这是用户在 2024-10-28 14:43 为 https://app.immersivetranslate.com/pdf-pro/65ea0ce4-70f9-4462-8fd8-d22813451b0c 保存的双语快照页面,由 沉浸式翻译 提供双语支持。了解如何保存?

Rising Top, Falling Bottom: Industries and Rising Wage Inequality
上升的顶端,下降的底端:行业与日益加剧的工资不平等

By John Haltiwanger, Henry R. Hyatt, and James R. Spletzer*
约翰·哈尔蒂旺格、亨利·R·海厄特和詹姆斯·R·斯普莱策*

Most of the rise in overall earnings inequality from 1996 to 2018 is accounted for by rising between-industry dispersion. The contribution of industries is right-skewed with the top 10 percent of four-digit NAICS industries dominating. The top 10 percent are clustered in high-paying high-tech and low-paying retail sectors. In the top industries, high-wage workers are increasingly sorted to high-wage industries with rising industry premia. In the bottom industries, low-wage workers are increasingly sorted into lowwage industries, with rising employment and falling industry wage premia. (JEL J23, J24, J31, L25, M52)
1996 年至 2018 年间,整体收入不平等上升的主要原因是行业间差异的增加。各行业的贡献呈右偏分布,四位数 NAICS 行业中前 10%的行业占主导地位。前 10%的行业集中在高薪高科技和低薪零售部门。在顶尖行业中,高薪工人越来越多地被分配到高薪行业,行业溢价不断上升。在底层行业中,低薪工人越来越多地被分配到低薪行业,随着就业增加和行业工资溢价下降。 (JEL J23, J24, J31, L25, M52)
A growing number of studies attribute increases in earnings inequality to rising between-firm dispersion. 1 ] 1 ] ^(1]){ }^{1]} We confirm this pattern with US matched employer-employee data from 1996 to 2018. Our contribution is to demonstrate that rising between-industry dispersion accounts for most of the overall increase in earnings inequality, and is driven by a relatively small number of industries in the tails of the industry earnings distribution. About 10 percent of four-digit NAICS industries account for virtually all of the increase in between-industry dispersion, while
越来越多的研究将收入不平等的增加归因于企业之间的差异扩大。我们通过 1996 年至 2018 年的美国匹配雇主-员工数据确认了这一模式。我们的贡献在于证明,行业之间的差异扩大占据了收入不平等总体增加的大部分,并且这一现象是由行业收入分布尾部的相对少数行业驱动的。大约 10%的四位数 NAICS 行业几乎占据了行业之间差异增加的全部。
accounting for less than 40 percent of employment. 2 2 ^(2){ }^{2} Remarkably, the remaining 90 percent of four-digit industries contribute little to rising between-industry earnings inequality. 3 3 ^(3){ }^{3}
占就业人数不到 40%。 2 2 ^(2){ }^{2} 值得注意的是,剩余的四位数行业的 90%对行业间收入不平等的上升贡献甚微。 3 3 ^(3){ }^{3}
The top 10 percent of industries that contribute to rising inequality include 19 that are high-paying (top quartile in pay). These industries account for 54.1 percent of the increase in between-industry inequality. Average worker earnings have surged in these industries accompanied by employment gains. The top 3 of these are high-paying, high-tech industries: Software Publishers (NAICS 5112), Computer Systems Design (5415), and Other Information Services (5191) - and, in total, 11 of these 19 high-paying industries are high-tech. As discussed in Oliner, Sichel, and Stiroh (2007) and Fernald (2014), these industries are characterized as the source of rapid technological advances over our sample period.
贡献于日益不平等的前 10%行业中,有 19 个是高薪行业(薪资的前四分之一)。这些行业占行业间不平等增加的 54.1%。在这些行业中,平均工人收入激增,同时伴随着就业增长。这三大行业是高薪高科技行业:软件出版商(NAICS 5112)、计算机系统设计(5415)和其他信息服务(5191)——在这 19 个高薪行业中,总共有 11 个是高科技行业。正如 Oliner、Sichel 和 Stiroh(2007)以及 Fernald(2014)所讨论的,这些行业被认为是我们样本期间快速技术进步的来源。
Eleven low-paying (bottom quartile) industries constitute the remainder of the top 10 percent in terms of their contribution to inequality. These industries, in combination, account for 44.1 percent of the increase in between-industry inequality. More than one-fourth of the increase is accounted for by just 3 of these 11: Restaurants and Other Eating Places (7225), Other General Merchandise Stores (4529), and Grocery Stores (4451). These industries have gone through substantial changes in recent decades, moving away from single-establishment firms to large, national chains, see Foster, Haltiwanger, and Krizan (2006) and Autor et al. (2020). Average earnings declines accompanied by substantial increases in employment in these low-paying industries have led to especially large increases in inequality. 4 4 ^(4){ }^{4}
十一种低薪(底部四分之一)行业构成了前 10%在不平等贡献方面的其余部分。这些行业的组合占行业间不平等增加的 44.1%。其中,仅有这 11 种行业中的 3 种:餐馆和其他餐饮场所(7225)、其他综合商品商店(4529)和杂货店(4451)就占据了超过四分之一的增加。这些行业在最近几十年经历了重大变化,从单一企业转变为大型全国连锁店,参见 Foster、Haltiwanger 和 Krizan(2006)以及 Autor 等(2020)。这些低薪行业的平均收入下降伴随着就业的大幅增加,导致不平等的特别大幅上升。
A distinctive feature of the dominant 10 percent of industries is that they exhibit a sharp increase in the share of employment at mega firms, which we define as firms with more than 10,000 employees. Strikingly, the remaining 90 percent of industries exhibit small declines in the share of employment at mega firms. For the low-paying dominant industries, there is a decline in the size-earnings premium. For the high-paying dominant industries, mega firms increased earnings substantially relative to both small firms in the same industry and to earnings of the average industry. Thus, we find that the rise in “superstar” firms (e.g., Autor et al. 2020) is concentrated in these dominant industries with accompanying systematic changes in the size-earnings premia.
主导 10%行业的一个显著特征是,它们在大型企业(我们定义为员工超过 10,000 人的企业)中的就业份额急剧增加。令人惊讶的是,其余 90%的行业在大型企业的就业份额上则出现小幅下降。对于低薪主导行业,规模-收入溢价有所下降。对于高薪主导行业,大型企业的收入相较于同一行业的小型企业和平均行业的收入大幅增加。因此,我们发现“超级明星”企业(例如,Autor 等,2020)的崛起集中在这些主导行业,并伴随规模-收入溢价的系统性变化。
The top 30 industries also exhibit distinctive changes in the education and occupations of their workers. Especially notable are large increases in the share
前 30 个行业在其员工的教育和职业方面也表现出显著变化。尤其值得注意的是,员工的教育水平和职业类型的比例大幅增加。
of workers with a bachelor’s or advanced degree in the top 19 high-paying industries. These industries also exhibit sharp increases in high-paying occupations such as Business and Financial Operations (SOC 13), Computer and Mathematical Science (15), and strong declines in Office and Administrative Support (43). The top 11 low-paying industries exhibit substantial increases in low-paying occupations such as Food Preparation and Serving Related (35) and Personal Care and Service (39). Strikingly, nearly all of the growth in the employment share of high-paying occupations occurs in the top 19 high-paying industries, and all of the growth in low-paying occupations occurs in the top 11 low-paying industries.
拥有学士或更高学位的工人在前 19 个高薪行业中的比例。这些行业还表现出高薪职业的急剧增加,例如商业和金融运营(SOC 13)、计算机和数学科学(15),以及办公室和行政支持(43)的显著下降。前 11 个低薪行业在低薪职业方面也出现了显著增加,例如食品准备和服务相关(35)以及个人护理和服务(39)。值得注意的是,几乎所有高薪职业就业份额的增长都发生在前 19 个高薪行业中,而所有低薪职业的增长则发生在前 11 个低薪行业中。
We provide further insights about rising between-industry inequality by quantifying the role of industry premia and between-industry sorting. By sorting, we refer to the frequency with which high (low) wage workers are employed in industries with high (low) wage premia, as well as the concentration of high (low) wage workers in particular industries. We follow Song et al. (2019) in using an Abowd, Kramarz, and Margolis (1999)—hereafter, AKM—decomposition of earnings as our initial method for quantifying these effects. We also characterize these effects using a standard human capital equation (see Hoffman, Lee, and Lemieux 2020) that relates earnings to age, education, occupation, and industry effects. This takes advantage of linked data from the Current Population Survey (CPS). Finally, we describe how earnings vary with industry and occupation using published aggregates from the Occupational Employment and Wage Statistics (OEWS). Regardless of our approach, we find that an increase in sorting plays the most important role. Rising dispersion in between-industry premia also make an important direct contribution in addition to the indirect contribution via sorting.
我们通过量化行业溢价和行业间排序的作用,进一步深入探讨了行业间不平等的上升。这里的排序指的是高(低)工资工人在高(低)工资溢价行业中的就业频率,以及高(低)工资工人在特定行业中的集中程度。我们遵循 Song 等人(2019)的研究,使用 Abowd、Kramarz 和 Margolis(1999)——以下简称 AKM——的收入分解作为量化这些影响的初步方法。我们还使用标准的人力资本方程(参见 Hoffman、Lee 和 Lemieux 2020)来描述这些影响,该方程将收入与年龄、教育、职业和行业效应相关联。这利用了来自当前人口调查(CPS)的链接数据。最后,我们使用职业就业和工资统计(OEWS)发布的汇总数据描述收入如何随行业和职业而变化。无论我们采用何种方法,我们发现排序的增加发挥了最重要的作用。行业间溢价的上升差异也在排序的间接贡献之外,做出了重要的直接贡献。
An increasing covariance between industry premia and earnings premia for worker characteristics for an industry underlies the dominant role of sorting. This increased covariance depends critically on the presence of industry premia which have been an active part of the debate in the literature since at least Krueger and Summers (1988). In this respect, our findings are related to but distinct from both the earlier and more recent literature on inter-industry earnings differentials. Especially related is the recent work of Card, Rothstein, and Yi (2024), who assess the relationship between industry premia and the cross-sectional dispersion in earnings using matched employer-employee data.
行业溢价与工人特征的收益溢价之间日益增加的协方差是排序主导作用的基础。这种协方差的增加在很大程度上依赖于行业溢价的存在,自克鲁格和萨默斯(1988)以来,这一直是文献中活跃的讨论主题。在这方面,我们的发现与早期和近期关于行业间收益差异的文献相关,但又有所不同。特别相关的是卡德、罗斯坦和易(2024)的最新研究,他们使用匹配的雇主-员工数据评估行业溢价与收益的横截面离散度之间的关系。
Our work is distinct from the inter-industry earnings differential literature in two important and related ways. First, we focus on the change in inequality rather than the level. 5 5 ^(5){ }^{5} Second, we distinguish the direct effects of industry premia from the indirect contribution from the covariance of industry premia and the earnings premia for worker characteristics in the industry. Taken together, the direct and indirect effects of industry premia account for most of the dominant role of rising between industry dispersion. Moreover, these effects are concentrated in the top 30 industries.
我们的研究在两个重要且相关的方面与行业间收入差异文献有所不同。首先,我们关注的是不平等的变化而非水平。其次,我们区分了行业溢价的直接效应与行业内工人特征的收入溢价的协方差所带来的间接贡献。综合来看,行业溢价的直接和间接效应占据了行业间差异上升的主导作用。此外,这些效应集中在前 30 个行业。
A rising between-industry covariance requires that industry premia are changing in the same direction as the earnings premia for observable characteristics such as education and occupation, as well as unobservable worker effects. A major finding of our paper is that this property holds using either the AKM approach (for
行业间协方差的上升要求行业溢价与可观察特征(如教育和职业)以及不可观察的工人效应的收益溢价朝同一方向变化。我们论文的一个主要发现是,这一特性在使用 AKM 方法时成立。
unobservable effects) or education and occupation effects (using the CPS-LEHD and OEWS data). We find, for example, that Other Information Services (5191, where internet portals and search engine companies are located) has increases in both the industry premia and earnings premia for worker effects of over 20 percent over our sample period.
不可观察的效应)或教育和职业效应(使用 CPS-LEHD 和 OEWS 数据)。例如,我们发现其他信息服务(5191,互联网门户和搜索引擎公司所在的行业)在我们的样本期间,行业溢价和工人效应的收入溢价均增加了超过 20%。
Our findings imply that understanding rising earnings inequality during the last several decades requires understanding the restructuring of how firms organize themselves in a relatively small set of industries. The importance of select industries is well-recognized in the study of productivity but has been neglected in the inequality literature, which has focused on changes in the relative demand and supply of workers with different skills that are used to accomplish different tasks. While our results are not inconsistent with that view, our findings highlight that the skill and task mix varies dramatically across firms and especially industries reflecting the different ways that production is organized. As such, when structural changes such as adoption of new technologies or globalization impact the relative demand of workers by skills and tasks this manifests itself through changes in the organization of production in distinct ways across detailed industries.
我们的研究结果表明,理解过去几十年收入不平等上升的原因需要理解企业在相对少数行业中如何重组自身的方式。特定行业的重要性在生产力研究中得到了广泛认可,但在收入不平等文献中却被忽视,后者主要关注不同技能工人的相对需求和供给的变化,这些工人被用于完成不同的任务。虽然我们的结果与这一观点并不矛盾,但我们的发现强调了技能和任务的组合在不同企业,尤其是行业之间的显著差异,反映了生产组织的不同方式。因此,当结构性变化(如新技术的采用或全球化)影响不同技能和任务工人的相对需求时,这种影响通过各个细分行业中生产组织的变化以不同方式表现出来。
Our findings provide a related but distinct perspective on rising polarization (see Autor, Katz, and Kearney 2006, 2008; Goos and Manning 2007; Acemoglu and Autor 2011). We find that the rise in overall earnings inequality from 1996 to 2018 is accounted for by the polarization of industry wage structures in key industries. In these key industries, high-wage workers are increasingly sorted to high-wage industries with rising industry premia and are increasingly working with each other. Low-wage workers are increasingly sorted into low-wage industries, with rising employment and falling industry wage premia and are increasingly working with each other. This role of industry was not detectable in the earlier literature given the limitations of industry codes in the household survey data that has been the workhorse of the inequality literature (see Haltiwanger, Hyatt, and Spletzer 2023). Using high-quality industry codes that are inherent in our administrative matched employer-employee data, we are able to identify and analyze this polarization of industry wage structures.
我们的研究结果提供了一个相关但不同的视角来看待日益加剧的两极化现象(参见 Autor, Katz 和 Kearney 2006, 2008;Goos 和 Manning 2007;Acemoglu 和 Autor 2011)。我们发现,从 1996 年到 2018 年,整体收入不平等的上升是由关键行业的行业工资结构两极化所导致。在这些关键行业中,高工资工人越来越多地被分配到高工资行业,行业溢价上升,并且他们之间的合作也在增加。低工资工人则越来越多地被分配到低工资行业,随着就业上升和行业工资溢价下降,他们之间的合作也在增加。由于家庭调查数据中行业代码的局限性,早期文献中无法检测到行业的这一作用,而这些数据一直是不平等文献的主要依据(参见 Haltiwanger, Hyatt 和 Spletzer 2023)。利用我们行政匹配的雇主-雇员数据中固有的高质量行业代码,我们能够识别和分析这种行业工资结构的两极化现象。

I. Data 一. 数据

A. LEHD Data and the Analysis Sample
A. LEHD 数据与分析样本

For our core analysis, we use Longitudinal Employer-Household Dynamics (LEHD) linked employer-employee data, which is created by the US Census Bureau as part of the Local Employment Dynamics federal-state partnership. The LEHD data are derived from unemployment insurance (UI) wage records and quarterly census of employment and wages (QCEW) data, see Abowd et al. (2009). Every quarter, employers who are subject to state UI laws-approximately 98 percent of all private sector employers, plus state and local governments-are required to report information on their workers (the wage records, which record the quarterly earnings of every worker in the firm) and their workplaces (the QCEW, which provides the industry and location of each establishment). The wage records and the QCEW data submitted by the states to the US Census Bureau are enhanced
对于我们的核心分析,我们使用了由美国人口普查局创建的纵向雇主-家庭动态(LEHD)链接雇主-员工数据,这是地方就业动态联邦-州合作的一部分。LEHD 数据来源于失业保险(UI)工资记录和季度就业与工资普查(QCEW)数据,参见 Abowd 等(2009)。每个季度,受州失业保险法约束的雇主——大约 98%的所有私营部门雇主,以及州和地方政府——必须报告其员工(工资记录,记录公司中每位员工的季度收入)和工作场所(QCEW,提供每个机构的行业和位置)信息。各州提交给美国人口普查局的工资记录和 QCEW 数据得到了增强。

with census and survey microdata in order to incorporate information about worker demographics (age and gender) and the firm (firm age and firm size). A job in the LEHD data is defined as a worker-employer match, and earnings is defined as the amount earned from that job during a given quarter.
通过普查和调查微观数据,以纳入有关工人特征(年龄和性别)以及公司的信息(公司年龄和公司规模)。LEHD 数据中的工作被定义为工人与雇主的匹配,收入被定义为在特定季度内从该工作中获得的金额。
Because states have joined the LEHD program at different times, the length of the time series varies by state. We use data from 18 states that have data from 1996:I through 2018:IV, which gives us data for 23 years. 6 6 ^(6){ }^{6} We create annual person-level data as Y t i = j q t Y q t i j Y t i = j q t Y q t i j Y_(t)^(i)=sum_(j)sum_(q in t)Y_(qt)^(ij)Y_{t}^{i}=\sum_{j} \sum_{q \in t} Y_{q t}^{i j}, which sums the earnings Y Y YY that worker i i ii receives from firm j j jj in any quarter q q qq during year t t tt. We use the federal employer identification number (EIN) as the firm identifier, and include only private sector jobs. 7 7 ^(7){ }^{7} We follow Abowd, McKinney, and Zhao (2018) and delete any worker with 12 or more jobs in a given year. A worker’s employer is the firm that contributes the most earnings in a given year. This yields 1.395 billion person-year observations (averaging about 61 million persons per year).
由于各州在不同时间加入 LEHD 项目,时间序列的长度因州而异。我们使用来自 18 个州的数据,这些州的数据覆盖 1996 年第一季度至 2018 年第四季度,提供了 23 年的数据。我们创建年度个人级别数据,计算工人在某一年中从某公司在任何一个季度获得的收入总和。我们使用联邦雇主识别号码(EIN)作为公司标识符,仅包括私营部门的工作。我们遵循 Abowd、McKinney 和 Zhao(2018)的做法,删除在某一年中有 12 个或更多工作的工人。工人的雇主是该年中贡献最多收入的公司。这产生了 13.95 亿个个人年观察值(平均每年约 6100 万人)。
We create our analytical dataset following the sample restrictions of Song et al. (2019). We restrict to persons aged 20-60 with annual real (2013 = 100 PCE deflator) earnings greater than $ 3 , 770 $ 3 , 770 $3,770\$ 3,770 ( = 13 = 13 =13=13 weeks × 40 × 40 xx40\times 40 hours per week × $ 7.25 × $ 7.25 xx$7.25\times \$ 7.25 minimum wage). We top-code annual earnings at the 99.999 percent value (for anyone with earnings in the top 0.001 percent, we replace their earnings with the mean earnings of the top 0.001 percent). Our dataset has 1.048 billion person-year observations (an average of 45.6 million persons per year). We use real annual log earnings y t i = ln ( Y t i ) y t i = ln Y t i y_(t)^(i)=ln(Y_(t)^(i))y_{t}^{i}=\ln \left(Y_{t}^{i}\right). We define three seven-year intervals (1996-2002, 2004-2010, 2012-2018), reducing the sample to 959 million person-year observations. 8 8 ^(8){ }^{8}
我们按照 Song 等人(2019)的样本限制创建我们的分析数据集。我们限制在年龄为 20-60 岁、年实际(2013 = 100 PCE 平减指数)收入大于 $ 3 , 770 $ 3 , 770 $3,770\$ 3,770 = 13 = 13 =13=13 × 40 × 40 xx40\times 40 小时每周 × $ 7.25 × $ 7.25 xx$7.25\times \$ 7.25 最低工资)的人。我们将年收入的上限设定为 99.999 百分位值(对于收入在前 0.001 百分位的人,我们用前 0.001 百分位的平均收入替代他们的收入)。我们的数据集包含 10.48 亿人年观察值(每年平均 4560 万人)。我们使用实际年对数收入 y t i = ln ( Y t i ) y t i = ln Y t i y_(t)^(i)=ln(Y_(t)^(i))y_{t}^{i}=\ln \left(Y_{t}^{i}\right) 。我们定义了三个七年区间(1996-2002 年、2004-2010 年、2012-2018 年),将样本减少到 9.59 亿人年观察值。 8 8 ^(8){ }^{8}
Again following Song et al. (2019), we restrict to firms that employ 20 or more persons in a given year. This reduces our sample to 763 million person-year observations. Due to Census Bureau disclosure rules (to avoid small samples), we further restrict the LEHD data to include firms that have at least one male and one female to allow for separate variance decompositions by gender. The final LEHD data used to create all our results contains 758 million person-year observations.
再次遵循宋等人(2019)的研究,我们限制在某一年雇佣 20 人或以上的公司。这将我们的样本减少到 7.63 亿人年观察值。由于人口普查局的披露规则(为避免小样本),我们进一步限制 LEHD 数据,仅包括至少有一名男性和一名女性的公司,以便按性别进行单独的方差分解。用于生成我们所有结果的最终 LEHD 数据包含 7.58 亿人年观察值。

B. CPS-LEHD, OEWS, and LBD
B. CPS-LEHD, OEWS 和 LBD

We consider additional datasets to supplement our analysis. The first integrates CPS data at the person-level into the LEHD data. 9 9 ^(9){ }^{9} The matched CPS-LEHD data provides estimates of changing earnings inequality and the contribution of industries that are very similar to the full LEHD when using LEHD earnings and industry codes. The advantage of this integrated data is that we can exploit observed
我们考虑额外的数据集来补充我们的分析。第一个数据集将个人层面的 CPS 数据整合到 LEHD 数据中。 9 9 ^(9){ }^{9} 匹配的 CPS-LEHD 数据提供了关于收入不平等变化的估计,以及与使用 LEHD 收入和行业代码时非常相似的行业贡献。这种整合数据的优势在于我们可以利用观察到的
education and occupation to help understand the role of rising between- industry dispersion to earnings inequality. One potential difference between the full LEHD sample and the CPS-LEHD sample is that we don’t make any minimum firm size restrictions. This feature enables us to show that this is not critical as results are robust in removing that restriction.
教育和职业有助于理解行业间收入差距上升的作用。完整的 LEHD 样本与 CPS-LEHD 样本之间的一个潜在差异是我们没有设定任何最低公司规模限制。这个特征使我们能够表明,这并不是关键,因为在去除该限制后,结果仍然稳健。
We also use published OEWS aggregates to explore the role of occupation in rising between-industry dispersion. We obtain a consistent time series for 22 occupation categories for 281 detailed (four-digit) NAICS industries for the years 2002-2016. 10 10 ^(10){ }^{10} As we will see, the top industries that contribute to rising dispersion in the OEWS strongly overlap with those in the LEHD data.
我们还使用已发布的 OEWS 汇总数据来探讨职业在行业间差异上升中的作用。我们为 2002 年至 2016 年间的 281 个详细(四位数)NAICS 行业的 22 个职业类别获得了一致的时间序列。 10 10 ^(10){ }^{10} 正如我们将看到的,导致 OEWS 中差异上升的主要行业与 LEHD 数据中的行业有很强的重叠。
We also use establishment-level microdata from the Longitudinal Business Database (LBD) to decompose the variance of real log earnings per worker across firms and industries. We use the LBD to assess the robustness of our LEHD-based analysis for coverage across states, our firm size restriction, and different business definitions (establishment and EIN). We also use the LBD to assess the role of between-industry changes in inequality in earlier periods (1980-1986 to 1996-2002).
我们还使用来自纵向商业数据库(LBD)的企业层面微观数据,来分解不同行业和企业之间实际工人收入对数的方差。我们利用 LBD 评估基于 LEHD 的分析在各州覆盖范围、企业规模限制和不同商业定义(企业和 EIN)上的稳健性。我们还使用 LBD 评估早期时期(1980-1986 至 1996-2002)行业间不平等变化的作用。

C. Industry Codes C. 行业代码

Industry codes play a fundamental role in our analysis. We define industry at the four-digit NAICS level. [ 1 ] [ 1 ] ^([1]){ }^{[1]} Our basic results use establishment-level industry codes from the BLS QCEW program. We assign each EIN its dominant industry. If an EIN has N > 1 N > 1 N > 1N>1 establishments with M M MM industry codes, where N M > 1 N M > 1 N >= M > 1N \geq M>1, the industry code with the maximum employment is chosen.
行业代码在我们的分析中发挥着基础性作用。我们在四位数字的 NAICS 级别上定义行业。我们的基本结果使用来自 BLS QCEW 项目的企业级行业代码。我们为每个 EIN 分配其主导行业。如果一个 EIN 有 N > 1 N > 1 N > 1N>1 个企业,且有 M M MM 个行业代码,其中 N M > 1 N M > 1 N >= M > 1N \geq M>1 ,则选择就业人数最多的行业代码。
Both BLS and the Census Bureau have strong incentives and extensive statistical programs to assign detailed and accurate industry codes at the establishment-level. For BLS, the QCEW program yields high quality industry codes from the Annual Refiling Survey as well as BLS business surveys. For the Census Bureau, periodic business surveys and the Economic Census provide rich sources of information on industry. BLS also shares their industry codes with the Census Bureau. The Census Bureau also obtains codes from SSA as part of the first step of identifying new businesses. The SSA industry code uses the information provided in the application for a new EIN (Form SS-4). While SSA industry codes are a useful first step, the Census Bureau has a clear hierarchy for industry codes in their business register and their business statistical programs, with the Economic Census (and related surveys) and BLS codes preferred (see Walker 1997).
美国劳工统计局(BLS)和人口普查局在为各个机构分配详细且准确的行业代码方面具有强烈的激励和广泛的统计项目。对于 BLS 而言,季度雇佣和工资报告(QCEW)项目通过年度重新登记调查以及 BLS 商业调查提供高质量的行业代码。对于人口普查局,定期的商业调查和经济普查提供了丰富的行业信息来源。BLS 还与人口普查局共享其行业代码。人口普查局还从社会保障局(SSA)获取代码,作为识别新企业的第一步。SSA 行业代码使用在申请新雇主识别号(EIN,表格 SS-4)时提供的信息。虽然 SSA 行业代码是一个有用的第一步,但人口普查局在其商业登记和商业统计项目中有明确的行业代码层级,经济普查(及相关调查)和 BLS 代码被优先考虑(见 Walker 1997)。
In complementary work, Haltiwanger and Spletzer (2020) show that the fraction of the variance of earnings accounted for by industry effects is very similar using either BLS or census codes but is much smaller using the industry codes census obtains from SSA. Moreover, Bloom et al. (2018) indicate that the same SSA
在补充研究中,Haltiwanger 和 Spletzer(2020)表明,使用 BLS 或人口普查代码时,行业效应所占的收入方差比例非常相似,但使用人口普查从社会保障局(SSA)获得的行业代码时则要小得多。此外,Bloom 等人(2018)指出,相同的 SSA
microdata used in Song et al. (2019) has missing industry codes for all new firms post 2002. Table 2 of Bloom et al. (2018) shows that the employment share of EINs with missing industry codes increased from 4 percent in 1980-1986 to 24 percent in 2007-2013 in their microdata. Our inference is that the high-quality industry codes from BLS and the Census Bureau yield a more accurate characterization of the role of industry variation in accounting for earnings dispersion. In the analysis that follows, the LEHD, CPS-LEHD, and OEWS data and analysis are all based on the QCEW sample frame with consistent industry codes. The LBD analysis uses the Census Bureau’s business register frame which differs from the QCEW frame. 12 12 ^(12){ }^{12}
在 Song 等人(2019)的研究中使用的微观数据在 2002 年后所有新公司的行业代码均缺失。Bloom 等人(2018)的表 2 显示,在他们的微观数据中,缺失行业代码的 EIN 的就业份额从 1980-1986 年的 4%增加到 2007-2013 年的 24%。我们的推断是,来自 BLS 和人口普查局的高质量行业代码能够更准确地描述行业变异在解释收入分散中的作用。在随后的分析中,LEHD、CPS-LEHD 和 OEWS 的数据和分析均基于具有一致行业代码的 QCEW 样本框架。LBD 分析使用的人口普查局的商业登记框架与 QCEW 框架不同。

II. Earnings Inequality within and between Firms and Industries
II. 企业和行业内部及之间的收入不平等

Unless otherwise stated, the empirical analysis uses the LEHD data infrastructure. When we turn to the CPS-LEHD, OEWS, and LBD analysis, we are explicit about the use of those data sources.
除非另有说明,实证分析使用 LEHD 数据基础设施。当我们转向 CPS-LEHD、OEWS 和 LBD 分析时,我们会明确说明使用这些数据源。
Letting i i ii index the worker, j j jj the firm, k k kk the industry, and t t tt the year, we can write the variance of real annual log earnings y y yy as
i i ii 表示工人, j j jj 表示公司, k k kk 表示行业, t t tt 表示年份,我们可以将实际年对数收入 y y yy 的方差写为

We estimate this variance decomposition separately by seven-year intervals denoted by p p pp. Note that, for a given interval p p pp, average earnings are represented by y ¯ p y ¯ p bar(y)^(p)\bar{y}^{p} for all workers, y ¯ j , k , p y ¯ j , k , p bar(y)^(j,k,p)\bar{y}^{j, k, p} for firm j j jj, and y ¯ k , p y ¯ k , p bar(y)^(k,p)\bar{y}^{k, p} for industry k k kk. Table 1 shows that for all workers, the variance of earnings increases from 0.794 in the first interval (1996-2002) to 0.915 in the third interval (2012-2018). Of this 0.121 increase, 0.018 ( 14.9 percent) occurs within firms, 0.028 ( 23.1 percent) between firms but within industries, and 0.075 ( 61.9 percent) between industries. These estimates state that between-industry variance growth accounts for 72.8 percent ( = 0.075 / ( 0.028 + 0.075 ) ) ( = 0.075 / ( 0.028 + 0.075 ) ) (=0.075//(0.028+0.075))(=0.075 /(0.028+0.075)) of the between-firm contribution to increasing inequality. 13 13 ^(13){ }^{13}
我们按七年间隔单独估计这一方差分解,间隔用 p p pp 表示。注意,对于给定的间隔 p p pp ,所有工人的平均收入用 y ¯ p y ¯ p bar(y)^(p)\bar{y}^{p} 表示,企业 j j jj 的平均收入用 y ¯ j , k , p y ¯ j , k , p bar(y)^(j,k,p)\bar{y}^{j, k, p} 表示,行业 k k kk 的平均收入用 y ¯ k , p y ¯ k , p bar(y)^(k,p)\bar{y}^{k, p} 表示。表 1 显示,对于所有工人,收入的方差从第一个间隔(1996-2002)的 0.794 增加到第三个间隔(2012-2018)的 0.915。在这 0.121 的增加中,0.018(14.9%)发生在企业内部,0.028(23.1%)发生在企业之间但在行业内部,0.075(61.9%)发生在行业之间。这些估计表明,行业间方差增长占企业间对不平等增加的贡献的 72.8% ( = 0.075 / ( 0.028 + 0.075 ) ) ( = 0.075 / ( 0.028 + 0.075 ) ) (=0.075//(0.028+0.075))(=0.075 /(0.028+0.075)) 13 13 ^(13){ }^{13}
It is important to distinguish between a cross-sectional variance decomposition versus a growth decomposition. At a given point in time, the majority ( 58.0 percent to 64.6 percent) of variance is within firms. However, this within-firm person component of earnings variance is becoming less important over time. Growth in the within-industry firm component is positive but much smaller than between-industry
区分横截面方差分解与增长分解是很重要的。在某一时点,大多数(58.0%到 64.6%)的方差存在于公司内部。然而,随着时间的推移,这种公司内部的个人收入方差成分变得越来越不重要。行业内部公司的增长是正的,但远小于行业之间的增长。
Table 1-Variance Decomposition, by SEVEn-Year Interval
表 1-方差分解,按七年间隔

区间 1:1996-2002 (1)
Interval 1: 1996-2002
(1)
Interval 1: 1996-2002 (1)| Interval 1: 1996-2002 | | :--- | | (1) |

区间 2:2004-2010 (2)
Interval 2: 2004-2010
(2)
Interval 2: 2004-2010 (2)| Interval 2: 2004-2010 | | :--- | | (2) |

区间 3:2012-2018 (3)
Interval 3: 2012-2018
(3)
Interval 3: 2012-2018 (3)| Interval 3: 2012-2018 | | :--- | | (3) |

增长:1 至 3(4)
Growth:
1 to 3
(4)
Growth: 1 to 3 (4)| Growth: | | :--- | | 1 to 3 | | (4) |
Panel A. Variance, in levels
面板 A. 方差,水平
Total variance 总方差 0.794 0.862 0.915 0.121
Within-firm 企业内部 0.512 0.532 0.531 0.018
Between-firm, within-industry
企业间,行业内
0.112 0.127 0.140 0.028
Between-industry 行业间 0.170 0.203 0.245 0.075
Panel B. Variance, as percent of total
面板 B。方差,占总数的百分比
Within-firm 企业内部 64.6 61.7 58.0 14.9
Between-firm, within-industry
企业间,行业内
14.0 14.7 15.3 23.1
Between-industry 行业间 21.4 23.6 26.8 61.9
Panel C. Other measures
面板 C. 其他措施
Sample size (millions) 样本大小(百万) 239.4 249.2 269.7
Number of firms (thousands)
企业数量(千家)
470 460 466
Number of NAICS industries
北美行业分类系统(NAICS)行业数量
301 301 301
"Interval 1: 1996-2002 (1)" "Interval 2: 2004-2010 (2)" "Interval 3: 2012-2018 (3)" "Growth: 1 to 3 (4)" Panel A. Variance, in levels Total variance 0.794 0.862 0.915 0.121 Within-firm 0.512 0.532 0.531 0.018 Between-firm, within-industry 0.112 0.127 0.140 0.028 Between-industry 0.170 0.203 0.245 0.075 Panel B. Variance, as percent of total Within-firm 64.6 61.7 58.0 14.9 Between-firm, within-industry 14.0 14.7 15.3 23.1 Between-industry 21.4 23.6 26.8 61.9 Panel C. Other measures Sample size (millions) 239.4 249.2 269.7 Number of firms (thousands) 470 460 466 Number of NAICS industries 301 301 301 | | Interval 1: 1996-2002 <br> (1) | Interval 2: 2004-2010 <br> (2) | Interval 3: 2012-2018 <br> (3) | Growth: <br> 1 to 3 <br> (4) | | :---: | :---: | :---: | :---: | :---: | | Panel A. Variance, in levels | | | | | | Total variance | 0.794 | 0.862 | 0.915 | 0.121 | | Within-firm | 0.512 | 0.532 | 0.531 | 0.018 | | Between-firm, within-industry | 0.112 | 0.127 | 0.140 | 0.028 | | Between-industry | 0.170 | 0.203 | 0.245 | 0.075 | | Panel B. Variance, as percent of total | | | | | | Within-firm | 64.6 | 61.7 | 58.0 | 14.9 | | Between-firm, within-industry | 14.0 | 14.7 | 15.3 | 23.1 | | Between-industry | 21.4 | 23.6 | 26.8 | 61.9 | | Panel C. Other measures | | | | | | Sample size (millions) | 239.4 | 249.2 | 269.7 | | | Number of firms (thousands) | 470 | 460 | 466 | | | Number of NAICS industries | 301 | 301 | 301 | |
Notes: Authors’ tabulations of LEHD microdata. Tabulations include workers with annual real earnings > $ 3 , 770 > $ 3 , 770 > $3,770>\$ 3,770 in EINs with 20 or more employees. See equation (1) for definitions.
注:作者对 LEHD 微数据的汇总。汇总包括年实际收入为 > $ 3 , 770 > $ 3 , 770 > $3,770>\$ 3,770 的 EIN 中拥有 20 名或更多员工的工人。有关定义,请参见方程(1)。

growth. The between-industry component grows substantially over time, from 21.4 percent in the first interval to 26.8 percent in the third interval. 14 14 ^(14){ }^{14}
增长。行业间的组成部分随着时间的推移显著增长,从第一个区间的 21.4%增加到第三个区间的 26.8%。

III. The Industries That Drive Increasing Inequality
III. 驱动日益不平等的产业

We have demonstrated that almost two-thirds of the growth in inequality occurs between rather than within industries. We now propose a measure of a particular industry’s contribution to inequality and assess how this varies across industries. Between-industry variance growth can be expressed as
我们已经证明,几乎三分之二的不平等增长发生在行业之间,而不是行业内部。我们现在提出一种衡量特定行业对不平等贡献的指标,并评估这种贡献在各行业之间的差异。行业间方差增长可以表示为
Δ var [ y ¯ k , p y ¯ p ] between-industry variance growth = industry k 's contribution to between-industry variance growth 301 Δ ( N k , p N p ) ( y ¯ k , p y ¯ p ) 2 employment share ative earnings , Δ var y ¯ k , p y ¯ p  between-industry   variance growth  =  industry  k  's contribution to   between-industry variance growth  301 Δ N k , p N p y ¯ k , p y ¯ p 2  employment   share ative   earnings  , ubrace(Delta var[ bar(y)^(k,p)- bar(y)^(p)]ubrace)_({:[" between-industry "],[" variance growth "]:})=sum_({:[" industry "],[k^(')" 's contribution to "],[" between-industry variance growth "]:})^(301)ubrace(Delta (ubrace(((N^(k,p))/(N^(p)))ubrace))(ubrace(( bar(y)^(k,p)- bar(y)^(p))^(2)ubrace))ubrace)_({:[" employment "],[" share ative "],[" earnings "]:}),\underbrace{\Delta \operatorname{var}\left[\bar{y}^{k, p}-\bar{y}^{p}\right]}_{\begin{array}{c} \text { between-industry } \\ \text { variance growth } \end{array}}=\sum_{\begin{array}{c} \text { industry } \\ k^{\prime} \text { 's contribution to } \\ \text { between-industry variance growth } \end{array}}^{301} \underbrace{\Delta \underbrace{\left(\frac{N^{k, p}}{N^{p}}\right)} \underbrace{\left(\bar{y}^{k, p}-\bar{y}^{p}\right)^{2}}}_{\begin{array}{c} \text { employment } \\ \text { share ative } \\ \text { earnings } \end{array}},
where N N NN counts worker-employer-year combinations (i.e., employment), N k , p N k , p N^(k,p)N^{k, p} is total employment in industry k k kk in interval p p pp, and N p N p N^(p)N^{p} is total employment in interval p p pp. We define industry k k kk 's contribution to between-industry variance growth as Δ ( N k , p N p ) ( y ¯ k , p y ¯ p ) 2 Δ N k , p N p y ¯ k , p y ¯ p 2 Delta((N^(k,p))/(N^(p)))( bar(y)^(k,p)- bar(y)^(p))^(2)\Delta\left(\frac{N^{k, p}}{N^{p}}\right)\left(\bar{y}^{k, p}-\bar{y}^{p}\right)^{2}.
在这里, N N NN 计数工人-雇主-年份组合(即就业), N k , p N k , p N^(k,p)N^{k, p} 是在区间 p p pp 内行业 k k kk 的总就业, N p N p N^(p)N^{p} 是在区间 p p pp 内的总就业。我们将行业 k k kk 对行业间方差增长的贡献定义为 Δ ( N k , p N p ) ( y ¯ k , p y ¯ p ) 2 Δ N k , p N p y ¯ k , p y ¯ p 2 Delta((N^(k,p))/(N^(p)))( bar(y)^(k,p)- bar(y)^(p))^(2)\Delta\left(\frac{N^{k, p}}{N^{p}}\right)\left(\bar{y}^{k, p}-\bar{y}^{p}\right)^{2}
There are a total of 301 four-digit NAICS industries in our LEHD data. The distribution of contributions and employment shares by industry are depicted in Figure 1. Industries are rank ordered in the contribution to the change in between-industry earnings dispersion from the least to greatest. Strikingly, this distribution is highly skewed
在我们的 LEHD 数据中,总共有 301 个四位数的 NAICS 行业。各行业的贡献和就业份额的分布如图 1 所示。行业根据对行业间收入差异变化的贡献从最小到最大进行排名。值得注意的是,这一分布高度偏斜。
Figure 1. Industry Contribution to Between-Industry Inequality and Intitial Employment Share
图 1. 行业对行业间不平等和初始就业份额的贡献

Notes: Authors’ tabulations of LEHD microdata. Tabulations include workers with annual real earnings > $ 3 , 770 > $ 3 , 770 > $3,770>\$ 3,770 in EINs with 20 or more employees. See equation (2) for definitions.
注:作者对 LEHD 微数据的汇总。汇总包括年实际收入 > $ 3 , 770 > $ 3 , 770 > $3,770>\$ 3,770 的员工,EIN 中有 20 名或更多员工。有关定义,请参见方程(2)。

to the right. 15 15 ^(15){ }^{15} There are thirty industries in the right tail of Figure 1 that account for more than 1 percent of inequality growth and five industries in the left tail that account for less than -1 percent. These industries account for 98.1 percent and -7.2 percent of the overall contribution, respectively. The top 30 industries account for almost 40 percent of employment while the bottom 5 industries account for about 5 percent of employment. 16 16 ^(16){ }^{16}
向右。 15 15 ^(15){ }^{15} 图 1 的右尾有三十个行业占不平等增长的超过 1%。左尾有五个行业占不平等增长的不到-1%。这些行业分别占整体贡献的 98.1%和-7.2%。前 30 个行业几乎占就业的 40%,而后 5 个行业大约占就业的 5%。 16 16 ^(16){ }^{16}
To summarize these patterns, Table 2 groups industries by their contributions to increasing inequality. There are five industries that each contribute more than 5 percent of between-industry variance growth, accounting for 40.7 percent of between-industry variance growth. These five industries have 8.8 percent of total employment. An additional 25 industries each contribute between 1 percent and 5 percent of between-industry variance growth, accounting for 57.4 percent of
为了总结这些模式,表 2 根据各行业对不平等加剧的贡献对行业进行了分组。有五个行业每个行业对行业间方差增长的贡献超过 5%,合计占行业间方差增长的 40.7%。这五个行业的总就业人数占 8.8%。另外 25 个行业每个行业对行业间方差增长的贡献在 1%到 5%之间,合计占 57.4%。
Table 2-Industry Contributions to Between-Industry Variance Growth, by Variance Contribution
表 2-按方差贡献划分的行业对行业间方差增长的贡献

行业间方差增长的行业份额
Industry share of
between-industry
variance growth
Industry share of between-industry variance growth| Industry share of | | :--- | | between-industry | | variance growth |
Number of industries 行业数量

总就业份额 (%)
Total
employment
share (%)
Total employment share (%)| Total | | :---: | | employment | | share (%) |

行业间方差增长的总贡献
Total contribution
to between-industry
variance growth
Total contribution to between-industry variance growth| Total contribution | | :---: | | to between-industry | | variance growth |

行业间方差增长的总份额 (%)
Total share of
between-industry
variance growth (%)
Total share of between-industry variance growth (%)| Total share of | | :---: | | between-industry | | variance growth (%) |
> 5 % > 5 % > 5%>5 \% 5 industries 五个行业 8.8 0.8 ( 3 ) ( 3 ) (3)(3)
1 % 1 % 1%1 \% to 5%  1 % 1 % 1%1 \% 到 5% 25 industries 25 个行业 30.5 0.031 40.7
0.05 % 0.05 % 0.05%0.05 \% to 1%  0.05 % 0.05 % 0.05%0.05 \% 到 1% 71 industries 71 个行业 21.8 0.017 57.4
0.05 % 0.05 % -0.05%-0.05 \% to 0.05%  0.05 % 0.05 % -0.05%-0.05 \% 到 0.05% 145 industries 145 个行业 19.3 -0.000 22.3
< 0.05 % < 0.05 % < -0.05%<-0.05 \% 55 industries 55 个行业 19.7 -0.015 -0.1
Overall 总体 301 industries 301 个行业 100.0 0.075 -20.3
"Industry share of between-industry variance growth" Number of industries "Total employment share (%)" "Total contribution to between-industry variance growth" "Total share of between-industry variance growth (%)" > 5% 5 industries 8.8 0.8 (3) 1% to 5% 25 industries 30.5 0.031 40.7 0.05% to 1% 71 industries 21.8 0.017 57.4 -0.05% to 0.05% 145 industries 19.3 -0.000 22.3 < -0.05% 55 industries 19.7 -0.015 -0.1 Overall 301 industries 100.0 0.075 -20.3| Industry share of <br> between-industry <br> variance growth | Number of industries | Total <br> employment <br> share (%) | Total contribution <br> to between-industry <br> variance growth | Total share of <br> between-industry <br> variance growth (%) | | :--- | :--- | :---: | :---: | :---: | | $>5 \%$ | 5 industries | 8.8 | 0.8 | $(3)$ | | $1 \%$ to 5% | 25 industries | 30.5 | 0.031 | 40.7 | | $0.05 \%$ to 1% | 71 industries | 21.8 | 0.017 | 57.4 | | $-0.05 \%$ to 0.05% | 145 industries | 19.3 | -0.000 | 22.3 | | $<-0.05 \%$ | 55 industries | 19.7 | -0.015 | -0.1 | | Overall | 301 industries | 100.0 | 0.075 | -20.3 |
Notes: Authors’ tabulations of LEHD microdata. Tabulations include workers with annual real earnings > $ 3 , 770 > $ 3 , 770 > $3,770>\$ 3,770 in EINs with 20 or more employees. Employment shares are calculated as the average of 1996-2002 and 2012-2018 employment shares. See equation (2) for definitions.
注:作者对 LEHD 微数据的汇总。汇总包括在员工人数为 20 人或以上的 EIN 中年实际收入为 > $ 3 , 770 > $ 3 , 770 > $3,770>\$ 3,770 的工人。就业份额是 1996-2002 年和 2012-2018 年就业份额的平均值。有关定义,请参见公式(2)。

between-industry variance growth. In total, the top 30 industries-about 10 percent of all four-digit NAICS industries-account for 98.1 percent of between-industry variance growth and 39.2 percent of employment.
行业间方差增长。在所有四位数 NAICS 行业中,前 30 个行业约占 10%,却占据了 98.1%的行业间方差增长和 39.2%的就业。
As nearly two-thirds of the growth in US earnings dispersion has occurred between rather than within industries, these thirty industries account for most of increasing inequality. We provide detail about these thirty industries in Table 3 (sorted by NAICS). The largest contribution is from Restaurants and Other Eating Places (7225), which alone accounts for 16.9 percent of between-industry variance growth. The second-largest contribution occurs among Other General Merchandise Stores (4529), which accounts for 6.8 percent. While the most important two industries to increasing inequality tend to offer low-paying jobs, the other three industries that account for more than 5 percent of between-industry variance growth are high-paying: Software Publishers (5112), Computer Systems Design and Related Services (5415), and Management of Companies (5511).
由于美国收入差距增长的近三分之二发生在行业之间而非行业内部,这三十个行业占据了大部分日益加剧的不平等。我们在表 3 中提供了这三十个行业的详细信息(按 NAICS 排序)。最大的贡献来自餐馆和其他餐饮场所(7225),其单独占行业间方差增长的 16.9%。第二大贡献来自其他综合商品商店(4529),占 6.8%。虽然对收入不平等增长影响最大的两个行业往往提供低薪工作,但其他三个占行业间方差增长超过 5%的行业则是高薪行业:软件出版商(5112)、计算机系统设计和相关服务(5415)以及公司管理(5511)。
The top 30 industries reflect a small number of industry clusters that are notable for undergoing structural transformations that have been the subject of independent analysis. Eleven of the 19 high-paying industries have been defined as high-tech industries in terms of STEM intensity by Hecker (2005). These innovative industries in combination account for about one-third of the between-industry increase in earnings dispersion. As discussed in Oliner, Sichel, and Stiroh (2007) and Fernald (2014), these industries are characterized as the source of rapid technological advances in information and communications technologies. The transformation of the retail sector accounts for another one-third of the increase. These include the large industries in retail that have undergone the most significant business model transformations towards large national chains (see Foster, Haltiwanger, and Krizan 2006). High-paying Financial Services and Health Care Sectors also play an important role. For the latter, there are also low-paying health care sectors (e.g., retirement centers) making a substantial contribution. Outsourcing is connected to the top 30 industries. As documented by Dey, Houseman, and Polivka (2010) and Dorn et al. (2023), occupations such as protective service,
前 30 个行业反映了一小部分行业集群,这些集群因经历结构性转型而受到独立分析的关注。在 19 个高薪行业中,有 11 个被 Hecker(2005)定义为在 STEM 强度方面的高科技行业。这些创新行业的组合大约占行业间收入差异增加的三分之一。如 Oliner、Sichel 和 Stiroh(2007)以及 Fernald(2014)所讨论,这些行业被认为是信息和通信技术快速技术进步的源泉。零售部门的转型占收入差异增加的另一个三分之一。这些包括在商业模式上经历了向大型全国连锁转型的零售大行业(见 Foster、Haltiwanger 和 Krizan 2006)。高薪的金融服务和医疗保健部门也发挥着重要作用。对于后者,还有一些低薪的医疗保健部门(例如,退休中心)也做出了重要贡献。外包与前 30 个行业相关联。 正如 Dey、Houseman 和 Polivka(2010)以及 Dorn 等人所记录的。(2023),如保护服务等职业,
Table 3-Industry Contributions to Between-Industry Variance Growth, Top 30 Industries
表 3-行业对行业间方差增长的贡献,前 30 个行业
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline \multirow[b]{2}{}{NAICS} & \multirow[b]{2}{}{Industry title} & \multicolumn{2}{|l|}{Employment share percent} & \multicolumn{2}{|l|}{Relative earnings} & \multirow[t]{2}{*}{```
\hline \multirow[b]{2}{}{北美行业分类系统} & \multirow[b]{2}{}{行业标题} & \multicolumn{2}{|l|}{就业份额百分比} & \multicolumn{2}{|l|}{相对收入} & \multirow[t]{2}{*}{```

Share of between 在之间的份额
industry variance 行业方差
growth (%) 增长(%)
(5)
\hline & & Average
(1) & Change
(2) & Average
(3) & \begin{tabular}{l}
Change \\
(4)
\end{tabular} & \\
\hline 2111 & Oil and Gas Extraction & 0.3 & \(-0.0\) & 1.012 & 0.247 & 1.8 \\
\hline 2131 & Support Activities for Mining & 0.5 & 0.3 & 0.374 & 0.191 & 1.4 \\
\hline 3254 & Pharmaceutical Manufacturing & 0.5 & \(-0.1\) & 0.799 & 0.203 & 1.6 \\
\hline 3344 & Semiconductor Manufacturing & 0.8 & \(-0.5\) & 0.556 & 0.299 & 1.4 \\
\hline 4234 & Professional Equipment Wholesaler & 0.7 & \(-0.0\) & 0.557 & 0.190 & 1.9 \\
\hline 4441 & Building Material and Supplies & 0.9 & 0.1 & \(-0.293\) & \(-0.180\) & 1.5 \\
\hline 4451 & Grocery Stores & 2.4 & 0.0 & \(-0.378\) & \(-0.194\) & 4.7 \\
\hline 4481 & Clothing Stores & 0.7 & \(-0.0\) & \(-0.607\) & \(-0.244\) & 2.6 \\
\hline 4529 & Other General Merchndse. Stores & 1.4 & 1.5 & \(-0.539\) & \(-0.051\) & 6.8 \\
\hline 5112 & Software Publishers & 0.5 & 0.2 & 1.009 & 0.186 & 5.6 \\
\hline 5182 & Data Processing Services & 0.3 & \(-0.0\) & 0.545 & 0.301 & 1.3 \\
\hline 5191 & Other Information Services & 0.2 & 0.3 & 0.798 & 0.699 & 5.8 \\
\hline 5221 & Depository Credit Intermediate. & 2.1 & 0.0 & 0.189 & 0.234 & 2.5 \\
\hline 5231 & Securities Brokerage & 0.5 & \(-0.1\) & 0.866 & 0.204 & 1.1 \\
\hline 5239 & Other Financial Investment Activity & 0.3 & 0.1 & 0.834 & 0.388 & 3.3 \\
\hline 5241 & Insurance Carriers & 1.6 & \(-0.4\) & 0.488 & 0.167 & 2.3 \\
\hline 5413 & Archt. and Engineering Services & 1.2 & 0.1 & 0.469 & 0.161 & 2.6 \\
\hline 5415 & Computer Systems Design & 1.7 & 0.9 & 0.663 & 0.012 & 5.6 \\
\hline 5416 & Mgmt. and Scientific Services & 0.9 & 0.6 & 0.381 & 0.069 & 1.8 \\
\hline 5417 & Scientific Research Services & 0.8 & \(-0.1\) & 0.741 & 0.244 & 3.3 \\
\hline 5511 & Management of Companies & 2.0 & \(-0.1\) & 0.471 & 0.201 & 5.0 \\
\hline 5613 & Employment Services & 3.9 & 0.6 & \(-0.685\) & 0.017 & 2.5 \\
\hline 5617 & Services to Buildings and Dwell & 1.1 & 0.3 & \(-0.493\) & \(-0.002\) & 1.1 \\
\hline 6211 & Offices of Physicians & 1.7 & 0.5 & 0.254 & 0.099 & 1.6 \\
\hline 6216 & Home Health Care Services & 0.8 & 0.4 & \(-0.525\) & \(-0.016\) & 1.7 \\
\hline 6221 & General Medical and Hospitals & 4.5 & 0.5 & 0.205 & 0.162 & 4.2 \\
\hline 6233 & Continuing Care Retirement & 0.6 & 0.4 & \(-0.493\) & \(-0.001\) & 1.2 \\
\hline 6241 & Individual and Family Services & 0.8 & 0.6 & \(-0.490\) & \(-0.155\) & 3.5 \\
\hline 7139 & Othr. Amusement and Recreation & 0.6 & 0.1 & \(-0.594\) & \(-0.106\) & 1.7 \\
\hline 7225 & Restaurants and Othr. Eat Places & 4.9 & 2.0 & \(-0.739\) & \(-0.027\) & 16.9 \\
\hline
\end{tabular}

Notes: Authors' tabulations of LEHD microdata. Tabulations include workers with annual real earnings \(>\$ 3,770\) in EINs with 20 or more employees. Average log earnings for industry \(k\) are relative to the economy average. The 1996-2002 and 2012-2018 intervals are averaged. Changes are the growth (or decline) from 1996-2002 to 2012-2018. See equation (2) for definitions. Top 11 low-paying industries in bold.
transportation and material moving, building and grounds cleaning and maintenance, and food preparation and serving are increasingly employed in industries that provide services to other firms. Important industries in the top 11 low-paying industries that fit this description include Employment Services (5613) and Services to Buildings and Dwellings (5617). As we will see below using the OEWS, we find patterns consistent with the shift of such occupations away from the top 19 high-paying industries.

What about the other 271 four-digit NAICS industries? Figure 1 highlights they make relatively little contribution. Using Table 2, there are 145 industries that each

Table 4-Industry Contributions to Between-Industry Variance Growth, by Average Earnings
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline \multirow[b]{2}{*}{Industry relative earnings} & \multirow[b]{2}{*}{Number of Industries} & \multirow[t]{2}{*}{Total employment share (1)} & \multirow[t]{2}{*}{Total contribution to between-industry growth
(2)} & \multirow[t]{2}{*}{Total share of between-industry growth} & \multicolumn{2}{|l|}{Shift-share} \\
\hline & & & & & \begin{tabular}{l}
Employment \\
(4)
\end{tabular} & Earnings
(5) \\
\hline Overall & 301 & 100.0\% & 0.075 & 100.0\% & \(14.0 \%\) & 86.0\% \\
\hline \multicolumn{7}{|l|}{Panel A. 30 industries with variance contribution \(>1 \%\)} \\
\hline High-paying & 19 & \(21.1 \%\) & 0.041 & \(54.1 \%\) & \(16.1 \%\) & 83.9\% \\
\hline Low-paying & 11 & \(18.1 \%\) & 0.033 & \(44.1 \%\) & \(68.3 \%\) & \(31.7 \%\) \\
\hline \multicolumn{7}{|l|}{Panel B. 271 industries with variance contribution \(\leq 1 \%\)} \\
\hline High-paying & 146 & \(34.9 \%\) & 0.001 & \(1.3 \%\) & & \\
\hline Low-paying & 125 & \(25.9 \%\) & 0.000 & 0.6\% & & \\
\hline
\end{tabular}

Notes: Authors' tabulations of LEHD microdata. Tabulations include workers with annual real earnings \(>\$ 3,770\) in EINs with 20 or more employees. Employment shares are calculated as the average of 1996-2002 and 2012-2018 employment shares. Industry \(k\) 's contribution to between-industry variance growth is specified in equation (2). The shift-share calculations for changing employment and earnings follow equation (3). Shift-share results are summed across industries and normalized by the total contribution so that the two components sum to 100 percent. The two rows for the 271 industries with variance contribution \(\leq 1\) percent have missing cells because the denominator for the shift-share decomposition is close to zero.
contribute approximately 0.0 percent (to be precise, greater than -0.05 percent and less than 0.05 percent) to between-industry variance growth. This says that almost one-half of all four-digit NAICS industries contribute essentially nothing to inequality growth. There are 71 industries that contribute between 0.05 percent and 1.0 percent, accounting for 22.3 percent of between-industry variance growth. These industries are basically offset by another 55 industries that have a negative contribution ( \(<-0.05\) percent), accounting for -20.3 percent of between-industry variance growth.

As seen in Table 4, the top 30 industries include 19 high-paying industries that account for 54.1 percent of between-industry variance growth, and 11 low-paying industries that account for 44.1 percent of between-industry variance growth. The other 271 industries that have small contributing and offsetting contributions to increasing inequality do not occur systematically among high-paying versus low-paying industries. 146 high-paying industries account for 1.3 percent of between-industry variance growth, and 125 low-paying industries account for only 0.6 percent of between-industry variance growth.

Changes in earnings and employment share determine an industry's contribution to growth in inequality. This is seen in the expression defining industry \(k\) 's contribution to between-industry variance growth: \(\Delta\left(\frac{N^{k, p}}{N^{p}}\right)\left(\bar{y}^{k, p}-\bar{y}^{p}\right)^{2}\). If an industry with relatively high earnings exhibits an earnings increase, then, ceteris paribus, inequality will increase. Analogously, inequality will increase if an industry with relatively low earnings exhibits an earnings decrease. In contrast, when average earnings in an industry moves closer to the overall average, inequality decreases.

Employment shares also determine industry-level contributions to inequality. An industry's earnings changes will have larger effects on inequality when its employment share is larger. Changes in an industry's employment share will have smaller effects on inequality when that industry's pay is more similar to the overall average.

Employment gains among very high- and very low-paying industries tend to increase inequality.

In Table 4, we report the relative importance of earnings changes versus employment changes using a shift share decomposition. Industry \(k\) 's contribution to between-industry variance growth is \(\Delta\left(\frac{N^{k, p}}{N^{p}}\right)\left(\bar{y}^{k, p}-\bar{y}^{p}\right)^{2}\). We can use a standard shift-share decomposition to express this change in terms of the components attributable to changes in employment versus earnings:
\[
\underbrace{\Delta\left(\frac{N^{k, p}}{N^{p}}\right)\left(\bar{y}^{k, p}-\bar{y}^{p}\right)^{2}}_{\begin{array}{c}
\text { industry } k^{\prime} \text { s contribution to } \\
\text { between-industry variance growth }
\end{array}}=\underbrace{\overline{\left(\bar{y}^{k, p}-\bar{y}^{p}\right)^{2}} \Delta\left(\frac{N^{k, p}}{N^{p}}\right)}_{\text {shift-share: employment }}+\underbrace{\frac{\overline{N^{k, p}}}{N^{p}} \Delta\left(\bar{y}^{k, p}-\bar{y}^{p}\right)^{2}}_{\text {shift-share: earnings }}
\]
where \(\overline{\left(\bar{y}^{k, p}-\bar{y}^{p}\right)^{2}}\) and \(\overline{\frac{N^{k, p}}{N^{p}}}\) are averages of intervals 1 and 3 . We do this for our top 30 industries, distinguished by high-paying and low-paying industries (we do not present the shift share estimates for the other 271 industries since the denominator of the shift share is very close to zero). Among the 19 high-paying industries, 83.9 percent of between-industry variance growth is accounted for by changing relative earnings, and the remaining 16.1 percent is accounted for by changing employment shares. Among the 11 low-paying industries, the relative importance of earnings versus employment is reversed: 68.3 percent of between-industry variance growth is accounted for by changing employment shares, and the remaining 31.7 percent is accounted for by changing relative earnings. These results highlight different explanations for why between-industry variance growth is increasing at the opposite tails of the earnings distribution. Inequality growth at the top of the earnings distribution is a story of increasing earnings, whereas inequality growth at the bottom of the earnings distribution is a story of increasing employment.

These two different explanations for increasing inequality among low- versus high-paying industries is evident in the earnings and employment changes of the thirty industries listed in Table 3. All of the 19 high-paying industries exhibit earnings increases during our time period. The most rapid growth is found in Other Information Services (5191), which had a 69.9 log point (101.2 percent) increase in relative earnings. \({ }^{17}\) Of the remaining high-paying industries, nine had earnings increases in excess of \(20 \log\) points ( 22.1 percent), six had increases between 10 (10.5 percent) and \(20 \log\) points, and three had increases less than \(10 \log\) points.

Most of the 11 low-paying industries exhibit earnings decreases, yet they are smaller in absolute value than the earnings increases among the high-paying industries. The only low-paying industry with a decline greater in magnitude than 20 \(\log\) points ( 22.1 percent) is Clothing Stores (4481), which had a 24.4 log point ( 27.6 percent) decrease in relative earnings. Of the remaining low-paying industries, four had earnings declines between 10 ( 10.5 percent) and 20 log points, and five had

\footnotetext{
\({ }^{17}\) We convert \(\log\) differentials to proportionate changes using the expression \(e^{x}-1\). For small differences, \(\log\) points are approximately equal to the percentage change.
}
earnings declines between 0 and 10 log points. One industry, Employment Services (5613), exhibited a relatively small increase in earnings.

On the other hand, changes in employment are more important for the 11 low-paying industries than for the 19 high-paying industries. Two low-paying industries in Table 3 stand out: Restaurants and Other Eating Places (7225) had a 2.0 percentage point increase in employment share, and Other General Merchandise Stores (4529) had a 1.5 percentage point increase in employment share. Eight of the other low-paying industries have smaller employment share increases (less than one percentage point), and one industry (Clothing Stores, 4481) had a declining employment share. Among the 19 high-paying industries, none had employment share increases exceeding 1 percentage point, ten had small employment share increases (less than 1 percentage point), and about one-half (9) of the high-paying industries had declining employment shares.

In the analysis that follows, we also use the CPS-LEHD integrated data, the OEWS and LBD data to provide further insights into the role of rising between-industry dispersion. It is worth highlighting that all of these alternative sources provide a similar quantification of the contribution of the top 30 industries listed above to rising between-industry dispersion. While more detail is provided below, the share of the between-industry increase in dispersion from 1996-2002 to 2012-2018 accounted for by the top 30 industries is 98.1 percent in the LEHD data, 105.5 percent in the CPS-LEHD data, 96.2 percent in the OEWS data, and 94 percent in the LBD data.

\section*{IV. Firm and Worker Composition in the Top 30 Industries}

\section*{A. Mega Firms}

Changes in the employment shares and size-earnings premia for mega (10,000+) firms play a critical role in accounting for rising between-industry earnings inequality. Table 5 shows descriptive statistics of employment and earnings in mega firms and non-mega firms in our four industry groups. One immediate result in Table 5 is that employment has shifted over time to the top 30 industries. The employment share of the top 30 industries increased by 8.2 percentage points, with most of this increase ( 6.0 percentage points) among the 11 low-paying industries. The employment share of the other 271 industries analogously declined by 8.2 percentage points, with most of this decline ( 6.8 percentage points) among the 146 high-paying industries.

The substantial increase in the employment share of the top 30 industries is driven by mega firms. This is evident in both Table 5 and Figure 2. Figure 2 shows the change in employment share by detailed size class for each of our four industry groups. \({ }^{18}\) The employment share of the 11 low-paying industries increased in every size class, with mega firms exhibiting the largest increase ( 2.5 percentage points). The 19 high-paying industries had a smaller increase in employment, but most of this increase ( 1.4 of a total of 2.2 percentage points) is accounted for by mega firms. Given the high average relative pay of mega firms in the high-paying

\footnotetext{
\({ }^{18}\) The corresponding employment share levels in the first interval (1996-2002) and in the third interval (2012-2018) are given in online Appendix Figure F1.
}

Table 5-Changes in Employment and Earnings, by Industry Earnings, Mega Firms versus Others
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline \multirow[t]{2}{*}{Industry relative earnings} & \multirow[t]{2}{*}{Number of industries} & \multirow[t]{2}{*}{Firm employment} & \multicolumn{2}{|l|}{Employment share} & \multicolumn{2}{|l|}{Relative earnings} \\
\hline & & & Average
(1) & Change
(2) & \begin{tabular}{l}
Average \\
(3)
\end{tabular} & Change
(4) \\
\hline \multicolumn{7}{|l|}{Panel A. 30 industries with variance contribution \(>1 \%\)} \\
\hline \multirow[t]{3}{*}{High-paying} & 19 industries & Any & \(21.1 \%\) & \(2.2 \%\) & 0.440 & 0.177 \\
\hline & & \(10,000+\) & \(3.8 \%\) & 1.4\% & 0.579 & 0.145 \\
\hline & & <10,000 & 17.3\% & 0.8\% & 0.410 & 0.174 \\
\hline \multirow[t]{3}{*}{Low-paying} & 11 industries & Any & 18.1\% & 6.0\% & -0.586 & -0.069 \\
\hline & & 10,000+ & 4.3\% & 2.5\% & -0.492 & \(-0.125\) \\
\hline & & \(<10,000\) & 13.8\% & \(3.5 \%\) & \(-0.613\) & \(-0.061\) \\
\hline \multicolumn{7}{|l|}{Panel B. 271 industries with variance contribution \(\leq 1 \%\)} \\
\hline \multirow[t]{3}{*}{High-paying} & 146 industries & Any & \(34.9 \%\) & \(-6.8 \%\) & 0.281 & 0.046 \\
\hline & & \(10,000+\) & \(3.9 \%\) & \(-1.2 \%\) & 0.646 & 0.042 \\
\hline & & \(<10,000\) & \(31.0 \%\) & \(-5.7 \%\) & 0.236 & 0.052 \\
\hline \multirow[t]{3}{*}{Low-paying} & 125 industries & Any & 25.9\% & \(-1.3 \%\) & -0.325 & -0.002 \\
\hline & & 10,000+ & 3.3\% & -0.5\% & -0.404 & \(-0.061\) \\
\hline & & <10,000 & 22.6\% & -0.9\% & \(-0.314\) & 0.006 \\
\hline
\end{tabular}

Notes: Authors' tabulations of LEHD microdata. Tabulations include workers with annual real earnings \(>\$ 3,770\) in EINs with 20 or more employees. Averages and changes use the employment shares and earnings from the 1996-2002 and 2012-2018 intervals. Average log earnings are relative to the economy average.
![](https://cdn.mathpix.com/cropped/2024_10_28_afef6c19081ade40df36g-15.jpg?height=714&width=1073&top_left_y=1186&top_left_x=325)

Figure 2. Change in Employment Share by Size Class, by Industry Group
Notes: Authors' tabulations of linked LEHD microdata. Tabulations include workers with annual real earnings \(>\$ 3,770\) in EINs with 20 or more employees. Changes in the employment shares are expressed in terms of percentage points. The denominator is total employment across all size classes and industry groups.
industries ( \(57.9 \log\) points, or 77.9 percent) and the low pay of mega firms ( -49.2 log points, or -63.6 percent) in the low-paying industries, these shifts
in employment to mega firms contributed to rising between-industry earnings inequality. \({ }^{19}\)

Mega firms also play a key role in the changing earnings of the top 30 industries. For the 11 low-paying industries, the relative pay of mega firms decreased by \(12.5 \log\) points ( 13.3 percent) compared to a decline of \(6.1 \log\) points ( 6.3 percent) for the non-mega firms. Both mega firms and non-mega firms in the 19 high-paying industries exhibit large earnings increases: \(14.5 \log\) points ( 15.6 percent) for mega firms and 17.4 log points ( 19.0 percent) for non-mega firms. Earnings at mega firms increased relative to the smallest firms in the top-paying industries but not by as much as the increase in relative earnings at large but not mega firms. \({ }^{20}\) In contrast, relative earnings increases at mega firms in the 146 remaining high-paying industries are modest ( 4.2 log points, or 4.3 percent) compared to \(14.5 \log\) points in the top 19 high-paying industries. Similarly, relative earnings declines at the mega firms in the remaining 125 low-paying industries are modest ( -6.1 log points, or -6.3 percent) compared to the \(-12.5 \log\) points in the top 11 low-paying industries.

\section*{B. Education and Occupation}

To shed light on the changing education composition of the top and bottom industries, we turn to the CPS-LEHD integrated data. Figure 3 shows the change in employment in the top 30 industries from 1996-2002 to 2012-2018. Both low- and high-paying industries had increases in the educational attainment of the workers that they employ. These changes were much more dramatic in the top 19 high-paying industries. The share of workers with bachelor's degrees at these high-paying industries increased by 7.0 percent, workers with advanced degree increased by 8.4 percent, and workers with a high school diploma declined by 8.7 percent. 21

For occupation, we turn to the OEWS published data. \({ }^{22}\) In Figure 4, we consider employment changes across all 22 occupation groups in the top 30 industries. The occupation groups are ranked from left to right by the changes in employment shares in the top 30 industries. There are substantial differences in the changing mix of occupations across the top 19 and bottom 11. The top 19 industries have large increases in Business and Financial Operations (13) and Computer and Mathematical Science (15) with accompanying large declines in Office and

\footnotetext{
\({ }^{19}\) Online Appendix Table H2 shows that the patterns of changes in employment shares by mega firms is robust to using the national and enterprise concepts available in the Business Dynamic Statistics (BDS). The BDS is derived from the LBD.
\({ }^{20}\) Online Appendix Figure F3 shows the cross-sectional size-earnings premia for the 1996-2002 and the 2012-2018 intervals. Among the top 19 high-paying industries, the size-earnings profile shifts upward, with increases in all size classes. Online Appendix F also provides more details of the changing patterns by firm size.
\({ }^{21}\) Online Appendix Figure D1 shows the distribution of employment by education group in the top 19 high-paying and 11 low-paying industries for the 1996-2002 and 2012-2018 periods. In the initial period, the top 19 high-paying industries had significantly higher shares of workers with a bachelor's or advanced degree while the top 11 low-paying industries had higher shares of workers with only a high school diploma or less. These differences became much more pronounced by 2012-2018 with sharp increases in the share of workers with bachelor's and advanced degrees in the top 19 high-paying industries and accompanying declines in the share of workers with a high school diploma or less.
\({ }^{22}\) Online Appendix Figure G1 shows that the CPS-LEHD yields similar patterns on changes in occupations by industry.
}
![](https://cdn.mathpix.com/cropped/2024_10_28_afef6c19081ade40df36g-17.jpg?height=726&width=1063&top_left_y=218&top_left_x=343)

Figure 3. Change in Employment Share by Educational Attainment, by Industry Group
Notes: Authors' tabulations of linked CPS-LEHD microdata. Tabulations include workers with annual real earnings \(>\$ 3,770\). Changes in the employment shares are expressed in terms of percentage points. The denominator is total employment in the respective industry group.
![](https://cdn.mathpix.com/cropped/2024_10_28_afef6c19081ade40df36g-17.jpg?height=721&width=1088&top_left_y=1212&top_left_x=330)

Figure 4. Change in Employment by Occupation and Industry Group
Note: Authors' calculations of published OEWS aggregates.

Administrative Support (43), a low wage occupation group. These patterns are consistent with the dramatic innovations in information and communication technology largely developed by top-paying industries. Related, much of the role of Office and Administrative Support (43) tasks in these high-paying industries are

Table 6-Changes in Employment and EarningS, by Industry EarningS, by Occupation Groups
\begin{tabular}{|c|c|c|c|c|c|}
\hline \multirow[b]{2}{*}{Industry group earnings} & \multirow[b]{2}{*}{Occupation group earnings} & \multicolumn{2}{|l|}{Employment share(\%)} & \multicolumn{2}{|l|}{Relative earnings} \\
\hline & & Average
(1) & Change
(2) & \begin{tabular}{l}
Average \\
(3)
\end{tabular} & Change
(4) \\
\hline \multicolumn{6}{|l|}{Panel A. 30 industries with variance contribution \(>1 \%\)} \\
\hline \multirow[t]{3}{*}{19 high-paying} & All & 16.3 & 1.3 & 0.405 & 0.107 \\
\hline & 12 high-paying & 9.7 & 2.0 & 0.707 & 0.067 \\
\hline & 10 low-paying & 6.6 & \(-0.7\) & \(-0.039\) & 0.029 \\
\hline \multirow[t]{3}{*}{11 low-paying} & All & 19.8 & 3.1 & \(-0.437\) & \(-0.010\) \\
\hline & 12 high-paying & 2.1 & 0.3 & 0.269 & 0.031 \\
\hline & 10 low-paying & 17.7 & 2.8 & \(-0.521\) & \(-0.013\) \\
\hline \multicolumn{6}{|l|}{Panel B. 251 industries with variance contribution \(\leq 1 \%\)} \\
\hline \multirow[t]{3}{*}{141 high-paying} & All & 30.0 & \(-2.8\) & 0.213 & 0.001 \\
\hline & 12 high-paying & 13.0 & \(-0.1\) & 0.496 & 0.005 \\
\hline & 10 low-paying & 17.1 & \(-2.7\) & \(-0.002\) & \(-0.036\) \\
\hline \multirow[t]{3}{*}{110 low-paying} & All & 33.9 & \(-1.6\) & \(-0.130\) & \(-0.011\) \\
\hline & 12 high-paying & 12.8 & \(-0.2\) & 0.238 & \(-0.017\) \\
\hline & 10 low-paying & 21.1 & \(-1.5\) & \(-0.353\) & \(-0.020\) \\
\hline
\end{tabular}

Notes: Authors' calculations of OEWS published aggregates. Changes in employment and earnings compare year intervals 2002-2003 with 2015-2016. Industry pay designations and contributions to variance growth follow the LEHD administrative records data.
now increasingly accomplished with adoption of information and communication technologies. Some of the reduction in Office and Administrative Support (43) may also reflect outsourcing of these tasks to service firms. Consistent with the latter, there have also been decreases in production and building and ground maintenance and cleaning. The employment share of the bottom 11 industries increases, with particularly strong growth in Food Preparation and Serving Related (35) and Personal Care and Service (39). These low-paying industries also exhibit a nontrivial decline in Management (11) occupations.

To get a sense of the contribution of occupations to industry-level wage inequality, we divide occupations into 2 broad categories, 12 that are high-paying relative to the overall average, and ten that are analogously low-paying (see Table 6). Occupational shifts appear to occur, at least broadly, at the industry level. Neanly all of the growth in the employment share of high-paying occupations occurs in the 19 high-paying industries, and all of the growth in low-paying occupations occurs in our 11 low-paying industries. Our 19 high-paying industries increase their employment share by 1.3 percent in the OEWS data. This reflects strong growth among our 12 high-paying occupations, whose share of employment grows by 2.0 percent. The employment share of our 10 low-paying occupations in these 19 high-paying industries contracts by 0.7 percent. The 11 low-paying industries increase their employment share by 3.1 percent. This mostly reflects an increase among our 10 low-paying occupations, whose share of overall employment increases by 2.8 percentage points.

Outside the thirty industries that contribute to inequality, low-paying occupations have a declining employment share. The employment share of our other 141 high-paying industries declines by 2.8 percentage points, and nearly all of this (2.7 percentage points) occurs among low-paying occupations. The employment
share of the 110 low-paying industries declines by 1.6 percentage points, which almost matches ( 1.5 percentage points) its decline in the share of low-paying occupations.

Table 6 also illustrates the role of industry-occupation pay differentials in rising inequality. Our 19 high-paying industries have a strong ( 40.5 log point) earnings differential. This reflects an even larger ( 70.7 log point) differential among our 12 high-paying occupations in these industries. These industries also had the highest earnings growth, both overall ( 10.7 log points), and especially among high-paying occupations ( 6.7 log points). Workers in our 10 low-paying occupations in these 19 high-paying industries had earnings gain as well ( 2.9 log points).

Earnings changed by relatively less among our top 11 low-paying industries. The overall change in earnings was a decline of only \(1.0 \log\) points. This reflects a gain in earnings of 3.1 log points among the relatively rate high-paying occupations. Low-paying occupations in these industries had an earnings decline of 1.3 log points. There were smaller changes among the other 251 industries (an increase of 0.1 log points among high-paying industries, and a decline of 1.1 log points among low-paying industries), with consistent declines in the earnings of low-paying occupations (of 3.6 log points, and 2.0 log points, respectively). \({ }^{23}\)

\section*{V. Inequality in Terms of Sorting and Pay Premia}

In this section, we present decompositions of earnings in the LEHD data using the AKM approach, in the CPS-LEHD data using a human capital approach, and in the OEWS data using a related but distinct approach given the limitation of only having occupation by industry data.

\section*{A. Sorting and Industry Premia Using AKM}

To understand the role of workers and firms in the generation of earnings inequality, we begin by using the linear model of AKM. We estimate our model separately for each of three seven-year intervals: 1996-2002, 2004-2010, and 2012-2018. Following Song et al. (2019), we assume that earnings \(y_{t}^{i, j, k, p}\) are the sum of the effect \(\theta^{i, p}\) of worker \(i\) in interval \(p\), a firm effect \(\psi^{j, k, p}\) when employed by employer \(j\) in industry \(k\) during interval \(p\), and a vector of time-varying observable characteristics \(X_{t}^{i, p}\) for worker \(i\) at time \(t\), which have distinct marginal effects \(\beta^{p}\) by interval \(p\). We express this as
\[
y_{t}^{i, j, k, p}=X_{t}^{i, p} \beta^{p}+\theta^{i, p}+\psi^{j, k, p}+\varepsilon_{t}^{i, j, k, p} .
\]

Our observable characteristics include a set of year dummies that capture calendar year effects on earnings. Following Card, Cardoso, and Kline (2016), we center age around 40 , include a quadratic and cubic transformation of worker age, and omit the

\footnotetext{
\({ }^{23}\) One potential issue is whether greater occupational detail may change the inferences in this section. However, there is a high degree of concentration of occupations across detailed industries. Online Appendix Figure G3 shows that for the two-digit occupations we use in the current analysis, occupations are highly concentrated in detailed industries. Online Appendix Figure G4 shows that using six-digit occupations, the median occupation has a top 20 -industry concentration ratio of 100 percent.
}
linear term. To solve this model, we implement the iterative method proposed by Guimarães and Portugal (2010).

The AKM approach to decomposing earnings has received substantial scrutiny in terms of the interpretation of the estimated person and firm effects. Recent research has highlighted the potential for limited mobility bias arising from the small number of transitions per firm. Bonhomme et al. (2022) find that limited mobility bias yields an upward bias in the variance of firm effects and a downward bias in the covariance between firm and worker effects. However, they find little bias on the contribution of the change in the role of premia and sorting for the change in inequality (and it is the latter that is the focus of our work). \({ }^{24}\) Our use of AKM makes our results directly comparable to Song et al. (2019) which enables us to highlight that the between-firm effects they identified are mostly occurring at the industry level. However, we also include (see Section VB) alternative decompositions based on the CPS-LEHD and OEWS data.

We use the AKM approach to decompose the between-firm components into those that occur within- and between-industries. \({ }^{25}\) To explore cross-industry differences, we calculate industry-level averages. For a given interval \(p\) (and hereafter omitting this superscript), we define the average worker effect in industry \(k\) as \(\bar{\theta}^{k}\), the average effect of observable characteristics as \(\bar{X}^{k} \beta\), and the average firm effect as \(\bar{\psi}^{k}\). Given this notation, it is possible to distinguish between how firm-level pay premia relate to within- versus between-industry earnings dispersion. This is given by
(5) \(\operatorname{var}\left[y_{t}^{i, j, k}\right]\)
\[
\begin{aligned}
& =\underbrace{\operatorname{var}\left[\theta^{i}-\bar{\theta}^{j, k}\right]+\operatorname{var}\left[X_{t}^{i} \beta-\bar{X}^{j, k} \beta\right]+2 \operatorname{cov}\left[\theta^{i}-\bar{\theta}^{j, k}, X_{t}^{i} \beta-\bar{X}^{j, k} \beta\right]}_{\text {within-firm person effect and observables }} \\
& +\underbrace{\operatorname{var}\left[\bar{\psi}^{k}\right]}_{\begin{array}{c}
\text { between-industry } \\
\text { pay premia }
\end{array}}+\underbrace{\operatorname{var}\left[\psi^{j, k}-\bar{\psi}^{k}\right]}_{\begin{array}{c}
\text { within-industry, } \\
\text { between-firm } \\
\text { pay premia }
\end{array}}+\underbrace{2 \operatorname{cov}\left[\bar{\theta}^{k}, \bar{\psi}^{k}\right]+2 \operatorname{cov}\left[\bar{\psi}^{k}, \bar{X}^{k} \beta\right]}_{\text {between-industry covariance }} \\
& +\underbrace{2 \operatorname{cov}\left[\left(\bar{\theta}^{j, k}-\bar{\theta}^{k}\right),\left(\psi^{j, k}-\bar{\psi}^{k}\right)\right]+2 \operatorname{cov}\left[\left(\psi^{j, k}-\bar{\psi}^{k}\right),\left(\bar{X}^{j, k} \beta-\bar{X}^{k} \beta\right)\right]}_{\text {within-industry, between-firm covariance }} \\
& +\underbrace{\operatorname{var}\left[\bar{\theta}^{j, k}-\bar{\theta}^{k}\right]+\operatorname{var}\left[\bar{X}^{j, k} \beta-\bar{X}^{k} \beta\right]+2 \operatorname{cov}\left[\left(\bar{\theta}^{j, k}-\bar{\theta}^{k}\right),\left(\bar{X}^{j, k} \beta-\bar{X}^{k} \beta\right)\right.}_{\text {within-industry, between-firm segregation }}] \\
& +\underbrace{\operatorname{var}\left[\bar{\theta}^{k}\right]+\operatorname{var}\left[\bar{X}^{k} \beta\right]+2 \operatorname{cov}\left[\bar{\theta}^{k}, \bar{X}^{k} \beta\right]}_{\text {between-industry segregation }}+\underbrace{\operatorname{var}\left[\varepsilon_{t}^{i, j, k}\right]}_{\text {residual (within-firm) }} .
\end{aligned}
\]

\footnotetext{
\({ }^{24}\) There has also been concern raised about exogenous mobility. Bonhomme et al. (2022) have highlighted that this is less of an issue than limited mobility bias. The reason is that what is required is that mobility is unrelated to the residual from the AKM model after controlling for person and firm effects.
\({ }^{25}\) Our between-industry decomposition builds on Song et al. (2019). In online Appendix E, we show the original Song et al. (2019) decomposition and compare our results to theirs.
}

Within-firm dispersion is given by the collection of terms in the first line of equation (5). Worker-level effects are given by
\[
\operatorname{var}\left[\theta^{i}-\bar{\theta}^{j, k}\right]+\operatorname{var}\left[X_{t}^{i} \beta-\bar{X}^{j, k} \beta\right]+2 \operatorname{cov}\left[\theta^{i}-\bar{\theta}^{j, k}, X_{t}^{i} \beta-\bar{X}^{j, k} \beta\right]
\]

Residual dispersion \(\operatorname{var}\left[\varepsilon_{t}^{i, j, k}\right]\) occurs within firms given the inclusion of firm effects.
The firm-level premia contributions can be decomposed into withinand between-industry components. Specifically, \(\quad \operatorname{var}\left[\psi^{j, k}\right]=\operatorname{var}\left[\bar{\psi}^{k}\right]+\) \(\operatorname{var}\left[\psi^{j, k}-\bar{\psi}^{k}\right]\), where \(\operatorname{var}\left[\bar{\psi}^{k}\right]\) reflects the between-industry dispersion in average firm effects, that is, industry-level pay premia. The remaining term \(\operatorname{var}\left[\psi^{j, k}-\bar{\psi}^{k}\right]\) captures the within-industry dispersion of firm-level pay premia.

In addition to pay premia, we can distinguish the within- versus between-industry components of a covariance contribution and segregation. The between-industry covariance contribution is defined as \(2 \operatorname{cov}\left[\bar{\theta}^{k}, \bar{\psi}^{k}\right]+2 \operatorname{cov}\left[\bar{\psi}^{k}, \bar{X}^{k} \beta\right]\), which reflects the extent to which highly paid workers are employed in industries with a high pay premium (and vice-versa). This is distinct from the within-industry covariance contribution \(2 \operatorname{cov}\left[\left(\bar{\theta}^{j, k}-\bar{\theta}^{k}\right),\left(\theta^{j, k}-\bar{\theta}^{k}\right)\right]+2 \operatorname{cov}\left[\left(\theta^{j, k}-\bar{\theta}^{k}\right),\left(\bar{X}^{j, k} \beta-\bar{X}^{k} \beta\right)\right]\). This is the component of the covariance contribution where relatively highly paid workers tend to work at high-paying firms, apart from industry-level differences. For example, workers and firms in the Restaurants and Other Eating Places (7225) industry may tend to have low worker effects, while those among Software Publishers (5112) may have high effects. The between-industry component reflects these industry level differences. The within-industry component reflects the extent to which relatively low- versus high-paid workers work for relatively low- versus high-paying firms in those industries.

Segregation also can be decomposed into its within- versus between-industry components. Between-industry segregation is given by industry-level average worker effects. Formally, this is expressed as \(\operatorname{var}\left[\bar{\theta}^{k}\right]+\operatorname{var}\left[\bar{X}^{k} \beta\right]+2 \operatorname{cov}\left[\bar{\theta}^{k}, \bar{X}^{k} \beta\right]\). This is the extent to which low- versus highly paid workers tend to work with each other. To continue with the previous example, Restaurants and Other Eating Places (7225) may employ workers with a low person effect, on average, while employers among Software Publishers (5112) may employ workers with a high average person effect. The extent to which this is related to the firm-level pay differences reflects the covariance component. The extent to which it reflects similar workers grouped together is segregation. Segregation that occurs within industries is expressed as \(\operatorname{var}\left[\bar{\theta}^{j, k}-\bar{\theta}^{k}\right]+\operatorname{var}\left[\bar{X}^{j, k} \beta-\bar{X}^{k} \beta\right]+2 \operatorname{cov}\left[\left(\bar{\theta}^{j, k}-\bar{\theta}^{k}\right),\left(\bar{X}^{j, k} \beta-\bar{X}^{k} \beta\right)\right]\).

While the covariance and segregation components are distinct in the above decomposition and potentially conceptually, in practice we find that the covariance and segregation components of the between industry increase in earnings inequality are closely related. For the complete set of four-digit NAICS industries, the correlation of the changes in the two components is 0.93 and for the top 30 industries it is 0.90 . This tight relationship reflects our finding that high-wage workers work increasingly in high wage industries and work together while low-wage workers work increasingly in low wage industries and work together. Given this tight relationship, we combine these covariance and segregation components into a combined sorting contribution and in turn for expositional convenience primarily focus our

Table 7-Industry-Enhanced Variance Decomposition
\begin{tabular}{lcccc}
\hline \hline & Interval 1: & Interval 2: & Interval 3: & Growth: \\
& \(1996-2002\) & \(2004-2010\) & \(2012-2018\) & 1 to 3 \\
& \((1)\) & \((2)\) & \((3)\) & \((4)\) \\
\hline Panel A. Variance, in levels & & & & \\
Total variance & 0.794 & 0.862 & 0.915 & 0.121 \\
Within-firm & 0.512 & 0.532 & 0.531 & 0.018 \\
Person effect and observables & 0.382 & 0.401 & 0.405 & 0.023 \\
\(\quad\) Residual & 0.130 & 0.131 & 0.125 & -0.005 \\
Between-firm, within-industry & 0.112 & 0.127 & 0.140 & 0.028 \\
\(\quad\) Firm sorting & 0.087 & 0.098 & 0.111 & 0.025 \\
\(\quad\) Firm pay premium & 0.025 & 0.029 & 0.028 & 0.004 \\
Between-industry & 0.170 & 0.203 & 0.245 & 0.075 \\
\(\quad\) Industry sorting & 0.137 & 0.162 & 0.201 & 0.065 \\
\(\quad\) Industry pay premium & 0.033 & 0.042 & 0.044 & 0.011 \\
Panel B. Variance, as percent of total & & & & \\
Within-firm & 64.6 & 61.7 & 58.0 & 14.9 \\
\(\quad\) Person effect and observables & 48.2 & 46.5 & 44.3 & 18.8 \\
\(\quad\) Residual & 16.4 & 15.2 & 13.7 & -3.9 \\
Between-firm, within-industry & 14.0 & 14.7 & 15.3 & 23.1 \\
\(\quad\) Firm sorting & 10.9 & 11.3 & 12.2 & 20.3 \\
\(\quad\) Firm pay premium & 3.1 & 3.4 & 3.1 & 2.9 \\
Between-industry & 21.4 & 23.6 & 26.8 & 61.9 \\
Industry sorting & 17.2 & 18.8 & 22.0 & 53.2 \\
Industry pay premium & 4.2 & 4.8 & 4.8 & 8.7 \\
\hline
\end{tabular}

Notes: Authors' tabulations of LEHD microdata. Tabulations include workers with annual real earnings \(>\$ 3,770\) in EINs with 20 or more employees. See equation (5) and discussion in text for definitions.
discussion on this combined contribution in the remainder of the paper. \({ }^{26}\) In addition, while often expositionally convenient to combine these terms, it is useful to emphasize the especially large contribution of the covariance component of what we denote as sorting to rising inequality. This covariance contribution highlights that rising dispersion in between industry premia contributes directly and indirectly via this rising covariance. We draw this point out in various places in the main text.

Table 7 exploits our industry-enhanced AKM decomposition to understand rising earnings inequality. \({ }^{27}\) The first three columns of Table 7 show results for our three intervals while the last column computes the terms underlying the change in inequality from our first to last intervals (1996-2002 to 2012-2018). 28

Between-industry dispersion accounts for 61.9 percent of the growth in inequality over the time period covered by our dataset. 53.2 percent of the total rise in

\footnotetext{
\({ }^{26}\) In the online Appendix (Appendix Table A4), we provide the separate contribution of these components. In using sorting for the combined contribution of the covariance and segregation components, we deviate from the labels used by Song et al. (2019) who denoted the covariance component sorting. We have a further discussion of the alternative decompositions and terminology in the literature in the online Appendix.
\({ }^{27}\) Online Appendix Table A1 presents estimates that use the Song et. al (2019) approach without reference to industry, and online Appendix Table A2 aggregates these estimates into firm-level segregation, pay premium, and covariance contributions. Online Appendix Table A3 presents estimates of equation (5) before aggregating them as done in Table 7 .
\({ }^{28}\) While their focus is on inequality in the cross section rather than its change over time, Card, Rothstein, and Yi (2024) demonstrate that there is substantial variation in industry premia not explained by worker sorting, which is consistent with our findings in Table 7.
}
inequality can be attributed to rising between-industry sorting. Rising dispersion in industry-level pay premia accounts directly for a smaller but still substantial 8.7 percent of the total rise in inequality in addition to its indirect contribution via sorting.

What firm-level inequality is left after we account for industry-level differences? Table 7 shows that, in the cross-section, less than one-sixth ( 14.0 percent to 15.3 percent) of earnings dispersion occurs between firms in the same industry. Looking at growth, we find that 23.1 percent of variance growth is between firms, within industries. Of this, sorting accounts for 20.2 percent. Rising within-industry, between-firm pay premia play a smaller role in rising inequality accounting for 2.9 percent of the increase in inequality. \({ }^{29}\)

Table 7 also describes within-firm inequality. In the cross section, most of the variation in earnings is within-firm rather than between-firm-but notably the share is declining from 64.6 percent in the first interval to 58.0 percent in the last. Although its share of overall earnings dispersion falls over time, rising within-firm inequality accounts for a modest amount ( 14.9 percent) of the growth in inequality. This mostly reflects an increase in the dispersion of worker effects (18.8 percent), as the residual has a relatively small role offsetting inequality growth ( -3.9 percent). \({ }^{30}\)

We next turn to two alternative methods of analyzing the critical role of industry in rising inequality in recent decades. These methods allow us to assess the robustness of our AKM approach and also provide detail on the mechanisms by which worker characteristics may lead to earnings dispersion.

\section*{B. Sorting and Industry Premia: Alternative Approaches}

We use the CPS-LEHD and OEWS to decompose rising inequality into related but distinct sorting and industry premia contributions. Details are in the online Appendix but we summarize the results here. \({ }^{31}\) Starting with the CPS-LEHD data, we estimate the contributions of age by education, occupation, and industry using the specification from Hoffman, Lee, and Lemieux (2020). This allows us to decompose the variance of earnings into the contribution of within-industry dispersion arising from observable (age by education and occupation) differences in earnings for workers in the same industry, between-industry pay premia, and between-industry sorting. The latter reflects the contribution of the covariance between average industry worker observable effects and industry premia along with the segregation of observable worker effects between industries.

Using the OEWS, we do not have person-level information but we can estimate a related decomposition using the industry by occupation by time interval data. First, we estimate a model in which each occupation has an additive effect on earnings, as does each industry. We then decompose the variance of earnings into within-industry occupation effects (and a within-industry residual), between-industry premia, and

\footnotetext{
\({ }^{29}\) Details about industry contributions to the between-firm, within-industry contributions are provided in online Appendix C.
\({ }^{30}\) These estimates are quite close to analogous results reported by Song et al. (2019) for a similar time period. We elaborate on this in online Appendix E.
\({ }^{31}\) The CPS-LEHD decomposition is described in online Appendix D and that of the OEWS is in online Appendix G.
}

Table 8-Variance Decomposition: AKM versus Human Capital versus Occupation
\begin{tabular}{lcccc}
\hline \hline Data source: \\
Specification: & \begin{tabular}{c} 
LEHD \\
AKM \\
\((1)\)
\end{tabular} & \begin{tabular}{c} 
CPS-LEHD \\
AKM \\
\((2)\)
\end{tabular} & \begin{tabular}{c} 
CPS-LEHD \\
Human capital \\
\((3)\)
\end{tabular} & \begin{tabular}{c} 
OEWS \\
Occupation \\
\((4)\)
\end{tabular} \\
\hline Variance, as percent of total \((\%)\) & & & & \\
Between-industry growth & 61.9 & 66.2 & 66.2 & 87.8 \\
\(\quad\) Sorting & 53.2 & 56.5 & 44.8 & 85.0 \\
\(\quad\) Pay premia & 8.7 & 9.6 & 21.4 & 2.8 \\
\hline
\end{tabular}

Notes: The first column is taken from Table 7. The second and third columns are from the linked CPS-LEHD dataset, including workers with annual real earnings \(>\$ 3,770\), see discussion in text and online Appendix equation (D2) for definitions. The fourth column is from the OEWS dataset, see discussion in text and equation (G2) for definitions. In the first 3 columns total variance growth is measured at the person level. In the OEWS, total variance growth is measured at the industry-occupation level.
between-industry sorting (reflecting the covariance between industry premia and average industry occupation effects and the segregation of occupation effects across industries). \({ }^{32}\)

Table 8 compares the decomposition of the contribution of industry using the full LEHD with AKM as in Table 7 with the linked CPS-LEHD analysis using the human capital equation approach. In taking advantage of the data infrastructure from both approaches, we can also integrate the AKM estimates from the full LEHD analysis into the linked CPS-LEHD. Also, included in the table is the decomposition using the OEWS data.

In comparing the columns 1 and 2 in Table 8 we find that the overall contribution of industry is very similar in the full LEHD and the linked CPS-LEHD. The same is true of the respective contributions of industry premia and sorting using AKM estimates. This finding is reassuring as the CPS-LEHD linked data replicates the core patterns from the LEHD. It also highlights that the firm size restriction, imposed in column 1 but not in column 2, has a limited impact.

In comparing columns 2 and 3 in Table 8, we find that, whether using AKM or human capital equation based estimates, the contributions of industry premia and sorting are broadly similar. By construction, the overall contribution of industry is the same in columns 2 and 3 . Industry premia appear to be especially important when we use the human capital approach. In contrast, when we use AKM, between-industry sorting becomes more important. The AKM approach attributes more of inter-industry earnings differentials to time-invariant worker characteristics, some of which are not captured by our human capital framework. \({ }^{33}\)

The last column in Table 8 presents the results using the OEWS. The overall magnitudes of the change in inequality are not comparable to those from the first three columns given that this reflects the changing between-industry variance from industry by occupation data rather than from person-level data. However, it is still striking that the overwhelming fraction (87.8 percent) of the increase in dispersion

\footnotetext{
\({ }^{32}\) See online Appendix G for details and formulas.
\({ }^{33}\) The covariance components from AKM versus the human capital approach are almost identical. It is the segregation component that is smaller with the human capital approach reflecting the smaller role of observable human capital variables compared to worker effects in AKM.
}

Table 9—Sources of Between-Industry Variance Growth: Top 30 Industries
\begin{tabular}{|c|c|c|c|}
\hline \multirow[b]{3}{*}{Industry group} & \multirow[b]{3}{*}{Total contribution to between-industry variance growth} & \multicolumn{2}{|l|}{Share of contribution explained by between-industry (\%)} \\
\hline & & Sorting & Premium \\
\hline & & (2) & (3) \\
\hline \multicolumn{4}{|l|}{Panel A. LEHD AKM decomposition} \\
\hline Top 19 high-paying & 0.041 & 86.2 & 13.8 \\
\hline Top 11 low-paying & 0.033 & 84.2 & 15.8 \\
\hline \multicolumn{4}{|l|}{Panel B. CPS-LEHD human capital decomposition} \\
\hline Top 19 high-paying & 0.039 & 63.5 & 36.5 \\
\hline Top 11 low-paying & 0.030 & 66.1 & 33.9 \\
\hline \multicolumn{4}{|l|}{Panel C. OEWS occupation-industry decomposition} \\
\hline Top 19 high-paying & 0.016 & 77.3 & 22.7 \\
\hline Top 11 low-paying & 0.008 & 81.8 & 18.2 \\
\hline
\end{tabular}

Notes: "LEHD AKM decomposition:" Authors' tabulations of LEHD microdata. Tabulations include workers with annual real earnings \(>\$ 3,770\) in EINs with 20 or more employees. See discussion in text and equation (5) for definitions. "CPS-LEHD human capital decomposition:" Authors' tabulations of linked CPS-LEHD dataset including workers with annual real earnings \(>\$ 3,770\), see discussion in text and equation (D2) for definitions. "OEWS occupation-industry decomposition:" Authors' tabulation of OEWS published aggregates, 281 four-digit industries and 22 occupations. See discussion in text and equation (G2) for definitions.
in earnings in the published OEWS data derives from increasing between-industry dispersion. Furthermore, the between-industry contribution is dominated by sorting (high wage workers based on occupation increasingly sorting into high wage industries and working together and low wage workers increasingly sorting into low wage industries and working together).

\section*{C. Sorting and Industry Premia in the Top 30 Industries}

We now focus our attention on the top 30 industries that contribute to rising between-industry inequality taking advantage of the decompositions from the LEHD, CPS-LEHD, and OEWS data. Table 9 presents the between-industry sorting and firm pay premia contributions from these sources for the top 30 industries. Both of these components contribute substantially to rising between-industry dispersion. However, it is apparent that sorting dominates. However, industry premia contribute directly and also via the covariance component of sorting. \({ }^{34}\) In interpreting Table 9, it is useful to recall the dominance of the top 30 industries in accounting for rising between industry dispersion from the alternative data sources and decompositions. The share of the between-industry increase in dispersion from 1996-2002 to 2012-2018 accounted for by the top 30 industries is 98.1 percent in the LEHD data, 105 percent in the CPS-LEHD data, and 96.2 percent in the OEWS data.

\footnotetext{
\({ }^{34}\) The covariance component is large in all approaches and datasets as shown in the online Appendix, e.g., online Appendix Tables A4, D1, and G3.
}

\section*{VI. Industry Premia and Worker Effects in the Top 30 Industries}

In Table 3, we show that over our sample period all of the top 19 high-paying industries exhibit earnings per worker increases relative to the mean and ten of the top 11 low-paying industries exhibit earnings per worker decreases relative to the mean. \({ }^{35}\) In this section, we show that this translates into rising industry premia and worker effects for the top 19 high-paying industries and falling industry premia and worker effects for the top 11 low-paying industries.

At the industry-level, the average change in earnings per worker is equal to the sum of the change in the average worker effect plus the sum of the change in the industry premia. For the AKM decomposition, we have
![](https://cdn.mathpix.com/cropped/2024_10_28_afef6c19081ade40df36g-26.jpg?height=159&width=922&top_left_y=758&top_left_x=411)

We use analogous expressions for the CPS-LEHD human capital decomposition and OEWS decomposition. \({ }^{36}\) We start with the LEHD AKM based decomposition of the change in average industry earnings, shown in Figure 5. For each of the 30 industries, we report the change in average worker effects and industry premia, where industries are ranked by the changes in the industry's average pay.

In 23 of the top 30 industries, the change in the average worker and industry premia have the same sign. The correlation in the top 30 industries between the changing industry premia and changing average worker effects is 0.53 . Notable contributors to this correlation are high-paying industries like Other Information Services with a very large increase its average worker effect of \(42.3 \log\) points ( 52.7 percent) and its industry premia of \(27.8 \log\) points ( 32.1 percent) and low-paying industries like Grocery Stores with a large decline in its average worker effect of 11.5 log points ( 12.2 percent) and its industry premia of \(7.9 \log\) points ( 8.2 percent).

In 8 of our 19 high-paying industries (Figure 5, panel B), the average worker effect contributes more than \(10.0 \log\) points ( 10.5 percent) to that industry's change in average earnings, while the average firm effect contributes less than 5.0 \(\log\) points ( 5.1 percent). These include the two manufacturing industries, as well as Professional and Commercial Equipment and Supplies Merchant Wholesalers (4234), and Management of Companies and Enterprises (5511).

We show that the same patterns hold on average for the top 19 high-paying industries and top 11 low-paying industries using the CPS-LEHD human capital and OEWS occupation decompositions in Figure 6. First, panel A of the figure shows the average changes from the AKM based decomposition consistent with Figure 6. Panel B shows the results using the CPS-LEHD data with the earnings decomposition

\footnotetext{
\({ }^{35}\) The one exception in the top 11 industries is the very low-paying Employment Services (5613) industry that on average has earnings that are more than 50 percent lower than average and has a modest increase of about 2 percent.
\({ }^{36}\) The exact formulas for these CPS-LEHD and OEWS decompositions are online Appendices D and G, respectively. The CPS-LEHD decomposition is given as online Appendix equation (D3) and that of the OEWS decomposition is given as online Appendix equation (G3).
}

Panel A. Top 11 low-paying industries
![](https://cdn.mathpix.com/cropped/2024_10_28_afef6c19081ade40df36g-27.jpg?height=687&width=1073&top_left_y=267&top_left_x=367)

Panel B. Top 19 high-paying industries
![](https://cdn.mathpix.com/cropped/2024_10_28_afef6c19081ade40df36g-27.jpg?height=679&width=1121&top_left_y=1013&top_left_x=321)

Figure 5. Change in Industry-Level Average Worker and Firm Effects
Notes: Authors' tabulations of LEHD microdata. Tabulations include workers with annual real earnings \(>\$ 3,770\) in EINs with 20 or more employees. See equation (6) for definitions. Average earnings are normalized by the overall average in each time interval. "Worker" combines the contribution of the worker effect with that of observable characteristics.
based upon observable human capital variables. The top 19 high-paying industries have positive changes on average in industry premia and human capital premia. The top 11 low-paying industries have negative changes on average in each of these components. Panel C depicts the patterns for the OEWS decomposition. Recall the magnitudes are not directly comparable to the LEHD AKM or CPS-LEHD human capital decompositions since the starting point is industry by occupation data. However, for the top 19 high-paying industries we find that on average increases
![](https://cdn.mathpix.com/cropped/2024_10_28_afef6c19081ade40df36g-28.jpg?height=1034&width=1298&top_left_y=220&top_left_x=230)

Figure 6. Worker and Employer Components of the Change in Industry Average Earnings
Notes: Panel A: Authors' tabulations of LEHD microdata, including workers with annual real earnings \(>\$ 3,770\) in EINs with 20 or more employees. "Worker" combines the contribution of the worker effect with that of observable characteristics. See equation (6) for definitions. Panel B: Authors' tabulations of linked CPS-LEHD microdata, including workers with annual real LEHD earnings \(>\$ 3,770\) who match to the CPS. See online Appendix equation (D3) for definitions. Panel C: Authors' tabulations of published OEWS aggregates. See online Appendix equation (G3) for definitions.
in industry premia and occupation effects. For the top 11 low-paying industries, we find on average decreases in occupation effects with modest offsetting industry premia effects. Notably the modest positive industry premia change effects on average in the top 11 low-paying industries are much smaller than the positive industry premia changes on average in the top 19 high-paying industries.

\section*{VII. Taking Stock}

Our findings provide a distinct perspective on the role of polarization in increasing inequality (e.g., Autor, Katz, and Kearney 2006, 2008; Goos and Manning 2007; and Acemoglu and Autor 2011). This literature highlights the increasing bifurcation of the occupational skill and pay structure. Our findings show that this manifests itself via low- and high-paying industries becoming more sorted on worker skill, becoming more skill-segregated across industries, and becoming relatively higher and relatively lower paid due to rising industry premia in high-paying sectors and falling industry premia in low-paying sectors.

Panel A. Employment change by initial premium
![](https://cdn.mathpix.com/cropped/2024_10_28_afef6c19081ade40df36g-29.jpg?height=477&width=660&top_left_y=264&top_left_x=224)

Panel C. Employment change by premium change
![](https://cdn.mathpix.com/cropped/2024_10_28_afef6c19081ade40df36g-29.jpg?height=480&width=648&top_left_y=817&top_left_x=228)

Panel B. Premium change by initial premium
![](https://cdn.mathpix.com/cropped/2024_10_28_afef6c19081ade40df36g-29.jpg?height=475&width=631&top_left_y=268&top_left_x=888)

Panel D. Worker effect change by initial premium
![](https://cdn.mathpix.com/cropped/2024_10_28_afef6c19081ade40df36g-29.jpg?height=477&width=628&top_left_y=816&top_left_x=892)

Figure 7. Changes in Employment and Earnings (Top 30 Industries)
Notes: Authors' tabulations of LEHD microdata. Tabulations include workers with annual real earnings \(>\$ 3,770\) in EINs with 20 or more employees. Graphs plot local polynomial estimates of the variable on the vertical axis on the horizontal axis. Shaded regions indicate 95 percent confidence intervals. Changes compare 2012-2018 with 1996-2002.

This industry polarization is concentrated in relatively small set of industries in the tails of the industry earnings distribution. Figure 7 provides a visual summary of pattern of polarization in the top 30 industries that dominate rising earnings inequality. Both the low industry premia industries and high industry premia industries in the top 30 exhibit (upper-left panel) increases in employment shares with especially large employment share increases in the low premia industries. High initial premia industries exhibit increases in industry premia while the low premia industries exhibit declines in industry premia (upper-right panel)—with high premia industries having especially large increases. Industries with declines in the industry premia (the low premia industries) and industries with increases in the industry premia (the high premia industries) have had increases in employment shares (lower-left panel). Increases in worker effects are large in the high initial premia industries and declines in worker effects are present in the low premia industries.

In short, the upper-tail component is due to rising industry premia without much change in employment; the lower tail is driven by falling industry premia and rising
employment. These findings suggest that different forces are at play. At the top, the patterns are consistent with a positive demand shift, yielding higher wages and slightly higher quantities. At the bottom, the patterns are consistent with a positive supply shift, yielding lower wages and higher quantities.

The results are thus related but distinct from the existing polarization literature, which argues that the growth of employment in low-wage sectors is in part driven by the decline of middle-skill blue-collar production and white-collar office-administrative occupations. Relatedly, Acemoglu and Restrepo (2022) argue that skill groups that have been displaced by automation of routine tasks have seen falling earnings as these groups are reallocated to activities where they have weaker comparative advantage. These findings and perspective have a natural mapping to our findings on the growth of low-paid service industries, which have substantially expanded even as their already-low pay premia have fallen further.

In emphasizing the role of industry, it is important to emphasize the prominent role for industry premia in our findings. Conceptually, industry premia (and more generally firm premia) are consistent with the presence of frictions or market imperfections (e.g., monopsony power) in labor markets. Industry and firm premia may also arise from other factors than market imperfections such as environments where the marginal product of labor depends on the unique product attributes of the firm or industry, see Tervio (2008). While we acknowledge there are issues for mapping such mechanisms to the estimated firm and industry effects, the robustness of our findings across different source data and methodologies suggests that it is important to understand the sources of the direct and indirect effects of the industry premia in accounting for rising earnings inequality. That is, whether we use AKM to control for worker effects, or a human capital approach to control for age, education, and occupation effects, or rich occupation data to control for occupation effects, we find a role for rising between-industry premia and especially a role for a rising covariance between those premia and alternative estimates of worker premia (AKM worker effects, education and occupation premia).

Viewed from this perspective, our findings rule out explanations of rising inequality that don't have a prominent role for industry premia through both direct and sorting effects. While we have noted our results are complementary to Acemoglu and Restrepo (2022), the structural model in their paper has competitive labor markets without frictions. Hence, their model implies that the prominent role of industry they find is associated with rising segregation of workers between industries without any role for industry premia and sorting associated with such premia. In other words, their model focuses on the changing composition of workers across industries without any interaction of that changing composition with industry premia. Our empirical results suggest that enhancing the structural model to permit such interactions is critical. One way to do this would be to consider the role of frictions and imperfect competition in labor markets for capturing this important aspect of variation in the data. This argument is not restricted to this specific mechanism (which is a form of skill biased technical change). Our results imply that any mechanism (whether driven by frictions, distortions, or some other mechanism) must account for the combined contribution of increased between-industry sorting and rising between-industry premia.

\section*{VIII. Conclusion}

Rising earnings inequality is dominated by rising between-industry inequality. High earnings workers are more likely to work in high earnings industries and with each other, and low earnings workers are more likely to work with in low earnings industries and with each other. This polarization of the industry earnings and employment structure is concentrated in a relatively small number of industries. About 10 percent of the 301 detailed four-digit NAICS industries account for almost all of rising between-industry dispersion, while accounting for less than 40 percent of employment. These thirty industries are drawn from the top and bottom of the earnings distribution in terms of industry-level averages. The top 19 high-paying industries exhibit increases in industry pay induced by both increases in industry premia and average worker effects. The top 11 low-paying industries exhibit decreases in industry pay induced by both decreases in industry premia and average worker effects. These inferences are robust to identifying industry premia and worker effects using the AKM approach, a standard human capital approach using worker observables including age, education, and occupation, and using high quality industry by occupation data.

The top 10 percent of industries that account for virtually all of rising between-industry inequality are not randomly spread across the distribution of industries but concentrated in specific industry clusters in the tails of the earnings distribution. At the high end, dominant industries are drawn from high-tech and STEM intensive industries, finance, mining, and selected industries in health. At the low end, dominant industries are drawn from selected industries in retail and health. The top 30 industries are in industry clusters that have exhibited structural transformations that have been the subject of independent study. Notably absent are the vast majority of industries in manufacturing. Part of the polarization story is the decline in production workers in manufacturing and the rise of employment and decline in earnings in key low pay service sectors.

The dominance of industry effects is closely linked to the rising importance of mega \((10,000+\) ) firms in the US economy. The increasing share of employment accounted for by mega firms is concentrated in the thirty four-digit industries that account for virtually all of rising between-industry dispersion. This rising share of employment at mega firms is accompanied by a declining size-earnings premium in the 11 low-paying industries. For mega firms in the 19 high-paying industries in the top 30, earnings premia rise sharply relative to other industries (albeit not as rapidly as other large but not mega firms in these industries).

We find there is a close connection between changes in the occupation distribution and our top 30 industries. Most of the increase in employment in top-paying occupations is accounted for by our 19 top-paying industries, while all of the increase in employment in low-paying occupations is in our 11 low-paying industries. For the top-paying industries, there is a substantial increase in earnings differentials for the top-paying occupations. Likewise there is a substantial decline in earnings differentials for the low-paying occupations in our low-paying industries.

In the top 19 high-paying industries there is a sharp increase in the share of workers with a bachelor's or advanced degree with accompanying declines in workers without such degrees. In the top 11 low-paying industries there is no such shift in
the distribution. On average, this distribution is dominated by workers with a high school diploma or less.

Our findings imply that understanding rising earnings inequality during the last several decades requires understanding the evolution of how firms organize themselves in a relatively small set of industries. Moreover, since it is the between-industry contribution that dominates, the common effects of reorganization across firms in the same industry matter. Many mechanisms such as changing technology, market structure, and globalization likely underlie rising earnings inequality. The focus of future research on the impact of such changes on rising earnings inequality should be on the uneven and concentrated impact of such mechanisms across industries. In addition, since between-industry sorting and industry premia play important roles, the mechanisms must interact with the factors that yield such effects including possibly frictions and imperfect competition in labor markets.

\section*{REFERENCES}
-Abowd, John, Francis Kramarz, and David Margolis. 1999. "High Wage Workers and High Wage Firms." Econometrica 67 (2): 251-333.
-Abowd, John M., Kevin L. McKinney, and Nellie L. Zhao. 2018. "Earnings Inequality and Mobility Trends in the United States: Nationally Representative Estimates from Longitudinally Linked Employer-Employee Data." Journal of Labor Economics 36 (S1): S183-300.
Abowd, John, Bryce Stephens, Lars Vilhuber, Fredrik Andersson, Kevin McKinney, Marc Roemer, and Simon Woodcock. 2009. "The LEHD Infrastructure Files and the Creation of the Quarterly Workforce Indicators." In Producer Dynamics: New Evidence from Micro Data, edited by Timothy Dunne, J. Bradford Jensen, and Mark J. Roberts, 149-230. Chicago: University of Chicago Press.
Acemoglu, Daron, and David H. Autor. 2011. "Skills, Tasks and Technologies: Implications for Employment and Earnings." In Handbook of Labor Economics, Vol. 4, edited by Orley Ashenfelter and David Card, 1043-1171. Amsterdam: Elsevier-North Holland.
-Acemoglu, Daron, and Pascual Restrepo. 2022. "Tasks, Automation, and the Rise in US Wage Inequality." Econometrica \(90(5): 1973-2016\).
-Autor David, David Dorn, Lawrence F. Katz, Christina C. Patterson, and John Van Reenen. 2020. "The Fall of the Labor Share and the Rise of Superstar Firms." Quarterly Journal of Economics \(135(2): 645-709\).
-Autor, David H., Lawrence F. Katz, and Melissa S. Kearney. 2006. "The Polarization of the U.S. Labor Market." American Economic Review 96 (2): 189-94.
-Autor, David H., Lawrence F. Katz, and Melissa S. Kearney. 2008. "Trends in U.S. Wage Inequality: Revising the Revisionists." Review of Economics and Statistics 90 (2): 300-23.
- Barth, Erling, Alex Bryson, James C. Davis, and Richard Freeman. 2016. "It's Where You Work: Increases in the Dispersion of Earnings across Establishments and Individuals in the United States." Journal of Labor Economics 34 (S2): S67-97.
-Bloom, Nicholas, Fatih Guvenen, Benjamin S. Smith, Jae Song, and Till von Wachter. 2018. "The Disappearing Large-Firm Premium." AEA Papers and Proceedings 108: 317-22.
Bureau of Economic Analysis. 2021. "Personal Consumption Expenditures Price Index." Bureau of Economic Analysis. https://www.bea.gov/data/personal-consumption-expenditures-price-index (accessed November 22, 2021).
Bureau of Labor Statistics. 2022. "Occupation Employment and Wage Survey Data." Bureau of Labor Statistics. https://www.bls.gov/oes/tables.htm (accessed July 26, 2022).
Bureau of Labor Statistics. 2022. "Quarterly Census of Employment and Wages." https://https://www. bls.gov/cew/ (accessed July 26, 2022).
Bonhomme, Stéphane, Kerstin Holzheu, Thibaut Lamadon, Elena Manresa, Magne Mogstad and Bradley Setzler. 2022. "How Much Should We Trust Estimates of Firm Effects and Worker Sorting?" Journal of Labor Economics 41 (2): 291-322.
Briskar, Juraj, José Rodríguez Mora, Edoardo di Porto, and Cristina Tealdi. 2022. "Decomposition of Italian Inequality." Unpublished.
- Card, David, Ana Cardoso, and Patrick Kline. 2016. "Bargaining, Sorting, and the Gender Wage Gap: Quantifying the Effect of Firms on the Relative Pay of Women." Quarterly Journal of Economics 131 (2): 633-86.
-Card, David, Jorg Heining, and Patrick Kline. 2013. "Workplace Heterogeneity and the Rise of West German Wage Inequality." Quarterly Journal of Economics 128 (3): 967-1015.
-Card, David, Jesse Rothstein, and Moises Yi. 2024. "Industry Wage Differentials: A Firm- Based Approach." Journal of Labor Economics 42 (S1): S11-S59.
Davis, Steve J., and John Haltiwanger. 1991. "Wage Dispersion Between and Within U.S. Manufacturing Plants, 1963-1986." Brookings Papers: Microeconomics 21: 115-20.
Dey, Matthew, Susan Houseman, and Anne Polivka. 2010. "What Do We Know About Contracting Out in the United States? Evidence from Household and Establishment Surveys." In Labor in the New Economy, edited by Katharine Abraham, James Spletzer, and Michael Harper, 267-304. Chicago: University of Chicago Press.
Dorn, David, Johannes Schmeider, James Spletzer, and Lee Tucker. 2023. "Domestic Outsourcing in the United States." Unpublished.
-Dunne, Timothy, Lucia Foster, John Haltiwanger, and Ken Troske. 2004. "Wage and Productivity Dispersion in United States Manufacturing: The Role of Computer Investment." Journal of Labor Economics 22 (2): 397-429.
Fernald, John. 2014. "Productivity and Potential Output before, during, and after the Great Recession." NBER Macroeconomics Annual 29 (1): 1-51.
-Foster, Lucia, John Haltiwanger, and C. J. Krizan. 2006. "Market Selection, Reallocation, and Restructuring in the U.S. Retail Trade Sector in the 1990s." Review of Economics and Statistics 88 (4): 748-58.
-Goos, Maarten, and Alan Manning. 2007. "Lousy and Lovely Jobs: The Rising Polarization of Work in Britain." Review of Economics and Statistics 89 (1): 118-33.
—Guimarães, Paulo, and Pedro Portugal. 2010. "A Simple Feasible Procedure to Fit Models with High-Dimensional Fixed Effects." Stata Journal 10 (4): 628-49.
-Haltiwanger, John, Henry Hyatt, and James Spletzer. 2023. "Increasing Earnings Inequality: Reconciling Evidence from Survey and Administrative Data." Journal of Labor Economics 41 (S1): S61-93.
Haltiwanger, John, Henry Hyatt, and James Spletzer. 2024. Data and Code for: "Rising Top, Falling Bottom: Industries and Rising Wage Inequality." Nashville, TN: American Economic Association; distributed by Inter-university Consortium for Political and Social Research, Ann Arbor, MI. https://doi.org/10.3886/E198751V1.
Haltiwanger, John, and James Spletzer. 2020. "Between-firm changes in earnings inequality: the dominant role of industry effects." NBER Working Paper 26786.
Haltiwanger, John, and James Spletzer. 2022. "Rising Between-Firm Inequality and Declining Labor Market Fluidity: Evidence of a Changing Job Ladder." In Measuring Distribution and Mobility of Income and Wealth, edited by Raj Chetty, John Friedman, Janet Gornick, Barry Johnson, and Arthur Kennickell, 45-67. Chicago: University of Chicago Press.
Hecker, Daniel. 2005. "High-Technology Employment: A NAICS-Based Update." Monthly Labor Review 128 (7): 57-72.
-Hoffman, Florian, David S. Lee, and Thomas Lemieux. 2020. "Growing Income Inequality in the United States and Other Advanced Economies." Journal of Economic Perspectives 34 (4): 52-78.
Hoffman, Florian, David S. Lee, and Thomas Lemieux. 2020. Replication data and code for: "Growing Income Inequality in the United States and Other Advanced Economies" Nashville, TN: American Economic Association; distributed by Inter-university Consortium for Political and Social Research, Ann Arbor, MI. https://doi.org/10.3886/E122201V1.
-Krueger, Alan, and Lawrence Summers. 1988. "Efficiency Wages and the Inter-Industry Wage Structure." Econometrica 56 (2): 259-93.
Oliner, Stephen, Daniel Sichel and Kevin Stiroh. 2007. "Explaining a Productive Decade." Brookings Papers on Economic Activity 37 (2).
Ruggles, Steven, Sarah Flood, Matthew Sobek, Daniel Backman, Annie Chen, Grace Cooper, Stephanie Richards, Renae Rodgers, and Megan Schouweiler. "IPUMS USA: Version 15.0 [dataset]." IPUMS. Minneapolis, MN: IPUMS, 2024. https://doi.org/10.18128/D010.V15.0.
-Song, Jae, David J. Price, Fatih Guvenen, Nicholas Bloom, and Till von Wachter. 2019. "Firming Up Inequality." Quarterly Journal of Economics 134 (1): 1-50.
- Tervio, Marko. 2008. "The Difference That CEOs Make: An Assignment Model Approach." American Economic Review 98 (3): 642-68.
US Census Bureau. 2021. "Current Population Survey (CPS) - Annual Social and Economic Supplements (ASEC) microdata." US Census Bureau. https://www.census.gov/data/datasets/time-series/ demo/cps/cps-asec.html (accessed November 15, 2021).
US Census Bureau. 2023. "Longitudinal Business Database (LBD) microdata." US Census Bureau. https://www.census.gov/programs-surveys/ces/data/restricted-use-data/longitudinal-businessdatabase.html.

US Census Bureau. 2023. "Longitudinal-Employer Household Dynamics (LEHD) micro- data." US Census Bureau. https://lehd.ces.census.gov/data/lehd-snapshot-doc/latest/.
Walker, Ed. 1997. "The Census Bureau's Business Register: Basic Features and Quality Issues." Paper presented at the Joint Statistical Meetings of the American Statistical Association, Anaheim, CA.