparison with RR or MR. One di¢culty in using an aggregate of manufacturing activities is the choice of a conditioning set: the Levine-Renelt variables pertain to aggregate GDP, and thus do not belong in (4). We choose to include initial conditions only, as well as country …xed e¤ects and period indicator variables. That choice is of course mainly governed by data availability: to our knowledge, there are no equivalent to the Levine-Renelt variables applicable to the secondary sector. Our hope is that time invariant …xed e¤ects capture other relevant conditioning variables.35 The results in table 6 stand in stark contrast with the disaggregated evidence. Whereas ¯ 1 is not signi…cantly di¤erent from zero in simple unconditional OLS, dynamic panel estimations (i.e. OLS-within and Arellano-Bond) give rise to signi…cantly negative estimates of ¯ 1 . As in 33 To be precise, Martin and Rogers use cross-region as well as cross-country evidence, and include a measure of sectoral shares to account for putative transitional e¤ects. They also include the share of government sp ending and a measure of political instability, apparently without much e¤ect (results without a share of government are not presented in the paper). 34 MR …nd a coe¢cient non-di¤erent from zero for a large sample of 97 countries; RR …nd a substantially lower (in absolute value) negative coe¢cient in a sample of 92 countries than in a sample of 24 OECD countries. 35 The next section o¤ers sensitivity analysis involving data from the OECD, covering all economic activities. This will b etter address the concern of the conditioning set. 17 RR and MR, this is particularly true in a reduced sample of OECD. As expected, initial income enters negatively, particularly in the extended sample inclusive of developing economies. Sargan tests suggest the Arellano-Bond estimator is better speci…ed in the reduced OECD sample, where ¯ 1 is most negative; it is also there that serial correlation is least present.36;37 We now turn to an account of the discrepancy. One immediate possibility is related with the components of aggregate volatility. Inasmuch as it is a sum of sectoral value addeds, the variance of GDP embeds both within sectors variances and cross-sectoral covariances. We …rst investigate this possibility, by performing the following approximate decomposition X X X ¢ ln yi;j;t ' !i;j;0 ¢ ln yi;j;t + (!i;j;t ¡ ! i;j;0) ¢ ln yi;j;t i i i + X ¢!i;j;t ln y i;j;t¡1 (5) i y with ! i;j;t = P i;j;t .38 There are two conceptually distinct components of i yi;j;t aggregate output growth in (5): the …rst term captures growth originating within each sector, weighted using the initial sectoral structure. The other two terms involve growth e¤ects of structural change, involving changes in the sectoral structure of the economy. In table 7, we decompose volatility in estimations of (4) making use of these two elements. P In particular, we investigate whether the estimated coe¢cient on VT ( i ! i;j;0 ¢ ln yi;j;t) di¤ers from P that on the complementary component in aggregate volatilP ity, VT (¢ ln i yi;j;t ) ¡ VT ( i !i;j;0 ¢ ln yi;j;t ), which embeds the e¤ects of structural change. The idea is to verify whether di¤erent components of aggregate volatility relate di¤erently with aggregate growth. The results are striking: estimates in the upper panel of table 7 suggest that, in dynamic estimations, even a weighted sum of within-sector volatilities only, still correlates negatively with aggregate growth.39 The lower panel of table 7 goes 36 We fail to reject the null of no serial correlation at any standard con…dence level in the OECD sample, but reject at 6% in the extended sample. 37 All estimates are also allowing for heteroskedasticity, for the number of sectors is di¤erent across countries. P P 38 We made use of the approximation ln i yi;j;t ' i !i;j;t ln yi;j; t. The correlation between the two measures in 0.997 in our data, and OLS of one on the other gives an R2 of 0.995. 39 ¯1 is zero in OLS conditional on initial aggregate value added. The Arellano-Bond estimator points to the presence of serial correlation in the extended sample, but, once again, not in the reduced one. 18 one step further, and takes country averages of sector by sector estimations, using initial relative outputs as a weighting device. Thus the measures of aggregate growth and volatility now both contain within sector developments only. In particular, we estimate the following using our cross-section, OLS-within and Arellano-Bond: X X X ! i;j;0 ln yi;j;T = ¯ 1 (! i;j;0 )2 VT (¢ ln y i;j;t ) + (¯ 2 + 1) !i;j;0 ln yi;j;T ¡1 i i i +®0j + (±0T ) + "0j;T (6) The results still point to non-positive estimates of ¯ 1 , if slightly less unanimously. This suggests the key to the discrepancy does not lie in a decomposition of aggregate growth or volatility into within- and between-sector components. Country (weighted) averages of sectoral growth and volatility tend to correlate negatively. This points to the possibility that cross-sectional heterogeneity, and failure to account for it in aggregate estimations, is a candidate explanation to the discrepancy. 4.2 Compositional E¤ects In this section, we establish the importance of heterogeneity in the estimates of ¯ 1. There are two ways in which it can matter for aggregate estimations. Firstly, assuming economic structure is identical across countries, ¯ 1 could be non-positive in (4) because of failure to account for the possibility that estimates are di¤erent across the sectors composing the aggregate economy of country j. Alternatively, assuming no heterogeneity across sectors, it could be that ¯ 1 should be allowed to di¤er across countries, simply because sectoral composition of aggregate output does as well. This question is easy to address, by implementing random coe¢cient estimations of (3), as introduced for instance by Hildreth and Houck (1968). This procedure estimates a typical relationship Yi = Xi Bi + ei where E(ei) = 0, E(eie0i ) = ¾2I and Bi = B +vi is allowed to vary randomly along the i dimension, E(vi ) = 0 and E(vivi0 ) = ¡. Hildreth and Houck show the model is equivalent to Yi = Xi B + ´ i 19 with E(´i ) = 0 and E(´i ´ 0i) = ¾2I + Xi ¡ Xi0 which can be estimated immediately using Generalized Least Squares. In what follows, we use this procedure in a panel context to obtain estimates of ¯ 1 in three di¤erent setups: (i) for comparison purposes, a simple estimation of growth on volatility without any allowances for heterogeneity, (ii) an estimation allowing for ¯ 1 to depend on i and (iii) an estimation allowing for ¯ 1 to depend on j.40 The results, reported in Table 9, are striking. Whereas estimates of ¯ 1 continue to be signi…cantly positive in a simple random e¤ects estimation, as well as, in most cases, when sectoral heterogeneity is allowed for, they become non signi…cantly di¤erent from zero when country-speci…c heterogeneity is allowed.41 We draw two inferences from these results: …rstly, the sectoral link evidenced in the previous section, although it is heterogenous between low and high productivity growth sectors, is a robust feature of the data. Secondly, the nature of the aggregate link between growth and volatility crucially depends on the sectoral composition of aggregate activity. The composition of aggregate output, and presumably the balance of low versus high productivity growth activities, are liable to have drastic consequences on the aggregate evidence on this question. We leave a detailed investigation of this possibility open for further research. 5 Extensions and robustness 5.1 An Alternative Dataset Here, we supplement the evidence based on UNIDO data with the International Sectoral DataBase (ISDB) published by the OECD, which covers all activities in fourteen OECD economies, but only at the two-digit level. The reduced coverage, focused on developed economies only, is an asset from the standpoint of robustness analysis, for it provides an alternative source of information on a homogenous group of economies. Furthermore, while ISDB does not supply the number of establishments per sector, it does actually contain indices of Total Factor Productivity (TFP). Time coverage is similar 40 The …rst estimation without heterogeneity is very similar to (3), although not exactly. In particular, it allows for random intercepts, and thus is not directly comparable with previous estimations. We report the corresponding results to ascertain the role of heterogeneity. 41 The table also reports Â2 statistics to test whether estimates of ¯1 are signi…cantly heterogenous in the dimension under consideration. For details, see Johnston (1984). 20 to the UNIDO sample, with complete data from 1970 to 1992. Summary statistics are reported in the lower panel of table 1, and the unconditional correlation between growth and volatility is plotted in …gure 3. We …rst perform our parametric estimations using the ISDB data, i.e. cross-sectional estimations of (2) and OLS-within, Anderson-Hsiao and ArellanoBond estimations of (3). The results are reported in table 9, and are less encouraging than those based on the UNIDO data. ¯ 1 is never signi…cantly di¤erent from zero at usual con…dence levels, and the point estimates are most of the time negative. However, the presence of heterogeneity across country-sectors is liable to bias the estimations, possibly more here than in the UNIDO dataset which is more homogenous in its coverage of sectors. In other words, non-parametric estimations are probably even more important here than in section 2. As a …rst pass, table 10 reports estimates where we split our cross-sections according to sectoral averages of both labour and total factor productivity growth. The upper panel reports the result for a split according to the median sectoral TFP growth. Results are somewhat more encouraging, as both the cross-sectional and the Anderson-Hsiao estimators suggest substantial heterogeneity across sub-samples, with ¯ 1 increasing with average sectoral TFP growth. Results for the Arellano-Bond estimators, however, are only reported for completeness, as the speci…cation is systematically rejected by the Sargan test in the low TFP growth subsample. The conclusions using a split based on the median sectoral labour productivity growth are simillar and reported in the lower panel of table 10. Estimates of ¯ 1 increase with sectoral productivity growth, except using the Arellano-Bond estimator, which is once again mis-speci…ed according to the Sargan test. Figure 10 through 15 report the non-parametric estimates of ¯ 1 corresponding to sub-samples of sectors ordered by increasing average productivity growth, using the estimator described in section 2. Figures 10 to 12 present the results when productivity growth is measured by the TFP indices, while …gures 13 to 15 use labour productivity growth.42 As the sample splits in table 10 suggested, the evidence that estimates of ¯ 1 increase with productivity growth is supported by both cross-sectional and within estimations. The Arellano-Bond estimator yields opposite results when using TFP 42 The total numbers of observations for the six …gures were 181, 179, 177, 226, 224 and 211, respectively. Unsurprisingly, there are more observations on labour productivity growth than on TFP indices. The bandwidthes were 100, 75, 75, 75, 100 and 75, and the step parameter set at one observation in all cases. Results are robust to alternative parameters choice. 21 indices, but, as in table 10, there is every indication that it is mis-speci…ed as overidentifying restrictions are actually rejected, once again particularly in sectors with low TFP growth.43 The Arellano-Bond does however lead to similar conclusions when using labour productivity, with an upward, if insigni…cant, trend in ¯ 1 . These estimates are however merely reported for completeness, as overidentifying restrictions are rejected in Figure 15 as well. Overall, results using the ISDB dataset are weaker, but not inconsistent with those based on manufacturing sectors only. Heterogeneity seems to be more of an issue when using data covering all sectors of the economy. Further, aggregation in the ISDB data is coarser to start with than in the UNIDO, leading to the possibility that disaggregated results there already su¤er from the bias we document in the previous section. Table 11 reproduces estimations of (4) using aggregate ISDB data; we have data on all sectoral activities in 14 countries, and thus can control for the variables Levine-Renelt showed to be robust determinants of GDP growth. Once again, a negative aggregation bias is apparent, as panel estimates of ¯ 1 clearly become signi…cantly negative, whether or not the standard control variables for economic growth are included. 5.2 Sensitivity Analysis To be completed. Clustered Standard Errors, ARCH and endogeneity, Random e¤ects. 6 Conclusion In this paper, we reconcile results spanning two literatures. Microeconomists, particularly students of industrial organization, often think of volatility in economic activity as a manifestation of change, for instance in the guise of technology adoption, and the entry and exit of …rms that accompanies it. Macroeconomists, however, have traditionally associated economic volatility with risk, liable to deter investment, hamper the adoption of new practices, or simply re‡ecting policy instability. In this paper, we show that the 43 For clarity, the actual values of the Sargan tests are not rep orted. It is a fair approximation, however, to consider all positive estimates of ¯ 1 in …gure 12 as corresponding to speci…cations rejected at the 5% con…dence level. 22 two approaches are not necessarily mutually exclusive. We use the same dataset to generate both a positive and a negative link between growth and volatility. We show that volatility and growth are particularly positively related in activities that display fast productivity growth, as implied by the Schumpeterian theory of creative destruction. We further establish the importance of accounting for sectoral heterogeneity when assessing the link between growth and volatility at the country level. 23 Appendices: A. Sectoral Coverage 1. UNIDO 3-Digit Classi…cation (28 sectors) 300 311 313 314 321 322 323 324 331 332 341 342 351 352 353 354 355 356 361 362 369 371 372 381 382 383 384 385 390 Total manufacturing Food products Beverages Tobacco Textiles Wearing apparel, except footwear Leather products Footwear, except rubber or plastic Wood products, except furniture Furniture, except metal Paper and products Printing and publishing Industrial chemicals Other chemicals Petroleum re…neries Miscellaneous petroleum and coal products Rubber products Plastic products Pottery, china, earthenware Glass and products Other non-metallic mineral products Iron and steel Non-ferrous metals Fabricated metal products Machinery, except electrical Machinery, electric Transport equipment Professional and scienti…c equipment Other manufactured products 2. OECD 2-Digit Classi…cation (20 sectors) 100. 200. 310. 320. Agriculture, hunting, forestry and …shing Mining and quarrying Food, beverages and tobacco Textiles, wearing apparel and leather industries 24 330. Wood and wood products, including furniture 340. Paper and paper products, printing and publishing 350. Chemicals, chemical petroleum, coal, rubber, plastic products 360. Non-metallic mineral products excl. products of petroleum & coal 370. Basic metal industries 380. Fabricated metal products, machinery and equipment 390. Other manufacturing industries 400. Electricity, gas and water 500. Construction 610+620. Wholesale trade and retail trade 630. Restaurants and hotels 710. Transport, storage and communication 810+820 . Financial institutions and insurance 830. Real estate and business services 910. Community and Social services 920. Personal services 25 B. Geographic Coverage Australiaa Austria Bangladesh Belgium a;b Canada a Chile Colombia Costa Rica Cyprus Denmarka Egypt Finlanda Francea;b Germanya Greece Hong Kong India Indonesia Ireland Italya;b Japana Jordan Kenya Korea Malaysia Mexico Netherlandsa Norwaya;b Pakistan Panama Philippines Portugal Singapore South Africa Spain Swedena USAa United Kingdoma Uruguay Zimbabwe a: in OECD dataset b: not in UNIDO data 26 References Aghion, Philippe and Peter Howitt (1996), Endogenous Growth Theory, Cambridge: MIT Press. Alesina, Alberto, Sule Ozler, Nouriel Roubini and Philip Swagel (1996), Political Instability and Economic Growth, Journal of Economic Growth, Vol.2, pp.189-213. Anderson, T.W. and Cheng Hsiao (1981), Estimation of Dynamic Models with Error Components, Journal of the American Statistical Association, Vol.76, pp.589-606. Arellano, Manuel and Stephen Bond (1991), Some Tests of Speci…cation for Panel Data: Monte Carlo Evidence and an Application to Employment Data, Review of Economic Studies, Vol.58, pp.277-297. Asplund, Marcus and Volker Nocke (2001), Imperfect Competition, Market Size and Firm Turnover, CEPR Discussion Paper No. 2625. Basu, Susanto and John Fernald (1997), Returns to Scale in U.S. Production: Estimates and Implications, Journal of Political Economy, Vol.105(2), pp.249-283. Bernard, Andrew and Bradford Jensen (2001), Who Dies? International Trade, Market Structure and Industrial Restructuring, NBER Working Paper No. 8327. Burnside, Craig, Martin Eichenbaum and Sergio Rebelo (1996), Sectoral Solow Residuals , European Economic Review, Vol.40, pp.861-869. Caballero, Ricardo and Mohammed Hammour (2000), Creative Destruction and Development: Institutions, Crises and Restructuring, NBER Working Paper No. 7849. Caselli, Francesco, Gerardo Esquivel and Fernando Lefort (1996), Reopening the Convergence Debate: A New Look at Cross-Country Growth Empirics, Journal of Economic Growth, Vol.1, pp.363-389. Cuñat, Alejandro (2000), Sectoral Allocation in the Process of Growth, Mimeo London School of Economic and Political Science. Davis, Steven and John Haltiwanger (1998), Gross Job Flows, in Orley Ashenfelter and David Card, eds. Handbook of Labour Economics. Forbes, Kristin (2000), A Reassessment of the Relationship Between Inequality and Growth, American Economic Review, Vol.90(4), pp.869-887. 27 Foster, Lucia, John Haltiwanger and C.J. Krizan (2001), Aggregate Productivity Growth: Lessons from Microeconomic Evidence, in Edward Dean, Michael Harper and Charles Hulten, eds. New Directions in Productivity Analysis, University of Chicago Press. Helpman, Elhanan and Manuel Trajtenberg (1998), A Time to Sow and a Time to Reap: Growth based on General Purpose Technologies, in Elhanan Helpman, ed. General Purpose Technologies and Economic Growth, Cambridge: MIT Press. Hildreth, C. and C. Houck (1968), Some Estimators for a Linear Model with Random Coe¢cients, Journal of the American Statistical Association, Vol.6, pp.158-162. Imbs, Jean and Romain Wacziarg (2002), Stages of Diversi…cation, American Economic Review, forthcoming. Johnston, J (1984), Econometric Methods, New York: Mc Graw Hill. Jovanovic, Boyan and Yaw Nyarko (1996), Learning by Doing and the Choice of Technology, Econometrica, Vol.64(6), pp.1299-1310. Jovanovic, Boyan and Rafael Rob (1990), Long Waves and Short Waves: Growth through Intensive and Extensive Search, Econometrica, Vol.58, pp.13911409. Levine, Ross and David Renelt (1992), A Sensitivity Analysis of CrossCountry Growth Regressions, American Economic Review, Vol.82(4), pp.942963. Martin, Philippe and Carol Ann Rogers (1997), Stabilization Policy, Learning by Doing and Economic Growth, Oxford Economic Papers, Vol.49(2), pp.1520-166. Martin, Philippe and Carol Ann Rogers (2000), Long-Term Growth and Short-Term Economic Instability, European Economic Review, Vol.44(2), pp.359-381. Pindyck, Robert (1991), Irreversibility, Uncertainty and Investment, Journal of Economic Literature, Vol.29(3), pp.1110-1148. Rajan, Raghuram and Luigi Zingales (1998), Financial Dependence and Growth, American Economic Review, Vol.88(3), pp.559-586. Ramey, Valerie and Garey Ramey (1991), Technology Commitment and the Cost of Economic Fluctuations, NBER Working Paper No. 3755, June. 28 Ramey, Valerie and Garey Ramey (1995), Cross Country Evidence on the Link between Volatility and Growth, American Economic Review, Vol.85(5), pp.1138-1159. Ventura, Jaume (1997), Growth and Interdependence, Quarterly Journal of Economics, Vol.112(1), pp.57-84. Young, Alwyn (1991), Learning by Doing and the Dynamic E¤ects of International Trade, Quarterly Journal of Economics, August, pp.369-405. 29 Table 1: Summary Statistics Extended UNIDO Sample (715 obs – 36 countries) Mean Std. Dev Min Average Yearly -0.016 3.385 -16.070 Relative Growth Average Yearly 3.521 4.738 0.047 Relative Variance Correlation = 0.144 Reduced UNIDO Sample (317 obs – 15 countries) Mean Std. Dev Min Average Yearly Relative Growth Average Yearly Relative Variance Correlation = 0.118 Average Yearly Relative Growth Average Yearly Relative Variance Correlation = -0.098 Max 17.635 37.558 Max -0.282 2.639 -16.070 9.995 1.205 2.188 0.047 21.051 ISDB Sample (232 obs – 14 countries) Mean Std. Dev Min Max -0.367 1.916 -6.130 5.331 0.296 0.375 0.007 2.905 Table 2: Country and Sector Effects A. Extended Sample (715 obs) (i) (ii) Vt (∆lnyi,j,t ) 1.974 (2.78) 1.966 (2.77) ln(yi,j,0) (iii) -0.156 (0.21) -0.201 (7.80) ln(yi,j,0 / Σ i yi,j,0) Comp. Adv. Intercept R-Square 0.072 (2.15) 0.011 0.115 (1.13) 0.071 (2.13) 0.013 B. OECD Sample (317 obs) (i) (ii) Vt (∆lnyi,j,t ) 4.893 (2.63) 5.007 (2.66) ln(yi,j,0) 0.223 (2.24) -3.864 (7.65) 0.091 (iii) 4.072 (2.16) -0.118 (3.00) ln(yi,j,0 / Σ i yi,j,0) Comp. Adv. Intercept R-Square 0.121 (3.68) 0.022 -0.059 (0.39) 0.133 (2.94) 0.022 0.102 (0.64) -2.425 (2.84) 0.049 (iv) -0.661 (0.91) -0.283 (9.46) 0.257 (2.63) 0.980 (9.69) 0.123 (iv) 3.832 (2.06) -0.204 (4.10) 0.062 (0.41) 0.797 (4.75) 0.072 Results for estimation of (2), with country and sector fixed-effects. yi,j,t is the real value added in sector i, country j at time t. The dependent variable is ln(yi,j,1992 /Si yi,j,1992) - ln(yi,j,1970/Si yi,j,1970 ). Initial values are measured in 1970, or the nearest data available. The comparative advantage variable (Comp. Adv.) is an interaction term between average aggregate capital- labour growth between 1970 and 1985, and measured capital share in 1970. t-statistics are reported between parentheses. Table 3: Dynamic Panel Estimations: OLS Within Volatility Initial Relative VA Comparative Advantage R-Square # Obs. AndersonHsiao Volatility Initial Relative VA Comparative Advantage First Stage R-Square # Obs. ArellanoBond Volatility Initial Relative VA Comparative Advantage Sargan Test # Groups Full Sample 0.863 (2.47) -0.967 (25.50) 0.515 641 OECD Sample 0.243 (0.62) -0.831 (19.07) 0.057 (1.45) 0.459 440 Full Sample 1.054 (2.19) -0.785 (2.52) 4.602 (5.28) -0.641 (11.74) 2.838 (2.13) -0.632 (10.80) 0.027 (0.56) 0.343 268 0.375 317 OECD Sample 9.655 (1.08) 1.417 (0.24) 0.028 -0.448 (0.79) -1.548 (5.57) 0.113 (2.09) 0.059 0.025 77.407 (0.13) 15.743 (0.14) -2.667 (0.14) 0.075 641 440 317 268 Full Sample 0.413 (3.20) -0.572 (4.62) ?2 (2)=0.58 651 OECD Sample 0.478 (2.93) -0.514 (4.29) -0.220 (4.56) ?2 (2)=3.97 607 2.751 (2.28) 0.545 (1.87) ?2 (4)=4.38 268 2.427 (2.92) 0.321 (2.66) -0.285 (2.61) ?2 (4)=3.55 268 The dependent variable is relative sectoral value added growth in [1970,1981] and [1982,1992] (boundary dates are 1970, 1976, 1982, 1988 and 1992 for the ArellanoBond estimator). Initial values are measured in 1970 and 1981 (1970, 1976, 1982 and 1988 for the Arellano-Bond estimator). Variances are computed over the corresponding two sub-periods (four for the Arellano-Bond estimator). AndersonHsiao uses ln(yi,j,70/Σ i yi,j,70) to instrument for ln(yi,j,81 /Σ i yi,j,81 )- ln(yi,j,70/Σ i yi,j,70). Arellano-Bond uses all available lags. t-statistics are reported between parentheses. The Sargan test of overidentifying restrictions is reported in the last row for the Arellano-Bond estimator; the number of lagged dependent variables used as instruments and the inclusion of time dummies were chosen to maximize the value of the Sargan test. Table 4: Sample Splits Full Sample Low Labour Productivity Growth High Labour Productivity Growth Vt ( ∆lnyi,j,t ) yi,j,0 / Σ i yi,j,0 Comp.Adv. Intercept R-Square Sargan Test CrossAndersonSection Hsiao (357 obs) (220 obs) -2.062 1.291 (2.36) (0.84) -0.237 -3.160 (6.06) (2.71) 0.202 0.302 (1.47) (1.75) 0.661 (4.94) 0.096 0.041* ArellanoCrossAndersonBond Section Hsiao (303 obs) (358 obs) (220 obs) 0.317 2.236 2.556 (1.44) (1.86) (2.15) -0.420 -0.311 -0.555 (2.71) (7.07) (2.17) -0.400 0.302 -0.016 (3.76) (2.23) (0.24) 1.285 (8.77) 0.176 0.183* ?2 (2)=2.37 ArellanoBond (303 obs) 1.099 (3.49) -0.251 (4.47) -0.166 (2.57) ?2 (2)=4.02 Reduced Sample Low Labour Productivity Growth High Labour Productivity Growth Vt ( ∆lnyi,j,t ) yi,j,0 / Σ i yi,j,0 Comp.Adv. Intercept R-Square Sargan Test CrossAndersonSection Hsiao (158 obs) (134 obs) 2.960 -0.410 (1.30) (0.16) -0.136 0.679 (1.91) (0.93) -0.101 -0.139 (0.44) (1.05) 0.488 (2.04) 0.042 0.076* ArellanoCrossAndersonBond Section Hsiao (134 obs) (159 obs) (133 obs) 1.469 5.391 3.286 (2.06) (1.73) (0.50) 0.127 -0.260 -0.934 (2.48) (3.94) (1.51) -0.232 0.343 0.036 (2.68) (1.78) (0.21) 1.059 (4.76) 0.133 0.202* ?2 (4)=4.16 ArellanoBond (133 obs) 4.072 (2.37) 0.035 (1.64) -0.277 (1.12) ?2 (4)=2.41 t-statistics are reported between parentheses. The cross-section is split according to the median sectoral labour productivity growth over [1970,1992]. The Sargan test of overidentifying restrictions is reported in the last row for the Arellano-Bond estimator; the number of lagged dependent variables used as instruments and the inclusion of time dummies were chosen to maximize the value of the Sargan test. *: first-stage R-square. Table 5: Intensive and Extensive Growth (Country-Sector Effects) Vt (∆lnyi,j,t ) yi,j,0 / Σ i yi,j,0 Comp. Adv. Intercept R-Square Full Sample (715 obs) Extensive Intensive Growth Growth 1.127 -0.200 (2.87) (0.44) -0.005 -0.088 (0.21) (3.47) 0.182 0.041 (2.50) (0.48) -0.106 0.488 (1.41) (5.57) 0.021 0.017 Reduced Sample (317 obs) Extensive Intensive Growth Growth 5.485 -1.434 (5.11) (1.26) 0.041 -0.036 (1.12) (0.86) -0.163 0.110 (1.34) (0.85) -0.180 0.265 (1.44) (1.99) 0.077 0.007 Extensive growth is measured by ln(Nesti,j,1992 /Nesti,j,1981 ), and intensive growth by ln(yi,j,1992 /Siyi,j,1992) - ln(yi,j,1981 /Siyi,j,1981) - ln(Nesti,j,1992 /Nesti,j,1981). Initial values are computed in 1981. t-statistics are reported between parentheses. Table 6: Aggregate Growth and Volatility Full Sample (36 countries) Vt (∆ lnSi yi,j,t ) OLS Within 24.683 (1.48) -14.991 (1.47) -0.320 (2.79) 0.279 Initial Income R Square / Sargan Test 0.097 ArellanoBond -1.910 (0.72) -0.459 (1.63) ?2 (5)=7.77 OECD Sample (15 countries) OLS Within -15.996 (0.22) -32.959 (2.28) -0.229 (1.46) 0.371 0.005 ArellanoBond -13.177 (2.05) 0.074 (3.30) ?2 (5)=1.15 The dependent variable is aggregate growth as measured by ∆lnSi yi,j,t between the relevant sub-periods (1970 and 1992 in OLS, 1970, 1982 and 1992 in the within estimation and 1970, 1976, 1982, 1988 and 1992 in the Arellano-Bond estimation). Aggregate volatility is computed over the corresponding sub-periods. Period dummy variables are included when meaningful. Huber-White heteroscedastic-consistent tstatistics are reported between parentheses. The reported value for the Sargan test corresponds to the homoskedastic case, since it is specified only under that assumption. Table 7: Aggregate Volatility Decomposition Full Sample (36 countries) Vt (Si? i,j,0 ∆lnyi,j,t ) Vt (∆Silnyi,j,t )-Vt (Si? i,j,0 ∆lnyi,j,t ) OLS Within 19.521 (2.32) -41.169 (2.33) -10.456 (1.35) -57.437 (2.56) -0.350 (3.01) 0.428 Initial Income R Square / Sargan Test 0.284 ArellanoBond -10.500 (2.98) 8.158 (2.29) -1.736 (1.34) ?2 (2)=3.08 Full Sample (36 countries) Si (? i,j,0 )2 Vt (∆lnyi,j,t ) OLS Within -0.815 (0.93) -40.445 (2.57) 0.037 (11.57) 0.403 Si ? i,j,0 lnyi,j,0 R Square / Sargan Test 0.020 ArellanoBond -1.179 (0.11) -1.685 (0.96) 2 ? (2)=2.67 OECD Sample (15 countries) OLS Within -14.090 (0.19) -58.271 (0.23) -32.240 (1.94) -41.563 (0.70) -0.242 (1.32) 0.372 0.007 ArellanoBond -12.852 (1.43) -14.915 (0.63) 0.073 (2.40) ?2 (5)=1.14 OECD Sample (15 countries) OLS 1.833 (0.23) 0.004 Within -55.674 (0.36) 0.037 (14.21) 0.480 ArellanoBond -123.86 (2.37) -0.556 (1.55) 2 ? (5)=4.92 In the upper panel, the dependent variable is aggregate growth as measured by ∆lnSi yi,j,t between the relevant sub-periods (1970 and 1992 in OLS, 1970, 1982 and 1992 in the within estimation and 1970, 1976, 1982, 1988 and 1992 in the Arellano-Bond estimation). The decomposition of aggregate volatility is computed over the corresponding sub-periods. ? i,j,0 is computed in 1981. In the lower panel, the dependent variable is aggregate growth as measured by Si ? i,j,0 ∆lnyi,j,t , computed over the same sub-periods. Period dummy variables are included when meaningful. HuberWhite heteroscedastic-consistent t- statistics are reported between parentheses. The reported value for the Sargan test corresponds to the homoskedastic case, since it is specified only under that assumption. Table 8: Random Coefficients Estimations Simple Random Effects Full Sample (36 countries) Within 2 Vt ( ∆ln yi,j,t ) Initial Income R Square Random Coefficients by Sector 1.359 (3.95) 0.009 (1.45) 0.034 Initial Income Identical Coefficients Random Coefficients by Country 2.185 (3.62) 0.021 (2.83) ?2 (81)=124.18 Initial Income Identical Coefficients Within 4 0.709 (3.22) 0.009 (2.42) ?2 (81)=134.44 Full Sample (36 countries) Within 2 Vt ( ∆ln yi,j,t ) 0.574 (4.64) 0.006 (2.09) 0.016 Full Sample (36 countries) Within 2 Vt ( ∆ln yi,j,t ) Within 4 Within 4 0.084 0.344 (0.06) (0.71) -0.047 -0.024 (1.89) (1.84) ?2 (102)=1264.14 ?2 (102)=950.89 OECD Sample (15 countries) Within 2 2.291 (2.50) -0.007 (0.84) 0.010 Within 4 0.562 (1.91) -0.003 (0.73) 0.003 OECD Sample (15 countries) Within 2 8.778 (2.42) 0.013 (0.58) ?2 (2)=284.79 Within 4 1.098 (0.78) 0.007 (0.77) ?2 (81)=327.65 OECD Sample (15 countries) Within 2 -0.906 (0.32) -0.003 (0.10) ?2 (39)=282.30 Within 4 0.135 (0.13) -0.013 (0.60) ?2 (39)=197.38 The dependent variable is sectoral growth as measured by ∆lnyi,j,t between the relevant sub-periods (1970, 1982 and 1992 in the within-2 estimation and 1970, 1976, 1982, 1988 and 1992 in the within-4 estimation). Period dummy variables are included when meaningful. Table 9: Estimations in ISDB Volatility Initial Relative VA Comparative Advantage R-Square # Obs CrossSection -11.597 (1.53) -0.117 (3.00) 0.241 (3.47) 0.074 230 OLS Within -4.452 (1.06) -0.354 (6.02) 0.143 232 -2.948 (0.68) -0.350 (5.81) -0.003 (0.10) 0.140 228 Anderson-Hsiao -5.515 (1.22) -0.128 (0.42) 0.045* 232 -2.105 (0.35) 0.494 (0.63) -0.082 (1.01) 0.059* 228 Arellano-Bond 1.572 (0.90) -0.407 (2.44) ?2 (2)=2.36 177 1.229 (0.74) -0.492 (2.53) -0.031 (1.15) ?2 (2)=2.89 177 Estimation of (3) with period dummies. The dependent variable is relative sectoral value added in 1970, 1981 and 1992 (1970, 1976, 1982, 1988 and 1992 for the Arellano-Bond estimator); variances are computed over the corresponding subperiods. Anderson-Hsiao uses ln(yi,j,70/Σ i yi,j,70 ) to instrument for ln(yi,j,81/Σ i yi,j,81 )ln(yi,j,70/Σ i yi,j,70 ). Arellano-Bond uses all available lags, and includes period dummy variables. t-statistics are reported between parentheses. *: first-stage R-square. Table 10: Sample Splits in ISDB: Low TFP Growth Vt ( ∆lnyi,j,t ) yi,j,0 / Σ i yi,j,0 Comp.Adv. Intercept R-Square Sargan Test CrossAndersonSection Hsiao (90 obs) (90 obs) -43.286 -8.115 (3.37) (1.19) -0.078 0.176 (1.29) (0.71) 0.326 -0.096 (3.35) (0.89) -0.034 (0.18_ 0.174 0.497* ArellanoBond (88 obs) 4.687 (1.58) -0.195 (6.92) -0.041 (1.11) yi,j,0 / Σ i yi,j,0 Comp.Adv. Intercept R-Square Sargan Test CrossAndersonSection Hsiao (90 obs) (90 obs) 3.372 15.224 (0.37) (2.02) -0.163 -0.141 (3.07) (0.33) 0.255 -0.072 (2.51) (1.65) 0.621 (3.70) 0.157 0.066* ?2 (2)=6.15 Low Labour Productivity Growth Vt ( ∆lnyi,j,t ) High TFP Growth CrossAndersonSection Hsiao (112 obs) (112 obs) -38.862 -10.826 (3.33) (1.71) -0.086 0.163 (1.62) (0.74) 0.200 -0.003 (2.38) (0.06) 0.092 (0.55) 0.118 0.244* ArellanoBond (88 obs) 0.099 (0.04) 0.005 (3.37) -0.072 (1.91) ?2 (4)=1.27 High Labour Productivity Growth ArellanoCrossAndersonBond Section Hsiao (105 obs) (113 obs) (112 obs) 0.446 3.323 14.578 (0.16) (0.34) (1.65) -0.203 -0.118 -0.351 (6.45) (2.25) (0.52) -0.031 0.317 -0.077 (1.55) (3.01) (1.20) 0.434 (2.59) 0.115 0.044* 2 ? (2)=5.39 ArellanoBond (105 obs) -0.141 (0.07) -0.804 (0.49) -0.007 (0.15) ?2 (2)=0.28 t-statistics are reported between parentheses. The cross-section is split according to the median sectoral labour productivity growth over [1970,1992]. The Sargan test of overidentifying restrictions is reported in the last row for the Arellano-Bond estimator; the number of lagged dependent variables used as instruments and the inclusion of time dummies were chosen to maximize the value of the Sargan test. *: first-stage R-square. Table 11: Aggregate Growth and Volatility in ISDB Full Sample (14 countries) OLS Levine-Renelt Controls No -3.829 (1.19) -0.027 (6.37) No R Square / Sargan Test 0.160 0.711 Vt (∆ lnSi yi,j,t ) 10.061 (3.17) Within Initial Income ArellanoBond -45.777 (1.80) -2.233 (3.64) No ?2 (2)=3.89 OLS Within 7.605 (0.10) 0.030 (1.44) Yes -52.226 (1.83) -0.496 (2.73) Yes ArellanoBond -26.405 (1.62) -1.785 (2.15) Yes 0.723 0.808 ?2 (5)=7.64 The dependent variable is aggregate growth as measured by ∆lnSi yi,j,t between the relevant sub-periods (1970 and 1992 in OLS, 1970, 1982 and 1992 in the within estimation and 1970, 1976, 1982, 1988 and 1992 in the Arellano-Bond estimation). Aggregate volatility is computed over the corresponding sub-periods. Period dummy variables are included when meaningful. Huber-White heteroscedastic-consistent tstatistics are reported between parentheses. The reported value for the Sargan test corresponds to the homoskedastic case, since it is specified only under that assumption. Figure 1: Unconditional Relative Growth and Volatility in UNIDO 0.4 Variance of relative growth rate 0.35 0.3 0.25 0.2 0.15 0.1 0.05 0 -4 -3 -2 -1 0 1 2 3 4 Relative growth rate Figure 2: Unconditional relative Growth and Volatility in reduced UNIDO 0.25 Variance of relative growth rate 0.2 0.15 0.1 0.05 0 -4 -3 -2 -1 0 1 2 3 Relative growth rate Figure 3: Unconditional Relative Growth and Volatility in ISDB 0.03 Volatility of relative growth rate 0.025 0.02 0.015 0.01 0.005 0 -1.5 -1 -0.5 0 Relative growth rate 0.5 1 1.5 Median Productivity Growth 1.03 1.01 0.99 0.96 0.94 0.91 0.89 0.86 0.84 0.83 0.81 0.79 0.78 0.76 0.74 0.73 0.70 0.68 0.66 0.64 0.63 0.61 0.59 0.57 0.56 0.54 0.53 0.51 0.49 0.47 0.44 0.42 0.39 0.37 0.35 0.33 0.32 0.29 0.27 0.25 0.23 0.22 0.9 7 0.9 4 0.9 2 0.9 1 0.8 9 0.8 7 0.8 5 0.8 4 0.8 3 0.8 2 0.8 0 0.7 8 0.7 7 0.7 6 0.7 3 0.7 3 0.7 1 0.6 9 0.6 6 0.6 4 0.6 3 0.6 1 0.5 8 0.5 7 0.5 6 0.5 4 0. 53 0. 51 0.4 9 0.4 8 0.4 6 0.4 2 0.4 0 0.3 9 0.3 6 0.3 4 0.3 2 0.2 9 0.2 8 0. 25 -0.5 0.20 Coefficient on Volatility Coefficient on Volatility 0.9 9 1.0 1 0.8 6 0.8 9 0.9 1 0.9 4 0.9 6 0.8 3 0.8 4 0.7 4 0.7 6 0.7 8 0.7 9 0.8 1 0.7 0 0.7 3 0.5 7 0.5 9 0.6 1 0.6 3 0.6 4 0.6 6 0.6 8 0.5 4 0.5 6 0.3 9 0.4 2 0.4 4 0.4 7 0.4 9 0.5 1 0.5 3 0.3 5 0.3 7 0.3 2 0.3 3 0.2 5 0.2 7 0.2 9 0.2 0 0.2 2 0.2 3 Coefficient on Volatility Figure 4: Coefficient on Volatility in Cross-Sectional Estimations, Increasing Labour Productivity Growth (UNIDO) 3 2 1 0 -1 -2 -3 Median Productivity Growth Figure 5: Coefficient on Volatility in Within Estimations, Increasing Labour Productivity Growth (UNIDO) 2.5 3 1.5 2 1 0.5 0 -1 Median Productivity Growth Figure 6: Coefficient on Volatility in Arellano-Bond Estimations, Increasing Productivity Growth (UNIDO) 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0 -1 -2 -3 Median Productivity Growth 1.06 1.04 1.03 1.00 0.98 0.96 0.93 0.92 0.91 0.88 0.86 0.85 0.84 0.83 0.83 0.81 0.78 0.77 0.76 0.75 0.73 0.73 0.71 0.70 0.68 0.66 0.64 0.63 0.63 0.61 0.58 0.56 0.54 0.54 0.52 0.51 0.49 0.47 0.43 0.41 0.39 0.36 0.34 0.32 Coefficient on Volatility 0.8 5 0.8 6 0.8 7 0.8 8 0.8 9 0.9 1 0.9 1 0.9 2 0. 93 0.9 4 0.8 2 0.8 3 0. 83 0. 84 0.8 5 0.7 6 0.7 7 0.7 7 0. 78 0. 80 0.8 1 0.7 1 0.7 3 0.7 3 0. 73 0.7 5 0.6 6 0.6 7 0.6 9 0. 70 0.7 1 0.5 3 0.5 5 0.5 6 0.5 7 0.5 9 0.6 2 0.6 3 0. 63 0.6 4 Coefficient on Volatility 0.9 6 0.9 2 0.9 3 0.8 8 0. 89 0.9 1 0.8 5 0.8 6 0. 84 0.8 5 0.8 1 0. 83 0.8 3 0.7 8 0.8 0 0.7 5 0. 76 0.7 7 0.7 3 0.7 4 0. 71 0.7 2 0.6 7 0. 69 0.7 0 0.6 4 0.6 6 0.6 2 0. 63 0.6 3 0. 58 0.6 1 0.5 4 0. 55 0.5 6 0.5 2 0.5 3 0.4 9 0.5 1 0.4 3 0. 46 0.4 8 Coefficient on Volatility Figure 7: Coefficient on Volatility in Cross-Sectional Estimations, Increasing Labour Productivity Growth (Reduced UNIDO) 8 7 6 5 4 3 2 1 0 Median Productivity Growth Figure 8: Coefficient on Volatility in Within Estimations, Increasing Labour Productivity Growth (Reduced UNIDO) 7 6 5 4 3 2 1 0 Median Productivity Growth Figure 9: Coefficient on Volatility in Arellano-Bond Estimations, Increasing Labour Productivity Growth (Reduced UNIDO) 6 5 4 3 2 1 0 -0.5 -1.5 Median TFP Growth 0.5 0 0.4 7 0.4 6 0.4 4 0.4 3 0.4 2 0.4 0 0.3 9 0.3 8 0.3 4 0.3 3 0.3 2 0.2 9 0.2 8 0.2 7 0.2 6 0.2 5 0.2 2 0.2 2 0.2 0 0.1 7 0.1 4 0.1 2 0.1 2 0.1 1 0.0 8 0.0 7 0.0 5 0.0 4 0.0 2 0.0 0 -0.0 2 -0.0 3 -0. 05 Coefficient on Volatility 0.5 1 0.5 0 0.4 6 0.4 6 0.4 4 0.4 3 0.4 1 0.4 0 0.3 9 0.3 6 0.3 3 0.3 4 0.3 1 0.2 9 0.2 8 0.2 7 0.2 6 0.2 4 0.2 2 0.2 2 0.1 8 0.1 7 0.1 4 0.1 2 0.1 1 0.0 9 0.0 6 0.0 8 0.0 5 0.0 3 0.0 1 -0. 01 -0. 02 -0. 05 -0. 04 Coefficient on Volatility 0.4 3 0.4 4 0.4 6 0.3 9 0.3 9 0.4 0 0. 41 0.4 2 0.3 3 0.3 4 0.3 4 0. 35 0.3 8 0.2 8 0.2 8 0.2 8 0. 29 0.3 2 0.2 4 0.2 6 0. 26 0.2 6 0.2 2 0.2 2 0.1 8 0. 19 0.2 2 0.1 4 0.1 7 0.0 9 0.1 1 0.1 1 0. 12 0.1 3 0.0 4 0.0 5 0.0 6 0. 08 0.0 8 0. 01 0.0 3 Coefficient on Volatility Figure 10: Coefficient on Volatility in Cross-Sectional Estimations, Increasing TFP Growth (ISDB) 20 10 0 -10 -20 -30 -40 -50 Median TFP Growth Figure 11: Coefficient on Volatility in Within Estimations, Increasing TFP Growth (ISDB) 30 20 10 0 -10 -20 -30 Median TFP Growth Figure 12: Coefficient on Volatility in Arellano-Bond Estimations, Increasing TFP Growth (ISDB) 7.5 6.5 5.5 4.5 3.5 2.5 1.5 0.5 Coefficient on Volatility -1 -2 -3 -4 -5 -6 Median Productivity Growth 0.83 0.82 0.78 0.77 0.76 0.75 0.74 0.73 0.70 0.70 0.67 0.67 0.65 0.64 0.61 0.60 0.58 0.57 0.56 0.54 0.52 0.52 0.51 0.49 0.47 0.45 0.44 0.43 0.42 0.38 0.37 0.35 0.34 0.34 0.31 0.31 0.28 0.27 0.25 0.23 0.22 0.21 0.18 0.7 5 0.7 6 0.7 7 0. 73 0.7 5 0. 70 0.7 1 0. 67 0.6 8 0. 65 0.6 6 0.6 1 0.6 2 0.5 6 0.5 7 0.5 8 0. 53 0.5 4 0. 51 0.5 2 0.4 8 0.5 0 0.4 3 0.4 4 0.4 6 0. 40 0.4 2 0. 36 0.3 8 0. 34 0.3 5 0.3 1 0.3 2 0.2 6 0.2 7 0.3 0 0.2 2 0.2 5 0. 20 0.2 1 -5 0.16 0.14 0. 15 0.1 7 Coefficient on Volatility 0.8 3 0.8 5 0.6 7 0.6 9 0.7 0 0.7 3 0.7 5 0.7 6 0.7 7 0.7 9 0.5 8 0.6 1 0.6 4 0.6 6 0.5 2 0.5 3 0.5 5 0.5 7 0.3 4 0.3 5 0.3 6 0.3 8 0.4 2 0.4 4 0.4 5 0.4 8 0.5 0 0.0 9 0.1 1 0.1 3 0.1 5 0.1 7 0.2 1 0.2 2 0.2 5 0.2 7 0.3 0 0.3 1 Coefficient on Volatility Figure 13: Coefficient on Volatility in Cross-Sectional Estimations, Increasing Labour Productivity Growth (ISDB) 20 10 0 -10 -20 -30 -40 -50 -60 Median Productivity Growth Figure 14: Coefficient on Volatility in Within Estimations, Increasing Labour Productivity Growth (ISDB) 20 15 10 5 0 -10 -15 -20 -25 Median Productivity Growth Figure 15: Coefficient on Volatility in Arellano-Bond Estimations, Increasing Labour Productivity Growth (ISDB) 4 3 2 1 0
© Copyright 2026 Paperzz