With respect to this aim of separate the role of demand for and supply of labor as drivers of sectoral reallocation, our work is, in fact, mostly related to Lee and Wolpin . Lee and Wolpin devised and structurally estimated a rich model to study the process of labor reallocation from manufacturing to services in the United States. We see our work as complementary to it, to the extent that we are interested in a similar question, but we tackle it from a different perspective. Specifically, our approach aims to impose the minimal structure to interpret the data, closer in spirit to the accounting literature. Relative to the three papers above, we also depart in extending our empirical analysis to as many countries as we could gather data for, rather than focusing only on the United States. Using multiple countries allows us to provide additional ways to identify the size of the mobility frictions, which is an important component of our analysis. More broadly, our work is related to a rich literature that studied the contribution of human capital to growth and development. This literature showed that the level of human capital is significantly correlated with consequent growth , Barro , Mankiw et al.. However, the effects of changes in human capital stocks have been much more elusive and Pritchett. Pritchett in particular, in a famous article, asked: “Where has all the education gone?”. In this respect, our work provides some encouraging answers: we show that growth of human capital stocks matters for explaining reallocation out of agriculture. Methodologically, we are in debt to the approach developed by the growth and developing accounting literature , Barro , Hsieh and Klenow , and more recently Gennaioli et al.. Relative to this literature, we show that observable variation across birth-cohorts can be used in an accounting framework,gallon pot and we introduced a way to measure the role of human capital without having to rely on prices.
In terms of the purely empirical contribution of this paper – that is, documenting novel cross country patterns in reallocation out of agriculture by cohort – our work relates to Kim and Topel , Lee and Wolpin , and Perez , which document sectorial reallocation by cohort but limit their focus to, respectively, South Korea and United States and Argentina; and especially to Hobijn et al. , which, in ongoing work, is also using the IPUMS dataset to document patterns on reallocation across sectors by cohorts, and use them to motivate a model linking demographic forces to structural change. Our model combines elements and insights already presents in Matsuyama , Lucas , and Herrendorf and Schoellman . To the best of our knowledge, we are the first to provide a tractable framework to analytically characterize reallocation within and across-cohorts in a context with general mobility frictions. Hsieh et al. also exploits year and cohort effects to calibrate a model of allocation of talent. It uses them to discipline the relative role, for the aggregate efficiency of the allocation of talent, of changes in frictions that affect human capital investment and frictions that distort the labor market. Relative to this paper, we focus on a simpler framework that allows us to consider fixed-cost-type frictions, which turn out to be crucial to correctly identify the role of human capital. Finally, our work relates to a growing literature that uses longitudinal wage data to reconsider the agricultural productivity gaps and that shows that these gaps are more consistent with sorting across-sectors than with large mobility frictions; , Herrendorf and Schoellman , and Hicks et al.. We contribute to this literature in two ways: we provide a model that highlights when wage data can be informative on frictions; and we show, without relying on wage data, additional evidence corroborating the sorting explanation and casting doubts on the presence of large mobility frictions.We next describe how we use data to quantify the role of human capital in labor reallocation out of agriculture. This is motivated by the results in Proposition 2, which provide an accounting framework to link within- and across-cohorts labor reallocation to the relative contribution of human capital vs. prices/productivity in reallocation out of agriculture. In this Section, we start by documenting a number of novel cross-country facts about reallocation by cohort using micro level data for a large set of countries.
Most of the cross-country evidence available to date only covers aggregate rates of reallocation. Our paper is among the first to provide micro level evidence on the behavior of different cohorts of workers in the process of structural transformation. We present the patterns descriptively to convey information on what the data say in a transparent format, focusing on the novelty of the findings rather than on their role in our approach. In Section 4 we will instead interpret the observed patterns through the lens of theory. There, we make explicit how Proposition 2 can be brought to the data to make inference on whether human capital matters for the movement of workers out of agriculture. Below we introduce the data and measurement approach, and then discuss the novel crosscountry findings on reallocation by cohort.We use micro level data from the Integrated Public Use Microdata Series 8. The data are either censuses or large samples from labor force surveys that are representative of the entire population. We include in our analysis all IPUMS countries for which we have available at least two ten-years apart repeated cross-sections with available information on age, gender, and working industry. This gives us a sample of fifty two countries covering about two thirds of the world population. For fifty one countries, the IPUMS data also include geographical information at the sub-national level which we use in our analysis as a source of additional variation.9 For twenty three countries, we observe four or more cross-sections, for seventeen we observe three or more. On average, we observe countries over a period of 28 years. For some countries, such as United States and Brazil, our data cover a long time span of half a century or more of labor reallocation. Table A.1 in the Appendix lists the countries in our sample, the income level of each country, in 2010, relative to the one of United States, the years of coverage, the agricultural employment shares, and the number of observed cross-sections. The countries in the sample comprise a wide range of income levels, from the United States to Liberia and El Salvador. Eight countries are high-income countries, twenty five are middle-income countries and the remaining nineteen are low-income10. Our sample also spans a large geographical area, covering Asia and Oceania , Africa , Central and South America , and Europe and North America . We focus on males and restrict our attention to those aged 25 to 59. This is meant to capture working age and identify the period after education investment is completed, which allows to consider human capital as constant.
We exclude women from the current analysis given the large cross-country differences in female labor force participation.In the model, we assumed that each cohort has equal size and that size is constant over the life-cycle. In the data, however, we observe that cohorts have different sizes,gallon nursery pot and that the size of a given cohort changes over time due to mortality. We may be concerned that these demographic compositional effects are relevant in explaining the cross-country variation in cohort and year effects. We here perform a series of exercise to show that, reassuringly, demographic composition does not mechanically drive our estimates. First, we compute the year effect weighting each cohort by its share in the population at time t + k rather than at time t. In Figure 8a, we show that this change does not make any difference. Second, we compute the year effect with the raw data – i.e. without smoothing the demographic distribution, which we did in the benchmark exercise to adjust for age-heaping. As is well know, Indian data suffer from extreme age-heaping. Consistent with this, we see in Figure 8b that only in India the year effect estimated with the raw data is different from the benchmark one. Third, we recompute the rate of labor reallocation out of agriculture keeping constant the demographic structure at time t – i.e. we compute LA,t+k weighting each cohort according to !t . Figure 8c, shows that the estimated rate of labor reallocation are almost identical to the benchmark ones. Finally – in Figure 8d – we perform the same exercise, but keeping constant the demographic structure at time t + k. Again, we conclude the demographic changes do not have relevant mechanical effects. In Figure A.3 in the appendix, we recompute the same exercise using sub-national units. We find similar results.This section introduces our empirical approach and discusses how we make use of the data and patterns described above to quantify, in an accounting sense, the relative contribution of human capital and prices/productivity to labor reallocation out of agriculture. Our main starting point is Proposition 2, which tells us that we can do so by leveraging within- and across-cohorts reallocation. Proposition 2 provides a mapping between two observable objects, year and cohort effects, and our two main objects of interest, the contribution of human capital and of prices/productivity to labor reallocation. The mapping, however, is made challenging by the possibility that labor mobility frictions bind, and by general equilibrium. The spirit of our empirical exercise in this section is to exploit – through multiple approaches – observable variation in year and cohort effects both across- and within-countries to discipline the size of the frictions. We then use further micro level data to calibrate the strength of the general equilibrium. We provide a range of estimates depending on parameter values, but the overarching conclusion, which we reach in Section 4.1, is that human capital explains approximately one third of total reallocation out of agriculture. In Section 4.2, we show that, at the same time, human capital has at most a minor role in explaining why some countries have faster reallocation rates than others.Next, we provide estimates for the size of the friction through a series of different exercises. The results of our estimates are summarized in Table 2. Conceptually, we follow two main alternative methods, each tied to a source of variation that can be exploited to back out the size of the frictions. Our first method builds on Proposition 3, which relates the size of labor mobility frictions to the reallocation rates of individuals of different ages. The fixed cost traps in agriculture individuals that would otherwise move to non-agriculture in a frictionless environment. This effect, however, is not symmetric across ages, and is instead stronger for older individuals: they benefit from future increases in non-agriculture wages for fewer years and hence, for given fixed cost, face a stronger constraint. This means that the presence of frictions causes old individuals – the constrained – to reallocate at a slower rate than young individuals – the unconstrained. Fact 3 in Section 3 showed that old and young individuals reallocate at similar rates, thus providing evidence against a sizable role for mobility frictions. The following Lemma shows that this intuition can be used to provide a direct estimate for the size of the frictions.In this section, we provide a proof of concept that it is possible to trigger reallocation out of agriculture through policies that successfully increase the educational attainment of the population. To do so, we identify the causal effect of schooling on labor reallocation out of agriculture in the context of a school construction program in Indonesia. Following the seminal work of Du- flo , we use the INPRES school construction program, which built 61,000 primary schools between 1974 and 1978, to provide quasi-experimental variation in schooling. While the intensity of the program, captured by the number of new schools per pupil, was not random, only somecohorts, those younger than 6 at the time the program started, were fully exposed to the program. Therefore, we can run a fairly standard difference-in-difference exercise: we compare cohorts fully exposed to the treatment to those not exposed to it, in districts with higher or lower treatment intensity.