Equity of access to fast food outlets in Victoria, Australia: a comparison of statistical methods

Abstract

OBJECTIVES: The hypothesis that food outlets are inequitably distributed across neighbourhoods of varying levels of disadvantage has been tested in numerous studies, with findings somewhat inconsistent. The statistical approaches utilised in these studies vary and while there is a need to consider spatial issues in these analyses, such as spatial autocorrelation, many studies do not use appropriate analytical techniques which can lead to incorrect inferences about associations. METHODS: Using data on the location of the four major chain fast food outlets (McDonald’s, Hungry Jacks, KFC, Red Rooster) in Victoria, Australia as an example, we examined the distribution of outlets by neighbourhood-level socioeconomic position (SEP) with and without adjustment for population size. Our aims were to compare commonly used statistical methods (one-way ANOVA, Kruskal-Wallis, Poisson and negative binomial regression) to identify whether the choice of technique affects the results obtained. Furthermore, we aimed to examine the spatial autocorrelation in the distribution of outlets, determining whether accounting for neighbourhood-level predictors accounted for any observed correlation. RESULTS: We identified discrepancies in the results dependent on which method was adopted. After adjustment for population size, all methods found evidence of an association between SEP and the number of fast food outlets at the 5% significance level, apart from the one-way ANOVA which found weak evidence of an association (p=0.07). From pair-wise comparisons, all studies, apart from the one-way ANOVA, found evidence of higher access in the most disadvantaged neighbourhoods compared to the least disadvantaged. Poisson regression identified more pair-wise differences than the other methods employed. We found weak positive spatial autocorrelation in the distribution of major chain fast food outlets in Victoria (ρ=0.06, pseudo p-value=0.03). Adjustment for neighbourhood-level variables resulted in no evidence of remaining residual spatial autocorrelation. CONCLUSIONS: In studies involving neighbourhood-level data, care should be taken to respect the nature of the outcome distribution in the analysis as well as any spatial structure in the data. Research using spatial data should examine and adjust for spatial autocorrelation where necessary. Failure to adopt appropriate methodologies and to account for spatial autocorrelation can lead to incorrect inferences and can cause problems when comparing studies.