Comparison of Fruit and Vegetable Intake Among Urban Low-Income US Adults Receiving a Produce Voucher in 2 Cities

Key Points Question Are produce food vouchers associated with different changes in diet among different populations, and what characteristics are associated with potential differences? Findings In this pre-post cohort study of 671 adult users of community-based produce voucher programs in 2 cities, use of vouchers was associated with less of an improvement in fruit and vegetable intake and a lower composite nutrition index in San Francisco than in Los Angeles. In a statistical transportability analysis conducted to assess reasons for the difference, lower income of the Los Angeles population appeared to be a primary factor in the difference. Meaning In this study, use of statistical transportability methods helped to identify that produce vouchers were associated with the greatest dietary improvements in the lowest income populations, concordant with principles of flat-rate rather than income-scaled benefit programs.


eAppendix Transportability methods.
Transportability methods seek to understand how and why estimates of the average treatment effect of an intervention may vary across study populations, given that effect modifiers (effect-modifying covariates) may differ among study populations.
In a regression analysis, the value that is estimated is the conditional average treatment effect (conditioned on the covariates), whereas in randomized trials and in causal inference, we estimate a marginal average treatment effect (i.e., we marginalize out the other covariates). The weighting estimator, in a sense, integrates the conditional average treatment effect over the empirical distribution of the covariates to get an estimate of the marginal average treatment effect. In this particular case, we integrate over the empirical distribution observed in the target population (Los Angeles).
In this study, we use the transportability method proposed by Josey and colleagues, 14 who use an entropy balancing approach that seeks to balance entire distributions of covariates (potential effect modifiers) across populations in a study, weighting one population to be more similar to another, to understand how much different covariates may explain differences in estimated average treatment effect. Entropy balancing is a preferred approach to transportability estimation because it has been shown to be doubly-robust (enabling either the weightestimating equation or the effect-estimating equation to be misspecified without introducing bias) and focuses on an entire distribution of covariates among participants rather than simply the mean values of a group.
The entropy balancing approach to transportability estimation is an extension of the methods of moments approach proposed by Signorovitch and colleagues, 21 where given xi ∈ x of measured covariates, the balance function used to model the moments for Si , Yi, and Zi is defined as ̃ = (2 − 1, )and the target covariate distribution ̂0 = [ | = 0] with ̃0 = (0,̂0). The method of moments estimator is defined by the Lagrangian dual problem: which is used to estimate the sampling weights: In the entropy balancing approach proposed by Josey and colleagues, 14 the Langragian dual problem is instead defined as: From which the sampling weights are computed as: The three main assumptions to consider when applying the transportability methods are (i) mean exchangeability, which means that the mean of potential outcomes are exchangeable among the populations, conditional on their covariates; (ii) sampling positivity, which means that the probability of participating in the study is given the covariates is not near zero or one; and (ii) strongly ignorable treatment assignment, which means that all participants could have the same potential outcomes regardless of their current treatment status. The statistical code for this computation is provided online with the overall code to reproduce results of this study, at https://github.com/sanjaybasu/vouchertransportability/.

Additional analyses related to sugar-sweetened beverage consumption.
A penny-per-ounce sugar-sweetened beverage tax was implemented in San Francisco in January 2018, while no such tax existed in Los Angeles. 22 This means that participants in the SF1 group and Los Angeles group were not subject to a tax, but those in SF2 were. As such, it is possible that in San Francisco, the sugary beverage tax might alter the effects of a fruits and vegetable voucher with respect to dietary quality, an effect that would also extend to the eTable 1. Descriptive statistics of the study participants, with the San Francisco population further subdivided between earlier versus later enrollment groups.