Discussion Draft Do Not Cite Without Permission

Exploring Differences in Employment between Household and Establishment Data∗

First draft December 17, 2007 This draft September 29, 2008

Katharine G. Abraham University of Maryland

National Bureau of Economic Research

John Haltiwanger

University of Maryland National Bureau of Economic Research

Kristin Sandusky

U.S. Census Bureau James Spletzer

U.S. Bureau of Labor Statistics

∗ This document reports the results of research and analysis undertaken in part by U.S. Census Bureau and U.S. Bureau of Labor Statistics staff. This work is unofficial and has not undergone the review accorded to official publications of these agencies. All results have been reviewed to ensure that no confidential information is disclosed. The views expressed in the paper are those of the authors and not necessarily those of the U.S. Census Bureau or the U.S Bureau of Labor Statistics. The authors are grateful for financial support from the Princeton Data Improvement Initiative. The current draft is a revised version of a paper circulated and presented at the January 2008 meetings of the American Economic Association. We thank Sasan Bakhtiari, Mary Bowler, David Johnson, Anne Polivka, and Greg Weyland for helpful discussions about data and conceptual issues. The paper has benefited from the comments of Kristin McCue, Bruce Meyer, and Martha Stinson on earlier drafts.

1

I. Background Each month, the Bureau of Labor Statistics (BLS) releases data on current employment. The data come from two different surveys – the Current Employment Statistics survey (CES, also referred to as the payroll survey) and the Current Population Survey (CPS, also referred to as the household survey). The CES is a monthly sample survey of 160,000 businesses covering approximately 400,000 worksites, and is designed to measure employment with industry and geographic detail. The CPS is a monthly sample survey of approximately 60,000 households, and is designed to measure employment and unemployment with demographic detail. The CES and CPS have several well known conceptual and scope differences.1 The CES measures the number of non-farm payroll jobs, whereas the CPS measures the number of employed persons. The jobs-based employment measure in the CES counts all jobs held by a given person, whereas the person-based employment measure in the CPS counts individuals with multiple jobs only once. The self-employed are excluded from the CES but included in the CPS. Employees of all ages are included in the CES, whereas the CPS only measures workers aged 16 and over. Other differences are that the CES excludes workers in the agricultural sector, private household workers, unpaid family workers, and workers on leave without pay, whereas these workers are included in the CPS. Another potentially important difference is that the reference period in the CES is the pay period that includes the 12th of the month, whereas the CPS reference period is the week that includes the 12th of the month. Despite their different definitions, samples, estimation procedures, and concepts, the CPS and CES employment series track well over long periods. There have been times, however, when their rates of growth and decline have differed significantly. Figure 1a graphs seasonally adjusted employment from the CPS and the CES over the years 1994 through 2006, along with a second CPS employment series that has been adjusted to be similar in concept and definition to the payroll survey employment series. 2 For the adjusted CPS series, agricultural and related employment, nonagricultural self employed, unpaid family workers, private household workers, and workers absent without pay from their jobs are subtracted from the usual CPS series, and the number of nonagricultural wage and salary multiple jobholders then is added back to approximate a jobs-based rather than a person-based series that is similar in concept to the CES series. Further details about these adjustments can be found in Bowler and Morisi (2006). The CES-CPS employment trend discrepancy that motivates our research is evident in Figure 1a. As can be seen in the chart, the adjusted CPS series matches the CES series well between 1994 and 1997. From 1998 through the onset of the 2001 recession, CES payroll survey employment grew faster than adjusted CPS employment. During and immediately following the end of the 2001 recession, CES employment declined while adjusted household survey employment remained relatively flat. By 2003, the CES and the adjusted CPS series matched 1 Bowler and Morisi (2006) document the differences between the two surveys. A summary comparison of the CES and the CPS also can be found in the document “Employment from the BLS household and payroll surveys: summary of recent trends” which is available at http://www.bls.gov/web/ces_cps_trends.pdf. 2 Figure 1a is from http://www.bls.gov/web/ces_cps_trends.pdf. The CPS employment series in Figure 1a smooths out the effects of population control revisions that occurred during 2003-2007. This smoothed series is available at http://www.bls.gov/cps/cpspopsm.pdf.

2

again in levels, and from 2004 to the end of 2006, employment as measured by both surveys has trended upward. Although the divergence between the CES and CPS employment series that occurred between 1998 and 2003 is unprecedented in size and duration, there have been past divergences. This is evident in Figure 1b, which provides a 60 year history of the ratio of CES employment to CPS nonagricultural wage and salary employment. In 2000, CES employment exceeded CPS nonagricultural wage and salary employment by more than 5.5 percent, a discrepancy that is, notably larger than the previous peak discrepancies of 4 percent in the early 1950s and 3.5 percent in 1969. Also evident in Figure 1b is the cyclical behavior of the CES to CPS employment ratio. Establishment survey employment typically increases relative to household survey employment during business cycle expansions, then falls in relative terms during recessions and the early part of the subsequent recovery period. Any story about the trend discrepancy between CES and CPS employment must account for this business cycle pattern in the difference between the two series. Much work has been done to try to explain the cause of the CES-CPS employment trend discrepancies, especially the discrepancy observed between 1998 and 2003. As noted by Bowler and Morisi (2006), “(a)lthough many theories about the discrepancies have been put forth, complete explanations have never been found for any of the divergences, despite a significant amount of research by the Bureau of Labor Statistics and by outside analysts.” Bowler and Morisi provide a thorough review of the work done on possible explanations for the divergence between the two series even after adjustment to make the CPS series more comparable to the CES series.3 The list of potential explanations includes sampling error in the two surveys; possible problems with the benchmarking of the payroll survey; possible problems with the population controls used in the household survey and especially problems with accounting correctly for net immigration in setting those controls; the fact that the CPS excludes employed persons under the age of 16 years; the fact that some multiple jobholders in the CPS hold more than two jobs; possible second jobs in the civilian sector held by persons in the armed forces; employment among the institutional population that is not covered in the CPS (for example, prisoners working outside the correctional institution where they are incarcerated); jobs held by Canadians or Mexicans who commute into the United States; problems with the treatment in the data of government-subsidized employment programs such as welfare to work; and the possibility of missed establishment births in the CES. None of these possible explanations appears to account for the CES-CPS employment trend discrepancy. Another perhaps more promising class of explanations for the observed CES-CPS employment trend discrepancy revolves around problems with the accuracy of the employer job reports collected in the CES and/or the accuracy of the employment status information collected from household members in the CPS. One such explanation rests on potential errors in the identification of workers as self-employed in the CPS. Measuring the self-employed has always been difficult, but these difficulties have been compounded by recent increases in the prevalence of small, self-directed sources of additional household income and in the prevalence of contracting out. These changes have made it less straightforward for survey respondents to 3 Earlier studies that have examined alternative explanations for the differing trends in the two series include Juhn and Potter (1999) and Nardone et al (2003).

3

answer standard questions about whether they or another household member are self-employed. Self-employment is subtracted from the CPS total employment series to produce the adjusted CPS employment series shown in Figure 1a. If the number of people in the CPS who are reported to be self-employed when in fact they are wage and salary workers tends to increase during expansions and to fall during recessions, this could explain the widening of the CES-CPS gap during upturns and the closing of the gap during downturns. Although this may be an explanation that merits further investigation, preliminary evidence reported by Bjelland, Haltiwanger, Sandusky and Spletzer (2006) suggests that reported self-employment in the CPS as compared to the number of self-employed people recorded in administrative data moves in the wrong direction over the business cycle to explain the observed CES-CPS discrepancy. Problems with the reporting of “off-the-books” employment offer another possible explanation for the CES-CPS employment trend discrepancy. Suppose that CPS respondents who hold off-the-books jobs report themselves as employed but that they do not appear on their employers’ payrolls and thus are not reflected in employer survey responses. If the amount of such off-the-books employment tends to fall in expansions and increase during recessions, this could account for the observed discrepancy between the two employment series. To our knowledge, this is not an explanation that has been explored in previous research. The existence of CPS respondents who fail to report on-the-books work they do for pay offers another candidate explanation for the discrepancy we seek to resolve. Jobs that are short-lived and/or associated with small amounts of earnings – positions that might be termed “marginal” – seem especially likely not to be reported during the CPS interview. If the number of “marginal” jobs that are not reported in the CPS tends to grow during expansions and fall during recessions, this could account for the observed trend discrepancy between the CES and CPS. This is another potential explanation that, to our knowledge, previous research has not explored. Increased turnover during business cycle upturns is a final way in which missing some wage and salary jobs in the CPS could help to explain the observed CES-CPS employment trend discrepancy. Job turnover is a potential source of discrepancy between the two employment series because an individual who leaves one job and starts another may be counted twice in the CES but can only be counted once in the CPS. This is because the CES counts all jobs with positive earnings during the payroll period that includes the 12th of the month, whereas the CPS counts employed people. Although it may be unlikely that two successive jobs held by the same individual both will be counted in the case where the two employers have weekly pay periods, this becomes more likely if one or both of the employers has a longer pay period. The most recent published statistics show that 45 percent of businesses have a weekly pay period, 43 percent a bi-weekly or semi-monthly pay period, and 12 percent a monthly pay period (U.S. Bureau of Labor Statistics 2004). If turnover tends to be procyclical, rising during expansions and falling during contractions, this could create a similar procyclical pattern in the CES-CPS employment gap. Previous research has concluded that the turnover hypothesis is not a major factor in explaining the CES-CPS employment trend discrepancy (U.S. Bureau of Labor Statistics 2004). Based upon questions about the turnover data used to reach this conclusion, however, a December 2005 report by the Federal Economic Statistics Advisory Committee

4

(FESAC) recommends re-examining the hypothesis using a turnover series based on UI wage records rather than the CPS job changer series. In summary, no one has yet put forward a convincing empirical explanation for the CES-CPS employment trend discrepancy that is so evident in Figure 1. Amongst the plausible explanations that might be offered but have not yet been quantified (or satisfactorily quantified) are those related to missing marginal jobs in the CPS, missing off-the-books employment in the CES, and pro-cyclical turnover that creates a wedge between the number of CES jobs and the number of CPS employed persons. In this current paper, we link CPS microdata to administrative UI wage records to draw inferences about the potential contributions of these factors to the divergent behavior of the CES and CPS employment series.4 The paper proceeds as follows. In section II, we provide a conceptual framework for exploring discrepancies in individuals’ employment situations in household-reported as compared to employer-reported data. Section III offers a detailed discussion of the integration of the CPS and UI wage record data on which our empirical analysis rests. In Section IV, we examine differences in the information about employment status (employed versus not employed) and number of jobs held (one job versus more than one job) recorded in the CPS and the corresponding information for the same individuals from the UI wage records. In addition to looking directly at how the magnitude of the discrepancies between the two data sources varies over time, we also examine the personal and job characteristics that are associated with disagreements about employment status or about the number of jobs held. The coefficients from models that relate the probability of disagreement between the two data sources are used together with information on changes in the distribution of worker and job characteristics over time to simulate changes in the number of people we would expect to see recorded as employed in the UI data but not the CPS and vice versa. These simulations yield useful insights into the differing trends in CES and CPS employment. We end with some concluding observations and suggestions for future research. II. Framework for Analysis In a purely accounting sense, discrepancies between CPS wage-and-salary job counts and CES job counts may arise either because people hold wage and salary jobs in one dataset but not the other or because the two datasets differ with regard to the number of wage and salary jobs that people hold. For this analysis, we have constructed a data set that contains CPS records matched with employer-reported information for the same individuals. As will be described in greater detail, the records were constructed in such a way that we have information from both sources about the number of jobs a person held over the first quarter of the year. Before turning to the detailed description of the data, however, we would like to explain briefly how we believe 4 Two recent papers have examined earnings discrepancies as measured in household survey versus and administrative data. Roemer (2002) compares both CPS and SIPP data to administrative data; Abowd and Stinson (2003) focus on quantifying measurement error in earnings in the SIPP versus administrative data. Both papers use the match of household survey records to the Detailed Earnings Records (DER) from the Social Security Administration. These papers focus on earnings rather than employment but are relevant to the present analysis because discrepancies in employment status and the number of jobs held are an important source of discrepancies in earnings.

5

this sort of data can be informative regarding the sources of discrepancy between CES (employer reported) and CPS (household reported) employment and the trends in employment as measured by the two data sources. One way that the levels of CES and CPS employment could differ is that different numbers of people could be recorded as holding wage and salary jobs in the two data sources. With respect to the classification of a particular individual at a particular point in time, as shown in the table below, there are four possibilities:

Wage and salary job reported by employers

No Yes

No

X1

X2

Wage and salary job reported in the CPS Yes

X3

X4

A person may be recorded as having a wage and salary job in neither data set (X1), as having a wage and salary job in the employer-reported but not the CPS data (X2), as having a wage and salary job in the CPS but not the employer-reported data (X3), or as having a wage and salary job in both data sets (X4). Another way that CES and the CPS employment could differ is that, among those the two surveys agree hold wage and salary jobs, some could have multiple jobs in one of the data sources but not in the other. With respect to the classification of those the two surveys agree are employed, as shown in the table below, there are again four possibilities:

Number of wage and salary jobs reported by employers

One Two plus

One Y1

Y2

Number of wage and salary jobs reported in the CPS Two plus

Y3

Y4

A person may be recorded as having a single wage and salary job in both data sets (Y1), as having one wage and salary job in the CPS but more than one in the employer-reported data (Y2),

6

as having more than one wage and salary job in the CPS data but just one job in the employer-reported data (Y3) or as having more than one wage and salary job in both data sets (Y4). 5 Different stories about observed discrepancies between CES employment and CPS wage-and-salary employment have different implications about how we would expect the number of people in the different cells to change over time in response to aggregate labor market conditions, as well as potentially different implications regarding the personal and job characteristics of those observed in the various cells. Although we do not expect to be able to quantify the precise importance of different explanations for the observed divergence between the CES and CPS employment series based on examination of data that classify individuals into the various cells over time, our expectation is that certain explanations will prove to be inconsistent with the data while support will be offered for others. The principal fact for which we are seeking an explanation is the faster growth of employer-reported CES employment compared to household-reported CPS employment during tightening labor markets and the reverse during economic downturns. One possible explanation for the CES-CPS trend discrepancy might be that, in tight labor markets, a growing number of marginal jobs are not reported in the CPS. As economic activity strengthens, employers may become more inclined to hire extra help to cover peak workloads, raising the number of short duration wage and salary jobs. For example, the owner of a retail store might decide to hire 5 temporary staff over the Christmas holidays rather than 2 or 3 such people. To the extent that short duration jobs are less likely to be reported by CPS respondents, either because they do not overlap the week of the 12th or because the respondent fails to report short duration jobs that were in progress during the reference week, this might lead us to expect an increase in X2 as the economy tightens (i.e., to expect that X2 will be procyclical). A similar dynamic might be in play for Y2. A second possible explanation for the CES-CPS trend discrepancy might be that it reflects differences in how the two sources capture informal or “off-the-books” employment. Suppose that, as economic activity strengthens and labor markets become tighter, people tend to leave informal jobs (jobs not recorded on employer payrolls) for formal jobs (jobs that employers report). Alternatively, during periods of stronger economic activity, employers might “regularize” more of their jobs, converting them from off-the-books positions to jobs for which the employer pays applicable employment taxes. These changes could be reflected in the job count matrices in two ways. First, as labor markets tightened, we might expect X3 to fall. In addition, to the extent that those affected are multiple job holders, we might expect Y3 to fall. Put another way, we might expect both X3 and Y3 to be countercyclical. A final possibility is that the frequency of job changing may increase in tight labor markets. In the CPS, someone who changes jobs outside of the week of the 12th should 5 There are a few individuals who hold multiple wage and salary jobs in one data source and are not employed in the other data source, but only a small fraction of multiple job holders fall into these categories. In the data we examine later in the paper pooled over the years 1996 to 2003, those who are recorded as employed in the UI data but not employed in the CPS (the X2 cell) account for less than 6 percent of those with multiple jobs in the UI data. Similarly, those who are recorded as employed in the CPS but not employed in employer-reported unemployment insurance (UI) data (the X3 cell) account for only about 10 percent of those with multiple jobs in the CPS using a restrictive definition and 15 percent of those holding multiple jobs in the CPS using a more lenient definition.

7

contribute only one job to the wage and salary job count; in the CES data, the same person might contribute two jobs to the wage and salary job count. Increases in the frequency of job changing in tighter labor markets could lead to increases in Y2 (i.e., Y2 may be procyclical). The table below summarizes the different possible stories that seem to us to have implications for the data we will be examining:

Story Implications for Different Employment Cells

Marginal (short duration or low earnings) jobs that are not reported by household survey respondents grow in number during business cycle expansions

X2 procyclical Y2 procyclical

Informal or “off-the-books” jobs that are not reported by employers shrink in number during business cycle expansions

X3 countercyclical Y3 countercyclical

Increases in the job-changing rate during business cycle expansions lead to increases in employer-reported job counts

Y2 procyclical

Information on the number of people in the various cells in the two data tables and on how those numbers have changed over time should in principle allow us to evaluate these different hypotheses. In considering the cyclical implications of this table, a more general interpretation is that X2 should increase relative to X3 in cyclical upturns, while X3 should increase relative to X2 in cyclical downturns. Likewise, Y2 should increase relative to Y3 in cyclical upturns, while Y3 should increase relative to Y2 in cyclical downturns. In our empirical work, we begin by creating annual estimates of the weighted number of people in each of the X1 - X4 and Y1 - Y4 cells. Our next step is to examine whether the variation over time in the number of people in the off-diagonal cells X2, X3, Y2, and Y3 is consistent with our hypotheses about the role played by marginal jobs, informal jobs and procyclical turnover in explaining the cyclical pattern of the gap between CES and CPS employment. In addition, we consider whether the demographic and job characteristics of those observed in the off-diagonal cells is consistent with expectations based on our three hypotheses. III. Data and Measurement

To examine the levels and cyclical behavior of the various X’s and Y’s in the cells of the data tables just described, we need direct individual-level comparisons of worker- and employer-based employment information. Both because the CES and the CPS are sample surveys, meaning that neither covers the entire population, and because the CES does not collect information about individual workers, no direct linkage can be made between the two surveys. Instead, we link CPS information to unemployment insurance (UI) wage record data. This is a reasonable strategy because the CES employment estimates are benchmarked to the Quarterly

8

Census of Employment and Wages (QCEW) employment counts, and, since the QCEW derives from reports employers must file along with payment of their UI taxes, the QCEW employment counts should be highly correlated with the UI wage record employment counts. Inferences based on comparisons of CPS and UI wage record data thus should carry over to understanding the differences in behavior of the CPS and CES employment series.

Analysis Sample

In order to make direct comparisons between the CPS and UI data, we must be able to

match workers in the two datasets. Each quarter, the states that administer the UI program collect reports from virtually all non-agricultural private-sector employers, state government units and local government units on the earnings in that quarter of all their employees. Small agricultural employers frequently are excluded from state UI systems and the federal government administers its own separate UI system. State UI records also do not reflect the earnings of unincorporated self-employed workers. Because we are attempting to mimic the CES series in the UI data, we in any case do not want to include agricultural workers, private household workers, or the self employed, all of whom are excluded from the CES. Our analysis therefore focuses on what we term in-scope employment – wage and salary employment in the private sector excluding agriculture and private household jobs, plus state and local government employment. Each worker experiencing in-scope employment is identified in the UI reports by his/her Social Security number (SSN). All files containing SSNs received by the Census Bureau are maintained and protected in a secure environment within an administrative records division. The administrative records division validates the SSNs and replaces the SSNs with a Protected Identity Key (PIK). The PIK is the person ID used internally at Census to process and integrate person-level data. The U.S. Census Bureau maintains PIK information for a large fraction of CPS respondents.6 A linked dataset including both UI and CPS information has been constructed by the Census Bureau’s Longitudinal Employer-Household Dynamics (LEHD) program.7 We use these data to perform our analysis.

Although the linked UI-CPS data permit a direct comparison of survey and administrative micro data on employment, there are some notable limitations of the linked data that must be addressed before the administrative and survey employment measures for these respondents can reasonably be compared. First, although both the CPS and the CES produce monthly information on employment, the UI data contain reports of employment and earnings over the course of a quarter and a monthly decomposition is not feasible. Thus, the monthly CPS employment data must be modified to generate quarterly employment measures for comparison with the UI-based measures. Because we must construct a quarterly employment measure, our attention is restricted to individuals for whom CPS reports were present in all three months of the first quarter of the year, that is, to those who responded in January, February and March. Given the sample rotation pattern for the CPS, this means our sample includes only individuals who were in their 3rd, 4th, 7th, or 8th month in the CPS sample in March of the year in question, cutting the sample otherwise available roughly in half. Second, all UI records contain a PIK but in the CPS the PIK is collected only for respondents to the Annual Social and Economic Supplement

6 Census compares demographic characteristics of CPS respondents to the SSA master SSN database in order to validate that the respondent provided the correct SSN. 7 A comprehensive description of the LEHD data infrastructure can be found in Abowd et. al. (2006)

9

(ASEC) conducted as part of the March CPS and is available for only about 70 to 80 percent of supplement respondents. Finally, although the LEHD program is approaching national coverage, data are available in all years from the mid-1990’s through the present only for the following 17 states: California, Colorado, Florida, Idaho, Illinois, Kansas, Maryland, Minnesota, Missouri, Montana, New Mexico, North Carolina, Oregon, Pennsylvania, Texas, Washington, and Wisconsin. In most states, 95 percent or more of the workers resident in the state also are employed in the state, but more than 15 percent of workers resident in Maryland (where they are coded in the CPS) actually work in another state or the District of Columbia (where their employers pay unemployment insurance taxes).8 We chose to drop Maryland from our analysis, leaving us with data for 16 states. These states account for roughly 50 percent of employment nationwide.

In addition to these constraints on the CPS records we can use, we restrict our attention to

CPS records for persons aged 16 years or older. This restriction makes the age range of our sample comparable to the age range of the population covered by published CPS employment estimates.

Measuring Employment and Job Characteristics As described in section two, we compare two primary measures across the CPS and the UI wage records. First, for each observation in our analysis sample in each year, we compare whether or not the person is an in-scope worker in the CPS with whether or not they are an in-scope worker in the UI data. Second, among those who are observed to work in both datasets, we compare the number of in-scope jobs held during the quarter. Measuring the number of unique in-scope jobs held by a respondent in each quarter in both the UI and the CPS data is not a straightforward task. The strategy we have adopted is outlined below, together with a brief description of other measures we have created of the characteristics of UI and CPS jobs held during the quarter. UI Variable Construction. For employers who are covered by a state’s unemployment insurance system, the UI data contain a variable identifying the employing business, a variable identifying the worker, and a record of the earnings the worker received at the business in each quarter.9 Thus, there is a UI record for each employer-employee match present in each quarter. Using this information, we create a first variable to indicate whether a person in the linked sample is observed with positive in-scope UI earnings during the first quarter and a second

8 According to American Community Survey data for 2005 obtained from staff in the Local Area Unemployment Statistics program at the Bureau of Labor Statistics, 2.33 percent of employed people in our 16 states lived in one state but worked in another. This percentage was above 5 percent only in Kansas (7.22 percent), Missouri (5.64 percent) and Pennsylvania (5.39 percent), and was very low in the three largest states in our sample, California (0.54 percent), Florida (1.24 percent) and Texas (1.16 percent). The corresponding percentage for Maryland was 16.74 percent. If the state where a CPS respondent works is included among the states for which we have UI wage records, we are able to link to that information even if a person lives and works in different states. Maryland is a particular concern because we do not have UI data for Virginia, West Virginia or the District of Columbia. 9 In the UI wage records, the employer ID is a state unemployment insurance account number for the business. For multi-establishment businesses, this employer ID is typically at a level above the establishment but below the firm, generally representing the activity of the firm at an industry-state level. For details about the UI wage record data, see Abowd et. al. (2006).

10

variable that is equal to the number of different in-scope employers from whom the worker received positive earnings in the quarter. We also construct a measure of earnings during the quarter on the worker’s primary job, defined simply as the highest-earning job, and, for those who hold more than one job during the quarter, a measure of earnings on any additional jobs. By using information from the preceding and subsequent quarters (the 4th quarter of the previous year and the 2nd quarter of the current year), we are able to construct additional variables pertaining to the jobs recorded in the UI data file. If we observe that a worker is employed at a particular business in the first quarter and also worked for that same employer in the preceding and/or following quarter, we code that job as a long-duration job. Using this information, we create variables to indicate whether an individual had at least one long-duration job and whether the individual held any secondary jobs that were of long duration.

CPS Variable Construction. Although the basic CPS data are collected monthly and thus

in some sense should be richer than the UI data, it is more difficult than might at first be apparent to construct from these monthly records quarterly measures of unique wage and salary jobs that can be used to make meaningful comparisons to the person records from the UI data file. Although the CPS records up to four jobs in each month, the class-of-worker information needed to determine whether a job is in-scope is collected as part of the basic CPS interview only for the main job. Class of worker information also is collected for the second job, if there is one, but only in the outgoing rotation months (CPS months-in-sample four and eight). No class of worker information is observed for other jobs.

We define a CPS respondent to have been an in-scope worker in the first quarter if the

worker’s main job in any of the three months is a non-agricultural private sector wage-and-salary job or a wage-and-salary job in state or local government.10 Constructing a count of the number of unique in-scope jobs held over a quarter requires knowing both the number of in-scope jobs held each month and the number of employer changes that may have occurred across months. Because we know the class of worker for second jobs only in the outgoing rotation months and have no class-of-worker information for additional jobs, we cannot be certain about the number of in-scope jobs held by those who report holding multiple jobs at any one point in time. Further, the monthly CPS questionnaire probes only for changes in main job, and even this information is not complete. Given these difficulties, rather than attempting to count the number of in-scope jobs that a worker holds during the quarter, we instead construct indicator variables for whether a worker holds one in-scope job or more than one in-scope job during the quarter.

As has been noted, a worker may hold more than one job during the quarter either by

holding more than one job at the same time and/or by changing jobs. We can identify neither of these situations in the CPS with certainty in all cases. We can, however, separate workers into broad groups within which we have greater or lesser confidence that the worker held more than

10 This definition excludes individuals whose primary job is out of scope but who have a second job that is in scope. Except in an individual’s outgoing rotation months, we cannot say with certainty whether a second job is in-scope. In data for the March outgoing rotation groups covering the years 1996 to 2003, adding those with out-of-scope primary jobs but in-scope second jobs to the weighted count of in-scope workers would raise the total number of in-scope CPS workers less than 1 percent on average.

11

one in-scope job at the same time or that the worker changed employers. By combining these groups in different ways, we are able to generate both a conservative and a more expansive version of the multiple job indicator variable for each worker, and thence both a conservative and a more expansive estimate of the total number of workers who held multiple in-scope jobs in the quarter. These then may be compared to a similar UI indicator variable or to counts of the number of people holding more than one in-scope job during the quarter from the UI data.

Simultaneous Jobholders. Some CPS workers can be identified with reasonable

confidence as having held two or more in-scope jobs simultaneously at some point during the quarter, but for others this is less clear. The complicating factor is that some of those who report holding multiple jobs may hold a mix of in-scope jobs and jobs in agriculture, jobs in private households, jobs in the federal government and/or self-employment jobs, and complete information about the type of job is not collected as part of the basic monthly interview. We consider two possible cases.

First, if a worker has two or more jobs in January, February and/or March, has two or

more jobs in his/her outgoing rotation month (March or April), and both of the jobs held in the outgoing rotation month are in-scope jobs, we can be reasonably confident that the worker held two simultaneous in-scope jobs at some point during the quarter. These people belong to multiple job group 1 (MJ1). In each year, roughly 2 percent of CPS workers fall into this group.

Second, if a worker has two or more jobs in January, February and/or March and has two

or more jobs in the outgoing rotation month, but class-of-workers detail is missing for at least one of these outgoing rotation jobs, the worker could have held simultaneous in-scope jobs at some point during the quarter, but this is less certain. The same is true of workers who have two or more jobs in January, February and/or March but only one job in the outgoing rotation group. These people are assigned to multiple job group 2 (MJ2). In each year, between 4 percent and 5 percent of workers fall into this category. Though many in this group may hold more than one wage-and-salary job, others may not.

Job Changers. Those who changed jobs during the quarter also can be identified with

greater or lesser confidence. The basic CPS questionnaire includes a question asked of those employed both in the prior month and in the current month that indicates whether the respondent has changed employer on their main job, but this question is not always asked even when one might think it should be.11 In addition, respondents are not queried about changes in employer on any jobs other than the main job.

Most respondents who were employed in both January and February were asked in

February if they were still working for the same employer on their main job. Similarly, most respondents employed in both February and March were asked this same question in March. We can be reasonably confident that those who reported a change of employer did indeed change

11 Roughly 7 percent of workers who are employed in adjacent months have a missing value for the variable that records the answer to the job change question. When we asked why this variable was missing for so many people, we were told that the question is a screener to determine whether information on industry and occupation needs to be updated and that interviewers have the discretion just to ask those questions if they have any reservations about the quality of the job information collected in the previous interview.

12

jobs. Respondents who report a change in job from one in-scope employer to another are assigned to job change group 1 (JC1). Depending on the year, between 2 percent and 3 percent of workers belong to this group.

A high fraction of workers either report no job change or were not asked the job change

question. A fair number of these workers, however, have employment patterns that suggest a job change may have occurred. If a worker holding an in-scope job either in both January and February or in both February and March does not report a job change directly but does report a change in industry, occupation, or class of worker between months, we consider him/her to be a possible job changer and the worker is assigned to job change group 2 (JC2). In addition, if a worker held fewer in-scope jobs in February than in both January and March and the worker was not on layoff in February, there is a basis for assignment to group JC2. Again depending on the year, between roughly 4 percent and 6 percent of workers belong to this group. Note that it is entirely possible for a CPS worker both to hold simultaneous in-scope jobs and to have changed from one in-scope job to another, though either alone would be enough for the person to be categorized as holding more than one in-scope job over the course of the quarter. For the results reported in the next section of the paper, we use the indicator variables to construct both a more restrictive and a less restrictive measure of which CPS respondents hold more than one in-scope job during the quarter. The more restrictive measure counts only workers assigned to groups MJ1 and/or JC1 as multiple job holders. The less restrictive measure includes all workers assigned to groups MJ1, MJ2, JC1 and/or JC2. We also use the MJ indicator variables in our investigation of which individuals classified as multiple job holders in the CPS are not so classified in the UI data. Other Job Characteristics. In addition to variables that tell us whether there is reason to believe that a CPS worker held more than one in-scope job simultaneously at any point during the quarter or changed from one in-scope job to another during the quarter, we construct several other job characteristic measures based on the CPS data. First, as a measure of employment instability, we construct an indicator of whether the person held no in-scope job in at least one month of the quarter. Similarly, for those we believe held more than one in-scope job in any month, we construct an indicator of whether the person held more than one job in all three months of the quarter. Information on rates of self-employment by industry and occupation was used to determine whether a person worked in a job in which the probability of being a contractor rather than a wage-and-salary worker is high. Finally, we constructed indicators to distinguish those who worked full time in at least one month from those who did not and to distinguish those who worked 16 hours or more on a second job in at least one month from those whose second jobs always involved fewer hours. Estimation Weights As noted above, because we require CPS records that can be matched with UI wage records for the same individuals, we must impose a number of restrictions on our CPS-based analysis sample. One consequence of imposing these restrictions is that we cannot use the standard CPS weights for our analysis. We use a two-step propensity scoring and weight

13

adjustment procedure to modify the CPS sample weights to make our analysis sample representative of the desired population. Step 1: Because the person identification variable that makes it possible for us to link the CPS to the UI data is available only on the March ASEC file, we restrict our analysis to those who responded to the March CPS and who reside in one of the 16 states for which we have access to UI wage records data. The March basic estimation weight is the foundation for our adjusted sample weights. Because we require quarterly data, we keep only those March respondents who also responded to the January and February monthly surveys. Starting with the full March sample aged 16 and above, we estimate the probability that the individual was interviewed in all three months of the first quarter as a function of age group, gender, race, education, marital status, foreign-born status, and an indicator for whether the person had in-scope employment in March. March basic estimation weights are used in estimating these models. For each observation in the three-month sample, we then construct a weight adjustment factor equal to the inverse of this predicted probability. Only about half of CPS respondents are in the rotation groups eligible to have been interviewed in all three of the months we care about, and others are lost because they moved or failed to respond in one or more months. The average value of this propensity score adjustment factor is roughly 2.5. Step 2: Our analysis also requires that a CPS record have available a PIK for linking to the UI data. Between 20 and 30 percent of March records do not have a PIK. For each record for a person age 16 and above living in one of our 16 states who was interviewed in January, February and March, we estimate the probability that the record has a PIK available as a function of the same traits listed above, but using an indicator for in-scope employment at any time during the quarter rather than during March. The sample weight used in the estimation of this stage is the adjusted weight constructed by applying the weight adjustment factor from Step 1 to the March basic estimation weights. For each record with a PIK, we then apply a second weight adjustment factor equal to the inverse of the predicted probability of having a PIK. This adjustment has a much smaller impact on the distribution of weights, adjusting the value of each weight upward by roughly 20 to 30 percent on average (again, the exact adjustment varies with respondent characteristics). By construction, the estimates of the population for the month of March obtained by applying the final weights just described to our CPS analysis sample match the published CPS population estimates. Estimates based on the first set of propensity-score-adjusted weights we produced yielded employment estimates for the month of March that were noticeably larger than those based on the full CPS sample. This was because these initial models did not account for the fact that employed people are more likely to have a PIK on their CPS records than people who are not employed. Adding a measure of employment to the propensity score models to take this higher probability into account reduced the final weight accorded to employed sample members. Applying these new weights to our analysis sample yielded employment estimates for the month of March that were very close to published CPS employment estimates.

14

Weighted Aggregates Given the various restrictions on our sample and the employment concept we must apply, an important question for our purposes is how the weighted employment series we end up with compare to those observed in published data. To answer this question, we begin with Figure 2a, which compares seasonally-adjusted employment for the month of March taken from Figure 1a. The CES employment series in Figure 2a-1 grows more steeply during the late 1990s than does the CPS employment series, the difference between the CPS and the CES figures is smallest in March 2001, and CES employment falls more steeply than CPS employment from 2001 through 2003 (indeed, CPS employment actually grows between March 2002 and March 2003). These patterns are more obvious in Figure 2a-2, which shows the March-to-March employment trends from the CES and CPS, with both indexed at 1996=100. In Figures 2b-1 and 2b-2, we display CES and CPS employment for our 16 states. The CPS estimates were calculated from the microdata, and the CES estimates were downloaded from the BLS website. The data displayed in Figure 2b are not seasonally adjusted, but using seasonally unadjusted data has no effect on the trends shown in Figure 2a. While there are some minor differences between Figure 2a and Figure 2b, the employment trends and the resulting employment trend discrepancy are essentially the same in our 16 states as in the national data. The effects of our sample restrictions begin to show themselves in Figure 2c, where we limit the scope of the estimates reported to jobs in the private sector, state government or local government (denoted as PSL), removing the self-employed, agricultural workers, private household workers, and Federal government workers from the CPS data and Federal government workers from the CES data. These restrictions are needed so that the scope of the CPS microdata we analyze conforms to the scope of the UI wage records. As is most apparent in Figure 2c-2, in contrast to the pattern in Figures 1, 2a and 2b, CES and CPS data for our 16 states for this measure of employment show essentially the same trend growth rate between 1996 and 2001. Because estimated CES employment starts in 1996 at a higher level than CPS employment, however, it is still the case that CES job gains exceed CPS job gains over this period, by more than half a million over the five years in question. Further, as in the earlier figures, CES employment falls off sharply after 2001, while CPS employment falls more modestly between 2001 and 2002, then rises between 2002 and 2003. The next step in the transition to our analysis sample is to switch from the published CES data to the UI wage records microdata. For the reasons already explained, using the UI data requires a switch from the familiar CPS monthly employment concept to a first quarter employment concept (employed during January, February, or March). As can be seen by comparing the numbers in Figure 2d with those in Figure 2c, switching to a quarterly employment concept increases the level of employment, for the somewhat obvious reason that more people are employed at some point during the quarter than in any one month. The growth trends in Figure 2d look very similar to those in Figure 1 that motivated our analysis. Most of our analysis rests on a sample of individual records for which we have both CPS and UI information. This linked sample includes approximately 12 to 15 thousand individuals each year. We use the adjusted March CPS basic estimation weights already described to create

15

16-state employment totals for both the CPS and the UI data. As discussed in the previous section, the CPS employment totals in Figure 2e-1 are very close to the CPS employment totals in Figure 2d-1. Applying these same weights to the UI data produces employment levels that are, on average, slightly below the actual level of employment in the full-sample UI data (for example, 58.0 million in our sample versus 59.9 million in the full sample in 1996). One likely contributing factor is that UI jobs held by individuals who reside outside of our 16 states are, by construction, not represented in these estimates. Another contributing factor seems likely to be missing or invalid SSNs, arising either because of transcription errors or deliberate misreporting, that cause us to miss valid linkages between our CPS sample and jobs in the UI database. It is very reassuring, however, that the employment trend in our linked-sample UI employment series is strongly similar to that in the full-sample UI data. In sum, all three of the key patterns of the CES-CPS employment trend discrepancy evident in Figure 2a – the faster growth of CES employment between March 1996 and March 2001, the larger employment decline in the establishment data between March 2001 and March 2002 and the divergent movements in the two series between March 2002 and March 2003 – are reproduced in the weighted employment estimates based on data for our linked CPS-UI sample. This gives us important confidence that explorations based on our linked sample microdata can be informative about the reasons for the different behavior of published CES and CPS employment estimates. IV. Micro data Comparisons of Differences in Employment Status and Multiple Job

Holding in the CPS and UI In this section, we present results comparing employment status (i.e., working or not working in an in-scope job) and multiple job holding status at the micro level in the matched CPS-UI data. The results are based on micro data comparisons for individuals in the matched CPS-UI data for the 1996-2003 period. A. Basic Patterns of Discrepancies in Employment Status and Multiple Job Holding We begin with tables that summarize the basic patterns of discrepancy in the categorization of individuals as employed versus not employed in an in-scope job – where an in-scope job is defined as a non-agricultural private-sector wage and salary job, a state government job or a local government job – and, among those who are employed, as having a single in-scope job versus more than one in-scope job across the two data sets. All of the statistics reported in these initial tables were calculated using the pooled micro data for the 1996-2003 period and the adjusted CPS estimation weights. We classify individuals into the categories described in section II using the measurement methodology described in section III. For employment status, we make use of the X1 through X4 categories outlined earlier. X1 workers, for instance, are workers who do not have an in-scope job in either the UI or the CPS. For multiple job holding status, we restrict our attention to X4 workers and in turn classify these workers into the Y1 through Y4 categories. Y1 workers, for instance, are workers who have only one in-scope job both in the CPS and in the UI data in a given quarter.

16

Table 1 presents a simple two-by-two matrix of employment status – whether or not the person has an in-scope job – in the CPS versus the UI data. Overall shares, row shares, and column shares are presented along with standard errors. Over the eight year period covered by our data, about 49 percent of individuals aged 16+ are employed as an in-scope worker in both the CPS and the UI data. Some 37 percent of individuals aged 16+ are not in-scope workers in either the CPS or the UI data (not employed in the UI is defined as not having a wage record). About 3 percent of individuals are in-scope workers in the UI but not in the CPS and about 11 percent are in-scope workers in the CPS but not in the UI. Given the large size of the pooled matched dataset, the standard errors of these estimates are low. The row and column percents are useful to illustrate conditional relationships. For example, about 6 percent of in-scope UI workers are not in-scope CPS workers while about 18 percent of in-scope CPS workers are not in-scope UI workers.12 In Tables 2 and 3, we explore discrepancies in multiple job status. Recall that, because of the difficulty of identifying workers in the CPS who held more than one job during the quarter, we define alternate indicator variables for multiple job status. Table 2 uses the more restrictive classification methodology.13 The at-risk group in Table 2 is individuals who are working as wage and salary workers both in the CPS and in the UI – in other words, the X4 group from Table 1. About 81 percent of such workers have just one job both in the CPS and in the UI, 5 percent have two or more jobs both in the CPS and in the UI, 10 percent have two or more jobs in the UI but only one job in the CPS and 4 percent have two or more jobs in the CPS but only one job in the UI. Although Table 2 is based on fewer observations than Table 1, the standard errors for the estimates reported remain small. Row and column percents are again instructive. About 69 percent of workers with two or more in-scope jobs in the UI have only one in-scope job in the CPS. Conversely, conditional on having two or more in-scope jobs in the CPS, about 45 percent of workers have only one in-scope job in the UI. In both cases, the conditional discrepancy is quite large. Table 3 repeats this exercise using the less restrictive criteria to identify multiple job holding in the CPS.14 By construction, using less restrictive criteria increases the share of those in the CPS who are categorized as multiple job holders. The less restrictive approach also tends to increase the share of those classified as a multiple job holder in the CPS who are not so classified in the UI data. Conditional on having two or more in-scope jobs in the CPS, this less restrictive CPS classification yields 58 percent of workers having only one in-scope job in the UI, as compared to 45 percent with the more restrictive CPS multiple jobholder classification. As it should, however, the less restrictive classification works the opposite way for multiple job 12 One measurement issue that might be important in the asymmetry in the size of the off-diagonal cells is that some CPS workers might be working in a state that is not included among the 16 states for which we have UI wage records. To explore the possible role of such effects, we tabulated the statistics in Table 1 using a restricted sample of three large states (California, Florida and Texas) in which very few residents are employed in another state. In this restricted analysis, we found very similar patterns to those reported in Table 1. For example, we found that 7 percent of in-scope UI workers are not in-scope workers in the CPS. Conversely, we found that 18 percent of in-scope CPS workers are not in-scope UI workers. We also found results very similar to those reported in Tables 2 and 3 for this more restricted sample. 13 As described in section III, the more restrictive classification includes only workers in categories JC1 and/or MJ1 as multiple job holders. 14 The less restrictive classification includes workers in categories JC1, JC2, MJ1 and/or MJ2 as multiple job holders.

17

holders in the UI – using the less restrictive definition, about 54 percent of workers who hold multiple in-scope jobs in the UI data have only have one in-scope job in CPS, as compared to 69 percent with the more restrictive CPS multiple jobholder classification. To summarize, we find substantial discrepancies in both employment status and multiple job holding status between the CPS and the UI at the micro level. Some 18 percent of in-scope CPS workers do not hold an in-scope UI job. About 6 percent of in-scope UI workers are not working in the CPS. Discrepancies are even larger for multiple job holding status. Conditional on being employed in both the CPS and the UI and having multiple jobs in the UI, between 54 and 69 percent are not multiple job holders in the CPS. Conditional on being employed in both the CPS and the UI and having multiple jobs in the CPS, between 45 and 58 percent are not multiple job holders in the UI. B Trends in Levels of Employment in Off-Diagonal Cells. Our interest lies not only with the average rate of agreement or disagreement between employment status measures derived from our two data sources but also with how the number of people for whom the two data sources disagree has varied over time. We turn to that question now. Figure 3 displays the trend both in the number of people estimated to be employed in the UI but not the CPS (X2) and in the number of people estimated to be employed in the CPS but not in the UI (X3). The X2 series is considerably smaller in magnitude and fluctuates only modestly over time. The X3 series is larger in magnitude and considerably more volatile. What matters for the gap between the employer- and household-based employment estimates is the relative trends in X2 and X3 over time. All else the same, any increase in X2 relative to X3 will be associated with growth in UI employment relative to CPS employment; conversely, any decrease in X2 relative to X3 will be associated with a decline in UI employment relative to CPS employment. To the extent that X2 and X3 grow or shrink together, the gap between the UI and CPS employment series will not be affected. Over the period from 1996 to 2001, X2 and X3 fluctuate relative to one another, but not in a consistent fashion. Over the 2001 to 2003 period X3 grew by about 900,000 workers while X2 fell by about 300,000 workers, both movements that would have contributed to the relative increase in CPS employment over this period. The modest decline in X2 is consistent with a shrinking number of marginal (short duration or low earnings) jobs and perhaps also with a slowing rate of job turnover. The larger increase in X3 is consistent with more marked growth in the number of informal or other nonstandard jobs. The combined swing of about 1.1 million jobs is substantial and is an important factor in the shrinking discrepancy between the household and employer job counts over this subperiod. Figure 4 displays the trend in the number of people categorized as holding more than one in-scope job in the UI data but a single in-scope job in the CPS data (Y2) and the trend in the number of people categorized as holding more than one in-scope job in the CPS data but a single job in the UI data (Y3). Recall that there are two different indicators for whether a person holds more than one job in the CPS and thus two different versions of both series. The top panel in the figure displays numbers based on the more restrictive definition and the bottom panel displays numbers based on the less restrictive definition. As expected, there are far fewer people assigned to the Y3 category using the more restrictive indicator, but in both cases the trend line for the Y3 series is relatively flat. The behavior of the Y2 series is more interesting. This series also is

18

affected by the choice of indicator for whether a person has a multiple job in the CPS data – there are fewer people who hold multiple jobs in the UI data but single jobs in the CPS data when using the less restrictive CPS multiple job indicator. Both trend lines, however, show that the number of people in the Y2 category grew markedly between 1996 and 1999 and leveled off thereafter.

Here again what matters for the gap between employer- and household-based employment estimates is whether the difference between Y2 and Y3 is changing over time. The gain in the number of people holding multiple jobs in UI but not in CPS (Y2) from 1996 to 1999 is substantial under both classification methodologies. Y2 grows by about 1.6 million people using the more restrictive method and by about 1.3 million people under the less restrictive method. In contrast, the number of people holding multiple jobs in CPS but not in UI (Y3) grows by about 200,000 under the more restrictive method and falls by about 150,000 under the less restrictive method, both relatively modest changes. Under either method then, the difference between Y2 and Y3 grows substantially over the 1996 to 1999 period — by about 1.4 million people in both cases – contributing significantly to the increase in the UI job count relative to the CPS job count over these years. Interpreted in the light of the framework outlined in Section II, the sizeable increase in Y2 is consistent both with growth in the number of marginal (short duration or low earnings) second jobs and with increases in the job turnover rate. The smaller decrease in Y3 is consistent with a shrinking number of informal second jobs.

Looking at Figures 3 and 4 together suggests that different components of the off-

diagonal elements of the employer and household data play a role in the different subperiods. One of the reasons that the employer job count grew so rapidly relative to the household job count over the 1996 to 1999 period appears to be that the number of employed people holding multiple jobs increased faster in the employer than in the household data. The years 1999 to 2001 are a time period in which there are fluctuations in the off-diagonal elements but with less systematic implications. Then in the downturn from 2001 through 2003, employment status plays a bigger role, with growth especially in the number of individuals identifying themselves as employed in the household data but not in the employer data.

C. Exploring Individual and Job Characteristics for Employment Status Discrepancies In Section II, we hypothesized that discrepancies between employment status as recorded for individuals in the UI data and employment status as recorded in the CPS data reflect the types of jobs held by the affected individuals. More specifically, our working hypotheses are that marginal but formal jobs are more likely to show up in the UI records but not the CPS data (and be counted in the X2 cell), while informal or non-standard employment relationships are more likely to show up as CPS jobs but not as UI jobs (and be counted in the X3 cell). To the extent that these hypotheses are valid, they should have implications for the characteristics of both the workers and the jobs found in the off-diagonal cells. For example, individuals still attending school or individuals who have retired from their career job may earn wage and salary income that shows up in the UI data. When asked whether they are working (or when someone in their household is asked whether they are working), however, this employment activity may not be reported because the individual considers him- or herself to be “a student” or

19

“retired.” We would expect, therefore, that young adults and people near or past retirement age should be more likely than others to be counted in the X2 category. In a related way, whatever the characteristics of the job incumbent, we would expect short term or low-earnings jobs that are present in administrative data to be more likely than other jobs not to be reported in a household survey, for the reason that the people in these jobs seem quite likely not to consider working to be their primary activity. Those holding this sort of job also should be more likely than others to show up in the X2 category. Analogously, but with opposite consequences, individuals who have earned income from an “off-the-books” or informal job or from working as a contractor should be classified as self-employed but may regard themselves as wage and salary workers since they report to an employer and perhaps work alongside other wage and salary workers. Such workers would be counted in the X3 category. Those more likely to appear in the X3 category might include, for example, workers with low education who may be more likely to work off the books, highly educated people who may be more likely to work as contractors, or anyone in an industry and occupation in which there are a large number of self-employed workers, suggesting the potential for confusion in the reporting of employment status. To the extent that informal employment also tends to be less stable or less intensive than more formal employment, measures of job continuity and hours or work also may help to explain individuals’ presence in the X3 cell.

The simplest way to identify the factors that affect the probabilities of being found in the off-diagonal X2 or X3 cells is to fit linear probability models containing an appropriate set of explanatory variables.15 We focus our attention on models that condition either on being employed in the UI records or on being employed in the CPS data. Table 4 presents results for models that examine the factors affecting the probability that an individual classified as an in-scope worker in the UI data holds no in-scope job in the CPS data. The models reported in Table 5 examine the other off-diagonal conditional probability – namely the factors affecting the probability that a person classified as an in-scope worker in the CPS data holds no in-scope job in the UI data. Results are reported both for models that include only demographic characteristics and models to which measures of job characteristics have been added. We interpret both sets of characteristic variables with reference to whether the characteristic seems likely to be associated with a job being either marginal (in the case of the X2 models) or informal or otherwise non-standard (in the case of the X3 models). In addition to the demographic and job characteristic variables, all of the models also include year dummies.

The same demographic controls are used in both tables – age classes (with 35-54 year

olds the omitted group), education classes (less than high school, high school graduate, some college, college graduate or completed post-graduate work, with high school graduates the omitted group), gender (female the omitted group), race (black, white or other, with white the omitted group), an indicator variable for whether the individual is the CPS respondent rather than someone else in the household, an indicator for whether the individual is married, and an indicator for whether the individual is foreign born.

15 In the next version of the paper, the linear probability models will be replaced with similarly specified probit models.

20

Among the job characteristics that might influence whether a person who has a UI wage record also is reported as employed in the CPS data (Table 4), we include a measure of whether the individual held any long UI job during the quarter (defined as a job that began before the start of the quarter or continued beyond the end of the quarter); an indicator for whether the individual held two or more UI jobs during the quarter; and indicators for the level of earnings on the individual’s primary UI job (earnings during the quarter of under $1,000, $1,000 to $2,500, $2,500 to $12,500, $12,500 to $25,000 or more than $25,000). Our expectation is that the stable UI jobs with higher earnings should be more likely to also be present in the CPS (i.e., to have negative coefficients in the model reported in Table 4). We have somewhat less good information about jobs reported in the CPS data than about jobs reported in the UI data; perhaps most importantly, we have no usable information about earnings on the CPS jobs.16 Among the job characteristics that might influence whether an individual reported as working in the CPS also has a UI job (Table 5), we include an indicator for whether the individual had a work discontinuity (i.e., was not employed at the time of one or two of the monthly CPS interviews during the quarter); an indicator for whether the respondent works in an occupation and industry with a high percentage of self-employed workers; and an indicator for whether any of the CPS jobs held during the quarter were full time jobs. Our prior is that workers with work interruptions, workers who are more likely in fact to be working as contractors rather than in wage-and-salary positions, and workers in part-time positions are more likely to hold informal or non-standard jobs and as such less likely to have UI wage records.

The first column of Table 4 reports results for models that include only demographic

factors. We find that, among those who hold an in-scope UI job, those who are very young or past the middle working years, less educated, female, black or foreign born all are less likely to hold an in-scope CPS job (i.e., more likely to be in the X2 category). Among the demographic variables, being age 65 or older has the largest effect – all else the same, being in this age group raises the probability of being in the X2 category by 14 percentage points. We interpret being in the X2 category – having a job that shows up in the administrative data but no job in the CPS – to indicate having a job that, at least from the holder’s perspective, is marginal. The demographic patterns we observe seem largely consistent with this interpretation. Older workers, for example, may not consider themselves regular wage and salary workers (e.g., they have retired from their main career job) but may nevertheless generate some earnings from a wage and salary job. When we add job characteristics to the model, some of these patterns change in ways that seem sensible given the relationship between the demographic and job characteristics. For example, we find that very low earnings workers are much more likely to be X2 workers. This is consistent with the notion that those holding low earning jobs are more likely to consider something other than paid work to be their primary activity. We also find that workers with long-lasting jobs and with more than one UI job are much less likely to be X2 workers. Controlling for job characteristics, however, weakens the effect of age and having low education on the probability of being an X2 worker. It makes sense that part of what the controls for these characteristics were picking up was the tendency to hold a job with low earnings or short duration.

Turning to Table 5, among in-scope CPS workers, controlling only for demographic

characteristics, those over age 65 are especially likely not to have an in-scope UI job. All else 16 Earnings data are collected only for the main job held in the outgoing rotation months.

21

the same, being in this age group raises the probability of being in the X3 category by 14 percentage points. Being male, non-black, or foreign born and having either a low or a very high level of education have smaller effects in the same direction. Most of these effects are robust to the addition of job characteristic variables to the model. The job characteristic variables included in the second column of the table add substantial additional explanatory power. CPS workers who have a work discontinuity (did not work all three months) or are in an industry and occupation with a high proportion of contractors have a substantially higher probability of being absent from the UI wage records. CPS workers who have a job during the quarter that is full time in at least one month are less likely to be absent from the UI wage records. As discussed above, we expect this off-diagonal to be populated by workers who have some type of informal job or non-standard employment relationship. The estimated demographic and job characteristic coefficients are largely consistent with this hypothesis. For example, older workers, less educated workers and very highly educated workers – all less likely to be found in the UI data conditional on reporting a job in the CPS – very plausibly are more likely to have either an informal job or a non-standard employment relationship. In terms of job characteristics, stable, full time jobs are less likely to be informal or non-standard jobs, and CPS workers holding such positions are more likely to be found in the UI records.

D. Exploring Individual and Job Characteristics for Multiple Job Holding Status Discrepancies

We now turn to an examination of the factors associated with discrepancies in multiple job holding status. As with our examination of employment status discrepancies, we focus on conditional models. In Tables 6 and 7, we report the results of models that seek to identify the factors impacting the probability that those with multiple UI jobs have only a single CPS job. In Tables 8 and 9, we report analogous results for the factors impacting the probability that those with multiple CPS jobs have only a single UI job. Tables 6 and 8 use the more restrictive classification of multiple job holding in the CPS; Tables 7 and 9 the more expansive version.

We do not have clear priors about the demographic characteristics we should expect to be

associated with holding either marginal or informal second jobs. Any demographic characteristic associated with higher job turnover, however, might result in people with that characteristic being more likely to show up in the Y2 cell. It also seems clear that second jobs with certain characteristics should be more likely to show up in the off-diagonal cells of the multiple job cross-tabulation. Thinking first about the Y2 off-diagonal, we would expect long-lasting and high-paying UI second jobs not to be considered marginal by those who hold them and thus to be more likely to be reported in the CPS. Similarly, thinking about the Y3 off-diagonal, we would expect that more stable CPS jobs and CPS jobs with longer hours should be more likely to be recorded in the UI data, since such jobs seem less likely to be off the books.

The first column of Table 6 reports the results of models fit for the conditional Y2 models

that contain demographic factors only. For those with multiple UI jobs, the likelihood of having only a single CPS job (the likelihood of being a Y2 worker) is lower for more educated workers and for married workers, and higher for black workers and foreign born workers. One possible interpretation of these findings is that they reflect relative likelihoods of job turnover. These demographic effects are generally robust to the addition of job characteristics to the model and,

22

in most cases, adding the job characteristic variables actually strengthens the effect of the demographic variables. In the model in the second column of Table 6, the job characteristic variables refer to the characteristics of the second UI job(s). Several of the job characteristic variables have large effects. Conditional on reporting multiple UI jobs, workers with any long second UI job are much less likely to be Y2 workers while workers with high second UI job earnings are much more likely to be Y2 workers. The effect of earnings is non-monotonic – workers with earnings on the second job in the range $1-$2.5K have the lowest probability of being a Y2 worker – but we did not expect the positive effect associated with the highest earnings levels.

In Table 7, we repeat the same models as in Table 6, but using the more expansive definition of multiple job holding in the CPS. The only noteworthy changes are that, in the Table 7 models, younger workers and workers with 3 or more UI jobs are less likely to be in the Y2 group.

Taken together, Tables 6 and 7 suggest that both demographic and job characteristics significantly impact the likelihood that a UI multiple job holder reports just one job in the CPS. It does not appear that a simple “one size fits all” story emerges but there are some interesting patterns. A surprising finding is that some workers with very high earnings on the second job do not consider the work they did to earn this money as a second job per se when reporting their employment to the CPS interviewer. It might be, for example, that individuals in some professions earn significant supplemental earnings that are not regarded by the worker as constituting a second job. This may be the case for those who work as actors, plumbers, electricians, or even doctors or lawyers who regard themselves as having one job but in fact have income that derives from multiple sources.

Tables 8 and 9 report results for models that estimate the probability that CPS multiple job holders hold only one UI job (the likelihood of being a “Y3” worker). The first column of Table 8 reports results based on a model that includes only demographic factors. Personal characteristics that raise the probability of being a CPS multiple job holder but a UI single job holder include being highly educated, age 55 or older (though the effect is not statistically significant for those age 65 and older), or male. The addition of job characteristics has little effect on the demographic variable coefficients, but raises the explanatory power of the model. Multiple job holders who have two simultaneous jobs in one of the months, simultaneous jobs in all three months, and work at least 16 hours a week on the second job(s) are much less likely to be Y3 workers. For both demographic and job characteristics, these same patterns hold when the more expansive definition of CPS multiple job holders is applied. These results are reported in Table 9.

The strongest findings to emerge from Tables 8 and 9 are those related to job characteristics. Interpreted in the context of the framework we have outlined, it would appear that second jobs recorded in the CPS data that are of short duration and involve relatively few hours are more likely to be “off the books” positions that are absent from the UI records. There is also some evidence that older and highly educated workers are more likely to hold “off the books” jobs. As with the X3 workers, it could be that the second CPS jobs reported by the Y3 workers reflect outside work that they do on a contract basis and for which they receive 1099

23

income, but that they or someone else in their household report as wage and salary work when queried by the CPS interviewer. E. Time Series Implications of Changing Composition of Workers and Jobs

The estimated effects from the linear probability models provide a great deal of information about the demographic and job characteristic correlates of discrepancies between employer- and household-reported employment and multiple-job-holding status. The patterns are largely consistent with our working hypotheses – factors that we would have expected to be correlated with a job being marginal are predictive of being found in the X2 and Y2 categories (people with jobs that are observed in the UI data but not the CPS data) and factors that we would have expected to be correlated with a job being off-the-books or non-standard in some other way are predictive of being found in the X3 and Y3 categories (people with jobs that are observed in the CPS data but not the UI data). The patterns documented in the linear probability models are of interest in their own right and may well have important implications for the analysis of labor market behavior at the individual level. Our primary interest, however, lies with whether the estimated relationships between demographic and job characteristics and the employment and multiple-job-holding status outcomes can help us to understand the aggregate time series behavior of the employer-based and household-based job count estimates reported in Figures 1 and 2. The differing behavior in the aggregate job series must, in an accounting sense, reflect the differing behavior of X2 and X3 and/or the differing behavior of Y2 and Y3, and we turn now to the question of whether changes over time in worker characteristics and/or job characteristics can explain the movements over time in the number of people in these off-diagonal cells.

To explore this question, we use information on the composition of workers and jobs in each year together with the estimated coefficients from the linear probability models to simulate aggregate values for X2, X3, Y2 and Y3. The simulated shares of employment falling into the various X and Y categories are computed using the intercept term and the average of the year dummy coefficients from the relevant linear probability model plus the vector product of the characteristic coefficient estimates and the annual characteristic values. To simulate the number of workers in the different X and Y categories, we multiply these simulated shares by the number of UI or CPS workers, as appropriate, or by the number of UI or CPS multiple job holders. Our simulated value for X2, for example, equals the predicted probability that a UI worker is not a CPS worker times the number of UI workers in the year in question. In each case, the coefficient estimates employed are taken from the model that includes both demographic characteristics and job characteristics. Another point to note is that, rather than using our linked sample to compute the shares of worker and job types by year, we use the possible largest dataset for each characteristic variable. Specifically, for the demographic characteristics and CPS job characteristics we use all of the records for people in our 16 states who completed CPS interviews in January, February and March in the year in question, rather than the more restricted CPS sample for which a PIK was available. For the UI job characteristics, we use all of the UI wage records available for our 16 states. The reason for using these larger databases is to reduce the sampling variation in the simulated X and Y values; one consequence is that the average of

24

the simulated values will not necessarily equal the average of the predicted values, as would be the case if we had used sample-based characteristic values.

Figure 5 displays the actual trends in X2 and X3 previously shown in Figure 3 together with the predicted values calculated as just described. The simulated and actual values need not behave especially similarly – it would be entirely possible for the factors that explain the variation in X2 and X3 over time to have been captured by the year dummies in the models rather than by the demographic and job characteristic variables. The simulated values in X2 and X3 are somewhat smoother than the actual values, which not especially surprising given that the simulation process eliminates random variations, but, importantly for our current analysis, clearly display the widening gap between X3 and X2 that we commented upon in our discussion of Figure 5. This implies that the widening gap is a function of changes in the workforce and job characteristics that we have argued are indicative of marginal and off-the-books or other non-standard employment.

One obvious question at this point is whether the movements in the simulated series are driven more by changes in demographic characteristics or changes in job characteristics. This question is explored in Figure 6, which rests on a closely related but different metric than Figure 5 – specifically the predicted share of X2 workers (panel A) and X3 workers (panel B). In Figure 6, we show simulated shares using all of the variation in worker and job characteristics over time, and then simulate separately the effects of changes in demographic characteristics alone and then the effects of changes in job characteristics alone, holding the values of the other set of characteristics constant at their average values over the whole sample period. For example, in examining how variation in the UI job characteristic values affect the share of UI workers who are not CPS workers (i.e., the predicted share of UI workers in the X2 category), the demographic composition of the workforce is kept constant at the average values over the 1996-2003 data.

Figure 6a shows that the relatively flat behavior of the predicted share of UI workers in the X2 cell over the 2001 and 2003 period is driven mostly by changes in the characteristics of the jobs held by these workers that offset what would have been an increase in X2 based upon changes in demographic characteristics alone. In Figure 6b, it is apparent that both job and demographic characteristics contributed to the increase in X3 from 2001 to 2002 and then that only the demographic characteristics contributed to the increase in X3 from 2002 to 2003, offsetting what would have been a decline if only job characteristics had been operative. The systematic increase in X3 dating from 1999 is also more readily apparent in Figure 6b than was the case in Figure 5.

The next several figures display results from our simulations of the aggregate time series behavior of Y2 and Y3 based on changes in the demographic composition of the workforce and the changes in the characteristics of second jobs as measured in the UI and CPS data. Figure 7 shows the actual vs. predicted number of workers in the Y2 and Y3 groups using the more restrictive classification of multiple jobs in the CPS. The increase in the simulated Y2 series closely mimics the increase in the actual Y2 series over the 1996 to 1999 period that we emphasized in our discussion of Figure 4. As noted there, the actual Y3 series is relatively flat over this period, and the same is true of the simulated series.

25

To provide insight into whether it is the demographic or the job characteristics that play a bigger role in the simulated movements in the Y series, Figure 8 shows the predicted Y2 and Y3 shares – the share of UI multiple job holders with a single CPS job and the share of CPS multiple job holders with a single UI job, respectively -- and the shares predicted with only demographic characteristics or only job characteristics are allowed to vary. It is apparent that most of the movement in the predicted shares of both Y2 and Y3 are associated with changes in the composition of the second jobs reported either in the UI or in the CPS database. Even taking both sets of characteristics into account, the predicted share of UI multiple job holders who are not multiple job holders in the CPS (the Y2 share) shown in Figure 8a exhibits only a mild increase between 1996 and 1999. Even this mild increase, however, yields a substantial gain in the number of Y2 workers over this period – as can be seen in Figure 7 – when combined with the rapidly growing number of UI multiple job holders. In Figure 8b, the predicted share of Y3 workers increased notably over the 1999 to 2001 period, due mostly to changes in the mix of characteristics of the second jobs of CPS multiple job holders.

Figures 9 and 10 show the analogous patterns for Y2 and Y3 workers using the more expansive definition of multiple jobs in the CPS. The increase in Y2 over the 1996 to 1999 period is also very apparent both in the actual and in the simulated data. Figure 10 shows that the predicted share of UI multiple job holders who hold only a single CPS job (the Y2 share) exhibits a systematic increase over this period of time and that it is the job characteristics that account for most of this increase.

One interesting feature of the simulations that can be seen in both Figure 8 and Figure 10 is that the predicted share of UI multiple job holders with only a single CPS job (the Y2 share) increased over much of the 1996 to 2003 period even though the most notable increase in the total number of Y2 workers is over the 1996 to 1999 period. The differences in the pattern for the shares and that for the numbers of workers is attributable to the rapid growth in the number of UI multiple job holders over the 1996 to 1999 period. Over the 1999 to 2003 period the number of UI multiple job holders actually declined but, because of the rising share of UI multiple job holders with only a single CPS job, the number of Y2 workers stayed about the same (with some year to year volatility). Put differently, over the 1996 to 1999 period the rising share and the rising number of UI multiple job holders were working in the same direction while over the 1999 to 2003 period, the rising share worked in the opposite direction from the falling number of UI multiple job holders. V. Concluding Remarks

Aggregate employment statistics from the CPS household survey and the CES payroll survey track each other reasonably well over long periods of time in spite of conceptual and scope differences. Even after making adjustments for conceptual and scope differences, however, there are shorter-term discrepancies between the two series that have a clear cyclical pattern. These discrepancies have been actively studied but researchers have had limited success in reconciling them. In this paper, we explore possible sources of these discrepancies using a person-level data set that contains both CPS and UI wage records data for the same set of individuals.

26

Our linked data permit us to identify individual workers who are employed in the

household data but not in the employer data, and vice versa. In like fashion, the linked data permits us to identify workers in the household data who have multiple jobs but only one job in the employer data, and vice versa. By focusing on the people in these "off-diagonal" cells, we are able to quantify in an accounting fashion the sources of the discrepancies in the aggregate time series. We find that the more rapid rise in employment in the employer data over the late 1990s is due primarily to growth in the number of multiple job holders in the employer (UI) data who report only one job in the household (CPS) data. In contrast, we find that the sharper decline in employment in the employer data relative to the household data after 2001 is associated with an increase in the number of CPS workers who cannot be found in the employer (UI) records.

Motivated by these interesting time trends for the "off-diagonal" elements, we explore

the role of demographic and job characteristics within a conceptual framework that suggests that UI jobs may not show up in the CPS when the job is in some way marginal (short lived, low earnings, not the main activity of the individual) and that CPS jobs may not show up in the UI wage records data if they are “off the books” or involve some other form of non-standard employment relationship. In models that include demographic and job characteristics that it is plausible to think are associated with a position being a marginal or an “off-the-books” job, we find evidence that supports these hypotheses. For example, we find that UI workers with short duration or low earnings jobs in the reference quarter are much less likely to be identified as employed in the CPS data. We also find that CPS workers with full time, stable jobs and in industry/occupation cells containing fewer self-employed people are more likely to have UI wage records and thus to be identified as employed in the UI data.

Using the micro based estimates of the factors influencing the likelihood of being an off-

diagonal worker (or being off-diagonal in terms of multiple jobs), we explore the implications of changes in the demographic and job characteristics of the workforce for the number of people for whom the employment status information reported in the UI data does not match that reported in the CPS data. We find that the changes in the demographic and job characteristics of the workforce over time do a reasonably good job of predicting values for the size of the off-diagonal cells that match the actual numbers of people in those same cells. Further, both job characteristics and demographic characteristics contribute to these predictions. The relative increase in the number of UI multiple job holders in the 1996-1999 period who are not multiple job holders in the CPS, however, is accounted for primarily by changes in the characteristics of the second jobs held by the UI multiple job holders rather than by changes in the demographic characteristics of the multiple job holders.

Taken as a whole, these findings are provocative. Even ignoring the time series dimension of the off-diagonal series we have examined, the evidence of particular problems with the reporting of employment for certain types of people and certain types of jobs is something that should be factored into analyses of labor market activity based on the data commonly used for this purpose. Further, consistent with our goal in embarking upon this research, the findings offer useful insights into the sources of the CES-CPS employment discrepancy. Some of the questions that our analysis leaves unanswered reflect limitations in the data with which we have

27

worked. It is possible that expanding the number of years of data available for analysis will permit more definitive conclusions at some future point. In the course of carrying out this analysis, it became apparent to us that better information about the characteristics of the jobs held by CPS respondents and about job changes they experience would have considerable value. That, too, is something to consider for the future.

28

References Abowd, John and Martha Stinson. 2003. “Estimating Measurement Error in SIPP Annual Job

Earnings: A Comparison of Census Survey and Administrative Data.” Unpublished working paper. September.

Abowd, John, Bryce Stephens, Lars Vilhuber, Fredrik Andersson, Kevin McKinney, Marc

Roemer and Simon Woodcock. 2006. “The LEHD Infrastructure Files and the Creation of the Quarterly Workforce Indicators,” LEHD Technical Working Paper No. TP-2006-01. January.

Bjelland, Melissa, John Haltiwanger, Kristin Sandusky and James Spletzer. 2006. “Reconciling

Household and Administrative Measures of Self-Employment and Entrepreneurship.” Unpublished working paper. April.

Bowler, Mary and Teresa L. Morisi. 2006. “Understanding the employment measures from the

CPS and CES survey.” Monthly Labor Review, February, 23-38. Juhn, Chinhui and Simon Potter. 1999. “Explaining the Recent Divergence in Payroll and

Household Employment Growth.” Federal Reserve Bank of New York Current Issues in Economics and Finance, December, 1-6.

Nardone, Thomas, Mary Bowler, Jurgen Kropf, Katie Kirkland and Signe Wetrogan. 2003.

“Examining the Discrepancy in Employment Growth Between the CPS and the CES.” Paper prepared for presentation to the Federal Economic Statistics Advisory Committee. October.

Roemer, Marc. 2002. “Using Administrative Earnings Records to Assess Wage Data Quality in

the March Current Population Survey and the Survey of Income and Program Participation.” LEHD Technical Paper #TP-2002-22. November.

U.S. Bureau of Labor Statistics. Undated. “Employment from the BLS household and payroll

surveys: summary of recent trends,” updated monthly and posted at http://www.bls.gov/web/ces_cps_trends.pdf (accessed November 2007).

U.S. Bureau of Labor Statistics. 2004. “Effects of Job Changing on Payroll Survey

Employment Trends.” August. Posted at http://www.bls.gov/ces/cesjobch.pdf (accessed November 2007).

U.S. Bureau of Labor Statistics. 2005. “Report of the FESAC Subcommittee on the

Discrepancy in CPS-CES.” December. Posted at http://www.bls.gov/bls/fesacp2120905.pdf (accessed November 2007).

U.S. Bureau of Labor Statistics. 2007. “Labor force and employment estimates smoothed for

population adjustments, 1990-2006.” February. Posted at http://www.bls.gov/cps/cpspopsm.pdf (accessed November 2007).

29

Figure 1a: Household and Payroll Survey Employment, Seasonally Adjusted, 1994-2006

30

Figure 1b: Ratio of Establishment Survey Employment to Household Survey Nonagricultural Wage and Salary Employment, 1948-2004

0.98

0.99

1.00

1.01

1.02

1.03

1.04

1.05

1.06

1948 1952 1956 1960 1964 1968 1972 1976 1980 1984 1988 1992 1996 2000 20040.98

0.99

1.00

1.01

1.02

1.03

1.04

1.05

1.06

Source: Bowler and Morisi, 2006

31

Figure 2: Effects of Sample Restrictions and Employment Concepts on CPS, CES, and UI Employment Trends, 1996-2003

Figure 2a-1: CPS - CES Employment TrendsNational, All Sectors, March Employment, SA

115,000

125,000

135,000

145,000

1996 1997 1998 1999 2000 2001 2002 2003CPS CES

Figure 2a-2: CPS - CES Employment TrendsNational, All Sectors, March Employment, SA

100

105

110

115

1996 1997 1998 1999 2000 2001 2002 2003CPS CES

Figure 2b-1: CPS - CES Employment Trends

16 States, All Sectors, March Employment, NSA

55,000

60,000

65,000

70,000

1996 1997 1998 1999 2000 2001 2002 2003CPS CES

Figure 2b-2: CPS - CES Employment Trends16 States, All Sectors, March Employment, NSA

100

105

110

115

1996 1997 1998 1999 2000 2001 2002 2003CPS CES

Figure 2c-1: CPS - CES Employment Trends

16 States, PSL, March Employment, NSA

50,000

55,000

60,000

65,000

1996 1997 1998 1999 2000 2001 2002 2003CPS CES

Figure 2c-2: CPS - CES Employment Trends16 States, PSL, March Employment, NSA

100

105

110

115

1996 1997 1998 1999 2000 2001 2002 2003CPS CES

Figure 2d-1: CPS - UI Employment Trends

16 States, PSL, 1st Quarter Employment, NSA

55,000

60,000

65,000

70,000

1996 1997 1998 1999 2000 2001 2002 2003CPS UI

Figure 2d-2: CPS - UI Employment Trends16 States, PSL, 1st Quarter Employment, NSA

100

106

112

118

1996 1997 1998 1999 2000 2001 2002 2003CPS UI

Figure 2e-1: CPS - UI Employment Trends

16 States, PSL, 1st Quarter Employment, NSA

55,000

60,000

65,000

70,000

1996 1997 1998 1999 2000 2001 2002 2003CPS (w eighted) UI (w eighted)

Figure 2e-2: CPS - UI Employment Trends16 States, PSL, 1st Quarter Employment, NSA

100

106

112

118

1996 1997 1998 1999 2000 2001 2002 2003CPS (w eighted) UI (w eighted)

32

Figure 3: Estimated Number of People in Off-Diagonal In-Scope Employment Cells, 1996-2003 (in millions)

0

3

6

9

12

1996 1997 1998 1999 2000 2001 2002 2003

X2: Employed in UI, not in CPS X3: Employed in CPS, not in UI

33

Figure 4a: Estimated Number of People in Off-Diagonal Multiple Job Status Cells, More Restrictive CPS Classification, 1996-2003 (in millions)

0

2

4

6

1996 1997 1998 1999 2000 2001 2002 2003

Y2: Multiple Job Holder in UI, 1 Job in CPSY3: Multiple Job Holder in CPS, 1 Job in UI

Figure 4b: Estimated Number of People in Off-Diagonal Multiple Job Status Cells,

Less Restrictive CPS Classification, 1996-2003 (in millions)

0

2

4

6

1996 1997 1998 1999 2000 2001 2002 2003

Y2: Multiple Job Holder in UI, 1 Job in CPSY3: Multiple Job Holder in CPS, 1 Job in UI

34

Figure 5: Actual and Predicted X2 (Employed in UI, Not Employed in CPS) and X3 (Employed in CPS, Not Employed in UI), 1996-2003 (in millions)

0

3

6

9

12

1996 1997 1998 1999 2000 2001 2002 2003

X2 (Actual) X2 (Predicted) X3 (Actual) X3 (Predicted)

35

Figure 6a: Predicted Shares of UI Workers Not Found in CPS (X2), 1996-2003

0.060

0.065

0.070

0.075

1996 1997 1998 1999 2000 2001 2002 2003

Demographic Characteristics Job CharacteristicsAll Characteristics

Figure 6b: Predicted Shares of CPS Workers Not Found in UI (X3), 1996-2003

0.170

0.175

0.180

0.185

1996 1997 1998 1999 2000 2001 2002 2003

Demographic Characteristics Job CharacteristicsAll Characteristics

36

Figure 7: Actual and Predicted Y2 (Multiple Job Holder in UI, Single Job in CPS) and Y3 (Multiple Job Holder in CPS, Single Job in UI), More Restrictive CPS Classification, 1996-2003 (in millions)

0

2

4

6

1996 1997 1998 1999 2000 2001 2002 2003

Y2 (Actual) Y2 (Predicted) Y3 (Actual) Y3 (Predicted)

37

Figure 8a: Predicted Shares of UI Multiple Job Holders with Single Job in CPS (Y2), More Restrictive CPS Classification, 1996-2003

0.64

0.65

0.66

0.67

1996 1997 1998 1999 2000 2001 2002 2003

Demographic Characteristics Job CharacteristicsAll Characteristics

Figure 8b: Predicted Shares of CPS Multiple Job Holders with Single Job in UI (Y3),

More Restrictive CPS Classification, 1996-2003

0.44

0.45

0.46

0.47

0.48

1996 1997 1998 1999 2000 2001 2002 2003

Demographic Characteristics Job CharacteristicsAll Characteristics

38

Figure 9: Actual and Predicted Y2 (Multiple Job Holder in UI, Single Job in CPS) and Y3 (Multiple Job Holder in CPS, Single Job in UI), Less Restrictive CPS Classification, 1996-2003 (in millions)

0

2

4

6

1996 1997 1998 1999 2000 2001 2002 2003

Y2 (Actual) Y2 (Predicted) Y3 (Actual) Y3 (Predicted)

39

Figure 10a: Predicted Shares of UI Multiple Job Holders with Single Job in CPS (Y2), Less Restrictive CPS Classification Definition, 1996-2003

0.50

0.51

0.52

0.53

0.54

1996 1997 1998 1999 2000 2001 2002 2003

Demographic Characteristics Job CharacteristicsAll Characteristics

Figure 10b: Predicted Shares of CPS Multiple Job Holders with Single Job in UI (Y3),

Less Restrictive CPS Classification, 1996-2003

0.61

0.62

0.63

0.64

0.65

1996 1997 1998 1999 2000 2001 2002 2003

Demographic Characteristics Job CharacteristicsAll Characteristics

Table 1: Discrepancies in Employment Status Between CPS and UI Data

Not In-Scope Worker in UI

In-Scope Worker in UI

Not In-Scope Worker in CPSOverall Share 0.371 0.034

(0.001) (0.000)Row Share 0.917 0.083

(0.001) (0.001)Column Share 0.779 0.064

(0.001) (0.001)In-Scope Worker in CPS

Overall Share 0.105 0.491(0.000) (0.001)

Row Share 0.176 0.824(0.001) (0.001)

Column Share 0.221 0.936(0.001) (0.001)

Note: Numbers shown are weighted shares of the CPS-UI overlap sample described in the text. Standard errors are shown in parentheses.The upper left quadrant corresponds to the X1 group, the upper right quadrant tothe X2 group, the lower left quadrant to the X3 group, and the lower right tothe X4 group.

Table 2: Discrepancies in Multiple Job Status Between CPS and UI Data(More Restrictive CPS Classification)

One Two Plus

Number of In-Scope Jobs in CPSOne

Overall Share 0.813 0.104(0.001) (0.001)

Row Share 0.887 0.113(0.001) (0.001)

Column Share 0.956 0.692(0.000) (0.005)

Two PlusOverall Share 0.037 0.046

(0.000) (0.000)Row Share 0.446 0.554

(0.007) (0.007)Column Share 0.044 0.308

(0.000) (0.005)

Note: Numbers shown are weighted shares of persons in the CPS-UI overlap sample described in the text who both sources agree have an in-scope job. Standard errors are shown in parentheses. The upper left quadrant corresponds to the Y1 group, the upper right quadrant to the Y2 group,the lower left quadrant to the Y3 group, and the lower right to the Y4 group.

Number of In-Scope Jobs in UI

Table 3: Discrepancies in Multiple Job Status Between CPS and UI Data(Less Restrictive CPS Classification)

One Two Plus

Number of In-Scope Jobs in CPSOne

Overall Share 0.754 0.081(0.001) (0.001)

Row Share 0.903 0.097(0.001) (0.001)

Column Share 0.886 0.539(0.001) (0.005)

Two PlusOverall Share 0.097 0.069

(0.001) (0.001)Row Share 0.583 0.417

(0.005) (0.005)Column Share 0.114 0.461

(0.001) (0.005)

Note: Numbers shown are weighted shares of persons in the CPS-UI overlap sample described in the text who both sources agree have an in-scope job. Standard errors are shown in parentheses. The upper left quadrant corresponds to the Y1 group, the upper right quadrant to the Y2 group,the lower left quadrant to the Y3 group, and the lower right to the Y4 group.

Number of In-Scope Jobs in UI

Age 16 to 24 0.0511 ** -0.0205 **(0.0033) (0.0032)

Age 25 to 34 0.0026 -0.0047(0.0025) (0.0024)

Age 55 to 64 0.0257 ** 0.0190 **(0.0036) (0.0034)

Age 65 plus 0.1442 ** 0.0879 **(0.0066) (0.0062)

Less than High School 0.0474 ** 0.0066 *(0.0034) (0.0032)

Some College -0.0028 -0.0003(0.0026) (0.0024)

College Graduate -0.0107 ** 0.0017(0.0030) (0.0029)

More than College -0.0003 0.0099 *(0.0041) (0.0040)

Black 0.0277 ** 0.0215 **(0.0035) (0.0033)

Other Non-white 0.0161 ** 0.0088 *(0.0047) (0.0043)

Male -0.0088 ** 0.0055 **(0.0020) (0.0019)

Non-proxy Interview 0.0044 0.0063 **(0.0024) (0.0022)

Married -0.0012 -0.0094 **(0.0024) (0.0022)

Foreign Born 0.0149 ** 0.0259 **(0.0031) (0.0029)

Any Long Jobs -- -0.2013 **(0.0061)

Two or More UI jobs -- -0.0175 **(0.0027)

Qtr UI earn < $1K -- 0.2803 **(0.0036)

$1K <= Qtr UI earn < $2.5K -- 0.0571 **(0.0031)

$12.5K <= Qtr UI Earn < $25K -- -0.0118 **(0.0031)

Qtr UI earn >= $25K -- -0.0001(0.0058)

Observations 56,027 56,027R-squared .024 .160

* significant at 5% level; ** significant at 1% level

Table 4: Effects of Person and Job Characteristics on the Probability that a UI Worker is not a CPS Worker (X2)

Note: Coefficients obtained from linear probability regressions using pooled data for all years 1996 to 2003 for respondents aged 16 and older. Year dummies included in both models. Standard errors in parentheses.

Age 16 to 24 0.0043 -0.0258 **(0.0050) (0.0051)

Age 25 to 34 -0.0167 ** -0.0201 **(0.0037) (0.0037)

Age 55 to 64 0.0336 ** 0.0285 **(0.0053) (0.0053)

Age 65 plus 0.1423 ** 0.1053 **(0.0096) (0.0096)

Less than High School 0.0345 ** 0.0179 **(0.0051) (0.0051)

Some College 0.0052 0.0055(0.0039) (0.0038)

College Graduate 0.0119 ** 0.0166 **(0.0044) (0.0044)

More than College 0.0521 ** 0.0547 **(0.0059) (0.0059)

Black -0.0164 ** -0.0188 **(0.0053) (0.0053)

Other Non-white 0.0124 0.0064(0.0068) (0.0068)

Male 0.0170 ** 0.0250 **(0.0030) (0.0030)

Non-proxy Interview -0.0061 -0.0035(0.0035) (0.0034)

Married -0.0040 -0.0075 *(0.0035) (0.0034)

Foreign Born 0.0360 ** 0.0395 **(0.0046) (0.0045)

Work Discontinuity -- 0.1520 **(0.0050)

Probability of Being a Contractor -- 0.0909 **(0.0079)

Any Full Time Jobs -- -0.0664 **(0.0051)

Observations 63,901 63,901R-squared .010 .027

* significant at 5% level; ** significant at 1% level

Table 5: Effects of Person and Job Characteristics on the Probability that a CPS Worker is not a UI Worker (X3)

Note: Coefficients obtained from linear probability regressions using pooled data for all years 1996 to 2003 for respondents aged 16 and older. Year dummies included in both models. Standard errors in parentheses.

Age 16 to 24 -0.0063 0.0023(0.0161) (0.0161)

Age 25 to 34 0.0176 0.0313 *(0.0133) (0.0131)

Age 55 to 64 0.0269 0.0350(0.0223) (0.0219)

Age 65 plus 0.0316 0.0621(0.0492) (0.0483)

Less than High School 0.0143 0.0125(0.0183) (0.0179)

Some College -0.0482 ** -0.0508 **(0.0137) (0.0134)

College Graduate -0.0344 * -0.0595 **(0.0161) (0.016)

More than College -0.0288 -0.0822 **(0.0224) (0.0225)

Black 0.0857 ** 0.0972 **(0.0160) (0.0156)

Other Non-white 0.0044 0.0027(0.0238) (0.0233)

Male 0.0171 0.0069(0.0108) (0.0107)

Non-proxy Interview -0.0186 -0.0255 *(0.0125) (0.0122)

Married -0.0383 ** -0.0278 *(0.0122) (0.0120)

Foreign Born 0.0487 ** 0.0574 **(0.0164) (0.0161)

Any Long 2nd Jobs -- -0.1103 **(0.0144)

Three or More UI jobs -- -0.0016(0.0115)

Qtr UI earn < $1K (2nd job) -- -0.0136(0.0139)

$1K <= Qtr UI earn < $2.5K (2nd job) -- -0.1572 **(0.0150)

$12.5K <= Qtr UI Earn < $25K (2nd job) -- 0.2097 **(0.0264)

Qtr UI earn >= $25K (2nd job) -- 0.2499 **(0.052)

Observations 7,442 7,442R-squared .015 .058

* significant at 5% level; ** significant at 1% level

Table 6: Effects of Person and Job Characteristics on the Probability that a UI Multiple Job Holder Has a Single CPS Job (Y2) (More Restrictive CPS Classification)

Note: Coefficients obtained from linear probability regressions using pooled data for all years 1996 to 2003 for respondents aged 16 and older. Year dummies included in both models. Standard errors in parentheses.

Age 16 to 24 -0.0230 0.0030(0.0175) (0.0173)

Age 25 to 34 -0.0229 -0.0026(0.0144) (0.0140)

Age 55 to 64 0.0312 0.0427(0.0241) (0.0234)

Age 65 plus 0.0359 0.0967(0.0532) (0.0517)

Less than High School 0.0355 0.0373(0.0198) (0.0192)

Some College -0.0507 ** -0.0555 **(0.0148) (0.0143)

College Graduate -0.0355 * -0.0733 **(0.0174) (0.0171)

More than College -0.0507 * -0.1217 **(0.0242) (0.0240)

Black 0.0531 ** 0.0689 **(0.0173) (0.0167)

Other Non-white 0.0092 0.0012(0.0258) (0.0250)

Male 0.0150 -0.0077(0.0117) (0.0114)

Non-proxy Interview 0.0046 -0.0051(0.0135) (0.0131)

Married -0.0409 ** -0.0203(0.0132) (0.0129)

Foreign Born 0.0454 * 0.0549 **(0.0178) (0.0172)

Any Long 2nd Jobs -- -0.1357 **(0.0154)

Three or More UI jobs -- -0.0656 **(0.0123)

Qtr UI earn < $1K (2nd job) -- -0.1259 **(0.0149)

$1K <= Qtr UI earn < $2.5K (2nd job) -- -0.2357 **(0.0160)

$12.5K <= Qtr UI Earn < $25K (2nd job) -- 0.2479 **(0.0283)

Qtr UI earn >= $25K (2nd job) -- 0.2937 **(0.0561)

Observations 7,442 7,442R-squared .013 .075

* significant at 5% level; ** significant at 1% level

Table 7: Effects of Person and Job Characteristics on the Probability that a UI Multiple Job Holder Has a Single CPS Job (Y2) (Less Restrictive CPS Classification)

Note: Coefficients obtained from linear probability regressions using pooled data for all years 1996 to 2003 for respondents aged 16 and older. Year dummies included in both models. Standard errors in parentheses.

Age 16 to 24 -0.0786 ** -0.1168 **(0.0228) (0.0225)

Age 25 to 34 -0.0139 -0.0331(0.0187) (0.0183)

Age 55 to 64 0.0873 ** 0.0868 **(0.0297) (0.0290)

Age 65 plus 0.0426 0.0124(0.0679) (0.0664)

Less than High School 0.0310 0.0143(0.0271) (0.0265)

Some College -0.0099 0.0036(0.0197) (0.0194)

College Graduate 0.0368 0.0450 *(0.0227) (0.0223)

More than College 0.0969 ** 0.1167 **(0.0294) (0.0291)

Black -0.0365 -0.0378(0.0258) (0.0253)

Other Non-white -0.0337 -0.0367(0.0341) (0.0333)

Male 0.0538 ** 0.0468 **(0.0153) (0.0150)

Non-proxy Interview -0.0571 ** -0.0403 *(0.0178) (0.0175)

Married -0.0252 -0.0178(0.0172) (0.0169)

Foreign Born 0.0450 0.0310(0.0238) (0.0234)

Simultaneous Multiple Jobs -- -0.1020 **(0.0210)

Multiple Jobs All Three Months -- -0.0601 **(0.0220)

16+ hours per week 2nd job(s) -- -0.1236 **(0.0220)

Observations 4,352 4,352R-squared .025 .070

* significant at 5% level; ** significant at 1% level

Table 8: Effects of Person and Job Characteristics on the Probability that a CPS Multiple Job Holder Has a Single UI Job (Y3) (More Restrictive CPS Classification)

Note: Coefficients obtained from linear probability regressions using pooled data for all years 1996 to 2003 for respondents aged 16 and older. Year dummies included in both models. Standard errors in parentheses.

Age 16 to 24 -0.1139 ** -0.1371 **(0.0164) (0.0157)

Age 25 to 34 -0.0614 ** -0.0759 **(0.0130) (0.0125)

Age 55 to 64 0.0466 * 0.0378(0.0208) (0.0198)

Age 65 plus 0.0888 * 0.0269(0.0429) (0.0410)

Less than High School 0.0069 -0.0133(0.0192) (0.0183)

Some College -0.0089 0.0110(0.0137) (0.0131)

College Graduate 0.0256 0.0472 **(0.0156) (0.0150)

More than College 0.0264 0.0667 **(0.0208) (0.0200)

Black -0.0679 ** -0.0744 **(0.0175) (0.0167)

Other Non-white -0.0119 -0.0303(0.0236) (0.0226)

Male 0.0522 ** 0.0454 **(0.0107) (0.0102)

Non-proxy Interview -0.0430 ** -0.0126(0.0126) (0.0121)

Married -0.0221 -0.0213(0.0121) (0.0116)

Foreign Born 0.0263 0.0023(0.0165) (0.0158)

Simultaneous Multiple Jobs -- -0.1685 **(0.0122)

Multiple Jobs All Three Months -- -0.1303 **(0.0153)

16+ hours per week 2nd job(s) -- -0.1389 **(0.0151)

Observations 8,646 8,646R-squared .024 .112

* significant at 5% level; ** significant at 1% level

Table 9: Effects of Person and Job Characteristics on the Probability that a CPS Multiple Job Holder Has a Single UI Job (Y3) (Less Restrictive CPS Classification)

Note: Coefficients obtained from linear probability regressions using pooled data for all years 1996 to 2003 for respondents aged 16 and older. Year dummies included in both models. Standard errors in parentheses.