Automation and Jobs - When Technology Boosts Employment Ch 1

AUTOMATION AND JOBS: WHEN TECHNOLOGY BOOSTS EMPLOYMENT Automation and Jobs: When technology boosts employment Abstract: Will industries use new information technologies to eliminate jobs? Sometimes productivity-enhancing technology increases industry employment instead. In manufacturing, jobs grew along with productivity for a century or more; only later did productivity gains bring declining employment. What changed? Markets became saturated. Using two centuries of data, a simple model of demand accurately explains the rise and fall of employment in the US textile, steel, and automotive industries. The model also predicts that computer technology should generate relatively greater job growth in nonmanufacturing industries today. Estimates show computer use is associated with declining employment in manufacturing industries, but not in other sectors. When an industry automates its production does its employment decline? There is widespread concern today that many jobs will be lost to new computer technologies. One recent paper concluded that new information technologies will put “a substantial share of employment, across a wide range of occupations, at risk in the near future” (Frey and Osborne 2013). The example of manufacturing decline provides good reason to be concerned about technology and job losses. In 1958, the US broadwoven textile industry employed over 300 thousand production workers and the primary steel industry employed over 500 thousand. By 2011, broadwoven textiles employed only 16 thousand and steel employed only 100 thousand production workers.1 Some of these losses can be attributed to trade, especially since the mid-1990s. However, overall since the 1950s, most of the decline appears to come from technology and changing demand (Rowthorn and Ramaswamy 1999). Of course, jobs lost in one industry may be replaced with new jobs in another, but many fear that deindustrialization has heightened economic inequality in the US or caused unemployment in Europe and caused political instability in both. “Premature deindustrialization” has also become a concern in developing nations (Rodrik 2016). Yet technical change does not always lead to declining employment in the affected industry. Figure 1 shows how textiles, steel, and automotive manufacturing all enjoyed strong employment growth during many decades that also experienced very rapid productivity growth. Despite persistent and substantial productivity growth, these industries have spent more decades with growing employment than with job losses. This “inverted U” 1 These figures are for the broadwoven fabrics industry using cotton and manmade fibers, SIC 2211 and 2221, and the steel works, blast furnaces, and rolling mills industry, SIC 3312. pattern appears to be quite general for manufacturing industries (Buera and Kaboski 2009, Rodrik 2016). Figure 1. Production Employment in Three Industries Why is automation associated with growing employment in some industries at some times, while at other times and in other industries jobs are lost? What determines whether technical change tends to increase employment in an industry or decrease it? The answer would seem to be important for understanding what affects the pace of deindustrialization in both the developed and developing worlds and for anticipating the impact of new information technologies. This paper uses century-long time series data on the US cotton textile, steel, and automotive manufacturing industries to explore what determines whether technology will increase or decrease employment. While a substantial literature has looked at structural change at the level of the manufacturing sector as a whole, the data for these individual industries allows a tighter identification of the interaction between technology, demand, prices, and income.2 Using these data, a very simple model predicts the rise and fall of employment in these industries with reasonable accuracy: the solid line in Figure 1 shows that prediction. I then explore implications of this model. One implication is that computer automation should have a more positive effect on employment in nonmanufacturing industries than in manufacturing. I find support for this hypothesis using data on computer use across industries from 1984 through 2007. Structural change The inverted U pattern in Figure 1 is also seen in the relative share of employment in the whole manufacturing sector, shown in Figure 2. Logically, the rise and fall of the sector as a whole in this chart results from the aggregate rise and fall of separate manufacturing industries such as those in Figure 1. Yet explanations of this phenomenon based on broad 2 Papers empirically analyzing the sector shifts include Dennis and Iscan (2009), Buera and Kaboski (2009), Kollmeyer (2009), Nickell, Redding, and Swaffield (2008), and Rowthorn and Ramaswamy (1999). Figure 2. Manufacturing Share of the Labor Force Sources: US Bureau of the Census 1975; BLS Current Employment Situation. Labor force includes agricultural laborers sector-level factors face a challenge because individual industries show rather disparate patterns. For example, employment in the automotive industry appears to have peaked nearly a century after textile employment peaked. Data on individual industries are needed to analyze such disparate responses. The literature on structural change provides two sorts of accounts for the relative size of the manufacturing sector, one based on differential rates of productivity growth, the other based on different income elasticities of demand.3 Baumol (1967) showed that the greater rate of technical change in manufacturing industries relative to services leads to a 3 Acemoglu and Guerrieri (2008) also propose an explanation based on differences in capital deepening. declining share of manufacturing employment under some conditions (see also Lawrence and Edwards 2013, Ngai and Pissarides 2007, Matsuyama 2009). But differences in productivity growth rates do not seem to explain the initial rise in employment. For example, during the 19th century, the share of employment in agriculture fell while employment in manufacturing industries such as textiles and steel soared both in absolute and relative terms. But labor productivity in these manufacturing industries grew faster than labor productivity in agricultural. Parker and Klein (1966) find that labor productivity in corn, oats, and wheat grew 2.4%, 2.3%, and 2.6% per annum from 1840-60 to 1900-10. In contrast, labor productivity in cotton textiles grew 3.0% per year from 1820 to 1900 and labor productivity in steel grew 3.0% from 1860 to 1900.4 Nevertheless, employment in cotton textiles and in primary iron and steel manufacturing grew rapidly then. The growth of manufacturing relative to agriculture surely involves some general equilibrium considerations, perhaps involving surplus labor in the agricultural sector (Lewis 1954). But at the industry level, rapid labor productivity growth along with job growth must mean a rapid growth in the equilibrium level of demand—the amount consumed must increase sufficiently to offset the labor-saving effect of technology. For example, although labor productivity in cotton textiles increased nearly 30-fold during the nineteenth century, consumption of cotton cloth increased 100-fold. The inverted U thus seems to involve an interaction between productivity growth and demand. A long-standing literature sees sectoral shifts arising from differences in the income elasticity of demand. Clark (1940), building on earlier statistical findings by Engel (1857) and others, argued that necessities such as food, clothing, and housing have income elasticities that are less than one (see also Boppart 2014, Comin, Lashkari, and Mestieri 2015, 4 My estimates, data described below. Kongsamut, Rebelo, and Xie 2001 and Matsuyama 1992 for more general treatments of nonhomothetic preferences). The notion behind “Engel’s Law” is that demand for necessities becomes satiated as consumers can afford more, so that wealthier consumers spend a smaller share of their budgets on necessities. Similarly, this tendency is seen playing out dynamically. As nations develop and their incomes grow, the relative demand for agricultural and manufactured goods falls and, with labor productivity growth, relative employment in these sectors falls even faster. This explanation is also incomplete, however. While a low income elasticity of demand might explain late 20th century deindustrialization, it does not easily explain the rising demand for some of the same goods during the nineteenth century. By this account, cotton textiles are a necessity with an income elasticity of demand less than one. Yet during the 19th century, the demand for cotton cloth grew dramatically as incomes rose. That is, cotton cloth must have been a “luxury” good then. Nothing in the theory explains why the supposedly innate characteristics of preferences for cloth changed. It would seem that the nature of demand changed over time. Matsuyama (2002) introduced a model where the income elasticity of demand changes as incomes grow (see also Foellmi and Zweimueller 2008). In this model, consumers have hierarchical preferences for different products. As their incomes grow, consumer demand for existing products saturates and they progressively buy new products further down the hierarchy. Given heterogeneous incomes that grow over time, this model can explain the inverted U pattern. It also corresponds, in a highly stylized way, to the sequence of growth across industries seen in Figure 1. Yet there are two reasons that this model might not fit the evidence very well for individual industries. First, the timing of the growth of these industries seems to have much more to do with particular innovations that began eras of accelerated productivity growth than with the progressive saturation of other markets. Cotton textile consumption soared following the introduction of the power loom to US textile manufacture in 1814; steel consumption grew following the US adoption of the Bessemer steelmaking process in 1856, and Henry Ford’s assembly line in 1913 initiated rapid growth in motor vehicles. Second, there is a general problem of looking at the income elasticity of demand as the main driver of structural change: the data suggest that prices were often far more consequential for consumers than income. From 1810 to 2011, real GDP per capita rose 30fold, but output per hour in cotton textiles rose over 800-fold; inflation-adjusted prices correspondingly fell by three orders of magnitude. Similarly, from 1860 to 2011, real GDP per capita rose 17 fold, but output per hour in steel production rose over 100 times and prices fell by a similar proportion. The literature on structural change has focused on the income elasticity of demand, often ignoring price changes. Yet these magnitudes suggest that low prices might substantially contribute to any satiation of demand. I develop a model that includes both income and price effects on demand, allowing both to have changing elasticities over time. The inverted U pattern in industry employment can be explained by a declining price elasticity of demand. If we assume that rapid productivity growth generated rapid price declines in competitive product markets, then these price declines would be a major source of demand growth. During the rising phase of employment, equilibrium demand had to increase proportionally faster than the fall in prices in response to productivity gains. During the deindustrialization phase, demand must have increased proportionally less than prices. Below I obtain estimates that show the price elasticity of demand falling in just this manner. To understand why this may have happened, it is helpful to return to the origins of the notion of a demand curve. Dupuit (1844) recognized that consumers placed different values on goods used for different purposes. A decrease in the price of stone would benefit the existing users of stone, but consumers would also buy stone at the lower price for new uses such as replacing brick or wood in construction or for paving roads. In this way, Dupuit showed how the distribution of uses at different values gives rise to what we now call a demand curve, allowing for a calculation of consumer surplus. This paper proposes a parsimonious explanation for the rise and fall of industry employment based on a simple model where consumer preferences follow such a distribution function. The basic intuition is that when most consumers are priced out of the market (the upper tail of the distribution), demand elasticity will tend to be high for many common distribution functions. When, thanks to technical change, price falls (income rises) to the point where most consumer needs are met (the lower tail), then the price and income elasticities of demand will be small. The elasticity of demand thus changes as technology brings lower prices to the affected industries and higher income to consumers generally. I fit the model to actual demand data for the three industries with a lognormal specification that allows for changes in both the price elasticity of demand and the income elasticity of demand. The model estimates per capita demand accurately using only a single independent variable: labor productivity. I use the demand estimates to make the predictions of the actual rise and fall of employment in the textile, steel, and automotive industries shown in Figure 1. This simple model of labor productivity and demand thus provides an explanation for the inverted U seen in manufacturing employment. The model implies that the characteristics of demand should determine whether the new technologies tend to increase or decrease employment. To test this, I compare the impact of computer technology on industry employment growth between manufacturing and nonmanufacturing industries. Inferring from the historical analysis that demand for manufactured products is less elastic on average than demand in other industries, the model suggests that computer technology should have a more positive impact on employment in nonmanufacturing industries. I find that from 1984 through 2007 computer technology use is associated with economically and statistically significant job losses in manufacturing industries but with a weak increase in employment in nonmanufacturing industries. This analysis supports Autor’s (2015) explanation of why automation has not eliminated employment. Model Simple model of the Inverted U Consider production and consumption of two goods—cloth and a general composite good—in autarky. The model will focus on the impact of technology on employment in the textile industry under the assumption that the output and employment in the textile industry are only a small part of the total economy. Production Let the output of cloth, 𝑞 = 𝐴 ∙ 𝐿, where L is textile labor and A is a measure of technical efficiency. Changes in A represent labor-augmenting technical change. Note that this is distinct from those cases where automation completely replaces human labor. Bessen (2016) shows that such cases are rare and that the main impact of automation consists of technology augmenting human labor. I initially assume that product and labor markets are competitive so that the price of cloth is (1) 𝑝 = 𝑤 𝐴, where w is the wage. Below, I will test whether this assumption holds in the cotton and steel industries. Then, given a demand function, 𝐷(𝑝), equating demand with output implies 𝐷 𝑝 =𝑞 =𝐴∙𝐿 (2) or 𝐿 = 𝐷 𝑝 /𝐴. We seek to understand whether an increase in A, representing technical improvement, results in a decrease or increase in employment L. That depends on the price elasticity of demand, 𝜖, assuming income is constant. Taking the partial derivative of the log of (2) with respect to the log of A, 𝜕 ln 𝐿 𝜕 ln 𝐷(𝑝) 𝜕 ln 𝑝 = − 1 = 𝜖 − 1, 𝜕 ln 𝑝 𝜕 ln 𝐴 𝜕 ln 𝐴 𝜖≡− 𝜕 ln 𝐷(𝑝) 𝜕 ln 𝑝 If the demand is elastic (𝜖 > 1), technical change will increase employment; if demand is inelastic (𝜖 < 1), jobs will be lost. In addition to this price effect, changing income might also affect demand as I develop below. Consumption Now, consider a consumer’s demand for cloth. Suppose that the consumer places different values on different uses of cloth. The consumer’s first set of clothing might be very valuable and the consumer might be willing to purchase even if the price were quite high. But cloth draperies might be a luxury that the consumer would not be willing to purchase unless the price were modest. Following Dupuit (1844) and the derivation of consumer surplus used in industrial organization theory, these different values can be represented by a distribution function. Suppose that the consumer has a number of uses for cloth that each give her value v, no more, no less. The total yards of cloth that these uses require can be represented as f(v). That is, when the uses are ordered by increasing value, f(v) is a scaled density function giving the yards of cloth for value v. If we suppose that our consumer will purchase cloth for all uses where the value received exceeds the price of cloth, 𝑣 > 𝑝, then for price p, her demand is ! 𝐷 𝑝 = ! 𝑓 𝑧 𝑑𝑧 = 1 − 𝐹 𝑝 , 𝐹(𝑝) ≡ ! 𝑓 𝑧 𝑑𝑧 ! where I have normalized demand so that maximum demand is 1. With this normalization, f is the density function and F is the cumulative distribution function. I assume that these functions are continuous with continuous derivatives for p>0. The total value she receives from these purchases is then the sum of the values of all uses purchased, ! 𝑈 𝑝 = 𝑧 ∙ 𝑓 𝑧 𝑑𝑧 . ! This quantity measures the gross consumer surplus and can be related to the standard measure of net consumer surplus used in industrial organization theory (Tirole 1988, p. 8) after integrating by parts: ! 𝑈 𝑝 = ! 𝑧 ∙ 𝑓 𝑧 𝑑𝑧 = ! ! 𝑧 ∙ 𝐷′ 𝑧 𝑑𝑧 = 𝑝 ∙ 𝐷 𝑝 + ! 𝐷 𝑧 𝑑𝑧. ! In words, gross consumer surplus equals the consumer’s expenditure plus net consumer surplus. I interpret U as the utility that the consumer derives from cloth.5 The consumer also derives utility from consumption of the general good, x, and from leisure time. Let the portion of time the consumer works be l so that leisure time is 5 Note that in order to use this model of preferences to analyze demand over time, one of two assumptions must hold. Either there are no significant close substitutes for cloth or the prices of these close substitutes change relatively little. Otherwise, consumers would have to take the changing price of the potential substitute into account before deciding which to purchase. If there is a close substitute with a relatively static price, the value v can be reinterpreted as the value relative to the alternative. Below I look specifically at the role of close substitutes for cotton cloth, steel, and motor vehicles. 1 – l. Assume that the utility from these goods is additively separable from the utility of cloth so that total utility is 𝑈 𝑣 + 𝐺(𝑥, 1 − 𝑙) where G is a concave differentiable function. The consumer will select v, x, and l to maximize total utility subject to the budget constraint 𝑤𝑙 ≥ 𝑥 + 𝑝𝐷 𝑣 where the price of the composite good is taken as numeraire. The consumer’s Lagrangean can be written ℒ 𝑣, 𝑥, 𝑙 = 𝑈 𝑣 + 𝐺(𝑥, 1 − 𝑙) + 𝜆 𝑤𝑙 − 𝑥 − 𝑝 ∙ 𝐷 𝑣 . Taking the first order conditions, and recalling that under competitive markets, 𝑝 = 𝑤 𝐴, we get 𝑣 = 𝐺! 𝑝 𝐺! = , 𝑤 𝐴 𝐺! ≡ 𝜕𝐺 . 𝜕𝑙 𝐺! represents the marginal value of leisure time and the second equality results from applying assumption (1). In effect, the consumer will purchase cloth for uses that are at least as valuable as the real cost of cloth valued relative to leisure time. Note that if 𝐺! is constant, the effect of prices and the effect of income are inversely related. This means that the price elasticity of demand will equal the income elasticity of demand. However, the marginal value of leisure time might very well increase or decrease with income; for example, if the labor supply is backward bending, greater income might decrease equilibrium 𝐺! so that leisure time increases. To capture that notion, I parameterize 𝐺! = 𝑤 ! so that (3) 𝑣 = 𝑤 ! 𝐴 = 𝑤 !!! 𝑝, 𝐷 𝑣 =1−𝐹 𝑣 . Elasticities Using (3), the price elasticity of demand holding wages constant solves to 𝜖=− 𝜕 ln 𝐷 𝜕 ln 𝐷 𝑣 𝜕 ln 𝑣 𝑝𝑓(𝑣) = = 𝑤 !!! 𝜕 ln 𝑝 𝜕 ln 𝑣 𝜕 ln 𝑝 1 − 𝐹(𝑣) and the income (wage) elasticity of demand holding price constant is 𝜌= 𝜕 ln 𝐷 𝜕 ln 𝐷 𝑣 𝜕 ln 𝑣 = = 1 − 𝛼 𝜖. 𝜕 ln 𝑣 𝜕 ln 𝑤 𝜕 ln 𝑤 These elasticities change with prices and wages or alternatively with changes in labor productivity, A. The changes can create an inverted-U in employment. Specifically, if the price elasticity of demand, 𝜖, is greater than 1 at high prices and lower than 1 at low prices, then employment will trace an inverted U as prices decline with productivity growth. At high prices relative to income, productivity improvements will create sufficient demand to offset job losses; at low prices relative to income, they will not. A preference distribution function with this property can generate a kind of industry life cycle as technology continually improves labor productivity over a long period of time. An early stage industry will have high prices and large unmet demand, so that price decreases result in sharp increases in demand; a mature industry will have satiated demand so further price drops only produce an anemic increase in demand. A necessary condition for this pattern is that the price elasticity of demand must increase with price over some significant domain, so that it is smaller than 1 at low prices but larger than 1 at high prices. It turns out that many distribution functions have this property. This can be seen from the following propositions (proofs in the Appendix): Proposition 1. Single-peaked density functions. If the distribution density function, f, !" has a single peak at 𝑝 = 𝑝, then ≥ 0 ∀ 𝑝 < 𝑝. !" Proposition 2. Common distributions. If the distribution is normal, lognormal, exponential, or uniform, there exists a 𝑝∗ such that for 0 < 𝑝 < 𝑝∗ , 𝜖 < 1, and for 𝑝∗ < 𝑝, 𝜖 > 1. These propositions suggest that the model of demand derived from distributions of preferences might be broadly applicable. The second proposition is sufficient to create the inverted U curve in employment as long as price starts above 𝑝∗ and declines below it. Econometric specification Below, I estimate the demand function using a lognormal distribution. To do this using aggregate data, the model, which describes the demand of an individual consumer, needs to be recast to describe aggregate consumer behavior. The distribution function can be recast as an aggregate distribution across both different consumers and different uses for each consumer. Also, an average wage now determines the equilibrium value of 𝑣, so the distribution also reflects dispersion of wages across consumers. Specifically, I estimate per capita demand, D, (4) 𝐷 =𝛾 1−Φ −ln 𝐴 + 𝛼 ln 𝑤 − 𝜇 𝜎 +𝜀 or (5) 𝐷 =𝛾 1−Φ ln 𝑝 𝑤 + 𝛼 ln 𝑤 − 𝜇 𝜎 +𝜀 where Φ is the standard normal cumulative distribution function and 𝜀 is an error term that captures, among other things, demand shocks and changing tastes. I estimate these equations using non-linear least squares (NLLS). I also estimate demand with a general polynomial form. Data Time series over a century in length often require combining data from different sources involving various adjustments. I describe the data sources and adjustments in detail in the Appendix. This section describes the main data series used in estimating employment in cotton textiles, steel, and automotive industries and in the computer technology analysis. Production and demand I use physical quantities to measure production and demand. For the textile industry, I measure output as yards of cotton cloth produced plus yards of cloth made of synthetic fibers from 1930 on. From 1958, I use the deflated output of the cotton and synthetic fiber broadwoven cloth industries (SIC 2211 and 2221). For the early years, I also included estimates of cotton cloth produced in households. For steel, I used the raw short tons of steel produced. For the motor vehicle industry I used the number of passenger vehicles and trucks produced each year. To estimate per capita demand or consumption, I add net imports to the estimates of domestic production and divide by the population. Note that these measures do not adjust for product quality.6 This approach avoids distortions that might arise from constructing quality adjusted price indices over long periods of time. It does mean that “true” demand and productivity are understated. However, this does not pose a significant problem for my analysis because I measure both without quality adjustments. The distribution function I estimate would, of course, be different if it were The deflators used from 1958 on in cotton are quality adjusted but the series closely matches the unadjusted output measure during the years when they overlap. 6 estimated with quality-adjusted data, but using unadjusted data allows for consistent predictions of employment. Employment, prices, and wages I count the number of industry wage earners or, from 1958 on, the number of production workers. For prices, I use the prices of standard commodities. For cotton textiles, I use the wholesale price for cotton sheeting. For steel, I use wholesale prices for steel rails. I do not have a similar commodity price for motor vehicles. The BLS does have a price index for the automotive industry, but this measure implicitly changes as the quality of vehicles improved. I need to use a commodity type price because my measures of output and consumption (cars and trucks) does not capture these quality improvements. For wages, I use the compensation of manufacturing production workers. This measure includes the value of employee benefits from 1906 on. Labor productivity I calculate labor productivity by dividing output by the number of production employees times the number of hours worked per year. I use industry specific estimates of hours if available and estimates of hours for manufacturing workers if not. Over the sample periods, each industry exhibited rapid labor productivity growth. From 1820 to 1995, labor productivity in cotton textiles grew 2.9% per year; in steel, it grew 2.4% per year from 1860 to 1982; in motor vehicles, it grew 1.4% per year from 1910 through 2007. Figure 3 shows labor productivity for each industry on a log scale over time. Each industry exhibits steady productivity growth over long periods of time. Textiles and especially automotive show initially higher rates of growth; steel exhibits faster growth since the 1970s, likely the effect of steel minimills that use recycled steel rather than blast furnace production of iron. Figure 3. Labor Productivity over Time Computer technology The data on computer use come from the Current Population Survey (CPS) of the Bureau of Labor Statistics. In select years, a supplemental survey asked whether the respondent “directly used a computer at work?”7 The universe for these questions was employed persons of 15 years of age or older. I use these survey items as the measure of computer use. Because these measures are self-reported, they may undercount cases where information technology is embedded in hardware. I use the Census 1990 industry classification scheme where IPUMS crosswalks have converted industry codes to a consistent set of 227 industries (Ruggles 2015). I calculate the share of work hours performed by workers using computer technology for each industry for each sample year. Note that I assume that the use of computers means that industries are subject to an ongoing stream of productivity-enhancing technical changes as new hardware and software applications are installed. 7 I use the surveys of computer use in October of 1984, 1989, 1993, 1997, and 2003 and in September of 2001.