Effects of Modern Food Retailers on Adult and Child Diets and Nutrition

In many developing countries, food environments are changing rapidly, with modern retailers—such as supermarkets—gaining in importance. Previous studies have suggested that the rise of modern retailers contributes to overweight and obesity. Effects of modern retailers on dietary quality have not been analyzed previously due to the unavailability of individual-level dietary data. Here, we address this research gap with data from randomly selected households in Lusaka, Zambia. Anthropometric and food-intake data from 930 adults and 499 children were analyzed to estimate effects of purchasing food in modern retailers on body weight, height, and dietary quality while controlling for income and other confounding factors. The food expenditure share spent in modern retailers was found to be positively associated with overweight in adults, but not in children. For children, a positive association between expenditures in modern retailers and height was identified. Modern retailers contribute to higher consumption of ultra-processed foods and calories. But they also increase protein and micronutrient intakes among adults and children, mainly through higher consumption of meat and dairy. The findings underline that modern retailers can influence diets and nutrition in positive and negative ways. Differentiated regulatory policies are needed to shape food environments for healthy food choices and nutrition.

also be used to control for unobserved heterogeneity in non-linear models, such as probit and Poisson models [3]. The CF was estimated as a two-stage model, whereby the first-stage regression was represented as: where is a vector of instrumental variables, and is a random error term. The other variables were defined as in the main paper. The second-stage model was the regression equation shown in the main paper with the individual diet and nutrition outcomes as dependent variable but including the residuals from the first-stage model as an additional regressor. If the residual term was statistically insignificant in the secondstage regression, the null hypothesis of no endogeneity bias could not be rejected, so the CF approach was not required; in that case, the regular one-stage models led to unbiased and more efficient estimates.
However, if the residual term turned out statistically significant, the CF approach was preferred to control for endogeneity bias.

Instrumental Variables
We used three instrumental variables for the CF models. First, distance from each household to the closest shopping mall, which was calculated using global positioning system (GPS) data collected during the survey. GPS-based data to measure distance to modern retailers were also used as instruments in several other studies [4][5][6][7]. Second, whether or not the household felt that modern retailers sell food of higher quality than traditional retailers, which was also used by Umberger et al. [4]. Third, the number of visits to a shopping mall of the household's closest neighbor in the sample. Beyond distance and accessibility, the neighbor's behavior may capture influence through local social networks [8]. Tests for instrument validity are discussed below.
All three instruments are significantly correlated with the household's own use of modern retailers, (Supplementary Table S21). As expected, distance to the closest shopping mall is negatively correlated, while the other two instruments are positively correlated with the food expenditure share spent in modern retailers. Moreover, the Wald test for the joint significance of the three instruments is statistically significant at the 1% level for both adults and children (Supplementary Table S21). This underlines that the instruments are relevant. A second important criterion for validity is that the instruments do not affect the dietary and nutrition outcome variables directly, other than through own use of modern retailers. We performed a simple falsification test following Di Falco et al. [9]. Results show that the three instruments are jointly insignificant in all models with dietary and nutrition outcomes as dependent variables (Supplementary Table S22). Hence, we conclude that the instruments are valid. Figure S1. Map of Lusaka City with sampled compounds/sections and households Table S1. List of main shopping malls with modern retailers in Lusaka City in 2018 The main shopping malls that were operating in 2018 are included. Very small shopping malls were not included. Likewise, malls that were still under construction in 2018 were not included. The list was compiled by the authors based on internet search, personal visits, and key informant interviews.  Saturated fats ≤8% of energy ≥16% of energy 10 a Ratio of poly-and mono-unsaturated fatty acids (PUFAs and MUFAs) to saturated fatty acids (SFAs). HEI components and weights based on [12][13][14].    Table S8. Effects of using modern retailers on dietary composition in terms of food processing levels (full model results for Table 3) Unprocessed Foods (expenditure share, %)
Modern retail use 1·109*** 1·049 -0·974 1·165*** 1·049 1·014   Coefficient estimates are shown with bootstrapped standard errors clustered at compound level in parentheses. Estimates for modern retail use can be interpreted as marginal effects of a 1 percentage point increase in the modern retail expenditure share. For Figure 3, coefficients and standard errors were multiplied by 10, to show effects of a 10 percentage point increase in the modern retail expenditure share. CF, control function approach. Bemba and Protestant are used as reference group for ethnicity-Chewa and Tonga, and religion status-Catholic, respectively. * p < 0·10, ** p < 0·05, *** p < 0·01.