__________CAPS )SS POLICY RESEARCH WORKING PAPER 1333 A Test of the International Tis model for estiniang an- economy's rate of Convergence Hypothesis convergence to its ow. steady btate uses a Using Panel Data neodasscAf Solow model and accounts for the presence of Norman V. Loayza cou c e.The esdtmated rMe of cunvergenre is 0.0494,. implng a halFHife of-about 14-years. The World Bank Policy Research Department Macroecoomics and Growth Division August 1994 I POLICY RESEARCH WORKING PAPER 1333 Summary findings Loayza, using a neoclassical Solow model, estimates an 0.0494, which implies a half-life of about 14 years. economy's rate of convergence to its own steady state. This es&mated rate of convergence is about two and a Using panel data for a sample of 98 countries, he applies half times highcr than thosc obtained by Barro and Sala- Chamberlain's (1984) estimation procedure to account i-Martin (1992) and Makiw, Romer, and Weil (1992). for the presence of country-specific effects resulting from Loayza claims that those estimates are biased toward zero idiosyncratic unobservable factors. This procedure also because they fail to account for country-specific effects. prevents the estimation bias due to measurement error in Finally, he estimates the capital share in production to GDP. be 0.347, which is very close to the accepted benchmark Controlling additionally, for the country's level of value. education, he estimates the rate of convergence to be This paper - a product of the Macroeconomics and Growth Division, Policy Research Department - is pam of a larger effort in the department to understandthc determinants of economic growth. Copies of the paperare available frec from the World Bank, 18 18 H StrectNW, Washington, DC 20433. Please contact Rcbecca Martin, room NI 1-043, cxtension 39026 (30 pages). August 1994- The Policy Resarch Workig Paper Seris disnmmates the finings of jork in progess to curm age the exchan of ideas about developmen issueAn objective of thescries is toget thefindins oat quiy, even if thepreseations arc Iess thanfly polished. be papers cany the nams of the auors and shoald be used and cited accordingl. The fmdings, iterpreetions, and condusions are the auor? own and sbould not be attributed to the World Bank. its Ezeciue Board of Directors, or any of its member coun ies Produced by the Policy Research Dissemination Center A TEST OF THE INTERNATIONAL CONVERGENCE HYPOTHESIS USING PANEL DATA Norman Loayza The World Bank . I. INTRODUCTION G. Mankiw, D. Romer, and D. Weil in their paper "A Contribution to the Empirics of Economic Growth" (1992) argue that the Solow neoclassical growth model, when augmented to include human capital, provides a very satisfactory guide to understanding the process of economic growth among nations. In fact, they report ftat 80% of the international variation in income per capita can be explained by the augmented Solow model. Manliw, et. al., provide convincing arguments that the empirical evidence is consistent with the predictions of the model in terms of the effects of investment, both in physical and human capital, and population growth on the level of output. They also point out that, when properly specified, the model predicts "conditional" convergence. This phenomenon has received much attention and is well documented in papers such as Barro and Sala-i-Martin (1992), De Long (1988), Dowrick and Nguyen (1988), and Easterlin (1981). We wish to add to this liteature; in fact, the main focus of this paper is the econometric study of the "conditional" convergence hypothesis. From the perspective of economic methodology, there is still one more reason why the Solow model is appropriate for the international study of growth. The Solow model is a "positive" theory of growth in the sense that its explanation uses variables that we can observe and measure, although not without considerable difficulty, directly from the real world. The model tkes as its primary variables the investment rate, the population growth rate, and the rate of technological change. In studies of the "conditional" convergence hypothesis, this "positive" feature of the Solow model is crucial because it informs us as to the variables that appropriately condition for the steady-state of each economy. -2- This paper uses the classical Solow model as a general guide. The Solow model considers an efficiency parameter in the aggregate production function. Most cross-sectional studies of growth and convergence (including Mankiw, et. al.) identify the efficiency parameter with the constant in their regression equations. In so doing, these studies assume that aU countries have the same level of efficiency in using the factors of production. If we consider that the efficiency parameter depends on elements such as fiscal and taxation policies, openness to trade, public infrastructure, political stability, and level of education, we cannot but reject the assumption that all countries have a common efficiency parameter. In this paper, we deal explicitly with the issue of different efficiency parameters by identifying them with county-specific factors, which can be accounted for by using panel data. In principle, we would like to obtain information about all those variables that constitute the efficiency parameter. Unfortunately, adequate information is unavailable for most of those variables. Chamberlain (1984) shows a method to avoid the omitted-variable bias that occurs when all or some of the elements of such country-specific factors are unavailable. We will use Chamberlain's proposed methodology. In the estimation section of the paper, we consider first the case in which no information as to the country-specific factors is available, and second, the case in which we have information concerning one of its elements, namely, the country's level of education. -3- II. THE MODEL We use a general version of the neclassical growth model. There are different treatments of this class of models in the literature. Barro and Sala-i-Martin (1992) take a utility maximization approach; thus, the resulting levels and growth rates, both in the transition period and in the steady state, are functions of the underlying parameters of the representative consumer's utility function and technology. Mankiw, et.al, (1990) argue that the Solow (1956) model, in which the savings and population growth rates are taken as fixed and exogenous, is a good guide to the study of the differences in growth performance across countries. In this approach, the resulting levels and growth rates are functions of the country's technology and "observable" variables such as the investment ratios in physical and human capital and the population growth rate. A common feature of all versions of the neoclassical growth model is that economies tend to converge towards their own long-run growth rates. As Rebelo (1991) points out, this is due to the assumption of decreasing returns in the set of reproducible factors in the production function. Furthermore, the above-cited versions of the neoclassical model predict "conditional convergence." In simple terms, conditional convergence means that if counties had the same preferences and technology, poor countries would grow faster than rich ones. The rapid recovery of most Western European countries after the destruction caused by World War II and their catch-up to the United States is a demonstration of what is meant by conditional convergence (Dowrick and Nguyen 1989). To summarize, convergence means that the growth rate of an economy is positively -4- related to the distance between its current level and its .ong-run goal. Mathematically, the concept of convergence can be expressed in the following equation: dlog = Y- (1) ___= P(logo-1ogd) dt where 9, and y' are the current and steady-state levels of output per effective worker (which adjusts for the trend of exogenous population growth and technological progress), respectively; and B is the convergence rate, which is a function of the underlying parameters of preferences and technology. Clearly, convergence toward the steady state is achieved if 8>0. Equadon (I) is the result of the linearization of the transition path of output per effective worker around its steady-state value. We are assuming that the population of workers grows exponenfially at rate, say, n, and that the available technology also grows exponentially at rate, say, g. Both rates are exogenously determined. They will dictate the growth rates in the steady-state; thus, the level of output will grow in the long run at rate n+g, and the level of output per worker, at rate g. Clearly, the level of output per effective worker will be constant in the long run (dtis is why we need to formulate the equation of convergence in terms of quantities per effective worker). Integrating equation (1) from (t-r) to t and expressing output in per-worker terms, we get logy, = (1 -e -')logy + e 'Iogy, + (1 -e -')gt + e -lgr + (1 -e -rP)IogA (2) -5- where y is output per worker and A0 represents the shifting parameter in the neoclassical production function. We assume that, conditional on 5', n3 is constant for all countries. Following Mankiw, Romer, and Weil (1992), we assume that the rate of technological progress, g, is the same for every country. As Barro and Sala-i-Martin (1992) point out, in cross-country empirical studies of convergence, it is crucial to hold fixed the steady-state levels of output per effective worke, y How can we do this? As it was said above, if we follow the utility-maximization approach, the steady-state levels are functions of the underlying parameters of the representative consumer's utility function and technology. It is almost impossible to directly obtain estimates of those parameters for most countries around the world. However, we can be informed as to the values of those parameters across countries by observing the variables that are determined by such parameters. Some of those variables are the investment ratios, the population growth rates, and the relive shares of the factors of production. The Solow model, using a Cobb-Douglas production function, gives a closed-form solution for y: log1r = --log(n +g+6)+-!alogs (3) 1-a 1E where s is the investment rate of the economy, 6 is the depreciation rate of the capital stock, and a is the capital share of output. Substituring equation (3) into (2), we obtain an expression for the evolution of output per worker in the transition path in terms of observable variables. Niformalizing r= 1, we get -6- logy, = -(I-c-P)(j 5)log(n+g+6) +(l-e )(Ije)logs + e-Plogy,_, + (I-e -P)gt +e -Pg + (I -e P)logAO In order to see more clearly the effect of each variable to the rate of growth in the transition to the steady state, we can rewrite equation (4) as follows: Iogy,-logytl = -(1-C e1)( Of)logtn+g e8)+(l-e1P)(1 )logs 1-4 )( 'M )og ~~~~~~~~(5) -(1 -etP)logy, + (1-e-P)gt + eg + (l-e-P)logA0 We will use this specification as a guideline but will not apply it literally. If the rate of convergence A is positive, we can predict the signs of the coefficients of each term in equation (5). Let us examine each of those terms in turn. The fir;; one indicates that, for given g and 6, the rate of growth of the worldng-age population, n, is negatively related to the growth of per capita output. The second term indicates that the more a country saves and invests, the more it grows. The third term tells us that countries grow faster if they are poor with respect to their potential. The fourth term suggests the presence of a time-specific effect in the growth equation. The fifth term tells us that increases in the rate of technological change bring about higher per capita growth rates. In the sixth term, the parameter A0 represents all those elements that -determine the efficiency of factors of production and available technology to create wealth; of course, the greater such efficiency, -7- the greater the rate of growth of the economy. Some of the elements that compose the AO parameter are govemment policies, natural resources, openness to foreign trade, and quality of education of the population. This term suggests the presence of a country-specific effect, which may well be correlated to the investment and population growth rates, as well as to the initial level of output per worker in each particular economy. The above interpretation of equation (5) suggests a natural regression to study the convergence hypothesis. Let us zewrite a more general form of equation (4) for a given county i: logy1. = O.log(nif+g+8) +631ogs4,, +(+y)logy;++ Ele,J | og(n11+,+b6),logs,,) .... og(n.,+g+8), logsI7)] - 0 for t = where E. and p1 represent the time-specific and the country-specific effects, respectively; and o., O., and -y are parameters to be estimated. The disturbance term ej., is assumed to be uncorrelated with aUl leads and lags of the independent regressors log(,t+g+±) and log(s.J); in particular, this implies that such regressors are not affected by the evolution of output, just as the Solow model assumes. Note that the disturbance ef., is not assumed to be i.i.d.. Thus, the model does not impose either conditional homoskedasticity or independence over time on the disturbances within each country. We want to allow for serial correlation in the error term because there may be some excluded variables that present short-run persistence; of course, the p1 component accounts for long-run persistence of excluded variables that may be correlated with the independent regressors log(,,±+g+5) and log(s;,). -8- Let us summarize what our working assumptions are. First, we assume that a log- linear specification for the regression equation is appropriate. This specification is quite popular in the growth literature both because it comes naturally from a Cobb-Douglas type production function and because it has proven to be relatively robust (Maddison, 1987). Second, we assume that conditional on the steady-state level of output per worker, 9r, the rate of convergence, B, is approximately equal across countries. Third, we assume that the working-age population growth rate and the ratio of investment in physical capital condition appropriately for 5'. A related assumption says that g, a, and 5 are approximately the same for all countries. Finally, we assume that the working-age population growth rate and the physical capital investnent ratio are strictly exogenous. The hypothesis of conditional convergence can be tested using regression equation (6). In fact, conditional convergence implies that the coefficient on log(yi*fr), (1+y), is less than 1. As it was said in the introduction, previous studies of convergence have used cross- sectional data. This forced the use of some rather restrictive assumptions in the econometric specification of the models. For instance, Mankiw, et.al, assume that logAo is independent of the investment ratio, the working- age population growth rate, and the initial level of output per worker. This amounts to ignoring country-specific effects; for example, their assumption implies that government policies regarding taxation and international trade do not affect national investment, or that the endowment of natural resources does not influence fertility. As Manliw, et. al. say, If countries have permanent differences in their production funcdons -that is, different Ao's- then these Ao's would enter as part of the error -9- term and would be positively correlatcd with initial income. Hence, variation in AO would bias the coefficient on initial income toward zero (and potentially would influence the other coefficients as well)' (p.424). Furthermore, since only one cross-section is considered, the time-specific effect becomes irrelevant. Fortunately, panel data for most variables of interest is available. We intend to use the additional information contained in panel data to analyze regression equation (6). -10- M. PANEL DATA ESTIMATION Let us rewrite equation (6) as follows: tsj = O'x11+(l+y)z111 + ,+ (t1 (7) where z1,t = log(y1,); xu., = (log(nj1+g+6), log(sj.))'; and 0 = (0, OX. We assume that the independent rcgressors, x, are well measured in the data. However, we allow for the possibility of errors in variables regarding the dependent variable, z. Observed output may not correspond to the model's output variable for two reasons. First, output may be poorly measured. Second, and most importantly, observed output has a business cycle and a growth (or trend) component. Since our worldng model explains only the latter, there is a potential estimation bias. Errors in the dependent variable are a potential source of bias because lagged output is one of the regressors. Let us consider the following estimation strategy. To account for the time effects we process the data by removing the time means from each variable. Then, we can ignore the Ct's and the regression can be fit without a constant (MaCurdy 1982). Least-squares estimation ignoring the country-specific effects and the errors-in- variables problem produces biased estimators. In particular the estimate of (I + 'y) is biased in an unknown direction: the measurement error biases the estimate downwards, and dth country-specific effect tends to bias it upwards. Using the "within" estimator (or any other panel-dam estimator based on time- differencing) to correct for the country-specific-effects bias is inappropriate. The specific- effects bias disappears, but the measurement-error downward bias tends to worsen; this is due to the reduction in "signal" variance brought about by time-differencing. Furthermore, given the presence of a lagged dependent variable, time-difference estimators by construction create an additional downward bias. Therefore, in general the "within" and other time- difference methods underestimate (1 +y). We will use the Il-matrix estimation procedure outlined in Chamberlain (1984). This procedure allows us to correct for both measurement-error and specific-effects biases. Chamberlain's 11-matrix estimation procedure consists of writing both the lag dependent variable and the country-specific effect in terms of the independent regressors, thus obtaining reduced-form regressions from which to obtain the coefficient estimates of interest. More specifically, the fl-matrix procedure consists of two steps: First, we estimate the parameters of the reduced-form regressions of the endogenous variable in each period in terms of'the exogenous variables in all periods; thus, we estimate a multivariate regression system with as many regressions as periods for the endogenous variables are available. Since, we allow. for group-wise heteroskedasticity and correlation between the errors of all regressions, we use the seemingly unrelated regression (SUR) estimator. As result of this first step, we obtain estimates of the parameters of the reduced-form regressions (these are the elements of the II matrix) and the robust (White's heteroskedasticity-consistent) variance-cuvariance matrix of sucin parmeters. Our working model implies some restrictions on the elements of the II matrix; or in other words, the parameters we are interested in are functions of the elements of the H matrix. Then, in the secor,d step of the procedure, we estimate the parameters of interest by -12- means of a minimum distance estimator, using the estimated robust variance-covariance of the estimated II as the weight matrix: Min(Vecdll A) (VecIll-)) where b is the set of parameters of interest, and n is the robust estimated variance- covariance of the 11 matrix. Chamberlain (1982) shows that this procedure obtains asymptotically efficient estimates. In order to use this method, we need to make explicit the restrictions that our model imposes on the fI matrix. After removing the time means, our basic model in equation (6) can be written as zj, = ejxi,+ a +ykt, L,-l+eu (8) Afe4Ix Jx.-.,.7=O] for t= 1,...,T By recursive substitution of the z, term in each regression equation, we have -13- (l+y)641 = + ( R + Zia = (+.Y)(,X,,1 + e'xi +(1 +Y)240 + 1 +(1+Y)]Pi + ,2 Z3= (i+y)2ex., +(l+y)eOX. +*Ox3, +(1+y)3Iz -{l+(l y) i(1+y)2] 1p I + 3 Z,T = (1 +... + (1 +Y)T+ ((a E ((j)i/I Xi...XIT)== (t= 1...,T and i=1,...,N) Chamberlain (1984) proposes. to deal with the correlated country-specific effect W and the initial condition (7c) by replacing them by their respective linear predictors (given in terms of the exogenous variables) and error terms, which by construction are uncorrelated with the exogenous variables. The linear predictors are given by E(z40 ( I |X(.2 ... x.T) = ;LxX1 + A42 +... + A E %LilX4 s aT Xi. I .ij2 + TA As Chamberlain points out, assuming that the variances are finite and that the distribution of (xj,,..., x.T, &). does not deped on i, using the linear predictors does not impose any additional restrictions. Now we are ready to write the II matrix implied by our working model. As we will -14- see in the next section, our panel data consists of 5 cross sections for the exogenous variables x and 6 cross sections for the variable z; the additional cross section for z is given by the initial condition ZD. Thus, the multivariate regression implied by our model is Z=IA z41. x. Zi.3 XO ~~~~~~~~~~~~(9) Z1.4~~~~~. II = [B + Cl' + where, o 0 0 0 0 0' ~~~0 0 0 0 (1~~y)8' 0') 0 0 0 (1I.-y)20' (i +flO' 0' 0 0 (1 ÷y)3e' (1÷7)20/ (l +y)0' 0' 0 (1 +yY4'0' (l+y)3 O' (1÷+y)20f Cl(IO 0'e 9 -15- (1 +Y) ,(1 +TY (rr = ] +1+Y 2 3y 0 1+(l+y)+(1+y2+(l+y9 1+(1+y) (l+( )2+(1l+y)+(1+y)4 As we said in the introduction, we would also like to consider the case in which we have some information as to one of the elements of the country-specific factors, namely, the country's level of education. In this case, we rewrite equation (8) as follows, Zjt Uxi, + (1 +Y)Zj, XC + Vi + Co (10) Nlej} lxj,l,.,x1 , =0 for t = l,.-,T where, e, is a proxy for the country's level of education (which is assumed to be constant through time), 0, is a constant coefficient, and v; is the new country-specific factor. By definition (Li = OAe + v'. In tis case, the associated multivariate regression is very similar to the one where no -16- information as to the country-specific effects is available. Worldng with recursive substitution and the appropriate linear predictors, as we did in the previous case, the multivariate regression associated with regression equation (10) is the following, 7-to .t Zt3 . 'A (11) z~~~~~X. H = [B + +*- where, 0 0 0 0 00 o0 0 0 0 0 0 (1+y)0' 0' 0 0 0 0 (1+y)20f' (0+y)0 0' 0 0 0 (1+y)30e (i+y)2e, (i +y)e' e' 0 0 (1+y)4Y0 (1+ y)3e (1+y)20' (I '-y)e0 0' 0 -'7- (}+y) = (i:y) [L 2 3 Al4 5 +y)S 0 4T = 1+(I+y)+(1+y)2 [T I2 3 T4 C + (c+)] 1 +(1 +y)+(l ay92i+(I +y)3 I+(l+y)+(1 +y)2+(l +y)3 +( +y)4 From the implied restrictions on the II-matrix (in particular those related to the coefficients on ej, note that we cannot separate r0 from °e: only (@c + O) is identified. Therefore, even though the level of education help condition for the country-specific factor, its precise effect on growth is not identified without further restrictions. -18- IV. DATA AND RESULTS The data source for all our variables, but the proxy for the level of education, is The Penn World Table (Mark 5), constructed by R. Summers and A. Heston (1991). This table provides annual information for a number of national accounts variables from around 1950 to 1988. However, data for most countries is available only for a shorter period of time, namely, 1960 to 1985. We work with regular non-overlapping intervals of five years each. Thus, our five cross sections correspond to the years 1965, 1970, 1975, 1980, and 1985. Let us explain each of the variables of the model in tum. The dependent variable is the natural logarithm of real GDP per worker, that is, log(y;65), .-., log(y.5)- When no information as to the country-specific factors is available, the regression equation has three explanatory varables (equation (8)). The first one is the natural logarithm of the working-age population average growth rate plus (g+c); we follow Mankiw, et. al. (1992) in assuming that (g+6)=0.05. The avenage of the working-age population growth rate is taken over the previous five-year interval; then we also have five observations of this variable far each country, that is, log(n.65+0.05), ..., log(ni,85+0.05). The second explanatory variable is the natural logarithm of the average ratio of real investment (including government investment) to real GDP. These averages are also taken over the previous five-year interval, so that we have five observations for each country, that is, l0g(s;65), ..-, 10g(Si.s)- The last explanatory variable is the natural logarithm of real GDP per worker lagged one period, that is, five years back; therefore, the observations for each country are, -19- 109(Yi.60), ...-, lOg(.80) We would also like to consider the case in which we have some information as to one of the elements of the country-specific effects, in particular, the country's level of education (equation (10)). The proxy we use for the level of education is taken from Mankiw, Romer, and Weil (1992). This is the percentage of the working-age population that is enrolled in secondary school, a measure that is approximated by the product of the gross secondary enrollment ratio times the fraction of the working age population that is of secondary school age (i.e., aged 15 to 19). Our sample consists of 98 countries (see Appendix B for a list of countries considered). These are the countnes for which data are available and for which oil production is not the primary economic activity. It is well known that standard growth models do not account for economies based on the extraction of natural resources and not on value-added activities. Excluding the countries for which data are not available may create sample selectivity problems given that these countries are frequently the poorest ones. Therefore, we will not presume that the results obtained here can be applied to those very poor economies. Equations (8) and (10) represent the cases we wish to study. Writing those two equations more explicitly, we have logy1, = 01og(nj'-O0.05;) + 0ogs, + (1 +y)logiy4 5 + ' ) (8) -20- logyi,, = O)og(n1+0O5) + elogs, +(l +y)logy4t,5 + 9"e + vi +e,, (10') Under the Solow growth model with a Cobb-Douglas aggregate production function (see equation (4)), we should expect the estimated values of -y to be negative, O., negative, and 0,, positive; furthermore, we expect O. and 05 to be approximately the same in absolute value. In tables 1-3, we provide test statistics for such hypotheses. Table I shows the estimated parameters of the simple Solow model in equation (8') using conventional procedures. The OLS and lst-differences estimators refer to least squares estimation in levels and 1st-differences, respectively. In order to use the information from the 5 available cross-sections, we ermiploy a system-regression procedure, considering parameter and covariance restrictions across the regrssions of the system.' As is well known, the Within estimator falls under the category of difference estimators. As explained in section DI, the OLS estimator ignores both errors-in-variables and country-specific effects, thus producing estimates for -y that are biased in an a priori unknown direction. The difference estimators control for country-specific effects but ignore the errors-in-variables problem and, by construction, create a correlation between the new error term and the differenced lagged dependent variable. Therefore, difference estimators produce downward-biased estimates for -y. Such downward bias is worse in the case of the lst-differences estimator than in the case of the Within estimator. Table 2 shows the estimated parameters of the simple Solow model in both equations 'Clearly, each regression in the system corresponds to one cross-section. -21- (10') and (8') (that is, with and without education as a regressor) using Chamberlain's II- matrix procedure. In each case we estimate both ignoring and accounting for country- specific effects. Given that tI;rough the fl-matrix procedure the endogenous variable, output, is not used as a regressor, its related errors-in-variables no longer produces estimation bias. Therefore, fl-matrix estimation assuming no country specific effects has no errors-in- variables bias but presents country-specific effects bias, which, as explained in section III, is an upward bias. Clearly, this bias is worse when the proxy for education, as an element of the country-specific effects, is not used as a regressor than when it is. fl-matrix estimation accounting for country-specific effects produces unbiased esdmates; when, additionally, the proxy for education is used as a regressor, the parameter estimates gain efficiency. In the context of the fl-matrix method, it is possible to test whether country-specific effects are important, in the sense that they are correlated to the independent regressors. Note that the absence of country-specific effects implies that the coefficients in the linear predictor of ; are all equal to zero, that is, H(: 1 .. = T ' = 0. As we can see in Table 2, the appiopriate Wald test for this hypothesis strongly rejects it. Controlling for country- specific effects is in fact quite important. From Tables 1 and 2, we learn that the estimates for -y obtained using various procedures agree with our predictions, in terms of how such estimates are related to consistent estimates. The lst-differences estimate for -y (-0.9786) is the most negative, followed by the Within estimate (-0.3457). Then we have the estimates using the H1-matrix accounting for country-specific effects (-0.2187 and -0.2686, with and without using the -22- education proxy, respectively). The OLS estimate (-0.0301) is next, showing that in this case the country-specific bias is stronger than the errors-in-variables bias. The highest (and only positive) estimate for -y is obtained using the fl-matrix procedure assuming no country- specific effects (0.0078), estimator which isolates the upward bias due to country-specific factors. Our consistent estimates for -y imply the following values for ,B, the speed of convergence: .0626 (not using the education variable) and .0494 (using it). These values are about two and a half and three times as high as those obtained in previous empirical papers (see in particular Barro and Sta-i-Martin (1992), and Mankiw, Romer, and Weil (1992)). A figure commonly provided in studies of convergence is the "half life," which is the time it tkes for an economy to move halfway to its own steady state. From equation (1), we find that the half life, T, can be calculated from an estimate of ,B as follows T = log2 Therefore, the fl-matrix method controlling for country-specific effects and using the education proxy as a regressor predicts a half life of about 14 years, while previous studies, suffering from errors-in-variables and specific effects biases, predict one of about 34.7 years. This could be interpreted as "good news' for poor countries. However, such interpretation would be inappropriate since the convergence occurs with respect to the country's own steady state level. As we will say later on, a higher rate of convergence is related to a low share of physical capital in the production function, which implies that decreasing returns set -23- in more quickly. Comparing the II-matrix consistent estimates for the coefficients on labor force growth and investment ratio (0, and 0,, respectively) with their OLS counterparts, we see that the consistent estimates are stronger (i.e., higher in absolute value) and more efficient (i.e., with a lower standard error). Comparing the two consistent estimators, we see that when the education proxy is used as a regressor, the effect of labor force growth and investment on output growth is somehow weaker. In Tables 1 and 2 we report a Wald test for the hypothesis that 0, and 0, have the same absolute value and opposite signs. From equation (4), we realize that if the restriction that 06 = -0° is imposed in the model, it is possible to retrieve an implied estimate for a, the capital share in the Cobb-Douglas production function. We impose such restriction and report the constrained estimation results in Table 3. Not surprisingly our OLS estimates for a and a are very close to those obtained bv Mankiw, et. al. (their estimates are 13 = 0.00606 and a = 0.7 to 0.8) when they use the simple Solow model. The II-matrix estimates ignoring country-specific effects but controlling for education are similar to those obtained by Mankiw, et. al. when they use their "human-capital augmented" Solow model (their estimates are 13 = 0.0137 and a = 0.48). Properly accounting for countiy-specific effects, we obtain estimates for a that are much closer to the accepted benchmark value?: 0.335 (not using the education proxy) and 0.347 (using it). Interestingly, Mankiw, et. al. argue that the simple Solow model performs well in their cross-sectional study except for the fact that their estimated ot is much bigger than the accepted benchmark value. 2Maddison (1987) estimates the share of non-human capital in production to be about 0.35. -24- V. CONCLUSION This study estimates the rate of convergence of an economy to its own steady state. Using panel data for a sample of 98 countries, we use Chamberlain's (1984) estimation procedure is applied to account for the presence of country-specific effects, which result from idiosyncratic unobservable factors. Furthermore, this procedure avoids the estimation bias due to measurement error in GDP. Controlling, additionally, for the country's level of education, we estimate the rate of convergence, f, to be 0.0494; which implies a half life of about 14 years. Also, we estimate the capital shaie in production, a, to be 0.347. We believe that our estimated rate of convergence (which is higher than in other studies) provides evidence in favor of the neoclassical Solow model, in which only physical capital can be accumulate. The Solow model predicts a rapid rate of convergence because it considers a production function with strong decreasing returns to capital, the factor that can be accumulated. In the simple Cobb-Douglas specification, such strong decreasing returns are produced by a low capital share (a, in our case). In fact, the Solow model with a Cobb- Douglas production function gives a closed-form solution for the rate of convergence, 0 = (n+g+-)(1l-a) Assuming that g+6 = 0.05, and using the average worldng-population growth rate for our sample, n = 0.022, we find that a value of 0.347 for ca implies a rate of convergence 3 of 0.047, which is very similar to our econometrically estimated rate of convergence. -25- REERECES, Barno, R. (1991), "Economic Growth in a Cross Secdon of Countries," The Quarterly Jounal of Economzics, May. Barro, R. and X. Sala-i-Martin (1993), 'The Neoclassical Growth Model,' Chapter 1 of unpublished manuscript. Barro, R. and X. Sala-i-Martin (1992), "Convergence," Journal of Political Economy, April. Barro, R. and H. Wolf (1989), "Data Appendix for Economic Growth in a Cross Section of Countries," unpublished, NBER, November. Becker, G., Murphy, K., and Tamura (1990), "Human Capital, Fertility, and Economic Growth," Journal of Political Economy, October. Chamberiain, G. (1982), "Multivariate Regression Models for Panel Data," Journal of Econometrics, 18, 5-46. Chamberlain, G. (1984), "Panel Data," Handbook of Econometrics, Vol. 2, edited by Z. Griliches and M. D. hntiligator. DeLong, J. B. (1988), "Productivity Growth, Convergence, and Welfare: Comment," Amercan Economic Reiew, 78, 1138-54. Dowrick, S. and D. Nguyen (1989), "OECD Comparative Economic Growth 1950-85: Catch-Up and Convergence," American Economic Revew, 79, December. Easterlin, R. (1981), "Why isn't the Whole World Developed," Journal of Economic iTstory, 41, 1-20. Greene, W. (1990), Econometric Analysis, Macmillan Publishing Company, New York. Gdiliches, Z. (1974), "Errors in Variables and Other Unobservables," Econometrica, 42, 971-998. MaCurdy. T. E. (1982), "The Use of Time Series Processes to Model the Error Structure of Eamings in a Longitudinal Data Analysis," Jourmal of Econometrics, 18, 83-114. Maddison, A. (1987), "Growth and Slowdown in Advanced Capitalist Economies: Techniques of Quantitative Assessment, " Joumal of Economic Literature, 25, June. -26- Mankiw, N. G., D. Romer, and D. N. Weil (1992), "A Contribution to the Empirics of Economic Growth," The Quarterly Journal of Economics, May. Mundlak, Y. (1978), "On the Pooling of Time Series and Cross Section Data," E:conometrica, 46, no. 1. Ramsey, F.P. (1928), "A Mathematical Theory of Saving," Economic Journal, 38, December. Rebelo, S. (1991), 'Long-Run Policy Analysis and Long-Run Growth," Journal of Political Economy, vol. 99, no.3. Romer, P. (1989), "Human Capital and Growth: Theory and Evidence,' mimeo, The University of Chicago, April. Sala-i-Martin, X. (1990), "Lecture Notes on Economic Growth (I)," NBER w.p. no. 3563, December. Summers, R. and A. Heston (1991), "The Penn Worid Table (Mark 5): An Ezxpanded Set of international Comparisons, 1950-1988," 7he Quarterly Journal of Economics, May. Solow, R. (1956), "A Contribution to the Theory of Economic Growth," The Quarterly Jounal of Economics, February. White, H. (1980), "A Heteroskedasticity-Consistent Covariance Matrix Estimator and a Direct Test for Heteroskedasticity," Economzetrica, 50, 483-499. -27- Table 1: Estimation of the Simple Solow Model Using Conventional Procedures | Parameters OLS Ist-Differences Widfiinl 7 -.0301 -.9786 -.3457 - ~~~(.0109) (.0741) (.0378) °n -.0756 -.0600 -.1140 (.0513) (.0364) (.0542) 9, .1055 .2184 .1745 (.0179) (.0363) (.0241) Implied .0061 .7689 .0848 (.0022) (.6925) (.0116) Wald Test for .2822 8.1713 2.1216 p-value .5953 .0043 .0488 -28- Table 2: Estimation of the Simple Solow Model Using Chamberlain's II-Matrix Procedure Parameters No Specific Specfic No Specific Specific Effects, Effects Effects Effects, Controlling for Controlling for Education Education -y .0078 -.2686 -.0670 -.2187 (.0051) (.0456) (.0077) (.0474) 0. -.0195 -.1220 -.0892 -.0948 (.0192) (.0250) (.0224) (.0244) as .0585 .1489 .0878 .1305 (.0091) (.0178) (.0077) (.0154) Implied -.0016 .0626 .0139 .0494 (.0010) (.0125) (.0016) (.0121) Wald Test for 2.4784 .7975 .0028 1.2980 0, a = -e p-ralue .1154 .3718 .9579 .2546 Wald Test for - 164.5277 - 134.1383 No Specific Effects p-value .0000 .0000 -29- Table 3: Estimation of the Simple Solow Model Imposing the Cobb-Douglas Restriction: 9 = -0 = ', Paraneters OLS fl-Matrix II-Matrix fl-Matrix No Specfic Specific Effects Specific Effects, Effects, Controlling for Controlling for education Education 7 -.0311 -.0669 -.2782 -.2262 (.0113) (.0075) (.0437) (.0458) 0 .1028 .0881 .1401 .1202 (.0165) (.0064) (.0147) (.0119) Implied 0 .0063 .0138 .0652 .0513 (.0023) (.0016) (.0121) (.0118) Implied a .7679 .5684 .3350 .3470 (.0472) (.0236) (.0418) (.0572) Wald Test for - 170.5039 129.3814 No Specific Effects p-value _.00 .0, . APPENDIX. List of Countries in the Sample. Algeria India Trinidad and Tobago Angola Israel United States Benin Japan Argentina Bostwana Jordan Bolivia Burlina Faso Korea, Rep. of Brail Burundi Malaysia Chile Caneroon Npl Colombia Central Afr. Rep. Pakistan Ecuador Chad Philippines Paraguay Congo Singapore- Peru Egypt Sri Tanka Uruguay Ethiopia Syrian Arab Rep. Venezuela Ghaa Thailand Austrlia Ivory Coast Austria Indonesia Kenya Belgium New Zealand Liberia Denmark Papua New Guinea Madagascar Finland Malawi France Mali Germany, Fed. Rep. Mauritania Greece Mauritius Ieland Morocco Italy Mozambique Netelands Niger Norway Nigeria Portugal Rwanda Spain Senega Sweden Sierra Leone Switzerland Somalia Turkey South Africa United Kingdom Sudan Canada Tanzania Costa Rica Togo Dominican Rep. Tunisia El Salvador Uganda Guatemala Zaire Haiti Zambia Honduras Zimbabwe Jamaica Bangladesh Mexico Burma Nicaragua Hong Kong Panama Policy Research Working Paper Series Contact Title Author Date for paper WPS1319 The Financial System and Public Ashl Demirgul9-Kunt July 1994 B. Moore Enterprise Reforn: Concepis and Ross Levine 35261 Cases WPS1320 Capital Structures in Developing Asfl DemirgOg-Kunl July 1994 B. Moore Countries: Evidence from Ten Vojislav Maksirnovic 35261 Countries WPS1321 Institutions and the East Asian Jose Edgardo Campos July 1994 B. Moore Miracle: Asymmetric Information, Donald Lien 35261 Rent-Seeking, and the Deliberaton Council WPS1322 Reducing Regulatory Barriers to Barbara Richard July 1994 M. Dhokai Private-Sector Participation in Latin Thelma Triche 33970 America's Water and Saniation Services WPS1323 Energy Pricing and Air Pollution: Gunnar S. Eskeland July 1994 C. Jones Econometric Evidence from Emmanuel Jimenez 37699 Manufacturing in Chile and Indonesia Uli Liu WPS1324 Voucher Funds in Transitional Robert E. Anderson July 1994 F. Hatab Economies: The Czech and Slovak 35835 Experience WPS1325 The Economics of Research and AnwarShah July 1994 C.Jones Development How Research and 37699 Development Capital Affects Production and Markets and Is Affected by Tax Incentives WPS1326 Banks, Capital Markets, and Gerhard Pohl July 1994 L. Hovsepian Corporate Governance: Lessons Slijn Claessens 37297 from Russia for Eastem Europe WPS1327 Is the Debt Crisis History? Recent Michael Dooley July 1994 S. King-Watson Private Capital Inflows to Developing Eduardo Femandez-Arias 31047 Countries Kenneth Kletzer WPS1328 The Use of New York Cotton Futures Panos Varangis July 1994 D. Gustafson Contracts to Hedge Cotton Price Risk Elton Thigpen 33714 in Developing Countries Sudhakar Satyanarayan WPS1329 The Regulation and Supervision of David H. Scott August 1994 K. Waelti Domestic Financial Conglomerates 37655 WPS1330 Revenue Uncertainty and the Choice Delfin S. Go August 1994 C. Jones of Tax Instrument during the Transition 37699 in Eastem Europe Policy Research Working Paper Series Contact Title Author Date for paper WPS1331 The Myth of Monopoly: A New View Annette N. Brown August 1994 M. Berg of Industrial Structure in Russia Barry W. Ickes 36969 Randi Ryterman WPS1332 Poverty and Household Size Peter Lanjouw August-1994 P. Cook Martin Ravallion 33902 WPS1333 A Test of the Intemational Norman V. Loayza August 1994 R. Martin Convergence Hypothesis Using 39026 Panel Data WPS1334 Taxation, Public Services, and the Juan Braun August 1994 R. Marfin Informal Sector in a Model of Norman V. Loayza 39026 Endogenous Growth WPS1335 Labor Regulations and the Informal Normnan V. Loayza August 1994 R. Martin Economy 39026