Earnings Inequality and Mobility Trends in the United States: Nationally Representative Estimates from Longitudinally Linked Employer-Employee Data

A. Related Literature

This paper makes two contributions to the literature on earnings inequality. First, we document the trends in earnings inequality using administrative data corrected using a consistent frame. Second, we analyze the sources of earnings inequality, with a particular emphasis on understanding the role of firms.

B. Date Regimes

States and the federal government joined the partnership that supplies input data to the LEHD program at different dates. When a state or the federal government joined, the data custodians were asked to produce historical data for as many quarters in the past, back to 1990Q1, as could be reasonably recovered from their information storage systems. As a result, the date that a data-supplying entity joined the partnership is not the same as the first quarter in which that entity’s data appear in the system. The start date for any state or the federal government depends primarily on the amount of historical data the state or federal government could recover at the time it joined. This potential ignorability (in the sense of Rubin 1987 or Imbens and Rubin 2015) of the start data for a segment of the LEHD data is the basis for our methods of constructing nationally representative estimates back to the early 1990s.¹⁸

To understand how state entrance into the LEHD program affects the trends in earnings inequality, we analyze four regimes that correspond to different epochs of data availability:

We did not create a separate date regime when the OPM data begin in 2000Q1 for two reasons. First, the proportion of reported federal jobs in 2000Q1 is small, no more than 4% of 2012Q1 UI-covered employment in a state and, on average, less than 1% except for the District of Columbia, where the proportion of OPM jobs is about 19%. Second, although the proportion of OPM jobs in the District of Columbia is high, jobs there are part of regime 4, which does not begin until 2004Q1.

C. Job and Person Sampling Frames

By 2004Q1, the LEHD data represent the complete universe of statutory jobs in the United States: all 50 states, the District of Columbia, and the federal government are reporting regularly. Before this date, LEHD data provide a complete frame for the states in each date regime (excluding the federal government before 2000Q1). After this date, the LEHD data provide a complete frame for the national population of UI-covered jobs, including federal employees. Studying job-level inequality, the task for which having a complete job frame is well suited, as a proxy for person-level inequality may be misleading because of the time-varying many-to-one assignment of jobs to workers. The number of employers per worker varies over the business cycle. Lower-earning workers tend to have more employers, complicating the interpretation of job-level results.¹⁹ Therefore, when studying inequality, it is preferable to have a person frame that covers a known population of interest, such as all persons legally eligible to work in the United States. For our analyses, we use two different approaches to creating frames of jobs and workers. The first approach relies only on the employer-employee links present in the UI data. This method captures all reported jobs. The second approach uses the US Census Bureau’s edited version of the SSA’s master SSN database (the Numident), capturing all reported employment-eligible workers but removing jobs associated with ineligible workers, as we elaborate below.

LEHD earnings records are reported quarterly by the employing firm. These records contain a nine-digit person identifier, typically assumed to be a SSN. However, at the time the report is received by the state UI office the nine-digit person identifiers are not verified, resulting in records both with and without a valid SSN. Using the Numident, we ascertain whether each earnings record is associated with a valid SSN. Records not associated with a valid SSN may have an alternate, valid person identifier, such as an Internal Revenue Service (IRS)–issued taxpayer identification number; nevertheless, we can distinguish only between valid and invalid SSNs. If the SSN is valid, we have access to demographic characteristics, such as sex and birth date, from the Numident and other US Census Bureau sources. We also have an employment history from the UI wage records. If the SSN is not valid, we only have access to the employment history.

Our first frame, the “all-workers frame,” contains earnings for all jobs reported on the UI data for each date regime in the relevant years from 1990 to 2013, as noted in figure 1. Our second frame, the “eligible-workers frame,” is also delineated by these date regimes but includes jobs only for the subset of the all-workers frame that meets the following criteria:

The purpose of the eligible-workers frame is twofold. The Numident data allow us to consistently identify a set of persons legally eligible to work each year while at the same time implicitly removing earning records from our analysis sample that are not associated with individuals in the covered population. We go a step further. We remove earnings records with valid SSNs where the available data strongly suggest that the SSN is not being used by the original owner.²⁰ These two types of suspect nine-digit person identifiers—invalid SSNs that do not match to the Numident and valid SSNs apparently being used by multiple persons and/or for whom the age of the person issued the SSN is inconsistent with labor market activity—we call “immigrant candidates.”

Table 1 presents basic counts of persons and jobs in the eligible-workers frame and the immigrant-candidate file. While we present some analysis of the immigrant-candidate jobs in appendix A, we do not have sufficient information to convert the collection of these jobs into an intertemporally consistent frame for this population of individuals. We have no plausible means of determining how many immigrant candidates are using each SSN in this collection of UI wage records.

Even accounting for the increasing coverage of each date regime, there is a clear, strong upward trend in the number of immigrant-candidate records until the Great Recession and a strong downward trend thereafter. Because there are data in the system for each of these records, the associated nine-digit person identifier represents at least one worker but may represent many. A great deal of supplemental research would be required to estimate the relation between how many jobs are reported for a “worker” in the immigrant-candidate records and the number of individuals employed. Even more research would be required to estimate their characteristics. We do not attempt such research here. We note that when we use the all-workers frame, each associated nine-digit person identifier counts as one individual, but there is no adjustment for being inactive; that is, we do not assume zero earnings when this “worker” has no reported earnings. When we use the eligible-workers frame, all immigrant-candidate records are excluded. Eligible inactive workers are assumed to have zero earnings when there is no reported activity in any job.

Appendix A contains additional detailed analyses of the construction of the eligible-workers frame. In particular, figure A1 plots the count of earnings records excluded from the eligible-workers frame each year, disaggregated by earnings records each year associated with individuals who have invalid SSNs who are either too young or too old and/or who report working 12 or more jobs in a year. Detailed counts are reported in Table A2.

C. Earnings Definitions and Coverage

In this section, we define our earnings measures for both the all-workers frame and the eligible-workers frame. Our primary measure of earnings is based on annual UI job-level earnings reports. We adjust nominal earnings to real earnings using the consumer price index for all urban consumers (CPI-U), with 2000 as the base year. Let y_ijt be the real earnings for worker i employed at firm j in year t. Person-level annual earnings sum all jobs for each worker in each year:

D. Estimation of the Earnings/Inactivity Distribution

We begin by calculating deciles of eligible-workers person-level earnings, e_it, pooled across the years 2004–13. Using these deciles, we create three earnings bins, as shown in Table 2. The bins are designed to capture the bottom, middle, and top of the earnings distribution over the entire 10-year period. For example, the first two columns in Table 2 show the results for bin 2, workers in the bottom 20% of the earnings distribution.²² Workers in this bin have a minimum annual earnings value of $2, a maximum value of $6,600, and a mean log earnings in real 2000 dollars of $7.473 (implied geometric mean real earnings, $1,760).

The low mean log earnings in bins 2 and 3 suggest that a large proportion of workers in these bins are employed for only part of the year. In Table 3 we present information about the labor force activity of workers in each of the three earnings bins. Each row in the table (except the residual category “All other”) represents a specific combination of quarters worked and number of quarters in the longest job, truncated at a maximum of 6 quarters. The number of quarters in the longest job takes on values from 1 to 6. A 5-quarter longest job is active in either the fourth quarter of the previous year or the first quarter of the subsequent year, while a 6-quarter longest job is active in both. Thus, a 6-quarter active job is active at the beginning of a calendar year, the end of a calendar year, and all quarters in the middle. The most prevalent pattern in each earnings category is listed first, followed by the next most prevalent and continuing until the table contains the patterns for approximately 80% of the workers in a typical year.

The dominant labor force attachment pattern varies substantially across bins: 31% of the workers in the bottom 20% of the earnings distribution work only 1 quarter. In contrast, for workers in the top 20% of the earnings distribution, the most common labor force status is employment with at least one firm for the entire year. Although there are almost certainly large differences in average wages for workers at the bottom, middle, and top of the earnings distribution, one of the primary reasons average annual earnings for workers in the bottom earnings category are so low is that they are employed for only a small portion of the year.

Finally, we combine the earnings bins discussed above with the active/inactive status information available for the eligible-workers frame to create four mutually exclusive earnings/inactivity categories:

A. Comparison of LEHD Data to the CPS and the ACS

To understand the differences between analyses using respondent-provided earnings data in large-scale household surveys and those using administrative data on earnings in the UI system, we constructed detailed analysis samples from the CPS-ASEC (1990–2004) and the ACS (2000–2013). We also linked the LEHD earnings data to the ACS data at the individual record level (2005–13). We used the linked data to study differences between individuals found only in the ACS data and those found in both sources.³⁰ Details on the data construction can be found in section C1 of appendix C. Details on the data linkage can be found in section C3 of appendix C.

In the household survey data, we defined “covered workers” to be individuals who worked in a job that should have appeared in the LEHD data. Table 4 shows that the estimated percentile values of the earnings distribution tend to be greater for covered workers than for all workers in the household survey data.³¹ For percentiles at or above the median, the values from the LEHD eligible-workers frame are close to the ones from the covered workers in household surveys. Below the median, however, the differences are greater, with the percentiles estimated from the household surveys being much greater than the percentiles estimated from the LEHD eligible-workers frame. For example, earnings associated with the 10th percentile in the CPS/ACS data are close to the 20th percentile in the LEHD data.³² We conclude that the differences between survey-reported and administrative earnings data at or above the median are minimal. Below the median, conclusions must be more nuanced. Household surveys seem to capture earnings not captured by the administrative data, but it is difficult to know what to conclude from that finding. One possibility is that respondents report earnings from activity that so closely resembles formal sector work we think we ought to find corresponding earnings reports but do not. Another plausible explanation is that the coverage of the low-earning population is very different in the ACS compared with the UI data. We consider reconciling the administrative and survey data at the bottom of the earnings distribution to be an open research question.³³

Figure 5 shows that the trends in earnings inequality in the household survey data are consistent with our findings from the eligible-worker frame in the LEHD data and are inconsistent with the findings from the all-worker frame. This conclusion also holds for the analysis of trends in the top and bottom separately.³⁴ The pattern of inequality ratios around the Great Recession also is similar to the one found in the eligible-workers frame and is dissimilar to the analysis based on the all-workers frame.³⁵ Additional results and discussion are provided in section C2 of appendix C.

To better understand the discrepancies at the bottom of the earnings distribution, we analyzed linked ACS and LEHD individual records. Eighty-five percent of the individuals we expect to find in both the ACS and the LEHD data are present in both with positive earnings. The other 15% are not found in the LEHD data and, therefore, have zero administrative earnings.³⁶ The result suggests that survey respondents are giving answers that imply that their job should be covered in the administrative data, but there are not corresponding administrative records to match the survey-reported income. The reported income of these individuals appears to be the reason for the discrepancy between the survey and administrative data at the bottom of the earnings distribution.³⁷

Our analysis of the trends in earnings inequality using the various samples of workers in CPS/ACS and the two worker frames in the LEHD data shows that, whether the data were designed or found, an understanding of the contributions of individuals included or excluded from the sample is essential. Conclusions regarding fundamental trends in inequality depend on these decisions.

B. Effects of Inactivity on Earnings Inequality

We used our household survey to confirm that the trends in the estimated employment-to-population ratio produced directly from the micro data match those of the official estimates. In particular, the employment-to-population ratio fell during the Great Recession and had not recovered through 2013.³⁸

We document that around 30 million eligible workers per year had no earnings in the current year but positive earnings in at least one of the past 4 years.³⁹ Their treatment materially affects the distribution of earnings in any given year. It also materially affects the cyclical features of inequality measures, such as the Gini coefficient.⁴⁰ Additional results and discussion are provided in section D2 of appendix D.

Although it is unclear which of the adjusted inequality measures correctly weights the inactive workers, it is worthwhile to consider adjusted measures that count at least some of the zero-earning workers as part of any general analysis of changes in earnings inequality.

A. Evolution of the Earnings/Inactivity Distribution

Beginning in 2004 our eligible-workers frame is complete, including all states, the District of Columbia, and the federal government. The adventitious timing of the start of the complete-data period presents us with the opportunity to study the evolution of the earnings distribution and the dynamics of nonemployment during three distinct epochs of labor market conditions. According to the National Bureau of Economic Research (NBER), the Great Recession began in December 2007 and ended in June 2009. Applying this definition to annual data, we have a prerecession epoch spanning 2004–7, a recession epoch running from 2008 to 2009, and a postrecession epoch beginning in 2010 and ending in 2013.

As we did in Section II.D, we simplify the earnings distribution by assigning each active worker to one of three earnings categories (bottom, middle, and top) and assign inactive but eligible workers to a fourth category. Using the estimated annual earnings/inactivity distributions, we start by comparing the distribution in 2005 with the distribution in 2004, repeating this process for each subsequent year until 2013. Each year, the earnings/inactivity distribution may change relative to the previous year. The extent of this change depends on the number of workers entering and exiting the eligible-workers frame, the number of workers moving between earnings/inactivity categories, and changes in average earnings within each category.⁴¹

From 2004 through 2013 there is relatively strong growth in the number of eligible workers, averaging about 1% per year through 2009 and declining to about half that rate after 2009, although growth within each category is uneven. For example, the largest growth in the number of workers occurs in the eligible to work but inactive category (i.e., no earnings in the indicated year), a category with a growth rate almost twice that of any other group. The growth rate of workers in the bottom and middle earnings categories is less than half the overall growth rate. When we examine total earnings, most of the growth is found in the top 20% of the earnings distribution, with relatively little growth in the bottom 80%. Given that average earnings are falling for the bottom 80% of workers, it should not be surprising that total earnings for this group fails to keep pace with the growth in the number of workers. The situation for workers at the top, however, is much brighter: the top 20% of workers have relatively strong earnings growth of 4.4% over the period, resulting in growth in total earnings (12.2%) that outpaces the growth in the number of workers (7.8%).

The relatively high growth in both the number of eligible workers and the average earnings in the top 20% of the earnings distribution has a strong effect on the distribution of total earnings. Figure 6 shows the share of total earnings attributed to each earnings category by year. We see substantial earnings inequality: earnings for the top 20% of workers are greater than those for the bottom and middle combined, with the relative share of the top increasing almost continuously except for a brief pause in 2008 during the height of the Great Recession. We also see the declining share of income accruing to the middle 60% of workers. Although the number of workers in the middle recovered after the Great Recession, average earnings continued to decline while higher growth in the number of workers at the bottom and top resulted in a declining share of earnings for workers in the middle 60% of the earnings distribution.

B. Earnings/Inactivity Distribution: Decomposition of the Changes

In the previous section, we discussed the changing structure of the earnings/inactivity distribution around the time of the Great Recession. Several key facts stand out. First, there is enormous growth in the number of eligible workers with no reported earnings from 83,200,954 in 2004 to 96,959,047 in 2010, remaining as high as 96,151,327 in 2013. Second, average earnings stagnate (bottom 20%) or decline (middle 60%) for active workers. Third, the growth in the share of earnings accruing to the top 20% results from both growth in the number of workers and average earnings in that category.⁴²

To better understand these changes, we turn our focus to the flows of workers moving between active and inactive status as well as between different earnings categories. When interpreting these results, we implicitly assume average earnings within each category are stable between 2004 and 2013. Although changes in average earnings have a role to play, the data suggest that we focus on worker flows because the impact of a change in the number of workers dominates the impact of a change in average earnings for each category.⁴³

Starting in 2005, each year we calculate the change in the number of workers between the current year and the previous year for the four earnings/inactivity categories. The year-to-year change in the number of workers in a specific category is driven by changes in the number of workers entering (inflows) and the number of workers leaving (outflows). Section E1 of appendix E provides the derivation of the flow accounting that we summarize here.

Figure 7 shows the change in the number of workers for each earnings/inactivity category by year. The data used to calculate the year-to-year differences can be found in the first panel of Table E1. Each line in the graph represents the year-to-year change in the number of workers for one of the four earnings/inactivity categories. For example, in 2009 there were 94,864,949 eligible workers with no reported UI earnings, while in 2008 there were only 88,245,425 workers in the same category, resulting in an increase of 6,619,524 workers. This increase in the number of eligible but inactive workers occurred during the height of the Great Recession and dwarfs the change in any other period, even the relatively large increases in the preceding and subsequent years. The area between the inactive line (blue) and the X-axis represents the cumulative change in the number of eligible workers with no earnings over the entire period of 12,950,373. We can clearly see that the recovery has largely failed to reduce the number of eligible workers with no earnings through 2013. In contrast, the middle 60% (green line) faced a large reduction in numbers during the Great Recession, but, unlike the inactive category, the number of workers in the middle has returned to prerecession levels, as reflected by the area between the middle 60% line and the X-axis.

Table 5 presents an overview of the flow analysis, which we decompose using the transition count matrix. The “Count $t-1$,” “Count t,” and “Net Change” columns are shown in Table E1 and figure 7. They are reproduced here for comparison with sums of the transition counts. The stayers (i.e., c₂₂) are also included, and while they do not directly affect the net change in the flows, they represent the number of workers who remain in a given category for at least 2 years—giving an indication of the earnings stability of the typical worker. The outflows, inflows, and net change columns show the results of using equation (E1). The difference between the inflows and the outflows equals the net change, which should also equal the difference in counts between the current period and the previous period. For example, returning to the eligible workers with no earnings in 2009, we can see that the inflows were 16,166,420 and the outflows were 9,546,896. The large increase in the number of workers in this category was due primarily to a large increase in the inflows relative to the previous year and a small decrease in the number of outflows. That is, a relatively large number of workers who had a job in the previous year were unable to find an employer in the current year, while a relatively small number of workers without a job in the previous year were able to find one in the current year or moved out of the eligible-workers frame.

The table also shows a relatively large increase in the number of eligible workers with no earnings for at least 2 years over the entire period (stayers in the “Eligible but No Reported UI Earnings” panel). The change in the number of stayers equals the difference between the inflows in the previous period (the candidates to become stayers) minus the outflows in the current period. For example, in 2008 there were 13,271,459 eligible workers with at least 1 year without reported earnings (inflows), and 9,546,896 of these workers transitioned to another category in 2009 (outflows), resulting in an increase in the stayers between 2008 and 2009 of 3,724,563 workers. The stayers are useful for understanding the short-term (2-year) volatility differences between each of the categories. For example, the bottom 20% of the UI earnings distribution has relatively few stayers compared with the middle and the top, consistent with the results presented earlier that most of these jobs are of relatively short duration. The results also imply that a large number of workers in the bottom 20% of the earnings distribution stay there for only a year or two before moving to another category, frequently inactivity.

Table 6 presents demographic characteristics for the 24 possible year-to-year transitions, excluding workers not eligible to work in both year $t-1$ and year t. The transitions labeled 0 represent workers moving into or out of the eligible-workers frame. The workers moving into the eligible-worker frame are typically young and predominately nonwhite, while the workers leaving the eligible workforce are typically older (60 years of age or more) and predominately white. One interesting transition group is the 0_4 (workers not eligible in $t-1$ who transition to the top 20% of the earnings distribution in t). These workers are predominately older than other newly eligible workers, male, nonwhite, and overwhelmingly not born in the United States. The remaining transitions have roughly similar characteristics, although older male workers are generally more prevalent in transitions associated with bin 4 (the top 20%), while female workers are more likely to be associated with transitions with bin 2 (the bottom 20%).

Next, we decompose the inflows and outflows further by using the transition matrix of counts. Figures 8–12 show the outflows and inflows for each earnings/inactivity category and the relevant transition counts by year. In particular, each year an individual can be in one of the following states: bin 0, not in the eligible-workers frame, or bins 1–4 as defined in Section II.D. We remind the reader that the cutoffs for the real earnings distribution bins are based on the pooled years 2004–13 and do not change over time. It is possible for substantially more or less than 20% of current earners to be in the top 20% bin, for example, in any given year.

The first subplot in each figure shows the gross outflows (blue line) and the gross inflows (red line) into an earnings/inactivity category. The difference between the two lines is the net inflows presented in the last column of Table 5. The second subplot shows the transitions to other categories associated with the gross outflows line in the first graph. Note that the sum of each series in the second subplot is equal to the gross outflows line in the first subplot. The third subplot shows the transition counts associated with gross inflows. Similarly, the sum of each of the series in the third subplot is equal to the gross inflows line in the first subplot.

Figure 8 plots the counts of workers moving out of the eligible-workers frame and those moving into the frame. Notice that gross inflows into eligibility are always greater than gross outflows from eligibility. Thus, net inflows into eligibility are always positive but decline and fail to recover after the Great Recession. This decline is primarily due to a decrease of inflows into eligibility, although there is a small increase in outflows as well. Notice that the outflows from eligibility come primarily from inactivity. Recall that inactive individuals are part of the eligible-workers frame but have no reported earnings. The first row under the “Inactivity” origin category in Table 6 shows that inactive workers moving to ineligibility (1_0) tend to be older, with an average age of 68. The inflows into eligibility tend to be young workers around the age of 20 moving into either inactivity or the bottom 20% of the earnings distribution. The exception are individuals moving into the top 20% of the earnings distribution (0_4), who tend to be older (average age is 33 years old), male (79.1% are male), and foreign born (only 12.7% are born in the United States). Prior to the Great Recession, on average 38.3% of workers moving into eligibility went into inactivity and 49.1% went into the bottom of the earnings distribution. This flipped during the Great Recession, with 47.5% of newly eligible workers moving into inactivity and 43.4% moving into the bottom of the earnings distribution.

Figure 9 plots the flows into and out of inactivity. Net inflows are generally close to zero, except during the Great Recession, when they became very positive. This spike in net inflows into inactivity was driven by both a large increase in inflows and a substantial decline in outflows. The increase in inflows is seen for every category except for the top of the distribution. The decrease in outflows during the Great Recession is primarily due to a reduction in workers moving to jobs in the bottom and middle of the earnings distribution; however, these flows return to their prerecession levels fairly quickly. The net result is a roughly symmetric increase in gross outflows and a decrease in gross inflows. Without either a very large relative increase in gross outflows or a relatively large decrease in gross inflows, little progress can be made toward reducing the almost 13 million person increase in the number of eligible workers with no reported earnings during the Great Recession.

Figure 10 plots the flows into and out of the bottom 20% of the earnings distribution. Compared with some of the other categories, the counts for the bottom 20% are relatively stable. The transitions show large changes occurring during the Great Recession. Outflows increased to nonemployment, while outflows to the middle of the earnings distribution fell (fewer workers moving up). At the same time, inflows from nonemployment decreased and inflows from the middle of the distribution increased. Workers who moved to nonemployment are being replaced (not at the job level but in terms of earnings) with workers from the middle of the distribution.

Figure 11 shows the flows into and out of the middle 60% of the earnings distribution. There is a large decrease in net inflows in 2009. This is largely due to an increase in the outflows of workers to the bottom 20% and nonemployment and a decrease in workers moving up from the bottom 20%. Although net inflows turn positive again after the Great Recession, these inflows are not large enough to halt the decreasing share of earnings accruing to workers in the middle 60%. There has also been a decline in workers moving from the middle to the top 20%, which peaked in 2007, implying a decrease in upward earnings mobility. A decrease in workers moving from the top to the bottom is also present, implying a decrease in downward earnings mobility. Postrecession, workers in the middle are more likely to stay in the middle, and workers at the top are more likely to stay at the top.

Figure 12 plots the flows into and out of the top 20% of the earnings distribution. Notice that there is a strong net inflow of workers to the top 20% prior to the Great Recession and a decrease during the recession. Although net inflows turn positive again in 2012, they do not return to the heights seen prior to the Great Recession. Workers at the top are relatively disconnected from the rest of the earnings distribution. The only substantial flows are to and from the middle, but the magnitudes of both of these flows appear to be declining after the recession.⁴⁴

Overall, mobility occurs most often between neighboring parts of the earnings/inactivity distribution. It is relatively rare to jump more than one earnings/inactivity category. For example, moving from the bottom to the top is a relatively rare event, while moving from the bottom to the middle is a common transition.⁴⁵

A. The Worker-Firm Earnings Decomposition

Having established (1) that the eligible-workers frame is more likely to be representative of the entire US labor market than uncorrected administrative record frames and (2) that it is complete and suitable for studying changes in the entire earnings distribution, including the movements into and out of activity, we now turn to the role of the employer as a source of earnings inequality. We use the statistical approach introduced to the labor economics literature in Abowd et al. (1999).

To understand the role of firms in the rise in earnings inequality, we estimate the following AKM model for the eligible-worker sample:

During the process of converting job-level firm effect estimates to person-level firm effect estimates, we move back and forth between dollars (levels) and logarithms as appropriate. We estimate equation (3) in semilogarithmic form due to the approximately lognormal conditional distribution of the dependent variable. However, a semilog specification returns estimated firm effects showing the log points of earnings attributable to the employer (approximately proportional for small effects). To combine the estimated job-level firm effects for workers who have multiple employers during the year into a single person-level firm effect, the relative firm effects must be converted to dollar values. For example, suppose we have a worker who earns $10,000 at job 1 and $50,000 at job 2, with fixed firm effects of 0.2 and 0.1, respectively, in logs. Although the estimated firm effect is twice as large in job 1, the earnings are only 20% of the earnings in job 2. To account for the earnings differences across jobs, we convert the estimated firm effects to dollars, aggregate the dollar amounts, and then convert the dollar value of the person and firm components back to logs.

First, we isolate the real dollar value of the firm component of earnings for a given job in a given year. Specifically, for individual i working at firm j in year t, the firm component of earnings y_ijt is defined as

Aggregating across jobs for each worker gives a decomposition of total annual earnings into two person-level components:

We also extract a component of log earnings that we can associate with the worker’s skill type. This component consists of the constant, which has been standardized to a year with average unemployment and full-year, full-time work; the estimated effects associated with labor force experience; and the estimated person effect. Thus, the skill component can represent the log earnings associated with the annual wage rate for a worker with a given person effect and labor experience.⁴⁷

We calculated both dollar and log estimates of the firm and nonfirm earnings components. We calculated only the log component for estimating the skill type. We used the skill component, which is logarithmic, to classify workers by bottom 20%, middle 60%, and top 20% of the skill distribution. When referencing the discrete distribution of the skill component, we refer to skill types. When referencing the value, we refer to the skill component of log earnings.

We similarly classified the firm component into three bins, and we also used the firm and nonfirm components to decompose earnings in the bottom, middle, and top bins. When referencing the discrete distribution of the firm component, we refer to firm types. We do not classify the cells of the discrete distribution of the nonfirm component of earnings. In all cases, distributions were estimated using the pooled person-level observations over the 2004–13 period. See Table 7 for the minimum, maximum, and average log values of all three components.

Our approach has three main benefits. First, all workers at a firm with only one job during the year are placed in the same firm-type bin. Second, the total earnings of the worker do not affect the firm-type bin assignment. Third, classifying the workers by their skill bin rather than the nonfirm bin controls for the state of the labor market and labor force attachment as well as eliminates the influence of the AKM residual. Using the log values of each component also allows us to study all possible mixes of worker skill types. If we had used the dollar-value bin assignments, the highest-paid workers would have dominated the top and bottom categories for each estimated component. For example, a worker with a very small log firm effect but high earnings would likely dominate a low earning worker with a large log firm effect.

Comparing the earnings bins in Table 2 with the firm and nonfirm component bins in Table 7, notice that, except for a relatively small number of extreme values, the distribution of log earnings and log nonfirm earnings component are similar within bins. This result is due to the small relative magnitude of the log firm component.⁴⁸ In spite of their relatively small magnitude in logs, the firm components can have a substantial effect, conditional on the size of the nonfirm component of earnings. The log firm component is about −0.716 for the typical firm in the bottom of the firm compensation distribution, 0.113 in the middle, and 0.807 at the top. All statistics are worker duration-weighted averages, implying that a worker at the middle-type firm receives about 11% more than would be expected given the worker’s characteristics. In comparison, the difference between the average log earnings value and the average log nonfirm earnings component for workers in the middle is 0.135 ($=9.938-9.803$), which is very similar to the 0.113 estimated for the worker duration-weighted average log firm component. Although this comparison is not a true worker-level comparison, it should be similar given that most workers are in the same nonfirm-type and overall-earnings bins.

There is some evidence of worker sorting. Taking the difference between the average log earnings for each overall-earnings bin (Table 2) and the average log nonfirm components for the same nonfirm bin (Table 2) shows that workers at the bottom of the overall-earnings distribution and the bottom of the nonfirm component distribution tend to work in lower-paying firms ($7.473-7.708=\mathrm{-0.235}$) compared with workers at the top ($11.236-10.976=0.260$).

Consider the magnitude of the average firm effect for each firm type and its potential effect on workers in different parts of the nonfirm component distribution. For example, the typical worker at the bottom of the nonfirm component distribution has average log earnings of 7.708. If this worker were employed at a firm in the middle of the firm component distribution, his log earnings would be greater by about 0.113, which is not enough to move him to the middle of the overall-earnings distribution ($7.708+0.113=7.821<9.938$). Even if this worker were able to transition to a firm at the top of the firm component distribution, ceteris paribus, his log earnings would be greater by 0.694 ($=0.807-0.113$) log points. The resulting log earnings of 8.515 ($=7.708+0.807$) would still not be large enough to move the worker to the middle overall-earnings bin. In comparison, for a worker in the top of the nonfirm component distribution, moving from a middle to a top firm results in an earnings increase large enough for the worker to transition from the middle to the top of the overall-earnings distribution ($10.976+0.807=11.783>11.236$). Although the relative effect is the same, the dollar value of the effect of working at a high-paying firm increases the greater a worker’s nonfirm component of earnings is.

Table 6 shows characteristics of workers associated with each of the 24 possible earnings and inactivity transitions. Given that the overwhelming majority of workers are in the same bin of the overall earnings and nonfirm component distributions, the characteristics of the workers in each corresponding nonfirm component bin will largely be the same. In Table 8, we show the characteristics of firms across each of the three firm-type categories. There are clear differences in the industry distribution by where the firm lies in the firm component distribution. Low-paying firms are highly concentrated in Trade, Transportation, and Utilities and Leisure and Hospitality, with more than 50% of workers at low-paying firms in these two industries. Firms in the middle of the pay distribution are not nearly as concentrated by industry, but nevertheless workers in these firms are prevalent in Trade, Transportation, and Utilities; Education and Health; and Manufacturing. Somewhat surprisingly, except for Leisure and Hospitality and Finance, the distribution of workers in high-paying firms across industries is relatively diffuse. Most industries have a substantial number of workers in high-paying firms, implying that, except for Leisure and Hospitality, it is not necessary to change industries to work at a high-paying firm. As found in other studies, low-paying firms tend to be both younger and smaller than high-paying firms (Haltiwanger et al. 2012, fig. 7).

B. Earnings and Mobility by Person and Firm Type

In this section, we use the AKM decomposition to explore how the three types of workers (bottom, middle, and top of the skill-type distribution) sort into the three types of firms (bottom, middle, and top of the firm-type distribution). The results for each worker type are presented separately. Tables 9, 10, and 11 present outcomes for workers in the bottom, middle, and top bins of the skill-type distribution, respectively.

The tables were created as follows. Bin types are based on the previous year’s data, that is, year $t-1$ classifications. Beginning in 2004 and ending in 2012, for every year that an eligible worker has positive earnings a single observation is added to one of the three tables. The appropriate table classification for each observation is determined by the skill type of the worker for that year, which can vary over time as workers accumulate experience. Within each skill type, the earnings record is further classified based on the firm type, resulting in each earnings observation being classified into one of nine possible cells.⁴⁹ Within each of the skill-type × firm-type cells, we break down the results by the three possible overall-earnings outcomes (bottom, middle, and top). There are, thus, 27 cells for which we present information on the number of workers, average earnings for the previous year $t-1$, and average earnings for the current year (t) by flow type.⁵⁰

To fix ideas, we will take a detailed look at two rows in Table 9. To be recorded in this table, the person must have been in the bottom (lowest) bin of the skill-type distribution in the previous year, that is, $t-1$.

Consider the first row of the table. This row is in the panel labeled “Bottom Firm,” indicating that this person is employed at a firm in the bottom bin of the firm-type distribution in $t-1$. Persons in this row are also in the bottom bin of the overall-earnings distribution in year $t-1$, and 0.782 is the share of such persons relative to those in the middle or top of the overall-earnings distribution. The flow labeled 2_0 is the movement from the bottom of the overall-earnings distribution (bin 2) to the ineligible (bin 0); that is, this is the flow out of the frame for persons at the bottom of the overall-earnings distribution. There were, on average, 39,565 such persons each previous year ($t-1$). They represent 0.7% of the flows from bin 2 of the overall-earnings distribution for low-skill workers in bottom-paying firms. Average earnings in $t-1$ were $1,921, of which −$2,285 are attributed to the firm component of our decomposition and $4,207 are attributed to the nonfirm component of our decomposition. There were no earnings in the current year (t) because the person has moved out of the frame in t.

Next consider the row labeled “Middle” in the “All Earnings” column in the “Middle Firm” panel with a 3_3 flow. All persons in this row were, once again, at the bottom of the skill component distribution in year $t-1$. Of all low-skill persons, 60% are employed by a firm in the middle of the firm-type distribution. Of all low-skill persons employed at middle firm types in year $t-1$, the proportion 0.608 were in the middle of the overall-earnings distribution. Among such persons, the 3_3 row shows those who remain in the middle of the overall-earnings distribution in the current year, t, of which there were, on average, 8,507,780 in the nine pairs of years for which the table was constructed. Those who stayed in the middle of the overall-earnings distribution represented 82.4% of all persons who were in the middle of the overall-earnings distribution, in the low-skill bin, and in a middle-paying firm in year $t-1$, on average. In year $t-1$, their earnings averaged $17,361, of which $2,337 are attributed to the firm component of our decomposition and $15,024 are attributed to the nonfirm component of our decomposition. In the current year, t, average earnings were $18,013, of which $2,619 are associated with the firm component and $15,394 are associated with the nonfirm component.

We use these tables to investigate worker sorting directly by looking at the interaction of the skill and firm type for each worker-year-earnings observation. If there were no sorting, the distribution of earnings observations across firm types would be similar for all three tables because outcomes would be unaffected by which part of the skill-type distribution an individual occupied, given his place in the overall-earnings distribution. This hypothesis is clearly not supported by the data and forms the basis of our major conclusion that the influence of the firm operates through channels that are, at least in part, different from the channels that intermediate the skill-type effect. For example, again using Table 9 showing the bottom of the skill-type distribution, about 27% of the earnings observations are in firms at the bottom of the firm-type distribution, 60% are in firms of the middle type, and only 13% are in top firms. By comparison, tables 10 and 11 show that persons in the middle and top of the skill-type distributions are much less likely to be employed at firms in the bottom type (17% and 21%, respectively) and are much more likely to be employed at top firms (21% and 25%, respectively).

Focusing on each skill type, we start with the earnings observations for low-skill types in Table 9. For workers at the bottom of the skill-type distribution, working at a higher-paying firm has two advantages: higher earnings than otherwise and a greater chance of moving to a higher bin in the overall-earnings distribution. For example, a worker at the bottom of the skill-type and overall-earnings distributions has a probability of moving to the middle of the overall-earnings distribution of 16.8% at a low-paying firm, 24.2% at a middle-paying firm, and 23.3% at a high-paying firm. Prior to the 2_3 transition, the average low-skill worker at a low-, middle-, and high-paying firm earns $3,398, $3,726, and $3,749, respectively.⁵¹ After the transition, the average low-skill worker at a low-, middle-, and high-paying firm earns $10,487, $11,739, and $15,047, respectively. Most of the additional increase in earnings for workers employed at a top-paying employer in the previous year is due to working at a top-paying employer in the next year.

Table 10 supports a similar conclusion. Middle-skilled workers in the bottom of the overall-earnings distribution also have a greater chance of moving to the middle of the earnings distribution the higher the firm type for which they work. When they transition, their current-year earnings will also be greater the higher the firm type for which they work. The vast majority (62%) of workers in the middle of the skill-type distribution are employed at middle-paying firms. The next most prevalent outcomes for such workers are employment at top- and bottom-paying firms, 21% and 17%, respectively. However, in spite of the majority of earnings observations being in the middle of the overall-earnings distribution, average earnings differ substantially across firm types. A middle-skill worker in the middle (bin 3) of the overall-earnings distribution who stays in bin 3 of the overall distribution (a 3_3 flow) at a bottom-type firm has $t-1$ earnings of $15,820, a middle-skill worker in a middle-type firm has $t-1$ earnings of $24,165, and a middle-skill worker at a top firm has earnings of $31,965. Most of the difference is due to a larger firm effect, although the nonfirm component declines as a middle-skill person is found in increasing firm types, giving back some of the gains. Another benefit of finding employment at a high-paying firm is a greater probability of moving to the top of the earnings distribution. For middle-skill workers in the middle bin of the overall-earnings distribution in the previous year, the relevant comparisons are as follows. The estimated probability of a 3_4 transition for workers in a low-paying firm is 0.6%. The estimated probability of the same transition for workers in a middle-paying firm is 2.6%. Finally, the estimated probability of a transition to the top of the overall-earnings distribution for middle-skill workers in a top-paying firm is 11.8%.

Table 11 shows that most of the high-skill workers are also in the top of the overall-earnings distribution, since the top bin of the overall-earnings distribution shows shares of 0.241, 0.585, and 0.891 for low-, middle-, and top-type firms, respectively. There is also a substantial minority in the middle of the overall-earnings distribution. Since transitions to the top of the overall-earnings distribution (3_4) are more likely at top-paying firms (32.7%) than low- or middle-paying firms (4.8% and 14.5%, respectively), we note that once again working at such a firm offers an advantage distinct from the worker’s skill type. In the high-skill category, the earnings differences between working at a middle-type compared with a bottom-type firm after making a 3_4 transition are relatively small, but the earnings gains from working at a top-paying firm are very large. While working at a top-paying firm is clearly preferred and the gains are large, a typical worker in any part of the skill distribution would also have a strong preference for working at a middle-paying rather than a bottom-paying firm. Although the dollar gains may be relatively small, the difference in earnings for bottom- and middle-paying firms is significant. For example, 78% of the low-skill persons employed at a bottom-paying firm are at the bottom of the overall-earnings distribution, while only 38% of the low-skill persons employed at a middle-paying firm are at the bottom of the overall-earnings distribution. Overall earnings within the bottom bin are not dramatically different in this case, but workers in the middle bin of the overall-earnings distribution have noticeably higher earnings at a middle-paying firm ($17,361) than at a bottom-paying firm ($12,342). Somewhat surprisingly, there are a relatively large number of top-skill workers at bottom- and middle-type firms. On average, these workers, especially in the middle, are employed at worse-paying firms than middle-skill workers.⁵²

Table 12 analyzes the earnings of individuals who do not move in the earnings distribution between consecutive years by firm and skill type. This table allows us to analyze the potential effects of redistributing the skill types across firm types or redistributing the firm types across skill types. For the bottom- and middle-skill types, there is almost no advantage to being employed in a higher-paying firm type, given their place in the earnings distribution. For example, a bottom-skill person in the top 20% of the earnings distribution earns $66,745 in a bottom-paying firm and $72,919 in a top-paying firm. But for a top-skill person, there is a big advantage to employment in a top-paying firm: $80,717 compared with $143,736 when in the top 20% of the earnings distribution. The skill pay premium for all firms is approximately the same when considering a bottom-skill as opposed to a middle-skill worker—for example, $12,342 versus $15,820 for a bottom firm and $17,361 versus $24,165 for a middle firm. This is also the case for considering a middle-skill versus a top-skill worker except for the top-paying firms, where that premium is much greater than for middle- or low-paying firms—$67,457 versus $143,296 for the top-paying firm compared with $61,625 versus $92,507 for the middle-paying firm.