Flaking Out: Student Absences and Snow Days as Disruptions of Instructional Time

1,723,606 student-year observations (math), 1,717,509 (ELA); 19,194-19,846 school-grade-year cells

Massachusetts public school students, grades 3-8 and 10, 2003-2010 school years, capped at 60 absences/year (top 1% excluded)

Three nested designs on the same data. (1) School-grade fixed effects. (2) Student fixed effects, identifying off within-student year-to-year variation in days missed. (3) IV exploiting a non-linearity in snowfall - moderate snow induces individual student absences while extreme snow (10+ inches) induces school closures - so absences and closures are separately instrumented. NBER working paper; never published in a peer-reviewed journal. Author explicitly flags that the IV LATE falls on marginal compliers who are "disproportionately disadvantaged", and that the exclusion restriction fails if snow also moves unobserved tardiness or teacher absences.

The headline 0.05 SD figure is the IV estimate and is 2.5x the paper's own school-grade FE estimate (-0.020) and ~6x its student FE estimate (-0.008). Within-student (Table 2, cols 2 and 4): one additional day missed reduces math scores by 0.008 SD and ELA by 0.008 SD; the STUDENT FE decomposition is own absence -0.008, peer absence -0.008, closure +0.002 (null in math, marginally positive in ELA). The SCHOOL-GRADE FE decomposition is different and larger: own absence -0.020 math / -0.016 ELA, closure -0.001 / -0.000. The IV (Table 6) puts each snow-induced absence at -0.052 SD (SE 0.025, two instruments) or -0.053 SD (SE 0.021, ten instruments) in math and each closure at essentially zero, ruling out closure effects larger than 0.01 SD/day in both subjects. Two things that qualify the headline: the IV effect of absences on ELA is a NULL (-0.014, SE 0.021), so the "absences hurt" result is a math result; and in POOR schools closures do significantly hurt (-0.006 SD/day school-grade FE, -0.014 SD/day IV in math, -0.016 SD/day IV in ELA), so the closure null is an average, not a universal. Goodman reads the closure null as Lazear-style coordination: teachers absorb synchronised disruption but not staggered disruption. The "-0.055 central estimate" appears only in footnote 19 and matches no table value.

Partially addressed, not eliminated. Student fixed effects purge time-invariant child and family traits, which is the main hereditarian worry, and the within-student estimate is the one that should be believed. The IV does not fix this as cleanly as it appears: whether a child is absent on a moderately snowy day when school is open is a family decision, and snow raises absence rates most for disadvantaged families - the compliers are selected on exactly the parental-organisation and child-health traits that independently predict achievement. The 6x gap between the within-student and IV estimates is consistent with selected compliers rather than a larger true dose-response.

Split record, which is why this is mixed rather than unreplicated. The absence half is independently corroborated by a much cleaner design — Liu, Lee and Gershenson (2021) identify off within-student between-subject variation and also find absences reduce achievement. The closure half CONTRADICTS the three papers that established the snow-day literature (Marcotte 2007; Marcotte and Hemelt 2008; Hansen 2013), all of which found bad-weather closures reduce achievement; Goodman argues they mis-identified because snowfall moves absences as well as closures, violating the exclusion restriction in a single-instrument design. No independent study has since adjudicated that dispute using his non-linearity.