Can Personalized Attendance Information Mitigate Student Absenteeism? Evidence from Six Randomized Field Trials

81,172 randomised PK-12 across six districts; 78,732 in the analysis sample after 2,440 withdrew between randomisation and launch

US public school students, preK-12, six urban and suburban districts of varying demographics (44% Black, 28% Hispanic, 17% white; ~25% chronically absent in the prior year)

Six randomised field trials of the Rogers/Robinson-style personalised parent messaging intervention, pooled and analysed together with Poisson models, block fixed effects and SEs clustered on the randomisation unit (student in District A, household in B-F). Crucially, the districts designed and implemented the intervention themselves rather than through a third-party research vendor, which makes this the closest thing available to an estimate of what the intervention delivers under real operating conditions. Districts chose their own mode (mail, backpack, text, email, robocall), content and frequency (mostly one message every 4-6 weeks; 3-4 messages total per student). All six pilots ran pre-COVID (2018-19 or 2019-20, concluding by early March 2020). Not fully arms-length: the districts were members of Proving Ground, the Harvard CEPR research-practice partnership the authors run, which advised on pilot design and ran the power analyses; funding from the Gates Foundation and Carnegie Corporation. The intervention itself is not the authors' creation, so this is still the replication test for the flagship results.

Personalised messages reduced student absences by 1.9% (95% CI -3.1% to -0.6%, p=0.003) in the preferred fully-adjusted model — roughly 0.19 fewer days missed against a control mean of 9.93 days, or a 0.22pp fall in the absence rate from a control mean of 11.64%. Effects appear separately for both chronically-absent and lower-absence students with no significant difference between them, suggesting Tier 1 rather than targeted use. The magnitude is the finding: the paper's own literature table puts prior single- or few-district trials at 2.4% to 8.3% reductions in days absent, so this six-site self-implemented pooled estimate sits at or just below the bottom of that range. Robustness is thinner than the headline: the intermediate specification (block fixed effects, no covariates) gives -1.2% and is NOT significant (p=0.213), and of the five delivery modes only mail is individually significant. No achievement outcome is reported — the authors say they were underpowered to detect effects on test scores or grades.

Low - randomised across six independent sites. The relevant caution is not genetic but about scale-up attenuation, and about the fact that no trial in this family has yet measured whether the attendance gain buys any learning.

Directionally consistent with the flagship messaging trials but at or below the bottom of their effect range. Table 1 of the paper puts the comparison set at -2.4% (Rogers 2017), -6.2%/-6.5% (Rogers & Feller 2018), -7.7% (Robinson 2018), -5% (Musaddiq 2020), -8% (Heppen 2020), -8.3% (Himmelsbach 2022) on days absent; this pooled six-site estimate is -1.9%. Called `mixed` rather than `failed` because the sign and significance replicate — it is the magnitude that does not.