Improving health and economic security by reducing work schedule uncertainty

Those who reported working in Seattle and whose employers were of the size and type that would be covered by the Secure Scheduling ordinance constitute the “Seattle treatment group.” Our sample includes 754 covered workers in Seattle at baseline, 747 Seattle workers surveyed in the first year of implementation, and 441 Seattle workers surveyed in the second year after the law took effect. Because turnover in the service sector is very high (6), we sought to reinterview baseline respondents but also replenished the sample at each subsequent wave. The workers in the first- and second-year surveys included a new cross section of workers employed by covered retail and food service employers and some workers who responded to a previous survey. All of our analyses adjust SEs to account for the nonindependence of observations for those surveyed more than once.

The comparison group consists of workers employed by the exact same set of companies, all large multistate employers, that were covered by the Seattle law but were employed in other large US cities. We restricted comparison cities to those that have a minimum wage that is higher than the federal minimum wage to align with Seattle’s progressive labor policy environment. We only include metropolitan areas with at least 1 million people so as to compare Seattle workers to those in other large urban areas. The comparison group consists of a pooled sample of workers in the 24 metropolitan areas that met those criteria: Baltimore, Boston, Buffalo, Chicago, Cleveland, Columbus, Denver, Detroit, Hartford, Jacksonville, Las Vegas, Los Angeles, Miami, Minneapolis-St. Paul, Orlando, Phoenix, Providence, Riverside, Rochester, Sacramento, San Diego, St. Louis, Tampa, and Washington, DC. The comparison sample includes 5,394 workers at baseline, 7,734 workers in the first year of follow-up, and 2,619 workers in the second year of follow-up.

Table 1 shows that the Seattle and comparison workers are similar in their work schedule attributes at baseline. Seattle workers and those in comparison cities averaged around 2.8 out of 6 possible types of schedule unpredictability. Workers in comparison cities were more likely than Seattle workers to report working back-to-back closing then opening (clopening) shifts. Seattle workers and those in comparison cities were well matched at baseline on well-being indicators. Our difference-in-differences research design addresses baseline differences between Seattle and comparison group workers by design, given that we are comparing changes over time for Seattle workers to changes over time for workers in comparison cities.

The Seattle and comparison samples are also closely aligned on their demographic characteristics at baseline (SI Appendix, Table S2). All of our model estimates control for individual-level demographic, socioeconomic, and job characteristics. As a robustness check, we also estimate a separate set of difference-in-difference models using samples of Seattle workers and workers in comparison cities who are matched on educational attainment, school enrollment, job tenure, industry subsector, age, race, gender, parenthood, and marital status using propensity score matching (for further details, see the SI Appendix).

The data for this study are from a nonprobability sample and as such may differ from the broader population of workers in Seattle and comparison cities. When we compare our sample characteristics to workers in the American Community Survey, we see that our sample resembles the broader population of service sector workers in terms of age and educational attainment but that our sample is disproportionately female and White, non-Hispanic. To align our sample with the broader population of workers, we construct and apply survey weights, drawing on data from the American Community Survey. We describe the construction of these weights in the SI Appendix along with our multiple imputation approach for addressing missing values due to item nonresponse.

The Seattle law contained a complex set of scheduling provisions designed to regulate aspects of job quality that are not reported on by workers in any existing large-scale data set (such as the Current Population Survey or American Community Survey) or in administrative data. A key innovation of our approach is to field a survey questionnaire that was tailored to align with the provisions of the Secure Scheduling ordinance. We focus here on six work schedule measures that were regulated by the Secure Scheduling ordinance, given our aim of estimating causal effects of schedule predictability on well-being outcomes. The six schedule measures are 1) short notice of work schedules (less than 2 wk), 2) last-minute schedule changes, 3) last-minute scheduling changes without pay, 4) working back-to-back closing then opening shifts, 5) being asked to be on-call for work, and 6) shift cancellations without pay. We sum these six measures of schedule unpredictability into a scale ranging from 0 to 6 types of schedule unpredictability. Our analysis first estimates the impacts of Seattle’s fair workweek legislation on these measures of schedule unpredictability. Then, we estimate the effect of this unpredictability on well-being outcomes by using the Securing Scheduling Ordinance as an instrumental variable that parcels out endogeneity in a two-stage least squares analysis.

We also examine four self-reported measures of worker well-being: happiness (very or pretty happy), sleep quality (good or very good), psychological distress (frequently feeling sad, restless, nervous, hopeless, that everything was an effort, and worthless), and experience of one or more material hardships from among six types of hardship related to hunger or food insecurity, housing instability, difficulty paying bills, and deferring needed medical care.

The survey data, collected from workers before and after the ordinance took effect in Seattle and in comparison cities, allow us to estimate the ordinance’s effects using a difference-in-differences approach. We compare changes over time for workers at the large companies covered by law in Seattle with the changes for workers at the same companies that are not affected because they work in other cities without fair workweek laws in place. In particular, we estimate the change in outcomes between baseline and follow-up for Seattle workers (the first difference) after parceling out any change in outcomes over the same period for workers in comparison cities (the second difference). The estimation approach is displayed in Eq. 1.where $Y$ is an outcome of interest for individual i at time t (where time t is either baseline, year 1, or year 2); $X$ represents a vector of individual-level control variables; $\mathit{Seattle}$ is a dummy variable that is 1 for “Seattle treatment group” and 0 for “Comparison group” workers; $\mathit{YearOne}$ is 1 for the year one follow-up period and 0 otherwise; $\mathit{YearTwo}$ is 1 for the year two follow-up period and 0 otherwise; $\mathit{Seattl}{e}_{it}\_\mathit{YearOne}$ and $\mathit{Seattl}{e}_{it}\_\mathit{YearTwo}$ are interaction terms that are 1 for treatment group members when the ordinance was in effect in year 1 and year 2, respectively, and 0 otherwise. The terms of interest are the $\mu$ and $\rho$ coefficients, which represent the effect of the fair workweek law on outcome $Y$ after 1 and 2 y following implementation.

The difference-in-differences models assume the trends in outcomes are parallel in the treatment and comparison groups prior to the intervention and would continue on a parallel trend were it not for the intervention. Because schedule unpredictability data are not available in historic survey or administrative data, we do not have pretrend data available. However, the large national sample from the Shift Project demonstrates that, in the absence of a legislative change regulating work schedules, workers in the retail and food service sectors reported a high and steady level of schedule unpredictability from Spring 2017 through Spring 2019. Examining survey reports from service sector workers in localities without scheduling legislation, we find that levels of schedule unpredictability remain uniform over time in the absence of a legislative intervention. In particular, SI Appendix, Fig. S2 shows that on the summative scale of schedule unpredictability ranging from 0 to 6, the average worker reports about 2.8 types of schedule unpredictability and that level remains stable across three points in time: spring of 2017, 2018, and 2019. In all, a key strength of our approach is that we ensure significant homogeneity in the comparison and treatment samples by design (38). By focusing on large multistate employers who have substantial homogeneity in their scheduling practices across geography absent regulation, we impose substantial constraints on the possibility of unobserved heterogeneity. In contrast, prior literature on minimum wage effects (39, 40) or on paid sick leave (41) lack firm-identified information and so must pool across many very different employers by geography. Notably, our design is most similar to the canonical minimum wage study by Card and Krueger (42) that similarly constrained the comparison and treatment samples to the same employers.

Using the difference-in-differences approach, we assess whether the “treated” group—workers in Seattle covered by the Secure Scheduling legislation—experienced significantly greater changes in outcomes between baseline and follow-up compared with comparison nontreated groups. We estimate these difference-in-differences models using linear probability models for dichotomous outcomes. Linear probability models were chosen for their ease of interpretation and have been found to generate results that closely align with marginal effects from probit and tobit models (43).

We control for a set of demographic characteristics including age, race/ethnicity, sex, educational attainment, school enrollment, marital status, and presence of children in the household. We also control for job tenure, managerial status, and industry subsector (retail apparel, cafe, casual dining, department or big box store, fast food, grocery, hardware, health and beauty, and miscellaneous retail). We adjust SEs to account for repeat observations of respondents.

The Seattle policy change represented an exogenous shock to work schedules that we exploit to examine the causal effects of schedules on well-being outcomes in an instrumental variables framework. We combine the bundle of scheduling conditions that were affected by the Secure Scheduling ordinance into a “schedule unpredictability scale” that ranges from 0 to 6 types of unpredictability. In the first stage, we treat working in Seattle after the fair workweek law went into effect as the exogenous treatment, which we use to predict the endogenous schedule unpredictability scale (Eq. 2). In the second stage, we use the predicted value for the schedule unpredictability scale, now purged of endogeneity, to predict well-being outcomes (Eq. 3).

The first stage estimation uses ordinary least squares (OLS) regression, and the second stage estimates a probit model for each of a set of dichotomous dependent variables. We estimate these models in Stata 16 using the ivprobit command. We then use the model estimates from Eq. 3 to predict well-being outcomes under two scenarios: with the unpredictable schedule scale set to its average value of 3 of 6 types of unpredictability and with the scale set to 0 of 6 types of unpredictability. These estimates are generated using Stata’s margins postestimation command to simulate the difference between observed average baseline unpredictability and hypothetical elimination of unpredictability.

The instrumental variables approach relies on two key assumptions (44). First, the treatment—exposure to the Seattle Secure Scheduling ordinance—must be strongly predictive of the endogenous predictor: schedule unpredictability. Indeed, the F-statistic measuring the strength of the relationship between treatment and the endogenous scheduling scale is 17.1, above the conventional, acceptable threshold (45). Second, the exclusion restriction requires that the Seattle law treatment influence well-being outcomes only via the endogenous schedule unpredictability scale and not through other channels. By bundling all six scheduling features together rather than examining each measure individually, we avoid the risk of violating the exclusion restriction because the omitted scheduling features would represent an alternative avenue through which the fair workweek law treatment could have influenced outcomes. The law may have influenced outcomes not only by mandating schedule predictability but also by changing the composition of the workforce in Seattle relative to comparison cities, perhaps by retaining workers or attracting new workers. We find that the law did not change the sex, race/ethnic composition, or educational attainment of workers. The law did lead to a slightly younger workforce with slightly less job tenure compared with workers in comparison cities. Because younger workers with short job tenure have lower well-being than their counterparts, these small compositional changes will bias our causal estimates downward.

In Seattle, service sector workers experienced a similar level of schedule unpredictability in the spring of 2017 as workers elsewhere before the Secure Scheduling ordinance took effect. Before the legislation went into effect, workers in Seattle and other cities reported an average of nearly 3 out of a possible 6 types of schedule unpredictability. After the legislation went into effect, workers in other cities continued to experience the same levels of schedule unpredictability, but Seattle workers experienced a 0.50-point reduction in unpredictability (an 18% reduction). These results are shown in Fig. 1A.

Fig. 1B displays the impacts that the Secure Scheduling ordinance had on the six specific types of schedule unpredictability that are summed into the unpredictability scale. The Secure Scheduling ordinance stipulated that workers should receive at least 2 wk notice of their schedule. The ordinance increased the share of workers who received at least 2 wk notice of their work schedules by 11 percentage points. Before the law went into effect, 57% of Seattle workers received less than 2 wk schedule notice (Table 1) and that share had declined by 11 percentage points by Year 2 (Fig. 1B). Meanwhile, in comparison cities, the share who received less than 2 wk notice was 55% at baseline (Table 1) and remained the same in Year 2 (Fig. 1B). The difference between the change for Seattle workers (11 percentage points) and the change for workers in other cities (0 percentage points) is the 11 percentage-point impact estimate.

The ordinance generated a large impact on last-minute shift changes without pay. This practice decreased by 19 percentage points for Seattle workers in 2019 compared with the period in 2017 before the law went into effect. Over that same time period, this practice decreased by 6 percentage points for workers in other cities. Thus, we estimate that the net impact of the law was 13 percentage points (19 percentage points minus 6 percentage points).

The Secure Scheduling ordinance was associated with modest declines in on-call shifts (7 p.p., P ≤ 0.07), back-to-back closing then opening (clopening) shifts (6 p.p., P = 0.09), and in cancelled shifts without pay (3 p.p., P = 0.23). Seattle workers experienced slight reductions in each of these work conditions, whereas workers in comparison cities reported a slight increase in these conditions.

Our estimates treat the passage and implementation of the Secure Scheduling ordinance as exogenous. However, the factors that shape city selection into legislative action could confound the effects we estimate. We propose a strategy to test this potential source of bias. We re-estimate our models, comparing Seattle workers to their counterparts in cities or states that had considered scheduling legislation but had not yet enacted it. As shown in the SI Appendix, Table S3, we find very similar results with this alternative estimation.

We also compare Seattle workers to their counterparts who worked just outside of Seattle city limits. Here, we find somewhat smaller impacts on schedule unpredictability. One reason for this attenuation could be that the effects of Seattle’s ordinance spilled over to affect workers in the surrounding area.

The SI Appendix, Table S3 contains the difference-in-differences estimates (the coefficients on the interaction of Seattle worker × postordinance period) for each scheduling dependent variable for each of the four comparison groups. This appendix table also contains results from a fifth comparison group derived using propensity score matching to closely align background characteristics between Seattle and the matched comparison group without applying survey weights. The impact estimates on scheduling outcomes derived from each of the five comparison groups are similar (SI Appendix, Table S3).

Although the central aims of the Secure Scheduling ordinance were related to work scheduling practices, by increasing schedule predictability the ordinance also had positive effects on workers’ well-being.

Fig. 2 shows the trends in levels of happiness for Seattle workers and workers in comparison cities. Reported happiness was similar for workers in Seattle and in comparison cities before the ordinance went into effect (Table 1). Happiness levels slightly improved for Seattle workers after the law took effect and declined for workers in comparison cities. On net, the law was associated with a 7 percentage point increase in happiness for Seattle workers relative to their counterparts (Fig. 2 and SI Appendix, Table S4). The ordinance did not have a statistically significant effect on Seattle workers’ reports of psychological distress.

The ordinance also led to improvements in sleep quality for Seattle workers. The impact estimate was an 11 percentage point increase in “good” or “very good” sleep quality as opposed to “fair” or “poor” sleep quality. This positive impact came about largely because reported sleep quality improved over time for Seattle workers while it stayed the same over time for workers in comparison cities.

The Secure Scheduling ordinance led to a 10 percentage-point decrease in reports of at least one of the material hardships described previously, such as food or housing insecurity. This improvement was driven by a reduction in hardships for Seattle workers and stability in hardships over time for workers in comparison cities.

In the SI Appendix, Table S4, we show that these results were consistent across most alternative comparison groups but somewhat weaker for the comparison group comprised of workers employed near Seattle but outside of city limits, which, as discussed previously, could reflect spillover effects of the ordinance.

The legislative change in Seattle provides an opportunity to estimate the causal effects of precarious work schedules on well-being outcomes using an instrumental variables approach. We generate causal estimates of the relationship between schedule unpredictability and well-being outcomes, using the Seattle Secure Scheduling ordinance as an instrumental variable. Table 2 contains probit coefficients from the second stage of the two-stage least squares modeling. The Table 2 estimates imply that increases in the number of types of schedule unpredictability cause statistically significant reductions in sleep quality and increase the share of workers experiencing at least one material hardship. Schedule unpredictability was associated with a reduction in happiness, but this relationship fell short of statistical significance (P = 0.08). Schedule unpredictability does not appear to increase psychological distress.

Table 3 draws on the two-stage least squares estimates and generates predicted values for the well-being outcomes for workers with an average level of schedule unpredictability and then simulates the expected level of well-being outcomes if schedule unpredictability were eliminated. The Table 3 estimates suggest that eliminating schedule unpredictability would improve happiness by 24 percentage points (from 68 to 92%), although this result should be interpreted with caution given that the IV estimate fell short of statistical significance (P = 0.08). The models also estimate large effects on sleep quality and material hardship. Eliminating all forms of schedule unpredictability would increase the share of workers reporting good or very good sleep quality by 49 percentage points (from 31 to 80%). The models also predict large reductions in the experience of at least one material hardship. Eliminating schedule unpredictability would reduce the share of workers experiencing at least one material hardship by 45 percentage points (from 64 to 19%).