Student Disadvantage Raises Teachers’ Stress-Related Sick Leave in Swedish Schools

Krzysztof Karbownik frames the study around a workplace feature that is often treated as background: the people workers must deal with. In service economies, and especially in “contact intensive” or face-to-face jobs, much of the work consists of managing interpersonal situations and responding to others’ needs, emotions, and behavior in real time. Teachers, nurses, doctors, home aides, flight attendants, social workers, lawyers, servers, and clergy all fit that description in different ways.

The familiar research question runs from provider to client: how much do teachers affect students, doctors affect patients, or service workers affect customer satisfaction? Karbownik asks the reverse: how do client characteristics affect provider outcomes, and specifically provider health?

That question is hard to answer because the data requirements are unusually high. A researcher needs to match providers to clients at a micro level, observe meaningful characteristics of the clients, and follow provider outcomes over time. Wages alone are not enough, because they may include compensating wage differentials. Health outcomes raise their own measurement problems: some conditions are persistent rather than contemporaneous, some absences may be strategic, and mobility can itself be a response to the exposure being studied.

Teaching in Sweden is useful because the registers allow the researchers to link teachers, schools, students, parental characteristics, earnings, mobility, and medically certified sick leave. The setting also limits, though does not eliminate, one common confounder: teachers’ wages are relatively regulated, so schools do not simply double pay to compensate teachers for taking harder assignments. Karbownik said there is some room for wage variation, and later showed earnings differences, but not enough to make the wage channel the whole story.

The study covers Swedish lower- and upper-secondary teachers, roughly grades 7 through 12, from the 2005–06 school year onward, with administrative health data extending to 2023 or 2024 depending on the register. The analysis focuses on middle and high school because the researchers can match teachers to students at the school-year level more reliably than in grades 1 through 6.

The broader claim is not that teachers are unique. Karbownik’s argument is that teaching provides a measurable case of a more general issue: as more employment shifts into contact-intensive work, the composition and behavior of clients may become a central determinant of worker health.

The preferred outcome is a consequential absence: a sick leave spell lasting longer than two weeks, certified by a doctor and recorded in Sweden’s Social Security Register. Short absences are not in the main data. If a teacher calls in sick for a single Tuesday, that teacher is coded as zero in this measure.

The study distinguishes several health outcomes. The broadest is any sick leave spell of more than two weeks. Mental health is measured through psychiatric diagnoses, any ICD-10 F code. Burnout is examined separately through ICD-10 F43, which Karbownik described using the WHO definition: “reaction to severe stress and adjustment disorders.” The researchers also use cardiovascular disease and cancer diagnoses as a placebo outcome, because those conditions should not plausibly respond in the same contemporaneous way to changes in student composition.

The study is not merely about whether teachers feel stressed. It is about doctor-certified absences long enough to disrupt schools. That matters for workers, who face forgone wages, scarring, and possible family spillovers, and for schools, which face productivity declines, sickness-leave costs, and potential workplace contagion. In Sweden, mental-health-related sick leave accounts for 46% of all sick leave, with 22% carrying a stress diagnosis; work burnout has been estimated to reduce national labor income by 3.6%.

Teachers are also important because absenteeism affects students. Karbownik pointed to existing literature on negative effects of teacher absence and on teacher shortages in hard-to-staff schools. His proposed mechanism adds another link to the inequality story: disadvantaged schools may not only have trouble recruiting and retaining teachers; their teachers may also be more likely to become sick, especially with stress-related diagnoses.

Audience questions pressed the definition of health. Karbownik answered that the researchers examine both general health and mental health, with particular attention to acute mental health and burnout. Short “mental health days” or one-day absences are outside the main administrative measure. The paper is about more severe, certified spells.

Student composition is not measured by one variable such as parental income. The researchers construct an index designed to summarize socioeconomic background predictive of academic performance. They begin with grade 6 GPA, which is predetermined relative to lower- and upper-secondary teacher exposure, and regress it on a broad set of student and parent characteristics.

The inputs include parental education, income percentile within cohort and sex, employment, welfare receipt, cohabitation status, country or region of origin, and foreign-born status. The model also includes interactions with foreign-born status. For children, it includes age at migration and country-region of origin. The predicted value is then aggregated to the school-by-year level.

Karbownik described the result as “a composite measure of socioeconomic background that is predictive of student achievement.” The index explains about 27% of GPA variation in lower-secondary schools and 30% in upper-secondary schools. A between-school standard deviation in the index is 0.3, and the study scales treatment effects by that amount.

The index is oriented so that higher values mean more advantaged students. In the presentation, Karbownik often described results as the effect of a one standard deviation increase in student advantage; the inverse is the effect of more disadvantage.

What “disadvantaged” means in Sweden is not identical to what it might mean in a lower-safety-net setting. Karbownik was cautious on that point. He said the comparison is presumably not about the same level of food insecurity or neighborhood crime one might imagine in the United States. When the index is decomposed, however, the variables that stand out are immigrant-family background and parental education. Conditional on those and other factors, parental income “doesn’t pop that much.” One-parent versus two-parent household status matters in univariate comparisons, he said, but not conditional on the other variables.

The relevant exposure also appears to be average composition, not a single disruptive outlier. Karbownik said the researchers treated that as an empirical question by including both the mean and dispersion of the student index. The effects are driven by the mean; within-school dispersion does not appear to drive the results.

At the school level, the descriptive pattern is stark. In both U.S. and Swedish evidence Karbownik showed, schools with more disadvantaged students have more teacher absenteeism. He used U.S. Office for Civil Rights data for context: the y-axis was the fraction of teachers absent more than 10 days in a year, and the x-axis was the share of students receiving free or reduced-price lunch. The slope grew over time from zero in 2010 to 0.043 in 2018, and the interquartile range in the raw measure of teachers absent more than 10 days was 27 percentage points.

He was careful about the limits of the U.S. comparison. The U.S. data show absences, not medically verified sickness. In Sweden, by contrast, the researchers can tell that the absence is a doctor-certified sick leave spell. Karbownik said that distinction should matter for policy: if part of the U.S. absenteeism gradient reflects health rather than adverse selection or shirking, the conversation about teacher absence in disadvantaged schools may need to change.

In Sweden, the school-level relationship appears across any sick leave, psychiatric sick leave, and burnout. A slide interpreting the school-level estimates reports that moving from the 25th percentile to the 75th percentile of student advantage is associated with reductions of 8.0% in any sick leave, 9.1% in psychiatric-diagnosis sick leave, and 8.7% in burnout. Expressed by a one standard deviation increase in student advantage, the same slide reports 4.2 fewer sick-leave cases per 1,000 teachers per year, 3.2 fewer psychiatric sick-leave cases per 1,000, and 2.5 fewer burnout cases per 1,000.

Karbownik repeatedly separated this descriptive equilibrium object from causal identification. The school-level relationship contains sorting, mobility, exposure effects, neighborhood characteristics, and correlated shocks. It still answers a policy-relevant question—whether schools serving more disadvantaged students have less healthy teachers—but it does not by itself show that student composition causes sickness.

The descriptive work also uses survey evidence to establish a “first stage.” Schools with better work environments have fewer sick teachers. Schools with more advantaged students have better reported work environments. The School Inspectorate survey index includes school climate, school safety, teacher support, adjustment to student needs, practices against abuse, and related dimensions. With school fixed effects included, a one standard deviation increase in student advantage is still associated with a 0.16 standard deviation increase in the work environment index, although much of the work environment is fixed across schools over time.

The core teacher-level analysis uses within-teacher variation. The question becomes: when the composition of students in a teacher’s school changes, does that same teacher’s health change?

Teacher fixed effects control for stable teacher characteristics, including latent health, resilience, and other fixed traits. Teacher-school, or spell, fixed effects go further by comparing a teacher to herself within the same school, thereby removing the variation created by moves across schools. Karbownik acknowledged that this can be an endogenous control if student composition itself causes a move, but it is useful for addressing concerns that unhealthy teachers sort into different schools.

The main identification assumption is that within-teacher changes in student composition are not correlated with unobserved time-varying determinants of teacher health. Karbownik described this as standard peer-effects-style identification, relying on cohort-by-school variation in student composition. He also stated the assumption more concretely: teachers with latent health issues should not systematically sort into schools that, in the future, will have a larger share of disadvantaged students.

The teacher-level table reported negative coefficients for any sick leave, psychiatric diagnoses, and stress diagnoses when comparing teachers to themselves over time. The preferred teacher-fixed-effects estimates imply that a one standard deviation increase in student advantage decreases sick leave cases by 3.9% per year, psychiatric sick leave cases by 5.1% per year, and burnout sick leave cases by 9.4% per year. The same table showed essentially zero effects for the placebo category of cardiovascular disease and cancer.

The evidentiary hierarchy matters. The school-level gradients show where sickness is concentrated. The teacher fixed-effects estimates ask whether the same teacher becomes more or less likely to have a certified sick-leave spell as student composition changes. The teacher-school fixed-effects estimates ask whether the same pattern holds within the same teacher-school spell, reducing the role of moves. The mover event study then asks whether changes in mental-health absence appear around the time teachers move into schools with different student composition.

When teacher fixed effects are added, estimates for any sick leave and psychiatric sick leave decline by about 45%, indicating that some sorting is present. For burnout, the decline is much smaller. Karbownik interpreted that as consistent with the acute nature of the F43 diagnosis: if burnout is more contemporaneous, it is harder for workers to sort on it in advance.

When the researchers replace teacher fixed effects with teacher-school fixed effects, the estimates do not disappear. For psychiatric sick leave, they become larger in the reported table; for burnout, they remain similar. Karbownik interpreted this as evidence that endogenous mobility is unlikely to be the main driver of the burnout result, and that selection may bias the simpler estimates toward zero rather than away from zero.

The placebo is central to his credibility claim. If the estimates were driven simply by “latently sicker” teachers being clustered in disadvantaged schools, cardiovascular and cancer diagnoses might move as well. Instead, those estimates are close to zero and reasonably precise. The 95% confidence interval is -0.17 to 0.30 cases per 1,000 teachers for the placebo, compared with reductions of 1.2 to 2.7 cases per 1,000 for the outcomes of interest.

The mover design restricts attention to teachers who switch schools. The treatment is the difference in average student composition between the destination and origin schools. A positive value means moving to a more advantaged school; a negative value means moving to a less advantaged school. The event study uses the timing of moves, comparing those who move earlier with those who have not yet moved.

The sample drops non-movers, teachers who exit teaching, and teachers who were not employed for at least three years before the move. The window spans three years before and after the move and includes roughly 6,500 distinct moves. Karbownik stressed the external validity caveat: movers may differ from stayers. But movers allow the analysis to ask whether health trends change at the time a teacher moves to a school with different student composition.

The event-study identifying assumptions are the usual ones: parallel trends, no anticipation, no co-timed shocks, and no interference that would invalidate the comparison. Karbownik said the pre-trends look flat. For psychiatric diagnoses, the graphs show an immediate drop after moving to a better school and then some bounce back. For stress diagnoses, the drop appears more persistent. He read that pattern as consistent with burnout being acute and responsive to a less stressful environment.

Asked whether any school change might itself generate stress, Karbownik said the effects appear symmetric. Teachers moving to worse schools show increases; teachers moving to better schools show decreases; teachers moving to similar schools show flat lines. The standard errors become large when the sample is split, but the pattern remains in that direction.

The mover evidence does not eliminate every concern. It applies only to movers, and the timing of moves must be treated as plausibly unrelated to other health shocks except through the school change. Its value in Karbownik’s account is that it points in the same direction as the fixed-effects estimates: mental-health sick leave changes when exposure to different student composition changes.

Karbownik described the robustness work as extensive and “surprisingly robust.” The checks include alternative constructions of the student index, different sample restrictions, excluding or including the COVID period, Stockholm-specific concerns, workplace contagion, correlated local shocks, reverse causality, and using only within-spell variation.

COVID came up because Swedish schools remained open in ways that differed from many other countries. Karbownik said the results are very similar whether COVID is included or excluded. In the raw Swedish data, burnout went down during COVID while other psychiatric outcomes such as anxiety went up, though the main identification uses within-teacher variation that controls for broad period changes.

Workplace contagion is another possible explanation. If one teacher’s absence raises the workload or stress of colleagues, the relationship between student composition and individual sickness might partly reflect peer absence rather than direct exposure to students. The researchers address this by looking at other teachers’ contemporaneous sick leave in the school, and it does not seem to matter.

Local shocks are also considered. A neighborhood economic shock could affect both the student composition of a school and the mental health of teachers living nearby. The researchers address this with municipality-by-year fixed effects and by controlling for sick leave among non-teachers in the teacher’s neighborhood. These controls do not materially change the estimates.

Reverse causality is tested directly. One table asks whether teacher sickness in period t predicts the student composition a teacher faces in period t+1. For psychiatric and stress-related sick leave, it does not. Karbownik presented that as evidence against the story that sick teachers are systematically moving into or being assigned to schools with different future student composition.

Moral hazard received explicit attention. One concern is that teachers exposed to more disadvantaged students may seek a diagnosis or use available leave channels to avoid those students rather than because the exposure made them sick. Karbownik’s response was twofold. First, the main sick leave measure requires a doctor-certified spell longer than two weeks, and the teacher loses income on leave. Second, the researchers examine leave for caring for sick children, which he described as less scrutinized and easier to use strategically. They do not find effects there.

The study also examines earnings and mobility. With teacher fixed effects, higher student advantage is associated with a lower probability of not working at the same school the following year and a lower probability of not working as a teacher the following year. Higher student advantage is also associated with higher annual earnings, with positive log-earnings coefficients in the table.

Karbownik earlier described the compensating wage differential effect as about half a percent. The main point he drew from the Swedish setting was institutional rather than numerical: salaries are regulated enough that disadvantaged schools are not simply offsetting health risk by paying dramatically more. There is some wage movement, but not a compensating differential large enough to make the health channel disappear.

Audience members pushed on the policy implications. If teachers face a higher probability of mental illness in disadvantaged schools, why not pay a compensating differential—“combat pay,” as one participant put it? Karbownik agreed that compensating wage differentials would be an obvious economic response and might make the system fairer, but he did not present it as sufficient. Paying people for misery does not necessarily remove the cause of the sickness.

He also resisted the idea of simply selecting teachers by mental-health resilience. Matching teachers who can withstand stress to harder environments may already occur through who remains in the profession, but he called that a blunt instrument. He said he would not want to “select out good teachers just because they cannot deal with the bad students.”

Inexperience matters. Karbownik said the largest negative health effects appear among rookie teachers. He connected that to a broader inequality mechanism: if inexperienced teachers are more likely to end up in disadvantaged schools, and if those teachers are more vulnerable to the health effects of difficult student composition, the assignment process may reinforce a vicious cycle.

The private-school result also stood out. Sweden has a large private school sector, and Karbownik said the effects are larger there. He noted that private schools are less regulated and that U.S. evidence is scarce because private-school data are limited, though anecdotal evidence suggests similar issues could arise.

The final mechanism analysis draws on three survey sources: the School Inspectorate’s Teacher Survey, the Swedish Work Environment Survey, and the Attitudes to School Survey. The goal is to distinguish direct client interaction from general organizational dysfunction. In the organizational-management framework, schools with disadvantaged students might have worse principals, weaker collegial support, lower autonomy, heavier workload, worse management, or less flexibility. If so, student composition would be proxying for organizational features.

Karbownik said the evidence points elsewhere. The mechanism slide separated survey measures into items that moved with student composition and items that did not. Measures tied to direct student interaction—respect by students, peacefulness, social interaction, student interaction, and related indicators of threats or conflict—showed the visible separation. Measures more familiar from organizational behavior and management—workload, management values, principal values, peer values, flexibility, satisfaction, and similar organizational features—clustered much closer to zero.

That distinction shaped his policy interpretation. If the main channel were dysfunctional school management, the natural intervention would be organizational reform: better principals, more autonomy, better staffing, better workflow. If the main channel is conflict and disrespect in the classroom, then the interventions point more toward student behavior, classroom climate, and reducing threats or abuse toward teachers.

Karbownik suggested behavioral interventions in disadvantaged schools aimed at conflict and “student bullying teachers.” He cited, as an analogy, work such as Becoming a Man in Chicago, involving relatively troubled students and violence or abuse in school settings, while noting that those studies did not link outcomes to teacher health in the way Swedish data allow.

One audience question raised the possibility that the Sweden result partly reflects cultural mismatch, especially where immigrant background is an important component of the student index. Karbownik did not dismiss the possibility, but he did not claim to identify it. He said a quasi-random immigrant placement policy in Sweden and Denmark would be a promising way to study that issue, but there are not enough observations in given locations to provide power. The survey evidence, he said, shows verbal abuse and conflict, including physical conflict. Whether those conflicts reflect culture or other dimensions of disadvantage is not something the study can resolve.

Karbownik’s central empirical claim is that the quasi-experimental evidence points to student composition as a determinant of teacher health, especially mental health and burnout. He repeatedly qualified that claim: there is no randomized experiment, no clean regression discontinuity, and no single design that eliminates every concern. Instead, the evidence comes from multiple strategies that point in the same direction: school-level gradients, within-teacher changes, teacher-school fixed effects, mover event studies, placebo outcomes, and robustness checks for mobility, local shocks, contagion, COVID, and moral hazard.

The narrowness matters. The study does not show what exact intervention would reduce teacher burnout in disadvantaged schools. It does not separate every possible dimension of disadvantage. It does not observe classroom-level matches between teachers and specific students, and Karbownik said aggregating to the school level means any within-school reallocations by principals are “baked in” rather than separately identified. It also does not yet fully address the health of teachers after they leave teaching; Karbownik said the team is working on imputations and bounds because they can observe health after exit but cannot observe continued treatment.

Still, the substantive implication is clear in Karbownik’s account. If disadvantaged schools have more teacher sickness partly because the student environment itself contributes to stress-related illness, then teacher absenteeism is not only a staffing or discipline problem, and not only an indicator of teacher quality or motivation. It is a workplace health outcome.

That reframes a familiar education inequality loop. Student disadvantage may worsen classroom conditions; classroom conditions may worsen teacher health; teacher sickness and absence may reduce school productivity and student learning; and the burden may fall especially on inexperienced teachers and hard-to-staff schools. In this account, the client composition of the classroom is not incidental to the job. It is part of the job’s health risk.