Re-examining Premature Deindustrialization
Sybren Deuzeman
s2203642
August 24, 2018
Abstract
Contents
1 Introduction 1
2 Literature Review 3
2.1 Structural Change and the Role of Manufacturing in Development . . . 3
2.2 Premature Deindustrialization . . . 5
2.3 Industries without Smokestacks . . . 7
3 Methodology 9 3.1 Dependent Variables . . . 9
3.2 Downward shift over Time in the Cross-Country Relationship between the Share of Industry and GDP per capita . . . 10
3.3 Change GDP per capita of Peak Industrialization . . . 10
3.4 Change GDP per capita of Peak Industrialization using a Cubic Relationship . . . . 11
4 Data 11 4.1 10-sector Database . . . 12
4.2 Maddison Database . . . 13
4.3 Modern Agriculture Employment Data . . . 13
4.4 Modern Agriculture Value Added Data . . . 14
5 Results 15 5.1 Downward Shift over Time . . . 15
5.1.1 Broad Modern Industries . . . 17
5.2 Change GDP per capita of Peak Industrialization . . . 20
5.2.1 Broad Modern Industries . . . 20
5.3 Change GDP per capita of Peak Industrialization using a Cubic Relationship . . . . 23
5.3.1 Broad Modern Industries . . . 26
6 Conclusion 27
A Detailed Data Description 31
Re-examining Premature Deindustrialization
Sybren Deuzeman
s2203642
August 24, 2018
1
Introduction
Historically, manufacturing has been one of the most important drivers for economic progress of nations. The industrial revolution caused a long period of sustained economic growth in Europe and the United States in the 19th century. Later, non-Western countries could catch-up through expanding their manufacturing sectors. Except some very resource rich countries, no country has been able to show a long sustained period of economic growth without it being driven by the manufacturing sector (Rodrik, 2014). Manufacturing is able to absorb large quantities of low-skilled labor into higher productivity activities, manufacturing goods are tradable which means that in an early stage already the larger global market and it is a dynamic sector that can drive future growth. Many countries still mired in poverty are looking towards manufacturing for their way toward prosperity.
Typically, manufacturing has followed a hump-shaped pattern over the course of economic de-velopment of which a stylized depiction is shown in figure 1. At low levels of GDP per capita, the manufacturing sector is very small. When GDP per capita rises, the manufacturing sector becomes larger and larger drawing labor from the agricultural sector. This is a process of industrialization. At some point in the economic development, however, the agricultural sector is so small that no more labor can be drawn from it. The service sector, however, continues to get larger which means that a process of deindustrialization starts and the manufacturing sector becomes smaller,
The importance of the manufacturing sector makes the findings by Rodrik (2016) on the change of this typical pattern of industrialization and deindustrialization worrisome. Using cross-country growth regressions, Rodrik (2016) shows that over the last couple of decades the relationship between the share of manufacturing and GDP per capita has shifted down and that the GDP per capita at which the share of manufacturing peaks has become lower. These results together are called premature deindustrialization. We show these results graphically in figure 1 and see that it entails a contraction towards the origin of the cross-country relationship between the share of manufacturing and GDP per capita. This means that developing countries nowadays will both have a smaller manufacturing sector when industrialization peaks and that this peak is also reached at an earlier point in its economic development. Therefore developing countries nowadays may need to discover new growth models that they might not be able to find.
Such a new growth model may be found in business services and modern agriculture, which is sometimes revert to as industries without smokestacks (Page, 2012). These sectors are both dynamic sectors with high productivity. An example for business services is the Philippines. The Philippines is home to the biggest call center industry in the world and Manila wakes during American business hours to answer calls from American customers. This provides better paying jobs and economic opportunities for many Filipinos1. Because of improvements in information and communication technologies, many business service tasks that previously had to be done close to the consumer, can now be done from a long distance. This provides opportunities for developing countries.
GDP per capita Share of man ufacturing emplo ymen t
Figure 1: The hump shape of the share of manufacturing employment and premature Deindustrial-ization: a downward shift of the peak and a peak at a lower GDP per capita
Another example of an industry without smokestacks is modern agriculture. Within agriculture, there are a wide range of ways to produce output. In the poorest countries agriculture is labor intensive and mainly produced for own consumption. In rich countries, agriculture is very capital intensive and advanced technology is used. The shift within agriculture based mainly on subsistence production towards modern production for markets should not be overlooked. Agricultural employ-ment for rich export markets can provide employemploy-ment and economic activity that shows many of the characteristics of manufacturing with regard to quality control and linkages with other sectors like the transport sector.
We will answer in this paper whether we still find the results by Rodrik (2016) on premature deindustrialization hold using the newest Maddison database. These results are that there is a downward shift in the relation between the share of manufacturing and GDP per capita and that the GDP per capita at which the share of manufacturing peaks is lower after 1990. We will discuss what the exact magnitudes and timing of these downward shifts over time tell us about the likely causes of these downward shifts. Next, we will answer whether these results still hold for broad modern industries, which we define as an aggregate sector of manufacturing, business services and modern agriculture.
To answer these questions, we follow closely methods used by Rodrik (2016) which we will describe in section 3. We use an additional method to estimate at which GDP per capita the share of manufacturing or broad modern industries peaks. Instead of a second degree polynomial, we will also use a third degree polynomial. A second degree polynomial imposes symmetry on the upward and downward sloping part of the hump shaped figure shown in figure 1. This means that the estimated form of this hump shape is a compromise between both the upward and downward sloping part. Using a third degree polynomial, we no longer impose this, which means that we are better able to find the exact form of the hump shape.
We confirm the finding by Rodrik (2016) that there is a downward shift over time in the cross-sectional relationship between the share of manufacturing and GDP per capita. We contribute these downward shifts over time to faster technological change through technology transfer and to a smaller degree to a shift in demand away from manufacturing. Using a third degree polynomial, we find that the GDP per capita at which the share of manufacturing employment peaks is lower after 1990, but this decrease is much smaller than estimated using a second degree polynomial. In contrast, we find that the GDP per capita at which the share of manufacturing nominal and real value added peaks has not changed.
We find that including business services mediates the downward shifts in manufacturing. The downward shift over time in the cross-sectional relationship between the share of employment and GDP per capita is smaller, while for the share of nominal and real value added these downward shifts disappear. Business services are able to replace only part of the employment lost in manufacturing, but the business services sector is more productive so it is able to replace the loss in nominal and real value added. Including modern agriculture, on the other hand, sharpens the downward shift found in the relationship between the share of employment and GDP per capita. If we broaden our concept of industry with modern agriculture, we observe a much larger downward shift in the relationship between the share of employment and GDP per capita, although we do not observe the same downward shift for nominal value added which may indicate that this sector has become more capital intensive.
This thesis adds to the literature in two ways. It is the first examination we know of on a cross-country basis whether industries without smokestacks replace the role of manufacturing in the process of economic development. Second, it introduces a new data set on modern agricultural employment and nominal value added, which can subsequently be expanded and used in other research.
The remainder of this thesis is structured as follows: in section 2 we will review the literature, in section 3 we discuss the methodology, in section 4 we discuss the data used, in section 5 we discuss the results and section 6 we offer some concluding remarks.
2
Literature Review
This thesis is related to two strands of literature, namely the literature on premature deindustri-alization and the literature on “industries without smokestacks”. Both are embedded within the literature on structural strange. We discuss the literature on structural change and the importance of manufacturing in this in subsection 2.1. We discuss the literature on premature deindustrializa-tion in subsecdeindustrializa-tion 2.2. We discuss the literature on industries without smokestacks in subsecdeindustrializa-tion 2.3.
2.1
Structural Change and the Role of Manufacturing in Development
In developing countries, the differences in sectoral output per worker are large (McMillan et al., 2014). The output per worker is generally lowest for agriculture and personal services and highest for the very capital intensive mining and public utilities sectors. There is a long tradition in economics that sees the movement of employment from the low productivity traditional sectors to the highly productive modern sectors, so called structural change, as a very important channel for growth (Kuznets, 1955; Lewis, 1954). McMillan et al. (2014) have estimated that if the economic structure of some African countries would match those of rich countries, while the labor productivity in all sectors remain the same, the economy-wide labor productivity could increase by more than 500% for Ethiopia and Malawi or even more than 1000% for Senegal. This shows us that the structural composition of an economy is important for the economy-wide labor productivity of a country and that structural change has an important role in the economic development of a country.
the share of employment becomes so small it can no longer decrease. The services sectors are small at low levels of GDP per capita and as GDP per capita increases, the services sectors become very large. For manufacturing we observe an inverted-U pattern as shown in figure 1. At low GDP per capita the manufacturing sector is small. When GDP per capita increases, the manufacturing sector becomes larger up until a certain point at which the manufacturing sector starts becomes smaller again (Herrendorf et al., 2014).
This pattern is explained in the literature mainly in two different ways. One explanation is based on demand-side factors. As income increases, demand shifts from agricultural to service products and the expenditure shares of agriculture decrease and of services increase (Kongsamut et al., 2001). An explanation based on demand can also involve trade and competitive advantage. As GDP per capita increases, the wages of the people living in the country increase as well. This means that these countries will loose their comparative advantage in labor intensive manufacturing. The internal demand for manufacturing products is now served by imports and external demand is no longer served by exports from the now high wage country.
Another explanation points towards different rates of technological change between sectors. If substitutability in consumption for the products between sectors is low, employment will move from sectors with high total factor productivity (TFP) growth towards sectors with low TFP growth (Ngai and Pissarides, 2007). Initially, higher productivity growth in agriculture than in manufacturing drives a movement of workers towards manufacturing and thus industrialization. Later, higher productivity growth in manufacturing than in services drives a movement of workers towards services and thus deindustrialization. These mechanisms are so called push mechanisms. Another model by Hansen and Prescott (2002) points towards pull factors in manufacturing. In this model, there are two different technologies, a traditional and a modern technology. The modern technology needs to get attractive enough before this technology is used. In this case, industrialization is led by technological change in the modern sector.
An increase of the manufacturing sector is seen as good for economic development of developing countries. This has several reasons. In developing countries the labor productivity in manufacturing is generally much higher than in agriculture and most services. Although the labor productivity in, for example, the mining sector is higher, this sector can not provide a lot of employment. Manufac-turing can absorb a large number of low skilled workers, exactly those workers developing countries have most of (Rodrik, 2016). This means that movement of labor from agriculture to manufacturing is able to cause structural change in developing countries. This argument can also be made dynam-ically, namely that manufacturing also has a higher productivity growth than agriculture. Before 1973 this indeed seems to be true for developing countries, however Szirmai (2012) shows that after 1973 productivity growth in real terms was generally higher in agriculture.
Closely related to this is the finding that manufacturing is the only sector that exhibits “uncon-ditional convergence” (Rodrik, 2013). The growth in productivity of manufacturing sectors is higher if productivity in that sector is initially low compared to the same sectors in other countries, which means that the productivity converges between the same sectors in other countries. This result is unconditional, in contrast to conditional GDP per capita convergence found using similar methods by Mankiw et al. (1992). To explain this, Rodrik (2013) points towards tradability of manufacturing goods and that manufacturing “can be rapidly integrated into global production networks, facilitat-ing technology transfer and absorption” (Rodrik, 2013, p.201-202). Thus competitive threat from abroad makes that industries needs to become more productive.
Another reason why manufacturing is seen as important is because it gives opportunities for capital accumulation. In developing countries, the capital intensity of the manufacturing sector is generally much higher than that in the agricultural sector, although this result is reversed in the developed economies. Especially in developing countries, where there is little physical capital, capital accumulation might be very important, because these countries start with very little capital per worker (Szirmai, 2012).
for learning increases if an economy is involved in certain activities, since learning “only takes place during activity” (Arrow, 1962, p.155) and the scope for learning is (historically) largest in the manufacturing sector. Related to these reason are the strong linkages and spillover effects from manufacturing. Growth in the manufacturing sector induces growth in other sectors as well, either because it uses inputs from other sectors or its products can be used as inputs in other sectors (Szirmai, 2012).
A very important reason why the manufacturing sector is seen as so important is that the output is tradable Rodrik (2016). Export markets provide larger markets than the home market alone. In many developing countries, the home market is still small, since consumers in those countries are poor. In these cases it helps if demand for products comes from outside the home market. In turn, this provides jobs with higher wages that are necessary to create larger internal demand. This means that manufacturing can solve the chicken-egg problem of market creation, where supply is needed for demand, but demand is also needed for supply. Tradability and the accompanying competition on exports markets is also at the core of unconditional convergence in manufacturing. Furthermore, firms that operate in markets with small internal demand can exploit economies of scale within manufacturing by trading on the world market.
2.2
Premature Deindustrialization
This importance of manufacturing in the process of growth makes results by Rodrik (2016) worrisome that the relationship between the share of manufacturing in employment, value added and real value added and GDP per capita has shifted downwards over time. These shifts are largest for employment and nominal value added and relatively small for real value added. Besides a downward shift in the relationship between the share of manufacturing and GDP per capita, Rodrik (2016) also finds that the GDP per capita at which the share of manufacturing peaks is lower after 1990 than before 1990. Hence, the size of the so important manufacturing sector is smaller throughout the economic development of developing countries nowadays than it was historically and the manufacturing sector starts to decline is at a much earlier stage in its economic development.
Rodrik (2016) is not so much concerned that there is a lack of economic growth in the developing world, but that the growth in many developing countries is not through the traditional channel of industrialization. Many growth booms are driven by capital inflows, external transfers, or commodity booms. If growth booms are not fueled by a large manufacturing sector, one can question whether these growth booms are sustainable once the reason for a growth boom disappear. Without the dynamics that the manufacturing sector creates that sustains growth for a longer period, the growth booms are not sustainable.
Rodrik (2016) finds that the downward shift in the share of manufacturing and GDP per capita is most profound in those countries without a comparative advantage in manufacturing, especially in Latin America and Sub-Saharan Africa. In Asia we do not observe a downward shift in the share of manufacturing and GDP per capita. To explain this regional pattern, Rodrik (2016) points towards a new explanation why deindustrialization happens, namely increased competition on global market, such that countries without a comparative advantage in manufacturing can no longer compete.
To show how demand, technology and trade shape the size of the manufacturing sector, Rodrik (2016) develops a simple model in which demand shifts, technology and trade forces interact. He shows the model both for a large economy in which prices are set by that country and a small, open economy in which the relative price of manufacturing goods is set on the world stage. He shows that faster technological change in manufacturing in the large economy leads to employment deindustrialization, while the share of real value added increases if demand substitution between manufacturing goods and non-manufacturing goods is low enough. The technological change causes the relative prices to decrease, which means that more manufacturing products are bought. However at the same time less employees are needed to fulfill the demand and the higher demand for man-ufacturing goods cannot completely mitigate this. A change in the export position of the country, which in this model is formally analogous to an adverse demand shock for manufactures, reduces both employment in manufacturing and real value added in manufacturing.
The relative price of the traded manufactured good is set on the global market, instead of the home market. If the relative price of manufactured goods is set on the global market, a technology shock actually means that there is an expansion of the share of manufacturing employment, since the manufacturing sector is now better able to produce manufacturing goods and provide these to the world market. However, if there is an external price shock, i.e. other countries are able to produce cheaper, one will observe that both the manufacturing employment share and real value added share decline.
The paper by Matsuyama (2009) makes a similar point more subtle: productivity growth in manufacturing does not necessarily lead to employment deindustrialization. In his model, two effects are competing: the income effect that an increase of productivity leads to lower employment in a sector and the trade effect that larger productivity growth in manufacturing than other countries give a larger comparative advantage and thus larger demand. Whether the share of manufacturing employment rises or falls depends on which of the two effects is stronger. For the foreign country, the effect however is clear, namely its share of manufacturing employment falls.
There are some important historical events that can be used as potential explanations of these findings. An explanation that Rodrik (2016) seems to implicitly use, is the paradigmatic change within governments and international organizations seeking economic development towards the so called Washington Consensus in the early 1980s. Broadly speaking, this paradigm recommends that governments should pursue macroeconomic stability, seek to open the economy up for world markets and also open up for capital and liberalize domestic markets (Gore, 2000). This followed a period in which many countries after independence tried to industrialize through large state interventions in which markets were protected from global market forces and thus are less able to develop.
Another strand of literature focuses more on the increased fragmentation of global value chains (Timmer et al., 2014). Firms in richer countries seek to reduce costs by outsourcing parts of their production processes to countries with lower wages. Baldwin (2013b, p.208) argues that since 1985 “most successful developing- nation industrializers joined the supply chains of firms from high- tech nations, especially the manufacturing giants of the 1980s - the United States, Japan, and Germany”. Joining a supply chain means that countries do not need to build up all the capabilities necessary to produce an end-product before it can start exporting. He points towards ICT technologies that lowered the costs of managing supply chains that are geographically dispersed as a reason. The technology and capital are provided by a firm from a rich country and linkages are with input-suppliers from other country. This means that it became easier to build up a manufacturing sector, since it is not necessary to build up all the linkages with the manufacturing sector. At the same time, this meant as well that industrialization through joining a supply chain means that there are less spill-overs and linkages with other sectors in the economy.
The growth of the Chinese economy is another important historical event. China has the largest population in the world and had a small GDP per capita for a long time. After pro-market reforms in 1979 the Chinese economy started a long sustained period of economic growth. At first, this was just growth of a relatively small economy and the effects for other countries were not yet large. Now the Chinese economy is the largest or second largest economy in the world. This means that the impact of it started to be felt in the late 1990s and 2000s. China concentrated in the production of consumer goods using labor intensive technologies. The exports of these consumer goods had adverse effects on other countries with a similar factor endowment to start export these products (Eichengreen and Tong, 2006). Exactly these products are excellent to start building a manufacturing sector, since it often does not require skilled labor and access to capital. Hence, other countries seeking to use labor intensive manufacturing as a development strategy are not able to do so.
To form our hypotheses, we expect to find the same results as Rodrik (2016) for manufacturing. We translate our discussion into the following hypothesis on the downward shift of the relationship between the share of manufacturing and GDP per capita.
Hypothesis 1 We expect that
• the cross-country relationship between the share of manufacturing nominal value added and GDP per capita has shifted downward over time, albeit less than the relationship between the share of manufacturing employment and GDP per capita.
• the cross-country relationship between real value added and GDP per capita has shifted down-ward over time, albeit less than the relationship between the share of nominal value added and GDP per capita.
Next, we formulate an hypothesis on the change of the GDP per capita at which the share of manufacturing peaks. Here we note that Rodrik (2016) has not estimated the change in the GDP per capita at which the share of manufacturing nominal value peaks. To formulate an hypothesis for the share of manufacturing nominal value added, we use an intermediate finding by Herrendorf et al. (2018) that, at least in their model, the change in the share of employment and the share of nominal value added should be the same. We thus expect to observe the same form and changes for employment and nominal value added.
Hypothesis 2 We expect that
• the GDP per capita at which the share of manufacturing employment peaks is lower after 1990 than before 1990.
• the GDP per capita at which the share of manufacturing nominal value added peaks – is at same GDP per capita at which the share of manufacturing employment peaks – and after 1990 this GDP per capita is lower than before 1990 and this decrease is equal
to the decrease of the GDP per capita at which the share of manufacturing employment peaks.
• the GDP per capita at which the share of manufacturing real value added peaks
– is higher than the GDP per capita at which the share of manufacturing nominal value added peaks,
– and after 1990 this GDP per capita is lower than before 1990.
2.3
Industries without Smokestacks
Page (2012) argues that industry does not need to have smokestacks and that the discussion around structural change is too much focused on the manufacturing sector alone. He argues, through falling transport cost and communications costs, there are now service and agriculture activities that resem-blance manufacturing. This idea has gotten traction and UNU-WIDER started an “Industries with-out Smokestacks” research project, culminating in a forthcoming book with case studies researching into the possibilities for industries without smokestacks in Sub-Saharan Africa (Newfarmer et al., 2018).
The working paper versions of the chapters in Newfarmer et al. (2018) provide some interesting observations. Ggombe and Newfarmer (2017), for example, show that in Rwanda after the civil war in the early 1990s its recovery and rapid growth was accompanied by a substantial expansion of Rwanda’s services sector first and only later followed by an expansion of Rwanda’s manufacturing sector. Another paper by Spray and Wolf (2017) using tax data and custom data shows that among the 10 most interconnected sectors with other sectors in each country eight are service sectors in Rwanda and six in Uganda. This points towards substantial possibilities for spillovers and linkage effects from service sectors. Another interesting finding from this paper is that among tax paying firms, the agricultural sector shows the highest average output per worker in Uganda and a higher average output per worker than manufacturing in Rwanda, showing that also in developing countries agriculture can have a high labor productivity.
services (ISIC rev3.1 sectors J+K) as part of the industries without smokestacks that we will discuss. Business services include financial services, real estate and renting services and business services of all sorts of varieties, like ICT services, data processing, legal activities, book-keeping and many other activities. These activities are also labor intensive and not necessarily skill intensive. It can thus accommodate structural change as well. Business services is the most tradable service sector and thus the benefits from trade apply also for business services. Business services accounted for 64% of the total worldwide trade in commercial services in 2006 (Francois and Hoekman, 2010, Figure 1) and business services are the most productive service sector and more productive than the manufacturing sector (McMillan et al., 2014, Table 2).
Already in 1984, Bhagwati (1984) saw the possibility that through information and communica-tion networks many services become disembodied in the sense that the presence of the provider of these services at the place of consumption is no longer necessary. He argued already that the more advanced developing countries with an abundant endowment of skills may find their comparative advantage in the provision of these disembodied services. Indeed, we observe that labor-intensive services like call-centers are at a large scale outsourced to the Philippines providing high paying jobs to a relatively well-educated population fluent in English2. Another often used example is the IT
sector in India.
Mishra et al. (2011) show that service export sophistication works in much the same way as earlier demonstrated in Hausmann et al. (2007). Exporting services that are associated with higher income countries is associated with higher growth rates in the future. They argue that services may provide an additional path besides industry towards economic growth. Ghani and O’Connell (2014) shows that even if one considers service sectors aggregated that services showed even faster unconditional convergence in productivity levels than observed in manufacturing. Their research, however, only considers highly aggregated sectors, while Rodrik (2013) shows unconditional convergence for more disaggregated sectors. Amirapu and Subramanian (2015) show that unconditional convergence is actually fastest in real estate and business services.
Eichengreen and Gupta (2011) show that there are two waves of service-sector growth. One observes first a strong expansion of the service sectors overall when a country moves from being a low income country to be being a middle income country and later at a higher GDP per capita one observes another strong expansion. They argue that the first wave is made up of primarily traditional services, while the second is made up of modern services that can use new information technology and that are tradable across borders. These modern services of Eichengreen and Gupta (2011) consist of financial, communication, computer technical, legal, advertising and business services, which are almost all included in our classification of the business service sector. They find that after 1990 the second wave of expansion of services occurs at a lower income level. This earlier second wave points towards the possibility that these business services have taken over part of the role of manufacturing in the process of development.
Rodrik (2016) acknowledges that “many services, such as IT and finance, are high productivity and tradable”, but points to the fact that these sector are often highly skill-intensive, making them unfit to provide employment for the many unskilled workers in developing countries. However, nowadays developing countries have a higher educated population than developed countries had when they had a comparable GDP per capita (Glewwe and Muralidharan, 2016). Their better educated populations mean that developing countries might be better able to engage in the higher skill-intensive tasks from business services than countries with similar a similar GDP per capita were historically. Besides this, an increase in trade in tasks means that developing countries do not need to be able to provide the whole service, but only those tasks that they are able to provide (Baldwin, 2006). Developing countries only need to master some of the skill-intensive tasks, like ICT in India, instead of all skills involved in the final product. Furthermore, many of the tasks within the business service sector like data-processing only require a limited amount of skills.
Agriculture in nowadays rich countries has changed from a sector in which a large part of the population was active into a sector in which only one percent of the employees is active. This points toward the large transformation that agriculture has gone through. From a labor intensive sector it
became one of the most capital intensive sectors. In many developing countries, agriculture is still labor intensive and large parts of the population are still active in this sector. This shows that there are many ways to be active in agriculture. Movement of labor within agriculture from traditional, subsistence modes of production to modern, commercial modes of production can be seen as an equally promising movement as the movement from agriculture to manufacturing.
What can be grown in certain places depends heavily on the soil and climate of an area (Alston and Pardey, 2014). We can not grow pineapples or bananas in the Netherlands without prohibitively large investments. This provides possibilities for countries to attract investments for these agri-cultural products that can not be grown in many places. To provide these products to Western consumers, one has to build supply chains and thus invest in these countries. At the same time, integration in these supply chains demand many capabilities and linkages. For example, transport to export markets require that products are kept cool during transport and can be relatively quickly transported. On global markets product requirements have become stricter, so one needs to invest in quality control. The mastering of these capabilities help also in developing other economic activities (Page, 2012). Further we observe that technological change in agriculture has been very rapid and is thus a highly dynamic sector.
Agricultural production has moved from developed countries towards developing countries. Among the largest producers of agricultural products are now large developing countries like China, India, Brazil and Indonesia. The agricultural production per capita in Latin America and Asia has risen very vast since the 1990s and is for Latin America now almost equal to the production per capita of high income countries. Also the use of fertilizer and capital like tractors in agriculture outside of high income countries has increased considerably (Alston and Pardey, 2014). This shows that modern agriculture may have moved away from the rich Western countries towards developing countries and thus that modern agriculture is larger in developing countries.
The literature on industries without smokestacks points towards other paths of development. We believe that in some countries these sectors may have taken over the role of manufacturing and thus that the downward shift in manufacturing is compensated by an equally sized upward shift in these industries without smokestacks. This leads us to the following hypothesis.
Hypothesis 3 We expect that hypothesis 1 and hypothesis 2 do not hold if we broaden our concept of industrialization to include industries without smokestacks, that is, if instead of manufacturing, we use an aggregated sector in which we combine manufacturing, business services and modern agriculture.
3
Methodology
In this section, we discuss the methodology used in this paper. We will use our various methods using different dependent variables. First we discuss the dependent variables used and afterwards we discuss the methods used.
3.1
Dependent Variables
We start with a discussion of the names of the dependent variables, so we can refer to these vari-ables throughout the text. The generic form of the dependent varivari-ables can be written down as sh sec var and has three elements. The first element sh indicates that it is a share. The second element sec indicates the sector, which can be manufacturing man, business services bs, modern agriculture modagr, an aggregate of the manufacturing and business services agg and an aggregate of manufacturing, business services and modern agriculture agg2. Finally, we have the element var, which indicates the variable. This can either be employment emp, nominal value added va and real value added va q5.
To proxy the share of modern agricultural nominal value-added in total value-added, sh modagr va, we use the share of agricultural production that is exported. We use the value of agricultural ex-ports, val exp, and the value of agricultural production, val prod, to calculate the share of exports in production as sh exp = val exp/val prod. We use this as an estimate of the share of mod-ern agricultural production in agricultural production. We calculate the value-added in modmod-ern agriculture as sh modagr va = sh agr va × sh exp, where we assume implicitly that the ratio of value-added to production is the same for non-exports and exports and thus that agricultural ex-ports do not use proportionately more or less intermediate products than non-exex-ports. If sh exp is bigger than 1, we assume that all the value-added in the agricultural sector is in modern agriculture, i.e. sh modagr va = sh agr va.
3.2
Downward shift over Time in the Cross-Country Relationship
be-tween the Share of Industry and GDP per capita
To test hypothesis 1, we mimic the methods used in section 3 of Rodrik (2016) for the appropriate dependent variables. This means we use the following regression equation
sh sec varct=α1ln(popct) + α2ln(popct)2+ β1ln(yct) + β2ln(yct)2
+ dc+
X
T
ϕTP ERT t+ εct (1)
where sh sec varct is the share of the variable,var, of sector sec in the total economy in country c
and year t, yctis the GDP per capita in country c and year t and popct is the population in country
c and year t. P ERT t is a dummy that is 1 if year t falls in period T , where the periods are 1960’s,
1970’s, 1980’s, 1990’s and the period after 2000, while we keep the period before the 1960’s as base. Finally, dcindicates a full set of country fixed effects. We add the size of population, since population
is an important factor in how large the manufacturing sector is (Chenery and Taylor, 1968). To account for demographic changes we therefore also add population to our regression analysis.
We expect to observe a hump shaped relationship between the share manufacturing and GDP per capita. This means that we expect β1 to be positive and β2 to be negative. For broad modern
industries, we do not have prior expectations of the form, except that it is increasing at low GDP per capita. This corresponds to β1 being positive, while we do not have expectations on β2.
To test hypothesis 1, we use the results from the regression of equation 1 with sh man var as dependent variable. Our hypothesis is confirmed if the estimates for ϕT are significantly negative
and if the estimates for sh man emp are more negative than for sh man va, and the estimates for sh man va are more negative than sh man va q5. To test hypothesis 3, we use the results from this regression with sh agg2 var. We say that our hypothesis is confirmed if the estimates of the period dummies from our regression with broad modern industries are either insignificantly negative or positive. To calculate the significance, we use heteroskedasticity robust standard errors as have been used in Rodrik (2016). We do not use clustered standard errors that are common for these kind of estimations, since this leads to higher standard errors, which would favor confirmation of our hypothesis.
3.3
Change GDP per capita of Peak Industrialization
To test hypothesis 2, we use the method from section 6 of Rodrik (2016) that looks at how the GDP per capita at which the share in manufacturing and broad modern industries peaks has changed. To do so, we first estimate the following regression for the share employment, nominal and real value value added for manufacturing and for broad modern industries.
sh sec varct=β1ln(popct) + β2ln(popct)2+ β3ln(yct) + β4ln(yct)2
+ β5ln(yct) × post90t+ β6ln(yct)2× post90t
where post90tindicates whether year t is after 1990 and other variables are as described in subsection
3.2.
Next, we find the log GDP per capita at which the share of the sector in the variable peaks before 1990 as ln(y\∗pre1990) = −βˆ3
2 ˆβ4
. We find the log GDP per capita at which the share of the variable in the sectors peaks after 1990 asln(y\∗
post1990) = − ˆ β3+ ˆβ5
2( ˆβ4+ ˆβ6)
, where a hat indicates that it is an estimate of the parameter.
3.4
Change GDP per capita of Peak Industrialization using a Cubic
Re-lationship
Since the exact location of the peak is dependent on the functional form we use, we will also find the peak from a cubic regression instead of a second degree polynomial. This gives more degrees of freedom to the relationship, which means that the precise form can better be estimated. To do so, we first use the following regression equation
sh sec varct=β1ln(popct) + β2ln(popct)2+ β3ln(yct) + β4ln(yct)2+ β5ln(yct)3
+ β6ln(yct) × post90t+ β7ln(yct)2× post90t+ β8ln(yct)3× post90t
+ dc+ εct, (3)
Next, we need to find the location of the peak from the parameters. To do so, let us consider the following third degree polynomial f (x) = ax3+ bx2+ cx + d. The following holds at a local
maximum (or minimum) of f
f0(x) = 3ax2+ 2bx + c = 0 ⇐⇒
x = −2b ± √
4b2− 12ac
6a . (4)
We confirm visually that we find the peak as a local maximum within the range of GDP per capita our sample. This means that we find the peak industrialization at the following log GDP per capita before 1990 at \ ln(ypre1990∗ ) = −2 ˆβ4− q 4 ˆβ2 4− 12 ˆβ3βˆ5 6 ˆβ5 (5) and after 1990 at \ ln(y∗ post1990) = −2( ˆβ4+ ˆβ7) − q 4( ˆβ4+ ˆβ7) − 12( ˆβ3+ ˆβ6)( ˆβ5+ ˆβ8) 6( ˆβ5+ ˆβ8) . (6)
We say that hypothesis 2 is confirmed two results hold. First, the GDP per capita at which the share of manufacturing employment and the share of manufacturing nominal value added peaks is the same and that this has decreased after 1990. Second, the share of manufacturing real value added peaks at a higher GDP per capita than both the share of manufacturing employment and the share of manufacturing nominal value added and this GDP per capita is lower after 1990. We say that the part of hypothesis 3 regarding hypothesis 2 is confirmed if the GDP per capita at which the share of broad modern industries does no longer decrease.
4
Data
Table 1: Summary Statistics
(1) (2) (3) (4) (5)
VARIABLES N mean sd min max
sh man emp 2,174 0.146 0.0773 0.00582 0.453 sh man va 2,111 0.200 0.0793 0.00979 0.416 sh man va q5 2,256 0.172 0.0710 0.00610 0.386 sh bs emp 2,173 0.0462 0.0438 0.000799 0.218 sh bs va 2,111 0.0690 0.127 -1.511 0.392 sh bs va q5 2,256 0.0772 0.0613 0.000314 0.351 sh modagr emp 1,317 0.103 0.0781 0.00405 0.404 sh modagr va 1,642 0.0330 0.0466 0.0000712 0.308 sh agg emp 2,173 0.193 0.101 0.0116 0.486 sh agg va 2,111 0.269 0.143 -1.128 0.503 sh agg va q5 2,256 0.249 0.0965 0.0171 0.516 sh agg2 emp 1,316 0.242 0.120 0.0261 0.514 sh agg2 va 1,642 0.302 0.138 -1.039 0.535
GDP per capita in 2011US$, 2011 benchmark 2,461 10,952 11,035 523 62,783
log GDP per capita 2,461 8.777 1.083 6.260 11.05
Population, mid-year (thousands) 2,461 82,026 196,566 497 1.344 million
log Population 2,461 10.08 1.527 6.209 14.11
4.1
10-sector Database
The main data source used is the 10-sector database of the Groningen Growth and Development Centre (GGDC). The 10-Sector Database provides long-run internationally comparable annual data on employment, value-added in current prices and value-added in 2005 prices for 10 broad sectors that encompass the whole economy. It covers 11 countries in Africa, 11 countries in Asia, 2 countries in the Middle East and North Africa, 9 countries in Latin-America, the US and eight European countries. The data set contains data from 1950 onwards to 2013 at the latest.
For the employment data, the 10-sector database uses population censuses as benchmarks. Pop-ulation censuses are held only sporadicly (typically every ten years). For the years in between census years the 10-sector database uses trends from labor force surveys and if labor force surveys are not available of establishment surveys. Population censuses are held over the entire population, so both the formal and the informal sector is included in the data. For the data on value-added (both real and nominal) the 10-sector database uses data from national accounts. All national statistical in-stitutes account for the informal sector in their estimates of value added, but the extent to which this is done and how successful this is differs from country to country. A concern with data from developing countries is the quality of this data, especially for Africa. The 10-sector database uses only data from developing countries in Africa with the strongest national statistical institutes and who have collected national accounts data already for some time. For a detailed description of the methods used in the creation of the 10-sector database, we refer to Timmer et al. (2015).
Data for Germany is only available in the 10-Sector Database before the reunification and value-added in constant prices is only available in 1991 prices. Since our interest is not in developed countries, we exclude West Germany.
In the summary statistics in table 1 we see something strange, namely that the minimum value for sh bs va, sh agg va and sh agg2 va is lower than 0. This is due to negative values for the business services value added in Peru before 1985 in the 10-sector database. We exclude Peru from our estimations if we use either sh bs va, sh agg va or sh agg2 va as a dependent variable.
dummies will pick up the differences in the share of the business service sector from this different classification among countries. In Indonesia, the business service employment was aggregated with other sectors in 1961, while in other years it was not. We therefore dropped this observation.
For the Netherlands, the 10-sector database only provides real value added data for the agri-cultural sector, the manufacturing sector and the transport sector before 1960. This leads to very high shares of manufacturing real value added in these years. We have therefore dropped these data points on real value added. We have also dropped data on agricultural employment between 1950 and 1958, since there was no employment data for other sectors than agriculture.
4.2
Maddison Database
We use the 2018 version of the Maddison Project Database for data on real GDP per capita and population size. The database provides annual data on real GDP per capita and population for 168 countries as far back as 1 AD, but for most countries covering the period from 1950 to 2016. The 2018 version contains two estimates of GDP per capita, one based on multiple benchmarks for purchasing power parities at the years of the different international comparison program (ICP) and one based on an benchmark from the 2011 ICP to obtain GDP levels in international dollars and afterwards uses real GDP growth rates. The first is mostly relevant to compare levels of income at certain points in time, while the second is mostly relevant to study growth over time. Without further notice, we use the second measure of GDP, since we are interested in how the cross-country relationship between GDP per capita and industrialization has changed over time and not how this compares to countries in the same period. We refer to Bolt et al. (2018) for more information on its construction.
4.3
Modern Agriculture Employment Data
Our estimate of employment in the modern agricultural sector are based on extracts from censuses and some other surveys obtained via IPUMS International (Minnesota Population Center, 2018). We have 88 benchmark points for specific country years for 25 countries. In the appendix, we included a list of benchmark country-year data points with the type of data and the original data producer. The data ranges back to 1960, although for most years, especially outside of Latin America, it starts in the 1980s or 1990s.
To find benchmark observations, we use the variable on class of worker harmonized by IPUMS. This variable records whether someone is working as a self-employed, wage/salary worker, unpaid worker or in another type of working relationship. We count as employees in modern agriculture everyone who works as a wage/salary worker in the agriculture, fishing and forestry sector. All other persons working in the agriculture, fishing and forestry sector are recorded as working in subsistence agriculture. We use the percentage of subsistence agricultural workers of total agricultural workers to estimate the percentage of subsistence workers in agriculture, i.e. the sh subs variable. In some cases, especially when IPUMS International provides data from household surveys, person’s weights are provided. Where applicable, we use these person’s weights. Between benchmark years, we use linear interpolation to estimate the share of subsistence employment. For years that are not between benchmark years, we extrapolate by using the benchmark observation that is closest to the given year.
Figure 2: Scatter plot with benchmark data for the share of export in agricultural production and the share of subsistence employment in agriculture. With regression line from fixed effect regression between both variables. Own calculation from IPUMS International and FAOSTAT, see text.
4.4
Modern Agriculture Value Added Data
Lastly, we have collected data on the share of exports in the agricultural production of a country. The data on both exports and production are from the the FAO (2018). For the agricultural production, we used the FAOSTAT Macro Indicators database. Here we used the gross output of Agriculture, Forestry and Fishing with its value in current US$. The data for agricultural exports is from the FAOSTAT Trade database. Here we used the Export Value in Base Price which is given in US$. The export data is available back to the 1960s, but the production data is only available from 1993 onwards. The share of exports, sh exp is given by dividing the agricultural exports by agricultural production. Between benchmark years, we use linear interpolation to estimate the share of exports. For years that are not between benchmark years, we extrapolate by using the benchmark observation that is closest to the given year. The share of exports can be larger than one due to re-exports. In those cases, we set the share of exports equal to one after we have interpolated and extrapolated for missing years.
We use data on exports, since exporting agricultural products means that for those products a country has mastered the necessary skills with regard to quality assurance, the transport of these products and the marketing to the market. This also relates closely to the view of Page (2012) on how modern agriculture can be a part of industries without smokestacks. This means that producing for export markets is a sign that the production is within modern agriculture.
In figure 2, we plot the baseline data of our estimates of sh exp against sh subs. We observe that as the share of subsistence employment in agriculture decreases, the share of exports increases. We thus indeed find that when the share of subsistence agricultural employment as we have defined it is high, the share of agricultural exports is lower. Besides, within countries, if there is a large share of subsistence workers in agriculture, the share of exports tend to go down as well.
access to micro data of this census via IPUMS International of the Minnesota Population Center (2018). Of all the workers we classify as “subsistence workers”, indeed 95.7% also is employed in “traditional or subsistence agriculture” as classified by the Botswanian national statistical institute. Of the agricultural workers in commercial agriculture according to the Botswananian national statis-tical institute, we classify 85.3% as modern agricultural employees. We see this as confirmation that our classification for subsistence workers has substantial overlap with the classification of traditional agricultural workers from the Botswananian national statistical institute and thus that it has overlap with the notion of traditional agriculture in a developing country.
5
Results
In this section, we discuss the results of our analysis. In subsection 5.1, we reexamine the downward shift over time in the cross-country relationship between the share of manufacturing and GDP per capita. We discuss to what we can contribute this downward shift and how these shifts change if we consider broad modern industries. In subsection 5.2, we discuss the change in the GDP per capita at which the share of manufacturing peaks after 1990. We discuss to what we can contribute this change and how this change when we consider broad modern industries. In subsection 5.3, we again discuss the change in the GDP per capita at which the share of manufacturing peaks, but now using our method using a third degree polynomial.
5.1
Downward Shift over Time
We start by re-examining the first claim by Rodrik (2016) that there is a downward shift over time in the cross-country relationship between the share of manufacturing employment and both nominal and real value added and GDP per capita. We start by using the same methods as Rodrik (2016) as we have described in subsection 3.2 and show the results in the first three columns of table 2. For the share of manufacturing employment, column 1 shows a strong shift downwards over time in its relationship with GDP per capita: from the period before 1960 to the period after the 2000 we estimate a shift downward of either 9.3 percentage points. This suggests that countries nowadays have around 9.3 percentage points less of their employees employed in manufacturing than countries before the 1960s at the same GDP per capita.
In column 2, we observe a smaller, but still very substantial shift downward in the share of manufacturing nominal value added and GDP per capita. We estimate that from the period before 1960 to the period after 2000, there was a downward shift of around 5.6 percentage points. Although this is a smaller downward shift than we found for the share of employment, it is still substantial. This means that countries that have a similar GDP per capita now as a country before the 1960s have 5.6 percentage points less of their nominal value added in manufacturing.
Finally, we observe in column 3 an even smaller shift downward in the relationship between the share of manufacturing real value added and GDP per capita. The shift downward is not significant from the period before 1960 to the period after 2000. However, we first observe a statistically significant upward shift until the 1970s. Afterwards, we observe a downward shift in the share of manufacturing real GDP after the 1970s. This coincides with the exit from widespread policies of import substitution and the introduction of the Washington Consensus (Baldwin, 2013a).
In the appendix, we show that these results are robust to sample selection and using a cubic regression. In table 9, we show our results were we use only those observations of a country-year for which we also have data on all three indicators. This only slightly changes our estimations. In table 10, we use a cubic regression. The magnitude of our estimates become smaller for the period dummies in our regression using the share of manufacturing employment. However, this does not lead to a qualitative change of our results
share of manufacturing employment and GDP per capita and the share of manufacturing nominal value added and GDP per capita, but we observe a much smaller downward in the share of real value added and this downward shift only started after the 1970s. A purely demand-based explanation is only valid if all three shares show a downward shift simultaneously. This means that for a large part of the downward shift, we need to consider the faster technological change in manufacturing than in the rest of the economy.
This does not mean that we should not look at changes in demand and globalization for part of the explanation. From faster technological change in manufacturing than in the rest of the economy we would also expect that the share of real manufacturing value added goes up or at least would not go down if substitutability in demand between manufacturing and other sectors is low enough. This means that we should also consider demand changes. These can stem from a preferences shift towards service, an overtake of production of manufacturing by one country with a very strong comparative advantage (for example China) or through countries that are not represented in the sample (for example, countries with very low GDP per capita), or activities that were previously done inside manufacturing firms that are now outsourced and performed by other companies, which means that these jobs are now reclassified as service jobs instead of manufacturing jobs.
Another interesting finding is that the relationship between the share of manufacturing employ-ment and GDP per capita shows a larger downward shift than the relationship between the share of manufacturing nominal value added. Technological change affecting total factor productivity should produce a downward shift in the share of employment equal to or smaller than that in nominal value added if productivity in manufacturing is larger than in the rest of the economy. We offer two possible explanations for this. The first explanation is that technologies have become more capital intensive, which means that less labor is needed, but that the relative price of manufacturing products have not dropped at the same rate. If one transfers the newest technologies, these tech-nologies might be adapted optimally for the factor prices in rich countries and thus be more capital intensive than would be optimal given factor prices in relatively poorer countries. A second expla-nation could hold that the downward shift over time might not be homogeneous for all GDP per capita. It is therefore possible that countries with a low GDP per capita for which manufacturing is much more productive than the rest of the economy, we actually observe an increase in the share of manufacturing employment and nominal value added. We leave this topic for future research. 5.1.1 Broad Modern Industries
We now turn to how the results on premature deindustrialization change if we consider broad modern industries. Our sample for modern agricultural employment and nominal value added and thus for broad modern industries is much smaller than our sample for both manufacturing and business services, because we did not have access for all countries to the micro data that is necessary to find this data. This potentially leads to sample selection issues. We also do not have data on the share of modern agricultural real value added, which means we can not analyze the share of broad modern industrial real value added. To deal with this, we first show the results of our regression for an aggregate of manufacturing and business services for the full sample in the last three columns of table 2. In table 3, we show results of our regressions for manufacturing, manufacturing and business services (so broad modern industries excluding modern agriculture) and broad modern industries using the sample for broad modern industries, so we can account for the changes due to sample selection.
1960 to the period after 2000 is not statistically significant.
Although business services are not able to replace all the manufacturing jobs for countries at a certain GDP per capita, the jobs it provide have a larger labor productivity and therefore the downward shift over time for value added disappears if we aggregate manufacturing and business services. This means that (developing) countries already have found a replacement for the loss of the high productivity, tradable manufacturing sector.
If we look at the share of manufacturing and business service employment, the only periods for which the change of the estimates period dummies were substantially smaller than those for manufacturing employment, were from the 1970s to the 1980s and from the 1980s to the 1990s. This indicates that the business service sector is actually less able to mitigate the downward shift over time in the share of manufacturing employment from the 1990s to the 2000s. From the narrative of industries without smokestacks, we would have expected that it would start to mitigate the downward shift in manufacturing in this period, because this was the time of the large, worldwide implementation of the internet. Note that the business services sector was able to mitigate the downward shift from the 1990s to the 2000s in manufacturing nominal value added and real value added. This indicates that the business service sectors have obtained a higher labor productive over this period through the large advances in ICT. This may point as well to increased worldwide competition in business services through the internet.
Next, we consider how including modern agriculture in broad modern industries affects our estimates. We present the results of this in table 3. Since there are potentially sample selection problems, we use the same sample we have for modern agricultural employment and nominal value added again for manufacturing and manufacturing and business services. For employment, we observe this does not qualitatively change the results. The downward shift in manufacturing and manufacturing and business services is less pronounced, but further it largely gives us the same results as before. However, when we include modern agriculture, we observe that the downward shift becomes even more pronounced than for manufacturing alone. From the period before 1960 to the period after 2000, the cross-country relationship between the share of employment and GDP per capita has shifted 11.5% downward. Technological change and transfer in agriculture has also been very fast (Alston and Pardey, 2014) and this technological change is probably fastest in modern agriculture. This means that the extra demand from opening up new markets may not have been fast enough to counteract the employment losses due to higher labor productivity. In modern agriculture we see the same fast technological change as in manufacturing and this leads to a downward shift in the share of employment.
For the share of nominal value added in the last three columns of table 3, we observe a large sample selection problem. The share of manufacturing nominal value added does not show a large downward shift from the period before the 1960 to after 2000s that was observed using the full sample, although we observe a downward shift after the 1970s. Including business services, we observe a shift upwards from the period before 1960 to the period after 2000 in the cross-country relationship between the share of nominal value added and GDP per capita, where we did not observe this using the full sample. If we also include modern agriculture, we observe an even larger shift upward. This change in upward shift from modern agriculture came mainly about from the 1960s to the 1970s. Although there are large sample selection problems, we still seem to find that modern agriculture helps to mitigate the downward shift in relationship between the share of manufacturing nominal value added and GDP per capita.
products with a higher price or that modern agriculture has become more capital intensive. It can also be that we find this result because we do not have benchmark observations before 1993 for modern agricultural value added.
5.2
Change GDP per capita of Peak Industrialization
Next, we examine the shift of the GDP per capita at which the share of manufacturing peaks. We use the methods described in subsection 3.3. The results are reported in the first three columns of table 4. We observe results that are very similar to those in Rodrik (2016). The peak of the share of manufacturing employment is at a much lower GDP per capita after 1990 than before 1990. This suggests that before 1990, the share of manufacturing employment started to decline at a GDP per capita of around 15,000 (2011 international US dollars), a GDP per capita comparable to Brazil in 2014. After 1990, the share of manufacturing employment peaks at a GDP per capita of around 8000 (2011 international US dollars), lower than the GDP per capita of Angola in 2016.
The GDP per capita at which the share of manufacturing nominal value added peaks after 1990 is at a lower GDP per capita, but this decrease is much smaller than for the share of employment. Before 1990, the share peaked at a GDP per capita of 16,500 (2011 international US dollars), which is roughly comparable to that of Bulgaria in 2016. After 1990, it peaks at a GDP per capita of 11,800 (2011 international US dollars), around the GDP per capita of Peru in 2016. We observe also that this GDP per capita is both before and after 1990 higher than the GDP per capita at which the share of manufacturing employment peaks.
For the share of manufacturing real value added, we observe a very large decrease in the GDP per capita at which the share of manufacturing peaks. However, the peak is outside the range of GDP per capita of our sample before 1990. After 1990, we have exactly one observation with a higher GDP per capita than the GDP per capita of the peak of manufacturing real value added.
The decrease of the GDP per capita at which the share of manufacturing employment peaks is very large. This suggests that not only was there a downward shift over time in the cross-country relationship between the share of manufacturing employment and GDP per capita, the GDP per capita at which this peaks is also substantially lower. This indicates that the relationship between the share of manufacturing employment contracts to the origin as indicated in figure 1 and thus that developing countries will obtain a smaller manufacturing sector in the course of their development and that this will be obtained at an earlier point in its economic development.
Before 1990, we observe that the peaks for the different measures of industrialization peak at the GDP per capita we expected. The share of manufacturing employment and nominal value added peak at approximately the same GDP per capita, while the share of manufacturing real value added peaks at a much higher GDP per capita. This is exactly the pattern that we would expect from the literature. After 1990, however, we observe that the share of manufacturing employment starts to peak much earlier than the share of manufacturing nominal value added. This is a curious change, since we expected these peaks be at the same GDP per capita.
5.2.1 Broad Modern Industries
Next, we see how GDP per capita at which the share of employment and nominal value peaks changes when we use a broader conceptualization of industry. In the last three columns of table 4, we show our results for manufacturing and business services. We observe that the GDP per capita at which our relationship peaks is much higher than for manufacturing alone for all three variables. The share of the business service sector increases with GDP per capita up until a much higher GDP per capita and thus the peak for an aggregate is also at a much higher GDP per capita. Here we note that the peak for employment before and after 1990 is at a much higher GDP per capita than for nominal value added.
value added shows a very large decrease, but these peaks are now well out of sample. This suggests that business services are able to mitigate the downward shift in the share of manufacturing, but are not able to mitigate the large decrease in the GDP per capita at which the share of manufacturing peaks.
If we consider how these results change when we include modern agriculture, we first need to redo the previous estimations using our sample for modern agriculture. We show these results in table 5. For employment, we observe that we estimate that the GDP per capita of the peak of the share of manufacturing employment is now much higher than using the full sample both before and after 1990. The decrease of this peak after 1990 is also very large, but still to a much higher GDP per capita than before. We thus observe that sample selection effects is very substantial. For the data on modern agricultural employment, we have focused on countries outside Europe and North America. Furthermore, we did not find data for countries like Hong Kong and Singapore. This means that the sample now only focuses on relatively poor countries and thus that we might be less able to observe the peaks, especially before 1990. If we include business services, of a local maximum, our method finds a local minimum. The minimum value found is not within the range of GDP per capita in our sample. This is also due to sample selection. Research in the near future can solve this by including countries for Europe and North America in our sample.
If we include modern agriculture, we observe a very large decrease in the GDP per capita at which the share of employment peaks. However, the estimated GDP per capita at which the share of broad modern industries peaks is astronomically high and is not statistically significant. Only after 1990, this peak becomes statistically significant at a higher GDP per capita than for manufacturing. Besides this, the GDP per capita of this peak is just below the GDP per capita of the USA in 2016. Hence, if we also consider both business services and modern agriculture, we observe a peak in the share of employment at a much higher GDP per capita and thus that these broad modern industries start to deindustrialize at a much later point in economic development.
Considering nominal value added, we observe that we estimate the GDP per capita of the peak of the share of manufacturing and the share of manufacturing and business services is lower due to sample selection. We also do no longer observe a decrease in the GDP per capita at which it peaks for both manufacturing and manufacturing and business services. The GDP per capita at which the share of broad modern industries peaks is slightly lower than the GDP per capita at which the share of manufacturing and business services nominal value added peaks. This indicates that the share of modern agriculture peaks at a lower GDP per capita. However, we still do not observe a decrease in the GDP per capita of this peak. Modern agriculture thus does not seem to affect our results for nominal value added.
5.3
Change GDP per capita of Peak Industrialization using a Cubic
Re-lationship
If we use our techniques to find the peak from a cubic relationship between GDP per capita, we observe different results. These results are reported in columns 1 to 3 of table 6. The share of manufacturing employment still peaks at a lower GDP per capita after 1990 than before 1990, however the decline is much smaller than we found using a squared relationship. Further, we observe that the share of manufacturing nominal value added even peaks at a slightly higher GDP per capita. For the share of real value added peaks we now observe peaks within our sample and the GDP per capita is also higher after 1990 than before 1990. However, we need to note that the standard errors of our estimations are large and therefore the confidence intervals of both estimations overlap substantially.
