Wage Inequality in American Manufacturing, 1820-1940 New Evidence

Wage Inequality in American Manufacturing, 1820-1940: New Evidence 1 Abstract: The consensus view among economic historians is that wage inequality in American manufacturing followed an inverted-U path from the early nineteenth century until just before World War Two. The previous literature, however, has been unable to fully document this path over time, or fully assess the role of explanatory factors such as changes in firm organization and technology. We provide fresh evidence that allow us to better document the inverted U as well as its causes. In the first part of the paper, we use the Department of Labor’s Hand and Machine Labor Study to show that wage inequality within manufacturing establishments rose over the nineteenth century. In the second part, we use information from Massachusetts on construct a new time series showing that wage inequality among production workers declined from the 1890s to the late 1930s. For both periods technology played a role – the shift to steam-powered factories over the nineteenth century raised wage inequality, whereas electrification after 1900 reduced it. JEL Codes: N31, N32, N61, N62 1.0 Introduction The historical evolution of economic inequality is a central topic in US economic history as well as public policy today. Much of the discussion of early inequality has focused on wealth and its generating mechanism. Here, we present new evidence on wage inequality in manufacturing from the early nineteenth century to just before World War Two given that wage labor was an economic opportunity open to most. This period is also particularly interesting as it covers the transition from the artisan shop with its skilled craftsmen to the steam-powered factory with its semi-skilled operatives—that is, the first industrial revolution; and, subsequently, the broad-based replacement of steam with electricity marking the second industrial revolution. 2 In other words, our period collectively embraces key features of long-term development and growth of the American manufacturing before the decline in her dominance. As Frederick C. Mills would remark in his introduction to Jerome’s study of mechanization, “mechanization … in all its countless manifestations reacts upon the volume of employment, the skills and the working methods of the human factors of production” (Jerome 1934, xxiii). The consensus view is that wage inequality in manufacturing followed an inverted-U path over our period of study (Lindert and Williamson 1976, especially Figure 1; Williamson and Lindert 1980). The rising portion of this inverted-U occurred from the early to late nineteenth century, and the decreasing from the turn of the twentieth century until just before World War Two (Atack et al. 2004; Goldin and Katz 1999; Goldin and Katz 2008; Katz and Margo 2014; Margo 2000). Rising inequality has been attributed to “hollowing-out” – growing relative demand for factory operatives in the lower tail of the wage distribution, and for white-collar nonproduction workers at its upper tail – at the expense of skilled blue-collar workers in the middle (Atack et al. 2004; Goldin and Sokoloff 1984; Katz and Margo 2014). The falling portion of the inverted-U is now explained by the rapid growth in the supply of skilled labor due to the “high school movement” at the top end and the emergence of “technology-skill complementarity” as electrification reorganized factory production, reducing the demand for low-paid factory jobs at the bottom (Goldin and Katz 1999; Goldin and Katz 1998; Goldin and Katz 2000; Goldin and Katz 2008; Gray 2013; Lafortune et al. 2019). Since wage labor figures prominently in explanations for both the rise and fall, economic historians should be able to document the inverted-U shape with comprehensive time series on the full distribution of manufacturing wages. Unfortunately, the available evidence falls short of the ideal, especially earlier in the nineteenth century. Instead, scholars have relied heavily upon 3 occupation-specific average wages to produce time series of “skill differentials” – ratios of wages of skilled blue-collar or white-collar to unskilled/semi-skilled workers – which broadly trace out an inverted-U (Autor et al. 2020; Goldin and Katz 1999; Goldin and Katz 2008; Goldin and Margo 1992; Lindert and Williamson 1976; Margo 2000; Williamson and Lindert 1980). While such series are highly useful, their correlation with temporal movements in overall wage inequality cannot be established directly for the pre-World War Two period. In recognition, there have been some attempts to go further using data from the federal censuses of manufacturing and related reports. For the rising portion of the inverted-U shape, Atack, Bateman, and Margo (2004) used data from samples of the 1850-1880 manufacturing censuses (Atack and Bateman 1999) to show that the distribution of the “establishment wage” – the average wage in individual establishments – became increasingly unequal over time, primarily because of an increase in the employment density of workers in establishments paying lower-than-average wages. While this pattern is consistent with rising wage inequality overall, it presumes that there were no offsetting changes in wage inequality within establishments, which cannot be tested with their data (Atack et al. 2004). For the falling portion of the inverted-U, Goldin and Katz (1999) estimate wage quantiles for a panel of ten industries in 1890 and ca. 1940, from which they construct standard inequality ratios (50-10, 90-50, and 90-10). These ratios are considerably smaller just before World War Two than in 1890, consistent with a reduction in overall inequality. However, as they point out, this inference presumes no offsetting changes in inter-industry wage differentials. Moreover, with just two endpoints, it is impossible to determine when the change occurred. In addition, existing studies provide, at best, limited direct evidence on the role of explanatory factors within manufacturing, such as the increasing division of labor and the declining importance of physical 4 strength with increasing use of inanimate sources of power such as steam earlier in the period (Atack et al. 2004) and electricity, later (Goldin and Katz 2008). Here, we present new evidence on wage inequality in historical manufacturing from two sources. The first source is recently digitized data from the US Department of Labor’s 1899 Hand and Machine Labor (hereafter, HML) study (U.S. Department of Labor. 1899) which provides extremely detailed data at the production operation level for two different modes, “hand” (artisan) v “machine” (mechanized factories) in the manufacturing of specific goods, called “units” in the study.1 In previous work (Atack et al. 2019; Atack et al. 2022) we have shown that empirical differences between hand and machine labor in the HML study faithfully reflect changes over time in the evolution of nineteenth century manufacturing from artisan to factory production. We use these data to document the impact of these changes had on wage inequality, something not previously possible. We begin our analysis of the HML study data by computing the standard deviation of (ln) labor cost (wage of the worker per standard unit of time) across all operations at the unit level (see above) separately for hand and machine labor production, which becomes the main variable of interest. Compared with hand labor, machine labor production utilized a greater division of 1 A unit in the HML study refers to two sequences of production operations making a precisely defined (in terms of characteristics, quantity, and quality) product, one using, one using the traditional, “hand” (aka artisan) methods, and the other the most advanced factory (“machine”) methods then available. Within this “unit,” the Commissioner of Labor staff trace and match production steps. While these data are not, strictly speaking, establishment level, they are sufficiently related to provide reliable evidence on the differences in, for example, labor productivity that resulted from the long-term shift from the artisan shop to the factory system (Atack et al. 2022). In order to make sure that these data were comprehensive, field agents actually secured parallel data from two different producers using each method selecting that they judged to be “the better and more complete” (U.S. Department of Labor. 1899, p. 13). 5 labor, which required more workers, and greater use of special purpose machinery, powered inanimately, mainly by steam (Atack et al. 2022). We sketch a simple model under which greater division of labor will lead to greater dispersion of wages within establishments, and hence a higher level of wage inequality. Wage inequality may increase further due to mechanization, as greater use of steam power raised the relative demand for unskilled – and hence, low-wage – labor since the machines being driven by steam required less operator skill than the use of traditional hand tools. Our main empirical result from the HML analysis is that wage inequality was approximately 90 percent higher, on average, within the machine labor units than the hand labor units, a difference that is economically and statistically significant. Regression analysis includes unit fixed effects, and so is equivalent to differencing between machine and hand labor, holding the product (the unit) constant. We find that wage inequality was increasing in the use of inanimate power and in the division of labor. On average, use of inanimate power was much higher under machine labor than hand labor, as also was the division of labor. However, in terms of explanatory power, greater division of labor was the more important factor, accounting for approximately twice as much of the higher wage inequality in machine labor production than mechanization. Our analysis of the HML study data suggest, therefore, that the transition from the artisan shop characteristic of the early nineteenth century to the mechanized factory of the late nineteenth century was associated with an increase in wage inequality, due to changes in the organization of production that occurred. Our second source of evidence is newly digitized data from state government reports on manufacturing wages in tables of so-called “classified” wages. These give the number of workers whose weekly earnings were categorized (“classified”) into bins – for example, from 6 $8.00 to $9.00 per week. These are the same type of data studied by Goldin and Katz (1999), except that our source and reference frame is different – the state vs. the federal government covering many years. Specifically, we use the most comprehensive series, those the state of Massachusetts which cover from the late 1880s to the late 1930s.2 We have also digitized data on electricity use in Massachusetts manufacturing which the state collected for select years. Our main analysis of the Massachusetts data is based on estimates of the 10th, 25th, 50th, 75th, and 90th quantiles of the classified distribution of weekly earnings for production workers. The quantile estimates are made following the general procedure outlined in Blalock (1960, p. 55; also used by Goldin and Katz 1999). We present the actual series along with smoothed estimates derived from non-parametric regressions. Our initial estimates show a marked reduction in wage inequality in Massachusetts manufacturing over the study period. This was present in the lower tail (50-10 ratio), upper tail (90-50), and the interquartile range (75-25), but quantitatively largest in the lower tail. Moreover, the decrease in lower tail inequality was modest in the 1890s but accelerated after the turn of the century. 2 The 1890 federal census of manufacturing (U.S. Census Office. 11th census 1895) reported such distributions, by industry, for establishments in select cities, which Goldin and Katz (1999) used in their work. The Massachusetts data come from the state’s “Annual Statistics of Manufactures” beginning in 1886 (An Act Relating to the Annual Collection of Statistics of Manufactures 1886, p. 135). In addition to the Massachusetts reports, we have located similar data in reports for ten additional states (California, Illinois, Iowa, Kansas, Massachusetts, New Hampshire, Ohio, Rhode Island, Texas, and Wisconsin). Unlike the Massachusetts reports, however, these cover much shorter periods of time (for example, New Hampshire in just 1916). In discussing the state of information regarding wages in the United States, Nearing (1914, p. 14) lamented “of the 10 leading industrial states, but three--Ohio, Massachusetts, and New Jersey, -furnish wage data, which merits … comment. The statistics for Ohio are excellent, but very diffuse and unconcentrated. The statistics for Massachusetts and New Jersey are, on the other hand, scientifically classified, accurately presented, and in every sense satisfactory and reliable.” 7 In explaining the decrease in inequality between their 1890 observation and their ca.1940 one, Goldin and Katz emphasize two factors: the “high school” movement, which compressed the top half of the wage distribution (by reducing the returns to skill, as embodied in education), and electrification, which eliminated many of the low wage jobs that were associated with the use of steam power such as shoveling coal (see also, Jerome 1934). The data for Massachusetts allow us to shed further light on the role of electrification. First, we generate a time series of the percent of horsepower used in Massachusetts manufacturing derived from electricity. While this series is not sufficiently frequent to use for time-series regression, it conclusively shows electrification began in earnest after the turn of the twentieth century, accelerating after 1910. In 1895 only about 1.5 percent of horsepower used by Massachusetts manufacturing was provided by electric motors. This share had increased to a third in 1920, and then to over 60 percent by 1938. To examine the impact of electrification on wage inequality, we have compiled an industry-level panel for Massachusetts for two years (1895 and 1920), which covers 42 industries. For each industry-year pair, we estimate the 50-10 ratio, which becomes the dependent variable. There are three independent variables – the male share of workers, electric horsepower per worker, and non-electric horsepower per worker. We estimate the regression in first differences, which is equivalent to two-way fixed effects (industry dummies and a year dummy for 1920). We find a statistically significant, negative effect of electricity on the 50-10 ratio, consistent with Goldin and Katz’s argument. The effect is large – at the sample means, we can account for nearly all (80 percent) of the decrease in the 50-10 ratio between 1895 and 1920. The remainder of the paper is organized as follows. Section 2 summaries the existing literature on the evolution of wage inequality in manufacturing in the United States over the 8 period of study. Section 3 presents our analysis of the HML study data. Section 4 offers our analysis of the data from the Massachusetts reports. Conclusions are presented in section 5. 2.0 Literature Review: Wage Inequality in Manufacturing, 1820-1940 In 1800 less than one percent of the American labor force was engaged in manufacturing. By the middle of the century, the manufacturing share had reached 15 percent, and it rose to 20 percent by 1900 (Atack et al. 2004, 172). The share continued to rise over the first half of the twentieth century reaching slightly more than a quarter of the labor force in 1950. It has declined since, so much so that by 2010, the share (10 percent) was lower than in 1850. The initial shift of labor into manufacturing coincided with massive changes in industrial organization. At the start of the nineteenth century, manufacturing took place in artisan shops (see, for example, Sokoloff 1982; U.S. Department of the Treasury and U.S. Congress House 1791). In such shops, the owner either worked alone or with a partner, and perhaps a few apprentices. Except for a few industries such as lumber and grist mills that used waterpower, capital was limited to basic, general-purpose hand tools, plus the building. In terms of numbers, these artisan shops would remain dominant at mid-century but, increasingly, the share of production shifted towards larger establishments—places termed factories (U.S. Census office. 10th census et al. 1883). These differed from artisan shops in three fundamental ways – a greater use of division of labor; more capital per worker; and, relative to the artisan shop, substantially greater use of inanimate power. Initially the power source was water but as the century progressed, steam power was used (Hunter 1979; Hunter 1985). By the end of the nineteenth century the factory system was dominant, and it continued to grow in importance during the first half of the twentieth century, but steam was replaced by electricity as the inanimate power source (Devine 1983; Hunter and Bryant 1991). Overall, the consensus view is that these changes in 9 industrial organization dramatically increased output per worker in manufacturing such that, by World War I, the United States was leading industrial economy in the world (Bairoch 1982; Broadberry and Irwin 2006). Although the success of American manufacturing at overtaking early industrializers like the British is undeniable, there is less agreement on how the shift from the hand labor of the artisan shop to the machine labor of the factory affected wage inequality in manufacturing. There are two strands to the literature. The first strand measures changes in manufacturing wage inequality using time series of skill differentials – ratios of wages of skilled to unskilled workers – which are interpreted through the lens of (relative) demand and supply. There is a substantial body of work documenting the evolution of skill differentials over the nineteenth century (Autor et al. 2020; Goldin and Katz 2008; Margo 2000; Williamson and Lindert 1980). The consensus view is that, relative to unskilled labor, the wages of skilled artisans were roughly stable from the early to late nineteenth century, while the relative wages of white-collar workers increased modestly. Katz and Margo (2014) document that, over the same period, the share of artisans in manufacturing declined, while the shares of operatives (including unskilled labor) and whitecollar workers staffing front offices increased. They interpret these shifts as evidence of “hollowing-out” – a shift in labor demand in manufacturing towards operatives and managerial labor and away from artisans. However, if this was the case, one might have expected the relative wage of artisans to have declined relative to unskilled labor, but it did not. Katz and Margo point out that, while relative demand decreased in manufacturing, the construction sector, which was intensive in the use of artisan labor, grew over the century and took up the slack. 10 Although one can craft a sensible narrative around the evolution of skill differentials, such time series offer, at best, a limited window on overall movements in inequality. Because of this limitation, research has turned to evidence from the various nineteenth century federal censuses of manufacturing, several of which collected data sufficient to compute the “establishment wage” – the average wage of workers employed in the establishment. Establishment-level samples from the original manuscript records are available for 1820, and for 1850-1880 (Atack and Bateman 1999; Sokoloff 1982). Unfortunately, samples are unavailable for other years either because relevant data were not collected (1830, 1840) or because the original returns have not survived.3 Atack, Bateman, and Margo (2004, hereafter ABM) use samples for 1850 and 1880 to construct measures of inequality in monthly establishment wages. In Table 1 we expand their original figures, by providing inequality statistics for 1860 and 1870 (monthly), and for annual wages in 1820, 1870, and 1880. These statistics are the 10-50 and 50-90 differentials in the log establishment wage, the coefficient of variation, and (the closely related) standard deviation of the log establishment wage. It is important to note that all measures in the table are weighted by the number of workers at the establishment, as was the case in the original ABM study. Sample sizes (number of establishments and number of workers) are also shown. As a further point of 3 The 1810 census is generally judged to be too poor in quality to be useful and, in any case, the original records were destroyed by the British in 1814 (Fishbein 1973). For 1820 we use the digital file originally created by Sokoloff (1982), which is a complete count of all establishments in 45 randomly chosen counties in the Northeast; this is compared with observations from the Northeast in 1870-80 (see Table 1). No census was taken in 1830 and the Census of 1840 collected no data on wages. Records of the 1890 and 1900 censuses were destroyed in the early twentieth century. The Treasury Department’s McLane Report for 1832 (U.S. Congress. House. 1833) contains wage information for manufacturing and were digitized by Sokoloff and Villaflor (1992) to estimate establishment wages in the Northeast, but their computer file is no longer extant. 11 comparison, we also show the coefficient of variation of annual establishment wages for 1919, based on a proprietary sample from the 1920 census analyzed by Brissenden (1929).4 As can be seen in the table in the columns labelled “Full Sample” (which have the largest sample sizes), there was a steady upward trend in inequality in monthly establishment wages from 1850 to 1880, most of which occurred due to a rise in the 10-50 differential, that is, a rise in inequality below the median wage. The evidence is much more limited for annual establishment wage inequality, but the general pattern remains an inverted-U, rising from 1820 to 1880, and then falling at some point after this, as indicated by the (much) lower level of inequality in 1919 than in 1880. To understand the increase in establishment wage inequality between 1850 and 1880, ABM analyzed the establishment wage data in a two-step procedure. In the first step, ABM estimated regressions of the log of the establishment wage on establishment characteristics. Their main findings were twofold – first, the establishment wage was a negative function of the number of employees; and second, that the establishment wage was a positive function of the log of capital per worker and of the use of steam power. ABM interpreted these patterns using a simple framework borrowed from Goldin and Katz (1998) – the negative correlation with the number of workers is consistent with a lower average skill level as the number of workers Brissenden’s estimates did not pertain to a nationally representative random sample of establishments but instead to a sample from 8 cities in approximately 20 industries. Due to sample size limitations the best match we can make to Brissenden back to 1850-1880 is to restrict the analysis to urban observations in the two-digit SIC industries which contain those studied by Brissenden. It is not possible to match Brissenden to 1820 for conceptual reasons (several of his cities did not exist in 1820) and because the available sample size is too small. 4 12 increases – de-skilling – whereas the positive effect of capital intensity and use of steam power can be read as both variables boosting the demand for skilled labor in these establishments. In the second part of their paper, ABM use the data to compute a decomposition of the overall change in establishment wage inequality between 1850 and 1880. This decomposition is performed on the 10-50, 50-90, and 10-90 differential in the log establishment wage, in terms of the portion explained by changes in the distribution of the independent variables in the regressions, the regression coefficients, and residual wage inequality. The regressions used for this purpose contains just establishment size dummies. The results show that, overall, establishment wage inequality increased substantially between 1850 and 1880 because of a growing concentration of employment in establishments with below average wages. These were relatively large establishments in terms of the number of workers and the reason they were low wage on average is because the establishment wage was a decreasing function of the number of workers. ABM argue that this is consistent with a relative demand shift in favor of less-skilled operatives, who increasingly dominated manufacturing employment as the factory system grew in importance. There are three significant limitations to ABM’s analysis. First, none of the extant samples from the federal censuses of manufactures provide direct evidence on the division of labor; the presumption is that such division was an increasing function of the number of workers employed by the establishment, but the census cannot be used to verify this directly. Second, while the increase in establishment wage inequality is clear enough, for wage inequality to also have increased across workers – as opposed to across establishments on average -- it is also necessary that wage inequality within establishments did not decrease. ABM argue this was so but have no direct evidence on the evolution of wage inequality within manufacturing 13 establishments. Third, while ABM show that average establishment wages were increasing in the use of steam power, the effect of the shift to steam on wage inequality remains unclear. The second component of the consensus view is that wage inequality in manufacturing decreased from the turn of the century to 1940. Here, the main analyses are provided by Goldin and Katz (1999; 2008). Specifically, Goldin and Katz (1999) compare adjusted wage distributions for manufacturing operatives in particular industries reported in the census of manufacturing data for 1890 (U.S. Census Office. 11th census 1895, Part 2) with similar distributions from BLS surveys in the late 1930s. These comparisons show a decrease in wage inequality within the given industries over the period. Goldin and Katz attributed some of the decrease in wage inequality to electrification, which reduced the demand for very low skilled workers in manufacturing. Annual times series of skill differentials over the period suggest that the bulk of the decrease in wage inequality occurred prior to 1920, is consistent with the timing of electrification. There are, however, two main limitations to Goldin and Katz’s argument. First, they acknowledge that their evidence on wage inequality is within-industry. It is thus possible that overall wage inequality in manufacturing was rising (or stable) if inter-industry wage differences were increasing, although they discount this explanation. Perhaps more importantly, the precise timing of changes in wage inequality cannot be determined from Goldin and Katz’s analysis of wage inequality with just a starting and ending date. Relatedly, Goldin and Katz present no evidence directly linking rising use of electricity within manufacturing to decreases in wage inequality across manufacturing workers. Our analyses in the next two sections are our proffer addressing the limitations of both ABM and Goldin and Katz. 14 3.0 The Hand and Machine Labor Study and Wage Inequality within Manufacturing Establishments The student of American manufacturing who wishes to study the evolution of wage inequality within establishments in the nineteenth century has very limited choices for evidence. The Weeks or Aldrich reports (U.S. Congress. House et al. 1883; U.S. Congress. Senate 1893), which have been previously used to construct wage series, are one possibility, as both are establishment based. However, while these reports do contain some within-establishment information on wages, the evidence is limited to occupation averages and thus does not fully capture the variation at issue.5 Another possibility is the 1900 census report on Employees and Wages (United States. Census Office. 12th census 1900. and Dewey 1903), which provides classified wage distributions (see the next section) at the establishment level in the late nineteenth century for a subset of industries in groups of states. However, this report, as well as the Aldrich or Weeks data, provides no direct evidence on the underlying causal factors at issue, namely the division of labor and mechanization. The source that we use, the U.S. Department of Labor’s Hand and Machine Labor Study (1899, hereafter HML) is quite different. This source contains information on wage inequality pertaining to production in large, mechanized factories – the typical production unit ca. 1900 – as well as wage inequality within small non-mechanized artisan shops. These were the typical production entity earlier in the century, although about a quarter of the observations were from businesses still in operation in the 1890s. Differences in wage inequality between these two 5 Wesley Mitchell (1903) reports an 1860 classified wage distribution in manufacturing. However, this distribution—based on underlying data from the Aldrich report pertaining to establishment-occupation averages—fails to capture fully the within-establishment variance. 15 types of production – “machine” (factory) vs. “hand” (artisan) labor provide important insights, we argue, into changes in wage inequality within establishments over the nineteenth century, as well as the causal factors behind them. Published in two volumes totaling almost 1,600 pages, the HML study detailed the production operations involved in the manufacturing of what the study termed “units” – specific quantities of precisely defined goods such as “50 dozen regular taper, triangular saw files, 4 inches long, tapering 23/64 inch” (U.S. Department of Labor. 1899, 1: 241-6 and 2: 1026-9). The report covered 672 units in various economic sectors including transportation services, mining and quarrying, and agriculture. Ninety-three percent of the units —626 (units 28-653) of 672 —were manufactures and are our focus here. The products in the study embrace almost the entire range of broadly defined manufactured goods (that is, 2-digit SIC codes 20-39 (United States. Executive Office of the President. Office of Management and Budget. 1987)), including those common in both the first industrial revolution as well as the second. The data, however, should not be viewed as a representative sample of manufacturing at the time (Atack et al. 2022, online Appendix Table 1). For each unit, the HML staff collected production data from multiple establishments that were using “hand labor” or “machine labor”.6 For confidentially reasons, the HML staff anonymized the information in the published report so that we do not know the names of the 6 The product descriptions were identical for the hand and machine labor goods for most units. Differences in the remaining descriptions varied from words like “dairy” v “creamery” or “30 ½” v “30 5/8” but, in all cases, though, the product served the same purpose. The HML staff collected data from two machine labor establishments and then two matching hand establishments, before selecting “the better and more complete” accounting of each mode of production for publication (U.S. Department of Labor. 1899, I, p. 1). In some units, the staff went so far as to trace a significant intermediate input purchased from elsewhere that was used in one type of production but manufactured in house in the other. 16 establishments or, except in few instances, their location.7 The HML staff was clearly aware of the widely held belief that the machine methods yielded a lower quality product than the hand methods and they expended great efforts to find units producing factory goods that were not of inferior quality to artisan products so this argument cannot be used to impeach our results.8 Our procedures in making these data useable are described in Atack, Margo and Rhode (2023; 2019; 2022). In our previous papers, we used the operations level data in the HML study to evaluate modern models of automation (Atack et al. 2019) and to estimate the impact of inanimate power use on the study’s measure of labor productivity, which is the amount of time that it took to complete the same operation – for example, polishing a piece of metal – in the production of the same product by hand and machine labor, but where, under machine labor, inanimate power powering a specific machine might be used (Atack et al. 2022). Specifically, Atack et. al. (2022) demonstrates that the average difference in labor productivity between the hand and machine methods was of a magnitude that accurately tracked the evolution of productivity growth in manufacturing over the century due to the long-term shift from one type of production to the other. 7 The original survey instrument (copies of blank forms are retained in Record Group 257, entry 8, box 3, in the US National Archives) included the name and address of the establishments but these are suppressed in the published study, except in a handful of cases (15) in which information was sought from outside the United States (these cases are excluded from our analysis). For domestically-based units, contemporary newspapers sometimes carried brief notices that Bureau agents were in town conducting the survey (see, for example, Evening Star 1895) but the establishments to be visited were not specifically named and the completed survey forms were subsequently destroyed (U.S. Congress. House 1906). . 8 The text of the HML study discusses quality differences, from which we were able to categorize whether the staff thought the quality was better for product when made by hand or vice versa; or no difference was detected, or no opinion was expressed; see Atack, Margo, and Rhode (2022). 17 The empirical analysis in Atack et al (2022) focuses on the subset of production operations that overlapped between hand and machine labor, thus excluding operations under hand labor that were abandoned as well as novel operations under machine labor. Here, instead we focus primarily on the unit, and the outcome of interest is the standard deviation of the logarithm of “labor cost” across operations within the unit by production mode. For each operation we know the wage of the worker(s) performing the activity, as well as the amount of time that the activity took; labor cost is the wage per unit of time in minutes. Using this information, we can compute the standard deviation, where each operation is weighted by its completion time. The operative assumption is that differences in wage inequality between hand and machine labor can inform the debate over the course of wage inequality within establishments as manufacturing shifted from the artisan shop to the factory, analogous to our previous analysis of labor productivity differences. A key advantage of the HML study for this paper is the information that it contains on potential explanatory factors. Of these, the two of greatest interest are division of labor and use of inanimate power. To motivate our use of this information in the regression analysis below, we sketch a simple framework linking both factors to within-establishment wage inequality. As a point of departure, imagine an artisan shop engaging in hand labor in which all production operations are performed by a single worker; that is, the artisan is a sole proprietor. There are N+1 such operations – N of which are actual production operations and a single overall task of “management” (non-production). We will assume that the opportunity cost of the artisan’s time, say a day’s worth of labor, is w. Because there is only one worker observed wage inequality in these firms is, by definition, zero. 18 Although the artisan is skilled in being able to complete all the myriad production operations, the artisan – like everyone else – will have a comparative advantage at some operations, and a disadvantage at others. Now, imagine that, instead of doing all tasks himself, the artisan specializes in one task – “management” -- and hires N workers, each of whom performs a single operation. We array each of these tasks in terms of the skill required to perform them, and let w(n), be the wage for task n = 1, … N. So long as at least some of the skill levels are strictly increasing (that is, the skill required to perform each of the tasks is not the same), wage inequality will increase through division of labor. In addition, wage inequality may have been affected by the shift to steam power. Here, there are two possible effects. First, the shift to steam may have induced greater division of labor (Atack et al. 2008) among existing tasks. Second, and likely more important, the shift to steam would have created new tasks associated with the operation of steam engines. Some of these tasks – for example, installation and maintenance of steam engines – were highly skilled, while others – moving raw materials (for example coal to fuel it) and intermediate inputs – around the shop floor were unskilled. Indeed, not surprisingly, there is a sharp increase in the need to move intermediate product between stations as the division of labor increased. In either case, wage inequality should have risen. 3.1 Regression Analysis Table 2 reports sample statistics for our regression analysis of the HML study data. We restrict attention to those units in which no inanimate power was used at any point in hand labor and where steam or waterpower was used in machine labor. 9 On average, the standard deviation 9 This sample restriction approximates the comparison that the HML study was attempting to make; see Atack, et. al. (2022). There are 496 units that meet these sample criteria. 19 of ln labor cost was about twice as high in machine labor production than in hand labor production, a level difference of 0.140 that is highly significant (s.e. = 0.013). About 55 percent of the production time under machine labor used inanimate power. On average, machine labor production employed more workers, allocated over many operations, implying a much higher degree of division of labor (Atack et al. 2022).10 27 percent of the hand labor units employed a single worker, compared with hardly any (1 percent) of the machine labor units. As noted, in single worker units, wage inequality was identically zero and there was no division of labor. Regression results are reported in Table 3. All regressions include unit fixed effects. For comparison purposes, the second column reports the coefficient of a dummy variable for machine labor, β = 0.140. This is the mean difference in the standard deviation of ln labor cost (see above). In column 3, we include a) the share of labor time devoted to operations that were mechanized using either steam or waterpower, b) the natural logs of the number of workers, and c) the number of operations. Collectively, these three variables “over-explain” the mean difference (by about 29 percent), as the coefficient on the machine labor dummy is now negative (β = -0.040). The signs of the variables are as expected and two of the three – labor time in powered, mechanized operations and the number of workers – are highly significant. Greater use of inanimate power is associated with greater wage inequality, as is greater division of labor -more workers and more operations. In column 5, we include a dummy variable for units in which a single worker performed all operations; in effect, this tests for a spline at exactly one 10 We include the natural logarithms of the number of workers and the number of operations separately as flexible controls for the division of labor. In a future draft, we plan to test the robustness of the analysis using the protocol described in Atack, Margo, and Rhode (2017) for estimating the proportion of total operations performed by the average worker, which can be computed from the operations level data in the HML study (as opposed to the unit level data used here). 20 worker. This coefficient is highly significant, which has the effect of reducing the magnitudes of the coefficients of the share of production time that was mechanized and the ln of the number of the workers, although the reductions are relatively small. Column 4 computes the “percent explained” of each variable, which is the coefficient multiplied by the difference between machine and hand labor in the mean value of the independent variable, divided by the mean difference in the standard deviation of ln labor cost. If we scale the total contribution of mechanization and division of labor variables by the sum of their respective percents explained, we can compare their relative importance. As can be seen, mechanization accounts for about 28-32 percent of the higher wage inequality of machine labor, depending on the inclusion of the dummy variable for single worker units. In sum, while mechanization contributed to greater wage inequality, division of labor was the relatively more important factor. 4.0. The Evolution of Wage Inequality in Manufacturing, 1890-1940: State Reports on Classified Wages In this section we make use of state level reports on so-called “classified wages” for manufacturing in Massachusetts. Wages are said to be classified because these reports give the number of workers whose wages – typically measured on a weekly basis – fell within given intervals. The first reporting of such statistics appeared in 1888 as a part of the 1885 Massachusetts Census (Massachusetts. Bureau of Statistics of Labor. 1888, pp. 233-62 (towns) and pp. 1115-26 (industries)). What or who motivated this inquiry is unclear, however, it would subsequently become a regular feature of the Massachusetts “Annual Statistics of Manufactures,” beginning with the state’s fifth compilation published in 1891 for 1890 but also covering 1889 21 (Massachusetts. Bureau of Statistics of Labor 1891, pp. 142-61).11 These annual statistics were intended to “fully portray the conditions of the industries” and “accurately show the trend of business from year to year” (Massachusetts. Bureau of Statistics of Labor 1891, p. xiii).12 In all, Massachusetts reported classified wages for 43 individual years between 1885 and 1938, likely inspiring periodic emulation by sister agencies in other states. It also seems plausible (but not certain) that the 1885 Massachusetts inquiry prompted its inclusion as a part of 11 Data collection were authorized by the Massachusetts legislature (An Act Relating to the Annual Collection of Statistics of Manufactures 1886, pp.135-7) beginning April 29, 1886. Data on twelve specific topics (stipulated in Section 1 of the Act) were to be collected by mail of every manufacturing establishment in the state by mid-December with returns due before the end of the year. While the Bureau complied with the law, the first data collected under it was not reported until 1889 (Massachusetts. Bureau of Statistics of Labor 1889). Reasons for the delay were spelled out Commissioner Wadlin (Carroll D. Wright’s successor at the Massachusetts Bureau) in his introduction to his first report where he also provides a summary history of the collection of statistical data by the state (Massachusetts. Bureau of Statistics of Labor 1889, pp. xi-xix). The first classified wage series appeared in 1890 with a column comparing the same establishments in 1889 and 1890. As noted in the text, it is unclear what motivated their collection—or even their precise source as the data that they represent was not among the authorized and stipulated questions (see, for example, Massachusetts. Bureau of Statistics of Labor 1889, pp.xii-xiii). Furthermore, Massachusetts law authorized destruction of the underlying returns on a regular basis—see for example Chapter 31 of 1889 Acts and Resolves. https://archive.org/details/actsresolvespass1889mass, p1365. The Commissioner indicated the intent was to simplify the reporting burden on manufacturers by superseding the state’s census in years ending in “5”. 12 The Chief of the Massachusetts Bureau of Statistics of Labor, Horace G. Waldin, reported that the data collection excluded the smallest (and most numerous) establishments (i.e. was not a census despite the wording of its legal basis) because “the condition of manufacturing in the commonwealth can be accurately portrayed by returns that do not include the small and comparatively unimportant concerns” (Massachusetts. Bureau of Statistics of Labor 1891, p. xix). His claim was apparently based on a comparison between the restricted Massachusetts data and that collected at the Eleventh Census. Access to these data presumably reflected the close collaboration that had long existed between personnel at the Massachusetts Bureau of Statistics and federal census officials beginning with the BLS’s second commissioner, Carroll D. Wright, who played a leading role in the Tenth Census before becoming the first U.S. Commissioner of Labor (North 1909). 22 the 11th Federal Census in 1890, appearing as item #6 of “General Schedule No. 3” of the Census of Manufactures to be asked of all respondents (U.S. Census Office. 11th census. 1895, p. 13; Wright 1900, p. 362). The federal census, however, made little use of these data. No summary tabulation of them by state was published and they only appear in the volume of statistics on cities (U.S. Census Office. 11th census 1895). The federal data did, however, provide the starting point from which Goldin and Katz (1999) estimated wage inequality statistics among adult males at the industry level for ten industries in 1890. These are matched by industry to analogous statistics reported by the U.S. Bureau of Labor Statistics in its Monthly Labor Review (1999, 31). For our purposes there are three advantages of the Massachusetts data. First, and foremost, their temporal coverage far exceeds that of any other state – effectively, the same length of the time, the 1890s to the late 1930s, covered by Goldin and Katz (1999). Second, contemporaries such as Nearing (1914) thought very highly of the quality of the Massachusetts data, which was not the case for all states that attempted to collect information on classified wages. Third, the data refer to production workers only, allowing a cleaner comparison over time, compared with Goldin and Katz’s (1999) use of federal census data.13 Fourth, we can construct an industry panel (see below) that allows us to explore the empirical impact of changes in electrification on wage inequality, as hypothesized by Goldin and Katz (1999). Against these advantages one must keep in mind certain limitations. The Massachusetts industrial structure 13 As Goldin and Katz (1999) discuss extensively in their paper, the federal census data from the 1890 census cover all production and non-production workers. Goldin and Katz make certain assumptions to eliminate the latter from the 1890 distribution, allowing an “apples-to-apples” comparison with the Department of Labor reports from the 1930s, which pertain to production workers. The assumptions are plausible but ideally it would be better to not have to adjust the data at all. 23 differed from the rest of the nation in that had more of the “first industrial revolution” staple industries (for example, textiles, boots and shoes) and less of those in the second, such as automobiles, though its industry data cover a broad swath of activities. While Massachusetts experienced electrification like the rest of the country in its broad contours and timing, the extent of coverage in manufacturing by the late 1930s fell somewhat below that of the rest of the country. Third, and related, a variety of factors local to Massachusetts – for example, the nature of its labor legislation, immigration, and labor strife, among others – may have affected wage inequality in the state’s industries in ways that could have differed from elsewhere.14 The Massachusetts data give the number of workers whose nominal weekly wages fall (are “classified”) into specified wage intervals. The first interval is always bounded below at zero (for example, “Under $3.00”) while the top interval is open, that is, for wages that exceed a certain amount there is no specified upper bound (for example, “$25.00 or more”). In between, there are closed intervals of a specified width -- for example, $8.00 - $8.99, or $12 - $14.99). Figure 1 shows an extract from the table in the 1907 report and a histogram of the same (Massachusetts. Bureau of Statistics of Labor 1908). The classified wage tables are of a kind found in published government documents, historical to the present, in which observational units are categorized into bins. Blalock (1960) is a standard reference for methods applied to binned data to estimate distributional statistics, such as the mean or variance, and quantiles. The approach used by Goldin and Katz (1999) follows Blalock and so we have also used it although it is not perfect. In particular, it assumes that the 14 As we point out in the text, the Massachusetts reports, while very detailed, do not contain sufficient information to allow us to construct a continuous time series for adult males (Goldin and Katz’s (1999) calculations of wage inequality pertain to adult males). That said, we do control for the male percent of workers in our analysis of the industry panel (see the text). 24 distribution of observations within bin intervals is uniform as opposed to, say, being distributed within them in a way that more closely approximates the distribution across bins (see, for example, von Hippel et al. 2017). Using Blalock’s approach, however, it is straightforward to calculate the cumulative distribution function because the total number of observations in the table is always known and the resulting inequality metrics usually compare favorably with those derived from more complex processes (von Hippel et al. 2017).15 Once we know the cumulative distribution function, we can determine which bin intervals include the various quantiles. If the interval containing the quantile is closed at both ends with a non-zero lower bound, the uniform assumption implies that we can calculate the quantile using linear interpolation within the bin. For the 10th quantile, there is the possibility that it will fall into the first interval, which is bounded below by zero. This, in fact, is the case for all years prior to 1906 which “$5.00 or below”. Strictly speaking, we could use linear interpolation, but it is not credible to assume that the support of the distribution is bounded by zero – that is, zero weekly earnings – from below. Instead, we assume that the lower bound of the support is $3.00 week for 1890-1906. The true lower bound of the support is less than $3.00 (though less so over time because of productivity gains and price movements); however, if we assume that, instead, it is $2.00, there is only a minor impact on the estimated values of q10 (the 10th quantile). For the 90th quantile, the estimated values fall into a closed interval in all years except for 1919 and 1920 (where wartime Von Hippel’s analysis (von Hippel et al. 2017) only shows a clear advantage to these other approaches if the true mean of the distribution is known—which it is not the case with the Massachusetts data. 15 25 and post-war inflation almost certainly played a role); for these two years we assume that the upper bound of the support is $50/week. For our base estimates, we focus on the distributions for all production workers, as these are reported for all years covered in the Massachusetts data, unlike for adult males. Figures 2-4 shows our estimates of the 50-10 (Figure 2), 90-50 (Figure 3) and 75-25 (Figure 4) ratios, each indexed at 100 to its respective value in 1890. Also shown are non-parametric polynomial smoothing regressions of the indexed ratios on observation year, along with the associated 95 percent confidence intervals around them. These figures provide clear and compelling evidence of statistically significant, declining wage inequality. The data suggest modest compression in the 1890s, which then accelerates after the turn of the century. There is also evidence of period effects during World War I and the onset of the Great Depression appear to disrupt the broader secular forces at work. By the late 1930s, the 50-10 ratio – the left tail of the distribution – had narrowed by approximately 20 index points, or about -0.22 log points. The narrowing in the 9050 ratio and the interquartile range was smaller, about 0.10 log points, or about 10 index points (from 100 to 90). 4.1 The Role of Electrification: Panel Estimates The shift to steam-powered production (from hand or water) was one of the central features of the transformation of manufacturing over the second half of the nineteenth century and is, of course, captured in the HML study that was the focus of section 3. Beginning in the late nineteenth century, the source of inanimate power began to shift from steam to electricity. The 1890 Census, for example, reported the use of over 15,500 electric horsepower in manufacturing with Massachusetts (#2), New York (#1 by a small margin), and Pennsylvania (#3) leading the way (U.S. Census Office. 11th census. 1895, p. 759). The shift began in earnest 26 the 1890s and then accelerated swiftly after the turn of the twentieth century (Du Boff 1966; Du Boff 1979, especially Table 3, p. 427). As Figure 5 shows, Massachusetts manufacturing followed this general pattern. By 1910 about 10 percent of horsepower used in Massachusetts manufacturing was generated by electricity. The electricity began rising steeply thereafter, reaching 60 percent in the late 1930s. One of Goldin and Katz’s (1999) hypotheses is that electrification reduced wage inequality in the lower tail of the distribution. Their argument is that electrification led to “technological-skill” complementarity in manufacturing production. Shop floors reorganized in response to electrification, reducing the demand for a wide array of low wage jobs on the factory floor (Devine 1983) as materials handling was more easily electrified and as the production process was “linearized” (Jerome 1934). They were unable, however, to test for an effect of electrification directly, because they did not have an annual time series of wage inequality statistics matched to a similar time series on electrification and their industry-panel consisted of just ten observations, too few for a regression analysis. Following Goldin and Katz (1999) lead, we estimate wage inequality at the industry level using the Massachusetts classified wage distributions. For two years, 1895 and 1920, we also have corresponding industry figures on power use (Massachusetts and Wadlin 1898, pp. 575-84; United States. Bureau of the Census. 1920, pp. 636-47). The industries are identified at the three-digit SIC level, and we have constructed a matched panel of 41 observations for 1895 and 1920. We follow the same protocols regarding cut-offs and procedures with these industry level data as for the aggregate data to estimate q10 and q50. 27 We specify a two-way fixed effects model, with fixed effects for year (1920) and industry. Because there are only two years, we estimate the model in first differenced form. The regression specification is: Ln (q50/q10) = α + β*∆ Elecwkr + γ*∆ Non-Elecwkr + δ*∆Pct Male + ε where Elecwkr = Electric horsepower per worker, Non-Elecwkr = non-electric horsepower (e.g. steam) per worker, and Pct Male = male percentage of workers. Because the industries differ greatly in employment, we weight observations by the average number of workers in the two years. The regression results are shown in Table 4. The hypothesis at issue is the sign of β. If it is negative, then electrification contributed to the compression in the lower half of the wage distribution. As can be seen, the coefficient is negative and is quite stable across the three columns, which add the other two independent variables. The magnitude of the coefficient, 0.128 in the last column, is quite large. If we multiply this coefficient by the mean value of Elecwkr, the predicted change in the dependent variable is -0.154, which accounts for 80 percent of the change (-0.192) in the dependent variable. This result supports Goldin and Katz’s (1999) argument that electrification contributed to the decrease in wage inequality in manufacturing before World War Two. 5.0 Concluding Remarks According to the conventional narrative, wage inequality in US manufacturing followed an inverted-U pattern from the early nineteenth century to just prior to World War Two, a period that encompassed the transition from the artisan shop to the steam-powered factory, and then electrification. This paper fills in two important gaps. For the rising portion of the inverted-U, 28 the previous literature (see, for example, Atack et al. 2004) was unable to measure changes in wage inequality within establishments, while for the falling portion, the precise time-series pattern of change from 1890 to 1940 could not be documented (Goldin and Katz 1999). For both parts of the evolution, a role for mechanization has been hypothesized, but previous work could not examine the role directly. Here, we show that for the rising portion of the inverted-U, we can use operations-level data from the US Department of Labor’s 1899 Hand and Machine Labor Study (U.S. Department of Labor. 1899) to study differences in wage inequality between “hand” (artisan) and “machine” (factory) production of specific manufactured goods. We show that wage inequality was much higher across operations in machine production than in hand production. Most of the higher inequality can be explained by a greater division of labor in machine production, however, not greater mechanization. Secondly, by digitizing and analyzing data from reports produced by the Massachusetts Bureau of Statistics of Labor, documenting so-called “classified wages” in manufacturing as well as the extent of electrification, we confirm a substantial narrowing of wage inequality across production workers starting in the 1890s. This accelerated after the onset of electrification. More concretely, we construct an industry panel for two years that allow us to estimate the impact of electrification on wage inequality directly, focusing on the lower half of the wage distribution. Consistent with Goldin and Katz (1999), we find a strong negative effect of electrification – as use of electric power increased, absolutely and relative to other sources of power, wage inequality in the lower tail compressed significantly. 29