The study uses a randomized control design at the child/parent level based on children in 7 schools (6 schools with treated and untreated children, and 1 pure control school). Target children (second to fourth grade, aged 7-10) are overweight or obese at baseline. Schools were identified at the outset for inclusion in the study from a set of schools that participated in an earlier child nutrition survey.
The statistical model will be a linear regression for all outcomes to estimate differences in means (intent-to-treat). We will present unconditional differences in means (model 1), as well as differences adjusting for baseline differences of a given outcome, as well as pre-determined child (age in months, gender), parental (survey respondent relationship to child, education, age, work status, height) and household (family size, home ownership, number of bedrooms) covariates (model 2). Robust standard errors will be presented.
Balance of missing outcome data will be presented by treatment status. If balanced by treatment status and if attrition is small (<10%), no adjustments will be made. If missingness is systematic and substantial, we will use bounds (such as Lee Bounds).
Subgroup analysis will include separate linear regressions by male and female children. We will also run quantile regressions [0.1;0.9] in case of anthropometric outcomes.
To address potential issues of spill-overs, we will use the following two-fold strategy:
(i) We will ask treated and control children about their 5 best friends at school in the baseline survey. We will also ask parents about the top 5 parents they interact with. In a robustness check we will then adjust for connections between control and treatment individuals or exclude heavily connected subjects.
(ii) We will also have a “pure control” school in which no kids receive the treatment package. This will allow us to conduct additional robustness checks and sensitivity analyses regarding the importance of treatment spillover effects. We will compare outcomes between control and pure control children and gauge (qualitatively) the sensitivity of baseline estimates to excluding and including pure control children.
Multiple hypothesis testing:
We will calculate p-values and false discovery rate corrected q-values (Benjamini & Hochberg,1995) by family of indicators.
Average standardized effects:
We will give average effects on standardized indicators (Kling, Liebman, and Katz 2007) by family of indicators.