Chronic absenteeism — does raising attendance raise achievement?

Verdict

"Students who attend more score higher" is the most heavily-cited confounded correlation in American education policy. Absence tracks illness, poverty, family chaos, parental follow-through and child externalising behaviour — every one of which independently predicts achievement, and most of which are substantially heritable. Truancy itself is 45% heritable with zero detectable household variance once you compare ordinary siblings rather than twins. The raw gradient is close to uninformative about the return to a marginal day of school.

Split the policy claim into its two steps, because they fail differently.

Step 1 — is there a causal achievement return to a day of attendance? Yes, and it is small. The best-identified designs converge on roughly 0.003–0.008 SD per day absent. That is real, it compounds over a chronically absent career, and Swedish administrative data following 1930s cohorts to retirement finds it reaches attainment and lifetime earnings. But it is an order of magnitude below what the advocacy framing implies.

Step 2 — do attendance interventions convert attendance into learning? On the evidence, no. Personalised parent messaging is a genuine, replicated, cheap effect on attendance. It is also tiny: 1.9% fewer absences across six trials and 78,000 students, 0.2–0.5 days across 47 districts. And the two randomised trials that moved attendance and then bothered to measure achievement both came back null — including Check & Connect, which cut absences by 22.9%, the largest behavioural effect in the literature, and moved neither test scores nor GPA.

The single most damaging fact for the seat-time model comes from inside its own flagship paper. Goodman finds that individual absences hurt while school closures — an entire day of instruction lost by the whole class — do nothing at all, with effects larger than 0.01 SD per closure day ruled out in both math and ELA. If losing a full day of school for everyone is undetectable, the story that one child's absence costs 0.05 SD of math is not primarily a story about lost instruction.

What the evidence shows

Source	Design	Grade	Key effect
Goodman 2014	snowfall IV + student FE, 1.7M student-years, MA	B	Own absence −0.008 SD math (student FE); IV −0.05 SD; school closures null (>0.01 SD/day ruled out)
Aucejo & Romano 2016	policy variation in calendar + flu IV + family-year FE, NC	B	10 absences = 0.055 SD math / 0.029 reading; 10 extra calendar days = only 0.017 SD math
Liu, Lee & Gershenson 2021	within-student, between-subject class-period absences	B	10 class absences = −0.03–0.04 SD tests but −0.17–0.18 SD grades; post-test-window absences = null
Cattan 2022	Swedish 1930s cohorts → lifetime registers	B	Absence reduces attainment and labour income across the life cycle
van der Aa 2009	4,835 twins + non-twin siblings	B	Truancy h² = 45%; shared environment 25% in twins, 0% in ordinary siblings
Gershenson 2017	observational, ECLS-K + NC admin	C	Modest linear association; unexcused absences "twice as harmful" as excused
Guryan 2020 (Check & Connect)	large RCT, Chicago, grades 1–8	A	Absences −4.2 days (−22.9%) in grades 5–7; no effect on test scores or GPA; null in grades 1–4
Bergman & Chan 2019	RCT, automated parent text alerts	B	Attendance +12%, course failures −27%, no impact on state test scores
Berger 2025	six RCTs, 78,000 students, district-run	A	Absences −1.9% (95% CI 0.6–3.1%); achievement not measured
Swanson 2026	RCT, 47 rural districts / 16 states	B	Absences −1.7% to −4.5% = 0.2–0.5 days, $4.07/student; authors urge caution
Rogers & Feller 2018	mailing RCT, 28,080 students, one district	B	Chronic absenteeism −10%+; achievement never measured
Robinson 2018	RCT, 10 districts, K–5	B	Chronic absenteeism −15%; achievement never measured
Eklund 2020	meta, 22 controlled studies	C	Attendance g = 0.25 between-group vs g = 1.04 within-group; achievement not pooled
Liu & Loeb 2019	teacher value-added on class attendance	C	Teacher effects on attendance only weakly correlated with their effects on achievement

The credible causal estimates cluster an order of magnitude below the headline. Liu et al.'s within-student, between-subject design — the same student's math absences versus their ELA absences, which no heritable trait can explain — gives 0.003–0.004 SD per class period. Aucejo & Romano's flu-instrumented estimate gives 0.0055 SD per day. Goodman's own student fixed effects give 0.008 SD per day. The widely quoted 0.05 SD is Goodman's IV, which he reports is 2.5× his school-grade FE estimate and which is ~6× his student FE estimate, and which he explicitly cautions is a LATE on compliers who are "disproportionately disadvantaged."

Attendance interventions move a behavioural outcome and stop there. Four of the six intervention studies here never measured achievement at all. The two that did — Bergman & Chan and Check & Connect — found nothing. That is not a power problem: Check & Connect was a large district-partnered RCT that produced a 22.9% absence reduction, and Bergman & Chan cut course failures by 27% in the same experiment where test scores did not budge.

The grades-versus-tests split is the clearest single diagnostic. Liu et al. find absences move course grades five times more than test scores (0.17 SD vs 0.035 SD). Bergman & Chan cut course failures 27% while moving standardized tests zero. Course grades reward showing up, handing work in, and being visible to a teacher; standardized tests measure what a student knows. Most of what attendance buys is credit, not knowledge — which matters for graduation, and matters much less for learning.

Scale-up attenuation is real. Single-district flagship trials reported 10–15% reductions in chronic absenteeism. The six-site, district-self-implemented replication of the same mechanism gives 1.9% fewer absences with a confidence interval whose lower bound is 0.6%. The 47-district rural trial translates its own result honestly into 0.2–0.5 days per student per year — which, multiplied by the best causal estimate of 0.005 SD/day, implies an achievement effect of 0.001–0.003 SD.

Hereditarian-lens assessment

Risk: medium. The verdict rests on randomised trials and on quasi-experiments with genuine identification, so the headline conclusions are not themselves products of the confound. But the literature they are correcting is almost entirely observational, and the correction is large.

Absence is not an assigned treatment — it is a child-and-family behaviour with a measured heritability. van der Aa et al. put truancy at h² = 45%, and their sibling extension is the detail that matters: the apparent 25% shared-environment component is twin-specific (same classroom, same cohort, same friends), and for ordinary siblings all environmental variance is non-shared. The household-level component that attendance policy implicitly targets is close to zero. Conscientiousness, externalising behaviour and childhood health are all heritable, all predict absence, and all predict achievement independently of it.

Gershenson's excused/unexcused finding is the cleanest demonstration that the observational literature is measuring type, not dose. An unexcused absence and an excused absence remove exactly the same amount of instruction. They differ in what they reveal about the child and the parent. The observational estimate says unexcused absences are twice as harmful — which is not a fact about instructional time, it is the confound announcing itself.

Goodman's snowfall instrument is weaker against this objection than it appears. It is exogenous at the level of whether it snowed, but conditional on the school staying open, whether a given child is absent that day is a family decision — and Goodman reports that snow raises absence rates most for disadvantaged students. The compliers are selected on precisely the parental-organisation and child-health channels the hereditarian lens flags. That is a plausible explanation for why his IV is 6× his within-student estimate, and it is a better one than the alternative that the true dose-response is six times larger than his own panel says.

Sizing the gap. Within-student and IV estimates land at 0.003–0.008 SD/day, while the raw policy rhetoric implies attendance can close achievement gaps. Even a sympathetic observational paper's own arithmetic deflates the claim: Swiderski et al. calculate that returning post-COVID absence rates all the way to pre-pandemic norms would recover 0.017–0.025 SD, 13–19% of what full academic recovery requires. That is the ceiling on a total elimination of the post-pandemic absence surge, estimated by authors who are trying to make attendance look important.

No genetically informative study of absence → achievement exists. There is no twin, adoption or sibling-difference estimate of the effect of absence on learning anywhere in this literature. Liu et al.'s within-student between-subject design is the closest substitute and is genuinely good, but it identifies a within-year, within-child contrast, not a family-controlled one. Aucejo & Romano's family-year fixed effects appear only as a robustness check. The absence of a behavioural-genetic design on the central question of a literature this large and this policy-active is itself a finding.

Boundaries & what critics say

• This is not an argument for tolerating chronic absenteeism. A student missing 30+ days is missing more than instruction, and Cattan et al. find real long-run attainment and earnings consequences. The claim is narrower: the marginal day is worth little, and the interventions we have do not deliver enough marginal days to matter academically.
• Attainment may move where test scores do not. Liu et al. find 10 ninth-grade absences reduce on-time graduation and college enrolment by 2% each, and Liu & Loeb find high-attendance-value-added teachers raise high school completion more than high-achievement-value-added teachers do. Per this archive's outcome taxonomy, attainment is scored separately from achievement — the credentialing case for attendance is stronger than the learning case.
• The steelman for larger effects is Goodman's congestion model: staggered absences disrupt a class more than a synchronised closure, and the effect is superlinear in the number of absent students. It is a coherent story and explains the closure null. It is also unreplicated, and it predicts effects that no attendance RCT has produced.
• Nulls could reflect insufficient dose rather than a broken mechanism. Check & Connect's 4.2 days is the exception and the strongest counter: that dose was large and still produced nothing.
• Attendance intervention effects are real, just small. Eklund's g = 0.25 is on attendance outcomes only; the g = 1.04 from within-group pre-post designs in the same meta-analysis shows what uncontrolled evaluations of these programmes produce, and should be treated as regression to the mean.

Practical guidance

• Buy the cheap thing and expect the cheap result. Personalised absence messages to parents cost ~$4/student, replicate across dozens of districts, and reduce absences by 2–4%. Worth doing; not worth a business case built on test scores.
• Do not fund intensive mentoring on an achievement rationale. Check & Connect is the most endorsed intensive attendance programme in US practice, it produced the largest attendance effect in this table, and it moved neither test scores nor GPA. The authors' own note on modest impact per dollar should be read literally.
• Ignore the excused/unexcused distinction as an achievement signal. It is a marker of family type, not of instructional loss. It is useful for triage, useless for estimating harm.
• Chase causes, not counts. Absence is downstream of illness, transport, housing instability and disengagement. School-based health access, for instance, moves chronic absence far more than messaging does. An attendance rate is a symptom dashboard.
• If you must forecast, use 0.005 SD per day recovered. Multiply by the days your intervention actually returns — usually well under one — before promising anything.
• Never cite the raw gradient. "Chronically absent students are X behind" is a statement about which children are chronically absent.

Open questions

• No behavioural-genetic study of absence → achievement exists. A sibling-difference or discordant-twin estimate is the obvious missing design and would be cheap in Nordic register data.
• Why do school closures do nothing while individual absences do? Either teachers absorb synchronised disruption (Goodman's reading) or the individual-absence estimate is contaminated. These have opposite policy implications and nothing distinguishes them yet.
• Does any attendance intervention move standardized achievement? Two trials have tested it and both failed. The literature needs the messaging RCTs — which are large, cheap and already running — to simply link to test-score files.
• Is the return to attendance non-linear at the chronic tail? All credible estimates are approximately linear averages; whether the 40th absence costs more than the 4th is untested with adequate design, and it is the case on which the entire chronic-absenteeism policy frame depends.
• Cattan et al.'s identification is unverified here (full text inaccessible), so the long-run earnings result should be treated as promising rather than established.