Automation and Jobs - When Technology Boosts Employment Ch 2

Empirical Findings Prices and Productivity In the stylized model, productivity equals wages over price. This is convenient for empirical analysis because it means the key independent variable can either by or. In practice, the productivity measure is preferable because it is less sensitive to demand shocks and because I have data over a longer time span for productivity and for motor vehicles. For various reasons, might diverge from. For purposes of my empirical analysis, labor productivity can be used as the independent variable as long as it is proportional to wage over price, not necessarily equal.8 Figure 4 shows that this is largely the case for textiles and steel. The correlation coefficients are 0.98 for textiles and 0.88 for steel (0.98 through 1974 only). The only major deviation appears for steel from 1974 to 1982 when, apparently, price inflation exceeded wage inflation. 8 This will be true if the labor share of output is constant or if technical change is Hicks neutral. Figure 4. Is Labor Productivity Proportional to Wage / Price? Demand Curves Before seeking to estimate the model, it is helpful to first examine the demand curves graphically. Figure 5 shows per capita demand (consumption) for each good against labor productivity, both on logarithmic scales. The solid line is simply a log version of equation (4) fit to the data. In Figure 5A and 5B, the circles represent observations where the measure of demand fails to capture the effect of imports of downstream products. Demand needs to take trade into consideration and so I have calculated demand by adding net imports to the amount of product produced domestically. However, for textiles and steel further adjustment is needed because these are intermediate goods industries. The ultimate consumption good is produced by another industry and that good can be imported as well. For example, the consumption of textiles in the form of apparel includes: 1) apparel produced in the US with US cloth, 2) textiles that were imported to the US and used by domestic apparel producers, and 3) apparel that were produced outside the US using cloth also produced outside the US. Even after adjusting for imports of textiles, my measure of consumption misses the cloth imported in apparel made abroad. For this reason, I can only estimate demand for those years where downstream imports are not too large. For textiles, I estimate demand through 1995; in 1996, imports comprised a third of apparel imports for the first time and have grown rapidly since. For steel, I estimate demand through 1982. After that, the largest steel-using industries, fabricated metal products and machinery excluding computers (SIC 34 and 35 excluding 357), show a large increase in import penetration. Between 1982 and 1987, the import penetration (net imports over domestic production) grew 10.5%. As the Figure shows, my measure of per capita consumption falls dramatically around these cutoff years. I also Figure 5. Per Capita Consumption conducted my estimates using different cutoff years as a robustness check. Prediction of employment and the general inverted U pattern were not sensitive to small changes in the cutoff year. Under the assumption that labor productivity is inversely proportional to price, the slope of the curve in the figure represents the price elasticity of demand. In each case, the price elasticity clearly decreases as labor productivity increases. As a first pass, Table 1 formally tests whether the elasticity of demand is constant by fitting the curve with a simple quadratic expression of the form (6) ln 𝐷 = 𝛼 + 𝛽 ∙ ln 𝐴 + 𝛾 ∙ ln 𝐴 ! + 𝑢. Column 1 uses labor productivity as the base independent variable. Column 2 uses price over wage; the labor productivity variable has a somewhat better fit. In each case, the coefficient on the quadratic term is negative and highly significant, rejecting the null hypothesis that the price elasticity of demand is constant. One concern is the possible endogeneity of price and labor productivity. A demand shock will affect both the equilibrium price and the error term, biasing the coefficient of price. A similar concern might relate to labor productivity if firms bring on less productive resources in response to positive demand shocks. In that case, the coefficient of labor productivity might be biased and, possibly, the predictions could be off as well. That problem might be mitigated for the data here because demand shocks are likely small compared to the large changes in demand over historical time frames. To test for endogeneity, Column 3 instruments the labor productivity variables in Column 1 using year and year squared as instruments in a GMM regression. Year should be independent of any demand shocks and the largely linear growth of log productivity seen in Figure 3 means that year should be a good instrument. The coefficients are quite similar for textiles and steel, but a bit different for auto. The table reports the probability value of a statistical test of the null hypothesis that all independent variables are exogenous.9 The null hypothesis that the independent variables are exogenous cannot be rejected at the 5% level of significance, but it can be rejected at the 10% level for auto and cotton. These tests thus raise some concern about endogeneity bias. However, my aim here is not unbiased coefficient estimates but accurate prediction. I compared the predictions of regressions of equation (6) with regressions that used year and year squared instead of the labor productivity variables. The predictions match fairly closely and are highly correlated. The correlation coefficients are 0.971 for textiles, 0.987 for steel, and 0.751 for automotive. The main disparity occurred in the automotive industry for 1910 when labor productivity was far below the 1920 value (thanks to Henry Ford). These tests provide some assurance that predictions from my model do not suffer substantially from the possible endogeneity of labor productivity. Model estimates Table 2 shows NLLS estimates of equations (4) in columns 1 and 2 and estimates of equation (5) for textile and steel in column 3. Columns 1 and 3 set 𝛼 = 0, excluding secondary income effects. All of the regressions have a good fit, although the regressions using labor productivity (columns 1 and 2) fit better than those using the ratio of prices to wages (column 3), probably because of the greater volatility of wholesale price data. Note that the model fits the data significantly better than estimates using a simple quadratic form in Table 1. None of the estimates in column 2 find a significant coefficient for 𝛼, suggesting The test is based on the difference in Sargan statistics; the specification used also shows a strong first stage regression; it is exactly identified. 9 that changes in the marginal value of leisure time are not important.10 The lines in Figure 5 represent the predictions based on the column 1 regressions. One concern is that the model assumes no substantial interference from close substitute products. That means that either there are no substitutes or the productivity growth in substitutes is sufficiently slow that the effect of substitution can be taken as constant. Each industry did have substitutes, especially during the early years. However, it seems that these substitutes were fairly static technologically and were quickly overtaken. Cotton cloth competed with wool and linen. However, wool and linen were mainly produced within the household (Zevin 1971) and did not directly compete in most markets. In urban markets where they did compete, wool tended to be substantially more expensive per pound and its price declined only slowly compared to cotton.11 During the early years of the Bessemer steel process, steel rails were much more expensive than iron rails, but steel rails lasted much longer, making the higher price worth it for many uses. By 1883, the price of steel rails fell below the price of iron rails, eliminating the production of this substitute (Temin 1964 p. 222). And cars and trucks competed with horse drawn vehicles during the early years. However, here, too, production of horse drawn vehicles collapsed very quickly.12 Since the 1970s, steel may have faced greater competition from aluminum and other materials for use in cars and cans (Tarr 1988 p. 177-8), perhaps contributing to the poorer fit of the model then. In any case, the close fit of the model overall suggests that substitution is not a significant problem. 10 Of course, leisure time increases dramatically over this historical period studied. For example, in Philadelphia in 1820, wool was $0.75 per pound while cotton sheeting was $0.15 (US Bureau of the Census 1975). 11 The production of carriages, buggies, and sulkies fell from 538 thousand in 1914 to 34 thousand in 1921; the production of farm wagons, horse-drawn trucks, and business vehicles fell from 534 thousand in 1914 to 67 thousand in 1921 (US Bureau of the Census 1975). 12 Using these predictions, I estimate the price elasticity of demand at each end of the estimation sample: Cotton Year Elasticity 1810 2.13 1995 0.02 Steel Year Elasticity 1860 3.49 1982 0.16 Automotive Year Elasticity 1910 6.77 2007 0.15 Demand was initially highly elastic, becoming highly inelastic. These predicted levels of per capita demand can also be used to estimate industry production employment by dividing domestic demand (total demand divided by 1 + import penetration) by the annual output per production worker. Measuring labor productivity as output per production worker-hour this is13 𝐷𝑒𝑚𝑎𝑛𝑑 𝑝𝑒𝑟 𝑐𝑎𝑝𝑖𝑡𝑎 ∙ 𝑃𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛 1 ∙ . 1 + 𝐼𝑚𝑝𝑜𝑟𝑡 𝑝𝑒𝑛𝑒𝑡𝑟𝑎𝑡𝑖𝑜𝑛 𝐿𝑎𝑏𝑜𝑟 𝑝𝑟𝑜𝑑𝑢𝑐𝑡𝑖𝑣𝑖𝑡𝑦 ∙ 𝐻𝑜𝑢𝑟𝑠 𝑤𝑜𝑟𝑘𝑒𝑑/𝑦𝑒𝑎𝑟 These estimate are shown as the solid lines in Figure 1. The estimates appear to be accurate over long periods of time. There are notable drops in employment during the Great Depression and excess employment in motor vehicles during World War II. Finally, employment drops sharply for the years when my measure of consumption fails in textiles (after 1995) and steel (after 1982). It appears that this simple model using a lognormal distribution of preferences provides a succinct explanation of the inverted U in employment in these industries. 13 For 1820 and before, I also subtract the estimate of labor performed in households. Implications Trade vs. technology in manufacturing job losses The model provides an estimate of the impact of technology on industry employment as mediated by demand. We can use this to understand how much technology contributed to the loss of manufacturing jobs compared to other factors including trade. Using predicted industry employment without the correction for imports, actual and predicted employment changes can be compared: Cotton & synthetic textiles Production workers (1000s) Steel Production workers (1000s) Year 1950 Actual 350 Predicted 336 Actual 550 Predicted 514 1995 120 134 2011 16 98 100 153 Job losses 1950 - 2011 334 238 71% 450 361 Job losses 1995 - 2011 104 36 35% Tech share Tech share In both textiles and steel, most of the jobs have disappeared since 1950 and most of the losses can be attributed to growth in labor productivity without compensating growth in demand. Other factors, including trade, the recession, and, perhaps changing tastes, can account for only 29% of the job losses in textiles and 20% in steel. That said, technology does not account for much of the more recent losses especially in textiles. Since 1995, only about one third of the job losses can be explained by growing labor productivity. Most of the loss in textile manufacturing jobs appears to be the result of the collapse of the domestic apparel industry in the face of heavy global competition. Note that the effect of trade on textile and steel manufacturing jobs does not appear to be mainly 80% about imports of textiles and steel—the import penetration in textiles was 8.9% in 2005 and in steel it was 5.2%. Instead, the main impact of trade appears to have come through its effect on downstream industries.