Associations of Tipped and Untipped Service Work With Poor Mental Health in a Nationally Representative Cohort of Adolescents Followed Into Adulthood

Abstract

Precarious work is concentrated in the service industry in the United States and is a risk factor for poor mental health. Service occupations in which workers receive tips are potentially more precarious due to unstable schedule and income, and lack of benefits. We tested hypotheses that individuals working in tipped service occupations have greater odds of experiencing poor mental health (as indicated by self-reported depression, sleep problems, and/or greater perceived stress) relative to individuals in untipped service and nonservice occupations, using cross-sectional data from wave IV of the National Longitudinal Study of Adolescent to Adult Health data set (2007–2008; age range, 24–33 years). To improve comparability of occupation types, propensity scores were computed as a function of childhood factors, then used to construct a sample of 2,815 women and 2,586 men. In gender-stratified multivariable regression, women in tipped service had greater odds of reporting a depression diagnosis or symptoms relative to women in nonservice work (odds ratio = 1.61; 95% confidence interval: 1.11, 2.34). Associations of similar magnitude for sleep problems and perceived stress were observed among women but were not statistically significant; all associations were close to the null among men. Additional research is necessary to understand the factors that underlie differences in poor mental health in tipped and untipped service versus nonservice workers.

An individual’s occupation can lead to differential exposure to physical, psychosocial, and environmental factors with the potential to influence their health (1). Individuals in service occupations, especially tipped service, may be particularly vulnerable; however, the potential health effects of these occupations are understudied.

Service work is precarious (2) and characterized by lack of control over hours and shift worked (3, 4), insufficient benefits (5–7), and lower wages. Service workers represent 42.5% of US workers earning the federal minimum wage and 78.1% of US workers earning less than minimum wage (8). Tipped work may be particularly precarious service work for several reasons. First, the normalization of tipping in certain service occupations in the United States has led to differential minimum wage standards; workers in tipped occupations can be paid a wage 71% lower than the federal minimum wage (9) with the expectation that highly unpredictable and inequitable tips from customers will make up the difference (7). On average, tipped workers are nearly twice as likely to live in poverty relative to untipped workers (7). Second, tipped workers are disproportionately exposed to last-minute scheduling practices (3, 5) and insufficient provision of benefits (7). Third, workers in tipped and untipped service occupations must frequently express or suppress certain emotions during interactions with customers (10–14) and manage sexualized or hostile customer behavior (15).

These aspects of tipped service work have direct consequences, such as physical harm, and indirect consequences for health, such as psychosocial stress, with the latter representing an important determinant of mental health (16, 17). A 2007 report from the National Survey on Drug Use and Health revealed the 12-month prevalence of depression among workers aged 18–64 years was highest among those in personal care and service (10.8%) and food preparation– and serving–related occupations (10.3%) (18). Similarly, prevalence of short sleep duration and of other sleep disturbances is highest within service-industry occupation categories (19) and in financially precarious employment in general (20). However, the potential health implications of tipped work have been minimally assessed and are limited to substance use (21).

In light of the dearth of research on the potential impact of working in a tipped service occupation on health, our objective was to test the hypotheses that 1) individuals in service occupations (tipped and untipped) have greater odds of experiencing depression, sleep problems, and/or stress relative to their nonservice counterparts; and 2) individuals in tipped service occupations are particularly vulnerable to these mental health outcomes.

METHODS

Participants

We used data from the National Longitudinal Study of Adolescent to Adult Health (Add Health), a nationally representative cohort of US adolescents followed for more than 14 years into adulthood (22). In Add Health, a stratified, school-based, clustered sampling design was used to ensure data were representative of the US adolescent, school population. Detailed information on the Add Health study design and procedures are described elsewhere (22). Briefly, a core subset of respondents and parents was randomly selected from within school and gender strata to participate in in-home interviews by trained interviewers (23). Of 20,745 baseline respondents, 76% (n = 15,701) completed interviews for wave IV of Add Health, when participants were aged 24–33 years. This study was exempted by the Oregon Health & Science University institutional review board.

Variables and measurement

Exposure: occupation type (nonservice, tipped service, untipped service)

Occupation type was classified from responses to 2 wave IV questions about current or recent occupation. The classification “service” was assigned to participants if their response to the question “When you see the list of categories, please tell me which best describes what you (do/did) at your (current/most recent) job?” was consistent with a service occupation according to the US Bureau of Labor Statistics industry classification system (e.g., “food preparation and serving occupation”) (Web Table 1; available at https://dbpia.nl.go.kr/aje) (24). The classification “nonservice” was assigned to participants reporting other occupation types. Tipped service was assigned to respondents classified as working in service occupations if their response to the question “Out of these categories, which one best describes this job?” was consistent with a predominantly tipped occupation according to the Economic Policy Institute (e.g., “waiters and waitresses”; Web Table 1) (7). The classification “untipped service” was assigned to respondents in service occupations reporting any other occupation.

Outcomes

The following variables developed by Add Health (25) were constructed in wave IV to be used as measures of the 3 mental health outcomes: 1) Depression or depressive symptoms (yes/no) were defined as self-reported diagnosis of depression and/or depressive symptoms within the past 7 days reported on the modified version of the Columbia Center for Epidemiological Studies Depression Scale-10; 2) sleep problems (yes/no) included self-reported difficulty falling or staying asleep and/or symptoms of sleep apnea over the past 4 weeks; and 3) perceived stress was a 3-level ordinal variable constructed from tertiles of the Cohen’s Perceived Stress Scale score (i.e., 0–3, 4–6, and 7–16). There are no broadly applicable score cutoffs and others advised within-sample comparisons (26).

Statistical analysis

Analyses were conducted in Stata/IC, version 14.2 (StataCorp LP, College Station, Texas), incorporating Add Health survey weights and sample design parameters to account for clustered sampling, attrition, and oversampling, thus approximating the target population of US adolescents in grades 7–12 in 1994.

The analytic sample was restricted to respondents who participated in waves I–IV and had an Add Health sampling weight for analysis (n = 9,421), reported a current or recent job in wave IV (n = 9,205), and had complete exposure and outcome data (n = 9,140). Participants with missing covariates were included; multiple imputation methods were applied as described later in Methods. Add Health sampling weights incorporate a nonresponse adjustment for nonparticipation in 1 or more wave of in-home interviews (27). The analytic sample thus contained 9,140 respondents (n = 4,996 women and n = 4,144 men) before application of propensity score (PS) methods.

Propensity score methods

Our analytic approach addressed 2 methodological challenges related to occupation stratification. First, there is gender-based stratification into occupational categories and specific occupations within those categories (28): 56.6% of all service workers (29) and 67% of tipped workers are women (7). Second, additional nonrandom assignment to occupational category resulting from social selection based on sociodemographic characteristics and other predisposing life experiences (30) may affect health. Therefore, we stratified all analyses by gender and used gender-specific PS to address residual structural confounding present in occupation-type assignment. For each gender, a single set of PS was generated to be used for all outcome models using variables for participant sociodemographics, parental characteristics, and childhood adverse experiences, behaviors, and health (Web Appendix 1) selected using our conceptual framework (Web Figure 1). Web Appendix 2 describes the process used to calculate PS, including the application of multiple imputation to address missingness of variables pertinent to PS development. Multinomial logistic regression was used to model occupation type as a function of these variables and predicted probabilities for each occupation type were computed (Web Appendix 2, Web Table 2). PS first served as decision aids for visually guided restriction of the analytic sample to satisfy the positivity assumption and provide support for exchangeability (Web Appendix 3) (31).

Multivariable analysis of outcomes

Multivariable analyses were conducted with the PS-restricted sample. Logistic regression was used for binary outcomes (depression, sleep problems) and ordinal logistic regression was used for the ordered 3-level categorical outcome (perceived stress), producing odds ratios. Though prevalence of study outcomes was high (>10%), it was not possible to estimate relative risk, because of model complexity with survey design parameters, multiple imputation, and inclusion of an ordinal outcome

Within the PS-restricted sample, PS regression adjustment was used in multivariable analyses to achieve models that were parsimonious and adequately adjusted (Web Appendix 4). In addition to PS regression adjustment, variables that remained unbalanced after sample restriction were included in our models for residual confounding adjustment.

Sensitivity analyses

We conducted 3 sensitivity analyses. First, the prevalence of childhood depression is disproportionately higher among individuals in service occupations and previous depression is a strong predictor of future depression (32). Therefore, to further account for the social selection of individuals with poor mental health into service occupations, we restricted analyses to respondents with no prior reported depression (i.e., those who had wave II and III Center for Epidemiological Studies Depression Scale scores ≤3). Similarly, we restricted assessment of the association between occupation type and sleep problems to respondents who reported never having difficulty falling or staying asleep or having difficulty “just a few times” in childhood (wave II). Measures of perceived stress in childhood were unavailable. Third, to examine the potential contributions of precarious work beyond the effects of underemployment (33), all outcomes were assessed with data restricted to those working full-time (≥35 hours/week).

RESULTS

Selected characteristics of 4,996 women and 4,144 men who reported a current or recent job during their wave IV interview are presented in Tables 1 and 2, respectively. Participants in the full analytic sample were, on average, 28 years old at wave IV (data not shown). Prior to PS-based sample restriction, women in service occupations tended to experience more adversity in early life, whereas women in nonservice occupations were more advantaged. For instance, parental income and educational attainment were highest for women in nonservice (mean income = $49,400; 36.1% college graduates) and lowest in untipped service occupations (mean income = $39,500; 19.4% college graduates). In contrast, parental incarceration was lowest among women in nonservice (14.0%) and highest among those in untipped service (23.8%) occupations. This trend was not as prominent in men. Among women and men, high educational attainment was most common among individuals in nonservice occupations. At the wave IV interview, women reported higher depression prevalence (across all occupation types: 25.6% in women vs. 13.5% in men).

PS distributions revealed 659 individuals with PS in regions where not all exposure levels were represented; an additional 3,080 individuals were below the fifth or above the 95th percentile of 1 or more PS distribution (Web Figures 2A–C, 3A–C). The PS-restricted analytic sample was thus reduced to 2,815 women and 2,586 men (Web Appendix 5). Compared with the full sample, women in the PS-restricted sample (Table 1) had parents with lower educational attainment (across all occupation types: 24.2% graduated college in PS-restricted sample vs. 32.1% in full sample) and household incomes (across all occupation types: $42,000 vs. $47,000), and had lower educational attainment themselves (across all occupation types 16.4% graduated college vs. 34.9%). Men in the PS-restricted sample similarly had lower educational attainment (Table 2). The following variables remained unbalanced after analytic sample restriction and were included as covariates in the multivariable models: participant educational attainment (for women and men), race (women only), and parental educational attainment (women only).

Occupation characteristics in the PS-restricted sample

The top 4 major occupation categories were as follows: sales and related, office and administrative support, food preparation and serving related, and construction and extraction occupations (Web Table 3); 60% of tipped workers were waiters, waitresses, and bartenders (Web Table 4). Job characteristics, such as shift type and access to paid leave, varied by broad occupation type (e.g., nonservice, untipped service, tipped service) (Web Table 5).

Multivariable models

Women in tipped service work had 61% higher odds of reporting depression diagnoses or symptoms relative to women in nonservice work (95% confidence interval: 1.11, 2.34) (Table 3). The association between untipped service work (vs. nonservice work) and reported depression diagnosis or symptoms was weaker and not significant (odds ratio = 1.25, 95% confidence interval: 0.93, 1.68). Associations for sleep problems and higher perceived stress tertile were not significant but of similar magnitude and direction. Although associations with depression, sleep problems, and perceived stress were not significant, they were all of greater magnitude for women in tipped relative to untipped occupations in women. Men exhibited an association similar to that seen in women for perceived stress, though it was weaker and not statistically significant.

Sensitivity analyses

Associations were stronger in samples restricted to women with no previous history of depression (for untipped service, odds ratio = 1.60; for tipped service, odds ratio = 2.98) (Table 4). Similarly, when analysis was restricted to women with no previous sleep problems, women in untipped service occupations had 72% higher odds of reporting sleep problems than did women in nonservice occupations. Stronger associations for depression or sleep problems were not observed in men. Associations for women and men were similar but statistically nonsignificant after sample restriction to full-time workers. Estimates obtained using the full analytic sample with logistic regression covariate adjustment were comparable in direction and magnitude (Web Table 6).

DISCUSSION

We investigated the association between occupation type and 3 adverse mental health outcomes within a nationally representative cohort of adolescents followed into adulthood. We observed cross-sectional associations between working in service occupations and poor mental health outcomes in women and men. Although only 1 of the examined associations was statistically significant—women in tipped service work had greater odds of reporting depression than did women in nonservice work—the magnitudes of associations were consistently highest among women in tipped service occupations. In men, associations were weaker and not statistically significant; these findings were confirmed in sensitivity analyses.

Our observation that service work was positively associated with depression is consistent with findings of previous analyses in which workers in the service industry were identified as having the highest prevalence of depression and the role of interpersonal conflict and encounters with difficult people were highlighted (18, 34). Observed associations with adverse mental health outcomes may further reflect the precarious nature of service work, which often entails lack of access to health-promoting benefits (5–7), low wages (35), and last-minute scheduling practices (3, 5). Associations with adverse mental health outcomes are also consistent with research in hotel employees, a subset of service workers that includes tipped and untipped workers in which job-related factors like low control, high psychological demands, and atypical work schedules are associated with a higher burden of morbidity (36–38).

Although observed associations for all outcomes were only significant for self-reported depression, they were of greater magnitude for tipped relative to untipped service work in women. In analyses restricted to women with no previous history of depression, women in tipped service occupations had greater odds of reporting depressive symptoms or a depression diagnosis than women in untipped service. This finding may reflect characteristics of tipped work that make it more precarious than untipped work, such as more unstable income (7) and greater emotional labor demands (39, 40). In a bivariable examination of job characteristics by occupation type, we found that a smaller proportion of study participants in tipped service had access to paid leave, health insurance, regular shift schedules, or freedom to make important decisions, compared with those in untipped service and nonservice occupations.

Higher odds of depression were not observed for men in tipped service. Women in tipped service occupations have more unstable income and earn less than men in tipped service occupations (wave IV household income: $48,500 for women vs. $60,000 for men). Nationwide, women in tipped occupations earn 6% less per hour than men (7). Part of this discrepancy is attributed to further gender-based occupational stratification. For instance, in our PS-restricted sample, women in tipped service occupations were largely restaurant wait staff (44.3% of all women tipped service workers vs. 30.8% of all men tipped service workers). However, even among wait staff, women are less likely than their male counterparts to work in fine dining establishments (41). Moreover, tipping practices are discriminatory. In 1 study, women were found to only have earned equivalent tips to men when their service was rated by the customer as “exceptional,” suggesting that women are being held to higher standards, especially by male customers (42). In another study, researchers found that male customers tipped more favorably if they found the female server attractive and/or she was wearing makeup (43). Discrepancies in the occupation category–depression association may reflect gender-based differential exposure to the discriminatory aspects of tipped service work. Researchers have also observed that the association between various psychosocial work exposures and poor mental health may differ by gender (44). In service occupations in particular, having to manage challenging customers may undermine gender-role authenticity, with detrimental mental health effects (45). As such, women may experience differential vulnerability to the emotional demands inherent to tipped and untipped service work.

Gender-based differences in observed associations may also be a product of gender differences in stratification into specific occupation types within the 3 categories we examined. Among the PS-restricted sample (which largely excluded professionals), 68% of the nonservice occupations filled by women were administrative, whereas 64% of the nonservice occupations filled by men were blue-collar physical labor–oriented occupations, with different job types providing a different constellation of physical, psychosocial, and environmental exposures. Our observations are consistent with data from the Department of Labor on gender segregation in the workforce (46) and may partially explain some of the weaker associations observed with the other outcomes in men.

Our analysis has limitations. First, because we restricted our analytic sample, generalizability is limited to individuals from lower- to middle-class upbringings without college degrees. However, use of PS enhanced internal validity. Although the unrestricted analysis yielded estimates that are nationally representative for adolescents followed into adulthood, these estimates were computed in a sample that contained individuals who were not exchangeable, due to social stratification, as evidenced by the nonoverlapping propensity distributions observed in Web Figure 2A. We further argue that given the bimodal distribution of the propensity for nonservice among women in service occupations, our sample restriction likely removed affluent, atypical individuals. For instance, women entering service work despite having a high propensity for nonservice work may be entering outlier occupations (e.g., fine dining establishments, high-end salons) and/or be selecting this type of work for the flexibility it affords. Second, our estimates may be biased due to unmeasured confounding, because PS only balance measured variables. We posit that estimates observed from our restricted sample are more conservative than those in unrestricted analysis to account for unmeasured confounding that may make an individual with a high propensity for entering 1 occupation still enter another. Third, exposure and outcomes are subject to misclassification error. Although it is likely individuals were appropriately classified as being in service-industry professions or not, some degree of nondifferential misclassification is expected on determination of whether the occupation was tipped. Notably, specific occupations for those reporting tipped service work were predominantly occupations that are less ambiguous in regard to tipping status in the United States. Here, we expected estimates would be biased toward the null. Also, self-report of study outcomes may introduce differential misclassification. The symptom-recall periods for measures of perceived stress and sleep problems were 1 month, whereas the recall period for depressive symptoms was 1 week. Researchers have observed a systematic bias in recall that is largest for those asked to reflect on a longer time (47). Regardless, the consistency of the magnitude and direction of the observed associations for these 3 mechanistically interconnected outcomes irrespective of differences in recall periods lends credibility to our observed associations. Furthermore, we leveraged the use of a prospective cohort initiated in childhood, which allowed us to minimize recall bias and ensure temporality of the items included in the study. Notably, our sensitivity analysis restricted to individuals with no history of depression yielded larger point estimates among tipped service workers. We posit that because individuals in untipped service work experienced disproportionately more childhood adversity relative to nonservice and tipped service workers and because of the relationship between childhood adversity and childhood mental health (48), this sensitivity analysis addressed further residual confounding related to childhood disadvantage. Fourth, 51% of the original Add Health sample did not complete all 4 waves and of those that did, 34% were missing 1 or more measure pertinent to PS development. We mitigated the potential introduction of selection bias by including sampling weights that accounted for loss to follow-up and multiply imputing in conjunction with PS development to address missing covariate data. Although there is concern that certain variables were not missing randomly, such as responses to childhood maltreatment questions, it is likely that any bias introduced would bias estimates to the null. Fifth, we were unable to perform risk-estimation procedures. Outcomes evaluated in this analysis are all prevalent in this population; however, we expected direction and magnitude of the associations to remain largely consistent, and this was observed for models we were able to evaluate (Web Table 6). Sixth, the incorporation of PS in our analysis was limited in that within our statistical software, we were unable to incorporate standard errors that accounted for multiple imputation. In addition, the 3-category exposure variable coupled with the complex survey design of the Add Health data limited our ability to incorporate PS beyond simple adjustment. Different approaches to the incorporation of PS in analyses can yield different results; however, our use of asymmetric trimming before performing analyses enabled us to procure estimates that are likely more similar to those that would be obtained through other PS methods (49). Finally, our analyses may have been underpowered to detect further significant differences between occupation categories due, in part, to small sample sizes, particularly among tipped service workers (50).

We conclude that the heightened precariousness of tipped service work may place individuals in these occupations, especially women, at increased risk of poor mental health. Optimal public policy and employment practices to alleviate the excess risk of depression in tipped service workers will depend on understanding the factors that underlie these differences in health status.

ACKNOWLEDGMENTS

Author affiliations: OHSU-PSU School of Public Health, Oregon Health & Science University, Portland, Oregon (Sarah B. Andrea, Lynne C. Messer, Miguel Marino, Janne Boone-Heinonen); and Department of Family Medicine, Oregon Health & Science University, Portland, Oregon (Miguel Marino).

The project described was supported by the Office of Research in Women’s Health and the National Institute of Child Health and Human Development, Oregon Building Interdisciplinary Research Careers in Women’s Health grant K12HD043488 (J.B.-H.) and National Institute of Digestive Disorders and Nutrition grant K01DK102857 (J.B.-H.).

For this research, we used data from the National Longitudinal Study of Adolescent to Adult Health, a program project directed by Dr. Kathleen Mullan Harris and designed by Dr. J. Richard Udry, Dr. Peter S. Bearman, and Dr. Kathleen Mullan Harris at the University of North Carolina at Chapel Hill, and funded by grant P01-HD31921 from the Eunice Kennedy Shriver National Institute of Child Health and Human Development, with cooperative funding from 23 other federal agencies and foundations. Special acknowledgment is due Dr. Ronald R. Rindfuss and Dr. Barbara Entwisle for assistance in the original design. No direct support was received from grant P01-HD31921 for this analysis.

Components of this work were presented by S.B.A. at the American Heart Association EPI Lifestyle Scientific Sessions on March 9, 2017, Portland, Oregon; and the Society for Epidemiologic Research Annual Meeting on June 22, 2017, Seattle, Washington.

Conflict of interest: none declared.

Abbreviations

 
• Add Health

  National Longitudinal Study of Adolescent to Adult Health

 
• PS

  propensity score.

REFERENCES