Abstract
There is growing interest in the situations in which incentives have a significant effect on positive behaviors, particularly in children. Using a randomized field experiment, we find that incentives increase the fraction of children eating a serving of fruits or vegetables during lunch by 80 percent and reduce the amount of waste by 33 percent. At schools with a larger fraction of low-income children, the increase in the fraction of children who eat a serving of fruits or vegetables is even larger, indicating that incentives successfully target the children who are likely to benefit the most from the increased consumption.
I. Introduction
Schools provide a large fraction of the meals that children eat. This fraction has increased with expansions in school breakfast programs, summer meal programs, and the addition of dinner programs in some districts. These programs provide a unique opportunity to encourage fruit and vegetable consumption in children, particularly among children from low-income families who consume fewer fruits and vegetables at home (Krebs-Smith et al. 1996; Muñoz et al. 1997). However, the potential benefits of these opportunities will only be realized if these children actually eat the fruits and vegetables being offered.
In this paper, we examine the degree to which the use of rewards can increase the fraction of children who eat fruits and vegetables as part of their school lunch. There is a growing body of research that examines the impacts of financial incentives on various positive behaviors in school-age children (Angrist and Lavy 2009; Bettinger 2012; Kremer, Miguel, and Thornton 2009; Fryer 2011). Research on the use of financial incentives to encourage healthy behaviors, however, has mostly focused on adults (Cuffe et al. 2011), with incentivized health behaviors including exercise, weight loss, smoking cessation, health screenings, and immunizations (Charness and Gneezy 2009; Cawley and Price 2011; Volpp et al. 2008; Volpp et al. 2006; Malotte, Rhodes, and Mais 1998; Moran et al. 1996).1
Research on other interventions designed to encourage fruit and vegetable consumption indicates that children may be particularly sensitive to even small changes when making food decisions. In one intervention, the fraction of children eating fruit increased by simply having the cafeteria workers provide a verbal prompt (Perry et al. 2004; Schwartz 2007). Other programs that include educational materials, food service changes, and parental involvement show modest but consistent improvements in fruit and vegetable consumption, though the effects of these programs are sometimes limited only to fruit or fruit juice consumption (Gortmaker et al. 1999; Perry et al. 1998; Cassady et al. 2006). We find that providing a reward increases the fraction of children eating a serving of fruits or vegetables by 28 percentage points (an 80 percent increase), and this increase occurs even on those days in which only vegetables are being offered.
We test the degree to which the response to incentives depends on the nature of the reward being offered. Past research suggests that physical prizes and cash rewards may have different effects, and that the timing of the reward matters as well (Kivetz and Simonson 2002; Laibson 1997; Bettinger and Slonim 2007). We randomly assign each school to one of five different incentives that vary in size (quarter or nickel), type (cash or prize), and timing (now or in a few weeks). Among the cash prizes we find that children discount the future (as expected) and are more responsive to larger prizes (with an implied elasticity of 0.57).
We also test whether the introduction of rewards produces a larger effect at schools with a larger fraction of children who receive free and reduced-price lunch (a proxy for the income level of the school). The schools in our sample come from a disparate range of income levels, with the fraction of children receiving a free lunch ranging from 17 percent to 77 percent. We find that the effect of the incentive program at the lowest-income schools was more than twice as large as the effect at the highest-income schools. Because consumption rates of fruits and vegetables are much lower among children from low-income households (Krebs-Smith et al. 1996; Muñoz et al. 1997), these children are likely to benefit the most from the additional fruits and vegetables. This suggests that this type of rewards program successfully targets the children who will benefit the most from the increased consumption of fruits and vegetables.
Finally, we examine some of the potential negative effects of providing incentives. Because we only provided an incentive to children for eating at least one serving of fruits and vegetables (with no additional rewards for eating two or more), we document that our incentives did not cause individuals who were already above the threshold to reduce their behavior back to the threshold. We also show that our incentives did not produce a negative rebound effect. Others have found that removing an incentive can lead to behavior returning to levels that are lower than the preincentive baseline rate (Lepper, Greene, and Nisbett 1973). We find no evidence of such an effect in the context of fruit and vegetable consumption.
II. Experimental Design
We conducted a field experiment at 15 elementary schools in two school districts in Utah. We randomly assigned each of the participating schools to one of six groups: (1) receive a lottery ticket for a prize immediately, (2) receive a quarter immediately, (3) receive a lottery ticket for a prize in two weeks, (4) receive a quarter in two weeks, (5) receive a nickel immediately, (6) no incentive (control). Our original randomization only included the first four groups and the control group (with three schools in each group). After preliminary results from a school in which students were rewarded with a quarter immediately, we added the nickel treatment to provide insight about whether children understand the value of money and, if so, how responsive they are to increases in the incentive.2 The final assignment includes three schools to each of the first three treatments, two schools to each of Treatments 4 and 5, and two schools as controls.
We decided to use a lottery for the prizes rather than awarding individual prizes because it was difficult to find a set of prizes that cost a quarter that would appeal to a broad set of ages and that we could easily hand out during lunch. By using a lottery we were able to award prizes related to some form of active recreation such as rip-sticks, tennis rackets, soccer balls, and swim goggles. The total cost of these prizes was equal to the number of kids who ate a serving of fruits or vegetables multiplied by 25 cents (so that the prizes would have the same expected cost as the quarter treatments). Since the prizes varied in value, we allowed the children to pick prizes in descending order based on the order in which their raffle tickets were drawn.
The treatment days occurred over five lunch periods spanning two to three weeks. On each treatment day there was a message in the morning announcements, read over the school’s public address system, about the reward students could receive by eating a serving of fruits or vegetables that day. For those schools where we provided prizes instead of cash, we displayed prizes on a table near where we were collecting data, visible to all students. For schools where students would receive prizes that same day we displayed the number of prizes we anticipated giving out that day. For the schools where students would receive the prizes in two weeks, we displayed the full set of prizes that we anticipated giving out at the end of the treatment (which included five days’ worth of prizes).
The children were informed that they would receive a reward if they ate at least one serving of fruits or vegetables as part of their school-provided lunch (with no additional rewards for eating more than one serving). Each of the schools in our study provide a main entrée and students are allowed to choose as many items as they want from a selection of fruits, vegetables, and other side dishes. Our experiment occurred prior to the current guidelines that require that every child have at least one serving of fruits or vegetables on their tray. However, during our experiment, many of the schools had a similar requirement in place.
Prior to the start of any treatments, we collected baseline data at each school for five days spread out over a two-week period. Our research assistants stood by the trash cans in each cafeteria and recorded the number and type of fruits and vegetables taken and the number and type consumed by each child (potatoes, corn, and fruit juices were not included as fruits or vegetables). During the treatment days, data was collected in the same way as during the baseline period with an additional research assistant present to distribute the prizes. The fruits and vegetables came in special cups or left behind a peel or core, allowing us to record the number of servings taken and consumed based on the number of full or empty servings left on the tray, reported in half serving increments. The gender and grade of each child were also noted as the child dispensed of their tray.3 The same data collection procedure was used during both the baseline and treatment periods as well as in both the control and treatment schools.
This data approach allowed us to quickly record information for a large number of children during a relatively short period of time without having any verbal interaction with the children. One disadvantage of this approach is that it does not allow us to record any identifying information about the child (aside from their grade and gender). Our approach also does not allow us to directly measure health outcomes. However, increasing consumption of fruits and vegetables should indicate that children are closer to meeting recommended dietary guidelines.
We combine our consumption data with descriptive statistics from the Common Core of Data. This dataset provides sociodemographic and performance information on all schools in the United States, allowing us to determine whether schools we included in our sample are similar to national averages. The schools in our study are very similar to the national average for the fraction of students who are eligible for a free lunch (42 percent compared to the national average of 39 percent), the pupil / teacher ratio (16.35 vs. 16.17), and the fraction of students who are Hispanic (23 percent vs. 24 percent) or Asian (4 percent vs. 6 percent). One characteristic that differs greatly for these schools is the fraction of students who are African-American (one percent vs. 13 percent). In addition, the schools in our sample are entirely from suburban neighborhoods and may not reflect the types of patterns that might occur at urban elementary schools.
III. Impact of Incentives
In Figure 1, we provide pre- and posttreatment information about the fruit and vegetable consumption patterns for each of our treatment groups. During our baseline data collection period, 33.2 percent ate at least one serving of fruits and this baseline rate was similar across all of our treatment groups. Among schools that provided an immediate reward, the change in the fraction of children eating at least one serving of fruits or vegetables was lowest when the reward was a nickel (a 15.4 percentage point increase) and highest when the reward was a quarter (a 38.5 percentage point increase).
Effect of incentives on the fraction of children eating at least one serving of fruits or vegetables.
Notes: The control group was split into a baseline and incentive period based on the first five days and second five days of data collection, even though nothing changed between the two periods. The sample sizes for group are provided in parentheses.
The regression results that follow test more precisely the overall effect of providing incentives and how this effect differs based on the nature and timing of the reward. When interpreting these regressions, it is important to note that the effect of providing incentives is the combination of responding to the reward and possibly responding to having to interact with an adult about their eating habits. Since the effect of interacting with an adult should be the same for the different treatment groups, differences in the rewards will drive the variation between the different treatment groups.
All of the regressions include school fixed effects, day of the week fixed effects, controls for the child’s grade and gender, and an indicator for whether only vegetables were offered that day (instead of vegetables and fruit). Since the treatment was randomized at the school level, we cluster all of the standard errors at the school level. In each regression, we examine the effect of the incentives on the fraction of kids that ate a serving of fruits or vegetables, the number of servings eaten, and the number of servings thrown away. It should be noted that the number of items that are placed on the student’s tray, consumed, and thrown away are jointly determined (in fact the three form an accounting relationship).
In the first panel of Table 1, we combine the different incentive types into a single indicator variable that provides the average effect of providing any incentive. We find that providing an incentive raised the fraction of children eating at least one serving of fruits and vegetables by 27.7 percentage points, which represents an 82 percent increase relative to the consumption rate in the baseline period of 33.6 percent.4 The number of servings being thrown away was also reduced by 0.11 servings per child (a 33 percent decrease). We have employed a linear probability model given the ease with which marginal effects may be reported and interpreted. Results do not change substantively when estimated using the appropriate discrete, count, or proportional methods.5
Effect of incentives on children’s fruit and vegetable consumption.
The second panel of Table 1 displays the results by the type of incentive, providing three basic insights into designing a rewards program for children. First, larger prizes produce a larger response (with an implied elasticity of 0.57).6 Second, immediate rewards produce a larger effect than delayed rewards. Third, cash prizes tend to produce a larger effect than a cash-equivalent lottery for prizes (though not statistically significant). The p-values for each of these differences using “ate a serving” as the outcome are 0.001 (quarter now vs. nickel now), 0.065 (quarter now vs. quarter later), and 0.191 (quarter now vs. prize now).7 Since we only had two to three schools in each of the treatment arms we mention these differences by incentive type only as being suggestive of how schools might structure the parameters of a rewards program.
While, from a statistical perspective, it is generally ideal to assign the treatment at the individual level, there are many situations in which randomization should occur at the school level. This is particularly true in cases where there are likely to be contamination effects, a potential resentment effect on the part of the control group, or a need to advertise the intervention through school-wide announcements or posters. The primary challenge of school-level randomization is that the cost of providing a treatment to an entire school can limit the number of schools that can be included in a single study.
Since our experiment was randomized at the school level we need to correct for general autocorrelations among the errors of students attending the same school (Bertrand, Duflo, and Mullainathan 2004). In the results presented in Table 1, we cluster all of the standard errors at the school-level. Since there are only 15 schools in our sample, the asymptotic properties of clustering may not be appropriate for our setting. As a result, we also provide two alternative tests of whether the incentives produced a statistically significant change in consumption behavior.
First, we collapse all of our data so that there are just two observations for each treatment school (one for the baseline and one for the treatment period) and one for each control school. We reestimate the results from panel A of Table 1, but now the only controls are a set of school fixed effects. Using this collapsed data approach with only 28 observations, we find very similar results and all of our main treatment effects continue to be statistically significant at the 1 percent level (we provide these results in Table A1 in the appendix).
Second, we test for a statistically significant effect of providing incentives using a Fisher Exact test. In Figure A1 (in the appendix) we plot the percentage change in the fraction of children who ate a serving of fruits and vegetables at each of the schools in our sample. Under the null hypothesis that there was no treatment effect we would expect about half of the schools to experience an increase in consumption rates and half the schools to experience a decrease in consumption rates. In fact, at every treated school we observe an increase in consumption rates. In contrast, both of the control schools had a lower fraction of children eating a serving of fruits and vegetables during the last five days of the observation period than they did during the first five days (one school had a decrease of one percentage point and the other school a decrease of six percentage points). The p-value for the Fisher Exact Test of whether the program had a positive effect is 0.0095.
IV. Additional Effects
One potential concern with our intervention design concerns the establishment of social norms. In our experiment, we only provided a reward to children for the first serving of fruits and vegetables that they ate each day. This aspect of our rewards program might have communicated a social norm that could cause those individuals already consuming multiple servings to decrease their consumption to just one serving. However, the data suggest that such is not the case. Figure 2 displays the distribution of the number of servings of fruits and vegetables that children consumed. There was a big drop in the fraction consuming zero and a large increase in the fraction consuming one serving. At the higher levels of consumption, the fraction of children eating two servings increases from 3.8 percent to 6.6 percent and the fraction of children eating three or more drops from 0.8 percent to 0.5 percent.
Changes in the number of servings of fruits and vegetables consumed.
Notes: This figure shows the fraction of children who ate that number of servings of fruits and vegetables during the baseline period and during the incentive period.
In addition to the direct cost of the rewards, the incentive program has an additional cost to schools in terms of the additional items of fruits and vegetables that they would have to serve. We find an increase in the number of items being served by the schools (about 0.17 servings per child) and a decrease in the amount of fruits and vegetables being thrown away (about 0.11 serving per child). Using administrative records from the school district, we find that the average cost of a serving of fruits and vegetables is about 20 cents. Thus the incentive program created an additional cost to the schools of about three cents per child because of the extra items that were served. This additional cost may be of less importance under the new School Lunch guidelines, which require each school lunch to contain a serving of fruit or vegetables to qualify for reimbursement.
When considering costs, it is helpful to compare the cost of the incentive program with the overall cost of lunch and the cost of other programs with a similar goal. The incentive program would cost an average of 24 cents per child per day, including 18 cents expended on incentives and six cents expended on the labor costs of administering the incentives. This constitutes roughly 8 percent of the cost of a school meal in this context.8 The federal fruit and vegetable program costs between $50 and $75 per year per student, providing each child with a serving of fruits or vegetables as a snack each day. By comparison, the incentive program we implemented would result in $42 of additional cost per year per student, and operates largely by reducing the amount of fruits and vegetables wasted.
Another concern is that the incentives may have affected the underlying sample of participants by encouraging students to purchase a school lunch. We find that the incentives increased the number of children consuming a school lunch each day by 11 lunches on average (a 4 percent increase) but increased the average number of children eating at least one serving each day by 80 (an 87 percent increase). If we made the extreme assumption that all of the children who switched to school lunches on the incentive days were already eating fruits and vegetables as part of their sack lunch, then we would need to scale our effect sizes by one eighth.9
We also examine whether the incentives had a larger impact on the schools with a larger fraction of low-income children (as measured by the fraction of children receiving a free or reduced-price lunch). The schools in our sample range from 17 percent of the students receiving a free lunch up to 77 percent. For the purposes of examining the interaction between incentives and socioeconomic status, we recentered the measure of those receiving free and reduced-price lunch by subtracting 0.17 (the lowest percentage observed) from each data point. Thus, the interaction can be interpreted as the marginal impact of increasing the percentage of students receiving free or reduced price lunch, and the coefficient on incentive can be interpreted as the impact at the schools in our sample with the lowest level of free and reduced-price lunch participation.
The coefficient on the incentive variable in panel C in Table 1 indicates that at the richest schools the fraction of children eating a serving of fruits and vegetables increased by 18 percentage points (33 percent smaller than the overall effect across all schools). The coefficient on the interaction term between the incentive variable and the free lunch rate indicates that an increase from a 0 to 100 percent free lunch rate would increase the effect of the incentive by 32 percentage points. After scaling this differential effect by the observed range of variation in the free lunch rate in our data (0.6), we find that the effect of the incentive on consumption rates at the lowest-income schools was more than twice as large as the effect at the highest-income schools.10
Part of the differential response could be due to the fact that poorer schools had slightly lower rates of fruit and vegetable consumption during the baseline period. When we fit a linear regression between the school’s free lunch rate and fruit and vegetable consumption during the baseline period, we find that a 10 percent increase in the fraction of children receiving a free lunch is associated with a one percentage point decrease in the fraction of children eating a serving of fruits or vegetables. Another potential reason for the differential response is that lower-income children may place a higher value on receiving the incentive.
There might be some concerns about an experimental design with only two control schools. The primary need for control schools in this setting is to measure the degree to which the treatment schools would have changed during the post period in the absence of the treatment. We address concerns about a small number of control schools in two ways. First, we varied the start date for the intervention from school to school, with start dates ranging over the entire course of the school year. Thus, there is less concern in this setting about the results being driven by time-varying factors such as seasonal variation or the timing of other programs that would have affected the schools in our sample. Second, we provide some evidence about the degree to which there was a Hawthorne Effect in which children alter their behavior in response to our research assistants being present in the cafeterias.
In Appendix Table A2, we provide results from a regression in which we examine the change in eating behaviors over the course of the five days of baseline observation that we conducted at all 15 schools. The results indicate that there was almost no change in eating behaviors during those first five days of observation and, if anything, there was a slight decrease in the fraction of children eating a serving of fruits or vegetables during lunch. We also estimate a similar time trend for the full set of days at our two control schools. Again, we find that there was no change in eating behaviors over the course of the observation period.11
Another concern with our data collection is that, as with any incentive program, there may be some fruit- and vegetable-averse children who will attempt to find ways to cheat. Possible opportunities to cheat include outsourcing (getting friends to eat fruits or vegetables), smuggling (hiding uneaten portions in milk boxes or under other uneaten food items), and premature disposal (dumping uneaten portions under the table). The difficulty of observing individual children in a full cafeteria during lunch makes monitoring of outsourcing very challenging, but we were able to collect some information about the latter two channels by measuring the amount of waste on the floor after lunch and doing a random check whether students were placing items in milk cartons. We found only minor levels of cheating relative to the overall increase in the consumption of fruits and vegetables.
Finally, there is often a focus in public health and nutrition about whether the effect of an intervention operates differently for fruits and vegetables. Many past interventions that have been successful in raising fruit or fruit juice consumption have had no effect on vegetable consumption (Krebs-Smith et al. 1996, Gortmaker et al. 1999; Perry et al. 1998; Cassady et al. 2006). We did not collect data on which item each child ate but rather simply recorded whether they had eaten at least one serving of fruits or vegetables. However, when we restrict our analysis to just the 20 days in which a vegetable was the only option available, we find that providing incentives increased the fraction of children eating at least one serving of vegetables from 16.4 percent to 45.4 percent (almost a threefold increase) with a p-value for this difference of 0.003.
V. Additional Experiment
Underlying much of the debate about the use of incentives is the question of whether removing the incentives might cause the child to return to a lower consumption level than prevailed prior to the start of the intervention (Lepper, Greene, and Nisbett 1973). This effect, called crowding out, can occur if the individual’s prior intrinsic motivation to complete the task is reduced when they begin to receive the incentive. For example, the incentive may communicate that the task is more onerous than previously thought, or the individual may believe that the incentive is likely to be provided at a future time if the task is not completed (see Gneezy, Meier, and Rey-Biel 2011 for a full discussion).
Crowding out has been found in many contexts. For example, donations decrease below original levels following a period of matching incentives (Meier 2007). However, Gneezy, Meier and Rey-Biel (2011) review many instances in which incentives do not crowdout intrinsic motivation. Notably, many of these examples are incentives given for educational performance or attendance. In these cases, individuals displayed superior academic performance, as compared to control groups, long after the suspension of incentives. This is attributed to the duration and the size of the incentives leading to the development of meaningful habits as the individual begins to value the action intrinsically. Crowding out tends to occur if the program is not in place long enough to create any meaningful amount of habit formation, or if the rewards are too small to be salient in encouraging the behavior in the first place.
Our original field experiment was focused only on the immediate effect of the incentives and we did not collect any consumption data after the removal of the incentives. To address this issue, we ran a second experiment in which we provided an immediate quarter reward at eight schools from the original sample. The second experiment occurred during the following school year using the protocols of the original field experiment (five days of baseline data and five days of incentives). This second experiment also provides an “out-of-sample” test of the main effects of the quarter now treatment documented in our original field experiment. In the second experiment we also collected data for the four weeks after the end of the incentive (with data collected two days each week for those four weeks).
In Table 2, we present the results of this second field experiment. The regressions in the first two panels mirror those presented in Table 1, except that now the only treatment that we examine is the quarter now treatment. For these first two panels we restrict the sample to just those observations that occurred during the baseline and incentive period (to provide the same test as in Table 1). We find that the incentive increased the fraction of children eating at least one serving of fruits and vegetables by 31.4 percentage points (a 95 percent increase). As before, the effects on the fraction of children that ate a serving were larger at the schools with a higher fraction of students who receive a free or reduced price lunch. The effect sizes for the quarter now treatment were smaller than in the first experiment but still represent a dramatic increase in the fraction of children eating fruits and vegetables.12 If we pool together the schools in the quarter now group from the first experiment with all of the schools from the second experiment we find a combined effect of 33.3 percentage points (with a standard error of 0.027) which represents a doubling in the fraction of children eating at least one serving of fruits and vegetables during lunch.
Impact of incentives on behavior after incentives are removed
In the final panel of Table 2, we include the full sample of observations from the second experiment (including the post-incentive observations). We divide the four-week post-incentive observation period into the first two weeks and the next two weeks. We find that the program produced a small effect on consumption during the two weeks after the incentive program (relative to the baseline rates for each school) but that these differences disappear within four weeks of the end of the incentive program. As with other studies in economics, though, we find no evidence of a boomerang or crowdout effect. In this case, the short-term use of incentives did not cause behavior after the removal of the incentives to revert to a level below that of the baseline period.
The rewards program that we implemented was not designed to be in place long enough for a meaningful habit formation process to take place. This analysis of the post-intervention period simply provides evidence that fruit and vegetable consumption was not any lower after the removal of the rewards than prior to the intervention. The ideal intervention would combine the large immediate effects that we document in this paper with an intervention that produces a slower initial effect, but is more likely to produce a lasting change in eating habits. The Food Dudes program is one such intervention that has been successfully implemented in the United Kingdom and combines a rewards program with a set of videos that depict a set of superheroes who enjoy eating fruits and vegetables with short video clips of pop stars who reinforce this message (Lowe et al 2004).
VI. Conclusion
We implemented a randomly assigned rewards program at 15 elementary schools that include over 47,000 student-day observations with information on the number of servings of fruits and vegetables that each child both took and ate. We find that providing rewards can lead to large increases in the fraction of children who eat fruits and vegetables as part of their school lunch. Providing rewards has an even larger effect at schools with a larger fraction of low-income children. We also find a significant increase in fruit and vegetable consumption when providing a reward as small as a nickel. This suggests that some of the concerns about providing rewards that are too small might not apply to young children (Gneezy and Rustichini 2000).
We allay some of the common concerns about the use of incentives. When we examine the distribution of the number of items that each child ate, we find little evidence that setting a benchmark of one serving reduced the consumption of fruits and vegetables for children who were already consuming above the benchmark. When we examine the change in fruit and vegetable consumption after we remove the rewards, we find that the positive change in behavior fades away quickly. However, children do not end up eating fewer fruits and vegetables than they did prior to the introduction of rewards—ruling out the rebound effects that have been a major concern documented in experimental psychology (Lepper, Greene, and Nisbett 1973).
Our incentive program produced only a small increase to the amount of fruits and vegetables that needed to be served but reduced the amount of fruits and vegetable items that were being thrown away by 33 percent. This suggests that schools can increase the cost-effectiveness of money they are already spending on fruits and vegetables by increasing the fraction of those items that are actually consumed. This reduction in waste might prompt schools to spend money on higher-quality fruits and vegetables, potentially producing an even larger increase in the consumption rates of these items.
While our experimental design provides key insights concerning the effect of incentives on children’s eating behavior, our experiment was not designed to measure the overall impact of a long-run incentive program. Our short two-to three-week intervention can only measure immediate effects of the incentive but cannot observe long term effects or habit formation. Even in our second experiment (where data were collected for four weeks after the treatment) the treatment period was too short for habits to form. Future experiments with longer intervention periods could more accurately measure long term effects of incentives on children’s health behaviors.
Another limitation of our study is that our data collection was only designed to observe changes in consumption that occurred at school. Other data collection approaches that include individual food consumption using 24-hour recall and food diaries have proven useful in past experiments and could yield beneficial results (Gortman et al. 1999; Cassady et al. 2006). In addition, although fruit and vegetable consumption is correlated with health, we were unable to measure the actual impact on health. Future experiments that combine long-run experimental designs with measures of changes in actual health measures would provide a fruitful area for future research. In the short term, however, incentives provide an effective means of influencing children not just to take, but to eat their fruits and vegetables.
Appendix
Percentage point change in the fraction of children eating a serving of fruits or vegetables at each school.
Notes: Each bar is a separate school and the schools are grouped based on the treatment they received.
Effect of incentives on children’s fruit and vegetable consumption
Change in behavior during the non-treatment period
Footnotes
David R. Just is associate professor of applied economics and management at Cornell University.
Joseph Price is associate professor of economics at Brigham Young University. The data used in this article can be obtained beginning May 2014 through April 2017 from Joseph Price. Address: Brigham Young University, Department of Economics, 162 FOB Provo, UT 8460. E-mail: joe_price{at}byu.edu.
↵1. Cuffe et al. (2011) is one of the few studies to examine the effects of incentives on health-related behaviors in children. Their results complement the findings of this paper by examining a program that incentivizes students to ride their bikes to school.
↵2. To include the nickel treatment, we reassigned one of the control schools and one of the quarter later schools to this treatment. These schools never knew about their initial treatment assignment.
↵3. In order to not influence the children’s behavior in the baseline observations, our data collectors did not mention fruits or vegetables when asked what they were doing, but rather replied that they were collecting data about school lunches.
↵4. All interactions between incentive and gender or grade were statistically insignificant, though the coefficients suggest that girls and older children are slightly more responsive to rewards, which is consistent with groups that had higher levels of consumption to begin with being more responsive to the rewards.
↵5. To check the robustness of our estimates we estimated a probit regression for the “ate a serving” column and a poisson regression for the other two columns. In each case the average marginal effect of the incentive was about 10 percent smaller in magnitude, though all coefficients are still statistically significant at the 1 percent level.
↵6. Elasticity is the percent change in the consumption rate divided by the percent change in value of the reward (using an arc-elasticity formula).
↵7. The prize now vs. prize later comparison provides an inappropriate test for time discounting because on the days that we provided immediate prizes, we set the items out for that day; whereas, at the schools where we did the delayed prizes, we set out all of the prizes that would be awarded after the experiment (which was five times as large as the prizes for any one day). Also, it is surprising that the quarter now treatment had a larger coefficient than prize now treatment since the prize now treatment had the added effect of providing students during lunch with a salient and visual image of the prizes that could be won.
↵8. The cost of providing the incentive per child is equal to the incentive, 25 cents, multiplied by the percentage of children who receive the incentive, roughly 70 percent. The cost of the school meal ($3.31) is assumed to be the sum of all federal and state support per meal. This includes the $2.57 reimbursed by the Federal government for each lunch provided to students qualifying for free meals, 20 cents in commodities provided by the Federal government per meal, and 54 cents reimbursed by the state government per meal.
↵9. Unfortunately our data collection approach does not work well for children with sack lunches. There is some evidence from past studies that sack lunches provide fewer fruits and vegetables than school lunches and often fail to include even one serving of fruits and vegetables (Rainville 2001; Gordon and McKinney 1995; Sweitzer, Riley, and Robert-Gray 2009), suggesting that the actual degree to which we would need to scale back our estimates is probably much less than one eighth.
↵10. The effect on number of servings discarded was about a third larger at the lowest-income schools, though this effect was not statistically significant. The lack of precision in this estimate is likely due to the fact that we can only proxy the income of the school rather than the individual child.
↵11. We also find similar results if we estimate a regression in which the time trend variable is based on calendar days rather than days in the experiment.
↵12. The overall incentive effect between the two experiments are not statistically different from each other. If we compare the second experiment with just the quarter now effect in the first experiment, the second experiment resulted in an effect that was about ten percentage points smaller increase for the fraction of children eating a serving of fruits and vegetables and 0.2 smaller decrease in the servings wasted (though neither of these differences were statistically significant).
- Received April 2012.
- Accepted October 2012.
