3.1 Defining nontraditional work

Defining nontraditional work is complicated – an agreed upon definition does not exist, and different definitions yield vastly different estimates. Much of the existing literature has defined this type of work based on the nature of the relationship workers have with employers. Using this approach, researchers have come up with a wide range of estimates. The narrowest definitions of nontraditional work are limited to workers in the ‘gig economy’ (e.g., Uber and Task Rabbit) or in short-term employment relationships.Footnote ⁷ These groups include just 1% and 2% of workers, respectively.Footnote ⁸

At the other extreme, the U.S. Government Accountability Office's (2015) broadest concept, which includes the self-employed and those in part-time jobs, covers 31% of the workforce.Footnote ⁹ In between these extremes is the definition of ‘alternative’ work used by the Bureau of Labor Statistics (BLS) – which has received considerable attention through research by Katz and Krueger (Reference Katz and Krueger2016, Reference Katz and Krueger2019). The BLS definition includes independent contractors and workers who are either with a temp agency, employed by a contract firm, or on-call. Under this definition, the prevalence of nontraditional work hovers around 10%. Another definition in between the two extremes is that of ‘1,099 workers’, as used in a 2019 study by Collins et al. These workers are self-employed individuals who work for firms that file 1,099 forms reporting the workers' pay (i.e., freelancers and ‘gig’ work) but do not fall under normal employment classification rules. Using this definition, nontraditional work would account for 11.8% of the workforce.Footnote ¹⁰

Regardless of how nontraditional jobs are defined, the common concern is that these jobs may lack even basic benefits, such as health insurance and retirement plans, and/or have volatile earnings and employment. For this reason, instead of looking at the employee–employer relationship – as the definitions above do – our analysis adopts a more direct measure of nontraditional jobs based on the presence of these job characteristics. Specifically, the analysis will define nontraditional work in two ways: (1) broadly as any job lacking both health insurance and retirement benefits; and (2) more narrowly as a job without these benefits that also has some measure of job instability.Footnote ¹¹

In the HRS, these characteristics can be identified for an individual's current ‘main’ job at the time of their HRS interview. Because these characteristics cannot be easily identified for jobs held in-between HRS interviews, the sequence analysis is limited to ‘snapshots’ of an individual's employment status at the seven interviews occurring between ages 50 and 62.Footnote ¹² Importantly, the job characteristics approach may disagree with the employee–employer-based definitions above in some instances. For example, a worker with stable hours at a contract firm that covers its workers with health insurance and a retirement plan would be labeled as in an ‘alternative’ job under the BLS definition, but as in a traditional job under our definitions.

Indeed, given the variance in the definitions, it is useful to see how a definition based on job characteristics like benefits compares to the employer–employee-based definitions from the existing literature. To ensure an accurate comparison, the analysis requires a dataset with questions on both the worker–employer relationships used in the other definitions presented above, and job characteristics such as the availability of benefits that are used in this paper. For this purpose, the BLS's Current Population Survey (CPS) is the best source.

For the comparison of the various definitions that exist in the literature to the one used in this paper, we first compute the share of workers aged 50–62 in 2017 who are in employer relationships under the standard BLS definition. Next, the share of workers in jobs without health insurance and retirement benefits is calculated – i.e., our broad ‘no benefits’ definition of nontraditional work (unfortunately, the CPS does not have the right variables to get at the measures of instability used in the narrow definition). Under the employee–employer-relationship-based BLS measure, 11% of workers in 2017 were in nontraditional jobs, compared to 20% under the no-benefits measure (Figure 1 also includes other definitions from the literature for additional context). The estimates in Figure 1 are for a single point in time. Figure 2 compares the two measures over 1994–2016, and it still finds a large and persistent gap.

Figure 1. Percentage of workers ages 50–62 in nontraditional jobs by definition.

Notes: ‘Gig’ definition as in Farrell and Greig (Reference Farrell and Greig2016) and covers all workers. ‘1099’ workers defined as in Collins et al. (Reference Collins, Garin, Jackson, Koustas and Payne2019) and applies to all workers. ‘Contingent’ (BLS), ‘Alternative’ (BLS), ‘No benefits,’ and ‘Contingent’ (GAO) were calculated by the authors and apply to workers aged 50–62.

Sources: Farrell and Greig (Reference Farrell and Greig2016); Collins et al. (Reference Collins, Garin, Jackson, Koustas and Payne2019); and Current Population Survey May Supplement (2017).

Figure 2. Percentage of workers in nontraditional jobs under different definitions, 1994–2016.

Note: ‘Alternative’ estimate comes from Katz and Krueger (Reference Katz and Krueger2019) and includes all workers.

Sources: Current Population Survey March Supplement (1995–2017); and Katz and Krueger (Reference Katz and Krueger2019).

Given the considerable gap between the two definitions of nontraditional work, the question is: which does a better job of picking up the vulnerable workers that researchers are concerned with – one based on the employer–employee relationship or one based on the presence of benefits? It turns out, compared to the workers defined as alternative under the BLS definition, the additional workers picked up by the no-benefits definition (i.e., workers with a more typical employer–employee relationship but who lack benefits) tend to have shorter job tenure and lower socioeconomic status (see Table 2).Footnote ¹³ The basic issue is that the majority of alternative workers under the BLS definition are independent contractors, and those individuals tend to have been employed in that type of work for a while with relatively high incomes, even though they may lack benefits.Footnote ¹⁴ By picking up many employees working without benefits instead of mainly individuals who lack benefits only because they employ themselves, the definition used in this paper and based on the presence of benefits picks up more vulnerable workers than the BLS measure, making it a better choice for this paper.

Table 2. Select characteristics of workers in nontraditional jobs, 2017

Note: The ‘no benefits minus BLS’ column consists of those workers without benefits who are not already captured under the BLS definition.

Source: Authors' calculations from Current Population Survey May Supplement (2017).

Given that the benefit-based definition used in this paper seems to appropriately capture the vulnerable workers researchers are concerned with, it is worth exploring how the definition looks in the longitudinal data needed to do sequence analysis – the HRS. Reassuringly, the HRS data show that the percentage of workers aged 50–62 in jobs with no benefits is generally similar to that using the CPS data despite a noticeable difference early in the period (see the solid versus dashed lines in Figure 3).Footnote ¹⁵ As noted above, the concern over nontraditional jobs stems not just from a lack of benefits, but also a lack of stability in earnings or employment. Some of the jobs that lack benefits will be stable, and these workers may be less vulnerable. Therefore, the paper also uses a more narrow definition that takes into account job stability. This definition will count a job as nontraditional if it lacks benefits and: (1) has hours that are variable at some point during the job; or (2) if the worker is self-employed with no benefits and with no employees.Footnote ¹⁶ Under this definition, the percentage of jobs that are nontraditional falls from 16.9% to 7.6%, somewhat lower than the standard BLS definition of nontraditional workers.

Figure 3. Percentage of workers aged 50–62 in jobs with no benefits by dataset, 1992–2010.

Sources: Authors' calculations from Health and Retirement Study (1992–2010); and Current Population Survey March Supplement (1995–2011).

4.1 Sequence analysis

Sequence analysis is a relatively novel technique in the social sciences; its strength is that the outcome of interest is an individual's entire employment history rather than employment status or job transition at a given age (Abbott, Reference Abbott1990; Aisenbrey and Fasang, Reference Aisenbrey and Fasang2010; MacIndoe and Abbott, Reference MacIndoe, Abbott, Hardy and Bryman2011). The goal of sequence analysis is to group together workers with similar employment statuses at similar times and in a similar order. Consider the hypothetical example below, which shows how three workers move between traditional work (T), nontraditional work (N), not-working (U), and retired (R).

Example 1. Employment Sequences for Hypothetical Workers

In this example, the sequence analysis will likely group workers A and B together, because they both started as traditional workers, used nontraditional work temporarily before returning to traditional work, and then retired at the same age. The only difference is small: how long they experienced nontraditional work. That experience differs distinctly from the pattern for worker C, who moved from traditional to nontraditional work at the same age as A and B, and retired at the same age, but never returned to traditional work.

In more technical terms, sequence analysis compares all of the sequences for sample members and constructs a matrix of how different each sequence is from the others. The difference between sequences is based on the minimum number of modifications needed to transform one sequence into another. A modification can take one of two forms. The first form is a substitution in which the state of one sequence is changed to match the state from another (e.g., changing work status at age 56 from nontraditional to traditional so that workers A and B have the same sequence). The second form is an insertion or deletion. An insertion occurs where a state is plugged into a sequence and every other state pushed back one wave to an older age. A deletion occurs when a state is removed and every subsequent state pulled forward to a younger age. Insertions and deletions typically happen simultaneously: a state is inserted and another state is deleted to preserve the number of observations.

To determine the difference between two sequences, the analysis follows the literature and uses optimal matching analysis (OMA) (MacIndoe and Abbott, Reference MacIndoe, Abbott, Hardy and Bryman2011; Calvo et al., Reference Calvo, Madero-Cabib and Staudinger2017). OMA requires that each substitution and insertion or deletion be assigned a ‘cost’ to calculate the difference between sequences. The simplest way to calculate these differences would be to add up the number of substitutions and insertions/deletions – in other words to assign a uniform cost of one – but this approach has several disadvantages (Aisenbrey and Fasang, Reference Aisenbrey and Fasang2010; Cornwell, Reference Cornwell2015). Most notably, it does not recognize that some substitutions reflect much bigger changes than others – e.g., substituting a traditional job for a person who is not working at all may be a bigger leap than substituting a status of retired. Simply assigning substitution costs based on theories of which transitions are more likely runs the risk of being highly arbitrary, so this project uses an intuitive metric (Cornwell, Reference Cornwell2015; Calvo et al., Reference Calvo, Madero-Cabib and Staudinger2017). Observed transition probabilities – transitions that are observed frequently in the data – e.g., from not working to retired – are assigned a lower substitution cost than those that are uncommon (Halpin, Reference Halpin2014; Cornwell, Reference Cornwell2015). Once substitution costs are assigned, this paper follows the approach commonly taken in the literature and sets the cost of insertions/deletions to one-tenth of the highest substitution cost (see Appendix Tables A1 and A2 for the substitution costs used under the broad and narrow definitions of nontraditional work) (Cornwell, Reference Cornwell2015).Footnote ¹⁷

The end result of OMA is a so-called pairwise distance matrix, which contains the sum of the costs of all substitutions and insertion/deletions required to transform each sequence into another. To group similar sequences together, a Hierarchical Cluster Analysis is used to detect groupings among the individual sequences with respect to their pairwise distances (Cornwell, Reference Cornwell2015; Calvo et al., Reference Calvo, Madero-Cabib and Staudinger2017). The last step is to determine the number of groups for the analysis to detect. To choose the number of groups, the process was run assuming 2 through 12 groupings, with the final choice reflecting the number that maximized the Caliniskin and Harabasz index such that the resulting sequences made theoretical sense (Cornwell, Reference Cornwell2015).Footnote ¹⁸

4.2 Regression methodology

With the sequences in hand, the next question is how individuals' employment patterns relate to their available retirement resources and their emotional well-being at age 62. The issue is that people experiencing different patterns of non-employment, retirement, traditional work, and nontraditional work in their 50s and early 60s will also have different initial characteristics that may cause them to fall into those sequences, and those initial characteristics are likely to affect both their preparedness and their emotional well-being.

For example, workers frequently doing nontraditional work may have less education and therefore contribute less to retirement accounts even when they have the resources. Failure to control for education would therefore exaggerate the negative role of a sequence showing frequent nontraditional work – i.e., these workers would indeed have less, but some of the effect would be due to their education level. Or, workers who spend their late careers in nontraditional jobs may have spent their early careers in these jobs too, leading to less pension coverage earlier in life or lower lifetime incomes and therefore lower retirement income. In this case, it would not necessarily be the late-career experience of these jobs driving poor outcomes, but earlier experiences as well. On the emotional health side, if workers who are able to maintain traditional employment throughout their 50s and early 60s were less likely to be depressed entering their late careers, then failure to control for age 50 emotional health would exaggerate the benefits of the traditional work sequence. Therefore, controlling for both demographic characteristics and an individual's initial state (i.e., prior to age 50) regarding financial or emotional health is crucial to understanding how the late-career use of nontraditional jobs impacts well-being approaching retirement.

The paper therefore estimates two sets of regressions in which the individual's assignment of a sequence group serves as the independent variable of interest. The first set of regressions use retirement income at age 62 as the dependent variable. These regressions are estimated as quantile regressions at the median to lessen the impact of outlier levels of retirement income and control for demographics and initial health, the availability of retirement plans, and lifetime income prior to late career. The complete equation to be estimated is:

$$R_{i, 62} = \beta _0 + \sum\limits_{\,j = 2}^K {\gamma _jS_{i, j} + {{X}^{\prime}}_{i, 50}\beta + \delta H_{i, 50} + \theta RP_{i, 50} + \rho LI_{i, 50} + \tau _i + \varepsilon _i}$$

where R _i,62 is the log of the individual's retirement income at age 62: defined benefit pension income, Social Security benefits, and annuitized defined contribution plan and other financial wealth.Footnote ¹⁹ The variable S _i,j is an indicator for whether person i was assigned to sequence group j. Therefore, γ _j is the predicted percentage point change in median retirement resources associated with being in sequence group j relative to the base sequence group, which is assigned as the one with the highest amount of traditional work. The vector ${X}^{\prime}_{i, 50}$ contains demographic characteristics that could ultimately affect an individual's preparedness, like education, gender, race/ethnicity, and their age-50 marital status. H _i,50 is an index of the individual's initial health at age 50 that is based on objective measures, with higher values indicating worse health, whereas τ _i enters year of sample entry fixed effects.Footnote ²⁰ Finally, RP _i,50 and LI _i,50 are coverage by retirement plans and lifetime income, respectively, at age 50.

The second set of regressions is similar to the first, except that the dependent variable here is the incidence of depression and the controls include depression prior to late career instead of lifetime income. These regressions are estimated as a probit, with average marginal effects reported in Section 5. The full specification to be estimated is:

$$D_{i, 62}^\ast \, {=} \, \beta _0 + \sum\limits_{\,j = 2}^K {\gamma _jS_{i, j} + {{X}^{\prime}}_{i, 50}\beta + \delta H_{i, 50} + \theta RP_{i, 50} + \rho D_{i, 50} + \tau _i + \varepsilon _i}$$

where $D_{i, 62}^\ast$ is the latent propensity of the individual to be depressed at age 62, D _i,50 is an indicator for depression at age 50, and the other variables are the same as defined above.

The hypothesis is that, even conditional on the initial characteristics described, sequences containing primarily traditional work with little interruption will be associated with higher retirement income and a lower incidence of depression at age 62. The next-best sequence will occur where nontraditional work is used sparingly as a stopgap, followed by long spells of nontraditional work. The sequences with the worst outcomes – i.e., lowest retirement income or highest incidence of depression – will be those associated with long spells of nontraditional work – in other words, weak attachment to the labor force – or very early retirement. This hypothesis means that relative to the base sequence of consistent traditional work, the coefficients γ _j will be increasingly negative for the retirement income median regressions, and increasingly positive for the depression probit regressions as they move from mostly traditional work to mostly nontraditional work and finally to unattached.