Data
We draw on unique data collected by the Shift Project, of which Dr. Schneider is the Co-PI. The Shift Project is an ongoing repeated cross-sectional survey that, beginning in 2017, collects two new cross-sections each year. The Shift Project uses Facebook/ Instagram as both quasi-sampling framing and a recruitment device, using Facebook’s sophisticated ad targeting system to construct “audiences” of workers at specific large named service-sector firms. The Shift Project then recruits these workers to the online survey by fielding paid advertisements that appear in workers’ newsfeeds on desktop and mobile. This approach is low-cost, very flexible, and allows for rapid-response data collection.
In this instance, the Shift Project approach allowed for the authors to undertake mid-COVID-19-pandemic collection of detailed data on qualifying events and leave-taking alongside economic outcomes for a large sample of vulnerable workers. To do so, we designed a new survey module on qualifying events and paid leave that we added to the Shift Project survey and that was fielded to 11,689 hourly service-sector workers at 119 large firms (a full list of these firms is provided in Additional file 1: Appendix Table A1) in two repeated cross-sections. A first group of respondents was surveyed between March and May of 2020 and a second independent group was surveyed between September and November of 2020.
Existing data on paid family and medical leave are much more limited than the data we draw on here. Few data sources, with the important exception of the ATUS Leave Module fielded in 2011 and 2017–2018, distinguish paid from unpaid leave, examine leave for reasons other than childbirth or bonding with a new child, or identify the source of pay (e.g., employer versus through a government program). The Shift Project data are the only large-scale data source that asks these detailed questions along with information on household economic security and worker subjective wellbeing.
However, these data are drawn using a non-probability sampling approach and Facebook-based survey data collection has a low response rate [24]. In this instance 12.2% of Facebook/ Instagram users who saw the ad, clicked-through to the survey and 10.4% of those who clicked-through contributed survey data, such that 1.3% of those to whom the ad was displayed contributed survey data. While this response rate is low, prior methodological work finds that univariate distributions and multi-variate associations in Shift replicate those in “gold-standard” data sources such as the NLSY and Current Population Survey [24] and these data have been used to examine the correlates and consequences of precarious job quality [7, 25, 26]. In these analyses, we weight the Shift Project sample to the characteristics of workers in the same occupations and industries in the American Community Survey on race/ethnicity, gender, and age and employ these weights in all of our estimates. Additional file 1: Appendix Table A2 contrasts the demographics of respondents in service-sector occupations with those of the weighted and unweighted analysis sample from the Shift Project data.
Measures
Leave taking
All survey respondents are asked if they experienced any of three types of events that would “qualify” them for paid leave under most existing state laws and company policies. Respondents were asked if they (1) “welcomed a new child into their family through birth, adoption, or foster placement” (not included in this analysis), (2) “had a serious health condition or illness, like recovering from a surgery or serious illness,” and (3) “have needed to care for seriously ill or injured family member.” For each item, respondents interviewed between March and May were asked about their experience with each event since January 1 of 2020; respondents interviewed between September and November of 2020 were asked about the prior 12 months. In this analysis, we focus on medical and caregiving events, dropping 368 respondents who reported a new child as their only qualifying event. At both waves, respondents could report more than one type of event, though respondents who experienced multiple events are asked about leave-taking for the combined event (i.e., “Did you take leave from your job to care for yourself or others?”). Reporting retrospectively on a reference period with a mean length of 7 months, 20% of these workers had experienced a need for medical or caregiving leave.
Respondents who reported at least one qualifying event were asked if they took leave from their job in response and, if they did, if they received pay from their employer, with options of receiving full pay, partial pay, or no pay. We draw on data from these two sets of measures to construct our key independent variable. We code respondents into four mutually exclusive categories: (1) did not experience any qualifying event, (2) experienced a qualifying event, but did not take leave, (3) experienced a qualifying event, and took unpaid leave, and (4) experienced a qualifying event, and took paid leave.
Economic insecurity
We construct four indicators of household economic insecurity. First, we measure if respondents currently find it “very difficult” (1) to cover expenses and pay bills versus finding it “somewhat” or “not at all” difficult (0). Second, we gauge if respondents reported that they could probably or certainly not come up with $400 in response to an unexpected need within the next month versus being probably able or certainly able to do so (0). Third, we distinguish respondents who did not pay the full amount of a gas, oil, or electric bill in the past month (1) from those who paid these bills in full (0). Finally, we measure hunger hardship if respondents report either receiving free food or meals because they didn’t have enough money or going hungry but not eating because they couldn’t afford enough food in the last month.
Wellbeing
In addition to these measures of economic security, we also measure two indicators of wellbeing. Respondents are asked “in general, how would you say things are these days?” We distinguish respondents who report being “very happy” or “pretty happy” (1) from those who are “not too happy” (0). We also measure respondents’ reported sleep quality during the past month, distinguishing those with “very good” or “good” sleep (1) from those with “fair” or “poor” sleep (0).
Controls
We measure and control for a set of demographic characteristics: gender; race/ethnicity (white, non-Hispanic; Black, non-Hispanic; Hispanic; other or multiple race/ethnicities, non-Hispanic); marital status (single, cohabiting, married); age; having children ages 0 to 4, ages 5 to 9, ages 10 to 14, and ages 15 to 18; current school enrollment; and educational attainment (< HS; HS/GED; some college; Associates degree; Bachelors degree; Masters degree or more). We also measure and control for a set of job characteristics: job tenure, union coverage, hourly wage, and number of usual work hours. Finally, we include a set of month and state fixed-effects. In a supplementary set of models, we also introduce controls for type of qualifying event (own health versus caregiving) and then for self-rated health.
Models/approach
To estimate the consequences of not taking paid leave when needed, we estimate ordinary least squares (OLS) regression models of leave-taking on a set of outcome measures that capture household economic insecurity and worker wellbeing.
Our analytic approach relies on exploiting the implicit time-ordering of events in these cross-sections. The leave-taking module asks about events in the recent past and we then gauge outcomes measured at the time of survey or in reference to the month prior. This survey structure allows us to correctly time-order qualifying events, leave-taking, and outcomes. The controls are measured at the time of survey.
Additionally, we note that recruiting workers employed at specific firms imposes a significant scope condition on the data. All respondents, in both Spring 2020 and Fall 2020, were employed at the time of survey. This means that our analyses only pertain to workers who returned to employment following a qualifying event. Workers who stopped working following a qualifying event are not included in the data.
These estimates are threatened by several potentially confounding processes. One possibility is that respondents who take paid leave may face the most severe health challenges or intensive caregiving responsibilities. That is, there may be negative selection into leave-taking. It is also possible that respondents who are able to take paid leave may be positively selected in being more knowledgeable about company policies or better able to navigate state systems. While the direction of bias is unclear, these models risk mistaking a spurious association between taking paid leave and worker outcomes for a causal one. In addition to the set of controls for possible confounders described above, we guard against these risks of confounding in several ways.
First, to guard against the risk of negative selection into leave-taking, we compare workers who took paid leave to those who took unpaid leave. We expect that workers who took paid leave will fare significantly better than both those who took unpaid leave and those with qualifying events who took no leave. However, we expect that workers without any qualifying events will fare best.
Second, as a robustness test, focusing only on workers with qualifying events, we control for the type of qualifying event, distinguishing medical and caregiving leave, and we then separately introduce a control for self-rated health as a conservative proxy for severity of event. While this risks “over-controlling” (in so far as not taking leave may make for worse health), it is also a powerful safeguard against negative selection into paid leave-taking.
Finally, to examine the risk of bias from different recall periods (i.e., two months for those interviewed in March 2020 to 12 months for those interviewed between September and November 2020), we conducted two additional sensitivity tests. We first control for implied maximum duration of the recall period. We then interact the maximum duration of recall period with our key independent variable and re-estimate all of the regression models.