Economic complexity and jobs: an empirical analysis

ABSTRACT This paper analyses the impact of economic complexity on the labour market using annual data on OECD countries for the period 1985–2008 and averaged data over the period 1990–2010 for 70 developed and developing countries with a large number of controls. We show that moving to higher levels of economic sophistication of exported goods leads to less unemployment and more employment, revealing that economic complexity does not induce job loss. Our findings remain robust across alternative econometric specifications. Furthermore, we place the spotlight on the link between products' embodied knowledge (sophistication) and labour market outcomes at the micro-level. We build a product-level index that attaches a product to the average level of unemployment (or employment) in the countries that export it. With this index, we illustrate how the development of sophisticated products is associated with changes in the labour market and show that the economic sophistication of exported goods captures information about the economy's job creation and destruction.


Introduction
Progressing to a more complex economy by developing and producing new sophisticated products is a process of creative destruction that directly affects the labour market by creating and destroying jobs. Although it is relatively straightforward to highlight positive and negative effects in particular cases, it is not that obvious to analyze and measure the overall (net) outcome of the economic complexity advancements throughout the economy. During the last 20 years, new sophisticated products have radically transformed sectors/industries, leading to a destruction of obsolete jobs and, in parallel, to the creation of new ones (Feldmann 2013). Product sophistication can displace labour by reducing or eliminating the demand for particular goods and/or services, in sectors which are specialized in routine activities (Dao et al. 2017). In addition, it can reduce employment within 'highly automatable' occupations, through the introduction of machines and robots (Acemoglu and Restrepo 2020; Graetz and Michaels 2018). On the other hand, 'automation' can increase the needs for job requirements of complex tasks and thus high skilled workers, availing employment (Autor and Salomons 2018). Thus if technological advances in production can lead concomitantly to both creation and destruction of jobs in the labour market, one can reasonably ask: ceteris in labour market outcomes (unemployment and employment rates) we would expect if a country were to modify its product mix by adding or removing a product.
The remainder of the paper is structured as follows. Section 2 discusses the data. Section 3 presents the methods used in the paper. Section 4 describes the econometric analysis for studying the impact of economic complexity on labour market outcomes, discusses the control variables and the instruments of ECI included in the model and presents and discusses the results. Section 5 introduces two indexes of the unemployment and employment rates expected for the producers and exporters of 773 different products in the Standard Industrial Trade Classification at the four-digit level . Using these indexes, we put the spotlight on the links between export's sophistication and the unemployment and employment rates, at the micro-level. We illustrate that the development of more complex products is associated with lower unemployment and higher employment rates. Finally, in Section 6, we draw our conclusion.

Data
Data on labour market are taken from the International Labour Organization (ILO) as reported in the World Bank's World Development Indicators database. Even though our main focus is on the overall unemployment and employment rates, in the econometric analysis we also experiment with the associated variables for subgroups of the population, i.e. young people and, men and women. To ensure comparability of our results in all instances, we use the same data source, i.e. the ILO national estimates.
To study the effect of ECI on the various labour market outcomes we use two different datasets. The first one includes OECD countries only, while the second includes a total of 70 countries both developed and developing. The OECD sample provides a more reliable dataset regarding the set of control variables (which is discussed in Section 4.1 and presented in Table 1) and its availability over a longer period of time. Specifically, given the availability of controls, the OECD sample covers the period 1985-2008 and includes the following countries: Australia, Austria, Belgium, Canada, Chile, Czech Republic, Denmark, Estonia, Finland, France, Germany, Greece, Hungary, Ireland, Israel, Italy, Japan, Korea Rep., Latvia, Mexico, Netherlands, New Zealand, Norway, Poland, Portugal, Slovak Rep., Slovenia, Spain, Sweden, Switzerland, Turkey, United Kingdom, United States. It should be mentioned, however, that some variables, although having the same number of observations, differ in the country/year observations. In contrast, for non-OECD countries we have a limited set of control variables. In order to maximize the number of countries included in the sample we consider a simple cross-section, taking averages for the period 1990-2010 to the detriment of having temporal variations but for much fewer countries (mainly for the OECD countries as above). In other words, the benefit of having more countries in the sample and available controls is counterbalanced by the absence of time variation. However, we feel that this additional analysis gives valuable insights and further robustification of our results. 2 We also use freely available international trade data from the Massachusetts Institute of Technology (MIT)'s Observatory of Economic Complexity (http://atlas.media.mit.edu). We chose the SITC-4 rev.2 dataset, which provides the longest time series, combining information from a dataset compiled by Feenstra et al. (2005) for the years 1962-2000 and the UN Comtrade dataset from 2001 to 2008 (https://comtrade.un.org), and details about the products exported by every country.
We measure economic complexity using the improved ECI (ECI+). ECI+ measures the diversity and sophistication of a country's export structure corrected by how difficult it is to export each product. It combines information on the diversity of a country, i.e. the number of products it exports, and the ubiquity of its products, i.e. the number of countries that export these products (Hidalgo and Hausmann 2009). ECI+ is estimated from data connecting countries to the products they export and is freely available at MIT's Observatory of Economic Complexity. The index is calculated by applying the methodology described in Albeaik et al. (2017b) to the international trade data from the MIT's Observatory of Economic Complexity (a brief description of this methodology is discussed in Section 3). Albeaik et al. (2017b) show that ECI+ outperforms the original ECI in its ability to predict economic growth and in the consistency of its estimators across different econometric specifications. ECI+ captures information about an economy's level of development that is different from what is captured by, for example, gross domestic product (GDP) growth or GDP per capita. ECI+ incorporates the idea that institutions, knowledge and technology are prerequisites for economic growth but, in contrast to other indexes of growth, ECI+ is measured with simple linear algebra techniques that determine the knowledge intensity of economies endogenously from the countries' export data (Albeaik et al. 2017b). In a very recent working paper, Albeaik et al. (2017a) show that the definition of ECI+ is equivalent to the Fitness Complexity metric proposed by Tacchella et al. (2012).

Methods
To calculate the improved measure of economic complexity (ECI+) used in this work, we rely on the methodology described in Albeaik et al. (2017b). In short, let us assume that we have trade information for l number of countries and k products. We can calculate the total exports of a country corrected by how difficult it is to export each product using , and normalizing X c at each iteration step by its geometric mean: where [C] is the number of countries in the sample. We estimate ECI+ as the total exports of a country corrected by how difficult it is to export each product, minus the average proportion that the country represents in the total exports of a product (which accounts for the size of a country's export economy): Likewise, but putting the spotlight on products rather than on countries, the improved product complexity index (PCI+) is defined as the following iterative map: with the initial condition X 0 p = c X cp X 0 c being the average proportion of product p in country c. Again, normalizing at each step X p by its geometric mean: where [P] is the number of products in the sample, we define the product complexity index, corrected by how difficult it is to export each product, where X p is total world trade of product p.
To summarize, ECI+ and PCI+ denote, respectively, the total exports of a country, corrected by how difficult it is to export each product, and the total trade in a product, corrected by how easy it is to export that product (Albeaik et al. 2017b). For simplicity of notation, we will hereafter call these measures ECI and PCI respectively.

Regression analysis
We study the effect of economic complexity, measured by the ECI, on various labour market outcomes, using the datasets described in Section 2.
According to Hausmann, Hwang, and Rodrik (2007) higher economic complexity is associated with higher productivity. The ECI ranks traded goods in terms of their implied productivity. This signals an important source of endogeneity in the relationship considered and an obvious problem of reverse causality. Furthermore, given that ECI is an alternative measure of structural transformations, i.e. the reallocation of factors of production from traditional to modern activities, its simultaneity with labour market outcomes cannot be neglected. In order to mitigate the endogeneity of the independent variables, we follow a fixed effects, two-stage least squares/instrumental variables (FE 2SLS/IV) strategy for both datasets (OECD panel sample and cross-section world sample).
We regress the baseline specification described by the following equation: Here, labour market outcomes for country i in period t are expressed as a function of the ECI, a set of control variables, time d t and country g i fixed effects, and a stochastic term u i,t . 3 The main dependent variable in all regressions is the overall unemployment rate. To examine the robustness of our results and to generalize our findings, we also replicate our analysis for the unemployment and employment rates, among young people and men and women separately. We suggest the use of lagged and differenced values of the main independent variable (ECI) for up to 4 years as instruments. The received literature on IVs (Arellano and Bover 1995;Blundell and Bond 1998;Roodman 2009) shows that lagged differences of the suspected endogenous variable can serve as appropriate instruments, provided that they pass the tests for relevance, weakness and overidentifying restrictions (Feldmann 2013). Regarding relevance, we argue that changes in the ECI over the previous 4 years are likely to have a direct impact on the level of ECI in the current year, since technological changes in productive structure are rather cumulative: higher ECI today is more likely to result in higher ECI next year. Statistically speaking relevance is shown in our results through statistical significance of the IVs. Regarding excludability, we argue that lagged ECI affects labour market outcomes in the current year only through the current level of ECI. Weakness and overidentification are addressed and discussed in Section 4.2.
Although relevance is hard to fail above, excludability could still be of concern. Indeed former levels of ECI could have implications to other aspects of the economy, related to and still be relevant for present economic growth. We propose the inclusion of the lagged ECI variables in the main equation, in an attempt to check indeed whether they pose a threat to identification. However, statistical validity of the suggested IVs avails in the opposite direction.

OECD panel sample
To correctly specify our regression model we use two broad groups of control variables out of the full set, which is listed in Table 1. The first group includes macroeconomic controls, i.e. Inflation, Imports and Output Gap. The inflation rate controls for the standard Phillips curve relationship (Wyplosz 2001). The Output Gap controls for the business cycle, whereas imports as proportion of GDP (Imports) controls for the effect of international trade. The second group of controls aims to capture the effect of labour market institutions. 4 More specifically, we use the average tax wedge, denoted Tax Wedge (Daveri and Tabellini 2000), and variables that may encompass key elements of the wage bargaining system as in Aidt and Tzannatos (2008) (namely Union Density, Coverage, Centralization and Coordination). Last, Replacement serves to pick up the generosity of the unemployment benefits system (Scarpetta 1996;Lichter 2016). 5 Table 2 presents our main results. In all cases except column (11), we report the IV results. Our intuition regarding IV appropriateness due to the cumulative nature of the ECI, is verified by the first-stage results (reported in the Appendix), since in all cases, the coefficients of the lagged and differenced ECI have a positive sign and are statistically significant at the 1% level (in only three cases, for the third lag, they are significant at the 5% level). In addition, in order to take into account Angrist and Krueger (2001)'s caution against blindly using lags as instruments, we run the baseline model including also the four lagged variables of ECI as independent variables. Our results, which are available upon request, favor our argument above, by showing that the variables considered as instruments do not belong in the main model. Hence the use of lagged and differenced ECI per se should not raise concerns on excludability.
On the bottom of Table 2, and respectively in all included Tables of IV results, we present relevant statistics, lending confidence to our model and method. The F-test is the F-statistic on the significance of the model, exhibiting relevance for the included controls. DWH-test is the Durbin-WU-Hausman test of endogeneity of the regressors, the results to which are prompting to the need for an IV regression. Weak-id gives the F-statistic for weak identification, which in all cases is larger than the minimum rule of thumb of 10 for strong instruments, suggested by Staiger and Stock (1997). The LM-Underid gives the Kleibergen-Paap Wald test of weak identification, with the null hypothesishere rejectedindicating that the model is weakly identified. Last, the Hansen test shows the p-value of the Hansen test of overidentification. Our p-values are suggesting failure to reject the null, especially in columns 2, 3 and 4, where the values are closer to 1. However, it is suggestive that Hansen test p-values are viewed with caution, due to lack of theoretical justification of the exact thresholds over 0.1 and below 0.25 in the literature (Roodman 2009). Brief explanation of the presented statistics is given at the bottom of each table of results.
Column 9 of Table 2 presents our baseline specification, whereas columns (1)-(8) report results from our robustness checks regarding the inclusion of additional variables. Column (10) repeats the specification of column (9) using a different IV and is discussed below. Column (11) provides the OLS estimates of column (9) for reference. Across specifications, the coefficient of interest on ECI is statistically significant, at least at the 10% level. Most importantly, and in line with Autor and Salomons (2018) for OECD and Ghodsi et al. (2020) for emerging and transition economies, the estimated effect is negative, suggesting that economic complexity associates with lower unemployment. Our OLS specification already reveals a substantial effect: a one standard deviation increase in ECI is associated with a one standard deviation decrease in the unemployment rate, or equivalently a decrease in the unemployment rate of about six percentage points (column 11). Such an effect should not be considered alone, as if in a vacuum. ECI is highly persistent and of rather rigid nature. Single unit increases require major structural transformations in the economy and would be neither simple nor swift. Where they to materialize however they would be groundbreaking enough to guarantee a vast impact on unemployment.
With regard to the rest of our controls, our results can be summarized as follows: higher inflation exerts a negative effect on unemployment. This is consistent with the view that in the presence of downward nominal wage rigidity, inflation allows for better wage adjustment, resulting in lower equilibrium unemployment (Wyplosz 2001). Inflation is negative and statistically significant at the 1% level across specifications. Imports are negative and statistically significant across specifications, highlighting the beneficial role of incoming products for unemployment. The negative sign of Imports is consistent with the view that higher competition from abroad results into more Table 2. Fixed Effects 2SLS, OECD sample. (1) ( efficient allocation of resources and lower unemployment (Felbermayr, Prat, and Schmerer 2011). An additional interpretation of the negative coefficient on imports could be the following: imports here could be essentially controlling for the ease of trade, lack of trade barriers, tariffs, quotas, etc. This result holds only for the OECD sample, which is heavily populated and essentially dominated by the presence of EU member states. This interpretation can be strengthened further by the fact that Imports turn positive and nonsignificant in the world sample, which contains 70both developing and developedeconomies, where is tough to argue that EU is dominant. In line with economic theory, OutputGap is found to be negative and statistically significant across, implying that unemployment is higher in a recession. 6 The estimated coefficient implies that a one standard deviation increase on the output gap, i.e. the difference between actual and potential output, is associated with a 0.5% decline in the unemployment rate. Concerning union variables, only two coefficients exhibit significance. UnionDensity turns positive and significant to unemployment, while higher bargaining Coordination results into lower unemployment (see also Di Tella and MacCulloch 2005). Coordinated bargaining induces unions to internalize the positive effects of wage moderation on unemployment.
Similarly, following previous research, we find that higher unemployment benefits, i.e. higher Replacement, are associated with higher unemployment. This effect might be attributed, for example, to lower job search intensity by the unemployed when receiving higher unemployment benefits (Bassanini and Duval 2006;Wyplosz 2006). However, this result loses significance and flips sign, exhibiting a nonrobust behavior. Finally, TaxWedge has a positive and statistically significant effect on unemployment, verifying the disincentives created by the heavier burden of employee taxation.
To further convince the reader about our main finding on the importance of economic complexity of exports on unemployment outcomes, in column (10) we adopt an alternative instrument of ECI, namely the (log) number of journal articles published in scientific and technical journals in a given year. This index calculates the total number of papers in the fields of physics, biology, chemistry, mathematics, clinical medicine, biomedical research, engineering and technology, and earth and space sciences. Higher values are associated with higher scientific effort and output, which are directly related to the intensity of process and product innovation in the economy. Hence, we naturally expect articles to influence economic sophistication as measured by the ECI. It is reasonable to assume that new knowledge appearing in scientific articles is materialized in more sophisticated products. Regarding the exogeneity of the instrument it is plausible to assume that changes in the number of journal articles do not have a direct impact on labour market institutions and outcomes. The relevant statistics at the bottom of Table 2 (column (10)) and the first-stage results reported in Table A3, column (10), in the Appendix reveal a strong instrument. The statistically significant coefficient of ECI on unemployment is confirmed in the second stage results presented in Table 2, column (10).
In Table 3, column (1), we include the baseline results (column (9) of Table 2) for reference. 7 In the other columns, we estimate the same econometric specification, however, we vary the dependent variable in the following way: unemployment rate for individuals aged 15-24 (column 2); male unemployment rate (column 3); female unemployment rate (column 4). We also use employment measures on the left-hand side, namely: the male (column 5), the female (column 6) and the total employment rate (column 7). 8 In all cases, the results of the baseline model are verified. Economic complexity has a negative effect on all types of unemployment and consistently a positive impact on all measures of employment. Female employment is the only case that ECI loses statistical significance. The rest of our explanatory variables remain in line with our baseline results, with the exception of Replacement, which, in columns (2)-(4), is negatively signed and statistically significant in contrast to our theoretical priors.
In Table 4, we present robustness checks to our baseline specification (column 9, Table 2), which we display in column 1 for reference. Columns (2) -(6) show that our main finding on ECI is robust regardless of the choice of controls included in each specification. We hereby discuss only results which differ compared to the baseline model as well as certain additional controls examined. As previously discussed, the OECD sample is heavily dominated by EU member states, and therefore Imports could reflect the ease of trade among them. EU member states could be driving this Table 3. Fixed Effects 2SLS, Unemployment and Employment in Specific Groups. (1) ( result because Imports lose statistical significance when EPL, Regulation and ALMP come into the picture (columns 2, 4, and 5 respectively). EPL, Regulation and ALMP could reflect EU regulations and institutions, i.e. labour policies, that Imports might have captured alongside the trade effect. In addition, UnionDensity and Replacement lose significance in certain specifications. In column 2, we introduce EPL, capturing the overall degree of employment protection in the economy. Next, column (3), we control for the minimum wage level (MinWage), while in column (4) we introduce regulation intensity of the product markets in the economy (Regulation), with higher values indicating lower regulatory intensity. In column (5), we include spending on active labour market policies as a proportion of GDP (ALMP). Finally, in column (6) we include Education (gross enrolment ratio, tertiary, both sexes, %) to capture the effect of human capital on labour markets. All coefficients on the aforementioned controls, with the exception of ALMP turn out with the expected sign, while all except for MinWage and Education are statistically significant.
In Table 5, we exhibit supplementary regressions, which focus on the dynamics of the relationship under investigation. The underlying hypothesis is whether the effect of our explanatory variables comes with a 1-year lag. Column (1) presents the estimates of Equation (9), where all variables are introduced with a time lag, while in column (2) ECI is considered without a lag. An additional benefit of the time lag hypothesis is that it circumvents potential endogeneity between controls and the ECI. In columns (3) and (4), we also consider the dependent variable with a time lag. In column (3), we estimate a simple ordinary least squares (OLS) panel fixed effects model, whereas in column (4) we employ the Arellano-Bond estimator. In all cases, our main variable of interest remains statistically significant and its magnitude does not change much, at least in columns (1) to (3). Interestingly, the value of the coefficient of ECI in column (4) increases to −3.710. 9

Cross-section, world sample
To generalize our findings from the previous section to a wider set of countries, we focus our analysis on a global sample of 70 developed and developing countries. 10 To maximize the number of countries used in the regression, we employ only a subset of the controls used in the previous section empirical exercise and use averages from 1990 to 2010. Data definitions and summary statistics for the world sample are given in Table 6. Applying a fixed-effects 2SLS/IV regression in a cross-section setting requires a set of exogenous instruments. We experiment with three instruments of ECI and examine the robustness of our results using different subsets of these instruments. First, we employ again the measure of the (log) number of journal articles published in scientific and technical journals in a given year (denoted Articles).
The second instrument considered is an index of genetic diversity. Following the comparative development literature (Ashraf and Galor 2013b), genetic diversity, predominantly determined during the prehistoric 'out of Africa' migration of humans, explains modern ethnic diversity and economic prosperity. Following the relevant literature, 'higher diversity therefore enhances society's capability to integrate advanced and more efficient production methods, expanding the economy's production possibility frontier and conferring the benefits of improved productivity' (Ashraf and Galor 2013b, p. 3). Therefore, the proportion of ethnic diversity explained by prehistoric diversity is expected to be correlated with economic complexity, without having a direct effect on contemporary unemployment and employment rates.
The third instrument used is the Secular Values Index, which is a 12-item measure of the distance from 'sacred' sources of authority in each country (Welzel 2013). It is a continuous scale in the [0,1] range, where 0 (1) denotes the less (more) secular position. Countries that hold high beliefs in 'sacred' sources of authority are expected to be less modernized and less prone to innovation and adoption of sophisticated methods of production.
The top part of Table 7 shows the output of the first-stage regressions for the exogenous instruments of the ECI. 11 The results indicate that ECI is positively associated with both the amount of research undertaken in an economy (Articles) and the Secular Values Index. On the other hand, it seems that higher levels of economic complexity are associated with lower genetic diversity. The hypothesis of weak identification is rejected in all instances, since the value of the relevant test (F-statistic of the first-stage estimation: Weak-id), is well above 10.
The second-stage regression results verify the negative and statistically significant relationship between ECI and unemployment, while the effect is slightly smaller than for the OECD sample. With respect to the rest of the (second stage) results, the main conclusions drawn from the OECD sample remain qualitatively intact.
To examine the robustness of our results in columns (2) and (3), we experiment using different subsets of the instruments employed in the main specification (column 1). Once again, the association of ECI with unemployment is negative and statistically significant. Finally, in columns (5), (6) and (7) we examine the effect of economic complexity on youth, male and female unemployment rates, respectively. Qualitatively, the results are similar to the ones obtained from the OECD panel dataset. 12

Products complexity and the labour market
The ECI methodology provides a useful toolbox that allows us to compute indexes that quantify economic sophistication, for both countries and products. For example, using the same methodology that computes ECI, but placing the spotlight on products rather than on countries, we can calculate the PCI (see Section 3). This index quantifies the sophistication of each product according to the amount of knowledge/know-how involved in its production, reflected by the countries that export the product (Hausmann et al. 2014). In other words, when a product is located in the center of the product space i.e. in the core of the international trade network of products, it ranks higher in the PCI because its production requires more knowledge/know-how. Recently, Hartmann et al. (2017), using the ECI methodology, introduced a measure that associates products with income inequality and showed how the development of new products is associated with changes in income inequality. However, the labour market effects are key to understanding national income disparities, since income differences are, by definition, based on differences in the labour productivity and/or employment level, among other factors. Here, we introduce a measure that links a product to the average unemployment and employment rates of the countries that export it. In this way, we illustrate how labour markets are affected by the level of export's sophistication and we quantify the influence of countries' level of economic complexity on their labour markets' outcomes. Genetic Diversity The expected heterozygosity (genetic diversity) of a country's contemporary national population, as developed by Ashraf and Galor (2013b) and Ashraf and Galor (2013a  Following Hartmann et al. (2017), we define the Product Unemployment Index (PUI) (resp. Product Employment Index, PEI) as the average unemployment rate (resp. employment rate) faced by the countries that export the focal product, normalized by the importance of this product to the total exports of the countries that export it. More precisely, we decompose the relationship between economic complexity and unemployment and employment rates into individual economic sectors, by creating product-level estimators of these rates that are expected for the countries exporting a given product.

Product unemployment and employment indexes
Assuming that we have trade data for l countries and k products, we can fill the l × k matrix M so that its matrix element M cp = 1 if country c has Revealed Comparative Advantage (RCA) for product p and zero otherwise. 13 For our case, the international trade data from MIT's Observatory of Economic Complexity contains information for 33 OECD countries and 773 products from 1985 to 2008, classified in accordance with the SITC at the 4-digit level. A visualization of matrix M for this dataset, which is used to calculate the ECI and the PCI is shown in Figure 1.
Every product p generates some value for the country c that exports it. Therefore, for every product p we can calculate the fraction s cp : where X cp is the total export value of product p when exported by country c, while p ′ X cp ′ is the value of all exports of country c. If U c (resp. E c ) is the unemployment (resp. employment) rate of country c, we can calculate the PUI p and the PEI p for every product, as where N p = c M cp s cp is a normalization factor. Utilizing the information we have for the unemployment and employment rates for the OECD countries we are able to calculate the above indexes. It is important to highlight at this point that the two indices, PUI and PEI, cannot capture differences between sectors on the link between 'produced value' and 'labour' because the index PUI (resp. PEI) is computed as the average unemployment rate (resp. employment rate) faced by the countries that export the focal product, normalized by the importance of this product to the total exports of the countries that export it. The two indices are product-level estimators that decompose the relationship between unemployment/ employment rates and economic complexity into individual economic sectors. In other words, they are tools that illustrate how labour-market outcomes are associated with the countries' RCA in exporting sophisticated products. For every year in the period 1985-2008, we calculate all product-related indexes, i.e. PCI, PUI and PEI, and we obtain their mean value for each product. Table 8 lists the averages of PUI across the sample and across 4-digit SITC4 categories of the 2-digit SITC4 industries. Industries are sorted in order of increasing PUI. Table 8 reveals that the industry group with the lowest average proportion of the total unemployment rate is 'Telecommunications and sound-recording and reproducing apparatus and equipment'. Similarly, the more sophisticated industry/product categories appear to have the lowest PUI. At the other end of the spectrum, the 'Coal, coke and briquettes' industry has the highest PUI. As the reader can easily verify, primary sector industries (with low product-sophistication), appear to be associated with higher rates of unemployment. This is also implied by Table 9 which lists the five products with the highest and lowest PUI and PEI values during the period 1985-2008.
In addition, we test the existence of a bivariate relationship between PCI and PUI and PEI. Thus we calculate Pearson's correlation coefficient for both pairs, i.e. PUI against PCI and PEI against PCI. If a relation exists, it should allow us to derive expectations of whether or not the products' complexity can be associated with the unemployment and employment rates. In the case of PUI against PCI, the correlation coefficient is r = −0.10 with p-value = 0.0061, while for the case of PEI against PCI it is r = 0.14 with p-value = 0.0002. In Figure 2, we present the scatter plots of PUI and PEI against PCI for all 773 products in our dataset together with the fitted linear models. The slopes of the linear fits are the corresponding correlation coefficients.
The statistically significant negative (resp. positive) correlation between PUI (resp. PEI) and PCI indicates that the sophisticated products are associated with countries that bear relative low unemployment rates (resp. high employment rates). This adds to our previous discussion about economic complexity at the country level, as it allows us to understand which sets of products are leading to more employment and less unemployment based on their sophistication.
Concluding, in Table 10 we run panel regressions between PCI and PUI and PEI. The results show that the relationship between PCI and PUI and PEI is the outcome of the correlations between products rather than within products. In other words, lower unemployment (and higher employment) is associated with increases in product complexity between products. It seems that change in economic complexity within products has no effect on the labour market. This suggests that the negative (resp. positive) effect on unemployment (resp. employment) rate is due to changes in the structure of the product space towards the creation of more sophisticated products, rather than increases in the sophistication of existing ones.

Conclusion
Our analysis illustrates that the labour market performance of a country is highly predicted by the mix of products that a country produces and exports. Both in a panel and in a cross-country setting we have verified that there is a robust negative (resp. positive) relationship between unemployment (resp. employment) and product sophistication. Moreover, the relationship between these two variables is verified by instrumental variables (IV) estimation techniques. Hence, the evidence presented in this paper suggests that a country's level of economic sophistication determines its labour market outcomes. Oil-seeds and oleaginous fruits 7.65  In detail, countries that produce more sophisticated products generally have lower unemployment rates and higher employment rates. As higher sophistication of exported goods results in higher growth rate, there seems to be a capitalization effect at work (Pissarides 1992;Bean and Pissarides 1993): the present value for firms creating new jobs is higher when product sophistication increases, and, according to our estimates, this effect is not symmetrical across industries. We built the PUI and the PEI, which associate exported products with the average level of countries' employment and unemployment rates, respectively. With these indexes we show how the development of sophisticated products is associated with changes in the labour market. This result is important from a policy perspective. Using the proposed indexes, it is possible to design sectoral reallocation policies and smart specialization strategies that promote activities/sectors that are associated with lower unemployment and higher employment. Adding to the above, our analysis provides additional insights for the ex post policy evaluation process. As many tax and subsidy policies are associated with sectoral reallocation, our indexes can provide a quantitative measure of the average unemployment cost (or gain) due to the implemented policy.
In sum, this study examined labour market outcomes at the macroeconomic level, but went beyond the standard institutional and economic factors to explain unemployment. We identified economic complexity as an explanatory variable of the observed differences in labour market outcomes across countries. An interesting way to build further on this would be to identify the exact inclusive institutions and technological capabilities that can have a mitigating effect on 'technological unemployment'.
Our study does not come without limitations. One is related to the fact that economic complexity does not capture differences across sectors/industries, making it impossible to capture where the improvements happen and matter. Second, economic complexity focuses only on exported goods, but not all goods produced in the economy, possibly not reflecting accurately the nature of the productive structure. Third, we have not been able to point to an exogenous instrument as an alternative to the lags, offering further robustness and validity to our results. Fourth, we acknowledge that what we learn from the above analysis is that in economies which export more sophisticated products, lower unemployment/higher employment takes place. However, this might be a consequence not only of economic complexity, but of other driverslike e.g. the quality of human capital available in the country. Considering measures of quality-adjusted educational attainment that have been proposed in the growth literature (Wößmann 2003;Lee and Barro 2001) and exploring their effect on economic complexity offers an important avenue for future research.

Notes
1. See Figure 1 in Hidalgo et al. (2007) for the network representation of the product space for 775 SITC-4 product classes exported in the 1998-2000 period. 2. Data definitions and summary statistics for this dataset are given in Table 6. 3. The cross-section model does not include country and time fixed effects. 4. For variable definitions, data sources and summary statistics see Table 1. 5. To examine the robustness of our results we also introduce a series of additional variables that capture the strictness of government regulation in the labour market. Specifically, we employ an index of employment protection legislation (EPL), an index that measures the strictness of regulation in the economy (Regulation), the proportion of public expenditures on active labour market programmes (as a percentage of government spending) and the variable Min Wage, which measures the generosity of the minimum wage scheme. 6. To address possible concerns of collinearity between OutputGap and ALMP, in addition to Uniondensity, Centralization, Coordination, and UnionCoverage (or EPL and Minimum Wage when they are used), we provide a correlation matrix in the Appendix. We also rerun all specifications excluding some regressors or at least not considering them (only) altogether. For the union related variables, due to the larger figures in the correlation matrix, we also estimate the Variance Inflation Factor for the panel data. Test statistics are way below the rule of thumb of 10, which is the level above which mutlicollinearity might be a problem (Hair et al. 2010). Our results appear robust and do not hint or raise concerns of multicollinearity. 7. We have chosen to present this specification as our baseline model due to the larger sample incorporated. We verify our baseline results through the inclusion of other variables that offer important insights to our empirical exercise. Our results remain robust. 8. All first stage results are shown in the Appendix, Tables A3-A5. 9. We should note that the coefficient represents the short-run effect. To save space, the first-stage results for the independent variables are not included in the Table. 12. To test for possible differences between developed and developing countries we introduce a dummy for developing countries and interact it with the ECI. The resulting coefficient turns out insignificant, providing no evidence of a difference between developed and developing countries in that regard. 13. RCA is the ratio between the share of a given product in a country's exports and the share of this product in the total global exports (Balassa 1965). According to the World Bank: Measures of RCA have been used to help assess a country's export potential. The RCA indicates whether a country is in the process of extending the products in which it has a trade potential, as opposed to situations in which the number of products that can be competitively exported is static. Notes: t-statistics in parentheses. * p<0.10, ** p<0.05, *** p<0.01.  Table 3.