ParkIndex dataset used for analyses.

This study compared marginal and conditional modeling approaches for identifying individual, park and neighborhood park use predictors. Data were derived from the ParkIndex study, which occurred in 128 block groups in Brooklyn (New York), Seattle (Washington), Raleigh (North Carolina), and Greenville (South Carolina). Survey respondents (n = 320) indicated parks within one half-mile of their block group used within the past month. Parks (n = 263) were audited using the Community Park Audit Tool. Measures were collected at the individual (park visitation, physical activity, sociodemographic characteristics), park (distance, quality, size), and block group (park count, population density, age structure, racial composition, walkability) levels. Generalized linear mixed models and generalized estimating equations were used. Ten-fold cross validation compared predictive performance of models. Conditional and marginal models identified common park use predictors: participant race, participant education, distance to parks, park quality, and population >65yrs. Additionally, the conditional mode identified park size as a park use predictor. The conditional model exhibited superior predictive value compared to the marginal model, and they exhibited similar generalizability. Future research should consider conditional and marginal approaches for analyzing health behavior data and employ cross-validation techniques to identify instances where marginal models display superior or comparable performance.