The Puzzling Pattern of Multiple Job Holding across U.S. Labor Markets

Transcription

1 The Puzzling Pattern of Multiple Job Holding across U.S. Labor Markets Barry T. Hirsch, Muhammad M. Husain, and John V. Winters January 2017 Revision, May 2017 Abstract Multiple job holding rates differ substantially across U.S. regions, states, and metropolitan areas. Rates decrease markedly with respect to labor market size. These patterns have been largely overlooked, despite being relatively fixed over (at least) the past twenty years. This paper explores explanations for these persistent differences. We account for roughly two-thirds of the mean absolute deviation in multiple job holding across local labor markets (MSAs). The results suggest that variation in multiple job holding across labor markets is driven by labor market differences in job opportunities and worker preferences. Most important in explaining variation in multiple job holding are MSA industry and occupation structure, ancestry shares, commute times, and, to a lesser extent, labor market churn. Keywords: Multiple job holding, local labor markets, city size and regional differences, commuting costs JEL code: J21 (labor force and employment), R23 (regional labor markets) Hirsch [contact author]: Department of Economics, Andrew Young School of Policy Studies, Georgia State University, Atlanta, Georgia , and IZA (Bonn). bhirsch@gsu.edu. Husain: Department of Economics, Johns Hopkins University, Baltimore, Maryland 21218, mhpavel99@gmail.com. Winters: Department of Economics, Oklahoma State University, Stillwater, OK , and IZA (Bonn). jvwinte@okstate.edu. We appreciate helpful comments from Peter McHenry, Tyler Ransom, referees, and audience members at university and conference presentations.

2 1. Introduction Working at a secondary as well as a primary job provides an important source of income, human capital accumulation, and job satisfaction for many workers. Understanding the determinants and geographic patterns of multiple job holding is thus important for researchers and policymakers. At a given point in time, roughly five percent of U.S. workers hold multiple jobs; a much larger share have held multiple jobs at some point in the past. Not widely recognized is that rates of multiple job holding differ substantially across regions and labor markets in the U.S., differences that have persisted over time. These geographic differences in multiple job holding rates have received minimal attention in the academic literature. 1 Multiple job holding is far more prevalent in Western North Central, Mountain, Northwest, and New England states than in states elsewhere. Rates are lowest in the South. As we will show, a similar pattern exists across metropolitan areas. Moreover, multiple job holding is found to be substantially higher in non-urban areas than in metropolitan area labor markets. Geographic differences in multiple job holding presumably reflect labor market differences in labor supply and demand. Economists have focused on workers labor supply (work hour) preferences coupled with demand-side constraints on hours worked. Previous studies have shown that total weekly work hours are on average higher for multiple job holders than for single job holders (e.g., Hipple 2010; Hirsch et al. 2016). That said, we do not find labor market (metropolitan area) multiple job holding rates to be positively correlated with market-level total (primary plus secondary job) average work hours. 2 Absence of a positive correlation between market-level multiple job holding and total work hours strongly suggests that supplyside preferences for work hours are not sufficient to explain the geographic patterns. Demand-related forces are likely to play an important role in determining the opportunity set facing workers and the resulting 1 The fixity of the regional pattern over time is documented here, but it can also be seen by comparing BLS annual news releases on multiple job holding by state across years (e.g., U.S. BLS 2015). 2 Based on 259 MSAs during the years (the sample used in our subsequent analysis), we find substantial negative correlations between MSA multiple job holding rates with respect to mean hours worked in first jobs (-0.22) and in second jobs (-0.61), but effectively zero correlation (-0.04) with mean total hours worked. 1

3 market-level rates of multiple job holding. For example, high (low) multiple job holding in a labor market may reflect a mix of primary jobs that produce substantive constraints (or opportunities) for lengthy work hours. Market-level factors that reduce the probability of good first-job matches (e.g., low rates of job churn) are likely to lead to high multiple job holding rates. Factors that lower the attractiveness of multiple job holding (e.g., high commuting costs) should lead to lower rates. Unfortunately, we know little from previous literature about the systematic differences in multiple job holding across U.S. labor markets, the reasons for these differences, or whether these differences reflect worker preferences or labor market constraints. The goals of this paper are two-fold. First, we identify largely unrecognized regional, labor market, and market size patterns in multiple job holding and show that these have been relatively stable over time. Second, we attempt to explain the systematic long-run differences in multiple job holding across metropolitan area (MSA) labor markets by accounting for standard worker and job measures, and for MSAlevel measures of job structure, commuting times, worker ancestry, and labor market churn. Based on the relationships found in our analysis, we assess (informally) the role played by multiple job holding in helping improve labor market outcomes and well-being. To preview results of our analysis, we account for about two-thirds of the cross-market (metropolitan area) variation in multiple job holding and offer several notable findings. Differences in industry and occupation structure, commute times, job churn rates, and ancestry patterns explain a significant share of the multiple job holding variation across MSAs. The findings for job structure and commute times suggest that some workers prefer to work longer hours, but are unable to do so via a second job because good second job matches in their area may be limited or time constraints from commuting make multiple job holding too costly. Our finding that multiple job holding is higher in markets with low job churn suggests that high turnover facilitates good primary job matches and lessens the need for multiple jobs. The correlation between ancestry shares and multiple job holding suggests that cultural norms and attitudes toward work affect employment outcomes, albeit in ways we cannot fully understand. Our descriptive evidence reveals substantive and relatively fixed differences in multiple job holding across U.S. regions and labor markets. 2

4 Although we cannot fully account for or explain such differences, it seems fair to conclude that variation in multiple job holding across labor markets reflects both differences in the opportunity set of primary and secondary jobs, the work preferences of the labor force, and the ease with which workers and jobs are matched. 2. Background and prior literature Multiple job holding (MJH) is typically treated by economists as an individual labor supply decision. Reasons for multiple job holding fall into two broad categories, one focusing on hour constraints and a second on job portfolios. In a widely cited paper, Shishko and Rostker (1976) provide a now-standard indifference curve diagram showing why workers may increase utility by taking a second job at a wage below one s wage at their hours-constrained primary job. Similar diagrams were provided earlier in two largely overlooked papers (Moses 1962; Perlman 1966). 3 Hour constraints on either a primary or secondary job can explain multiple job holding. If a worker s primary job (i.e., the job at which an individual works the most hours) has the higher wage but constrained work hours, workers may increase utility by taking a lower paying second job. Second jobs may be shortterm. For example, workers with constrained hours may take a second job because of temporary financial or family circumstances, expecting that their preferred long-run match is a single primary job. Workers not facing hour constraints on the primary job might take a higher paying second job that has constrained hours; say, a temporary job or one with limited hours per week. Unlike jobs with hourly pay, salaried jobs do not have explicit hour constraints, but do have an earnings constraint that can work in a similar way, leading some salaried workers to take a second job in order to increase their earnings. 4 We adopt the phrase job portfolio from Renna and Oaxaca (2006), who develop a model of 3 In contrast to Moses (1962) and Perlman (1966), Shishko and Rostker (1976) provide empirical analysis estimating the labor supply responsiveness to multiple jobs. Perlman (1966, p. 242) cites BLS reports of economy-wide multiple job holding rates of 5.3 percent in 1957 and 5.2 percent in 1964, the same order of magnitude as current rates, but for what was then a mostly male labor force. 4 Hirsch et al. (2016) make this point. They find a multiple job holding rate for salaried workers that is about half a percentage point lower than for hourly workers. 3

5 multiple job holding based on personal preferences for job differentiation. 5 We include several explanations for multiple job holding under this category. First, workers may prefer diversity in job tasks, being happier dividing time in two different jobs or occupations. Second, workers may work in a second job as a form of insurance, say diversifying one s human capital or because of employment or income uncertainty in a first job. Third, workers wanting to switch occupations or employers due to a poor match can use a second job to obtain job training that might facilitate a utility-enhancing move. Along these lines, Panos et al. (2014) have examined skill diversification and mobility among British multiple job holders. Using SIPP data, Conway and Kimmel (1998) emphasize the importance of heterogeneous jobs and conclude that it is an important reason for dual jobs, in addition to work hour constraints. Their paper also addresses the importance of accounting for multiple job holding when estimating labor supply elasticities. They conclude that multiple job holders have higher wage elasticities than do single job holders, but failure to account for multiple job holders has little bias given that they are a small portion of the workforce sample. An important contribution to the literature on multiple job holding was Paxson and Sicherman s (1996) focus on dynamic job holding. Secondary jobs are often short-term jobs. Although relatively few workers hold multiple jobs during any given week, a substantial number of persons have been multiple job holders at some point in the past. The authors use the Panel Study of Income Dynamics (PSID), which asks about jobs that provide earnings in addition to one s primary job over the previous calendar year. Using this definition, they find that 21 percent of men and 12 percent of women held dual jobs at some point during the previous calendar year (these figures were averaged over the PSID survey years ). These average annual rates were roughly double the rates that Paxson and Sicherman estimate to be comparable weekly rates based on the PSID definition of multiple jobs. Recent literature has returned to the theme that multiple job holding may be similar to short-term jobs. As noted by Abraham et al. (2013), there are differences in measuring multiple job holding based on 5 The job portfolio terminology had been used but not highlighted in prior literature (e.g., Paxson and Sicherman 1996). 4

6 household (CPS) data versus establishment data. Using a data set matching individual worker information with administrative employer-reported data indicated that establishment measures of multiple jobs within the same quarter sometimes fail to coincide with CPS worker reports of multiple job holding. Such discrepancies need not indicate reporting error. A worker with two (or more) jobs within a quarter may not have held multiple jobs during the CPS reference week. Likewise, reports of multiple job holding in the CPS sometime fail to show up as two jobs in administrative payroll records. Although we cannot rule out reporting error in the CPS, such discrepancies will exist if earnings from either the primary or secondary job are not reported to tax authorities (i.e., off the books). There also exists a literature examining multiple job holding and the business cycle. Amuédo- Dorantes and Kimmel (2009) provide a thorough summary of this literature, identifying studies that report a diverse set of results, some finding multiple job holding to be cyclical, some countercyclical, and some acyclic. Their own analysis using data from the NLSY79 finds mixed results. Using state-level employment growth as their business cycle measure, the authors conclude that multiple job holding among men is largely acyclic, whereas multiple job holding among women switched from countercyclical during the 1980s and early 1990s to procylical by A recent paper by Hirsch et al. (2016) uses a large CPS data set for to examine how multiple job holding in U.S. labor markets (MSAs) varies with respect to local unemployment rates and employment growth. Theory is ambiguous. Labor supply can be cyclical or countercyclical depending on the strength of income and substitution effects. Even if income effects were dominant, leading to a desire for multiple jobs when unemployment is high, it does not follow that such jobs are available given that markets need not clear in recessions. Absent MSA fixed effects, the authors obtain small but precisely estimated negative coefficients on the unemployment rate (or tiny positive coefficients using employment growth), reflecting higher multiple job holding in low unemployment labor markets. Once labor market fixed effects are added, however, coefficients effectively go to zero. Similarly, using two-year CPS panels of workers within MSAs, transitions into and out of multiple jobs over a year are uncorrelated with changes in 5

7 unemployment. The authors conclude that multiple job holding in the U.S. is effectively acyclic. As compared to the U.S., relatively few studies focus on multiple job holding in Europe. Zangelidis (2014) examines European evidence using a large micro-level dataset, the European Union Labour Force Survey (EU-LFS) for The MJH rate across 28 EU countries is lower than in the U.S., 3.2 percent in the EU versus about 5 percent in the U.S. Just as the U.S. displays substantial differences in MJH across states and regions, Zangelidis finds large variation in MJH rates across countries in the EU, ranging from less than 1 percent to 9 percent. 6 Livanos and Zangelidis (2012) document large differences in MJH rates across regions of Greece with rural areas with large primary sectors having the highest rates, likely due to low labor demand and weak primary job opportunities in those areas. Livanos and Zangelidis (2012) also find that MJH in Greece is procyclical. 7 A paper by Partridge (2002), who uses U.S. state level data to examine the determinants of multiple job holding, recognizes the substantial differences across states in multiple job holding rates. He provides a map of the U.S. showing the strong regional patterns in BLS multiple job holding rates (similar to our Figure 2, presented subsequently). His estimation sample includes the lower 48 states for the years (n = 48 x 4 = 240). Partridge includes state and year fixed effects in his MJH models. In contrast to our analysis, the goal of Partridge s paper is to estimate the determinants of multiple job holding behavior and not the fixed differences in MJH across states. Partridge nets out time-invariant differences in state MJH by including state 6 Zangelidis also finds that mean weekly hours on the second job average 12.9 hours across the continent, with little variation across countries. Zangelidis introduces the concept of second job intensity, measured by the percentage of total work hours due to the second job. Among multiple job holders, the intensity measure averages 26.7 percent across all 28 countries and displays limited variation (values range from 22 to 34 percent). Hirsch et al. (2016, p. 25) calculate a second job intensity measure for the U.S. that is just slightly higher than the EU country mean. 7 Our summary of the literature is not exhaustive. Other papers using the CPS include Hipple (2010), who provides extensive descriptive evidence (means) on multiple job holding rates for various groups of workers, and Lalé (2016), who provides detailed evidence on worker flows, transitions into and out of multiple job holding, and the relationship of multiple job holding and part-time primary jobs. Analyses of U.S. multiple job holding also have used the Survey of Income and Program Participation (SIPP), the Panel Study of Income Dynamics (PSID), and the 1979 National Longitudinal Survey of Youth (NLSY). For example, Conway and Kimmel (1998), Kimmel and Conway (2001), and Krishnan (1990) use SIPP; Paxson and Sicherman (1996) use the PSID (and CPS); and Amuédo-Dorantes and Kimmel (2009) use the NLSY79. These longitudinal data sets provide a rich set of covariates and enable researchers to examine worker transitions over lengthy time periods. Because of its large size and geographic coverage, the CPS is far better suited to examine multiple job holding patterns across labor markets than are these alternative data sets. 6

8 fixed effects. He notes the importance of these fixed effects, but does not attempt to account for their variation. A principal goal of our paper is to identify and explain the fixed differences in market-level multiple job holding, measured primarily at the MSA rather than state level. Because Partridge used state data for multiple job holding and its covariates, he could not observe the substantial urban/rural and city size differences in multiple job holding uncovered in our work. Most studies focusing on the determinants of multiple job holding have used individual level data and primarily (or exclusively) individual-level covariates. Geographic and market-level factors affecting U.S. multiple job holding have been largely ignored. We provide analysis that includes a rich set of individual level covariates, but also emphasize market level determinants. It is reasonable to expect that market-level forces can influence rates of multiple job holding. On the demand side, for example, some types of industries or occupations may prefer workers who work limited hours, are temporary, or work non-standard hours. The presence of such employers is likely to increase the availability and pay for second jobs, thus increasing multiple job holding. These same industries and occupations may also hire primary job workers who are offered limited hours, thus increasing the labor supply of hours-constrained workers desiring second jobs. On the supply side, labor market workforces may differ (conditional on individual covariates) in their preferences regarding total work hours and the types of jobs that they hold. Such labor supply differences will produce differences in multiple job holding. Differences in work preferences may reflect long-standing cultural and historical norms passed through generations within a labor market, or reflect norms acquired through ancestry rather than location per se. If market-level differences in desire for work hours were the only force at work, we would expect to see a positive correlation across labor markets in multiple job holding rates and total work hours (the sum of usual weekly hours on primary and second jobs). As stated previously (footnote 2), there is a negative (but near zero) correlation between market-level multiple job holding and total work hours. To the extent that many workers in rural or small urban markets have strong local preferences, we might see high rates of multiple job holding due to the difficulty in finding a good primary job match. That said, it seems unlikely 7

9 that differences in preferences and other supply-side factors can by themselves explain the substantive differences we find in market-level multiple job holding. Demand and other market-level factors are also likely to be important. In what follows, we first document systematic but largely unrecognized geographic patterns of multiple job holding across regions and with respect to labor market size. We then examine labor market differences in industry and occupation structure, along with several other market level determinants of multiple job holding that can affect the attractiveness and desire for multiple jobs. For example, traffic congestion and long commute times may decrease the willingness of workers to take second jobs. Labor markets with low levels of churn (turnover) may produce imperfect worker-job matches that lead to a desire for second jobs, while at the same time reducing the ease of finding such jobs. 3. Measurement of multiple job holding using the Current Population Survey We use the U.S. Current Population Survey (CPS) to measure multiple job holding (MJH). The CPS began continuous collection of information on multiple job holding in 1994 as part of the survey s major redesign. Prior to 1994, occasional CPS supplements included information on multiple job holding. Since 1994, all employed individuals are asked the question: Last week, did you have more than one job (or business), including part-time, evening, or weekend work? If they answer yes, they are then asked how many jobs (or businesses) they had altogether and how many hours they worked each week at all their jobs. The primary job is defined as the one at which the greatest number of hours were worked. Using monthly CPS data, the U.S. Bureau of Labor Statistics (BLS) defines a multiple job holder as an individual who: (a) holds wage and salary jobs with two or more employers; (b) combines a wage and salary job with self-employment; or (c) combines a wage and salary job with one as an unpaid family worker. In our analysis, multiple job holding is defined similarly, with the exception that our sample includes only those workers whose primary job is a wage and salary job and we restrict the sample to non-students ages (versus all workers ages 16+ by BLS). These sample criteria produce an estimation sample similar to those commonly seen in wage and employment analyses. MJH rates are affected little by these differences. 8

10 The exclusion of year olds and workers over 65 raises MJH rates by roughly two-tenths of a percent. Exclusion of full-time students raises MJH rates by less than one-tenth of a percent. Exclusion from the sample of those with a primary self-employment job typically reduces MJH rate estimates by one-tenth of a percent. Overall, our sample has MJH rates in most years that are one- or two-tenths higher than published BLS rates, with the gap being a bit larger in the earliest years and at most one-tenth of a percent in recent years. 8 In this paper, we use all rotation groups of the monthly Current Population Survey (CPS) data files from January 1994 through December The CPS reports work hours, detailed occupation, and detailed industry for both the primary and second jobs. Earnings are reported only for the primary job. We first focus on differences in multiple job holding with respect to urban versus non-urban markets and then turn to differences across states and metropolitan areas. Our initial analysis focuses on urban/nonurban differences in multiple job holding using a CPS sample of 13,448,612 workers, 9,914,287 or approximately three-fourths (0.737) of whom live in MSAs identified in the CPS, with the remaining 3,534,325 residing in non-urban areas. As we later discuss, households in non-urban areas are oversampled and have lower sample weights, while those in large urban areas have higher weights. This sample excludes CPS files during June-August 1995 (n=163,499), in which there were no metropolitan area identifiers, thus precluding identification of urban versus nonurban areas. 8 Fewer than 1 in 20 workers counted as multiple job holders by BLS in 2009 held a primary self-employment job (Hipple, 2010, Table 4). It is common for multiple job holders to have secondary self-employment jobs; these workers are included in our sample. Data restrictions influenced our sample criteria. We use the IPUMS CPS monthly files for the analysis, but these include neither an edited BLS field on multiple job holding nor the class of worker for the second job that would show whether workers with a primary self-employment job hold a wage and salary second job. 9 Households are in the CPS survey for eight months. They are interviewed four consecutive months (rotation groups 1-4), then out the next eight months, and then reenter the following four months (rotation groups 5-8). An initial working paper version of our paper used the quarter sample outgoing rotation groups (rotation groups 4 and 8) in order to have information on individuals earnings on the primary job and union status. We switched to all rotation groups once we determined that rotation group bias existed, with the highest rates of multiple job holding reported in rotation groups 1 and 5 and lower rates reported the longer one is in the survey (see the note by Hirsch and Winters 2016). This pattern is identical to that reported by Krueger et al. (2014) for unemployment rates, but the bias for multiple job holding is larger than for unemployment. That said, the basic conclusions of our analysis have changed little. In the prior analysis, individual earnings and union status were not important in explaining labor market differences in multiple job holding. Our current analysis measures earnings and union concentration at the MSA level. 9

11 Analysis of combined urban/nonurban samples at the national and state level include all months of Key analyses in our paper focus on the relatively fixed differences in multiple job holding across MSAs and with respect to market size. Analyses focusing on urban-nonurban and metro size differences in multiple job holding use a sample from September 1995 through December Our beginning date corresponds to the introduction of new MSA definitions in the CPS. Our ending date is prior to MSA definition changes in 2015 (we account for changes introduced during 2014). This CPS sample includes 11,962,560 workers, with about three-fourths (0.739) of the sample (8,838,400) living in MSAs identified in the CPS, and the remaining portion of the sample (3,124,160) residing outside these MSAs. 10 Subsequent analysis focuses exclusively on the 259 MSAs present in the CPS for the period September 1995 through December 2014 (202 present over the entire period and 57 small MSAs present in some but not all years). Unless otherwise stated, all analyses in the paper use survey weights. To illustrate the (substantial) difference weighting has on descriptive statistics, it is useful to compare weighted and unweighted mean multiple job holding rates. Over the entire period the national, urban, and non-urban weighted mean MJH rates are 5.6, 5.3, and 6.7 percent. The comparable non-weighted sample means are 6.1, 5.6, and 7.5 percent. In order to enhance reliability, the CPS oversamples households in less populated markets and undersamples in large markets. Because MJH rates systematically decline with size, it is essential that we use Census survey weights to provide unbiased descriptive statistics for representative populations. Because multiple job holding behavior may be heterogeneous, weighted regressions provide coefficient estimates representing roughly average effects across heterogeneous groups (see Solon et al. 2015). Figure 1 provides national evidence on the trends over time in multiple job holding. National annual rates (triangles) have trended down over time, from 6.3 percent in 1994 (and a high of 6.7 percent in 1996) to 10 The CPS does not identify all MSAs, excluding those that are small, roughly 100,000 or below in size (MSAs must have a central city of at least 50,000). Every 10 years Census adds and removes smaller MSAs based on population changes. What we refer to as our non-urban group includes both workers living outside of an MSA, plus those in small MSAs not identified in the CPS. 10

12 an eventual 5.0 percent in Of particular interest for our analysis is the largely unknown difference in multiple job holding rates between those in non-urban (squares) versus metropolitan areas (circles), rates being substantially higher in non-urban areas. The downward trends in MJH rates are similar in urban and non-urban areas, although estimates for the latter are more volatile. Although not shown in Figure 1, downward trends in MJH have been stronger among men than among women. Men s MJH rates between 1994 and 2015 declined from 6.4 to 4.6 percent, whereas women s rates fell from 6.2 to 5.3 percent. The sharper decline among men than women occurs in both the metropolitan and non-metropolitan samples. The secular downward trend cannot be accounted for by macroeconomic conditions. Multiple job holding is weakly cyclical, but the relationship is close enough to zero to characterize it as acyclic (Hirsch et al. 2016). 4. Systematic differences in multiple job holding across regions, states, and metropolitan areas Multiple job holding rates differ substantially across regions, states, and labor markets. These differences have substantial fixity over time. Neither the geographic differences in MJH nor the stability of these differences over time is widely recognized. In this section, we provide descriptive evidence on each of these patterns. We first use our CPS data set to show regional and state differences in multiple job holding over time. We then examine evidence on MJH differences across non-urban versus urban areas and show how MJH decreases with metropolitan area size. MJH differences across metropolitan areas display the same regional pattern seen for states. The stability of state MJH is shown through comparisons of MJH rates and relative rankings in versus A similar analysis is shown for metropolitan areas based on MJH rates in versus (we begin with 1996 and end with 2014 due to changes in MSA definitions in 1995 and 2015). Figure 2 provides shade-coded maps of relative multiple job holding rates among U.S. states in , , and Given the downward trend in MJH rates over time, we grouped the states into quartiles, states with the highest MJH rates coded in black, the next quartile in dark grey, the next in 11 As noted previously, our MJH rates differ slightly from official BLS rates because we require that workers have a primary wage and salary job, limit our sample to ages 18-65, and exclude full-time students. 11

13 light grey, and the lowest in white. Readily evident is the substantial similarity of the shade codes over time, with blocks of black (high MJH) among states in the north central region and northern New England, and blocks of white (low MJH) in the southeast, southwest, California, Nevada, New York, and New Jersey. In the top half of Figure 3 we show a scatterplot of the 51 state MJH rates (D.C. included) in (y-axis) and rates 20 years earlier in (x-axis). The same pattern is shown in the bottom half of the figure, where the scatterplot is based on MJH rankings rather than rates. MJH rates are closely related over the time period. A weighted OLS regression of MJH state rates on rates has an R 2 of 0.63 and a coefficient of 0.67 on the rates. A similar regression using MJH rankings had an R 2 of 0.58 and a coefficient of 0.79 on the rankings. Coefficients on MJH rates are expected to be below 1.0 given the secular decline in multiple job holding. Measurement error in MJH rates due to sampling may attenuate coefficients in both the rate and rank equations. The principal analysis in this paper focuses on multiple job holding differences across urban labor markets based on metropolitan areas identified in the CPS. Evident here are the same regional differences seen previously for states, plus differences in MJH rates by market size. We offer several pieces of evidence. Tables 1a and 1b provide lists of MSAs with the highest and lowest levels of MJH averaged over September December 2014 for the 202 MSAs continuously included in the CPS over this period. Here we see regional patterns similar to those seen in the state maps. Relatively high rates are observed for north central MSAs, a few of which are home to large state universities (Table 1a). MSAs with low MJH rates are concentrated in the south, along with several California cities and the large NYC-NJ MSA (Table 1b). Table 1a shows labor markets with high MJH dominated by relatively small MSAs (the mean sample size across MSAs is 34,498). In contrast, Table 1b showing low MJH markets includes several large MSAs (e.g., New York, Houston, and L.A), and has an average sample size of 73,027, more than double the size for high MJH markets (Table 1a). Comparison of sample sizes understates differences in population given that small (large) markets are oversampled (undersampled). More directly, there is strong within-state correlation between metro and non-metro MJH rates. Over the entire time period, the within-state correlation between 12

14 MSA and non-msa multiple job holding rates is The equivalent correlations for our earliest years ( ) and most recent years ( ) are 0.60 and 0.77, respectively. Figure 4 provides further evidence showing that the state and regional differences in MJH, seen previously in Figure 3, are not driven entirely by the 75 percent of workers who reside in MSAs. In Figure 4, we provide a scatterplot of within-state mean MJH rates for residents living in metropolitan areas (shown on the vertical axis) and of within-state mean MJH rates for residents living outside CPS designated metro areas. Both sets of means use sample weights and are calculated over the entire period. Clearly evident in Figure 4 is a strong correlation between within state urban and non-urban MSA rates. The stark regional differences seen in multiple job holding reflect both urban and non-urban differences. In order to examine fixity in MSA multiple job holding over time, in Figure 5 we show a scatterplot similar to that seen previously for states. The vertical axis measures the MSA multiple job holding rates calculated for while the horizontal axis shows the rates for Three-year averages are used to reduce sampling error, a concern for smaller cities. As evident in the figure, there is a relatively high degree of similarity in relative rates between the years. A weighted OLS regression of the rate on the rate produced an R 2 of 0.32 and a coefficient of 0.46 on the rate. The coefficient is likely less than 1.0 due to both the secular decline in MJH rates and attenuation bias from measurement error in the MSA rates. As seen in Figure 5, the range of metro area MJH rates is lower in than in This is due to the secular decline in MJH rates; the coefficient of variation is slightly higher in In addition to there being regional patterns and considerable fixity over time, multiple job holding also varies with respect to labor market size. In Table 2, we show the average MJH rates over the September (n = 11,962,560) period for both non-urban and metropolitan areas of varying sizes. In column 1, we show the mean MJH rates among workers residing in non-urban areas, those areas of the country either 12 Excluded are DC, plus a tiny number of states either with no non-msa households or with no designated MSAs in some or all years of the CPS. 13

15 outside of an MSA or in a small MSA (typically less than a 100,000 population and not identified in the CPS), followed by six groups of MSAs of increasing population size. The mean (weighted) MJH rates over systematically decline with size, ranging from 6.7 percent for the non-urban areas down to 4.4 percent among workers in MSAs 5 million plus. Little of the difference by size is accounted for by standard covariates. Adding a detailed set of worker and job attributes (listed in the note to Table 2), the spread between the unadjusted non-urban and largest urban markets decreases only slightly, from 2.3 to 2.1 percent (columns 1 and 2). 13 Of particular interest is column (3), however, where we add measures of labor market commute times. 14 Controlling for average commute times, the substantive differences in MJH rates previously seen for large versus small MSAs are no longer evident. We provide further evidence on commuting time costs in the next section of the paper. 5. What might explain metropolitan area differences in multiple job holding? The discussion and evidence in the prior section established that there is considerable variation across U.S. labor markets in rates of multiple job holding and that these differences are relatively stable over time. An obvious question arising from such evidence is: What explains these labor market differences in multiple job holding? We consider several possible explanations below, some that can be measured directly, some that can be imperfectly captured through proxy measures, and some that cannot be readily measured. Our strategy is to begin with the raw differences in multiple job holding rates for our 259 metro labor markets (202 of which were measured continuously over the entire September 1995 through December 2014), and then see to what extent these differences are reduced as we introduce various covariates. The CPS contains detailed measures of individual worker demographics and job types. We first control for differences in worker demographics and human capital (measured by schooling and potential 13 As indicated in the note to Table 2, the MJH dependent variable is coded 100 rather than 1.0 for multiple job holders and zero otherwise. This allows the coefficients shown in Table 2 to be interpreted as percentage rates. Regressions in Table 2 have standard errors clustered by MSA and by non-urban state areas. 14 Average commute times for MSAs and nonurban areas are calculated from the 2000 Census and the American Community Survey (ACS), as described subsequently in the next section of the paper. Mean commute times for non-urban areas are based on same-state residents not residing in one of our designated MSAs. 14

16 experience). We then add measures of worker job attributes on the primary job (hours worked in primary job and job sector as measured by public, industry, and occupation dummies). We then add MSA level measures based on data from the CPS, the decennial Census, the American Community Surveys (ACS), the Quarterly Census of Employment and Wages (QCEW), and the Local Area Unemployment Statistics (LAUS). These measures include commute times, labor market size, market measures of mean earnings, housing values, rental rates, union density, industry and occupation shares, percent foreign born, ancestry, market level job churn (turnover), employment growth, and unemployment. 15 These controls account for a substantial share of the dispersion in multiple job holding across markets, but some variation remains unexplained. As discussed at the outset, an economic-based explanation for multiple job holding is that MJH results from hours constraints on the primary job. Hours constraints may be more likely in labor markets with slow rates of labor demand and employment growth, while being less constrained in high growth labor markets. Of course, we cannot easily distinguish between employment growth driven by labor demand versus labor supply. Hirsch et al. (2016) provide clear-cut evidence that labor market MJH rates are not correlated with either local unemployment rates or employment growth after accounting for MSA fixed effects. Absent fixed effects, multiple job holding is weakly procyclical, consistent with a dominant labor demand effect and/or ambiguous labor supply effects due to competing income and substitution effects. An additional explanation for residual differences in labor market MJH is the degree of labor market dynamism or churn, although theory here is ambiguous. Recent literature has noted that the U.S. is exhibiting a gradual decline in overall labor market turnover, possibly reflecting a lower degree of dynamism in the U.S. economy (Decker et al., 2014). Similar patterns and concerns have been noted with respect to worker mobility. Internal migration within the U.S. has shown a gradual but steady decline since the early 1980s, raising further concerns that labor mobility and economic dynamism have fallen (Molloy et al. 2011, 2014). 15 Given the fixity of MJH differences across areas, we measure all of our MSA level variables either as an average over time or for a single point in time. Most of these are computed from the pooled 2000 Census and ACS. The exceptions are labor market size (computed from 2006 Census population estimates), union density (CPS), churn (CPS), employment growth (QCEW), and unemployment rate (LAUS). 15

17 A typical argument is that high (but not too high) rates of turnover reflect and make possible desirable matching and sorting in the labor market. If that is the case, we would expect high rates of churn to be associated with good primary job matches in which hours are not constrained, and thus lower rates of multiple job holding. This argument complements evidence found by Bleakley and Linn (2012), who show that MSA-level churn, which they measure by worker-specific changes in industry and occupation, is lower in densely populated areas. They find this result for both voluntary separations and involuntary worker displacements (the latter based on evidence from CPS Displaced Worker Survey supplements). Part of the productivity (wage) advantage of large markets is that workers acquire high-quality matches. Using this same logic, good primary job matches should reduce multiple job holding. This is supported by our evidence of substantially lower rates of MJH in urban labor markets versus non-urban areas, as well as our subsequent finding of lower MJH in metropolitan areas with higher churn rates. That said, there may be forces that work in the other direction. Hyatt and Spletzer (2013, Forthcoming) find that secular employment losses are associated with fewer short-term (one-quarter) jobs. Thus, it is possible that the gradual decline in multiplejob holding might be associated with lower churn and fewer short-term jobs. In the analysis that follows, we examine whether MJH rates are related to the level of churn. Rates of turnover at the MSA (and state) levels are constructed from the full September CPS files (i.e., all rotation groups) based on individual monthly individual transitions between employment and non-employment and job changes among those employed in consecutive months. Another possible explanation for MSA variation in multiple job holding is that low commuting costs in a labor market will be associated with higher MJH rates, and vice-versa. This is a natural extension of the work by Black et al. (2014), who find that metropolitan areas such as Minneapolis, with low commute times, have higher rates of female labor force participation than do labor markets such as New York City with long commute times. A quick glance at state rates of multiple job holding show Minnesota (and surrounding states) with among the highest multiple job holding rates, while New York has a relatively low rate multiple 16

18 job holding rate as compared to other northern states. The New York and several California MSAs are among the few non-southern metro areas in the list of MSAs with low multiple job holding. Multiple job holding decisions could be particularly sensitive to congestion costs. Commuting is largely a fixed cost of employment and hours worked at second jobs are far lower than in primary jobs, so the relative costs of commuting are high in second jobs, leading to a negative relationship between MJH and commute times. Possibly working in the opposite direction, high commute costs might lead to a poorlymatched primary job and thus increase demand for a second job. Census data for 2000 and, for later years, the American Community Survey (ACS), provide data on commute times. Given that city size is inversely related to multiple job holding, coupled with evidence seen previously in Table 2, we expect that commute times will explain some portion of the residual differences in multiple job holding across labor markets. The high rates of multiple job holding in the north central states give rise to a fourth possible explanation for systematic regional differences. Ethnic, religious, and cultural differences may affect labor market outcomes, including MJH. The north central region of the U.S. has a large number of households whose members are Lutheran and/or of German and Scandinavian heritage. Data on religion by area is not provided by Census or other governmental statistical agencies. The CPS, which provides data on MJH, includes little information on ethnicity, apart from identifying those who are Hispanic. The CPS provides no information on ancestry, with the exception that it records country of origin among those who are foreign born. Data on ancestry, however, is available in the decennial census long form survey in 2000 and the American Community Survey (ACS). We compile metro area measures of ancestry combining the 2000 Census with the pooled ACS. These measures allow us to demonstrate whether ancestry differences across U.S. labor markets are correlated with long-run differences in multiple job holding. Finally, the industrial and occupational structure of a metropolitan area could affect multiple job holding. Some types of primary jobs more naturally lend themselves to second job holding than do others, e.g., because of the physical demands, time demands, start times, and flexibility. Such differences can be loosely accounted for with detailed occupation and industry controls for one s primary job. Job structure 17

19 effects (measured by industry and occupation shares), however, extend beyond the impact of one s own primary job. Industries with strong labor demand for temporary, seasonal, or part-time workers, for example, can provide attractive second job opportunities for amenable workers. Moreover, we cannot condition on individual-level occupation and industry for second jobs since these variables are only observed for workers holding multiple jobs. Instead, we use the 2000 Census and ACS to measure metropolitan area industry and occupation shares to assess the effects of job structure on MJH differences across areas. 6. Evidence on multiple job holding differences across labor markets In this section, we examine why multiple job holding differs across markets, focusing on the explanations offered in the previous section. Our approach is to examine the extent to which controlling for a variety of detailed worker, job, and city attributes can account for differences across labor markets in multiple job holding. To describe the magnitude of MSA differences (dispersion) in multiple job holding, we calculate the mean absolute deviation (MAD) of MJH across our 259 labor markets based on estimates from increasingly dense individual worker OLS multiple job holding equations using the urban sample (n=8,832,284). 16 MAD is calculated as follows. For each multiple job holding regression, where m ijt is multiple job holding for individual i in MSA j during time period t, multiple job holding is coded 0 or 100 (rather than 1, so rates are measured as percentages). We estimate MJH regressions (models 1 through 11) with an increasing number of controls. m ijt = X ijt β + ε ijt For each model we extract the residuals, e ijt, and calculate their weighted means, e j, by MSA j for each of the 259 MSAs. We then compute their absolute deviation from zero, e j (zero being the weighted full 16 This count is slightly lower than that reported for the full urban sample used in Table 2 due to exclusion of a small number of small MSAs newly added to the CPS mid-year in

20 sample national mean of residuals, by construction). Finally, we calculate the weighted mean absolute deviation of multiple job holding across these 259 MSAs. 17 That is, MAD = w j e j Σw j where w j is the total sample weight for MSA j, summed over all individuals and time periods within each MSA. Our analysis examines the extent to which these weighted mean absolute deviations are reduced as we introduce increasingly dense controls. Table 3 shows the values of MAD using eleven increasingly dense MJH equation models. We first enter individual level attributes from the CPS. We then enter MSA level attributes compiled from several datasets. Of course, the contribution of each variable(s) to the measure of dispersion is not independent of the order in which variables are entered. We have examined multiple orderings; conclusions regarding the relative importance of particular sets of variables are reasonably insensitive to order. In subsequent analysis, we show the effects of each set of the attributes both when they are entered first and entered last into the MJH equations. Rankings of attribute importance are highly similar in these alternative approaches. 18 The first specification in Table 3 includes only year and month dummies, which we interpret as an unconditioned measure of the dispersion in MJH rates across labor markets. The second model adds controls for worker demographic and human capital characteristics sex, race/ethnicity, foreign-born, marital status, presence of young children, and detailed education and age dummies. The third adds job-level 17 For MJH covariates measured at the individual level, our approach is conceptually similar to including MSA fixed effects and measuring their mean absolute deviation. A fixed effects approach has two disadvantages. First, all variation from MSA level measures (e.g., commute times) is absorbed by fixed effects. Second, the approach is computationally more time and memory intensive, which is relevant given the large sample sizes and dense specifications. 18 Gelbach (2016) discusses how order of variable inclusion can affect coefficient decompositions and related accounting exercises. He outlines an order-invariant method to measure the partial effects of multiple covariates on a coefficient(s) of key interest. His approach is not directly applicable here given that that our outcome of primary interest is not the magnitude of coefficients but the dispersion in MJH rates across areas as measured by the mean absolute deviation of MSA mean residuals, for which the contribution of individual regressors cannot be estimated in an orderinvariant manner. 19

21 controls measuring public employment and hours worked on the primary job, while the fourth adds workers broad occupation and industry. The remaining models add MSA level measures. The fifth model adds MSA commute times, the sixth city size dummies, and the seventh MSA mean earnings, housing values, rental rates, and union density. The eighth model adds MSA-level industry and occupation shares (individual-level industry and occupation were included previously). The ninth model adds MSA measures of ancestry and percent foreign born, the tenth a measure of MSA turnover (churn), and the eleventh measures of MSA log employment growth and unemployment rates averaged over Given the large number of covariates and model specifications, we do not provide a full set of regression coefficient estimates. In the data appendix, we include a partial set of coefficient estimates for model 11, our most dense specification. 19 As seen in Table 3, line 1, the weighted mean absolute deviation of MSA multiple job holding absent controls (apart from year and month dummies) is 0.99, a 1 percentage point average absolute difference between MSA rates and the mean rate of multiple job holding in urban areas. 20 This average deviation from the mean is 18 percent the size of the 5.3 percent mean level of multiple job holding. The second specification, which adds control for worker demographic and human capital characteristics, reduces MAD from 0.99 to The third, which adds job-level controls measuring public/private sector employment and hours worked dummies on the primary job, reduces MAD from 0.82 to Addition of occupation and industry dummies of the primary job in model 4 slightly increases MAD to In short, controlling for individual worker and job measures available in the CPS accounts for a rather modest portion (20 percent) of the dispersion in multiple job holding across markets (from 0.99 to 0.79). 19 Note that this analysis examines MJH differences across urban labor markets (MSAs) and not differences between urban and non-urban areas. Regression results shown previously in Table 2 address urban/non-urban differences. 20 MJH rates are higher in winter months (February is highest) and lower in summer months (August is lowest). Monthly differences are highly similar unconditioned and in dense specifications. Although we do not examine the relationship between temperature and MJH, it is worth noting that not only is MJH lowest in summer and highest in winter, state MJH rates are highest in northern states and lowest in southern states, as seen previously in Figure 2. An opportunity cost of working at a second job may include foregoing outdoor leisure activities. Prior literature using the American Time Use Survey (ATUS) has shown an effect of high temperatures (Zivin and Neidell 2014) and rain (Connolly 2008) on work hours and leisure. High temperatures reduces work outdoors; rain decreases leisure. 20

22 Beginning with model 5, we introduce variables measured at the MSA rather than individual level. The model 5 specification adds MSA mean commute times (average minutes for a one-way trip from home to work), calculated from the pooled 2000 Census 5% PUMS and ACS for those persons who work outside the home. Inclusion of this measure sharply reduces unexplained deviations in multiple job holding, with MAD falling from 0.79 in line 4 to 0.64 in line 5. In model 6, we add city size dummies, which has a minimal effect, reducing MAD from to Of course, commute times and city size are highly correlated. When we reverse the order in which we introduce these two measures (not shown in Table 3), we find that adding city size dummies to line 4 reduces MAD from 0.79 to 0.70, a much weaker impact than seen for commute times. The addition of commute times following conditioning on city size reduces MAD substantially, from 0.70 to Our conclusion from these results is that commute times play an important role in determining multiple job holding rates. Across a wide range of specifications, the coefficient on mean transportation times for a daily commute (in minutes) is about -0.1 and always statistically significant at the 0.01 level. This implies that a 5 minute increase in a labor market s average commute time from home to work (the standard deviation across MSAs is 4.3 minutes) is associated with a 0.5 lower MJH rate (half a percentage point). In work not shown, we fail to observe substantive differences between women and men in the relationship between MJH and commute times, in contrast to findings in Black et al. (2014) that show married women s labor force participation to be particularly sensitive to commuting costs. In model 7, we introduce measures reflecting metro area income and housing values mean hourly earnings, housing values, and rental values plus union density (which can affect earnings). The first three of these measures are compiled from the 2000 Census and the ACS. Annual union density measures by MSA have been calculated from the CPS outgoing rotation group files by Hirsch and Macpherson (2003, with annual updates online), available at Unionstats.com. We use the average union density over the full sample period. This set of variables accounts for a small amount of MSA dispersion in multiple job holding, reducing MAD from 0.64 to

23 Although individual level industry and occupation on the primary job (model 4) did not help account for labor market differences in multiple job holding, accounting for an MSA s overall industry and occupation structure (model 8) substantially reduces MAD, from 0.62 to Among industries, the employment share of agriculture, forestry, and fishing has a substantial positive effect on MJH, which is consistent with the regional MJH pattern seen in Figure 2. The seasonality of work hours in these primary jobs is likely to create off-season demand for secondary jobs. Shares in business and repair services, mining, and public administration are associated with substantively lower MJH. Among occupations, high rates of MJH are found in MSAs with large employment shares in executive, administrative, and managerial occupations and in administrative support. Large shares in sales occupations are associated with low MJH rates. Although measures of industry and occupation at the individual and market levels need not coincide, we are somewhat surprised by the large effect of the latter absent substantive effects from the former. The regional patterns seen in state multiple job holding rates (Figure 2) prompted us to examine the effects of ancestry, as measured in the 2000 Census and annual ACS. Individuals are asked What is your ancestry or ethnic origin? This is followed by examples such as Italian, Jamaican, African American, Ukrainian, and so forth. The Census Bureau codes up to two answers for an individual. They do not provide codes for answers that are rare. We use responses on first ancestry (and ignore second measures) and tabulate and include in our MJH regression the percent of a MSA s workers who identify their ancestry as German, Irish, Italian, Nordic, Other Western Europe, Eastern Europe, Latin America, Northern Africa and the Middle East, Sub-Saharan Africa, Asian & other Pacific, Native American, and other white ancestries; English ancestry is the excluded group. In addition, we include a measure of MSA share foreign born. Recall that we already include individual worker measures of race (including Asian), ethnicity, and foreign born (separately for citizen and non-citizen) from the CPS. As seen in line 9, introduction of the percent ancestry variables accounts for a substantive share of the labor market variation in MJH, reducing MAD from 0.44 to Relatively high MSA multiple job holding is seen in labor markets with large shares of workers whose ancestry are (beginning with highest) Nordic, Asian-Pacific, English, German, other Western Europe, Italian, 22

24 Latin American, and Eastern Europe. Low MJH rates are associated with ancestry shares of (beginning with lowest) Northern Africa and the Middle East, Irish, Native-American, and Sub-Saharan Africa. 21 Scholars have found strong relationships between ancestry and economic performance across countries (Putterman and Weil 2010) and U.S. counties (e.g., Fulford, Petkov, and Schiantarelli 2015). 22 Putterman and Weil (2010) examine cultural variables (e.g., trust and obedience) and measures of institutions. They emphasize that economic outcomes of countries today are strongly correlated with the institutional development of its current population s ancestors from some 500 years ago, more specifically their agricultural and political development in Fulford et al. (2015) likewise find that the historical cultural values and institutions of the ancestors of today s populations in U.S. counties are associated with higher rates of county GDP growth. Interestingly, they find that on average people whose ancestors lived in high income countries disproportionately reside in poorer U.S. counties. This association is driven by the early settlement of the English who remain disproportionately in the South, coupled with later migrations of Italians to large U.S. cities and the southern out-migration of African Americans to northern cities. The authors discuss pathways through which culture can affect economic outcomes. They also conclude that diversity of ancestries in counties increases growth, as long as cultural attitudes are similar. It is difficult to interpret relationships between ancestry and multiple job holding since we do not regard particular levels of MJH as inherently good or bad outcomes. Areas of the country displaying high multiple job holding rates, in particular the North Central states, have large Nordic and German populations. These high MJH rates may in part reflect a work ethic and cultural values associated with their ancestors. But high MJH rates in these states also reflect low urbanization, low employment turnover, and in some cases weak economies. Cities such as New York, Los Angeles, and Houston, which have among the lowest MSA multiple job holding rates, have diverse populations and vibrant economies. Our analysis clearly points to 21 The ordering of ancestry shares stated in the text is based on model 9. In the appendix, we provide regression results from model 11 (our most dense model). The orderings based on models 9 and 11 differ slightly. 22 This literature is not independent of earlier work in sociology on religion, ethics, and capitalism (Weber 1905). 23

25 strong correlations between ancestry and market-level multiple job holding. Our ability to interpret such correlations, however, is limited. We next introduce a measure of monthly labor market churn (turnover) based on our calculations using all rotation groups of the CPS from September 1995 through December We examine transitions of all individuals ages from the current survey month and prior survey month for six rotation group pairs: 1-2, 2-3, 3-4, 5-6, 6-7, and 7-8 (rotation groups 4-5 are excluded given the 8 month interval between these interviews). For each individual-month pair we measure whether individuals transitioned from not employed to employed (NE), from employed to not employed (EN), employed both months in the same job (EE-same job), employed both months but switched employers (EE-job switch), and not employed in either month (NN). Our measure of monthly churn (turnover) is calculated for each individual and then summed to an MSA measure. Included in the numerator are the number of hires plus the number of separations (NE counts as 1, EN as 1, EE-job switch as 2, and EE-same job and NN as 0), divided by 2 times MSA employment. This measure corresponds closely to the standard turnover measure used in the literature with establishment level data, wherein the numerator is the sum of hires plus separation and the denominator two times employment (e.g., Decker et al. 2014). In studies using quarterly establishment data, a worker leaving one establishment and joining another within a quarter is counted twice in the numerator, whereas those transitioning into and out of employment across quarters are each counted once. As seen in line 10, the introduction of the turnover measure accounts for a modest amount of the dispersion in multiple job holding across labor markets, reducing the MAD measure from 0.35 to MJH regressions show that multiple job holding is negatively related to the level of turnover. This result supports our earlier argument that churn helps lubricate search and enables good primary job matches with respect to hours and other job attributes, implying that good job matches mitigate desire for second jobs. The result is not inconsistent with the alternative argument that second jobs are frequently short-term and that high rates of churn are often associated with short-term jobs. But it is clear from our data that either short-term primary 24

26 jobs differ in some way from second jobs, or the association between short jobs and second jobs is not sufficiently strong to produce a positive relationship between multiple job holding and labor market churn. The final two measures we address are long-run employment growth and unemployment, each of which are added tomodel 11. Log employment growth is measured over the period, calculated from the Quarterly Census of Employment and Wages (with county data aggregated to the MSA level). The unemployment rate is constructed from the Local Area Unemployment Statistics (LAUS) data base provided by BLS. As seen in line 11 of Table 3, the MAD measure remains unchanged at 0.33 after accounting for employment growth and unemployment. Employment growth and the unemployment rate are not strongly correlated across labor markets (e.g., Rappaport 2012). Neither business cycle measure has a substantive effect on multiple job holding, consistent with previous findings in Hirsch et al. (2016). Values of MAD provide simple summary measures of how well the various models perform in accounting for labor market differences in multiple job holding. An alternative way to compare models is to view the entire distribution of the 259 MSA mean residuals from both sparse and dense MJH models previously reported in Table 3. We provide this comparison in Figure 6, based on residuals from MJH model 1 (year and month dummies only), model 4 (model 1 plus the full set of CPS individual-level covariates), and model 11 (model 4 plus all MSA-level covariates). Readily evident is that the distribution of MSA mean residuals from model 11 is far steeper and more concentrated around zero than are the mean MSA residuals from models 1 and 4. As compared to models 1 and 4, model 11 results in a much larger share of MSAs having predicted MJH rates within plus or minus one percentage point of their actual rates. Because changes in MAD are not invariant to the order in which variables are entered, assessing the relative importance of various worker and MSA attributes in accounting for market differences in MJH is not straightforward. In order to judge the relative impact of our measures we estimate the impact of each set of variables on MAD using two alternative methods of analysis, presented in Table 4. The last-in approach measures the impact of each set of variables on MAD when entered last (i.e., the 11 th entry, as was the addition of employment growth and unemployment in Table 3). Being a last entry, the observed reduction in 25

27 MAD represents the effect of unique variation in the added variables, variation uncorrelated with previously entered variables. For example, entering commute times last enables us to observe its effect on MJH differences not already explained by other covariates, some of which (e.g., city size) may be highly correlated with commute time. The first-in approach measures the impact of each set of variables when entered first (apart from year and month dummies). Reduction in MAD from a first-in set of variables reflects the contribution of both unique variation and correlation with other omitted covariates. Entering MSA city size first, for example, captures a large share of the variation in commute times across MSAs. Comparing results from a last-in and first-in analysis is informative and, in this case, reinforces the conclusions reached previously based on the results in Table 3. In the top half of Table 4 we show the seven sets of variables that produce the largest marginal reduction in MAD when entered last. In the bottom half of Table 4 we show the seven sets that produce the largest reduction in MAD when entered first. As it turns out, the same four sets of variables produce the largest MAD in both the last-in and first-in analyses. These are MSA industry and occupation shares, percentage foreign and ancestry shares, mean commute times, and labor market churn (turnover). These same sets of variables were previously determined to be important based on our prior analysis seen in Table 3. In results not shown, we also ordered the importance of variables based both on reductions in the mean squared errors and the increases in R 2 from the various MJH models. The resulting orderings of the MJH covariates were highly similar to those shown in Table 4 based on changes in MAD. 7. Concluding Remarks For many workers and households, multiple job holding provides highly-valued income, human capital accumulation, or worker satisfaction. Multiple job holding may also reflect work hour constraints or poor matches in primary jobs. Although a relatively small share of workers hold multiple jobs at a given point in time, many workers have held multiple jobs at some point in their lives. A largely overlooked labor market pattern is that multiple job holding rates differ substantially across regions, states, and labor markets (MSAs). The relative differences across geographic areas display substantial fixity over the 20 year span of 26

28 our data. We document these persistent difference in regional and market-specific multiple job holding across the U.S. and explore alternative explanations for the differences. The mean absolute deviation in multiple job holding absent covariates, apart from year and month dummies, is 1 percentage point, relative to a mean urban rate over the period of 5.3 percent. Although individual characteristics help explain what types of workers choose to take second jobs, the geographic differences in multiple job holding cannot readily be accounted for by individual worker attributes. Rather, systematic differences in labor market characteristics help us understand the forces leading to differences in MJH rates. Controlling for a broad set of covariates measured at the worker and labor market levels reduces our measure of MJH dispersion by two thirds, indicating substantive progress in understanding the correlates of MJH differences across labor markets. Most important in accounting for multiple job holding differences are MSA-level variables measuring industry and occupation shares, differences in population ancestry shares, commute times, labor market churn, and to a lesser extent, individual level measures of primary job work hours and public sector employment. The city size gradient, which shows far lower levels of multiple job holding in highly-populated markets, appears to be the result of better primary job matching in large markets, lessening the need for second jobs, coupled with longer commute times that make such jobs less attractive. 27

29 References Abraham, Katherine G., John Haltiwanger, Kristin Sandusky, and James R. Spletzer Exploring differences in employment between household and establishment data. Journal of Labor Economics 31, no. 2, pt. 2:S129 S172. Amuédo-Dorantes, Catalina, and Jean Kimmel Moonlighting over the business cycle. Economic Inquiry 47, no. 4: Black, Dan A., Natalia Kolesnikova, and Lowell J. Taylor Why do so few women work in New York (and so many in Minneapolis)? Labor supply of married women across US cities. Journal of Urban Economics 79: Bleakley, Hoyt and Jeffrey Lin Thick-market effects and churning in the labor market: Evidence from US cities. Journal of Urban Economics 72: Connolly, Marie. (2008). Here comes the rain again: Weather and the intertemporal substitution of leisure. Journal of Labor Economics, 26(1), Conway, Karen and Jean Kimmel Male labor supply estimates and the decision to moonlight. Labour Economics 5, no. 2: Decker, Ryan, John Haltiwanger, Ron S. Jarmin, and Javier. Miranda The secular decline in business dynamism in the U.S. Unpublished manuscript, University of Maryland and U.S. Census Bureau. Fulford, Scott L., Ivan Petkov, and Fabio Schiantarelli Does it matter where you come from? Ancestry composition and economic performance of U.S. counties, IZA Discussion Paper No. 9060, May. Gelbach, Jonah B When do covariates matter? And which ones, and how much? Journal of Labor Economics 34, no. 2, pt. 1: Hipple, Steven F Multiple jobholding during the 2000s. Monthly Labor Review (July): Hirsch, Barry T., Muhammad M. Husain, and John V. Winters Multiple job holding, local labor markets, and the business cycle. IZA Journal of Labor Economics 5:4, Hirsch, Barry T., and David A. Macpherson Union membership and coverage database from the Current Population Survey: Note," Industrial and Labor Relations Review 56:2: Data updated annually and accessed at Hirsch, Barry T. and John V. Winters Rotation group bias in measures of multiple job holding. Economics Letters, 147, October: Hyatt, Henry R., and James R. Spletzer The recent decline in employment dynamics. IZA Journal of Labor Economics 2:5, Hyatt, Henry R., and James R. Spletzer. Forthcoming. The recent decline of single-quarter jobs. Labour Economics. Kimmel, Jean, and Karen Conway Who moonlights and why? Evidence from the SIPP. Industrial Relations 40, no. 1: Krishnan, Pramila The economics of moonlighting: A double self-selection model. Review of Economics and Statistics 72, no. 2: Krueger, Alan, Alexandre Mas, and Xiaotong Niu. Forthcoming. The evolution of rotation group bias: Will the real unemployment rate please stand up? Review of Economics and Statistics. Lalé, Etienne The evolution of multiple jobholding in the U.S. labor market: The complete picture of gross worker flows. IZA Discussion Paper 10355, November. Livanos, Illias, and Alexandros Zangelidis Multiple job-holding among male workers in Greece. Regional Studies 46, no. 1:

30 Molloy, Raven, Christopher L. Smith, and Abigail Wozniak Internal migration in the United States. Journal of Economic Perspectives 25, no. 3: Molloy, Raven, Christopher L. Smith, and Abigail Wozniak Declining migration within the U.S.: The role of the labor market. NBER Working Paper Moses, Leon N Income, leisure, and wage pressure. Economic Journal 72, no. 286: Panos, Georgios, Konstantinos Pouliakas, and Alexandros Zangelidis Multiple job holding, skill diversification, and mobility. Industrial Relations 53, no. 2: Partridge, Mark Moonlighting in a high growth economy: Evidence from U.S. state-level data. Growth and Change 33, no 4: Paxson, Christina H., and Nachum Sicherman The dynamics of dual job holding and job mobility. Journal of Labor Economics 14, no. 3: Perlman, Richard Observations on overtime and moonlighting. Southern Economic Journal 33, no. 2: Putterman, Louis and David N. Weil Post-1500 population flows and the long-run determinants of economic growth and inequality. Quarterly Journal of Economics 125, no. 4: Rappaport, Jordan Why does unemployment differ persistently across metro areas? Economic Review, Federal Reserve Bank of Kansas City, Second Quarter, Renna, Francesco, and Ronald L. Oaxaca The economics of dual job holding: A job portfolio model of labor supply. IZA Discussion Paper No Shishko Robert, and Bernard Rostker The economics of multiple job holding. American Economic Review 66, no. 3: Solon, Gary, Steven J. Haider, and Jeffrey M. Wooldridge What are we weighting for? Journal of Human Resources 50, no. 2: U.S. Bureau of Labor Statistics Multiple job holding in states in Monthly Labor Review August, accessed at: htm. Weber, Max The Protestant ethic and the spirit of capitalism (translated into English by Talcott Parsons, 1930). Zangelidis, Alexandros Labour market insecurity and second job-holding in Europe. July. Available at SSRN: Zivin, Joshua G., & Neidell, Matthew. (2014). Temperature and the Allocation of Time: Implications for Climate Change. Journal of Labor Economics, 32(1),

31 MJH Rates Figure 1: Annual Multiple Job Holding Rates for U.S., Metro, and Non-Metro Areas, U.S. Metropolitan Areas Non-Metro Areas 30

32 Figure 2: Quartile Rankings of State Multiple Job Holding Rates, , , and