Abstract
Does working in a low-wage job lead to increased opportunities for upward mobility, or is it a dead-end that traps workers? In this article, we examine whether low-wage jobs are “stepping-stones” that enable workers to move to higher-paid jobs that are linked by institutional mobility ladders and skill transferability. To identify occupational linkages, we create two measures of occupational similarity using data on occupational mobility from matched samples of the Current Population Survey (CPS) and data on multiple dimensions of job skills from the O*NET. We test whether work experience in low-wage occupations increases mobility between linked occupations that results in upward wage mobility. Our analysis uses longitudinal data on low-wage workers from the 1979 National Longitudinal Study of Youth (NLSY) and the 1996 to 2008 panels of the Survey of Income and Program Participation (SIPP). We test the stepping-stone perspective using multinomial conditional logit (MCL) models, which allow us to analyze the joint effects of work experience and occupational linkages on achieving upward wage mobility. We find evidence for stepping-stone mobility in certain areas of the low-wage occupational structure. In these occupations, low-wage workers can acquire skills through work experience that facilitate upward mobility through occupational changes to skill and institutionally linked occupations.
Does working in a low-wage job lead to increased opportunities for upward wage mobility, or is it a dead-end that traps workers? Understanding mobility out of low-wage jobs is important because low-wage workers are a sizable component of the U.S. labor market, representing 25 to 40 percent of the overall labor force, depending on the threshold used to define low wages (Henderson 2022; Ross and Bateman 2019; Schultz 2019), with notable demographic disparities as women, African Americans, and Latinos are substantially overrepresented (Ross and Bateman 2019). Low-wage workers are also highly concentrated in certain occupations (retail, clerical, and personal-service occupations [Escobari, Seyal, and Meaney 2019]) that vary in their average rates of upward mobility (Bihagen and Ohls 2004; Pomer 1984; Schultz 2019) as well as the types of higher-wage occupations to which they lead (Cheng and Park 2020; Escobari et al. 2019; Villarreal 2020). Analyzing the process of upward mobility out of low-wage jobs is crucial for evaluating the magnitude of the problem posed by low-wage employment (Andersson, Holzer, and Lane 2006; Carnevale and Rose 2001) and for designing successful policies that promote advancement (Escobari et al. 2019).
The question of low-wage mobility is a specific case of a larger enduring goal in research on social stratification, which is to explain how careers and intragenerational mobility are shaped by the dynamic interplay between labor market structures and workers’ individual characteristics and resources (Ishida, Su, and Spilerman 2002; Le Grand and Tåhlin 2002; Manzoni, Harkonen, and Mayer 2014; Rosenfeld 1992; Shin 2007; Sørensen 1974). Occupational structure is a central focus in this literature, and social scientists have long sought to identify and explain how occupations are unequally rewarded and linked to each other, thereby providing pathways of potential upward movement (Breiger 1981; Jarvis and Song 2017; Kalleberg and Mouw 2018; Rosenfeld 1992; Sørensen 1974). Recent research on occupational mobility, for example, attempts to detect the underlying network of mobility channels and barriers connecting occupations (Cheng and Park 2020; Lin and Hung 2022; Toubøl and Larsen 2017; Villarreal 2020). However, a key challenge remains: to develop models that can identify the interaction between structural and individual-level factors using data on the career dynamics of workers over time.
The low-wage labor market is a key location to analyze the role of occupational structure on upward mobility because of sharply divided views on the effect of working in a low-wage job, which we characterize as the “dead-end” and “stepping-stone” perspectives. Both of these perspectives have to make sense of the fact that a substantial minority of workers do move out of low-wage jobs (Boushey 2005; Carrington and Fallick 2001; Dickens 2000; Long 1999; Vornovytskyy 2011), even in the fast-food industry (Connolly, Gottschalk, and Newman 2003; Newman 2006), and that the rate of mobility differs based on individual characteristics such as education, experience, race, and gender (Even and Macpherson 2003; Schultz 2019). The question is who moves up and why.
The “dead-end” perspective argues that work experience in low-wage jobs does not lead to the accumulation of skills that can be transferred to jobs in higher-paying occupations (Grün et al. 2009; Knabe and Plum 2013; Truss et al. 2013). This perspective argues that dead-end jobs are disconnected from the rest of the occupational structure because of the absence of skill-based mobility pathways connecting low-wage to higher-wage occupations (Escobari, Seyal, and Contreras 2021). According to this view, the upward mobility that does occur out of low-wage work is not due to skill development but results from a sorting process where workers are hired for higher-wage jobs based on their inherent productivity (Jovanovic and Nyarko 1997) or due to discrimination based on demographic or status characteristics (Pager and Karafin 2009).
The “stepping-stone” perspective, in contrast, argues that some low-wage jobs promote the accumulation of skills that can be transferred to higher-wage occupations (Dahl, DeLeire, and Schwabish 2009; Scherer 2004; Van den Berg, Holm, and Van Ours 2002), by means of career ladders or mobility pathways that connect low-wage jobs to higher-wage ones (Escobari et al. 2021; Fitzgerald 2006). Conceptually, these pathways build on the idea of occupational internal labor markets (OILMs), where skill development, job-ladders, and closure mechanisms create cross-firm mobility clusters linking several occupations (Althauser and Kalleberg 1981; Smith 1983). The OILM perspective is complemented by a recent literature in labor economics that argues workers develop task-specific skills that can be transferred to occupations with similar task or skill requirements (Gathmann and Schönberg 2010; Gibbons and Waldman 2004).
In this article, we examine the interaction between worker characteristics and the occupational structure in explaining the wage and occupational mobility of low-wage workers. Our analyses are based on longitudinal data from the 1979 National Longitudinal Survey of Youth (NLSY) and four panels of data from the Survey of Income Program and Participation (SIPP). We use the NLSY because it provides complete career histories of a cohort of workers. In contrast, the large sample size of the combined SIPP panels allows us to analyze more precisely variation in upward mobility at the occupational level.
Our analysis proceeds in two stages. First, we analyze the effect of work experience in low-wage jobs on the rate of upward wage mobility using a logit model to adjudicate between the competing perspectives on mobility out of low-wage jobs: the dead-end perspective contends there is no effect of work experience on mobility, whereas the stepping-stone perspective argues the effect will be positive. Second, we analyze how variables measuring aspects of the occupational structure moderate the process of upward wage mobility that we observe in the first stage. We estimate multinomial conditional logit (MCL) models (Breen 1994) where characteristics of origin and destination occupations are included in the analysis and the data structure allows us to include all occupations as potential destination occupations for each worker. In particular, our models include cross-level interaction terms between individual low-wage work experience and measures of skill-transferability and institutional linkages between occupations. This enables us to test the structural mechanism predicted by the stepping-stone perspective. We test whether low-wage work experience leads to increased rates of upward wage mobility when transferred along a mobility pathway to a linked occupation.
Our analyses build on a structural perspective of the labor market based on the concept of OILMs (Althauser and Kalleberg 1981), where the skill and institutional linkages between occupations provide the pathways for the upward mobility of low-wage workers. We identify the pathways using unique measures of skill-transferability and institutional linkages between occupations, and we demonstrate that these pathways help explain the pattern of upward mobility that unfolds over low-wage workers’ careers.
Stepping-Stone Jobs and the Occupational Structure
An important structural argument for why low-wage jobs are “dead-ends” or “traps” comes from the literature on dual and segmented labor markets (Doeringer and Piore 1971; Leontaridi 1998). The dual labor market approach divides the labor market into a “primary” labor market consisting of good jobs and a “secondary” labor market with bad jobs. According to this view, secondary labor market jobs pay low wages and do not provide much on-the-job training or skill development, making the experience–wage profile relatively flat (Dickens and Lang 1985). This results in fewer opportunities for upward mobility compared to jobs in the primary labor market (Haller et al. 1985). This segmentation perspective is frequently cited as an explanation for why low-wage or temporary jobs may be “dead-ends” that constrain upward mobility (Bukodi and Dex 2010; Fuller and Stecy-Hildebrandt 2015; Kovalenko and Mortelmans 2014; Scherer 2004, 2005; Toledo, Pablo, and Usabiaga 2014). Nonetheless, evidence of considerable mobility between segments belies the dual labor market perspective (Boston 1990; Cain 1976; Jacobs 1983), suggesting that, at the very least, there are multiple segments that vary in the extent to which they are connected to higher-wage sectors (Hudson 2007).
Occupations are often used to distinguish labor market segments (Sacchi, Kriesi, and Buchmann 2016; Spenner, Otto, and Call 1982; Spilerman 1977; see Kalleberg and Mouw 2018). As descriptive evidence of the degree to which the low-wage labor market is segmented at the occupation level, Table 1 presents a summary of selected low-wage occupations in terms of their size and likelihood of upward mobility using data from the 2010 to 2019 Current Population Survey (CPS). The variable %Low in column 1 shows the proportion of low-wage workers (defined as earning less than $13 per hour in 2019 dollars) in the occupation, and the size variable in column 2 shows the percentage of all low-wage workers in the labor market who are in the occupation. Columns 8 and 9 show overall rankings (out of 500 three-digit occupations) based on the proportion and size measures. Dishwashers, for example, rank high in proportion of low-wage workers (#2) but low in terms of size (#32) compared to large occupations such as retail sales, janitors, and nurse’s aides.
Top Low-Wage Occupations, Current Population Survey Data 2010 to 2019
Note: The table uses the three-digit Census occupation codes for 2000. A full list is available at https://usa.ipums.org/usa/volii/occ2000.shtml. We use a definition of low wages as earning less than $13 per hour (in 2019 dollars). Between 2010 and 2019, 26.5 percent of the workforce earned less than $13 per hour, on average (authors’ calculations based on CPS data). CPS respondents are interviewed four consecutive months in the same calendar year, then they are off for eight months, and then interviewed for four more consecutive months in the next year. Wage questions are asked in the 4th and 8th month of interviews, allowing for one-year measures of wage changes by matching individuals across years. Full details of this CPS analysis are included in Part A1 of the online supplement. The variables in Columns 4 to 7 use CPS data from 2000 to 2019 (rather than the 2010 to 2019 data used in Columns 1, 2, and 3) to provide a larger sample size of mobility at the occupational level.
% Low is the proportion of workers in the occupation earning low wages (less than $13 per hour in 2019 dollars).
Size is the percent of all low-wage workers in the labor force who work in the occupation.
This is the average wage.
The monthly rate of occupational mobility.
The proportion of occupational moves that are to occupations with average wages >$16 per hour.
The one-year upward wage mobility rate (>$13 per hour).
Estimated occupation random effect on a linear probability model of upward mobility (see Part A of the online supplement for more details).
Rankings (out of 500 three-digit occupations): %Low = ranking based on %Low, size = based on size.
The CPS uses dependent coding of occupations for successive monthly interviews in the same calendar year, which substantially reduces measurement error in occupational mobility. See Forsythe (2018) and Moscarini and Thomsson (2007) for details and similar results.
Column 4 (%Occ. mob.) of Table 1 shows the monthly rate of mobility out of the occupation. On average, for example, 3.5 percent of cashiers change occupations each month, compared to 1.6 percent of truck drivers and 2.2 percent of nurse’s aides. Projected over 12 months, this shows enough occupational mobility to reject the notion that low-wage workers are literally “stuck” in their jobs, but it does not rule out the possibility that this mobility just reflects circular movement among a set of related low-wage occupations. Column 5 (%Up. Occ.) gets at this question by showing the proportion of movers whose destination occupation had an average wage of at least $16 per hour—for example, 23.5 percent of dishwashers and 52.6 percent of cashiers who changed occupations moved to higher-wage occupations based on this measure. Overall, columns 4 and 5 of Table 1 show there is substantial mobility out of some low-wage occupations, consistent with critiques of the segmented labor market perspective based on cross-segment mobility (Jacobs and Breiger 1988). Nonetheless, the limitations of the segmentation perspective do not rule out the possibility that there are structural effects of low-wage occupations on workers’ probability of achieving upward wage mobility, which is the focus of this article.
Additive Effects Models of Occupational Structure
A common approach used to assess occupation-level structural effects on wage mobility is to estimate an additive model with occupational characteristics or occupation dummy variables controlling for individual demographic and human capital characteristics. The coefficients on the occupational characteristics or dummy variables are then interpreted as the structural effects of occupations. Additive effects models of occupational structure are a specific case of the model for structural effects discussed by Farkas, England, and Barton (1988:94). Existing research using occupation dummy variables has found evidence of significant occupation-level variation in the upward mobility of low-wage workers controlling for individual characteristics (Even and Macpherson 2003; Mosthaf, Schnabel, and Stephani 2011; Pomer 1984; Schultz 2019). Columns 6 and 7 of Table 1 provide similar evidence using data on wage mobility from individuals matched across successive years of the CPS. Column 6 shows the overall proportion of upward mobility ($Up. > 13) for low-wage workers by occupation: 8.6 percent of low-wage cashiers, for example, earned more than $13 per hour a year later, compared to 14.3 percent of retail sales workers and 13.2 percent of janitors. Column 7 of Table 1 shows the results of an additive multilevel linear probability model of upward wage mobility for low-wage workers; 1 we report point estimates of occupation random effects controlling for human capital and demographic characteristics. 2
Interpreting the results of additive effects models with occupational characteristics, dummy variables, or random effects as the “effect” of these occupations on mobility is problematic, however, and open to criticism from both the human capital and occupational internal labor markets (OILMs) perspectives. Human capital theorists emphasize the role that variation in workers’ human capital and productivity-related characteristics plays in explaining why some workers move up (Carrington and Fallick 2001; Even and Macpherson 2003); they argue that the observed level of upward mobility reflects a sorting process where workers with more desirable characteristics (from the employer’s perspective) experience upward mobility, leaving “less desirable” workers behind (Jovanovic and Nyarko 1997). According to this view, the sorting process includes characteristics that are observed by employers but not by researchers, such as productivity-related traits like motivation or “ability” (Le Grand and Tåhlin 2002), as well as the effect of discrimination based on aspects of social distance not captured by categories of race, ethnicity, and gender recorded on conventional surveys and included in typical analyses (see Monk 2022).
The possibility that upward mobility at least partially reflects sorting on unobserved individual characteristics is a source of frequent criticism of the evidence on dual and segmented labor market perspectives (Cain 1976; Lang and Dickens 1988; Leontaridi 1998), and it means the estimates of occupation-level effects from additive effects models will be biased if unobserved individual factors affect the sorting of workers into and out of low-wage occupations (Dickens and Lang 1993). For example, a high rate of upward mobility of individuals who are employed as waiters could be due to the accumulation of skills in that occupation (i.e., the “effect” of working as a waiter), or it might be due to the fact that many young people tend to work temporarily in these low-wage occupations at the beginning of their careers (Abbott 2004). Thus, a high mobility rate could simply reflect the transitory movement of many workers through brief spells as a waiter, rather than something beneficial about working as a waiter per se. Consequently, variation in the rates of upward mobility across low-wage occupations in these additive models may reflect individual characteristics and not structural effects.
Career Lines, Sequence Analysis, and Network Approaches
The additive effects approach to studying structure has also been criticized for not considering job changes within career ladders or mobility pathways between detailed occupations (Lin and Hung 2022). The idea of career ladders is derived from research on occupational work histories (Sacchi et al. 2016; Spenner et al. 1982; Spilerman 1977; see also Kalleberg and Mouw 2018). In a foundational analysis, Spilerman (1977) used synthetic cohorts constructed from 1970 Census data to analyze trajectories of occupational careers, identifying “portals” as entry-level positions that are common starting points on career lines, and “dead-ends” as locations on a career line that lead to few higher-level positions. Spenner and colleagues (1982) showed that mobility rates between occupations vary greatly; some occupations are strongly connected by high rates of mobility and others are not (see also Sicherman 1993). In certain low-wage occupations, upward mobility might involve climbing to the next rung of a career ladder, and some higher-wage occupations may serve as points of entry and absorb relatively large numbers of upwardly mobile workers. However, studying career lines using detailed occupational categories can result in overwhelming complexity, leading some to suggest the categories should be aggregated (Kerckhoff 1995). In the process of aggregation, however, key details about the connections between specific occupations are likely to be lost.
The use of sequence analysis to analyze career histories is similar to the career-lines approach, in that it attempts to find common patterns of occupational mobility using longitudinal data (Abbott 1995; Aisenbrey and Fasang 2017; Biemann, Zacher, and Feldman 2012; Blair-Loy 1999; Dlouhy and Biemann 2015; Fuller and Stecy-Hildebrandt 2015; Halpin and Chan 1998; Joseph et al. 2012; Kovalenko and Mortelmans 2014). For our purposes, this is promising because it might be possible to identify significant links between occupations associated with stepping-stone mobility. Fuller and Stecy-Hildebrandt (2015), for example, use sequence analysis to test a stepping-stone model of the mobility of low-wage temporary workers in Canada, and they find patterns of mobility missed by studies that focus on single spells of temporary work. Nonetheless, sequence analysis remains essentially descriptive, as it is hard to model the conditional probabilities of individual and structural factors on mobility as they unfold over time (Hollister 2009; Manzoni et al. 2014; Wu 2000), making it difficult to adjudicate between the stepping-stone and dead-end perspectives on the mobility of low-wage workers.
An alternative empirical approach to identifying mobility pathways comes from research on occupational mobility networks, which uses clustering algorithms to detect features of occupational structure based on the mobility rates between pairs of occupations (Cheng and Park 2020; Lin and Hung 2022; Toubøl and Larsen 2017; Villarreal 2020). 3 This research has a strong affinity with one of the approaches we take here to identify occupational similarity (our key structural variable) based on relative mobility rates between occupations (Toubøl and Larsen 2017). In an occupation-level analysis of the mobility of low-wage workers, for example, Escobari and colleagues (2021) use a network approach to identify “skyway” occupations that connect low-wage workers to higher-paying sectors of the labor market. A criticism of the network approach is that it abstracts away from the mobility process of individual workers because the analysis is based on aggregate flows of workers between occupations, and thus it cannot incorporate individual-level characteristics into the analysis except as the average characteristics of (sending or receiving) occupations themselves (e.g., Lin and Hung 2022).
OILMs and Task-Specific Human Capital
A difficulty with all the approaches discussed above is the failure to specify the structural mechanisms that facilitate or inhibit the translation of work experience into upward mobility. This gap is filled by the theory of internal labor markets, which emphasizes how job ladders within firms or occupations enable workers to acquire greater levels of training and experience, thus allowing them to progress upward on ladders characterized by the progressive development of skills and knowledge (Althauser and Kalleberg 1981).
The concept of occupational internal labor markets (OILMs) incorporates institutional features of occupations or sets of occupations that account for why the incumbents of some occupations experience upward mobility without necessarily being tied to a particular firm (e.g., Eyraud, Marsden, and Silvestre 1990). According to this view, some occupations are able to establish mechanisms of social closure (e.g., via union practices or licensing; see Redbird 2017; Weeden 2002) that restrict competition from non-occupational members and facilitate upward mobility within the occupation or the transfer of skills between “families” of occupations that enable occupational incumbents to acquire greater skills through work experience. Groups or clusters of occupations that are linked in sequences permit people to acquire skills that may be specific to these contexts. Organizational, occupational, and economic sociologists, as well as labor economists, tend to adopt this view and assess the effects of movement within as opposed to between employers 4 or occupations as sources of upward trajectories in earnings (see Kalleberg and Mouw 2018). The theory of OILMs has also been used by occupational network analysts to explain the underlying clustering of occupations (Lin and Hung 2022; Villarreal 2020). Our analysis builds on this perspective by identifying the linkages between occupations based on skill similarity and observed patterns of mobility.
The recent literature in labor economics on task-specific human capital draws conclusions on the mobility links between occupations that are similar to those based on OILMs. Although the standard human capital model is often interpreted as treating workers’ education and training as a single dimensional measure of skill-related productivity (Mincer 1974), research on task- and occupation-specific human capital suggests a more realistic way of thinking about how the accumulation of specific skills increases the mobility rate along skill-based ladders between occupations (Gathmann and Schönberg 2010; Gibbons and Waldman 2004; Kambourov and Manovskii 2009). The basic idea is that skill development is often quite specialized and may transfer only to occupations that have similar task and skill requirements. For instance, the skills that make someone a successful courtroom lawyer will not directly translate into success as a heart surgeon. Studies by labor economists build on this notion of task-specific human capital to model careers as a process of skill formation (Gathmann and Schönberg 2010; Gebicka 2010; Holmes and Tholen 2013; Pavan 2011; Toledo et al. 2014; Yamaguchi 2010, 2012).
Conceptually, task-specific human capital operating at the individual level is consistent with the idea of patterns of occupational mobility at the structural level in the sociological literature, described in research on career lines and occupational internal labor markets (Althauser and Kalleberg 1981; Spenner et al. 1982; Spilerman 1977). The idea of OILMs complements the task-specific human capital models by identifying and explaining the structures within which task-specific skills are created (and about which studies on task-specific skills are relatively silent). Combining the task-specific skills and career lines approaches allows us to see how low-wage workers’ mobility may result from the interplay between individual and structural effects: working in a particular occupation enables one to accumulate specific skills; but the occupational division of labor, the skill relatedness among occupations, and the channels or barriers affecting mobility between pairs of occupations are all structural factors that affect whether workers’ accumulated skills facilitate their upward mobility in occupational labor markets.
Hypotheses
Based on the preceding review of the literature, we argue that the effect of low-wage work experience—that is, the effect of the length of time an individual has spent working in low-wage jobs—can help us adjudicate between the dead-end and stepping-stone perspectives on mobility. We define a stepping-stone occupation as one in which there is skill growth over time that can be transferred to related, higher-paying occupations. This will result in an increased rate of upward wage mobility as occupational experience increases. In contrast, the “dead-end” perspective claims there is little accumulation of transferrable experience-related skills in low-wage jobs, arguing that the upward mobility that does occur is largely the result of a sorting process. In this section, we outline two hypotheses related to how the interplay between low-wage work experience and the occupational structure affects upward wage and occupational mobility.
Work Experience Hypothesis
As discussed above, a criticism of the additive effects approach is that the estimated coefficient could represent either the effect of the occupation on upward mobility or the result of sorting on unobserved individual characteristics. This makes it difficult to adjudicate between the stepping-stone and dead-end perspectives on upward mobility. For example, a positive coefficient on the dummy variable for waiters could be the result of the positive “effect” of working in the occupation (consistent with the stepping-stone perspective), or a reflection of the higher unobserved ability of people working in it (consistent with the dead-end perspective). In contrast, Equation 1 depicts a model of the occupation-level effect of work experience on upward mobility with diverging predictions from these two perspectives. Let Y be a 1/0 measure of upward wage mobility, and p(Y = 1) represent the probability that worker i experiences upward mobility at time t:
where αj is a constant term for occupation j, Xi represents a set of observed individual variables such as education and training as well as basic demographic variables, and expij represents the length of the spell in the low-wage job in occupation j. In Equation 1, the coefficient on low-wage work experience is a measure of duration dependence. If then we have positive duration dependence—meaning the wage mobility rate is increasing with experience; the reverse is true if . Finally, let vi indicate an error term representing unobserved heterogeneity among workers—unobserved factors that might affect the probability of upward mobility such as ability or ambition.
Using Equation 1, we can formulate our first hypothesis based on the effect of work experience on upward mobility, using the sign of to differentiate between dead-end and stepping-stone jobs:
Hypothesis 1: According to the stepping-stone perspective, low-wage work experience increases the rate of upward wage mobility.
According to this work experience hypothesis, the stepping-stone view for a particular low-wage occupation j would be supported if the accumulation of job skills in occupation j increases the chances of upward mobility over time, resulting in a positive effect of occupational work experience, so . In contrast, according to the dead-end perspective, low-wage jobs are traps where there is no skill accumulation that can be transferred to other, higher-paying occupations. According to this view, any mobility that does occur is the result of a sorting process where workers with preferred human capital levels and demographic traits (the Xi term in Equation 1) and higher levels of unobserved ability (the vi term) have higher rates of upward mobility. 5 According to the dead-end perspective, however, there is no skill-building benefit to low-wage work experience, so . 6 Baert, Cockx, and Verhaest (2013), Scherer (2004), Cockx and Picchio (2012), Gash (2008), Pavlopoulos and Fourage (2010), and Knabe and Plum (2013) provide similar descriptions of dead-end and stepping-stone effects.
Mobility Pathways Hypothesis
According to the theoretical model of stepping-stone mobility based on OILMs and task-specific human capital, work experience is beneficial because it increases the likelihood a worker will obtain higher wages by changing jobs or occupations that have similar task requirements within an OILM. For the purposes of our analysis, we call this OILM linkage or task similarity a mobility pathway, that is, a pair of occupations—origin occupation A and potential destination occupation B—with a high degree of skill transferability between them, as measured by the continuous occupational similarity variable θab, which we discuss in the Data section. We use the terms mobility pathway and “linked” occupation interchangeably to indicate occupations with high levels of θab.
Our mobility pathways hypothesis (Hypothesis 2) argues that the effect of work experience on upward wage mobility between occupations A and B will be moderated by the degree to which accumulated occupational-specific experience can be transferred between them. We test this by including an interaction term we call the “experience × occupational-similarity interaction” (EOI) = expia × θab in our models. A positive coefficient on the EOI term indicates that the effect of work experience on the rate of upward mobility from A to B increases as the strength of the θab linkage increases. In our analysis, we calculate separate measures of EOI for each worker for all potential destination occupations B, and we use a discrete choice model to analyze the effect of variation in EOI on the probability of upward mobility to these different occupations. This results in our second hypothesis, which has two parts:
Hypothesis 2: According to the stepping-stone perspective: (1) the coefficient on the EOI term will be positive, , indicating that the effect of experience on upward mobility is moderated (with a positive slope) by the occupational similarity measure θab, and (2) the combined effect of experience—including the baseline effect and the interaction term—on upward mobility between linked occupations will be positive.
To test Hypothesis 2, we need two things. First, we need a definition of a mobility pathway between linked occupations based on the similarity measure θab, which we discuss in detail in the Data section. Second, we need a statistical model that will allow us to jointly analyze the effect of occupational experience on wage and occupational mobility, which is what we turn to next.
Methods and Analysis Plan
Table 2 outlines our analysis plan, which connects Hypotheses 1 and 2 to the data and methods we use to test them. Column 1 of Table 2 shows the statistical model we use, and column 5 describes predictions based on the relevant hypothesis. In our analysis, we use longitudinal data from the NLSY and the SIPP, which allows us to model the process of wage and occupational mobility for low-wage workers across their careers. In this section, we discuss the methods first because it affects our detailed discussion of the data to follow.
Analysis Plan Illustrating the Connection Between Our Statistical Models, the Hypotheses, and the Data
The “experience x occupational-similarity interaction” (EOI) is an interaction term between the strength of the mobility pathway (θab, the occupational similarity measure) and the worker’s low-wage occupational experience (EOI = EXP × θab) (see Equation 6).
is the measure of occupational similarity, which indicates the strength of the mobility pathway. We define a “linked” occupation as a potential destination occupation where θab ≥ 0.8 for the purposes of testing the combined effect of EXP and EOI in the MCL models. See Tables 4 and 5 for examples of linked occupations. However, because θab is a continuous variable between 0 and 1, the linkage between occupations in the models is a matter of degree.
As we discuss in the NLSY Data section, the set of potential destination occupations includes the worker’s actual destination occupation plus a random sample of 20 unchosen occupations. The sampling of unchosen alternatives is a central feature of the conditional logit model (Bruch and Mare 2012), and we select a different random sample of unchosen occupations for each worker at each time period. Compare Models 2A and 3A in Table O1.5 of the online supplement for an illustration of how sampling from the unchosen alternatives does not affect the results of MCL models.
In Row 1 of Table 2, we depict the use of a logit model to test the effect of low-wage work experience on upward mobility (Hypothesis 1) based on Equation 1. The logit model has random effects for individuals and dummy variable fixed effects for origin occupations. 7 The logit model is useful for testing Hypothesis 1, but it is limited in that it cannot analyze occupational mobility (except as a single 0/1 variable) or incorporate features of the occupational structure in the analysis (except as aggregate measures that vary among individuals rather than across potential destination occupations). These limitations are important because they prevent the logit approach from being able to test Hypothesis 2, about the role of occupational mobility along mobility pathways (i.e., the strength of the θab linkages between occupations) on upward wage mobility.
We use a multinomial conditional logit (MCL) model to overcome the limitations of the logit model because it allows us to incorporate features of the occupational structure directly in the analysis. The MCL is best explained as a combination of a logit model (0/1 for upward wage mobility) and a conditional logit (for occupational mobility). In the existing literature, researchers have used conditional logit models (CL) to analyze occupational or geographic mobility. A key feature of the CL model (and our corresponding MCL model) is that the data consist of a set of potential choices for each person (i.e., a “choice set” of potential destination neighborhoods or occupations), including choices that were and were not actually selected (Boskin 1974; Bruch and Mare 2012; Davies, Greenwood, and Li 2001; Quillian 2015). This allows the analysis to include destination-specific variables (e.g., the average wage of occupations or the racial composition of neighborhoods) as explanatory variables to predict which possibility was chosen.
In a recent example, Hsiung (2020) uses CLs to analyze the effect that characteristics of potential destination occupations (e.g., the occupation’s size, average wages, and detailed skill requirements) have on differential mobility rates of workers by gender. Nonetheless, a shortcoming of conventional applications of the conditional logit model is that they are conditional on mobility—that is, they are estimated on the subset of workers who change occupations, and thus the analysis excludes workers who are not mobile. 8 Furthermore, it is difficult for CLs to jointly analyze individual wage and occupational mobility. Although upward wage mobility in CLs can be proxied through moving to an occupation with a higher average wage, the large wage dispersion within low-wage occupations (Hunt and Nunn 2019) means that moving to a higher-wage occupation is no guarantee of individual wage mobility.
The MCL approach in row 2 of Table 2 is an extension of the CL model that allows us to model wage and occupational mobility simultaneously and include all workers—mobile and nonmobile—in the analysis. Our use of MCLs builds on earlier work that describes these models as a way to include categorical and continuous explanatory variables in the analysis of occupational mobility tables (Breen 1994; DiPrete 1990; Logan 1983), combining the benefits of regression-based status attainment models with the detail on origin and destination mobility tables from a loglinear approach (Hendrickx and Ganzeboom 1998; Treiman and Ganzeboom 2000; Wanner 2005). 9 The MCL has also been called a conditional multinomial logit (Dessens et al. 2003; Erola, Härkönen, and Dronkers 2012) or a “mixed model” (Hoffman and Duncan 1988; Powers and Xie 2000); the name itself is perhaps a misnomer, because any multinomial logit can be fit by an appropriately specified conditional logit model (Hendrickx 2000).
Tables 3a and 3b describe the data structures for our logit and MCL models. Table 3a shows the structure of the logit model, and Table 3b illustrates the data structure of our MCL model, which allows us to combine the analysis of wage and occupational mobility. The data structure of the MCL is similar to a conventional CL model (e.g., Bruch and Mare 2012; Quillian 2015) in the sense that each row in Table 3b represents a possible outcome or destination that a particular individual could wind up in. For example, rows 2 to 5 represent specific destination occupations that worker 1 could move to, and characteristics of those occupations (columns 5 and 8) are key variables in the analysis. The difference between the MCL and CL models is that in Table 3b the MCL adds an additional row to the data (row 1) that indicates staying in low wages (D1 = 0 in column 4). Aside from row 1, the other rows represent combinations of a specific destination occupation (column 2) with upward wage mobility (column 4), and the dependent variable in column 3 is whether a particular row was the outcome the specific worker experienced. As shown in Table 3b, worker 1 achieves upward mobility in occupation 2, so the dependent variable is 1 in row 3 and 0 in the rest of the rows of column 3.
Example of Data Structure for a Logit Model
Note: In this case, individual ID 1 achieves upward wage mobility (Y = 1) and has an experience value of 2. For the wage mobility outcome Y, 0 = low wages, 1 = upward mobility. The logit model has one case per person per wave. The dependent variable (Y) is a 1/0 variable indicating upward wage mobility. The key independent variable is log low-wage experience (EXP).
Example of the Data Structure for the MCL Model with Four Possible Destination Occupations
Note: D1 is a dummy variable indicating the Y = 1 category of the wage mobility outcome variable Y, and “*” indicates an interaction term. In this case, individual ID 1 achieves upward wage mobility (Y = 1 in column 4) and moves to destination occupation 2 (column 2). Column 3 is the dependent variable (DV), which is 1 for the outcome that actually happened and 0 otherwise. The MCL models have a different data structure than the logit models. Each “case” (i.e., each row) represents a potential outcome for individual ID 1. The row where the DV = 1 in column 3 is the outcome that actually occurred. In this example, there are four potential destination occupations (rows 2 to 5). For each wave, there is a case for each person for upward wage mobility to each potential destination occupation (rows 2 to 5) plus a case to indicate the possibility of remaining in low wages (row 1). The low-wage outcome includes staying in low wages but changing occupations (to include multiple low-wage jobs in the duration of the low-wage spell). In our analysis, for each person and time period, we randomly sample 20 occupations from the set of unchosen potential destination occupations. Our conditional logit models in Table D1 in the online supplement have the same structure as the MCL model here, except they only include upwardly mobile workers (D1 = 1 in row 4), and they do not have row 1.
The occupational similarity measure (θab) in Row 5 is a continuous 0–1 variable that measures the degree of similarity between origin occupation A and destination occupation B.
The experience-occupational interaction (EOI) in row 8 is an interaction term between θab in row 5 and low-wage experience in row 7.
Our setup of the MCL allows us to jointly model wage and occupational mobility because the dependent variable (in column 3 of Table 3b) is a combination of both variables. To see this, note that we start by setting staying in low wages (D1 = 0 in column 4) as the baseline category and include dummy variable interactions between the dichotomous wage mobility variable and all the explanatory variables (i.e., the interaction terms with the dummy variable for wage mobility, D1, in columns 7 and 8 of Table 3b). As a result, we are modeling not the probability of winding up in occupation B per se, but the joint probability of an individual being in occupation B and having high wages (D1 = 1). Importantly, all workers are in the analysis: we include workers who stay in the same occupation while achieving upward wage mobility (in which case their destination occupation is the same as their origin occupation) and workers who do not achieve wage mobility (i.e., nonmobile cases in row 1 with D1 = 0).
The MCL approach adds more complexity to our analysis than the logit approach. However, in row 3 of Table 2 we argue that the MCL is a better approximation to the combination of choice and constraint that characterizes the way individual workers actually experience job search and mobility (e.g., Logan 1996): workers apply for jobs in specific occupations, and the characteristics of those occupations, combined with individual-level characteristics, affect the probabilities of which occupation(s) (if any) could provide a worker upward wage mobility. 10 The MCL approach is flexible, and the data structure described in Tables 3a and 3b could be expanded in future research to include job openings in firms nested within the broader context of occupational labor markets using administrative worker–employer datasets (see Eliason et al. 2023), as well as including other aspects of the occupational structure such as gender composition and gender queues (Reskin 1991).
The MCL model is formally depicted in Equation 3, where we are estimating the probability Piab that individual i achieves upward wage mobility (Y = 1) while moving from occupation A to occupation B. The MCL combines a logit model of wage mobility with a CL model of occupational mobility; in the MCL, the probability of experiencing both upward wage mobility and moving to occupation B is modeled as a function of a set of individual characteristics xi (e.g., gender) and a set of variables ziab that vary across the possible destination occupations (and which may be specific to the pair of occupations A and B), and where J represents the set of possible destination occupations (see Powers and Xie 2000:243):
In Equation 3, the logit part of the MCL consists of the terms with β coefficients along with the individual random effect ui, and the conditional logit component consists of the terms with α coefficients. In particular, the terms
An additional feature of our methodological approach is the inclusion of a nonparametric random-effect term ui in the MCL model to incorporate individual heterogeneity in the rate of upward wage mobility using a latent class conditional logit approach (LCCL, Train 2008). In contrast to parametric random effects that are assumed to follow a normal distribution, nonparametric random effects are less sensitive to misspecification in the underlying distribution of the random effects (Agresti, Caffo, and Ohman-Strickland 2004). Part C of the online supplement describes the LCCL approach in detail, and online supplement O1 provides an extended example of estimating MCL models and an illustration of the approach’s ability to estimate the effects of occupational structure on mobility using individual-level simulated data.
Unobserved Heterogeneity
A substantively important methodological challenge in testing Hypotheses 1 and 2 is the presence of the unobserved heterogeneity term vi in Equation 1, which represents anything that affects mobility but is unobserved by the researcher. 11 In our analysis using the NLSY, we use models that incorporate random effects to allow for variation in unobserved heterogeneity (vi) across individuals, which builds on an existing literature on modeling hazard rates in morbidity and mortality with random effects using frailty models (Balan and Putter 2020; Hougaard 1995). As described in detail below, all our models also include occupational fixed effects. Because the NLSY data allow us to follow complete spells of low-wage work, we do not have the “initial conditions” problem (Heckman 1981) encountered by researchers using data with ongoing spells that began at an unspecified point in the past, as is often the case with state-dependence models, which often only have data on the most recent period of time (e.g., Cappellari and Jenkins 2004).
For our analysis, an important consideration in testing Hypotheses 1 and 2 is that unobserved heterogeneity will tend to cause downward bias in estimates of duration dependence (i.e., the effect of work experience during the low-wage spell, βEXP). This is because individuals with higher levels of unobserved heterogeneity (vi in Equation 1) are more likely to experience upward mobility first, which causes the expected level of vi among the remaining low-wage workers to decline over time. This is well established in the literature (Blossfeld and Hamerle 1989; Chamberlain 1979; Heckman 1985; Jost 2022; Lancaster 1990; Lopes 2021; Nicoletti and Rondinelli 2010; Trussell and Richards 1985; Van den Berg 2001; Van den Berg and Van Ours 1996; Zorn 2000). In general, evidence of negative duration dependence is widespread in sociological research on upward mobility and promotions (Becker and Blossfeld 2017; Bukodi and Dex 2010; Holmes and Tholen 2013; Hultin 2003; Ishida et al. 2002; Sacchi et al. 2016). Van den Berg (2001:3407) refers to negative duration dependence as the result of a process of “weeding out” or “sorting” that occurs in duration models, and Blossfeld and Hamerle (1989:131) note that “failure to control for unobserved variables leads to a bias towards negative duration dependence.” 12
While it is important to be cautious regarding claims about the effect of unobserved heterogeneity in general, we argue that in this specific case, unobserved heterogeneity that is not captured by our random effects will tend to cause a downward (negative) bias in our estimates of the effect of work experience, resulting in more conservative tests of the expectation of a positive effect of work experience on mobility in Hypotheses 1 and 2. The expectation of downward bias on estimates of duration dependence (i.e., the coefficient on βEXP) can be contrasted with the indeterminate (+ or –) effect of unobserved heterogeneity on the coefficients on occupational dummy variables (discussed earlier in the review of the additive effects approach) and on the coefficient on the occupation-specific constant term (αj) in Equation 1. 13 Part A2 of the online supplement provides computer code to allow readers to simulate data and explore this result in more detail by testing the effect of unobserved heterogeneity on estimates of duration dependence in models with and without random effects. 14
Correlated Random Effects
In our analysis of the SIPP data in Tables 10 and 11, we use a correlated random effects (CRE) approach (Bianconcini and Bollen 2018; Rabe-Hesketh and Skrondal 2013; Schunck 2013) to account for the problem posed by unobserved heterogeneity in Equation 1 with left-censored low-wage spells that would otherwise cause problems for a conventional random-effects approach. As we will discuss, the SIPP data include ongoing low-wage spells (and miss spells that started and finished in the past), and the NLSY data include all low-wage spells. Our approach builds on a recent literature on the use of correlated random effects for dynamic nonlinear panel models with state dependence (Cutuli and Grotti 2020; Klee 2013; Plum and Knies 2019; Wooldridge 2005, 2013).
In terms of Equation 1, the problem for unbiased estimation of the effect, βEXP, with incomplete work history data is that some low-wage spells are ongoing when the researcher first observes the data, resulting in left-censoring. This means the ongoing spells are a selective sample of spells that began at a particular time in the past, as workers with higher levels of vi are more likely to have exited from a low-wage spell prior to the first period of data. This results in a negative correlation between vi and the length of the spell at the first observation of the data (
where
Part A3 of the online supplement presents the results of a Monte Carlo simulation that illustrates the ability of CRE logit models based on Equation 5 to differentiate between the effect of unobserved heterogeneity and an actual positive effect of work experience.
Occupational Similarity Measures
Constructing the Occupational Similarity Measures
We adopt two approaches to measure the skill similarity between occupations: information on multiple dimensions of job skills from the O*NET, and data on occupational mobility from matched samples of the Current Population Survey (CPS). In each case, the strength of the linkage between occupations (θab) is the basis for our definition of a mobility pathway. For the purposes of testing Hypothesis 2, we define a “linked” occupation as one where θab ≥ 0.8, although our results show that a broad definition of a linked occupation is sufficient to support Hypothesis 2. 15
Skill Similarity in the O*NET Data
First, we use data from the Occupational Information Network (O*NET), which was developed by the U.S. Department of Labor and provides information on the skill requirements of different occupations. For our purposes, the advantage of the O*NET data is they allow us to calculate the relationship between pairs of occupations across a broad range of heterogeneous skill and ability requirements. The 2010 version of the O*NET data (ONET15) provides information on 120 variables measuring different components of the abilities, knowledge, and skill levels of each occupation (Hilton and Tippins 2010).
We measure the skill similarity between occupations in the O*NET using an approach intended to preserve as much information as possible about the relationship between pairs of occupations. First, we convert the skill ratings data from the O*NET occupational codes to the three-digit Census occupation codes using the crosswalks provided by the O*NET. We then standardize each of the 120 ratings variables by converting them to variables with a mean of 0 and a standard deviation of 1. Finally, we calculate the “skill similarity” (
Occupational Mobility in the CPS Data
The validity of the O*NET as a measure of skill similarity between pairs of occupations depends on how well the occupational analysts’ observations capture the heterogeneous and idiosyncratic skills that define expertise across different occupations and, in turn, affect workers’ ability to transfer skills between occupations. Considering each occupation separately with abstract skill categories can lead to the appearance of high levels of skill similarity between occupations that, in practice, rely on very different types of skill.
As an alternative to the O*NET data, we use occupational mobility data from the Current Population Survey (CPS) to measure the similarity between pairs of occupations. Although the CPS data do not provide a direct measure of skills, they may pick up patterns of mobility that are missed in the expert evaluations in the O*NET data. This assumes a high level of mobility from occupation A to B (relative to the size of each) is likely only if they have similar skill requirements, or if occupation B builds on the experience and training of occupation A.
Shaw (1984, 1987) and Sicherman (1993) use data on the relative level of mobility between one- and two-digit occupation codes as an indirect measure of occupational similarity. Our approach is similar except we use a larger CPS dataset to calculate the degree of mobility between three-digit occupations (see also de Toledo et al. 2014). We use data on occupational changes across successive months of the CPS from 1994 to 2016.
17
The resulting dataset contains 258,684 cases of occupational mobility. We divide these data into two periods (1994–2003 and 2000–2016) based on the occupational codes that are used, because the CPS switched from the 1990 codes to the 2000 codes in 2003, and both codes are available using double-coded occupational data for 2000 to 2003. This results in separate measures of
We then construct a 0–1 measure of the relative mobility between occupations A and B as
where
At first glance it might seem tautological to use occupational mobility data from the CPS to model occupational mobility in the NLSY and the SIPP, but the effect of
Table 4 provides a comparison of the CPS and O*NET similarity measures. A key finding is that the CPS measure seems to do a better job than the O*NET at identifying occupations that are linked through mobility that might be thought of as a natural progression based on the development of task-specific skills and knowledge that would help incumbents either manage workers from the initial job (i.e., hairdressers → supervisors of personal service workers), or perform more complex tasks that incorporate skills learned in the origin occupation (food preparation workers → chefs, and taxi drivers → dispatchers). The stronger ability of the CPS similarity measure to identify linked occupations (and the more robust results using this measure in our analyses) is likely due to it capturing other institutional linkages between occupations, including OILMs, not reflected by the O*Net skill measure. 18
Comparison of the CPS and O*NET Similarity Measures
Note: “Code” refers to the 2000 Census three-digit occupation code.
There is substantial variation among low-wage occupations in terms of their links to higher-wage jobs, as reflected in the CPS data (see Table 5). Here we return to our list of low-wage occupations from the descriptive analysis in Table 1, but now we rank them in terms of the strength of their links to higher-wage jobs. Column 3 of Table 5 shows the number of linked occupations for each low-wage origin occupation, defining a link based on
Low-Wage Occupations Ranked by Size of Their Mobility Pathways to Linked Occupations
Note: Selected low-wage occupations (including all the occupations from Table 1), ranked by total size of their mobility pathways (“Tot. size”) based on the size of the linked destination occupations above the “high” wage threshold. 2000 to 2019 CPS data.
A link is defined as an occupation pair where the similarity measure ≥ 0.8. This defines a dichotomous measure of a “mobility pathway” between two occupations. In our analysis, we use the continuous measure to represent the strength of the pathway connecting occupations. See note b of Table 2.
The “total size” is the percentage of all workers in the labor market who earn above the high-wage threshold ($14 in 2012 dollars) who work in the linked occupations.
Nlsy Data and Results
NLSY Data
We first use longitudinal data from the 1979 National Longitudinal Study of Youth (NLSY). The NLSY work history data allow us to measure the timing and duration of low-wage spells, as it provides weekly data on respondents’ main job and employment status over their whole career. While the job and employment data are in a weekly array, data on wages are updated once per interview. For our measurement of wages, we keep one case per interview for each respondent, which represents the wage of the job held at the time of the interview, or the most recently held job if the respondent is not currently employed. We define entry into low-wage spells as working at a job with wages of less than $11 per hour (in inflation adjusted 2012 dollars), with exit from low wages defined as earning more than $14 per hour. In our analysis, we also estimate additional models using $12 and $16 as the threshold for exiting low-wages to show that our results are robust to alternative definitions of low wages. 20
Upward wage mobility is the key dependent variable in all our analyses, and low-wage spells continue until individuals exit to higher wages. The low-wage spell can include multiple job and occupation changes, and work experience in all jobs during the spell is included in the analysis (see Equation 6). Individuals can achieve upward wage mobility in their current occupation or via mobility to other occupations. To prevent measurement error in wages from resulting in spurious low-wage entries and exits, we require two observations of low wages to start a spell, and two observations of high wages to exit a spell. 21 The duration of the low-wage spell is measured as the cumulative number of weeks of work in the work history data starting with the week of the second observation of low wages (i.e., <$11 per hour) and ending with the first week of a successful exit to higher wages (>$14 per hour).22, 23
Table 6 shows the mobility outcomes for the NLSY 1979 to 2016 low-wage panel using $11 per hour as the threshold for entry into a low-wage spell, and $14 per hour as the indicator of upward mobility. For each year, column 1 shows the number of low-wage workers in the panel, and columns 2, 3, and 4 indicate whether they are existing cases, new cases that are joining the panel, or cases that are returning after an absence from the data (due to either non-employment or a non-interview). Individuals who attrit from the panel still contribute cases for all years they are in the data. In 1986, for instance, there are 4,016 cases in the low-wage panel, 354 are new spells of low-wage work (column 2), and 484 cases exited the panel into higher-wage employment during the year (column 8). 24
Mobility Outcomes by Year for the NLSY79 Data, 1979 to 2016
Note: “New cases” indicates the number of cases that are joining the low-wage panel for the first time. “Returning cases” are returning to the sample after a temporary absence; “attrition” indicates cases leaving the sample because they are not employed, not in the sample, or it is the end of the panel; and “temp. absence” indicates cases that are not observed in the next wave but return to the low-wage panel later. “Upward mobility” exits the low-wage panel into higher wages.
To change the structure of the NLSY data into the format needed for the MCL models, we start with the event history data on upward mobility from Table 6 (which has one case per respondent per time period, as illustrated in Table 3a), and we expand the upward mobility outcome category Y = 1 (column 8 of Table 6) to include a random sample of 20 three-digit Census occupations in which the respondent could have achieved upward mobility (as illustrated in Table 3b). A different random sample of unchosen alternative occupations is selected for each individual case and time-period. 25 Sampling from the set of unchosen outcome categories is a basic feature of the conditional logit model (and, by extension, our MCL models), and it does not affect the consistency of the coefficient estimates (Bruch and Mare 2012; McFadden 1978). 26 In our case, this is of practical importance because we use the Census three-digit occupational codes that are available in the NLSY data (as well as the SIPP), which results in 500 destination occupations (with the 2000 Census codes) and quickly becomes impractical with large sample sizes. In our MCL models, we select a different random sample of 20 occupations for each worker for each time period. The coefficient on the constant term for Y = 1 (e.g., row 37 of Table 8) absorbs the difference in the number of cases, but the coefficients on all the other variables are unaffected by the sampling of alternatives. 27 Doing this shifts the structure of our data from one case per respondent per time period to multiple cases for each respondent for each time period, as was discussed earlier in reference to Table 3b. 28
Table 7 shows summary statistics for the 1979 to 2016 NLSY sample, with Panel A listing individual-level variables. Our key measure of duration is the log cumulative weeks of work experience during the low-wage spell in row 7. We use the log duration in the analysis because it fits better than alternative measures, including a linear measure and other nonlinear measures. Table D4 in the online supplement presents additional models using experience squared (and experience cubed), and Figure D1 shows the effect on the mobility rate is very similar. In addition to work experience, we also use measures of the log cumulative weeks unemployed (row 8) and the log cumulative weeks out of the labor force (“OLF”, row 9) during the low-wage spell. These measures allow us to compare the relative effect of working in a low-wage job compared to not working. 29 Row 10 shows the measure of log weeks of tenure with the current employer. This differs from the duration of the low-wage spell (the log experience measure in row 7) because low-wage spells continue across multiple employers until the worker exits low wages (as described above in the definition of the low-wage spell). In row 11, the median initial wage for the sample was $8.44, and the median final wage for these workers was $15.38 at the point they exited the panel—either into high wages, extended non-employment, attrition, or the end of the panel.
Summary Statistics for the NLSY Data
This is the choice set of potential destination occupations, which is a random sample of the 500 three-digit Census occupations for each wave of the data for each respondent. See the text for details on the sampling of choice sets in conditional logit models.
As a descriptive indicator of the importance of mobility pathways, row 14 of Table 7 shows that 21.5 percent of the 4,306 cases of upward mobility observed in the data were moves to linked occupations based on the definition
In Panel B of Table 7, rows 16 to 21 show variables that are cross-level interaction terms between characteristics of potential destination occupations and worker- or origin-occupation variables. We emphasize that the ability to include these interaction terms between individual and (potential destination) occupation variables is a central feature of the MCL model. In our discussion of the data structure for an MCL model in Table 3b, rows 2 to 5 are potential destination occupations, and the analysis attempts to predict which one (if any) was the actual destination. In Table 7, rows 16 and 17, for example, are the CPS and O*NET measures of occupational similarity between the worker’s current occupation and the potential destination occupation. 31 Rows 18 and 19 show measures of the experience occupational-similarity interaction term to potential destination occupations (“EOI”) using the CPS and O*NET measures of occupational similarity.
The EOI variables are the key part of our analysis, as they allow us to test the mobility pathways hypothesis that work experience increases the rate of upward mobility to skill-similar destination occupations. For each potential destination occupation B for worker i, this measure is calculated in Equation 6 as an interaction term where occupational experience is multiplied by the level of skill transferability to occupation B based on the CPS and O*NET data discussed above:
where is the duration of the low-wage spell at time t, excluding experience where the origin and destination occupations are the same (which we include as a separate variable). By taking the natural log of
Next, row 21 is the size of the potential destination occupation above the high-wage threshold, measured using CPS data from 1994 to 2012 as the proportion of all workers in the economy who earn more than $14 per hour (in 2012 dollars) who work in that occupation. Finally, rows 23 to 42 are dummy variables for the worker’s current origin occupation. Inclusion of these dummy variables means we are estimating occupation-level fixed-effects models for all our analyses. In contrast to the additive effects models discussed above, we are not interpreting the coefficients on the occupation dummy variables as structural effects; they are intended as controls for unobserved occupation-level differences in mobility rates that are subject to the same critique of the additive models, that the effects could be due to sorting on unobserved characteristics of workers across occupations. The occupation categories are based on the Census 1990 occupation codes, and they include large three-digit low-wage occupations, such as cashiers, janitors, and waiters, as well as a combination of related smaller three-digit low-wage occupations. 33
NLSY Results
Table 8 shows results for models with the NLSY79 data using all spells of low-wage employment that occur between 1979 and 2016. 34 We begin by estimating a random-effects logit model (M1) to test the effect of low-wage work experience on upward mobility (Hypothesis 1). As discussed earlier, according to the stepping-stone perspective, the effect of work experience should be positive, as it leads to skill development that can be transferred to higher-wage jobs. In contrast, the dead-end perspective argues that upward mobility is based on sorting and there is no effect of work experience on upward mobility. In model M1, the key variable is the measure of log low-wage experience in row 4, which has a positive effect on upward wage mobility (0.309, p < 0.001). This result supports the work experience hypothesis (Hypothesis 1) described in column 5 of Table 2. Longer employment in a low-wage job increases the rate of upward wage mobility. In contrast, log weeks of unemployment during the low-wage spell has a negative effect on upward mobility (−0.052, p < 0.001), and log weeks out of the labor force (OLF) has no effect (0.007, not significant at the p < 0.05 level). 35 These results are useful to contextualize the relative effect of low-wage work experience on upward wage mobility vis-à-vis unemployment or being out of the labor force. Note that M1 includes dummy variables for the origin occupation, but it does not include the variables that measure characteristics of the destination occupations or the potential skill transferability between occupations (in rows 1–3 and 8–11), which reflects the limitations of the logit approach.
Logit, Conditional Logit, and MCL Models of Wage and Occupational Mobility, 1979 to 2016 NLSY Data
Note: Standard errors are in parentheses. “RE” and “MCL” refer to random effects and multinomial conditional logits, respectively. The shaded rows indicate the key independent variables, EOI and log low-wage work experience.
p < 0.05; **p < 0.01; ***p < 0.001 (two-tailed tests).
In Models M2 to M5 we estimate random-effects multinomial conditional logit (MCL) models of upward mobility.
36
As we described in reference to Table 3b, the advantage of the MCL model compared to the logit model is that it enables us to incorporate the role of the occupational structure directly in the analysis because potential destination occupations are included as “choices” in the data, and the analysis focuses on modeling which (if any) destination occupation was the one where upward wage mobility was achieved. This allows us to test the mobility pathways hypothesis (Hypothesis 2), which focuses on the effect of the experience-occupation interaction (EOI) variable that represents the cross-level interaction between the structural-level measure of occupational similarity
The key finding in the MCL models using the CPS measure of occupational similarity (Model M2) and the O*NET measure (Model M3) is that there is strong evidence of “stepping-stone” mobility based on the effect of the low-wage experience and the EOI variables. In Model M2, the coefficient on the EOI variable is
A similar set of results holds for Model M3, which uses the O*NET measure:
In addition to the key results on the EOI and experience variables, the MCL models in Table 8 also estimate the effect of additional characteristics of potential destination occupations. Rows 8, 9, and 10 are dummy variables for the average occupational wage of the potential destination occupation. The destination occupation size (in row 11) is the proportion of all workers who earn more than the high wage threshold ($14 in 2012 dollars) who work in the occupation. Just as including characteristics of potential destination neighborhoods is fundamental to the use of discrete choice models in the study of residential mobility (Bruch and Mare 2012; Quillian 2015), including characteristics of potential destination occupations is a key feature of our MCL analysis because it allows us to incorporate measures of the occupational structure directly into our analysis. The positive coefficient in row 11, for example, indicates that the size of the potential destination occupation above the $14 wage threshold is an important determinant of upward mobility to that occupation, and the negative coefficient on the dummy variable in row 10 for Model M2 indicates that upward mobility to occupations with an average wage of more than $20 per hour is less likely. 37
The results in Models M2 and M3 are robust to alternative wage thresholds and sample definitions. In Model M4, for example, we use an alternative definition of upward mobility based on a $16 per hour threshold (rather than $14 per hour) and find similar results to the other MCL models. Finally, in Model M5 we estimate a separate model for women based on the expectation that women face gender-based discrimination embedded in organizations and occupations (Acker 2012; England et al. 1988) that will constrain upward wage and occupational mobility. Our conclusions in model M5, however, are the same regarding the mobility pathways hypothesis (i.e.,
Figure 1 highlights key features of the MCL analysis by illustrating how the results from Model M2 of Table 8 can be translated into predicted probabilities of upward mobility.
38
Figure 1 shows the predicted probability of upward wage mobility for janitors and food service prep workers to selected destination occupations that vary in terms of the strength of their occupational similarity to the origin occupation. The probability of upward wage mobility from occupation A to occupation B (
Predicted Probability of Upward Wage Mobility to Different Destination Occupations, Janitors and Food Prep Workers
For both janitors and food prep workers, the graphs in Figure 1 show that while the predicted probability of upward wage mobility to non-linked occupations (i.e., sales managers for janitors [
Sipp Data and Results
SIPP Data
To complement the NLSY data, we use longitudinal data from the 1996, 2001, 2004, and 2008 panels of the Survey of Income and Program Participation (SIPP) to analyze the mobility patterns of low-wage workers. Each panel of the SIPP is a representative sample of the U.S. population consisting of three waves of data per year. 39 An advantage of the SIPP is that each panel consists of about 50,000 households, which provides a large sample of workers allowing us to test for occupation-level variation in stepping-stone effects on upward mobility.
Beginning with the 1996 panel, the SIPP asks a question, EOCCTIM1, about respondents’ overall experience in their current occupation: “Considering [your] entire working life, how many years [have you] been in this occupation or line of work?” 40 We use this variable to measure the length of work experience in the respondent’s current occupation prior to the first wave of the survey. Our process of creating the analytic sample for the SIPP data parallels that of the NLSY data as much as possible; additional details on the construction of the SIPP sample are provided in Part E of the online supplement.
Table 9 presents summary statistics for the variables used in our analysis of mobility in the SIPP. Panel A of Table 9 presents the individual-level variables, and Panel B shows the variables that are either at the occupational level or a combination of individual and occupational characteristics. In row 8, the log months of experience in the current low-wage spell is the key measure of duration. Overall, 3,318 (12.1 percent) workers achieve upward mobility during the SIPP panel (row 21), and 1,132 (4.3 percent) do so in the same occupation they started the panel with. 41
Summary Statistics for the SIPP Data
This is the choice set of potential destination occupations, which is a random sample of the 500 three-digit Census occupations for each wave of the data for each respondent. See the text for details on the sampling of choice sets in conditional logit models.
In addition to the skill similarity and EOI variables in rows 22 to 25 discussed above with respect to the NLSY data, Panel B of Table 9 also shows additional variables used in the SIPP analysis to measure structural characteristics of the potential destination occupations. One advantage of the SIPP (compared to the NLSY) is that we can match it to yearly changes in the occupation structure using the large sample size of the American Community Survey (ACS), which is available after 2000. In Table 9, the variable “occupational size” is the proportion of all workers earning above the specific wage threshold who are in the potential destination occupation, and “high wage” employment growth is calculated as a three-year moving average of the absolute change in the size of the occupation measured as a proportion of the labor force earning above $14 per hour. Both variables are calculated using yearly data from the ACS for the years 2000 to 2014 and the CPS prior to 2000.
Next, in row 28 of Table 9 we include a measure of the role that destination occupations play as “entry portals” for upwardly mobile workers from low-wage occupations (see Spilerman 1977). We calculated this variable from the 1994 to 2016 CPS occupational mobility data we used to estimate the mobility-based measure of skill similarity described earlier. The entry portal variable is a measure of the number of upwardly mobile low-wage workers who enter specific destination occupations. It is calculated as
SIPP Results
Tables 10 and 11 present the results using the SIPP data. We first replicate the main models from the NLSY analysis before turning to models that utilize the larger sample size of the SIPP to estimate occupation-specific tests of Hypotheses 1 and 2. First, in the CRE logit model in M6, there is a large effect of log low-wage experience on upward mobility (0.367, p < 0.001), which is consistent with the work experience hypothesis (Hypothesis 1). Next, Model M7 shows that the mobility pathways hypothesis (Hypothesis 2) is confirmed in the MCL model using
Logit, CL, and MCL Models of Wage and Occupational Mobility, 1996 to 2008 SIPP panels
Note: Standard errors are in parentheses. “CRE” and “MCL” refer to correlated random effects and multinomial conditional logits, respectively. The shaded rows indicate the key independent variables, EOI and log low-wage work experience. The sample size is smaller in models M8 and M9 because they only use workers in origin occupations with median wages less than $15 per hour (in 2019 dollars).
p < 0.05; **p < 0.01; ***p < 0.001 (two-tailed tests).
Occupation-Specific Coefficients for Models M8 and M9
Note: Standard errors are in parentheses. Columns 3 to 7 are occupation-specific coefficients. Column 1 is the number of unique individuals in the occupation, and column 2 is the number of upward moves.
p < 0.05; **p < 0.01; ***p < 0.001 (two-tailed tests).
In addition to the tests of the key variables in rows 1 and 4, Table 10 also shows the effects of variables related to other aspects of occupational structure in rows 6 to 12. While there are positive effects of potential destination occupation size and growth, a key variable in Models M7 and M9 is the measure of entry portals for upwardly mobile workers in row 7. As described earlier, this variable is constructed from the CPS occupational mobility data as a measure of the degree to which higher-wage occupations serve as destinations for upwardly mobile workers from low-wage occupations. The fact that the entry portal variable is positive and statistically significant (at the p < 0.001 level) in M7 and M9 suggests some occupations serve as portals to higher-wage jobs. That we can include these features of potential destination occupations along with the origin-destination interaction effects based on occupational similarity between pairs of occupations in rows 1 and 2 indicates the flexibility of the MCL approach to incorporate structural effects in the analysis of occupational and social mobility.
Origin-Occupation-Specific Analysis
Models M8 and M9 in Tables 10 and 11 allow the effect of EXP and EOI to vary by occupational category. This enables us to test whether specific low-wage occupations are “dead-ends” or “stepping-stones” based on the effect that occupation-specific experience has on the rate of subsequent upward wage mobility, rather than generic low-wage experience. In these models, the measure of low-wage occupational experience and the EOI variable are calculated separately for each of the 20 occupational categories listed in Table 11. 43 Because this results in large numbers of additional variables, the occupation-specific terms are listed vertically in columns 3 to 7 of Table 11.
In Table 11, column 1 indicates the number of unique individuals who work in each of the occupational categories at some point during the SIPP panel. As discussed earlier, the categories are based on the three-digit Census occupational codes and are a combination of specific large three-digit occupations (e.g., cashiers) as well as groups formed by related smaller occupations (e.g., the low-wage sales occupations in row 3). 44 The practical reason for using these broader occupational groups is the limitation in the number of cases (column 1) and the number of individuals achieving upward mobility from that occupation (column 2). Much of the statistical power in the MCL models comes from the destination occupations of upwardly mobile workers (i.e., the number of cases in column 2), so the results in Table 11 should be considered illustrative. An ideal dataset for future research would have large numbers of workers in each three-digit origin occupation, permitting the results to be disaggregated even further.
The occupation-specific results for the correlated random-effects (CRE) logit model (M8) are presented in columns 3 and 4 of Table 11. As discussed in the Methods section, we use CRE models to account for the left-censoring of spells in the SIPP data. The key result is the estimate of column 3, which tests for the effect (
Nonetheless, as discussed earlier, a shortcoming of the logit approach in Model M8 is that it cannot test our proposed mechanism for upward mobility based on mobility pathways (Hypothesis 2). In contrast to the MCL model, the logit model cannot incorporate information on the strength of the pathways connecting low-wage occupations to potential destination occupations. This is important because Table 5 provides evidence that low-wage occupations vary in the degree to which they are connected to higher-wage occupations. As a result, it is possible that the null effects for experience in some occupations, such as teacher’s aides in Model M8, reflect structural differences in the strength of their mobility pathways rather than a failure of the proposed mechanism of upward mobility based on those pathways (Hypothesis 2).
When we move from the logit model in M8 to the CRE MCL model in M9 of Table 11, we extend our occupation-specific analysis to incorporate the role of these institutional and skill-based linkages between occupations as a test of the mobility pathways hypothesis (Hypothesis 2). In these models, the effect of occupational work experience in origin occupation A on the rate of upward wage mobility to potential destination occupation B is
The key finding in the MCL model in M9 in Table 11 is that all 20 occupations provide evidence consistent with the mobility pathways hypothesis (Hypothesis 2) (cf. row 2, column 5 of Table 2). For example, for teacher’s aides (row 2 of Table 11),
We can reconcile the apparent divergence in the occupation-specific results in columns 3 (logit) and 6 (MCL) of Table 11 by emphasizing that column 6 (for the MCL model M9) is testing the mechanism of stepping-stone mobility (Hypothesis 2) to linked occupations, 45 and the existence (or absence) of occupational linkages between occupations is a structural feature. Column 6 of Table 11 shows that evidence in favor of the proposed mechanism (Hypothesis 2) is supported in all 20 occupations. In contrast, the results for the logit model in column 3 of Table 11 are reduced form estimates that combine the effect of the stepping-stone links with occupational variation in the number (and strength) of links to higher-wage jobs. The descriptive evidence of variation in the number and size of mobility links for different low-wage occupations presented in Table 5 is consistent with the finding in column 3 of Table 11 that the overall effect of experience on upward mobility will vary by occupation.
Discussion and Conclusions
A key characteristic of inequality in the United States is the relatively large proportion of workers in low-wage jobs (Henderson 2022; Ross and Bateman 2019), particularly compared to other wealthy industrialized countries such as Denmark and France (Gautié and Schmitt 2010). However, the composition of the low-wage workforce is not static, and the available evidence indicates that a substantial proportion of low-wage workers move up over time (Boushey 2005; Carrington and Fallick 2001; Escobari et al. 2021; Schultz 2019). 46 A key question is who moves up and why. We argued that a central issue in understanding upward mobility in low-wage labor markets is analyzing the role that the occupational structure plays in facilitating or impeding the mobility prospects of workers.
We sought to understand how individual workers’ wage mobility is affected not only by their individual-level characteristics but also by variables that measure specific aspects of the occupational structure. We identified and tested two hypotheses. The work experience hypothesis (Hypothesis 1) posits that the accumulation of job-related skills as measured by work experience increases the rate of upward wage mobility out of low-wage jobs. The mobility pathways hypothesis (Hypothesis 2) predicts that the combined effect of work experience on upward wage and occupational mobility will be positive, with the strength of the experience effect conditional on the degree of similarity between the origin occupation and the potential destination occupation. The mobility pathways hypothesis provides a stronger test of the mechanism proposed by the stepping-stone perspective: the accumulation and transfer of work experience across occupational linkages in OILMs. We used data from the O*NET and the CPS to measure the skill and mobility linkages between pairs of occupations.
To test these hypotheses, we utilized longitudinal data from the NLSY79 and four panels of the SIPP that provide information on the wage and occupational mobility of low-wage workers. The NLSY includes the complete work histories of a cohort of workers, which allows for precise measures of cumulative occupational experience and enables us to test our hypotheses across all low-wage occupations. The larger size of the SIPP data allows us to test occupation-specific versions of the hypotheses. We used multinomial conditional logit (MCL) models (Breen 1994; Hoffman and Duncan 1988; Powers and Xie 2000) that enable us to jointly model wage and occupational mobility and test the mobility pathways hypothesis (Hypothesis 2). A key aspect of the analysis is the experience-occupation interaction (EOI) term (defined in Equation 6), which is a cross-level interaction between the strength of the occupation-level mobility link and individual-level occupation-specific work experience.
Overall, our analysis with the NLSY and the SIPP indicates clear support for a stepping-stone perspective on the mobility of low-wage workers. We find support for the work experience hypothesis (Hypothesis 1) using the logit model approach with both the NLSY (Model M1) and the SIPP (Model M6). Our preferred approach, however, is the MCL analysis because of its greater flexibility to incorporate variables representing the role of the occupational structure. In all the MCL models, including all 20 of the occupation-specific results in M9 of Table 11, there is consistent evidence of a positive effect of occupational work experience on the probability of upward wage mobility along mobility pathways to skill-linked destination occupations, indicating strong support for the mobility pathways hypothesis (Hypothesis 2). These findings suggest that, in addition to individual variables such as education, gender, and race, the degree to which a worker’s occupation is connected to higher-wage occupations influences workers’ prospects for upward wage mobility.
The finding that low-wage jobs can provide stepping-stones to upward wage mobility is consistent with the idea that skill accumulation occurs even in jobs that correspond to the popular conception of stigmatized “dead-end” jobs. The reality is, of course, that the promise of upward mobility is a deferred assurance based on the possibility of eventually moving out of one’s current job.
In terms of the methodological challenge of analyzing structural effects in labor markets, the use of a discrete-choice modeling framework in MCL models is valuable for understanding how careers and intragenerational mobility are shaped by the interplay between labor market structures and workers’ individual characteristics and resources. The MCL approach allows us to study workers’ careers by incorporating wage and occupational changes into the same model and including characteristics of potential destination occupations as explanatory variables. In general, we argue that MCL models permit a more realistic incorporation of the structural features of labor markets, and they can be extended to include additional aspects of mobility and the job matching process. For example, a researcher using employer–employee administrative data could include firms in their models as well, where individual mobility would depend on firm-specific job openings nested within broader occupational labor markets. The growing availability of such datasets around the world provides researchers with chances to model how occupations and organizations combine to shape labor market outcomes.
In our analysis, we focus on one structural effect related to workers’ careers, accumulating experience and moving across occupational linkages, as motivated by the literature on low-wage jobs as dead-ends or stepping stones. Future research on careers could profitably address several important sources of heterogeneity that we have not considered explicitly in this article. First, there is undoubtedly variation in regional and local labor markets in opportunities for movement into higher-wage occupations. This is especially true in the case of low-wage labor markets—which are often locally-based—but is likely to hold for labor markets generally. Second, business cycles and secular trends in increasing demand for certain occupations and skills create better prospects for movement into these (likely high-paying) activities. Third, occupations and the OILMs that are linked to them differ in the length of their job ladders and the speed with which persons can climb them. Structural features of occupations, such as their degree of inequality in wages and skills, provide both opportunities and constraints on the likelihood of upward mobility. OILMs may also represent different types of mobility structures, such as contest and tournament models of mobility. Finally, heterogeneity in the structures and processes of careers and intragenerational mobility reflect compositional differences in workers’ characteristics, such as gender, race, age, education, and family, among others. We have discussed some of the basic patterns, but further specification of the links between individuals’ characteristics and labor market structures is needed to understand more fully processes of socioeconomic achievement and intragenerational inequality.
Supplemental Material
sj-pdf-1-asr-10.1177_00031224241232957 – Supplemental material for “Stepping-Stone” versus “Dead-End” Jobs: Occupational Structure, Work Experience, and Mobility Out of Low-Wage Jobs
Supplemental material, sj-pdf-1-asr-10.1177_00031224241232957 for “Stepping-Stone” versus “Dead-End” Jobs: Occupational Structure, Work Experience, and Mobility Out of Low-Wage Jobs by Ted Mouw, Arne L. Kalleberg and Michael A. Schultz in American Sociological Review
Footnotes
Acknowledgements
We would like to thank the anonymous ASR reviewers for their comments and suggestions.
Funding
This research was supported by a grant to Mouw and Kalleberg from the Russell Sage Foundation (G-6477, “Stepping Stones and Ladders: The Sources of the Mobility of Low Wage Workers in the United States”). Schultz was supported by grant P2CHD042849 awarded to the Population Research Center at The University of Texas at Austin by the Eunice Kennedy Shriver National Institute of Child Health and Human Development. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.
Notes
). His most recent book is Precarious Asia: Global Capitalism and Work in Japan, South Korea and Indonesia (with Kevin Hewison and Kwang-Yeong Shin; Stanford University Press 2022). He served as President of the American Sociological Association in 2007–8 and is currently the editor of Social Forces: An International Journal of Social Research.
References
Supplementary Material
Please find the following supplemental material available below.
For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.
For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.
