The short- and long-run impacts of secondary school absences

Administrative panel of one large urban California district, school years 2002-03 through 2012-13, class-period-level attendance. Analytic samples: 112,711 student-year observations for math test scores, 117,445 for ELA, 333,624 in the stacked math+ELA model; 25,189 students for the 9th-grade long-run analysis and 31,163 for the 10th-grade one.

US middle and high school students (grades 6-12) in one large urban California district; 47% Asian, 22% Hispanic, 13% Black, 9% White, 25% ever-ELL, 8% chronically absent. Base rates for the 9th-grade cohort - 70% graduate on time, 62% ever enrol in college.

The cleanest identification in this literature FOR THE SHORT-RUN OUTCOMES, and it is important not to extend that credit to the long-run ones. Attendance is recorded by date AND class period, so for test scores and course grades the authors use within-student, between-subject variation - the same student's math absences versus their ELA absences in the same year - which differences out every student-level and family-level shock that hits the whole schedule. Two strategies give similar answers: a "proxy" model that conditions on total annual math+ELA absences, and a stacked model with student-by-year fixed effects. Adding the proxy HALVES the naive value-added estimate (-0.082 to -0.042 in math), which is itself a finding about the wider absence literature. Falsification test: absences occurring AFTER the annual state testing window have no effect on test scores, as they should not if the design is capturing lost instruction rather than student type - while the same absences still move course grades, so the placebo is not just a power failure. THE LONG-RUN ATTAINMENT ESTIMATES USE A DIFFERENT AND WEAKER DESIGN: the between-subject trick is unavailable there (total absences are the treatment, and there is no within-student variation in whether you graduated), so those come from selection-on-observables - school-by-year and neighbourhood-by-year fixed effects plus lagged achievement - shored up with Altonji/Oster bounds showing that selection on unobservables would have to be 2-5x that on observables to explain the estimates away. Grade B is earned by the short-run half. Two further caveats: it is a single district, and the paper does NOT estimate excused versus unexcused effects separately - it only reports that more than 90% of part-day (i.e. subject-specific) absences are unexcused while only about half of full-day absences are, so the identifying variation is overwhelmingly discretionary class-skipping rather than illness. NOTE FOR FUTURE READERS: the PDF held at education-papers/att-liu-lee-gershenson-2021-secondary-absences.pdf is the September 2019 IZA working paper (DP 12613), whose headline figures differ from the published ones - it reports 7% of an SD for SPRING-semester absences and 8%/7% attainment effects. The numbers in this record are from the published/accepted version (EdWorkingPaper 19-125, May 2021, identical abstract to JPubE 199:104441), where the headline is the pooled annual estimate.

Missing 10 class periods in a subject reduces that subject's state test score by 3-4% of a standard deviation and course grades by 17-18% of a standard deviation. Ten total absences in 9th grade reduce on-time graduation, immediate college enrolment, ever enrolling in college, four-year enrolment and two-year enrolment by roughly 1.3 percentage points each - about 2% relative to base rates of 70% and 62%. The test-score effect is roughly five times smaller than the grade effect, which is the expected signature of grades partly rewarding seat time and compliance rather than learning - and the timing analysis sharpens this: absences AFTER the testing window still cut course grades hard (-0.475 math, -0.608 ELA per 10) while having no effect at all on test scores. Timing dominates everything else: absences during the testing window are about 5-6x more harmful to test scores than absences earlier in the year. Effects are approximately linear with NO discontinuity at the chronic-absenteeism threshold, which leads the authors to conclude that "chronic absenteeism indicators, which are widely used in education policy-making, are arbitrary". Heterogeneity is almost entirely absent - no differences by gender, race or school poverty - with one exception the earlier extraction missed: ELA absences have NO effect in middle school (+0.001, SE 0.015) and the whole ELA effect comes from high school (interaction -0.048, P < 0.01).

The within-student between-subject design is the strongest available answer to the hereditarian objection: a heritable trait such as low conscientiousness or poor health cannot explain why a student is absent more from period 3 than from period 5 in the same year. Remaining worry: subject-specific absence is partly a choice (skipping the disliked class), and the classes a student skips are plausibly the ones they were already doing worst in, which would bias the estimate away from zero. The authors address this by showing that math and ELA absences have near-identical relationships with lagged math achievement (if skipping were preference-driven the ELA line would be flat) and that between-subject absence differences are small and stable across grades. The post-test-window placebo partially, but not fully, answers it. The long-run estimates do NOT inherit this protection - they are selection-on-observables, and the Oster bounds are the only defence.

The paper claims novelty and the claim looks right: "To our knowledge, this is the first credible evidence of the long-run harms attributable to secondary school student absences." What exists is corroboration from adjacent settings rather than replication of this design - Gershenson et al. 2017 obtained similar magnitudes for grades 4-5 in North Carolina using related methods, and Gottfried & Kirksey 2017 found the same late-year-absences-matter-more pattern in an elementary California district. Nobody has re-run the within-student between-subject strategy on another secondary district's class-period data, and it takes an unusually rich dataset to do so.