Earnings Inequality in the Great Depression

Transcription

1 Earnings Inequality in the Great Depression Felipe Benguria University of Kentucky Chris Vickers Auburn University Nicolas L. Ziebarth Auburn University and NBER November 27, 2017 Abstract We study earnings inequality during the Great Depression using establishmentlevel information from the Census of Manufactures. There is a large increase in the ratio in establishment-level average earnings for blue collar workers between 1929 and 1933, concentrated in durable goods industries, while the ratio declines. The white collar distribution is far less changed. We decompose changes in the earnings distribution into changes in the distribution of observable characteristics, the earnings associated with these characteristics, and residuals. The change in earnings for blue collar workers is driven almost entirely by changes in the industry-level effects. At the same time, regional differences in blue collar earnings become less pronounced. We thank participants at the NBER SI 2015 DAE, Census Bureau, William & Mary, UW-La Crosse, Florida State, Gettysburg College, Northwestern University, EBHS 2017, and Washington Area Economic History Seminar for useful comments. We thank Dave Donaldson, Rick Hornbeck, and Jamie Lee for providing county-level data for the 1929 Census of Manufactures. We thank Miguel Morin for providing the transcription of some of the published totals for the 1935 Census of Manufactures. The NSF and the University of Iowa provided funding. 1

2 1 Introduction Income inequality has returned as a central topic in economics. Much of this research, such as Piketty and Saez (2003), has examined long-term trends in U.S. income inequality. One key episode in the historical evolution is the so-called Great Compression, a period that experienced a sharp decline in the skill premium starting, at least, by 1940 and continuing through the 1960s (Goldin and Margo, 1992). More recently, economists have looked at the effect of business cycle fluctuations on inequality. For example, Guvenen et al. (2014) document larger losses in earnings for lower-income households during recent U.S. recessions, while Parker and Vissing-Jorgensen (2010) find that as the share of income accruing to the top income percentiles has risen, so too has that group s cyclical sensitivity. We bring together the historical perspective and the study of business cycles by examining the Great Depression. This was the the largest business cycle shock in American economic history, and it happened to take place on the eve of the Great Compression. Little is known about earnings inequality during the Depression because of the scarcity of data on earnings for different groups of the population. The Population Census was conducted only once per decade and did not include wage data until The Bureau of Labor Statistics (BLS) conducted occasional surveys during the Depression, but they do not allow for systematic measures of wage inequality across industries and states. Data drawn from tax records is useful only to compare those at the very top, the highest-earning 1 percent or at most 10 percent, relative to the rest of the population, since the vast majority of the population did not pay income taxes at this time. Instead of relying on scant individual level data on the earnings distribution, we turn our focus instead to the demand side of the labor market; that is, employers. We use the establishment-level schedules of 25 industries from the Census of Manufactures between 1929 and These industries, while not randomly chosen, cover a wide swathe of manufacturing, from durable goods to consumer products to high tech industries. The samples covers around 10% of all manufacturing establishments and around 20% of 2

3 total wages and value added. A key feature of this Census is its frequency: it was conducted every two years during the 1920s and 1930s. We have schedules from 1929, 1931, 1933, and 1935, allowing us to study income dynamics from the beginning of the Depression through its nadir and the first part of the recovery. 1 This dataset allows us to study earnings disparities across establishments as well as inequality within establishments across worker types. In each establishment, we observe the earnings of wage earners and of salaried workers, which we refer to as blue versus white collar workers. This distinction is very close to that between production and non-production workers commonly used today in working with manufacturing establishment data, e.g., Davis and Haltiwanger (1991). To support the claim that this distinction is related to educational differences, we show that there is a positive relationship between educational attainment and the share in white collar employment using data from the 1940 Population Census. The first contribution of this paper is to document patterns in the distribution of average wages across establishments along a number of dimensions, such as industry and geography. We find little change in dispersion measures for white collar workers between 1929 and For blue collar workers, there is an increase in the ratio and a decrease in the ratio. That is, the very highest paying establishments for blue collar workers were more highly paid relative to the median in 1933 than in 1929, but the lowest paid were closer to the median. This change is largely reversed by We then show that this difference is concentrated almost entirely in industries producing durable goods, which were the most severely affected by the Depression. One possible explanation for this pattern is the role of hours worked in total earnings. We also document the important role of establishment characteristics in understanding earnings differences. We focus on total revenue, incorporation status, and whether an establishment is part of a multiplant firm as our key establishment characteristics. There is a strong positive correlation 1 We exclude 1931 because of a data issue we discuss later. 3

4 between revenue and earnings for both white and blue collar. This relationship is stable over the years in our sample in the pooled regression. At the same time, we find little correlation with incorporation or multiplant status over and above the size effect captured by controlling for total revenue. Having documented these patterns in the between-establishment wage distribution, our second contribution is to quantify the role of these factors industry, geography, observable characteristics of the establishments in explaining changes in the wage distribution using the decomposition of Juhn et al. (1993). This technique decomposes changes in the distribution of earnings into changes in observables, changes in the passthrough or prices of those observables, and changes in the residuals. Davis and Haltiwanger (1991) use this method to study the role of establishment characteristics in postwar U.S. manufacturing. Pooling both blue and white collar workers, we show that changes in the common trends in various percentiles of the earnings distribution are driven by changes in the price component, while the residual component is crucial for understanding dispersion across percentiles of the earnings distribution. Changes in the quantities, or relative representation of observable characteristics, play basically no role. If we break out blue collar workers, changes in the distribution of earnings are explained by substantial changes in the earnings premiums associated with observable characteristics. In particular, changes in regional effects push towards the equalization of incomes from 1929 to 1933, as the penalties associated with being in the lower income regions in the southern United States become less severe. While these regional changes are persistent into 1935, the changes in industry effects reverse by The residuals are associated with a decrease in inequality across establishments, a trend partially reversed by That is, the residuals for the lowest percentiles increase from 1929 to 1933, suggesting that the inequality across establishments that cannot be explained by differences across regions or industries decreases during the first years of the Depression. The importance of regional differences in understanding the Depression has also been emphasized 4

5 by Rosenbloom and Sundstrom (1999). To summarize the results from the decomposition, the changes in the income distribution from 1929 to 1933 were concentrated in blue collar workers in durable goods manufacturing. While there were changes in the between region distribution of earnings leading to compression that remained at least until 1935, the changes associated with particular industries were less long lasting. Moreover, the unexplained portion of the income distribution resulted in compression from 1929 to 1933, but this shock was transitory. 1.1 Related Literature Most of the literature on inequality has focused on worker characteristics such as educational attainment in explaining earnings inequality. This has led to a relative dearth of studies about the role of demand side for labor, firms, in generating earnings dispersion. Earlier work by Groshen (1991) argues that employers are a key determinant of wage differences. Davis and Haltiwanger (1991) document that over half of wage variance in manufacturing can be accounted for by dispersion across establishments in mean earnings. Moreover, dispersion across establishments accounted for almost half of the overall growth in wage dispersion in manufacturing between 1975 and More recently, Song et al. (2015) use tax records to show that within-firm wage differentials have changed far less than between-firm ones over the last 30 years. Barth et al. (2016) find that the increase in inequality in recent decades is mostly a between-firm phenomenon. Using German linked employer-employee data, Card et al. (2013) show that changes in inequality in recent decades are due to larger dispersion of worker effects, larger dispersion of firm effects, and increased assortative matching between workers and firms. As for earlier work on changes in inequality during the Great Depression, most earlier work has focused on the geographic dimension mainly for data availability reasons. For example, Hanna (1954), Schmitz and Fishback (1983), and Creamer and Merwin (1942) all consider changes in earnings across states during this period. Rosenbloom and Sund- 5

6 strom (1999) document that different regions of the country do better or worse during the Depression, at least as measured by manufacturing outcomes, even after controlling for differences in industry composition. Hausman (2016) shows that much of the geography of the 1937 recession is explained by industry differences, particularly the sharp decline in the automobile industry. Drawing on BLS data, Wallis (1989) documents variation in state level employment outcomes, which he argues are not mainly driven by industry differences. In an earlier paper, Borts (1960) also documents the regional variation in business cycles across the country. Much of the other work on inequality (Kuznets, 1953; Goldsmith et al., 1954; Mendershausen, 1946; Tucker, 1938) uses the published statistics on income based on income tax returns and attempts to infer some measure of inequality from that and aggregate income. These methods are limited to only focusing on inequality driven mainly by the upper income percentiles. There are also a handful of papers that focus on industry differences during the Depression, including the classic work by Bernanke (1986). He studies a sample of eight industries with information on employment and wages, drawing on data first studied by Beney (1936). Hanes (2000), following work by Shister (1944) and Dunlop (1944), examines industry characteristics and their relationship to wage rigidity during the Depression and two other downturns in 1893 and Goldin and Margo (1992) have some data on the skill premium in a handful of industries that cover the period of the Great Depression. Clearly, differences in wage rigidity with respect to declines in labor demand will generate cyclical changes in inequality, though this aspect is not drawn out in that work. The source for the data in that work is a BLS survey of establishments. Other work by Wachter (1970) has examined more generally the cyclical variation in the distribution of wages across industries. There is no literature that we are aware examining the role of establishment or worker level characteristics for understanding inequality during the Great Depression. Our work also contributes to an emerging literature on inequality during the Great Recession. Mian and Sufi (2016) survey the literature on the distributional conse- 6

7 quences of recessions, and provide further evidence on the role of pre-recession household debt as a determinant of consumption losses. Heathcote et al. (2010) and Guvenen et al. (2014) show that in all U.S. recessions in recent decades earnings fall substantially more for those at the bottom quantiles of the pre-recession earnings distribution, widening inequality. 2 Saez (2015) also documents earnings growth in the U.S. during the Great Recession for various percentiles of the income distribution. 2 Data and Data Issues 2.1 Data Source: Census of Manufactures We employ the Census of Manufactures (COM), for which the original, establishmentlevel schedules are available from 1929, 1931, 1933, and The COM was taken for other years, including during the Depression, but the establishment-level schedules from the first half of the 20th century do not exist other than for these four years. The schedules provide a wealth of detail beyond simply identifying information and include a breakdown of outputs and inputs into quantities and values. For our purposes, there is also information on labor use broken down by type of worker, number employed, and total earnings. This work draws on a sample of 25 industries summarized in Table 1. 3 In Table 2, we compare the totals in the data used in this paper to the national published tables along various dimensions. The sample comprise over 20% of the total value of output in manufacturing, approximately 10% of the establishments, and about 20% of both the total wage bill. While not chosen to be representative, these industries do cover a wide variety of manufactured products from durables to consumer goods to high tech goods like ra- 2 An exception occurs at the top percentile, which also sees very large losses during the Great Recession. 3 In addition to our own collections, the sample combines industries already used in published research as well as a number of industries collected by Timothy Bresnahan and Daniel M.G. Raff (but to our knowledge never exploited) and a few new industries collected by Miguel Morin. 7

8 dios. In addition, there are large differences across the sample in terms of the average size of the establishment and the share incorporated to name just two. 4 In Figure 1, we consider the geographic coverage of the our sample. We plot the percent of manufacturing wage earners covered by our sample by county in 1929 compared to the national totals from the 1930 Population Census (Haines, 2010). Our data oversamples in the Carolinas and Georgia because of the dominance of textiles in that region, but overall it covers a reasonable percentage of manufacturing for a wide swathe of the country. It is useful here to compare the establishment level variables which are available in the 1930s COM to those in the modern COM used in, for example, Davis and Haltiwanger (1991). They consider regional and industry differences, size by number of employees, age, ownership, energy costs, product specialization, and capital intensity. In the COM data we employ, information about region, industry, and size are all available. The information for age is considerably scantier. The form for the 1929 census asks if the plant had begun operation before January 1, 1928, but other than that there is no information on age. None of the other censuses ask such a question. 5 We also have information on the incorporation status of an establishment as well as the ownership variable used in Davis and Haltiwanger (1991). This latter variable is whether the establishment is part of a single-plant or multiplant firm. This is available in the Census of Manufactures, in that the name of the operating firm is on the census form. As for other variables emphasized by Davis and Haltiwanger (1991), we have information for energy sales for 1929 and 1935 only, but this question is not asked in 1931 and Davis and Haltiwanger (1991) employ a measure of product specialization, which is the fraction of plant shipments accounted for by the largest five-digit SIC product class. The most im- 4 One detail to point out is that for a few of the industries, we do not have the revenue variable directly. Instead, we have to impute revenue by adding up the value of all products. The problem with that is due to limitations in how some data was originally collected, we do not necessarily have information on all of the products. To allay concerns here, we have compared our imputed value of total industry revenue to the value in the published volumes. In addition, to the extent that we are missing a common fraction of revenue across all establishments and years within a year, then this will be swept up in the industry fixed effects. 5 Of course, for establishments which are linked across censuses, we at least know a lower bound for age. That is, an establishment in both the 1931 and 1929 censuses must be at least 2 years old in

9 portant omission from the COM data is a measure of capital, which is by and large absent. There are, in some cases, industry-specific measures of physical capital stock; for example, the amount of compression capacity in the manufactured ice industry. However, nothing is systematic enough to use in regressions pooling across industries and years. 2.2 Skill Categories and the Comparability Across Years A key question for us is how we measure skill. The forms do not provide any information on, for example, educational attainment. Instead, we will distinguish between wage versus salaried employees, or what we will call blue versus white collar jobs. White collar workers in this classification were clerks, administrative officers, and office workers in general. Blue collar workers were presumably mainly production workers on the factory floor, but they could also include hourly janitorial staff or workers on the loading dock. In the modern COM, the labor force breakdown is into production and non-production workers. Dunne et al. (1997) defend the use of non-production workers as a measure of skilled workers. Our breakdown is quite similar to this. All of our salaried workers would fall into the category of non-production workers. However, there are surely some hourly workers such as janitors that would be classified as non-production workers in the modern taxonomy. We show below that this white versus blue collar distinction carries considerable information about a worker s educational attainment. For this reason, our results speak to changes in the skill premium, at least in the context of manufacturing. There are some additional details worth mentioning for how we construct these employment variables. For blue collar workers, there are two different choices for the employment number. In each of the years 1929, 1933, and 1935, the census asks for the total number of wage earners on a particular date in December, with a further breakdown in 1929 between men and women. It additionally asks for the number of wage earners employed in each month. The income measure comes from a question asking for total amount paid to wage earners. So in computing blue collar average earnings, we take the 9

10 total amount paid divided by the average of the monthly employment figures to compute the average income. Note that blue collar here does not refer to strictly speaking unskilled labor. The census forms in 1929 and 1933 ask establishments to include skilled and unskilled workers of all classes, including engineers, firemen, watchmen, packers, etc. along with foremen and overseers in minor positions who perform work similar to that done by the employees under their supervision. The question for these group in 1935 was somewhat different, asking respondents to include all time and piece workers employed in the plant, not including employees included in the other enumerated categories of officers, managers, clerks, and technical employees. We would emphasize that this average earnings variable does not take into account possible changes in hours worked. As noted by Rosenbloom and Sundstrom (1999), the Census itself cautioned about how to interpret the wage earner employment number, our measure of blue collar employment. In particular, they listed some possible reasons why the employment numbers are inflated relative to the true number of full-time equivalent employees. First, the establishments might have reported part-time workers as well. Second, workers that were laid off might have stayed on the payroll for a time and still been potentially counted as working at the establishment. While these may potentially bias the overall level of average earnings, identifying systematic biases they would create in cross-regional or cross-industry comparisons is more difficult. For white collar workers, the computation is somewhat more complex. We ignore the distinction made in 1929 between male and female workers, both for consistency with other years which do not ask this question and because the lack of a breakdown in pay between men and women makes the question difficult to exploit. In all three years, the census asks for the number of proprietors and the number of officers of the corporation employed on a particular date in December. There is no information about income for proprietors, but each census asks about the amount paid to officers during the year. We 10

11 exclude officers and proprietors from the calculations. For the remaining white collar workers, in 1929, the census asks about the number of managers, superintendents, and other responsible administrative employees; foremen and overseers who devote all or the greater part of their time to supervisory duties; clerks, stenographers, bookkeepers, and other clerical employees on salary, as well as the total amount that this group was paid. In 1933, this category is split between the managers and clerks reported separately, along with their total wage bill, with no mention of foremen. 6 For 1935, the same three white collar categories of officers, managers, and clerks are reported, with the total number of clerks being reported for four separate months. For this year, there is also an entry for the number of technical employees including trained technicians, such as chemists, electrical and mechanical engineers, designers, etc., which is not asked in any other year, along with the income they are paid. We do not include technical employees in the calculations since they are not included in the previous years. Unfortunately, the 1931 census form does not contain any information about white collar workers, so we will drop this year in our analysis. 2.3 What is the White vs. Blue Collar Distinction Measuring? An important question is what our distinction between salaried, white collar, and wage earners, blue collar, captures. We interpret it as reflecting educational attainment, commonly referred to as high versus low skill in the literature. To provide supporting evidence, we turn to the 1940 Population Census. While the COM did not collect information on the educational attainment of workers, the occupational description and industrial affiliation of these white collar and blue collar worker categories can be matched to those in the Census of Population to obtain information on the characteristics of workers in these different types of jobs. In particular, IPUMS has coded occupations into 3 digit categories, 6 In both 1929 and 1933 the census specifically includes foremen in minor positions who perform work similar to that done by the employees under their supervision in the wage earner category. 11

12 with 1 through 300 as corresponding to white collar and the rest corresponding to blue collar save for occupation 999, which is the code for missing or unclassified. Overall, pooling together workers in all manufacturing industries and in all states, we find that in 1940 only 17.2 percent of blue collar workers had an educational attainment corresponding to a high-school degree or more. In contrast, 61.6 percent of white collar workers had at least a high-school degree. The median white collar worker has an educational attainment of 12 years of school (exactly a high-school degree) while the median blue collar worker had 8 years of schooling. Table 3 puts this into a regression framework and predicts educational attainment based on the white versus blue collar distinction. We consider two definitions of educational attainment: (1) high school graduate and (2) some college, employing one specification with no controls and one with controls for sex, race, industry and state fixed effects. We find a statistically and economically significant relationship education and white-collar employment. Figure 2 shows that this relationship between educational attainment share in white collar is nearly monotonic, with only a slight flattening for only the most educated where the share in white collar is already approaching 1. 7 While not as good as directly observing educational attainment of a worker, this shows that the color of a worker s job carries considerable information about his educational attainment. 3 Changes in the Earnings Distributions We begin by focusing on the distribution of earnings across skill groups and time. We calculate employment and establishment level distributions of earnings. The difference statistically speaking is that the worker distribution weights establishment-level earnings by the number of workers, whereas the establishment distribution is unweighted or, more 7 Educational attainment is coded according to IPUMS standard where a value of 6 corresponds to high school graduate and a value of 10 or higher means college graduate. These values are not simply years of schooling. 12

13 precisely, weights each establishment equally. This is the relevant distribution for a search model where a worker samples from the distribution of establishment earnings (assuming the worker ignores the within establishment wage heterogeneity). The employment distribution is not the true aggregate distribution of worker earnings, since we are implicitly treating all workers at a given establishment as earning the same amount. This will surely mask variation in earnings within an establishment and skill group, so the inequality measures based on this distribution will be lower than the measures from the distribution where we observed every individual employee s earnings. However, the Davis and Haltiwanger (1991) decomposition for modern data shows over half of the variance in manufacturing wage inequality can be accounted for by betweenestablishment differences in average earnings. This suggests a substantial fraction of changes in the income distribution can be captured with the COM data. Figures 3 and 4 display these distributions over time across skill groups for the different weighting choices. All values are reported in thousands of 2015 dollars and we winsorize the 1% tails. First for blue collar earnings in Figure 3, there is a clear shift in the distribution leftward between 1929 and 1933 and then a shifting back in 1935 in the case of establishment weights. A similar pattern is present if we focus on employment weighted results. In Figure 4, for white collar workers there is much less of a shift and, if anything, the distribution shifts to the right. For both the blue and white collar earnings distributions, the whole distribution shifts to the right compared to equal establishment weights, implying that larger establishments tend are associated with higher average earnings. Tables 4 and 5 show some measures of inequality over time for the white and blue collar skill groups, respectively. For blue collar earnings at the establishment-level, there is a spike in both the and inequality measures. This pattern does not carry over when we weight by employment. In this case, the measure shows a sharp decline in 1933 and a continued fall in 1935, while the shows a similar decline in 1933 and then recovery in In either case, there is a large increase in the differential from 13

14 1929 to It is simply that for the employment weighted case, there is an even larger decline in the differential leading to the decline in measure of inequality. For white collar earnings, we observe less dramatic changes in the inequality measures than for blue collar workers, but the same overall patterns for both types of approaches to weighting. This is independent of whether we look at the establishment or worker levels. Turning to the blue collar earnings distribution, we observe a clearer inverted U-shape pattern in the establishment-level distribution here, with inequality spiking in 1933 and then returning towards its 1929 value in This is true if we consider the standard deviation, the difference, or the difference. The changes appear to be due to both the upper and lower tails of the distribution moving away from the median as evidenced by changes in both the and differences. Note that winsorizing does not affect these statistics. From a statistical point of view, the key difference in comparing the distributions over time and across the weights applied comes from establishments that shrink but do not altogether exit. For the set of establishments where average earnings do not change, these establishments will have no effect on the establishment-weighted distribution while there will be effects when weighting by the number of workers at the establishment. To get some intuition for this disconnect between results weighted by employment and by establishment, consider a very simple case where there are only two levels of earnings and two establishments. Normalize the smaller value of earnings to 0 while denoting the larger one by w and the fraction of workers that are paid this wage be p. Then the unweighted, or establishment weighted, standard deviation is equal to w/2 while the weighted, or employment weighted, one will be equal to w p(1 p). So for the establishment weighted measure to increase while the employment one to fall, it must be that the weights p and 1 p change in such a way to offset the direct increase from changes in the variance of establishment earnings. One way to interpret the fact that the employment weighted distribution is shifted to the right of the establishment weighted one is that p > 1/2. So the 14

15 decline in employment weighted dispersion must be due to an additional increase in p or such a large decrease that more than half of all workers end up earning the high wage. Given the long-run perspective, it is perhaps reasonable to ignore the hours dimension of earnings inequality and its possible effect on welfare inequality where consumption is adjusted for hours worked. The fact is that while average hours worked has become more unequal across skill groups over time, this change does not seem quantitatively relevant for earnings inequality overall. 8 This ignoring of the hours margin is perhaps not as tenable in our setting of the Great Depression when hours fell dramatically in a short period of time. Since we are not adjusting for this intensive margin hours, these changes in earnings inequality probably overestimate the changes in inequality of welfare, earnings adjusted for hours worked. On the other hand, because our data come from the employer side, our measure of inequality is a conditional one in that it only reflects earnings inequality for those that remain employed. In this sense, changes in our earnings inequality underestimates the changes in welfare inequality by excluding these extensive margin changes in employment. 4 Locating the Changes in the Earnings Distributions We now attempt to identify the location of these changes in the earnings distributions. We consider three types of characteristics: (1) cross-sectional, or the initial earnings percentile of an establishment in 1929; (2) industry characteristics, particularly whether or not the industry produces a durable good; and (3) establishment characteristics in the form of productivity and size. 8 In addition, when focusing on income inequality at the very top, capital income is a more important piece of overall earnings and there is no real hours margin there. 15

16 4.1 Cross-sectional Characteristics: Earnings Percentile in 1929 Where in the earnings distributions are these changes coming from? Following Song et al. (2015), we attempt to locate the source of changes in inequality in the earnings distribution. To do this, we sort establishments based on their workers average earnings in 1929 and 1933, then calculate the growth rate of earnings for these percentile groups between 1929 and 1933 or So if say the top percentiles show faster growth, then this implies that inequality has increased. Note that these growth rates are not within the same set of establishments. Those establishments in the top 1%, say, in 1929 need not be the same as the top 1% of establishments based on average earnings in While somewhat counterintuitive, this is what most studies are reporting when they talk about growth in the income of the top 1%. One feature of this approach is that it handles entry and exit by simply focusing on the set of establishments extant in 1929 and 1933 without having to make an assumption on the growth rate of earnings for establishments that exit or those that enter. Figure 5 shows that between 1929 and 1933, there is a positive relationship between initial percentile and subsequent growth. This implies a pulling apart of the earnings distribution. Between 1929 and 1935, there is an inverted U shaped relationship with the top and bottom percentiles seeing the slowest (most negative) growth. This is consistent across both skill groups. This implies a narrowing of inequality at the top and pulling away at the bottom. This compares to the result in Song et al. (2015), where they find growing inequality over a longer period of time driven by inequality between firms. 4.2 Industry Characteristics: Durable Goods While revealing, the cross-sectional characteristics do not identify economic variables that predict these changes. One possible economic source of these changes is the difference in cyclical sensitivities across industries. It is well known that employment in durable industries are much more sensitive to fluctuations in aggregate demand. With in mind, 16

17 Figures 6 and 7 plot the distributions of skill premiums, separating industries that produce a durable good from those that do not. Strikingly, only for blue collar workers in durable industries do we observe shifts in the earnings distribution over these years. This finding casts doubt on one interpretation of all these results that they are simply driven by some quirk in how the 1933 COM was collected. That theory could not explain why the differences in 1933 are only there for durable industries. Remember that we are measuring earnings inequality, which reflects both differences in wage rates and hours worked. One explanation for these differences by durability and skill group is that hours worked for blue collar workers, or, more to the point, hourly wage earners, in durable goods industries are more sensitive to demand fluctuations. Therefore some fraction of these changes in earnings inequality reflect differences in hours worked rather than differences in wage rates. This is not to say that non-durable and white collar workers are left totally unaffected. While there were surely job losses for white collar workers, at least under this interpretation of the results, there does not seem to have been a large changes in hours worked for those that remained employed. 4.3 Establishment Characteristics: Productivity and Size Another possible source of the changes in the distribution is changes in the relationship between establishment characteristics and earnings over time. We focus on two characteristics often discussed in the literature: labor productivity and size as measured by revenue in our case. Figure 8 shows the relationship between these two establishment-level characteristics and earnings for white and blue collar workers, disaggregating by year controlling for industry fixed effects. First, focusing on the relationship with productivity, we find strong positive correlations across all census years and both types of workers. The elasticity of blue collar earnings on labor productivity is slightly larger in 1933 relative to the other years. The elasticity for white collar earnings is actually lower though earnings at each point in the labor productivity are higher. Note that labor productivity is 17

18 measured as total revenue divided by total employment, blue plus white collar workers. Turning to the relationship between size as measured by revenue and average earnings, we see similar patterns with strong positive correlations for both collars. The elasticities do not change very much over time though there is some shifting of the whole curve for blue collar earnings. There are, of course, clear economic mechanisms for this positive relationship and copious modern evidence that more productive firms pay higher wages. There are also noneconomic, more mechanical reasons for this relationship. In particular, note that our measure of productivity is simply total revenue divided by the total number of workers. We are scaling total earnings of some type of worker by the same quantity. So if there is any measurement error in that employment variable, that will drive a spurious correlation between these two. The fact that these correlations are so stable across years suggests to us that these have more economic foundations. 5 Decomposing the Changes in the Earnings Distributions After identifying some potential drivers of changes in the earnings distribution, we now decompose the sources of those changes using the procedure developed by Juhn et al. (1993) (JMP). First, we write the earnings equation as y it = X it β t + u it where y it is the log average real income (in 2015 dollars) for either blue collar or white collar workers at establishment i in year t, X it a vector of establishment characteristics, and u it the wage component that is left unobserved. Let the distribution function of the residuals be denoted by F t (.). Then define the establishment s percentile in the distribution of residuals 18

19 as θ it. After estimating the regression, we can then invert the residual as u it = F 1 t (θ it X it ) where F 1 t (θ it X it ) is the inverse CDF of the residuals in year t and characteristics X it. Changes in the distribution can come from changes in average establishment characteristics X it, changes in the relationship between these characteristics and incomes represented by β t ( prices in the language of JMP), or changes in the distribution of residuals F t. If we define β = ˆβ 1929 and F 1 (θ it X it ) = ˆF (θ it X it ), we can write: y it = X it β + [ X it (β t β) + F 1 (θ it X it ) ] + [ F 1 t (θ it X it ) F 1 (θ it X it ) ] With this identity in mind, we first predict the earnings and construct residuals by using the fixed coefficients from 1929, β. If the distribution of wages were fixed and the coefficients β are fixed, then earnings are given by y 1 it = X it β + F 1 (θ it X it ) If the β coefficients can change as well as the value of the covariates X it, then earnings are y 2 it = X it β t + F 1 (θ it X it ) If we further allow the distribution of residuals to change, then earnings are given by y 3 it = X it β t + F 1 t (θ it X it ) which is the actual level of earnings. Changes in the distribution of y 1 it are those at- 19

20 tributable to changes in the characteristics of the establishments. Any additional change in the distribution of y 2 it are changes in the returns to the quantities, and finally additional changes to y 3 it = y it can be attributed to changes in the residuals. It is important to remember that this is just a decomposition and is not structural. We use the language of prices and quantities following JMP, 9 but, of course, there is no reason to think that these can be neatly separated. In specifying the regression, we pool both blue and white collar workers in estimating this specification, meaning we include a control for white or blue collar status and weight each establishment- color by the number of workers. This is similar in spirit to what JMP did using individual level data from the CPS. Following the literature, we will assume that F t is independent of X. For the vector of characteristics, we include a binary variable for whether or not the firm is incorporated or not, whether or not it is part of a multiplant firm, controls for regions of the country, 10 for the industry classification, and the log of employment. Table 6 shows the results from the regressions for each year pooling both types of workers weighted by the number of employees with robust standard errors. First, and not surprisingly, we find a large skill premium of 0.34 log points in 1929, increasing to 0.73 points in 1933, and falling back to 0.55 log points. These large swings in the skill premium we would argue can at least partly attributed to changes in hours worked across the skill groups. We find no statistically significant effect of working at an incorporated firm or a multiplant firm in any of the three years. There is also a significant size premium across all three years, consistent with modern literature on this question. The effects for the regional categorical variables (New England is the excluded category) are as expected, with a large negative earnings effect associated with being located in the South. We now turn to the decomposition using these regressions results. In Figure 9, we show the values from the JMP decomposition for all workers, taking 1929 as the base year 9 In fact, they attempt to identify a demand shifter for skilled labor. 10 These are the standard Census region definitions. 20

21 and defining changes relative to that year. We perform the decomposition at the 90th, 75th, 50th, 25th, and 10th percentiles of the distribution of log average income. The four graphs show the four parts of the decomposition: (1) Totals, (2) Quantities, (3) Prices, and (4) Residuals. The Totals replicates the pattern observed in the summary statistics of a fanning out of the earnings distribution in 1933 and then a return to the pre-depression spread by All the percentiles in Quantities show basically no change over this period and for Prices, all the percentiles move in the same direction, which will not be useful for explaining changes in earnings inequality. Instead, while changes in prices are crucial for explaining the total changes in average earnings, changes in the residual component are important for explaining the dispersion in total changes. For example, the reason why the 10th percentile of the earnings distribution does not fall as much as the other percentiles between 1929 and 1933 is because the residual component for this percentile actually increases over this time period. We emphasize that these results were not foreordained. For example, in the original paper by JMP, they argue that much of the increase in wage inequality between 1963 and 1989 is due to increases in the returns to skill. In our case, it could have turned out that changes in the composition of blue collar workers were important for the overall changes in the average earnings, perhaps due to changes in the location of these workers. It could have turned out that changes in the prices of these explanatory variables were important, perhaps due to changing loadings on incorporation. Neither of these explain the changes in the distribution. 5.1 Decomposition for Blue Collar Workers Because the sharpest changes in the distribution of earnings were for blue collar workers, and particularly in the durable goods industries, we perform the same JMP decomposition for blue collar workers only, separately by durable and nondurable industries. We plot the results for non-durable goods industries and durables in Figures 10 and 11 respectively. 21

22 First, for non-durables, recall the overall distribution was relatively unchanged from 1929 to 1933, with a slight increase in the number of relatively high paying establishments. This is reflected in the plot in the upper left of Figure 10, with relatively more of an increase in the upper parts of the distribution. The quantities are little changed, and the prices show the same pattern as the overall results. In the bottom right, the residuals are actually negative for the 90th percentile. That is, the distribution at the top would have been more compressed, but the wage premium associated with observable characteristics pulled the distribution in the opposite direction. In Figure 11, we perform the same exercise with durable goods. Here, residuals have very little effect on changes in the distribution. That is, virtually all of the common changes in the distribution are caused by changes in the wage premiums or penalties associated with the observable characteristics of particular establishments. The fact that the total change in the 90th percentile between 1929 and 1933 is smaller in magnitude than the other is almost completely due to differences in changes in the quantities component. It is important to be careful in comparing these results to those for the pooled sample. In particular, it should not be surprising that residuals are relatively less important here since we are using the same set of explanatory variables to effectively explain a smaller amount of variation. To examine how the prices are affecting the distribution, we plot the coefficients on the fixed effects for the industries in Figure 12; the omitted category is the macaroni (pasta) industry. 11 The coefficients are sorted by their 1929 values. In the upper part of the figure, comparing the regression coefficients from 1929 and 1933, we observe a number of industries associated with particularly large changes from 1929 to A number of primarily local industries have large positive changes in their coefficients, such as manufactured ice, ice cream, and beverages. In addition to being local, these are characterized by having a large number of relatively small establishments. In contrast, a number of large, heavy manufacturers, such as aircraft, motor vehicles, and blast furnaces, have falls in their wage 11 For clarity of the graph, both linoleum and sugar cane, which are small industries with large negative coefficients, are omitted. 22

23 premiums. Comparing to the bottom half of Figure 12, which compares 1929 to 1935, we see that most of these changes are reverted by Ice cream, manufactured ice, and beverages all have coefficients in 1935 within the confidence interval for We can compare this change in industry premiums to the changes in the other main explanatory variable we consider, geography. We plot the changes in the coefficients on the region from the same specification in Figure 13. New England is the omitted category. From 1929 to 1933, there is an increase in the coefficients associated with being in the lowest paid regions in 1929, the East South Central and South Atlantic regions. That is, in terms of pay for blue collar workers, there was a regional equalization during the depths of the Depression. This finding echoes the result of Rosenbloom and Sundstrom (1999), who find that the Depression was comparatively more mild in these areas. Unlike the changes for the industry coefficients, the regional coefficients do not revert to the 1929 values by Instead, the lowest paid regions are still considerable higher relative to their 1929 values in To summarize the above analysis, for earnings as a whole, there is an increase in inequality across establishments from 1929 to The unexplained (residual) component of establishment level incomes contributes to this widening, becoming more negative for lower parts of the income distribution. Drilling down to blue collar workers, where we see larger changes in the income distribution, we see that there are dramatic changes in the premiums associated with particular industries, with large and positive relative wage premiums for industries such as manufactured ice and ice cream. These industry level changes reverse by In contrast, the regional coefficients point towards an equalization of income between 1929 and 1933, a pattern which does not change by

24 6 Conclusion We have documented a number of dimensions of earnings inequality during the Great Depression drawing on establishment-level data from the COM. The wage distribution of white collar workers was little changed, while for blue collar workers in durable goods industries, there was a substantial increase in the ratio. We also show that withinfirm inequality increased substantially between 1929 to When decomposing these results, we find changes in the wage premiums associated with observable characteristics explain the bulk of the changes, particularly for the blue collar workers in durable goods industries. Going forward, it would be interesting to relate changes in within establishment earnings inequality to policy changes and other external drivers. For example, the 1934 Act creating the Securities Exchange Commission mandated executive pay disclosure and Mas (2015) finds that this had effects on executive compensation. In addition, there are a myriad of policies under the National Industrial Recovery Act such as minimum wages and mandated work sharing that might have affected inequality as well. We leave these for future work. 24