Automation and Jobs - When Technology Boosts Employment Ch 3

The direction and rate of job changes A naïve view holds that more rapid productivity growth will be more likely to create job losses. The analysis in this paper suggests, to the contrary, that demand determines whether productivity growth eliminates or increases jobs. Moreover, if demand is elastic, then more rapid productivity growth will actually lead to faster employment growth, all else equal. While the sign of the employment effect depends on the elasticity of demand and not on the rate of productivity growth, the rate of change depends on both. For instance, although productivity grew faster in cotton textiles than in auto, employment grew faster in auto; the distribution of preferences for motor vehicles was much more concentrated (small 𝜎) than the distribution for textiles. One might expect that computer technology would have a different effect on job growth across industries depending on industry demand elasticity. Assuming that the historical process of deindustrialization means that manufacturing industries have less elastic (more satiated) demand than most other industries on average, then computer technology should have a relatively more negative impact on employment in manufacturing industries, all else equal.14 To the extent that information technology allows manufacturers to create new products, then the technology might tap into new sources of demand and thus be more elastic. There is some evidence that information technology is used to create new products (see for example Bartel, Ichniowski, and Shaw 2007). New product varieties might provide a reason that the association between IT and employment growth in manufacturing is only weakly negative. 14 Table 3 performs a simple test of this hypothesis. From equation (2) it follows that, all else equal, ∆ ln 𝐿 = (𝜖 − 1) ∙ ∆ ln 𝐴. I assume that labor productivity growth in industry i at time t is proportional to the share of workers using computers in that industry, 𝐶!" . Then, estimating ∆ ln 𝐿!" = 𝛽!"!#$!% ∙ 𝐶!" ∙ 𝐼 𝑛𝑜𝑛𝑚𝑎𝑛𝑢. + 𝛽!"#$ ∙ 𝐶!" ∙ 𝐼 𝑚𝑎𝑛𝑢. + 𝛿! + 𝜀!" the hypothesis can be tested as a test of 𝛽!"!#$!% > 𝛽!"#$ . Column 1 shows that the share of the industry workforce using computers is positively associated with employment growth in nonmanufacturing industries, but is significantly associated with job loss in manufacturing. The difference in the coefficients is highly significant. The bottom panel of the table calculates the effect at the sample mean; computer use is associated with about 1% per annum faster employment growth on average in nonmanufacturing industries, but is associated with a loss of jobs of about 3% per year on average in manufacturing. These estimates do not take into account the possible impact of trade as was done above for the historical analysis. To consider the impact of trade, I repeated the regression after adjusting the dependent variable for the rate of change in the log import penetration for manufacturing industries.15 The coefficients were very close to the original estimates. Column 2 adds two additional controls for occupational characteristics that might be related to industry growth. Industries that send work offshore would tend to grow more slowly. Jensen and Kletzer’s (2010) developed an index for the likelihood that the work of occupations can be sent offshore. I use the weighted average over the occupations in each industry. Also, industries might grow at different rates depending on the relative supply of 15 That is, I added ∆ ln 𝑜𝑢𝑡𝑝𝑢𝑡 𝑜𝑢𝑡𝑝𝑢𝑡 + 𝑖𝑚𝑝𝑜𝑟𝑡𝑠 − 𝑒𝑥𝑝𝑜𝑟𝑡𝑠 . different type of labor they require. The Bureau of Labor Statistics O*NET database (version 17.0) rates occupations on the extent to which they require a college diploma or advanced degree. I include a weighted average of these ratings over the occupations in each industry. The additional controls are significant, but coefficients of computer use are similar. Perhaps other industry characteristics might be associated with both computer use and industry growth. Column 3 repeats the regression but with a full set of industry fixed effects. The difference between the coefficients is still highly significant, although both coefficients are now negative. The standard errors are also substantially larger, not surprising, given the short panel length. Computer use might also be endogenous if, say, faster growing industries tend to also use computers more. For example, perhaps rapidly growing industries have greater difficulty managing resources and they rely more on computer systems to help manage. To check for endogeneity, I instrument computer use with a measure of how sedentary each industry’s workforce is. The notion is that desktop computers can be more easily applied to sedentary jobs. I obtain the variable from the 1977 edition of the Dictionary of Occupational Titles.16 This variable is correlated with industry computer use in the sample (correlation coefficient of 0.43). However, because the 1977 estimates were made before computers were widely adopted, the instrument is likely independent of computer use; that is, computers could not have had much influence on whether jobs were sedentary then. The measure also appears to be independent of other unobserved variables that might be associated with employment growth. I performed t-tests comparing employment growth rates between Experts rated each occupation for the level of physical exertion involved ranging from 1 (sedentary) to 5 (very heavy). England and Kilbourne (2013) have mapped the 14,000 occupations in the DOT to Census detailed occupation codes, averaging them to this higher level of aggregation. I classified each occupation with a mean score less than 2 as sedentary and then calculated weighted averages across all occupations for each industry in the 1980 Census 5% sample. 16 sedentary and non-sedentary occupations for each decade. During the 1960s and 1970s, the difference in employment growth between these groups of occupations was not significant (probability values of 0.409 and 0.286 respectively). Thus there does not appear to have been a significant association between the instrumental variable and employment growth before computers became widely adopted. After 1980, however, the picture changes and the sedentary occupations grow significantly faster during the 1980s, 1990s, and from 2000 through 2013 (probability values of 0.000, 0.009, and 0.009 respectively). These tests suggest that while the instrument has been correlated with employment growth since computers were widely adopted, it was not correlated before then. This pattern suggests that sedentary work does not influence employment growth except through computer use, satisfying the exclusion restriction. Columns 4 and 5 perform separate instrumental variables estimations for nonmanufacturing and manufacturing industries using this measure of sedentariness as well as year dummies as instruments.17 For nonmanufacturing industries, computer use is associated with industry employment growth both economically and statistically; in manufacturing, the association is not significantly different from zero. Given the larger standard errors, however, the difference between these sectors is not statistically significant (P = .124). Also, test of the null hypothesis that computer share is exogenous cannot be rejected for the nonmanufacturing sample (P = .267), but can be rejected for manufacturing (P = .005). In summary and consistent with my hypothesis, computer use is associated with lower and typically negative employment growth in manufacturing industries. In other sectors, computer use is associated with employment growth in most estimates. These results I estimate the two samples separately in order to have a sufficient number of exogenous variables relative to endogenous variables. 17 correspond to several recent papers finding that information technology increases employment for some groups and does not appear to reduce net employment, including some causal studies. Gaggl and Wright (2014) find that ICT tended to raise employment in wholesale, retail, and finance industries, but had no statistically significant effect on other sectors, including manufacturing. Akerman, Gaarder, and Mogstad (2015) find that Internet technology increased employment of skilled workers and had no effect on unskilled. Bessen (2016) finds that computers tend to increase occupational employment modestly overall, with job losses in low wage occupations. Autor, Dorn, and Hansen (2015) find that local markets susceptible to computerization are not more likely to experience employment loss. Conclusion Productivity-enhancing technology will increase industry employment if product demand is sufficiently elastic. Technical change reduces the labor required needed to produce a unit of output, but it also reduces prices in competitive markets. If the price elasticity of demand is greater than one, the increase in demand will more than offset the labor saving effect of the technology. Understanding the responsiveness of demand is thus key to understanding whether major new technologies will decrease or increase employment in affected industries. This paper proposes that industry employment dynamics in the face of extensive productivity growth can be analyzed by deriving demand from a distribution of preferences. For many distribution functions, the elasticity of demand declines as price declines and productivity grows. In particular, a parsimonious model using a lognormal distribution fits the demand curves well for cotton textiles, steel, and motor vehicles over long periods of time. This model generates an industry life cycle explanation for the inverted U pattern of industrialization/deindustrialization seen in manufacturing employment. At high initial prices, industries have large unmet demand that is highly elastic. Productivity improvements give rise to robust job growth. Over time and with ongoing productivity gains, prices progressively decline until most demand is met and the price elasticity of demand is quite low. Then further productivity gains bring reduced employment. This view implies that major new technologies today should increase employment if they improve productivity in markets that have large unmet needs. That appears to be the case with computer technology. I find, consistent with the analysis, that computer technology use is associated with employment growth in nonmanufacturing industries, but in manufacturing, where a couple centuries of rapid productivity growth mean that demand is more satiated, computer technology is associated with employment declines. This model challenges a popular view that faster technical change is more likely to eliminate jobs. Some people argue that because of Moore’s Law, the rate of change will be fast in new information technologies and this will cause unemployment (Ford 2015). However, my model shows that if demand is elastic, faster technical change will, instead, create faster employment growth. Faster technical change will, however, also hasten the day when demand is no longer so elastic and deindustrialization sets in. This suggests that although new information technologies are not likely to create major job losses in the near future, job losses are a very real concern in the long run. On the other hand, as long as there are large unmet consumer needs, automation will raise demand in the relevant industries sufficiently to create job growth. Appendix Propositions To simplify notation, let the wage remain constant at 1. Then 𝜖 𝑝 = 𝑝 𝑓(𝑝) 1 − 𝐹(𝑝) so that 𝜕𝜖 𝑝 𝑓 !𝑝 𝑓 !𝑝 = + 1−𝐹 𝜕𝑝 1−𝐹 ! + 𝑓′ 𝑓 𝑓 1 =𝜖 + + 𝑓 1−𝐹 𝑝 1−𝐹 Note that the second and third terms in parentheses are positive for 𝑝 > 0; the first term could be positive or negative. A sufficient condition for !" !" ≥ 0 is (A1) 𝑓′ 𝑓 + ≥ 0. 𝑓 1−𝐹 Proposition 1. For a single peaked distribution with mode 𝑝, for 𝑝 < 𝑝 , 𝑓 ! ≥ 0 so that !" !" ≥0. Proposition 2. For each distribution, I will show that 𝜕𝜖 ≥ 0, 𝜕𝑝 lim 𝜖 = 0, !→! lim 𝜖 = ∞. !→! Taken together, these conditions imply that for sufficiently high price, 𝜖 > 1, and for a sufficiently low price, 𝜖 < 1. a. Normal distribution 𝑓 𝑝 = 1 𝜑 𝑥 , 𝜎 𝐹 𝑝 =Φ 𝑥 , 𝜖 𝑝 = 𝑝 𝜑 𝑥 , 𝜎 (1 − Φ 𝑥 ) 𝑥≡ 𝑝−𝜇 𝜎 where 𝜑 and Φ are the standard normal density and cumulative distribution functions respectively. Taking the derivative of the density function, 𝑓! 𝑓 𝑥 𝜑 𝑥 + =− + . 𝑓 1−𝐹 𝜎 𝜎 1−Φ 𝑥 A well-known inequality for the normal Mills’ ratio (Gordon 1941) holds that for x>0,18 (A2) 𝑥≤ 𝜑 𝑥 . 1−Φ 𝑥 Applying this inequality, it is straightforward to show that (A1) holds for the normal distribution. This also implies that lim!→! 𝜖 = ∞. By inspection, 𝜖 0 = 0. b. Exponential distribution 𝑓 𝑝 ≡ 𝜆𝑒 !!" , 𝐹 𝑝 ≡ 1 − 𝑒 !!" , 𝜖 𝑝 = 𝜆𝑝, 𝜆, 𝑝 > 0. Then 𝑓! 𝑓 + = −𝜆 + 𝜆 = 0 𝑓 1−𝐹 so (A1) holds. By inspection, 𝜖 0 = 0 and lim!→! 𝜖 = ∞. c. Uniform distribution 𝑓 𝑝 ≡ 1 , 𝑏 𝐹 𝑝 ≡ 𝑝 , 𝑏 𝜖 𝑝 = 𝑝 , 𝑏−𝑝 so that 𝑓! 𝑓 1 + = > 0. 𝑓 1−𝐹 𝑏−𝑝 By inspection, 𝜖 0 = 0 and lim!→! 𝜖 = ∞. 18 I present the inverse of Gordon’s inequality. 0<𝑝<𝑏 d. Lognormal distribution 𝑓 𝑝 ≡ 1 𝜑 𝑥 , 𝑝𝜎 𝐹 𝑝 ≡Φ 𝑥 , 𝜖 𝑝 = 1 𝜑 𝑥 𝜎 1−Φ 𝑥 , 𝑥≡ ln 𝑝 − 𝜇 𝜎 so that 𝜕𝜖 𝑝 𝑓′ 𝑓 1 1 𝑥 𝜑 1 =𝜖 + + =𝜖 − − + + . 𝜕𝑝 𝑓 1−𝐹 𝑝 𝑝 𝑝𝜎 𝑝𝜎(1 − Φ) 𝑝 Cancelling terms and using Gordon’s inequality, this is positive. And taking the limit of Gordon’s inequality, lim!→! 𝜖 = ∞. By inspection lim!→! 𝜖 = 0. Historical data sources I obtain data on production employees for cotton and steel from Lebergott (1966, see also US Bureau of the Census 1975) through 1950, and from 1958 on from the NBERCES manufacturing database for SIC 2211 and 2221 (broadwoven fabric mills, cotton and manmade fibers and silk) and SIC 3312 (primary iron and steel). The former measures the number of wage earners while the more recent measure production employees. I find that these series are reasonably close for overlapping years. For 1820 in cotton, I estimate 5,600 full time equivalent workers producing in households, using estimates of household production and Davis and Stettler’s (1966) estimates of output per worker. For the auto industry, I use the BLS Current Employment Statistics series for motor vehicle production workers from 1929 on. For 1910 and 1920, I obtained the number of employees in the motor vehicle industry from the 1% Census samples (Ruggles et al. 2015) and prorated those figures by the ratio of BLS production workers to Census industry employees for 1930. Weekly hours data for motor vehicles also come from the BLS from 1929 on. For earlier years and for cotton and steel before 1958, I use Whaples (2001) before 1939, linearly interpolating for missing year observations. From 1939 to 1958 I use the BLS Current Employment Statistics series for manufacturing production and nonsupervisory personnel. In cotton and steel, I use the NBER-CES data for production hours from 1958 on (this comes from the BLS industry data). For cotton production, I begin with Davis and Stettler’s (1966, Table 9) estimates of yards produced per man-year for 1820 and 1831 multiplied by the estimate of the number of cotton textile wage earners for those years (I assume productivity was the same in 1830 and 1831). For 1820, I estimate that an additional 9.6 million yards were produced in households based on data from Tryon (1917). From 1830 on, Tryon’s estimates indicate little cotton cloth was produced at home. From 1840 through 1950, is use estimates of the pounds of cotton consumed in textile production time three yards per pound (US Bureau of the Census 1975 and Statistical Abstracts, various years). This ratio is the historically used rule of thumb, but I also found that it applies reasonably well to a variety of twentieth century test statistics. While some cotton is lost in the production process (5% or less typically), these losses changed little over time. From 1930 on, I also include the weight of manmade fibers consumed in textile production. From 1958 on I found that the deflated output of SIC 2211 and 2221 in the NBER-CES tracked the pounds of fiber consumed closely for the ten years when I had measures of both. I used this ratio to estimate yards of cloth produced based on the NBER-CES real output from 1958 on. For steel, my output measure is the short tons of raw steel produced (Carter 2006). From 1913 through 1950, I measure motor vehicle production using the NBER Macrohistory Database series on passenger car and truck production. I obtained a figure for 1910 production from Wikipedia.19 From 1951 on, I use car and truck production figures from the Ward’s Automotive Yearbook, prorated to match the NBER series. For consumption of motor vehicles, I use the Ward’s Automotive series on sales of passenger cars and trucks. For cotton and steel, I add net imports to domestic production. For cotton from 1820 through 1950, I use the net dollar imports of cotton manufactures divided by the price of cloth. From 1820 through 1860, I use Sandberg’s (1971) estimate of the price of British imports; from 1860 through 1950, I use the price of cotton sheeting (see below). From 1958 on, I use import penetration ratios from Feenstra (1958 though 1994) and Schott (1995 on). For steel, I use Temin’s (1964, p. 282) estimates for steel rail imports from 1860 through 1889. I use the Feenstra and Schott import penetration estimates from 1958 on; I ignore imports between 1890 and 1957. For prices, I use the series on cotton sheeting from 1820 through 1974 (Carter 2006, Cc205); for steel I use series for the price of steel rails, splicing together separate series for Bessemer, open hearth, standard, and carbon steel (Carter 2006, CC244-7). Tables Table 1. Regressions on Log Per Capita Demand 1 2 3 Cotton Ln labor productivity (Ln labor productivity)2 1.05 (.04)*** 1.04 (.04)*** -0.13 (.01)*** -0.13 (.01)*** Ln p/w -0.96 (.06)*** (Ln p/w)2 -0.19 (.02)*** Ln real GDP/capita Number observations Adjusted R-squared 52 31 52 0.981 0.954 0.981 Exogeneity P-value 0.096 Steel Ln labor productivity (Ln labor productivity) 2 -2.89 (.29)*** -3.19 (.37)*** -0.65 (.05)*** -0.69 (.06)*** Ln p/w 1.44 (.60)** (Ln p/w)2 -0.25 (.06)*** Ln real GDP/capita Number observations Adjusted R-squared 35 34 35 0.977 0.968 0.977 Exogeneity P-value 0.134 Automotive Ln labor productivity -3.55 (.78)*** -5.26 (1.22)*** (Ln labor productivity)2 -0.45 (.07)*** -0.62 (.12)*** Ln real GDP/capita Number observations Adjusted R-squared 61 61 0.904 0.891 Exogeneity P-value 0.073 Note: Robust standard errors in parentheses. ***= significant at 1%; ** = significant at 5%; * = significant at 10%. Constant term not shown. Column 3 is an instrumental variables GMM estimation using year and yearsquared to instrument the labor productivity terms. The reported probability value is for the null hypothesis that these variables are exogenous. Table 2. Regressions of Per Capita Demand 1 Independent variable 2 Labor productivity 3 Labor productivity Price / wage A. Cotton cloth, 1820 - 1995 𝜇 𝜎 𝛾 𝛼 Observations R-squared -1.74 (.10)*** -1.51 (2.35) 1.29 132.94 (.15)*** 1.36 (.69)* (3.2)*** 133.09 -0.06 (3.48)** -1.49 (.79)* 2.04 (.58)** 184.42 (49.15)** (.58) 52 52 37 0.993 0.993 0.990 B. Raw steel, 1860 – 1982 𝜇 𝜎 𝛾 𝛼 Observations R-squared 3.24 (.14)*** 2.31 (.80)*** 3.60 (.86)*** 0.77 (.18)*** 0.56 (.23)** 1.46 (.40)*** 0.68 (.05)*** 0.69 (.05)*** 1.32 (.64)** 0.26 (.22) 35 35 116 0.982 0.982 0.958 C. Motor vehicles, 1910 – 2007 𝜇 𝜎 𝛾 𝛼 Observations 5.30 (.06)*** 7.32 (1.63)*** 0.49 (.10)*** 1.61 (1.17) 64.60 (4.55)*** 81.13 (24.79)*** -0.80 (.68) 61 61 R-squared 0.984 0.985 Note: Non-linear least squares estimates of equation (4) in columns 1 and 2 and equation (5) in column 3. Robust standard errors in parentheses; ***= significant at 1%; ** = significant at 5%; * = significant at 10%. Table 3. Annual Growth Rate of Industry Hours Worked Sample Computer use x nonmanufacturing Computer use x manufacturing Manufacturing 1 2 3 4 5 OLS OLS FE IV IV All All All Non-mfg. 3.00 (1.89) 2.30 (2.87) -9.09 (4.37)** 2.73 (1.82) Offshoreability -10.99 (4.70)** -5.23 (5.10) Mfg. 6.05 (1.71)*** 0.49 (3.19) -23.54 (5.89)*** 4.09 (1.97)** -2.57 (1.10)** College required 7.00 (2.43)*** No. observations 1299 1299 1299 799 R-squared 0.024 0.028 0.138 0.016 Prob 𝛽!"!#$!% = 𝛽!"#$ 0.010 0.006 0.003 464 0.012 0.124 Percentage contribution to annual employment growth Nonmanufacturing Manufacturing 1.27 0.97 -2.22 -3.48 -4.20 -9.00 2.53 0.19 Note: Errors, clustered by industry, in parentheses. ***= significant at 1%; ** = significant at 5%; * = significant at 10%. Data are for 1984, 1989, 1993, 1997, 2001, and 2003 (extending to 2007 for the dependent variable). The dependent variable is the annual growth rate in hours worked from the observation year to the next year in the sample trimmed of 1% tails. Year dummies not shown. The fixed effects regression is over 227 detailed industries. The regressions in columns 4 and 5 are instrumented using year dummies and the share of workers in sedentary jobs based on the Dictionary of Occupational Titles of 1977 (see text) and the 1980 Census. Probability values are shown for F tests that the manufacturing coefficient equals the nonmanufacturing coefficient.