Dissertations and Theses

Date of Degree

6-3-2026

Document Type

Dissertation

Degree Name

Doctor of Philosophy (Ph.D.)

Department

Epidemiology and Biostatistics

Advisor(s)

Heidi Jones

Committee Members

Denis Nash

Katarzyna Wyka

Kevin Konty

Subject Categories

COVID-19 | Epidemiology | Longitudinal Data Analysis and Time Series | Multivariate Analysis | Public Health | Statistical Methodology | Statistical Models

Keywords

New York City, K-12 public school students, Childhood obesity and body mass index, Health disparities, Multiple imputation, Mixed-effects models

Abstract

BACKGROUND

The COVID-19 pandemic profoundly disrupted children’s daily routines through school closures, social isolation, altered dietary patterns, reduced physical activity, and heightened stress – conditions highly conducive to weight gain. Understanding how these disruptions affected childhood body mass index (BMI) trajectories, and whether they exacerbated pre-existing racial/ethnic disparities, is essential for public health planning and pandemic preparedness.

Children with obesity face elevated risks of cardiometabolic complications, psychosocial consequences, and strong tracking of elevated weight into adulthood. Yet shifts in weight trajectories carry meaningful health implications across the entire BMI distribution. Rapid weight gain during childhood is associated with insulin resistance, elevated blood pressure, dyslipidemia, and altered brain structure and executive function – even among children remaining within “healthy” weight categories.

This dissertation examined the pandemic disruption period’s impact on BMI among students in New York City (NYC) public schools – the largest and most diverse school system in the United States. Data were drawn from school-based FITNESSGRAM assessments conducted as part of the standard physical education curriculum, linked to administrative enrollment records. This linkage supports longitudinal tracking of individual trajectories (Aim 1), explicit examination of selective attrition and bias correction for the full baseline cohort (Aim 2), and cross-sectional surveillance of enrolled populations (Aim 3).

The pandemic disruption period is defined as the period when NYC public schools suspended in-person learning – from March 2020 through September 2021. School closures eliminated structured daily routines regulating eating patterns and sleep schedules; nutrition programs ensuring access to meals often more nutritious than home-based alternatives; physical education and recess; active commuting; and school nurse health monitoring. The loss of this supportive environment occurred alongside broader disruptions including stay-at-home orders, social isolation, food insecurity, increased sedentary behavior, and psychological stress.

METHODS

Three complementary analyses addressed the overarching research question. The primary outcome was the adjusted logarithmic percentage from median age- and sex-specific BMI (ALPM). BMI z-scores suffer from compression at extreme values, reducing sensitivity among children with elevated BMI. Alternative metrics improve sensitivity but share limitations: right-skewed BMI distributions produce asymmetric deviations, and the widening range of BMI values with age means a given percentage deviation has different meaning at age 5 versus 17. ALPM addresses these limitations through logarithmic transformation (producing symmetric deviations across the full spectrum) and age-specific adjustment (enabling meaningful cross-age comparisons). The CDC recommends ALPM for longitudinal BMI assessment, yet it remains underutilized. For interpretability, ALPM values were transformed to percentage deviations from median BMI using [(exp(ALPM/100)-1)×100]; this transformation is exact for absolute values (e.g., intercepts, predicted means) and approximate for coefficients representing differences, with high accuracy for effect sizes observed in this study. All transformed coefficient values reported hereafter are approximations.

Aim 1 used longitudinal interrupted time series (ITS) analysis with linear mixed-effects models to assess within-individual BMI trajectories among 425,983 students continuously enrolled from 2018-19 through 2022-23. The repeated measures design controlled for time-invariant confounding by using each student as their own comparison across time; student-level random intercepts modeled individual baseline differences and accounted for correlation among repeated measurements. This design estimates the causal effect of the disruption period on individual BMI trajectories, isolating within-person changes from compositional shifts in the student population. Multiple imputation addressed missing BMI measurements under missing-at-random (MAR) assumptions. Reduced measurement rates occurred during partial-collection years (2019-20: 56% due to mid-March closures; 2021-22: 76% due to delayed assessment resumption).

Aim 2 applied pattern-mixture modeling (PMM) to evaluate and correct for selective attrition among 613,344 students enrolled in grades K-8 in 2018-19, regardless of subsequent enrollment status. This aim addressed whether conclusions from Aim 1 generalize beyond the continuously enrolled population to the broader baseline cohort. Separate imputation models for each enrollment transition group (continuously enrolled, system departure, charter school, homeschool, special education, continuing education) accommodated missing-not-at-random (MNAR) assumptions. The PMM framework both evaluates whether selective attrition biases estimates of disruption effects and provides bias-corrected estimates for the full baseline population, distinguishing true disparities from selection artifacts.

Aim 3 employed repeated cross-sectional ITS analysis to characterize population-level BMI burden among K-8 students from 2016-17 through 2022-23 (excluding 2020-21 when data collection was suspended). Each school year represented a distinct cross-section of the enrolled K-8 population. Survey-weighted regression models accounted for clustering at both the student level (for repeated observations of students across years) and school level (for within-school correlation). All primary analyses used the additive scale: linear regression for ALPM and linear probability models for dichotomous obesity and severe obesity outcomes. Non-response weighting ensured observed measurements represent enrolled populations. This cross-sectional approach captured the combined effects of within-individual changes and enrollment dynamics (out-migration, in-migration, aging out of K-8, and transitions between educational settings), reflecting actual population burden that schools and public health systems address each year. Scale sensitivity analyses assessed whether effect modification findings were robust across additive and multiplicative specifications, using quasi-Poisson regression (for ALPM) and quasibinomial logistic regression (for dichotomous outcomes) on the multiplicative scale, addressing whether conclusions about disparities depend on measurement scale choice.

RESULTS

All three aims documented statistically significant immediate BMI increases following the pandemic disruption period, followed by substantial recovery as in-person schooling resumed. Immediate post-disruption shifts ranged from 2.47 to 4.40 percentage points above age-sex-specific median BMI; post-disruption slope changes ranged from -1.64 to -1.87 percentage points per year relative to pre-pandemic slopes, yielding net annual declines of -1.22 to -1.80 percentage points per year.

Within-person longitudinal analyses (Aims 1-2) revealed that by 2022-23, individual students’ observed BMI values fell below counterfactual projections – indicating that pandemic-induced weight gain not only reversed but overcorrected, with students tracking below where pre-pandemic slopes would have placed them. Immediate shifts among continuously enrolled students (Aim 1) were 2.65 pp, with net post-disruption slopes of -1.35 pp/year. PMM (Aim 2) demonstrated that these core effects remained robust even after bias-correcting for selective attrition across six enrollment pattern groups, indicating the surge-and-recovery pattern was not an artifact of differential enrollment transitions.

At the population level (Aim 3), the cross-sectional analysis showed substantial but incomplete recovery. By 2022-23, population-level BMI approached but remained slightly above counterfactual projections, with severe obesity prevalence 18.8% above pre-disruption levels and obesity prevalence 6.2% above. The cross-sectional analysis produced larger immediate shifts (4.40 pp) and steeper recovery slopes (-1.80 pp/year net) than longitudinal analyses, reflecting both analytic differences and the combined effects of within-individual changes and enrollment dynamics (out-migration, in-migration, aging out of K-8, and transitions between educational settings).

Substantial racial/ethnic disparities existed before the pandemic and persisted throughout. Hispanic students tracked 7-8 percentage points above median BMI compared to White students; Black students tracked 5-7 percentage points higher; Asian students showed levels similar to or slightly below White students. The pandemic widened the Hispanic-White gap through differential immediate impacts. In the continuously enrolled longitudinal analysis (Aim 1), Black and Hispanic students showed differential immediate effects of 1.16 and 1.58 percentage points beyond White students. However, after bias-correcting for selective attrition (Aim 2), only the Hispanic differential effect (+1.22 percentage points) remained statistically significant; the Black differential effect (+0.75 percentage points) attenuated below significance, revealing it was partly a selection artifact. Black and Hispanic students who departed had lower BMI than those who remained, while Asian students who departed had higher BMI. At the population level (Aim 3), Black and Hispanic student populations experienced additional immediate shifts of 2.03 and 2.10 percentage points respectively beyond White populations.

A critical methodological finding from the cross-sectional analysis (Aim 3) revealed that conclusions about effect modification depend fundamentally on measurement scale choice. For ALPM, immediate post-disruption effect modification remained robust across both additive and multiplicative scales (P< .001), achieving 88% concordance. For dichotomous outcomes (obesity and severe obesity), effect modification was detected on the additive scale (Black and Hispanic populations experienced larger immediate shifts; all P< .001) but not on the multiplicative scale – indicating substantial scale dependency (55-56% concordance). This pattern indicates that while all groups showed similar proportional effects, Black and Hispanic populations experienced larger absolute burden increases – information directly relevant for intervention targeting and resource allocation.

Comparing continuous and dichotomous outcomes (Aim 3) revealed different patterns of scale robustness. ALPM demonstrated high scale robustness with effect modification persisting across both scales (88% concordance), while dichotomous outcomes showed substantial scale dependency (55-56% concordance). This divergence suggests that pandemic impacts among Black and Hispanic populations were distributed across the full BMI spectrum rather than concentrated at obesity thresholds.

CONCLUSIONS

The COVID-19 pandemic disruption period caused immediate BMI increases among NYC public school students that subsequently reversed as in-person schooling resumed, demonstrating both the vulnerability of children’s health behaviors to major disruptions and their capacity for recovery when institutional supports are restored. The longitudinal analyses revealed that within individual students, BMI not only returned to pre-pandemic trajectories but overcorrected below projected levels by 2022-23. However, at the population level, substantial but incomplete recovery occurred, with obesity and severe obesity prevalence remaining elevated above pre-disruption projections. This distinction between within-person overcorrection and population-level incomplete recovery reflects enrollment dynamics and underscores the value of examining both longitudinal and cross-sectional perspectives. While the study design cannot isolate school closures as the sole causal factor, the pattern of results provides population-level evidence consistent with the structured days hypothesis and suggests that institutional supports schools provide are influential for children’s weight trajectories.

The pandemic disruption period widened the Hispanic-White BMI gap through differential immediate impacts not offset by differential recovery. PMM revealed that apparent Black-White differential effects observed in the continuously enrolled population partly reflected selective attrition rather than true differential health impacts, demonstrating the importance of addressing selection bias when population mobility is substantial. Yet persistent baseline disparities that predated the pandemic, briefly widened during the disruption period, and remained afterward remind us that recovery is not equity. Addressing childhood obesity disparities requires sustained attention to structural conditions that produce and maintain racial/ethnic inequities – conditions that the pandemic disruption exposed but did not create.

This dissertation contributes methodological insights alongside substantive findings. First, the comparison of longitudinal and cross-sectional approaches – and the distinction between within-person overcorrection and population-level incomplete recovery – demonstrates that both perspectives provide essential but distinct information for comprehensive understanding of disruption effects. Second, comparing continuous (ALPM) and dichotomous (obesity, severe obesity) outcomes demonstrates that both are essential for comprehensive BMI surveillance: ALPM provides greater scale robustness and captures changes across the full BMI distribution, while dichotomous classifications identify clinically meaningful threshold crossings for intervention targeting. Third, scale sensitivity analyses examining both additive and multiplicative specifications demonstrate that conclusions about effect modification can depend fundamentally on measurement scale, with additive scales identifying populations experiencing the largest absolute burden increases relevant for resource allocation. Fourth, the PMM framework provides a principled approach for distinguishing true disparities from selection artifacts during periods of population instability. These findings have direct implications for pandemic preparedness planning, which should weigh collateral health consequences of mitigation measures – including effects on childhood weight trajectories – alongside infection control benefits.

Download

Available for download on Saturday, October 16, 2027

Share

COinS