Absorptive capability and economic growth: how do. countries catch-up?

Transcription

1 Absorptive capability and economic growth: how do countries catch-up? Mark Rogers Harris Manchester College Oxford University Telephone (01865) Fax (01865)

2 Abstract This paper empirically investigates the importance of technological catch-up in explaining productivity growth in a sample countries since the 1960s. New proxies for a country's absorptive capability based on data for students studying abroad, telecommunications and publications are used. Various regression models are estimated, with both GDP per worker growth and multifactor productivity growth as dependent variables. A specific focus is on influential observations and the robustness of results. The results indicate that absorptive capability is a factor in explaining growth, with the most robust finding that countries with relatively high numbers of students studying science or engineering abroad experience faster subsequent growth. However, the paper also indicates that the significance of coefficients varies across specifications and samples, suggesting caution in focusing on individual results. Key words: technological catch-up, absorptive capability, economic growth J.E.L. classification: F43, O40 2

3 1. Introduction Are countries that are better at absorbing technology from overseas are able to experience faster productivity growth? While many may answer yes, the issue of the strength of the link, and the determinants of absorptive capacity, are less clear. If countries can catch-up technologically, and there is a link between the level of technology and GDP per capita, we should observe convergence in GDP per capita levels for those countries. The issue of convergence has received considerable attention (see, for example, the collection of articles in the Economic Journal, Vol. 106, no. 437). However, there appears to be a lack of work on the underlying mechanisms of convergence and, in particular, the issue of technology. Bernard and Jones (1996) state: The empirical convergence literature envisages a world in which the presence or lack of convergence is a function of capital accumulation. This focus ignores a long tradition among economic historians and growth theorists which emphasises technology and the potential for technology transfer. We suggest that this neglect is an important oversight: simple models which incorporate technology transfer provide a richer framework for thinking about convergence. (Bernard and Jones, 1996, p.1037) There is empirical cross-country work based on the technological catch-up approach, dating from the work of Abramovitz (1986), Baumol (1986) and Dowrick (1989) (see Fagerberg, 1994, for a review). Recent developments have sought to test the idea that greater trade openness, more inward foreign direct investment and higher levels of human capital foster 3

4 technology flows and boost economic growth. 1 This paper adds to this literature by considering a set of new variables based on students studying abroad, telecommunications and publications data. The new variables aim to proxy various aspects of a country's ability to access and learn technology from overseas. By examining such data the paper aims to provide greater insight into the mechanisms of technological catch-up. The paper is structured as follows. In section 2 we consider some simple mathematics of technological catch-up. This analysis shows that the existence of a technology gap is a necessary but not sufficient condition for the process of technological catch-up. A country must also have sufficient absorptive capability defined as the capability to access, learn and absorb relevant overseas technology. The concepts of the technology gap and absorptive capability are therefore central to understanding the process of technology diffusion and economic growth. Section 3 defines absorptive capability more closely. Section 4 contains empirical results from cross-country regressions for a sample of up to 84 countries, to test the partial correlation of the new variables with GDP per worker growth over Use of this approach requires attention to two critical issues. First, the selection of the appropriate baseline regression specification. Second, ensuring that any results are robust to 1 For example, Hansson and Henrekson (1994) consider trade openness and human capital, Balasubramanyam et al (1996) consider foreign direct investment, while Barro and Sala-i-Martin (1995) consider human capital. There is also a related literature on testing for R&D spillovers between OECD countries (e.g. Coe and Helpman, 1995, Frantzen, 2000) 4

5 variations in the sample and outliers. Section 5 uses multifactor productivity growth ( ) as a dependent variable and section 6 concludes. 2. A simple model of technological catch-up Assume that an economy's level of technology and its absorptive capability can be expressed with scalars. A common equation to describe the growth of technology is da dt T A / A = φ (.) A, [1] where A represents the level of technology, T the world level of best practice technology and φ(.) is a function representing absorptive capability. This equation originates in work by Nelson and Phelps (1966), and has been used by Hansson and Henrekson (1994) in the context of technological catch-up. 2 Assuming that T grows at an exogenous constant rate g, the system of differential equations (i.e. equation [1] and dt/dt=gt) can be solved to show that, in the long run, growth of A must equal g. Intuitively, unless the growth rate of A equals the growth rate of T, the right hand side of [1] must be changing, which in turn means the growth rate of A must be changing. Figure 1 illustrates this result graphically. 2 Other models of catch-up include Gomulka (1971, 1990) and Verspagen (1991). 5

6 Figure 1 A monotonic technological gap model Figure 1 shows that a follower country which starts with a technology ratio (A/T) below the long run equilibrium level (φ/(φ+g)) will experience rapid growth (relative to g). Conversely, countries which start with a technology ratio above φ/(φ+g) will experience growth rates lower than g. The functional form of [1] is essentially arbitrary. In general, any functional form that satisfies the condition that knowledge growth is zero when the technology gap is zero could be considered. An alternative functional form, used in the empirical section, is da T / A = ln φ(.) dt A. [2] 6

7 At a theoretical level this specification is inferior since it is undefined if φ=0, however, it exhibits the same general properties as [1] and has the empirical advantage of allowing φ and the technology gap to be entered separately in a regression. The simple model highlights the importance of absorptive capability and the technology gap. Note, however, that the model describes the rate of absorption of technology from overseas: there is no allowance for the generation of internal technology. Also, the model considers the growth of technology which is not necessarily equivalent to the rate of economic growth. There are, however, a number of models which suggest that the growth rate of per capita output would equal the growth of technology or knowledge. For example, in the steady state of the Solow-Swan neoclassical model per capita GDP grows at the same rate as technology growth. This is based on a aggregate Cobb-Douglas production function with labouraugmenting technological change (i.e. Y=K β (AL) 1-β ). This is not to say that capital accumulation is unimportant without capital accumulation the economy would grow at (1-β)g only that technology is the fundamental driving force. 3 Similarly, in a number of endogenous growth models the rate of growth of GDP per capita is given by the rate of growth of technology or knowledge (e.g. Grossman and Helpman, 1991). 3 Taking the production function in the text, growth of GDP per worker (y) is given by (1-β)g + β(dk/dt/k), where k is the capital to labour ratio and g is growth of technology (A). In steady state the growth of k equals g, hence y also grows by g. 7

8 3. Absorptive capability 3.1 What determines absorptive capability? Absorptive capability can be considered to have three major elements: accessibility to overseas technology, learning ability, and the incentives or barriers to implementing new technologies. The term 'absorptive capability' is similar to the concept of 'social capability' used by Abramovitz (1986) 4. Also, in the distinct, but related, literature on national systems of innovation there is a focus on the capability of the economy to adopt and develop new technology (see Mowery and Oxley, 1995, who also use the term absorptive capability ). Lastly, similar concepts are found in theoretical models, for example, Goodfriend and McDermott (1998) use the concept of a country s familiarity with a foreign economy as a parameter in an endogenous growth model with technological catch-up. Accessibility to overseas technology depends on various factors including: business, educational, and social links; the level of trade in goods and services; and foreign direct investment. Countries with relatively high linkages would be thought of as having the prerequisites for technological catch-up. The ability to learn and understand new technology depends on a wide range of factors. At a basic level, literacy is clearly important, as will be 4 See Abramovitz and David (1996) for an updated treatment of technological catch-up and social capability. 8

9 foreign language skills. 5 More specialised 'language' skills, such as used in engineering, are also required in order to communicate technical information. These type of skills might be referred to as human capital, but there are obviously specific skills that are vital in certain contexts. The last aspect of absorptive capability is the incentive to implement new technologies. The existence and size of these incentives depends on a range of economic, social and political factors. Predominant among these are the existence of property rights, the rule of law and the degree of corruption. Government policies with regard to taxation, competition and finance, especially the consistency of their implementation, are also important. Equally, incentives are likely to be very low when political or civil disruption creates high levels of uncertainty. The macroeconomic and monetary factors emphasised by Abramovitz (1986), especially the banking and credit system, are also clearly vital. Health and life expectancy may also have an impact on the capacity and incentive to invest in economic change. Many of these factors will also determine the ability of a country to generate and implement its own new ideas. This means in empirical work it will be difficult to separate the effects of certain variables (e.g. to what extent education acts on growth through absorptive capability and the technology gap, as opposed to domestic invention and innovation). 5 An unpublished paper by Hall and Jones (1996) includes such a foreign language variable (the fraction of a country's population that speaks an international language). The authors find this variable yields a positive and significant coefficient in a regression on output per worker. 9

10 3.2 New variables that proxy the ability to access and learn overseas technology The previous literature has considered three categories of variables linked to the country s ability to absorb new technology: trade openness, foreign direct investment and human capital measures. In this paper we expand on these variables by considering study abroad, telecommunications and publications data. Table 1 shows the new variables considered. Table 1 New absorptive capability proxies Description Year(s) Study abroad in higher education based Students studying overseas p.t Students studying in USA p.t Students overseas studying natural sciences p.t Students overseas studying social sciences p.t Students overseas studying engineering p.t Telecommunications based International outgoing telegrams p.t Telex subscribers p.t Publications based Newspaper circulation p.t Periodicals published p.t. 1960, 65 Science books published p.t Volumes in libraries p.c Abbreviations: p.t. = per thousand, p.c. = per capita. Year(s) indicates the time period over which the variables has been averaged. For data sources and summary statistics see Appendix 1. An important characteristic of these variables in Table 1 is that they are all from the 1960s. In the empirical analysis below we assess the partial correlation of these variables with productivity growth over the 1960 to 1995 period. This method has the advantage of reducing any endogeneity problems (e.g. that rapid growth leads to higher income which allows more 10

11 students to study abroad). However, the method does assume that either the impact of the variables extends over a long period, or that the absorptive capability of a country is relatively stable over time. 6 Both of these assumptions may be true to a limited extent. For example, in the case of time lags, it may be that students sent abroad in 1965 will have positive benefits on the economy for the next 20 or 30 years. The students studying overseas variables are intended to proxy knowledge or technology flows into the domestic economy. The technology or knowledge could originate from their formal education, or from observation of, and interaction with, the host country. An obvious concern with using students in higher education overseas data is that students may fail to return home after their study period (a "brain drain" effect). In particular, differences in the return rate of students across countries implies finding a positive (partial) correlation if one exists between students studying abroad and subsequent economic growth less likely. Unfortunately, data on the return rate of students are not available and it is not possible to control for this potential bias. 7 It should be noted that even if some students remain abroad they may assist, to 6 Although, as a referee pointed out, this also means that disguised endogeneity may account for any regression results. 7 Some individual country data on return rates are available. The World Bank (1991, p94) asserts that, after completing their education, 63% of Korean, 49% of Jordanian and 33% of Greek students remained in the US between 1962 and The report gives no source for these data. Hugo (1996, p216) contains a reference to an estimate that of the 222,000 Chinese students who went abroad to study since 1979 only 75,000 have returned. 11

12 some extent, their home economy by passing back information, money or acting as agents for exports. A referee has also stressed that even returning students may not add to absorptive capability if they do not engage in productive activities. This is an important issue in most analyses of human capital where schooling is used as a proxy, since schooling measures the quantity supplied not the uptake of the human capital in productive activities. 8 The telecommunications variables are indicators of the level of infrastructure present in a country that is principally concerned with the transfer of information. 9 Previous studies have included the number of telephone lines per capita (see below), here the telegrams and telex based variables are included as proxies for the extent of international communication. Turning to the publication based proxies, the newspaper circulation and the periodical variables are intended to proxy the degree of literacy and knowledge dissemination within an economy. Temple and Johnson (1998) use a variable for mass communications, based on newspaper circulation and the number of radios per head, finding this variable a significant explanator of subsequent economic growth. Their interpretation of this result is that this variable is a good 8 For example, Baumol (1990) discusses how different societies offer entrepreneurs different incentives to engage in non-productive activities (e.g. crime or speculation). Barro (2001) notes that human capital created from educating females may be underutilised in some societies. 9 Dudley (1999) discusses theory, and historical evidence, on the link between information transmission and economic growth. 12

13 proxy for the "strength of civic communities, as reflected in trust and membership in associations" (p.988). 10 The publications of science books variable is a proxy for the level of specialist knowledge within a country, and also for the degree of demand for such knowledge. We might hypothesise that the greater the level of specialist knowledge, and the higher the demand for such knowledge, the greater the publications of such books. Book publications include new books, second editions or translations. The library based measure is a proxy for a number of related factors. First, countries with large numbers of volumes per capita may devote substantial effort to human capital formation. Second, the variable may proxy literacy rates. Third, the numbers of volumes in libraries may reflect the effort and importance devoted to the dissemination of knowledge within the country. 3.3 Previously used variables As noted above, the previous literature has used a number of variables that are related to the ability of a country to access and learn new technology. Many of these variables are included in our analysis to enable a comparison with the new variables discussed above. By including such variables we can shed more light on the findings of previous studies, since different 10 Romer (1990) uses the log of newsprint consumption, and the number of radios per thousand population, as instruments for the initial level of income per capita. He does not, however, enter these as separate explanatory variables, or discuss them as proxies for absorptive capability. 13

14 samples and methods are used here. For example, pre-growth period values are used for variables that have previously been investigated using contemporaneous data. Measures of trade openness have received considerable attention in the literature as proxies for technology diffusion. One argument is that some traded goods embody new technology, for example, capital goods or equipment. Another is that trade flows reflect other links between countries business, scientific, educational or social which, in turn, reflect knowledge flows. 11 Of course, trade openness may have other beneficial effects on growth that are unrelated to knowledge diffusion (see Tybout, 1992, and Edwards, 1998, for a review of the effects of trade on productivity). To empirically assess the trade openness and growth link some authors have used proxies for trade distortions (e.g. ratio of price index of tradables in country i to USA, Dollar, 1992). Others have used a range of variables to proxy the overall trade openness (Sachs and Warner, 1995). Edwards (1998) provides a summary and extensive testing of different proxies. These studies have found that trade openness and growth are positively correlated, something which has become a widely accepted economic truth. However, recent contributions by Rodriguez and Rodrik (2000) and Rodrik (1999) have criticised the robustness of this finding. Their main 11 For example, Helliwell (1992, p8-9) states, "[f]or a given country size, larger trade flows might be a good proxy for the freer flows of ideas, on the presumption that whatever pattern of access and opportunities gave rise to the trade flows would also apply to knowledge transfers. This is over-simplified reasoning, but it may provide a sensible starting point." 14

15 points concern the imperfect nature of the proxies, the role of influential observations, the fact that altering the regression specification changes results and the issue of causality. In a similar way, a theme of this paper is to question the findings concerning trade openness and growth. One of the ways of doing this is to use disaggregated trade flow data to investigate possible reasons for broad correlations. De Long and Summers (1993) do this by using the ratio of imports of equipment to GDP ratio as an explanatory variable and find it to be positively correlated with economic growth (the ratio is averaged over the growth period considered, and is used as measure of real investment). In the analysis below, the ratio of imports of equipment to GDP is used as an explanatory variable, although the value at the beginning of the growth period is used to lessen endogeneity issues. A further link between trade and technology flows may originate from export activity. Hobday (1995), in a case study of Korea, Taiwan, Hong Kong and Singapore, stresses the importance of the technology supplied by firms in export markets (i.e. the importing firms in developed countries assist with the design and production technology). To test this idea the ratio of manufactured exports to GDP, and the ratio of equipment exports to GDP, are included as explanatory variables. In both of these ratios, exports are defined as only those to developed countries. One of these ratios the ratio of manufactured exports to GDP was used by Pack and Page (1994) as an explanatory variable in a cross-country growth regression. Alwyn Young, in a comment at the end of their paper, casts doubt on their results. Two of Young's main points are the unreliable nature of the data (Pack and Page use World Bank data) and the fact that the export share is taken from the growth period (raising doubts about endogeneity). Indeed, Young provides 15

16 some regressions that show no link between pre-growth period values and subsequent growth. Since UN trade data are used below, and pre-growth period values, our results contribute to this debate. The role of foreign direct investment (FDI) in international technology and knowledge flows has also been the subject of study. In particular, a central issue is whether the presence of FDI leads to knowledge spillovers or other positive externalities with the host country. As might be expected, the existence of externalities depends on the nature of FDI and the host economy s characteristics ( Kokko, 1992, Ito and Krueger, 2001). Empirical analysis at the country level has shown that the FDI to GDP ratio has a positive and significant association with GDP per capita growth (Blomstrom et al, 1992, Balasubramanyam et al, 1996, and Borensztein et al, 1998). This suggests that FDI is likely to be a mechanism by which technology flows between countries. In this study the intention is to augment the existing literature. In the case of FDI use of pre-growth period data would be desirable, however, data are not available for many countries which prevents this approach. For this reason a variable based on FDI is omitted from the empirics below in the knowledge that the existing literature supports its importance. Two further variables linked to absorptive capability are included in the empirical analysis. One is the number of telephones per 100 population which is an indicator of domestic infrastructure in telecommunications. Easterly and Levine (1995) use data for telephones per worker in regression work and find a significant positive partial correlation with growth rates. They use the log of telephones per worker averaged over the growth period, hence, the variable can be criticised as being endogenous. In our analysis below we use the pre-growth value of this 16

17 variable. The second variable is the number of patent applications per thousand workers. Again, to avoid endogeneity concerns we use pre-growth values. Applications are defined as both those from domestic and overseas inventors, therefore, this variable has the potential to capture both domestic and foreign technology absorbed into the economy. Higher levels of patents per worker should indicate higher levels of (past) absorptive capability. Patent data have been used in a variety of studies relating to economic growth. Fagerberg (1987) uses the growth rate of external patents in explaining the growth of GDP (finding evidence of a positive link); Gould and Gruben (1996) consider an index of intellectual property rights as an explanatory variables in a cross-country regression study (finding its coefficient positive and marginally significant); and Sedgley (1998) uses patents per worker in a study of growth of GDP per capita in US states (finding a positive partial correlation). 4. Cross-country regression analysis 4.1 Empirical specification Initially, to assess the importance of the variables discussed above we use a cross-country regression framework that uses the growth rate of GDP per capita as the dependent variable. Since the mid-1980s there has been a rapid increase in such studies, investigating a huge number of variables, with the majority using Penn World Table data (Summers and Heston, 1991). There are a number of difficulties with such analysis, and the previous section has highlighted the issues of measurement error, influential observations and spurious correlation (see Temple, 1999, for a full discussion of empirical difficulties). A further issue concerns 17

18 which variables should be 'core' explanatory variables (i.e. the omission of which might seriously bias the results). Once a set of core variables has been defined, the new variables can be added to this model to assess their importance. Two main approaches to finding such a core set have been proposed. Levine and Zervos (1992) use a form of Leamer s extreme bounds analysis. They, a priori, select a core set of four explanatory variables (investment to GDP ratio, initial GDP per capita, initial secondary school enrolment and average annual population growth). Then each additional variable is tested in a series of regressions and the (statistical) lower and upper bounds on its coefficients are found. The series is defined by all combinations of three other explanatory variables from a total of around 50. The results suggest only investment and trade openness are robust correlates. Sala-i-Martin (1997) tackles this issue in a slightly different way in a paper entitled I just ran two million regressions. The regression number is from using 62 possible explanatory variables, with 3 core explanators (initial GDP per capita, initial primary school enrolment and initial life expectancy), and all others in combinations of 4 (i.e. there are almost 2 million combinations using this procedure). Summary statistics of which variables did well (assessed by considering the distribution of estimated coefficients) suggest that 22 variables are robustly linked to growth. There are various problems with both approaches, for example, the choice of core variables is made a priori. As Sala-i-Martin notes, there are 3.4 billion regressions if one allows any combination of 7 variables out of 62. Many more combinations are present if interaction terms or log transforms are added. 18

19 In practice, therefore, there is no universal agreement over which are the core explanatory variables or how to investigate new variables. This paper uses two recent, and influential, empirical specifications as baseline models for our analysis. Initially, each of the new variables is added to the baseline models. The first baseline model is from Sachs and Warner (1997). A key feature of this specification is the relatively high adjusted R 2 that the regression yields (around 0.84). The specification also has a strong trade openness focus, which is commonly thought of as a good proxy for technology flows. Specifically, the Sachs-Warner regression contains an openness index (the share of years which a country was adjudged to be 'open' between 1965 and 1990), as well as the interaction term of this index with initial income (since openness may facilitate convergence). The openness index is based on data for tariff and nontariff barriers, black market premia, whether the state has a monopoly on major exports, and if the country has a socialist system. A threshold value for each aspect is then (subjectively) defined which leads to a country being adjudged open or not open in any particular year. This methodology for assessing trade openness has been criticised (see below), but the Sachs- Warner model is widely perceived as focusing on trade openness. Given this, the new variables listed in Table 1 must 'compete' with a specification that is adjudged to incorporate an important aspect of technology flows. The second baseline model is a modified form of a regression used by Barro and Sala-i- Martin (1995). This baseline model has a strong focus on human capital, with explanatory variables for the stock of primary and secondary school education (stocks are expressed as average number of years received in population over 25 years of age). Since some of the 19

20 variables in Table 1 are included as proxies for specific forms of human capital it seems important to assess their importance alongside existing human capital variables. Data for both the Sachs-Warner and Barro-Sala-i-Martin specifications are taken from the authors' original data sets. Appendix 2 shows both the baseline models and the country samples used. In short, the results for the Sachs-Warner specification suggest that poorer countries grow faster conditional on good institutions, macroeconomic stability and trade openness. The Barro model also finds that poorer countries grow faster but this time conditional on high levels of human capital and smaller government size. Both models also use dummy variables to condition on geographic location of countries, and Sachs-Warner condition on reliance on natural resource exports and life expectancy. 4.2 Sachs and Warner specification Our initial investigation involves adding each of the variables in Table 1, and the variables discussed in section 3.3, to the baseline regressions models. This method of assessing the importance of new variables is common in the literature (e.g. Temple and Johnson, 1998). From a practical point of view, adding each proxy individually avoids multi-collinearity problems (since some of the proxies are highly correlated) and also allows an assessment of the individual importance of the variable. Each of the variables is added in log form. In the regressions, the maximum number of countries is 84 (defined by the data in the Sachs-Warner data set), although the sample size is less than this in many cases due to data limitations on many of the new variables. 20

21 The results of adding the variables to the Sachs-Warner baseline specification are shown in Table 2. For brevity, only the coefficients and t-statistics for the variable of interest are shown (full regression results are available from author). The first column of coefficients and t- statistics ( full sample ) shows the results from adding the variable to the Sachs-Warner regression. Two further columns are shown: the 'robust' and 'restricted sample' (defined below). Considering the results in the 'full sample' column in Table 2, a number of conclusions can be drawn. The coefficients on the students studying in higher education abroad variables are all positive and significant (10% level), apart from the coefficient on students studying in the USA. These results imply that countries with relatively higher numbers of students studying abroad in the early 1960s also had higher rates of growth in the next 30 years. Although causality is never proven by such regression results, reasons for such an association include (1) students do gain benefits from overseas study and these provide long term assistance to their home economy, and (2) a country that sends relatively high numbers of students abroad may have a set of other characteristics that are conducive to growth. However, the coefficient on the students studying in the US variable is not significant, suggesting this relationship is not robust. Such a result may be due to the failure to account for the return rate of students (i.e. a brain drain effect) which may be particularly critical for US-based students. Still focusing on the Sachs-Warner 'full sample' results, we can see that the telecommunications variables, although positive, are not significant. The results for the publications variables are poor: there appears to be no significant association with subsequent economic growth. In 21

22 contrast, the patent based variable has a positive and significant coefficient. One explanation for this is that this coefficient is reflecting the ability of the country to generate domestic technology and to benefit from foreign technology. However, it may also be that the coefficient is reflecting a more general institutional factor: countries that have relatively high numbers of patent applications also have efficient legal and scientific infrastructures. Note, however, that the Sachs-Warner baseline regression already includes an index of institutional quality (based on sub-indexes for "the rule of law, bureaucratic quality, corruption and alike" (Sachs and Warner, 1997, p186)). The results on the trade measures are interesting. Even though the Sachs-Warner specification includes a trade openness index we still find that the basic trade openness ratio, as well as the ratios of manufactures and equipment exports to GDP, have positive and significant associations with economic growth. 12 These results give wider empirical support to Hobday's observations concerning the role of exports in enabling a reverse flow of technology. Equally, the positive and significant (10% level) coefficient on the ratio of basic manufactured exports to GDP supports the Pack and Page (1994) findings. Lastly, in contrast to the results of De Long and Summers (1993), the pre-growth period value of the ratio of imports of equipment to GDP appears to have no explanatory power. Table 2 Results from including new variables in baseline regressions 12 For reference, in the regression with our measure of trade openness included, the coefficient on the Sachs-Warner measures of trade openness remain significant with little change in magnitude. 22

23 Explanatory variable Sachs-Warner specification Coefficient (t-stat) obs Full sample Restricted sample Robust Regression All variables entered in log form Study abroad based Students studying o'seas p.t (1.9) 0.49 (0.9) 0.47 (1.5) Students studying natural sciences p.t (1.7) 6.19 (2.1) 2.87 (2.0) Students studying social sciences p.t (1.7) 3.80 (1.9) 3.72 (3.6) Students studying engineering p.t (2.4) 5.32 (2.5) 1.82 (1.6) Students studying in USA p.t (1.0) 0.58 (0.5) 0.12 (0.1) Telecommunications based Main telephone lines per 100 population (1.4) 0.34 (1.2) 0.23 (0.9) International outgoing telegrams p.t (1.2) 0.23 (1.7) 0.04 (0.4) Telex subscribers p.t (1.4) 0.19 (1.5) (-0.2) Publications based Newspaper circulation p.t (-0.7) (-0.7) 0.01 (0.1) Periodicals published p.t (-0.3) (-0.3) 0.00 (-0.0) Science books published p.t (-0.6) (-0.5) (-1.0) Volumes in libraries p.c (0.6) 0.14 (0.7) (-1.0) Additional variables Patent applications p.t. workers (2.2) 0.35 (2.0) 0.27 (1.6) Trade openness (X+M/GDP)* (2.5) 0.19 (1.0) 0.31 (2.0) Imports of equipment / GDP (-0.1) (-1.0) (-0.1) Exports of basic manufactures / GDP (1.7) 0.10 (1.5) 0.08 (1.7) Exports of equipment / GDP (2.6) 0.12 (2.6) 0.12 (2.5) Notes: Results show coefficient (t-statistic) from adding each variable separately to baseline regression in log form (where zero values occur in a variable (x), the variable is transformed to log(1+x)). Dependent variable is growth of GDP per capita The t-statistics in columns 1 and 2 are White's corrected. Approximate critical values are 2.0 (5% level) and 1.7 (10%). Number of observations refers to full sample. Restricted regression and robust regression results are explained in text. 23

24 An important issue in empirical growth analysis is the potential for influential observations to drive the results on certain coefficients. An inspection of the data shows a number of the new variables have a few countries with relatively high values. Influential observations within other explanatory variables can also, of course, affect regression results something often not stressed in the literature (Temple, 1999). A simple way to assess the potential role of influential observations is to exclude countries with high 13 values for each of the absorptive capability proxies from the regression. The results in the second column of regression results in Table 2 ( restricted sample ) show the impact of this method. A more rigorous method is to use robust regression techniques. The robust regression estimates have the following procedure. Initially, the residuals from the OLS regression are analysed to give each observation a weight based on its residual s relative magnitude. A regression using the weighted data is then run. This procedure of re-weighting continues until no large residuals exist. 14 The results from this robust regression method are shown in the third column of regression results. 13 A high value is defined as above the 95 th percentile of the distribution. 14 More specifically the procedure is: (1) calculate Cook s D for each observation from OLS (2) observations with values greater than one are given zero weight (3) re-run regression (4) calculate M=med( e i -med(e i ) ), where e i is the residual (5) any observation with an absolute residual greater than 2M receives a downweight of 2M/ e i (called Huber weights) (6) repeat procedure until maximum change in weights drops below 0.01 (called convergence ) (7) based on final regression in step 6, repeat procedure using biweights, which are downweights given by [1-(7e i /M) 2 ] 2, until convergence. The procedure is described in more detail, with appropriate references, in STATA 6.0 reference manual under rreg ( 24

25 Looking at second and third column in Table 2 the significance levels of the coefficients are similar to the full sample regressions. Notable exceptions are the 'students studying overseas' coefficient (not significant in columns 2 and 3), and also the trade openness coefficient (not significant in restricted sample). However, the magnitude of some of the coefficients on the specific subjects of overseas study have risen, implying that the outlier countries for these variables tend to reduce the strength of association with economic growth. 4.3 Barro and Sala-i-Martin Specification The above analysis was also carried out using the Barro and Sala-i-Martin specification as the baseline model. The results are shown in Table 3. 25

26 Table 3 Results from including interaction terms in baseline regressions Explanatory variable Barro-Sala-i-Martin specification Coefficient (t-stat) obs Full sample Restricted sample Robust regression All variables entered in log form Study abroad based Students studying o'seas p.t (0.7) 1.17 (0.9) (-0.0) Students studying natural sciences p.t (1.4) 8.94 (1.6) 8.73 (3.2) Students studying social sciences p.t (0.1) 6.77 (1.6) (-0.1) Students studying engineering p.t (1.3) 6.39 (1.3) 1.78 (0.8) Students studying in USA p.t (-0.5) 2.78 (0.7) (-1.4) Telecommunications based Main telephone lines per 100 population (1.8) 1.25 (2.2) 0.81 (1.5) International outgoing telegrams p.t (-1.9) (-0.6) (-1.9) Telex subscribers p.t (3.6) 0.52 (3.9) 0.48 (2.7) Publications based Newspaper circulation p.t (3.0) 1.20 (2.8) 1.52 (4.6) Periodicals published p.t (2.2) 0.52 (1.9) 0.38 (1.5) Science books published p.t (1.3) 0.42 (1.7) 0.23 (0.8) Volumes in libraries p.c (2.2) 0.53 (1.8) 0.62 (2.1) Additional variables Patent applications p.t. workers (1.8) 0.70 (1.9) 0.29 (0.9) Trade openness (X+M/GDP)* (0.3) (-1.0) 0.00 (0.0) Imports of equipment / GDP (-0.6) (-0.4) (-1.0) Exports of basic manufactures / GDP (0.8) 0.09 (0.9) 0.06 (0.6) Exports of equipment / GDP (3.7) 0.34 (3.8) 0.32 (3.8) Notes Dependent variable is growth of GDP per capita The t-statistics in columns 1 and 2 are White's corrected, approximate critical values are 2.0 (5% level) and 1.7 (10%). Number of observations refers to full sample. The results in Table 3 show that the students studying overseas variables, although often having positive coefficients, are not significant at conventional levels in the full sample regressions. The significance of some coefficients improves in the restricted and robust regressions but, overall, the results contrast strongly with the Sachs-Warner baseline results. One possibility here is that the inclusion of the primary and secondary education explanatory variables in the Barro and 26

27 Sala-i-Martin baseline removes some of the significance from the students overseas variables. Considering the telecommunications variables, the Barro and Sala-i-Martin full sample results show that telephone lines and telex subscribers have a positive association with growth, while the telegrams based variable has a negative association. A further difference with the Sachs- Warner results are the results on the publications variables. In the Barro and Sala-i-Martin specification the coefficients on these variables are all positive and those on the periodicals, newspaper circulation and volumes in libraries variables are significant. These results are generally maintained in the restricted sample and robust regression. These results support Temple and Johnson's (1998) findings of an association between mass communications and economic growth, although our interpretation differs. Here we suggest that these variables proxy a country's ability to diffuse and learn new technology which, conditional on other factors, raises economic growth. Comparing the Sachs-Warner and Barro-Sala-i-Martin results for the additional variables in the lower section of Table 2 we again see some variation in the results. The coefficient on the patent variable maintains its significance in the full and restricted sample (its magnitude is also twice as high in the Sachs-Warner specification). Trade openness appears unimportant in the Barro-Sala-i-Martin regressions, but important in the Sachs-Warner specification. Equally, the ratio of basic manufacturing exports to GDP loses its significant coefficient in the Barro-Sala-i- Martin specification. Lastly, the ratio of equipment exports to GDP is positive and significant 27

28 coefficient across both specifications (although the coefficient is over twice as high in the Barro and Sala-i-Martin specification). An indication of the economic importance of the coefficient estimates reported above can be made using the summary statistics in Appendix 1. Using these we can calculate, for example, that a one standard deviation increase in students studying natural sciences abroad (per thousand population) is associated with a 0.3% rise in average annual economic growth. Similarly, a rise of one standard deviation in the ratio of patent applications to workers is associated with a 0.6% rise in average annual growth (both of these examples use the full sample coefficient estimates in the Sachs-Warner specification). As always, these calculations in no way imply causality. 4.4 An encompassing specification The previous two sections have used two existing models as a method of assessing the importance of the new variables. Adding each of the new variables in turn provides an indication of their influence, however, there is a desire to assess which variables may be most important. One method is to add all the variables to a single regression and then reduce the model according to the significance of the coefficients. This method is precluded due the large numbers of variables, the low number of observations for some of the variables and the high 28

29 correlations between sub-sets of the variables. 15 A further issue is that the two baseline models are substantially different (see Appendix 2). Both of these issues suggest the need to think further about a reduced form specification. The subsequent paragraphs explain how a more parsimonious specification can be justified. The main difference between the Sachs-Warner and Barro and Sala-i-Martin specifications is the inclusion of the constructed openness variable in the former and the schooling variables in the latter. Rodriguez and Rodrik (2000) have criticised the Sachs-Warner openness variable since most of its explanatory power comes from the black market premia component (which is also a proxy for macroeconomic conditions) and the state export monopoly component (which is very similar to a dummy variable for sub-sahara African countries). Given this, in the reduced form regressions shown in Table 4 the initial black market premia (average 1960 to 1964) is added directly. There is also an issue concerning using geographic dummy variables in regression analysis. While these variables have explanatory power they also offer no insight for any underlying mechanisms that might drive growth. To avoid this possibility the regressions shown in Table 4 contain no regional dummies. The Sachs and Warner specification includes life expectancy (both log of level and square of log of level) which is intended as a proxy for human capital. In contrast, the Barro and Sala-i- 15 For example, trade openness has correlation coefficients between 0.01 and 0.48 with the other variables; the telephone per capita variable has correlation coefficients of over 0.74 with the publication based variables. 29

30 Martin specification use the average years of secondary schooling as a human capital proxy, which would appear a more direct proxy. In view of this, in the regressions below life expectancy is replaced with the secondary schooling stock (although the quadratic form suggested by Sachs-Warner is retained). Finally, the Sachs-Warner regression contains the budget balance averaged over the growth period, which can be viewed as endogenous, is omitted. Instead, the specification below includes the pre-growth period ratio of real government expenditure (net of defence and education) to GDP as in the Barro and Sala-i- Martin model. The resultant regression based on these changes is [1] in Table 4. This regression contains the expected signs on coefficients. It is worth noting that the coefficients on the quadratic expression for the influence of secondary schooling imply that such schooling has its maximal impact at around 1.55 years of schooling (the sample maximum is 3.2). The results in Table 2 and 3 show that two of the new variables are generally significant in all the different specifications, namely those based on students study natural science abroad and exports of equipment. In addition, there is some suggestion that the variables based on main telephone lines, engineering students abroad, newspaper circulation and patents applications may have importance. Since the students studying natural science abroad variable is positively correlated with the variable based on engineering students (correlation coefficient 0.65), these variables are merged into a single variable. Regression [2] in Table 4 adds these five variables. Regression 2 shows that none of the new variables are significant. Part of the reason is likely to be multicollinearity (the last four variables in Table 4 have correlation coefficients between 0.59 and 0.93). A further drawback of including all these variables is that the sample size 30

31 drops to 47. Various additional regressions were undertaken dropping combinations of the new variables, the results indicate that only the study abroad variable showed a strong partial correlation. Regression [3] shows the full results when only the log of students studying natural science or engineering abroad per thousand is included. The adjusted R 2 for [3] is 0.726, as compared to if the specification in [1] is used on this sample. Note also that the inclusion of the study abroad variable causes the coefficient on the black market premia to become insignificant. The coefficient magnitude suggests a 1 standard deviation difference in the study abroad variable in 1965 is associated with a 0.3% increase in average annual growth

32 Table 4 Reduced form regressions [1] [2] [3] Log of GDP per worker in (-8.52) (-3.30) (-7.75) Black market premia ( ) (-1.88) (-0.18) (-0.78) Growth of economically active population less population growth (6.53) (2.22) (5.43) Ratio of real government expenditure (net of defense and education) to GDP (1960 to 1964) (-2.42) (-2.38) (-2.42) Institutional quality index (1980) (5.54) (3.12) (6.52) Share of natural-resource exports in GDP, (-3.55) (-1.72) (-3.57) Average years of secondary schooling in the total population over age 25 (1965) (4.42) (1.45) (3.48) Square of secondary schooling variable (-4.64) (-1.63) (-3.38) Students studying natural sciences or engineering abroad p.t (1966) (0.73) (2.45) Newspaper circulation p.t.( ) (-0.43) Patent applications p.t. workers ( ) 0.06 (0.17) Exports of equipment / GDP ( ) (-1.52) Main telephone lines per 100 population (1965) (1.18) Observations Adjusted R-squared

33 Notes: Dependent variable is trend growth in GDP per worker All t-statistics based on White's robust s.e.'s. A robust estimator, as defined earlier, performed on the sample in [3] shows very similar results. 5. Multifactor productivity In this section the dependent variable is a measure of multifactor productivity (MFP) growth. The theoretical arguments in section 2 suggest that absorptive capability and technology gap determine growth of technology directly. Many authors use MFP growth as an indicator of technology growth, hence this methodology is adopted here. 16 The data for MFP are taken from a study by Beck, Levine and Loayza (2000). MFP growth is defined as d ln MFP dt d ln y d ln k = 0. 3 dt dt where y equals GDP per capita and k equals capital stock per capita, and the growth period is 1960 to The capital stock is based on Penn World Table investment data (excluding residential investment), using a perpetual inventory model with a 7% rate of depreciation (see Beck et al, 2000). In Beck et al they analyse the link between the level of financial development within countries and MFP growth ( ) using, at least initially, a simple model with two explanatory variables: the initial GDP per worker; and the log of average years of schooling in population over 25 years of age in This is the approach adopted here 16 Some authors note that MFP may not reflect the full importance of technology growth (Nelson and Pack, 1999). 33