How steeply does hardship rise with schedule unpredictability?

The survey behind this finding covers 37,263 hourly retail and food-service workers at 127 of the largest US firms, collected between 2017 and 2019. The sample skews the way the sector does: 73% female, mean age 33, median wage $11 per hour. Exposure is measured on a 4-point additive scale built from four employer practices: schedules posted with under two weeks notice, on-call shifts, shift cancellations, and last-minute timing changes. Each practice a worker is subject to moves that worker one step up the scale.

Across that scale, four separate hardship measures move in the same direction. Adjusting for wages, income, and hours, and moving from the most predictable to the most unpredictable schedules, the predicted probability of hunger hardship rises from 23% to 50%, residential hardship from 12% to 28%, medical hardship from 18% to 42%, and utility hardship from 24% to 44% (Schneider & Harknett, 2021). Any material hardship rises from about half to 70-75%.

What makes the pattern worth reporting is that the four domains draw on different budgets. Hunger, residential, medical, and utility hardship are recorded as distinct measures rather than folded into one headline index, and the gradient shows up in each of them. For base rates, across the full sample 30% report hunger hardship, 25% medical, and 31% utility.

These figures are model-adjusted predicted probabilities, not counts. Read the 50% as what the model predicts for a worker at the top of the unpredictability scale once wages, income, and hours are held constant, never as the share of workers who experienced hunger hardship. That distinction is not pedantry, it is the entire reason the finding carries weight.

The obvious explanation for hardship among hourly service workers is low pay, and at a median wage of $11 per hour that explanation has real force. The gradient persists anyway. Whatever produces it is not accounted for by unpredictable schedules being concentrated among the lowest earners, because earnings and hours are already in the model. One reading consistent with that structure is that volatility, rather than level, is doing the work: income that swings week to week is hard to budget against, and a canceled shift lands as a hole in one month's rent rather than as a slightly lower annual figure.

This is a cross-sectional survey on a non-probability sample. Workers were recruited rather than randomly sampled, and each respondent is observed once, so the design measures association and nothing stronger. Nothing here demonstrates that granting a worker two weeks notice would move that worker down the hardship gradient. That is a different claim, and it requires a different design. Holding wages, income, and hours constant strengthens the finding by removing the most obvious competing explanation, but adjustment is not identification: a control variable tells you what the gradient is not explained by, never which variable came first.

The defensible summary is that workers reporting more unpredictable schedules also report substantially more hardship, and that the association holds after pay and hours are accounted for. Treat this as evidence about the size and shape of a problem, not as a measured return on a scheduling change. The causal question, whether changing schedules changes outcomes, is what the randomized retail experiment covered separately in this library was built to answer, and that is where a business case should source its causal claim.

Hardship in these four domains does not live outside the schedule. The measures sit close to the mechanics of attendance. A worker under strain in the hunger, residential, medical, or utility domain is a worker whose ability to arrive on time and finish a shift is under strain too. Residential hardship in particular changes commute distance and transport reliability with no notice to the employer at all.

Be precise about what is inference here. The survey measures hardship. It does not measure attendance, turnover, or output, and it makes no claim about any of them. The step from hardship to reliability is an operational reading, not a result in this paper. Saying so plainly is what keeps the argument standing when a skeptical colleague opens the study.

The practical value of an additive exposure scale is that it names four things an operation already controls. Notice period, on-call practice, cancellations, and last-minute timing changes are separately countable in most scheduling systems, and most operations are worse at one of them than the others. Counting them per employee per month produces a local version of the study's exposure variable, which is a more useful management object than a generalized engagement score.