Food accessibility at the walkable distance scale: a spatial statistical analysis using the Bayesian two-part log-normal model

Keywords: Food accessibility; Bayesian analysis; multi-level spatial modeling; two-part model; zero-inflation
Abstract Type: Paper Abstract

Abstract

In the past decade, the topic of food accessibility has gained tremendous attention from scholars in multiple disciplines. Various measures have been proposed to quantify geographical food accessibility across spatial scales based on assumptions of different transport modes. When food accessibility is measured at the walkable distance scale (e.g., 1200m) using continuous metrics such as the enhanced two-step floating catchment (E2SFCA) area approach, the values are highly skewed with many zeros. This issue poses challenges for subsequent statistical analysis in examining the association between food accessibility and potential risk/protective factors including socioeconomic indicators. If not properly addressed, it could result in biased statistical inferences.
This study proposes a Bayesian spatial multi-level two-part log-normal model to analyze food accessibility scores (i.e., on-site restaurant accessibility measured with E2SFCA at the community level) in Nanjing, China. Zeros and non-zero values are separately modeled. Covariates at both the community- (residential population density, % of male population, % of senior population, and urbanicity) and street block levels (% of rural immigrants and % of low-income population) are included in the model. The model also includes two sets of spatially structured and unstructured random effects at the community level and street block level, respectively, accounting for the remaining variations not explained by the covariates. Posterior predictive checking indicates that the proposed model, which explicitly accounts for excess zeros in the semicontinuous food accessibility scores as well as spatial autocorrelation at two spatial scales, fits the data reasonably well.