Their mechanism was neat. Shoppers' purchases at a supermarket were tracked via loyalty card; participating households also received a debit card. Depending on the treatment group, they either got a 10% rebate on all food purchases or a 10% wedge between healthy and unhealthy foods that was either framed as a tax on unhealthy foods, a subsidy for healthy foods, or as both: in all of those treatments, the real effect was a 15% discount on healthy foods and a 5% discount on unhealthy foods. Groups were evaluated relative to their baseline readings before treatment began, during which they all received a 10% rebate on all purchases; they then can run difference-in-difference to get treatment effects.
What did they find? No significant effect on actual purchases, but a lot of perceived effects.
In the survey conducted after the treatment concluded, subjects were also asked whether or not participating in the study influenced their shopping. The unconditional means by group are reported in Table 12. Those in the treatment groups (all pooled) expressed greater agreement with the statements that they were buying more starred (nutritious) foods, more healthier foods, and a higher percentage of healthier foods, but the difference between the treatment and control groups is not statistically significant in any of those cases.
There are significant differences in the mean response to these questions by frame. Specifically, those in the tax/subsidy frame tend to express greater agreement that the study led them to buy more nutritious foods, buy healthier foods, and buy a higher percentage of healthier foods, relative to those in the subsidy frame. Notably, we did not see such a difference in our data in the actual expenditures and quantities purchased.Some of the lack of effect is due to low power in a sample of about 200.
But they did run a neat permutation analysis afterwards checking to see whether the effects they did find across the groups made sense: they re-ran everything over a thousand iterations in which each participant was randomly assigned to having been in a different treatment than the participant actually was in. And they there found that 70% or more of the random-draw assortment showed larger effects than the ones found in the main estimates that looked at the effects of the group the person was actually in.
What do the point estimates say in a low(ish) power study? Treatment groups spent $1.11 more on nutritious foods and $1.55 less on less nutritious foods per week. If you look solely at lower income households, they spent much more on both nutritious ($7.03 per week) and less nutritious ($7.11) foods: the income effect of the subsidy mattered. Households in the treatment group increased the share of expenditures on nutritious foods by about a percent. That effect was smaller among poorer households because they spent a lot more on both categories.
The biggest effect they find is when they split things out by treatment frame and by income category: low-income households told that the debit card provided them a subsidy for nutritious food substantially increased purchases of less nutritious food ($21.23 per week), with no statistically significant (but still a large point estimate) increase in purchases of nutritious food of $11.58 per week. We should be careful on this one though as low powered tests will only find large effects to be significant.
Taxes on energy-dense foods are arguably the most commonly-advocated anti-obesity policy. The results of this paper have several implications for such policies to promote more nutritious diets. First, taxes may need to be large to change behavior. In the U.S., taxes on soda pop and snacks average one to four percent (Chriqui et al., 2014), but we find no significant impact on expenditures or purchases from a 10 percent relative price change. Second, price changes may have different impacts by income; we find that subsidies for nutritious food may lead low-income households to buy more of all food, including more of the less nutritious food that the policy is attempting to discourage....while giving the usual caveats about drawing inferences from studies that might not have sufficient power.