Predicting virologically confirmed influenza using school absences in PA

Talia Quandelacy, Shanta Zimmer, Chuck Vukotich, Rachel Bieltz, Kyra Grantz, David Galloway, Justin Lessler, Yenlik Zheteyeva, Amra Uzicanin, Hongjiang Gao, Derek Cummings

Abstract


ObjectiveTo determine if all-cause and cause-specific school absencesimprove predictions of virologically confirmed influenza in thecommunity.IntroductionSchool-based influenza surveillance has been considered forreal-time monitoring of influenza, as children 5-17 years old play animportant role in community-level transmission.MethodsThe Allegheny County Department of Health provided virologicallyconfirmed influenza data collected from all emergency departmentsand outpatient providers in the county for 2007 and 2011-2016.All-cause school absence rates were collected from nine schooldistricts within Allegheny County for 2010-2015. For a subset ofthese schools, in addition to all-cause absences, influenza-like illness(ILI)-specific absences were collected using a standard protocol:10 K-5 schools in one school district (2007-2008), nine K-12 schoolsin two school districts (2012-2013), and nine K-12 schools from threeschool districts (2015-2016). We used negative binomial regressionto predict weekly county-level influenza cases in Allegheny County,Pennsylvania, during the 2010-2015 influenza seasons. We includedthe following covariates in candidate models: all-cause school absencerates with different lags (weekly, 1-3 week lags, assessed in separatemodels using all other covariates) and administrative levels (county,school type, and grade), week and month of the year (assessed inseparate models), average weekly temperature, and average weeklyrelative humidity. Separately, for the three districts for whichILI-specific and all-cause absences were available, we predictedweekly county-level influenza cases using all-cause and ILI-specificabsences with all previously stated covariates. We used several cross-validation approaches to assess models, including leave 20% of weeksout, leave 20% of schools out, and leave 52-weeks out.ResultsOverall, 2,395,020 all-cause absences were observed in nineschool districts. From the subset of schools that collected ILI-specificabsences, 14,078 all-cause and 2,617 ILI-related absences werereported. A total of 11,946 virologically confirmed influenza caseswere reported in Allegheny County (Figure 1). Inclusion of 1-weeklagged absence rates in multivariate models improved model fits andpredictions of influenza cases over models using week of year andweekly average temperature (change in AIC=-4). Using grade-specificall-cause absences, absences from lower grades explained data best.For example, kindergarten absences explained 22.1% of modeldeviance compared to 0.43% using 12thgrade absences in validation.Multivariate models of week-lagged kindergarten absences, week ofyear, and weekly average temperature had the best fits over othergrade-specific multivariate models (change in AIC=-6 comparingK to 12thgrade). The utility of ILI-specific absences compared to totalabsences is mixed, performing marginally better, adjusting for othercovariates, in 2 years, but markedly worse in 1 year. However, theseresults were based on a small number of observations.ConclusionsOur findings suggest models including younger student absencesimprove predictions of virologically confirmed influenza. We foundILI-specific absences performed similarly to all-cause absences;however, more observations are needed to assess the relativeperformances of these two datasets.

Full Text:

PDF


DOI: https://doi.org/10.5210/ojphi.v9i1.7701



Online Journal of Public Health Informatics * ISSN 1947-2579 * http://ojphi.org