Is Automation Labor-Displacing in the Developing Countries, Too?: Robots, Polarization, and Jobs∗

This paper uses global census data to examine whether the labor market polarization and labor-displacing automation documented in the advanced countries appears in the developing world. While confirming both effects for the former, it finds little evidence for either in developing countries. In particular,the critical category corresponding to manufacturing worker, operators and assemblers has increased in absolute terms and as a share of the labor force. The paper then uses data on robot usage to explore its impact on the relative employment evolution in each sample controlling for Chinese import penetration. Trade competition appears largely irrelevant in both cases. Robots, however, are displacing in the advanced countries, explaining 25-50 percent of the job loss in manufacturing. However, they likely crowd in operators and assemblers in developing countries. This is likely due to off-shoring that combines robots with new operators in FDI destination countries which may, for the present, offset any displacement effect. Some evidence is found, however, for incipient polarization in Mexico and Brazil.


Introduction
Advanced country labor markets have sharply polarized over the last two decades. For the US, Katz et al. (2006); Autor (2010); Autor and Dorn (2013) document expanding job opportunities in both high-skill, high-wage occupations and low-skill, low-wage occupations, coupled with contracting opportunities in middle-wage, middle-skill white collar and bluecollar jobs. Of particular interest, job opportunities are declining in middle-skill, blue-collar production, craft and operative occupations. Goos et al. (2014) document that this phenomenon has appeared in each of 16 European countries from 1993 to 2006. Even growth optimists, such as Brynjolfsson and McAfee (2014) predict major shifts in the composition of labor and the need for compensatory social policies to offset the resulting inequality.
Leading explanations include the ongoing automation and off-shoring of middle-skilled "routine" tasks that were formerly performed by workers with moderate education. Routine tasks as described by Autor et al. (2003) are sufficiently defined that they can be carried out by a computer executing a program or alternatively, by a comparatively less-educated worker in a developing country who carries out the task with minimal discretion, such as repetitive assembly tasks. 1 Generally, the literature has emphasized automation change over trade forces. Autor argues that the general wisdom by the end of the 1990s was that trade flows were simply too small to explain the vast changes in skill demands and wage structures and Acemoglu and Autor (2011) suggest this empirically as well. David et al. (2013) for instance, specifically measure the impact of the rise in China and find that, while not negligible, it accounts for only 25% of the fall in manufacturing employment in the US. Though, recent work (Acemoglu et al., 2016) suggests larger impacts than previously thought, a major focus remains on automation and, in particular, robots. automation eliminates routine manufacturing type jobs, or as it permits 'reshoring' tasks, we may see a short circuiting of the traditional forces generating the "flying geese" pattern where stages of the value chain are passed down from advancing to follower countries and it is unclear whether developing countries have the necessary complementary skills to attract the parts of the chain that still require workers. 4 Though Hallward-Driemeir and Nayyar (2019) find that increased automation in advanced countries in general has not led to a declining growth rate in outward oriented FDI, there is some incipient evidence for this effect. For instance, Faber (2018) and Artuc et al. (2019) find Mexican exports to the US declining with increased US robot use and a concomitant fall in manufacturing employment most susceptible to automation, although the latter finds no overall decline in manufacturing employment.
The present paper uses global census data to explore whether patterns of polarization are visible in the developing world and the role of automation, proxied by robot adoption, in driving the patterns in both groups of countries. Section 2 discusses why we might find differing patterns between the advanced and developing countries and Section 3 discusses the data sources. Section 4, broadly following Autor (2010), tracks job categories across time for the advanced countries and 21 developing countries in Africa, Latin America and Asia and confirms the polarization patterns for the former, but not the latter. Previous work, any other country by 2017. Part of this investment may reflect the dramatic fall in robot prices. The payback period for a welding robot in the Chinese automotive industry, for instance, dropped from 5.3 years to 1.7 years between 2010 and 2015, and by 2017 was forecast to shrink to just 1.3 years. However, in addition, both the Chinese and Korean governments now subsidize the introduction of robots. See http://www.bloombergview.com/articles/2015-04-09/robots-leave-behind-chinese-workers. 4 While China has complemented this trend with investment in training for more complex jobs, recent college graduates report having problems finding employment and 43% consider themselves over-educated for their positions, much as Beaudry et al. (2013) suggest is happening in the US. "That might not be a problem if the Chinese economy were generating plenty of higher-skill jobs for more educated workers. The solution, then, would simply be to offer more training and education to displaced blue-collar workers. The reality, however, is that China has struggled to create enough white-collar jobs for its soaring population of college graduates. In mid-2013, the Chinese government revealed that only about half of the countrys current crop of college graduates had been able to find jobs, while more than 20 percent of the previous years graduates remained unemployed. According to one analysis, fully 43 percent of Chinese workers already consider themselves to be over educated for their current positions. As software automation and artificial intelligence increasingly affect knowledge-based occupations, especially at the entry level, it may well become even more difficult for the Chinese economy to absorb workers who seek to climb the skills ladder". See http://www.nytimes.com/2015/06/11/opinion/chinas-troubling-robot-revolution.html. broadly following Goos et al. (2014), WorldBank (2016) and using ILO Kilm data finds evidence that middle skilled occupations intensive in routine cognitive and manual skills have also decreased across the developing world as a share of the workforce with the exception of China, Ethiopia, Argentina and Nicaragua. Our picture is more mixed, offering less evidence for polarization, either in absolute levels of employment or share of the workforce with the exception, in the middle skilled category, of crafts and related occupations. Manufacturing jobs, captured in the Operators and Assembler category (from here on, OA) in fact, expand in both levels and shares. Section 5 uses data on robot stocks to confirm that robot adoption indeed displaces manufacturing jobs in the advanced countries, but in the developing world, they seem to be complementary.
2 Should we expect to see polarization and labordisplacement in developing countries as well?
The way in which off-shoring and automation technologies play out in developing economies may differ from their advanced counterparts for several reasons: Differing initial occupational distributions: Potential polarization dynamics are layered on very different initial occupational structures and positions in the demographic transition.
Most mechanically, in many developing countries the sector of middle income workers engaged in codified tasks is small in the first place-in Ghana, for instance, 90% of the workforce is informal and engaged in low skilled services and artisanal production (see, for example Falco et al. (2015)) and this is representative of many low-income countries. Hence, we would expect to see little in the way of displacement of these types of jobs.
More limited feasibility of automation? The degree to which automation is adopted depends heavily on a country's technological absorptive capacity, the skill of the workforce, ability to mobilize resources for large capital investments, capacity for maintenance, and attention to tolerances which may make it less easy to substitute away from labor in many poorer countries. Such factors contribute the the slower rate of technological diffusion, including robot use, in general to developing countries (Comin and Mestieri, 2018).
Recipients of off-shored jobs: Off-shored jobs from advanced countries are precisely moving to developing countries and hence we would expect to see a complementary expansion of the middle-a "de-polarization" of the wage distribution in at least some host countries.
Since multinational assembly operations will often included state of the art plants, including robots, it is possible that we may see a positive comovement of robots and manufacturing employment. That said, to the degree that newer arrivals to off-shoring, such as China or Vietnam, compete with established destinations such as Mexico, the net effect of diversion vs. increased total off-shoring is unclear. Hanson and Robertson (2008)   Rather than using the occisco variable, Autor (2010) and Autor and Dorn (2013) map these 4-digit categories into a distinct set of skill sets listed in Figure A-1 to better capture "routine" tasks in the US. Hence, in the original ISCO categorization, operators of machines in manufacturing appear in "Plant and Machine Operators, and Assemblers" (category 8) but manufacturing workers who don't operate machinery appear in " elementary occupations" (category 9). Both may be more routine than, for instance, food preparation or personal care, also found in category 9, which require potentially less skill, but which are also less easy to automate.
Using the occisco variable allow us to work with numerous countries with varying degrees of disaggregation and sometimes inconsistent or ambiguous categorizations across time that IPUMS has standardized into uniform categories. As we show below, for the US, the conclusions under both methodologies does not change appreciably. Robots: Data on robots are collected by the International Federation of Robotics (IFR) whose statistical department is the primary global resource on robot installation. They are collected from nearly all industrial robot suppliers worldwide and supplemented with information from several national robot associations by type, country, industry and application. The industrial robot is defined as an "automatically controlled, re-programmable multipurpose manipulator programmable in three or more axes" and a service robot as one "that performs useful tasks for humans or equipment excluding industrial automation applications." The service life is estimated at 12 years and hence, assuming immediate withdrawl thereafter, the reported stocks are the sum of installations over that period.
Import Competition: Import competition is the other hypothesized driver of operator displacement in the literature. As a proxy, we employ imports from China as a fraction of domestic output, GDP and imports.
Other controls: All regressions include a time trend to capture other trending unobserved factors over the same period. We further allow for differential trends by also including time interacted with pre-sample (1980s and 1990s) averages of population, openness as measured by (X+M)/GDP as from World Bank Development Statistics and import competition defined as imports from China as a share of domestic production. Table 1 reports descriptive statistics. Figure  To illustrate these aggregate tendencies, Figure  Tables 2 and 3 confirm these visible trends with the full panel of our countries. Specifically, we estimate the equations:

Polarization
where L it is the log-level (or the share) of each of the major categories in the International any differences are statistically significant. The time dummies capture differential changes by job category after the break point between advanced and developing country groups relative to the pre-breakpoint period. 7 We report cluster standard errors at the country level (see Bertrand et al. (2004)). Table 2 presents the results for the log of absolute employment as the dependent variable and Panel B, the share of employment, each by category. The presence of country fixed effects means that the dummies are measuring the average of country log level changes in employment (shares) by group relative to their pre-2000 levels and not relative to some third category. These broadly approximate the growth rate of the second period relative to the first.

Panel A in
Several regularities merit note. First, in absolute numbers, the Technicians, Professionals and Legislators categories are growing at similar rates in both the advanced and developing countries. However, as a share of the market, they are growing much faster in the advanced countries.
Second, the OA and Crafts categories in the advanced countries are stagnant, with growth rates insignificantly different from zero. However, in developing countries, these categories are expanding especially OA which is the third fastest growing. As a share of the workforce (panel B) OA is increasing almost as much as professionals. Craft workers are, however, decreasing and that may yield some ambiguity about the trends in the middle segment measured as shares found in WorldBank (2016).
Together, these distinct relative movements of the middle and upper segments of the market in absolutes and shares lead to the polarization found in the advanced countries. However, the expansion of the OA category more or less at pace with the technicians, professionals and legislators and managers category dampens the polarizing dynamic in developing countries.
In the bottom segment, the contribution to polarization is more ambiguous. The advanced countries see rates of growth in, for instance Services and Sales growing relatively quickly Developing countries see average growth for Elementary Occupations and high growth in Services and Sales that increases shares of the latter importantly. This is partly counterbalanced by a more rapid loss in share by skilled Agricultural and Fishery workers by a dramatic 11 percent relative to 5 percent in the advanced countries.
Annex tables 2 and 3 estimate equation 2 and explicitly test for differing evolution of each job category across the advanced and developing country samples by including an interactive variable for developing countries. In sum, in the advanced countries, we do see stagnation in the categories associated with the displacement of codifiable tasks and in particular in the operators and assemblers category, mainly relative to the surge in higher end employment. However, in developing countries, the picture is more ambiguous. In the middle segments, the Crafts segment has continued to grow, but at rate leading to a relative decline. However, the critical OA category continues to grow at rates similar to the professional categories and gains share of the labor force.
Since the OA category is of such import in the polarization and narrative and policy debate, and because it behaves so distinctly across advanced and developing countries, the rest of the paper will focus on the drivers of its evolution over the two sets of countries.

Robots
What drives these differences between the experience of developing and advanced countries?
Again, the literature highlights trade competition and automation as the prime suspects. In this section we focus primarily on the latter as proxied by the arrival of robots but controlling where possible for increased trade competition as proxied by Chinese import penetration.
The solid line in Figure 3 shows the total global robot stock as aggregated by IFR and documents a dramatic increase in robots over recent decades. There were 3,000 indus- Information at the country level is available since 1993 and for our core sample, we use the data as tabulated. However, we are also able to expand the sample with the assumptions that 1. before 1965, the global stock of robots was zero so we impute that to all countries; 2. if a country shows zero robots in 1993, we assume that was the case from 1965-1993 and 3. for countries with a strictly positive amount of robots in 1993, a back-cast regressing log (1+robots) on a polynomial of degree five in time offers a reasonable approximation to unobserved values. As a rough test of the reasonableness of these assumptions, Figure 3 plots the aggregate of our imputations and show it tracks the IFR aggregate extremely well except for the brief 1990-91 period. 8 To explore the robustness of this initial picture, we estimate where, again, L is either the log of operators and assemblers or the share, and robots is the log of the stock.  only an indicator variable on DCs to test the significance of this difference when linear and quadratic trends are included respectively, and shows that, in fact, the difference between the two samples is strongly significant and of large magnitude leaving the impact of robots in developing countries for both employment and share being positive (1, 2) or weakly negative (3). Columns 4-6 introduce year dummies which, while offering the most flexible form of control for other time varying factors, only permit estimating the differential effect and show again, that effect to be strongly significant.
In Columns 5-6, the sample is given by the countries that have observable information in pre-determined controls (variable trends generated by interacting pre-2000 averages of population, trade openness, and penetration of Chinese imports with a time trend). The inclusion of this new trends reduce the coefficient in less than 5%. Parallel with the employment trends documented above, there is a large negative effect of robots for the advanced countries that is robust to the inclusion of a variety of controls for other trending factors, 14 that is not shared by the developing world. Table 5 estimates Equation 4 which exploits individual country variation in robots stocks.
The cross country variation also allows us to explicitly test for the impact of trade competition as well. As in table 4, Column 1 establishes a negative significant impact on OA level for the whole global sample, although no significant impact in shares. Column 2 adds the interactive variable for LDCs that, again, allows us to reveal the heterogeneity in the sample: there has been a significant decrease in both number and share of the workforce in operators as a result of national robot adoption in the advanced countries. However, the interactive variable suggests, again, a significant difference with the advanced countries leaving the compound coefficient being positive or close to zero. In column 2, the P values of the total effect suggest that for the level of employment, this effect is not significant although it is for the share at the 10 percent. Column 3 includes the varying trends and preserves the previous results in levels although the negative effect for advanced countries disappears for shares.
Further, though the difference between the two samples remains strongly significant, the P value on the compound effect in developing countries is now insignificant.
Column 4 shows no remotely significant impact of trade competition for either sample.
The negative effect of robots on the advanced countries remains strong although the interactive effect for the developing countries is now only significant at the 10 percent level for both levels and shares and, again, the P values suggest that the compound effects is insignificantly different from zero. These results hold for whatever normalization of Chinese imports we employ: imports/GDP, imports/population, and imports over total imports.
Column 5 replicates these last two specifications but with the extended (40 percent larger). Again, import competition is not remotely significant. What reemerges is the strong negative effect for the advanced countries and a significant difference with the developing countries. For the latter, in the levels specification, the compound effect is insignificantly different from zero. However, in the share specification, it is strongly positive and significant at the 10 percent level without the China trade proxy, and the 5 percent level with.
The magnitudes of these effects are large and similar to those of previous studies. As a back of the envelope calculations, Autor (2010) Table 5 implies a reduction in operators of 11.2 percent or roughly half of the variation of the fall in OA in that period. Repeating the exercise for the previous decade roughly halves the effect.
With all relevant caveats, these are of the magnitude found by Dauth et al. (2017) (2015), all of which are bases for foreign assembly of cars and electronic devices. In a sense, then the positive impact of robots, and the lack of polarization, is importantly driven by outsourcing.

Conclusion
This paper has used global census data to explore to what degree findings of polarization in the advanced world can be found in the developing world and how much labor displacement by automation drives these patterns. We confirm previous findings of polarization for the advanced world. However, despite evidence that similar dynamics may be at work in China and Mexico, we find only limited evidence for polarization in developing countries. The key category-machine operators and assemblers-does not show absolute or relative decrease in most developing countries across the last decades. In fact, they show relatively strong growth leading to an increase in share of employment.
We then explore the causes of these differential effects, focusing on the growth in the use of robots and controlling for the impact of trade, in particular Chinese import penetration. We find little impact of trade competition on either advanced or developing countries.   1968−1975 1975−1982 1985−1990 1990−1999 1999−2006       S a l e s T e c h n i c i a n s P r o f e s s i o n a l s M a n a g e r s 1979-1989 1989-1999 1999-2007 2007-2009 Note: (Autor, 2010)