Long-Term Barriers to the International Diffusion of Innovations

This is the authors' abstract. We don't add key points for this paper.

Previous articleNext article FreeLong-Term Barriers to the International Diffusion of InnovationsEnrico Spolaore and Romain WacziargEnrico SpolaoreTufts University, NBER, and CESIfo and Romain WacziargUCLA, NBER, and CEPRPDFPDF PLUSFull Text Add to favoritesDownload CitationTrack CitationsPermissionsReprints Share onFacebookTwitterLinked InRedditEmailQR Code SectionsMoreI. IntroductionDifferences in technology adoption lie at the heart of differences in long-term economic performance and standards of living across societies. Total factor productivity (TFP), an indirect measure of technology based on the Solow residual, accounts for a vast share of differences in income per capita across countries (e.g., Caselli 2005; Hsieh and Klenow 2010). Comin, Hobijn, and Rovito (2008) have provided evidence of large lags in the usage of specific technologies, implying that technology usage disparities might account for a large part of cross-country TFP differentials. Technological innovation has also played a paramount role in the study of economic growth and development, both theoretically and empirically (e.g., see Aghion and Howitt 2008). In this literature, the question of why the adoption of new productivity-enhancing technologies takes place so slowly and unequally across different societies remains a central puzzle.In this paper, our goal is to shed light on the barriers and obstacles that prevent or delay the adoption of specific technologies across countries. We focus in particular on human barriers, holding fixed geographic barriers. We formulate and test the hypothesis that rates of technology adoption depend on measures of long-term historical relatedness between populations; that is, on their degree of kinship. The main idea of this paper is twofold: (1) on average, populations that are more closely related tend to be more similar with respect to traits (habits, customs, beliefs, values, etc.) that are transmitted with variation from one generation to the next; and (2) similarity in intergenerationally transmitted traits tends to reduce the barriers to technology adoption; that is, populations that share more similar intergenerationally-transmitted traits face lower costs when imitating each other's innovations.We argue that long-term genealogical distance works as a barrier to the diffusion of innovations across populations. It is important to emphasize that we are not assuming that populations would necessarily inherit traits that make them directly more productive. In fact, in our theoretical framework, inherited traits have no direct effect on a population's productivity. What matters, in our model, is that random historical divergence introduces different customs, habits, and norms across populations, and that these differences, on average, tend to decrease their ability to learn from each other. Even relatively trivial differences in attitudes, appearance, or behavior between groups may lead to misunderstanding or discrimination, and may create significant barriers to communication and social interactions, reducing opportunities for learning and imitation.1 In sum, our central hypothesis is that differences in technology adoption across societies can be explained by barrier effects inversely related to the degree of relatedness between human populations.Building on our previous work on the diffusion of economic development (Spolaore and Wacziarg 2009), we capture historical relatedness between populations using genetic distance. Genetic distance is calculated as a summary measure of the difference in allele frequencies across a set number of genes, or loci. As human populations split from each other over the course of history, they developed distinct genetic markers from random mutations. Most of these genetic mutations did not confer any environmental advantages and diffused as the simple result of randomness (they were not selected), so they are called neutral genes (Kimura 1968). Genes used to compute common measures of genetic distance are chosen from among neutral genes. For such genes, random mutations arise and diffuse at a constant rate, so measures of genetic distance based on a large enough number of neutral genes reflect the time separating two populations from a common ancestor, as in a molecular clock. We argue that human traits that create barriers between populations on average diverge to an extent proportional to the separation time between two populations.Thus, genetic distance is an ideal measure of the degree of historical relatedness between populations. An intuitive analogue is the concept of relatedness between individuals: two siblings are more closely related than two cousins because they share more recent common ancestors—their parents rather than their grandparents. Our key hypothesis is that technological innovations are more likely to be adopted across siblings than across more distant cousins: separation times between populations are associated with similarity in a wide range of traits, transmitted either genetically or culturally from parents to children, that matter for the ease of adoption of productive innovations.Figure 1 (from Cavalli-Sforza, Menozzi, and Piazza 1994, 78) is a phylogenetic tree illustrating how different human populations have split apart over time. Such phylogenetic trees, constructed from genetic distance data, are the population analogs of family trees for individuals.Fig. 1. Genetic distance among 42 populations. Source: Cavalli-Sforza et al. (1994)View Large ImageDownload PowerPointIt must be emphasized that genetic distance is based on neutral change, and therefore is not meant to capture differences in specific traits that may directly matter for fitness and survival. Genetic distance is a general measure of long-term relatedness, and is associated with the whole set of traits that are transmitted "vertically" from parents to children, biologically and culturally. It is therefore a general metric for the average difference in traits passed on across generations over the very long run—that is, over the time horizon along which populations have split apart. In this paper we call "vertical traits" all characteristics that are passed on intergenerationally with variation within a population over the very long run.This paper is the first to empirically document the effects of genealogical relatedness on the adoption of specific technologies across countries. To do so, we use historical and contemporary data on usage rates of a wide range of technologies, dating back to the year 1500. We use two recently developed databases of technology adoption: the Cross-Country Historical Adoption of Technology (CHAT) data set, which covers a wide range of detailed technologies going back to the year 1800 (Comin and Hobijn 2009), and the database of historical technology adoption used in Comin, Easterly, and Gong (2010) to study technological persistence since antiquity (this database goes back to 1000 BC, but we only make use of data since 1500 AD). We compare the empirical effects of the bilateral genetic distance between populations to that of genetic distance relative to the technological frontier, finding that the latter trumps the former as a determinant of bilateral differences in technological adoption rates. This empirical test is consistent with a barrier effect of long-term historical distance, whereby societies that are more distant from the technological frontier tend to face higher imitation costs. We find large and statistically significant effects of relative frontier distance on technology use differences. These large effects at the level of individual technologies can help explain current differences in total factor productivity and income per capita across countries.This paper is organized as follows. Section II presents a stylized model of the diffusion of technologies as a function of differences in vertically transmitted traits across human populations, and ultimately as a function of the degree of genealogical relatedness between them. Section III presents our data and empirical methodology. Section IV describes the empirical results, and Section V concludes.II. A Theory of Relatedness and Technology AdoptionIn this section we present a simple framework linking long-term relatedness (measured by genetic distance), intergenerationally-transmitted traits, and the adoption of technologies across populations. The main ideas are that (1) genetic distance measures the degree of long-term genealogical relatedness of different populations over time, (2) on average, populations that are more closely related tend to be more similar with respect to traits that are transmitted with variation from one generation to the next, (3) similarity in intergenerationally-transmitted traits tends to reduce the barriers to technology adoption: populations that share more similar intergenerationally-transmitted traits face lower costs when imitating each other's innovations. These hypotheses have testable empirical predictions: intensive and extensive margins of technology adoption should be a function of different populations' genetic distance relative to the technological frontier, rather than the bilateral distance between them. The empirical section then tests this prediction.A. Genetic Distance and Vertically Transmitted TraitsAs discussed earlier, genetic distance measures the difference in allele distributions between two populations, where the chromosomal loci under consideration are neutral—that is, they change randomly and independently of selection pressure. When two populations split apart, random genetic mutations result in genetic differentiation over time. The longer the separation time, the greater the genetic distance computed from a set of neutral genes. Therefore, genetic distance captures the time since two populations have shared common ancestors—that is, the time since they have been the same population. Overall, genetic distance is a general metric for average differences in traits passed on across generations over the very long run (i.e., over the time horizon along which populations have split apart). We call "vertical traits" all those traits that are passed from one generation to the next within a population, with variation, over the very long run.These concepts can be illustrated with a simple analytical example. For simplicity, let all vertical traits of a population be summarized as a point on the real line. At time 0 ("the present") population i has vertical traits vi(0), where vi(0) is a real number. In general, populations inherit their vertical traits from their ancestor populations with variation. Specifically, population i living at time 0, and descending from ancestral population ai living at time −T, will have traits given by:(1)vi(0)=va(i)(−T)+εi(−T,0),where va(i)(−T) are ancestral population a(i )'s vertical traits at time −T, and ϵi(−T,0) measures random variation between time −T and time 0.Consider the simplest possible mechanism for variation: vertical change as a random walk, whereas ϵi(−T,0)=ϵ>0 with probability 1/2 and ϵi(−T,0)=−ϵ <0 with probability 1/2. In addition, consistently with a "neutral" view of intergenerational change, we assume that such shocks are independent across different populations—that is, shock ϵi(−T,0) is independent of shock ϵj(−T,0) for j≠i. By the same token, at time −T population a(i)'s vertical traits are inherited from its ancestor population a(a(i)), living at time −T'<−T , according to the following equation:(2)va(i)(−T)=va(a(i))(−T′)+ϵa(i)(−T′,−T)where εa(i)(−T′,−T)ε′>0 with probability 1/2 and εa(i)(−T′,−T)−ε′ >0 with probability1/2. Again, shocks are independent across populations.Now, consider three populations (i=1,2,3). Population 1 and population 2 descend from the same last common ancestor population a(1)=a(2) ≡a(1⁢&2) living at time −T. In contrast, population 3 only shares common ancestors with populations 1 and 2 going back to time −T' : a(3)≠a(1&2), and a(a(3))a(a(1&2))≡a(1⁢&2&3). The phylogenetic tree of the three populations is illustrated in figure 2.Fig. 2. Population treeView Large ImageDownload PowerPointBy construction, genetic distance g(1,2) between population 1 and population 2 is smaller than genetic distance between population 1 and population 3 (and smaller than genetic distance between population 2 and population 3). Formally: (3)g(1,2)F<g(1,3)g(2,3)=F'.Vertical traits in each population are given by the following equations:(4)v1(0)va(1&2)(−T)+ε1(−T,0)(5)v2(0)va(1&2)(−T)+ε2(−T,0)(6)v3(0)=va(3)(−T)+ε3(−T,0)(7)va(1&2)(−T)=va(1&2&3)(−T′)+εa(1&2)(−T′,−T)(8)va(3)(−T)va(1&2&3)(−T′)+εa(3)(−T′,−T).Let dv(i,j)≡ |vj−vi | denote the distance in vertical traits between population i and population j. The expected vertical distance between population 1 and population 2, which share their last common ancestors at time −T, is:2 (9)E dv(1,2)=ε.Clearly, all variation between populations 1 and 2 is given by change that took place between time −T and 0. In contrast, expected distance between population 1 and population 3 (and between population 2 and population 3) will reflect shocks that took place both between time −T' and −T and between time −T and time 0. On average, such shocks bring about a larger expected distance:3(10)E[dv(1,3)]E[dv(2,3)]max{ε′2+ε,ε′+ε2}>ε=E[dv(1,2)].Thus, larger genetic distance, on average, is associated with larger distance in vertical traits. Of course, this relation is not deterministic. Some pairs of populations that are genealogically more distant may end up with more similar vertical traits than two more closely related populations, but that outcome is less likely to be observed than the opposite outcome. On average, genetic distance and vertical distance go hand in hand.B. Relative Vertical Distance from the Frontier and Technology AdoptionOver the millennia, populations have diverged in a series of traits transmitted vertically from parents to children. These vertical traits include physical and cultural characteristics. Most of these traits have diverged randomly, and do not need to have a direct effect on survival and fitness. However, this divergence has led to differences across populations. Our key hypothesis is that such differences may act as barriers to the diffusion of technological innovations across populations in modern times. It is reasonable to expect that, on average, populations that share a more recent common history and are therefore closer in terms of intergenerationally-transmitted traits face lower costs and obstacles to adopting each other's innovations. The literature on the diffusion of innovations (Rogers 1995) is consistent with a major role for subjective barriers between groups and populations. As Rogers points out, summarizing the lessons from decades of research, most people depend upon a subjective evaluation of an innovation that is conveyed to them from other individuals like themselves who have previously adopted the innovation. Overall, historical relatedness and similarity is likely to reduce imitation costs, while a higher vertical distance should be associated with higher imitation and adoption costs. In sum, our main hypothesis is that long-term divergence in vertically transmitted traits act as a barrier to more recent "horizontal" diffusion of innovations across societies.1. Extensive MarginThese ideas can be formalized very simply in a one-period model. At time 0, one of our three populations—say, population 1—introduces a continuum of innovations of mass N.4 An innovation k ∈0,N] will be adopted by population i =2,3 if and only if the innovation's benefits bk are higher than its costs cki.5 Our key hypothesis is that adoption costs for each population are a function of vertical distance from the frontier (population 1):(11)cki=χk+ϕ dv(i,f )where χk>0 and ϕ>0. Different innovations come with different costs and benefits. To fix idea, assume that the ratio (bk−χk)/φ is distributed uniformly between 0 and ω>0, with innovation 0 having the highest ratio (b0−χ0)/ϕ=ω and innovation N having the lowest ratio (bN−χN)/ϕ=0. In general, the ratio for innovation k∈0,N] is:(12)bk−χkφ=ω[1−kN].Population i will adopt Ni∗ innovations, where, for each adopted innovations k∈0,Ni∗], bk≥cki. Hence, the marginal Ni∗ is defined by:(13)bNi∗=χNi∗+φ dNv(i,f ),which, by using equation (12), can be rewritten as:(14)bNi∗−χNi∗φ=dv(i,f )=ω[1−Ni∗N],which implies:6 (15)Ni∗=N[1−dv(i,f )ω].This equation shows that the number of innovations adopted by each population is decreasing in its vertical distance from the frontier. If vertical traits change according to the simple random-walk process illustrated in the previous subsection, the expected rate of adoption for population 2 (which is genetically closer to the frontier population 1: g(1,2)=F ) will be larger than the expected rate of adoption for population 3 (which is genetically more distant from the frontier g(1,3)=F′>F):7 (16)E{N2∗}=N[1−E[dv(1,2)]ω]=N[1−εω]>E{N3∗}=N[1−E [dv(1,3)]ω]=N[1−εω−ε′ 2ω].Let dN(i,j )≡ |Nj∗−Ni∗ | denote the difference in rates of technology adoption (extensive margins) between population i and population j. Our basic framework implies that such difference will depend on relative vertical distance from the frontier, and hence, on average, on relative genetic distance from the frontier population:(17)dN(i,j )≡ |Nj∗−Ni∗ |=1ω|dv(j,f )−dv(i,f ) |.Therefore, we predict a positive correlation between differences in rates of technology adoption dN(i,j ) between two populations i and j, and their relative frontier distance, defined as |g(i,f )−g(j,f ) |.This framework can also account for a positive correlation between differences in rates of technology adoption dN(i,j ) and bilateral genetic distance g(i,j ), as long as |g(i,f )−g(j,f ) | and g(i,j ) are positively correlated (as they indeed are in the data). However, our theory also implies that relative frontier distance is a better predictor of technology adoption, because it is a more accurate measure of relative distance from the frontier in terms of vertical traits. For example, if one were to use bilateral genetic distance as a proxy for expected vertical differences, on average one should expect that differences in rates of technology adoption between population 2 and population 3 ought to be identical to differences in technology adoption between population 1 and population 3, because g(1,3)=g(2,3). This would be possible if and only if E |N3∗−N2∗ |=E |N3∗−N1∗ |=E |N3∗−N |=N[(ε/ω )+(ε′/2ω)], which would hold if and only if E{N2∗}=N. But, as we have seen earlier, on average population 2 does not adopt all innovations, because it faces barriers associated with its own vertical distance from the frontier: E |N2∗−N |=N(ϵ/ω)>0. Hence, using bilateral genetic distance g(2,3) as a measure of barriers to technology adoption between population 2 and population 3 would lead us to overestimate the adoption gap between population 2 and population 3 (or, equivalently, to underestimate the adoption gap between population 1 and population 3). In general, our theory predicts that bilateral genetic distance is a biased proxy for the fundamental determinant of comparative adoption of innovations across societies. This fundamental determinant is the relative distance in vertical traits, which can be measured more accurately by relative frontier distance, as defined earlier.8 For further discussion of the relation between relative frontier distance and bilateral genetic disatance, see Spolaore and Wacziarg (2009, 477). We test the prediction that relative frontier distance should trump bilateral genetic distance as a determinant of technology adoption in the empirical section.2. Intensive MarginAn analogous illustration of the main idea can be provided with respect to intensive margins of adoption. This is important because much of our technology data focuses on the intensive margin. Let Zki denote the level at which a given innovation k is used by population i. Assume that benefits and costs of adoption are now given as follows:(18)Bk=βkZk(19)Ck=γk2[φ0−φ1dv(i,f )]Zk2.Hence, net benefits for population i are maximized at:(20)Zki∗=βk[φ0−φ1dv(i,f )]γk.Again, the difference in adoption rates (this time at the intensive margin) will be a function of relative vertical distance from the frontier and hence, on average, relative genetic distance from the frontier population:(21)dZk(i,j )≡|Zkj∗−Zki∗ |=βkφ1γk|dv(j,f )−dv(i,f ) |.The above results illustrate our hypothesis that technology adoption is a function of relative distance in intergenerationally-transmitted traits. As we have already discussed, such relative distance between populations can be captured empirically by their relative genetic distance from the frontier.C. A Dynamic ExampleIn the rest of this section we illustrate our general ideas in a dynamic framework that will allow us to establish a more explicit and microfounded link between the adoption of innovations, imitation costs, distance in vertical traits, and total factor productivity. Changes in vertical traits tend to take place much more slowly and at a longer horizon than the spread of technological innovations, especially if we focus on modern technological progress and growth in a post-Malthusian world. Therefore, in this simple dynamic example we assume for simplicity that populations do not change over time (their size is fixed and normalized to one), and that their inherited vertical traits do not change over the relevant time horizon (additional small random shocks to vertical traits would not affect the basic results, while they would greatly complicate the algebra). Hence, consider our three populations i=1,2,3, with vertical traits v1, v2, and v3, inherited from their ancestral populations as described earlier, and unchanged in modern times (i.e., for t≥0). Time is continuous. At each time t ≥0 consumers in each economy i maximize:(22)Ui(t)=∫s∞lnCi(s)e−ρ(t−s)dsunder a standard budget constraint, where Ci(t) is consumption, and ρ>0 is the subjective discount rate. The Euler equation for the optimal growth rate of consumption holds:(23)dCidt1Ci(t)=ri(t)−ρ,where ri(t) is the real interest rate in society i (we assume that the three societies are not financially integrated). At time t, in each society i, there is a continuum of intermediate goods, measured on the interval [0,Ai(t)]. Each intermediate good is produced by a local monopolist. In each society i final output Yi(t) is produced according to:(24)Yi(t)=0Ai(t)[Xki(t)]αdk, 0<α<1,where Xki(t) is the quantity of intermediate good of type k employed at time t in economy i. We assume that society f is at the technological frontier, which means that Af(t0)>Ai(t0) for all i ≠f. Innovation at the frontier economy takes place endogenously, as in Romer (1990) and related literature. In particular, as in Barro and Sala-i-Martin (1997 and 2003, chapters 6 and 8), we assume that the inventor of intermediate good k retains perpetual monopoly power over the production of that input in society f, and henceforth sells it at price Pk=1 /α, earning the profit flow π=(1−α)α(1+α)/(1−α) at each time t.9 For a detailed derivation see Barro and Sala-i-Martin (1997 and 2003). The cost of inventing a new intermediate good at the frontier is λ units of final output. Free entry into the innovation sector implies that the real interest rate rf(t) must be equal to π/λ. We assume π/λ>ρ, which implies that consumption grows at the constant rate:(25)g≡πλ−ρ>0.Consequently, output Yf(t) and the frontier level of intermediate goods Af(t) will also grow at the rate g.The other societies cannot use the intermediate goods invented in economy f directly, but, as in Barro and Sala-i-Martin (1997), must pay an imitation cost μi in order to adapt the intermediate goods to local conditions. Our key assumption is that such imitation costs are increasing in the distance in vertical traits between the imitator and the frontier. Specifically, we assume that economy i's imitation cost is:(26)μi(t)=λeθdv(i,f)(Ai(t)Af(t))ψ.The above specification is consistent with our main hypothesis: differences in vertical traits act as barriers to adoption and imitation. The parameter θ captures the extent to which dissimilarity in vertical characteristics increases imitation costs. For a given distance in vertical characteristics, an imitator in society i faces lower imitation costs when there is a larger set of intermediate goods available for imitation (that is, when Ai(t) /Af(t) is low). The rationale for this assumption is that the intermediate goods that are easier to imitate are copied first. The parameter ψ>0 captures this advantage from technological backwardness. Wh