A Randomized Controlled Trial to Address Consumer Food Waste with a Technology-aided Tailored Sustainability Intervention

Abstract

There is growing concern whether pro-environmental behavioral interventions can generate sufficient reductions in carbon emissions to address climate change. While many have suggested enhanced tailoring of interventions to increase effect sizes, and while individual tailoring is common among health interventions, little is known about how individual tailoring may impact effect sizes for pro-environmental behavioral interventions. Using a novel technology-aided delivery and measurement approach, we conduct a randomized controlled trial featuring an individually tailored intervention focused on reducing the amount of food wasted by participants over approximately one week in their normal living conditions. We find large significant effects for the focal area of food wasted during dining (a 79% reduction), a null effect on food wasted over all household stages (preparation, dining and clean outs), and desirable or null effects for critical antecedent (e.g., waste during preparation, continued purchases of fresh produce), concurrent (e.g., food selection and consumption), and attendant behaviors (e.g., waste from storage clean outs, avoiding waste deposits in landfills).

Introduction

There is growing concern whether pro-environmental behavioral interventions can generate sufficient reductions in carbon emissions to address climate change.^1-4 Such interventions typically feature a single approach (e.g., information, social comparison, appeals), often chosen before project implementation to advance empirical understanding of hypothesized behavioral theories; offer the intervention uniformly to a large pool of potential participants; and measure a single focal behavior.^4-5 The pro-environmental behavior encouragement literature suggests that, while many of these interventions yield statistically significant effects,^{1, 6-10} the urgency of climate change necessitates that future interventions are designed to maximize potential impacts by targeting applicable populations, focusing on high-impact practices, and ensuring spillovers to other behaviors with environmental implications are beneficial or at least neutral.^{4-5, 10}

This paper introduces the concept of technology-aided tailored sustainability interventions (TATSIs), an approach that leverages technology to individually tailor interventions to achieve pro-environmental behavior change. TATSIs are a nascent subset of Tailored Lifestyle Interventions, which involve trained counselors working with individuals to develop and implement personalized strategies to change behavior. By adjusting messaging and treatment approaches to the needs of individual patients, Tailored Lifestyle Interventions have been shown to be more effective in many settings than non-differentiated interventions.^11-13 Technology-aided tailored interventions have proliferated in the health and wellness space, and have been shown to be efficacious in changing a variety of behaviors including weight loss,¹⁴ physical activity,¹⁵ pharmaceutical adherence,¹⁶ preventative procedure uptake,¹⁷ dietary adherence,¹⁸ smoking,¹⁹ and drug and alcohol intake.^20-21 Recent research suggests that when such interventions leverage technologies such as mobile phones to complement or replace human touchpoints, efficacy can be maintained if the technology-enhanced strategy is informed by pertinent behavioral theories and delivered with appropriate expertise and intensity.²² This has led to the development of Just-in-Time Adaptive Interventions (JITAIs), which provide individuals with the right type and amount of personalized support only when necessary by adapting intervention delivery to the current needs and environment of the participant.²³ JITAIs have moderate to large effect sizes across a number of behaviors and JITAI efficacy is associated with additive efficacy when tailoring occurs with human coaches and an algorithm.²⁴ As more than 95% of the U.S. population have some type of mobile phone,²⁵ including 70% of Medicaid recipients,²⁶ and as staffing and travel costs constitute some of the greatest costs and barriers to implementing tailored lifestyle interventions, integrating such technologies into these interventions holds the potential for cost-effectively scaling programs across a broader population.

Our primary research question is whether a TATSI deployed with participants in their everyday living conditions can reduce household food waste. From a social ecological perspective, food waste, and the ability to alter it, is influenced not only by personal factors, but also factors associated with a person’s group (e.g., household), community, and the physical and social milieu in which person lives.²⁷ We report results from a randomized controlled trial that evaluates a TATSI that relies on the theory of planned behavior and the theory of reasoned actions.²⁸ This is accomplished by facilitating an environment that supports behavior change in the context of the household’s environment, inclusive of factors that support and limit modifying food waste. Behavioral goals are clearly defined and self-efficacy is promoted. Additionally, a behavioral toolbox approach can be used by trained staff to suggest food waste reduction techniques based upon data and interpersonal feedback received from the participant. Toolbox approaches have been utilized in successful behavior change interventions that focus on food intake, namely weight management interventions, and they facilitate internal validity and treatment fidelity.^29-31 In the context of the recent literature on identifying the most impactful approach to inducing sustainable consumer behaviors,³ this approach emphasizes individually matching intervention approaches to participants rather than testing average treatment effects for a particular intervention. The flexibility afforded by the toolbox approach is a critical as research has identified 160 specific food waste drivers that operate through three primary mechanisms (motivation, opportunity and ability) across multiple points within household food routines (purchasing, preparation, serving, consumption, clean up).³²

Reduction of food waste is a part of the United Nation’s Sustainable Development Goals³³ because approximately one-third of all food produced globally is never consumed, and this loss and waste of food creates negative impacts on financial well-being, food security, natural resource stocks, and environmental quality.³⁴ As the final link in the food supply chain, consumers take on special importance, as reductions in waste at the consumer level can reduce the demands on the remainder of the food system and the resource base upon which it rests. In the United States, where this study takes place, consumer food waste reduction interventions are of particular interest since 53% of wasted food in the United States occurs in either residential or foodservice settings.³⁵ However, the ability to formulate implementable recommendations is hampered by a paucity of high-quality evaluations of household food waste reduction interventions.^32,36

While several in-home food waste reduction interventions have been rigorously evaluated, to the best of our knowledge none feature tailored interventions.^37-39 Romani et al.’s³⁷ treatment group received information on the importance of meal planning and a meal planning tool. The result, as assessed via weighed food-waste diaries, was a significant (24%) reduction in waste relative to the control group one week following the intervention compared with a control group. Van der Werf et al.’s³⁸ treatment group received a standard multi-component intervention aimed at building waste prevention literacy. The intervention was associated with a significant (29%) reduction in avoidable food waste compared to a control group as measured by weighing curbside food waste collections. Soma et al.’s³⁹ study included three treatment arms: one sharing passive information (booklet and a refrigerator magnet with food storage tips); another with the passive information materials plus a series of four community workshops with interactive learning; and the third with the passive information and online quizzes featuring points accumulation and prizes. No significant post-campaign differences across the control and treatment groups were identified in the weight of food waste collected at the curb. Thus, out of five previous non-tailored interventions assessed in these three studies, only two yielded statistically significant reductions in waste. Among the larger number of food waste intervention studies that feature less rigorous designs³² (e.g., largely pre-post intervention comparisons without a control group or evaluations that rely on retrospective or intended waste reports), none use tailored treatments.

The activities targeted by consumer food waste reduction interventions plausibly alter behavioral antecedents such as food shopping (amounts purchased and frequency of shopping), meal patterns, food preparation techniques and menu planning;^40-41 concurrent behaviors such portion size selection and dietary intake;⁴² and attendant behaviors such as saving uneaten food, food storage cleanouts and the tendency to divert wasted food from landfill through behaviors such as composting.^43-45 Such antecedent, concurrent and attendant behaviors are of interest to an array of audiences including nutritionists, food retailers, waste managers, and policy makers. In terms of antecedents, the types of food purchased can hold major implications for sustainability.⁴⁶ Concurrent decisions that affect food consumption can affect obesity, which has implications not only for health outcomes⁴⁷ but also for environmental sustainability.⁴⁸ In terms of attendant behaviors, diverting wasted food and food scraps from landfills via composting can alter waste hauling patterns and limit greenhouse gas emissions.⁴⁹ While some literature recognizes the potential for how an intervention targeting one behavior (e.g., beginning composting, implementing water conservation) can affect untargeted behaviors (e.g., energy and water conservation^10,50) measurement of such spillovers, particularly to actions that plausibly affect health, remains rare in the evaluation of pro-environmental behavior change interventions.^1,6-10

To our knowledge no previous work explores tailored food waste reduction interventions or food waste interventions that track behaviors across multiple domains of activity. Rather, evaluations of such interventions have largely focused on a narrow evaluation of the amount of wasted food. This may reflect a lack of systems thinking within the literature evaluating food waste reduction interventions.³² The narrow focus of evaluations is likely further hampered by the methods available to assess household food waste, which include surveys, food waste diaries, or audits of household waste. Audits of collected waste do not permit inference of behavioral antecedents as the audit relies solely upon physical analysis of waste and is typically independent of interaction with household members. Food waste diaries and surveys could be augmented to gather information concerning other relevant household behaviors, but we are unaware of any such examples in the literature. Given that surveys and diaries can require considerable effort on behalf of study participants,⁵¹ researchers may eschew expansion of these instruments to reduce participant burden and enhance the quality of the food waste data.

This paper adds to the limited stock of studies that evaluate in-home food waste reduction interventions. The study contributes to the literature on food waste in two distinct ways. First, the intervention implemented in this study relies on the theory of planned behavior and the theory of reasoned actions²⁸ while considering the social ecological context of the household and leveraging a behavioral toolbox approach.^27,52 Further, the intervention features a tailored rather than a standardized intervention that leverages collaboration with a coach, rapid feedback, and goal setting techniques. This approach aligns with observations within the food waste literature that argue that, in order to maximize effectiveness, prevention measures must be individualized to the practices of a particular household and focus on decisive moments within daily routines.⁵² Second, the evaluation method captures highly granular data on both food waste during dining and key antecedent, concurrent and attendant behaviors, including food shopping, food preparation waste, dietary intake, saving of uneaten food for later consumption, food storage waste, and landfill diversion activity using the FoodImage™ smartphone app, a technology developed by our team that has been validated as providing waste measurements equivalent to directly weighed waste measures during food preparation and eating occasions.⁵³ This methodological advance not only permits the evaluation of interventions in multiple spheres (e.g., food waste, nutrition, food acquisition) for multiple food related behaviors (shopping, preparation, consumption and waste) but also enables rapid feedback necessary to support just-in-time, tailored interventions.

Methods and Materials

Forty-four adults (Age 18 – 63, mean = 36.3; 40 women; 65.1% Caucasian; mean BMI = 28.3 with no significant differences between treatment and control for any demographic characteristics) from the Baton Rouge, Louisiana area were recruited. Data collection occurred between July and September of 2018, with 19 and 23 participants randomly assigned to the treatment and control groups, respectively (two participants dropped out during baseline assessments prior to being randomized to treatment). A total of 40 (18 treatment, 22 control) participants completed the protocol and provided sufficient data for analysis (two participants, one from each group, had no baseline data and were not included in the analysis). Sample size was determined by an a priori power analysis conducted to secure study funding (see supporting information) with the resulting number of participants per group similar to the median of 24.5 reported in a recent meta-analysis of similar interventions in the health and wellness space.⁵⁴ Inclusion criteria included: age 18-65 years, body mass index 18.5 – 50 kg/m² (based on self-reported height and weight), conducted some of the household’s food shopping and preparation, possessed an iPhone and operable Apple ID/password, and affirmed willingness to complete all study procedures. Exclusion criteria included: persons who were severely immunocompromised or pregnant. Recruitment occurred via electronic mailing list (listserv), social media posts and ads on Pennington Biomedical Research Center’s (PBRC’s) recruitment webpage. No data is available concerning the number of eligible individuals that chose not to participate as recruitment materials included open invitations.

Relevant internal institutional review boards approved the research design. The study was registered with the National Clinical Trials registry (NCT03309306) prior to study commencement. Respondents provided written informed consent before enrollment. Data collection followed guidelines for ethical treatment and good clinical practice. Participants received $175 for the study.

All participants first attended an individual session at PBRC where research staff verified proper installation of the FoodImage app and provided training until mastery was demonstrated in using the app to capture and send data. Specifically participants were shown how to use the app to take photos of receipts, food and waste items, which included instructions to place a standardized visual reference card in each photo involving food (see supporting information including figures S1-S4 in Roe et al. 2020).⁵³ Participants were told to verify photo quality, to add informational tags to photos concerning the source, quantity, destination, reason for waste, and any details concerning reason for waste, and to ensure data was transmitted.

Upon leaving the training session all participants recorded up to 7 continuous days of data, which included photos of food shopping receipts, food preparation waste, food consumption (photo of plates prior to and upon completion of each eating occasion both at home and while dining out with consumption being the difference), and purges of uneaten food from storage. For participants in multi-person households, data concerning food shopping, preparation and storage purges may capture household-wide activity, while eating occasion data (food selection, consumption and plate waste) are always data from only the participant. Within the app, additional information entered by the participant (meal occasion, waste disposition) was appended to the photo and securely sent to PBRC servers automatically by the app upon completion of the participant’s interaction with the app, so that the staff could monitor if participants were regularly sending data and follow up should participants stop collecting data. Furthermore, staff could access data to qualitatively assess participants’ waste patterns and trends to inform individual support sessions for the treatment group. After 7 days, half of the participants were randomly assigned to the food waste reduction intervention (independent of behavior during baseline monitoring) with the remainder assigned to the intensity-matched (number of topics, contacts and materials) control intervention (stress management).

The treatment group met individually with a trained coach who presented a structured set of materials that included an introduction to food waste and ways that food waste could be reduced over time. The coach and participant then engaged in a lifestyle interview in which the coach asked multiple open-ended questions to gather information about the participant’s typical eating, shopping, preparation, storage, and meal habits, including leftover and disposal tendencies. Questions covered the number of people in the household, number of shopping trips, meal planning (or lack thereof), etc. After basic information on food waste management was introduced and the coach acquired an in depth understanding of current habits, the intervention was tailored to target actions with high marginal impact with regard to current food waste. The coach and participant reviewed instructional information including the Survival TIPPS (Thriftiness, Inventory, Plan, Prepare, and Store). The coach and participant used specific goal setting techniques to make a plan that reduces food waste in the immediate future (i.e., daily in the next week). The coach and participant collaborated to find a technique that the participant found logistically and financially feasible and efficacious (high self-efficacy) with coach support. Specific goal setting techniques such as SMART (Specific, Measurable, Attainable, Reward, Timebound) goals were tailored to ensure that the participant had a concrete idea of what to do to enact the plan in their daily life. The participant then received regular (e.g., three to four per week) semi-structured tips via text, email or call (based on participant preference) and follow up interactions with the coach who assessed participant comprehension of the materials and the efficacy of the chosen strategy. Strategies were modified or added based on gathered data and delivered to assist the participant in following through with plans to reduce food waste with the coach providing feedback and encouragement to promote self-efficacy and motivate change. See the supplemental information for an appendix with more details and examples of coaching interactions.

The control group received a similarly tailored intensity-matched intervention focused on stress management.

Data from the FoodImage app were prepared for analysis using methods detailed in Roe et al. (2020).⁵³ Images captured through the app were viewed by PBRC’s nutrition staff who were blinded from a participant’s group (treatment vs. control) and did not interact with participants. Trained staff members identified nutrient matches (Standard Reference 28 and the Food and Nutrient Database for Dietary Studies, 6.0)^55-56 and estimated the mass of portions of food prepared, selected, consumed, returned, and discarded at each stage based upon these images using methods as previously validated.⁵³ The food waste data includes the sum of unavoidable (e.g., chicken bones), possibly avoidable (e.g., chicken skin) and avoidable (e.g., chicken meat) waste. Served food that was uneaten but saved for later consumption (leftovers) was tracked separately and not counted as waste until, if not consumed, it was discarded. Food selection, consumption and waste amounts are expressed in both weight (grams) and energy (Calorie) units. Participants denoted the destination for each discarded food item (garbage, compost, pet, etc.). Staff also documented the number of food shopping trips and the number of food items purchased from different categories of foods from images of store receipts with, e.g., each line of a receipt corresponding to a unique item of produce. Because receipts do not consistently identify the weight and exact identity of different foods, mass and energy data for food purchases is not available. For food purchase, mass and energy figures, measures were taken repeatedly for each participant and summed across the relevant period (pre- or post-treatment). The validity and accuracy of using FoodImage to measure consumer food waste in this manner has been documented previously.⁵³ These resulting figures were then used in the data analysis.

Data analysis was conducted in SAS (version 9.4) and Stata (version 14). The efficacy of the intervention is assessed via a difference in differences analysis, i.e., by comparing the differences in average daily waste recorded pre- and post-intervention between the treatment and control groups via least squares regression using a mixed effect linear model. The approach tests the null hypothesis that the behavior changes (food waste, food purchase, food consumption) attributable to tailored food waste reduction intervention are no different to the behavior changes due to the control stress management intervention. Key requirements for a difference in difference analysis to yield internal validity for such a hypothesis test include that (1) baseline outcome levels do not affect assignment to treatment, (2) there is no spillover between the treatment and control groups and (3) that baseline outcome levels feature parallel trends prior to the treatment period.⁵⁷ Random assignment to treatment satisfies requirement (1) while the individual structure of staff interactions with both treatment and control groups ensures requirement (2). Requirement (3) is validated as detailed in the Discussion section.

Waste was assessed across all domains (preparation, eating, and discard settings) and for each setting individually. Secondary analysis included assessment of the intervention on food consumption, including total energy (Calorie) consumed and Food Pattern Equivalent Data (FPED), and food purchases. For each waste outcome a compound symmetric covariance matrix was used for the random subject effect. The only covariates included were fixed effects, which included time, treatment, and the interaction of these factors. The only exception was for meal-level waste modeling, where mealtime (breakfast, lunch and dinner) and the three-way interaction of these covariates were also included. The only exception to the use of the mixed effects linear model is for the assessment of the intervention on the consumption of food types and macronutrients, which was assessed with a Wilcoxon signed-rank test because the normality of the data was rejected via a Kolmogorov-Smirnov test. The Wilcoxon test was applied to the difference between the mean observation from the pre-treatment and post-treatment periods. Outliers were defined as an observation that has a studentized residual value greater than 3. Each outlier was individually checked to ensure it was a correctly recorded value and not a reporting error. All outliers identified in the data were verified as correctly reported, and thus no data was removed from any analyses.

Results

Overall summary statistics are provided in Table 1 while the key treatment effects are summarized in Figure 1, which reveals that the tailored intervention focused on reducing wasted food yielded both direct and indirect effects. The intervention’s effectiveness in reducing avoidable food waste was consistently revealed in the reduction of the amount of plate waste created compared to the control group (−78.8%), while other sources of wasted food (during meal preparation and storage clean outs) were not significantly altered by the intervention. When aggregated across all possible sources, the effect size remains large (−46.6%, see Figure 2, panel b), but there is considerable heterogeneity in the effect, which results in statistical insignificance (p = 0.26). Within the significant plate waste finding, the largest effects occur during the dinner meal followed by lunch with insignificant effects at breakfast and during snacks (Figure 3).

Table 1.

Summary Statistics

Variable	Mean	Std. Dev.	Min	Max	N
Avoidable Waste	741.5	984.6	0	5888.4	80
Preparation Waste	26.8	54.1	0	269.5	80
Plate Waste	278.1	314.0	0	1412.6	80
Plate Waste – Breakfast	76.9	158.8	0	951.6	79
Plate Waste - Lunch	101.9	169.4	0	807.8	79
Plate Waste – Dinner	99.8	142.6	0	689.4	80
Plate Waste – Snack	34.5	100.0	0	630.0	66
Food Selection	9064.8	3582.4	3287	18675.3	80
Food Intake	8409.4	3412.1	3201	17164.8	80
# FV Items Purchased/trip	4.1	5.0	0	22.0	66
Item Discarded to Landfill/Sewer*	0.90	0.3	0	1	647

Notes: all units in grams/week unless otherwise noted.

^*coded =1 if a discarded food item was sent to landfill or sewer/drain and =0 otherwise.

Figure 1. Summary of Treatment Effects.

Effect size is calculated as the difference in difference regression coefficient as a percent of the treatment group’s average at baseline. ‘No sig. change’ is listed if the difference in difference coefficient is not statistically significant as determined by the p-value from a two-sided t-test. All presented effects are significant at the 5% level. Fruit and Vegetable Purchase units are in number of individual fruit and vegetable items purchased per week. Waste figures refer to avoidable food waste (excluding unavoidable waste, such as pits and bones). Food selection is measured in edible grams selected per week, food consumption is in grams consumed per week, plate waste saved for later is in g/week stored for intended future consumption. Waste sent to landfill refers to the percent of uneaten items not saved for later consumption that are sent to landfill or sewer per week. Detailed results for Fruit & Veg Purchase are presented in Table 2, column 4. See supplemental information (Tables SI-1 to SI-7) for all other detailed results.

Figure 2. Box Plot of (a) Plate Waste and (b) Avoidable Food Waste by Study Time Period for Control and Treatment.

Avoidable food waste includes all waste sources (preparation, plate waste, cleanouts). The upper and lower ends of the box represent the 75^th and 25^th percentile of the raw data; the solid horizontal bar within the box depicts the median; and the whiskers depict the smallest and largest observations within the inner fence (the 25^th percentile less 1.5 times the interquartile range (IQR)) and outer fence (the 75^th percentile plus 1.5 times the IQR) of the data. Any observation greater than the upper fence is depicted as an individual dot. The difference in difference regression coefficient value as a percent of the treatment group’s baseline (pre) waste is 78.8% for food waste (p = 0.001) and 46.6% for all avoidable food waste (p = 0.26). p – values are from two-sided t-tests of the random effects least squares estimate of the difference in differences coefficient ([post-control – pre-control] - [post-treatment – pre-treatment]) where positive numbers indicate greater decreases in waste by the treatment group. Detailed regression results are in Tables SI-4 and SI-8.

Figure 3. Difference in Difference Regression Coefficients (g) for Avoidable Food Waste by Functional Domain.

Error bars are 95% confidence intervals. **, * denotes the regression difference-in-difference coefficient for this domain is statistically significant at the 5% and 10% level as determined by a two-sided t-test. Positive numbers indicate greater decreases in waste by the treatment group. Detailed results for preparation, clean outs, plate waste, breakfast, lunch, dinner and snack are in Tables SI-1, SI-6, SI-4, SI-9, SI-10, SI-11 and SI-12.

Concurrent and antecedent and behaviors

Several concurrent behaviors were also significantly affected by the food waste reduction treatment, including food selection and food consumption (Figure 4). When measured in grams, the decline in food selection by the treated group is statistically significant while the decline in consumption is marginally significant. However, when measured in Calories (see table SI-13 and S-14), the pre-treatment group averages are more similar (within 3%) and the difference-indifference coefficients are no longer statistically significant, suggesting that treatment affected the types of foods selected and consumed as well as simply the mass of food. Analysis of food consumption by food category and macronutrient group (see tables SI-15) reveals that the treated group experienced marginally larger declines in the consumption of grains and cereals (13.5%, p = 0.10) and marginally larger reductions in carbohydrate consumption (18.2%, p = 0.08). No other food categories (protein foods, fruits, vegetables, dairy, oils, fats, added sugars) or macronutrients (protein, fats) revealed any significant differences.

Figure 4. Box Plot of (a) Edible Food Selection and (b) Food Consumption by Study Time Period for Control and Treatment Groups.

The upper and lower ends of the box represent the 75^th and 25^th percentile of the raw data; the solid horizontal bar within the box depicts the median; and the whiskers depict the smallest and largest observations within the inner fence (the 25^th percentile less 1.5 times the interquartile range (IQR)) and outer fence (the 75^th percentile plus 1.5 times the IQR) of the data. Any observation greater than the upper fence is depicted as an individual dot. The difference in difference regression coefficient value as a percent of the treatment group’s baseline (pre) waste is −19.2% for selection (p = 0.005) and −14.9% for consumption (p = 0.032). p – values are from two-sided t-tests of the random effects least squares estimate of the difference in differences coefficient ([post-control – pre-control] - [posttreatment – pre-treatment]) where negative numbers indicate greater decreases by the treatment group. Detailed regression results are in Tables SI-2 and SI-3.

In terms of antecedents, we analyzed shopping patterns (Table 2) and waste created during food preparation (Figure 3). We found no significant treatment effects on the number of fruit and vegetable items purchased (total or fresh), the number of shopping occasions, nor the amount of food waste created during preparation.

Table 2:

OLS Regression Estimates - Weekly Shopping Outcomes

Variables	(1) Any Food Shopping	(2) # Shopping Occasions	(4) # FV Items	(5) # Fresh FV Items
Post	−0.278 (0.195)	−0.556* (0.283)	−2.111* (1.056)	−2.222** (0.926)
Treatment*Post	0.344 (0.248)	0.489 (0.400)	1.178 (1.912)	1.489 (1.804)
Constant	0.788*** (0.064)	1.212*** (0.101)	3.909*** (0.463)	3.455*** (0.433)
Observations	66	66	66	66
R-squared	0.081	0.120	0.096	0.112

Notes: Post refers to change in the outcome from baseline to post-intervention. FV = fruits and vegetables. Treatment*Post is the coefficient of interest with statistical insignificance suggesting no effect of the food waste treatment (versus the stress reduction control) on the shopping behavior listing in the column heading. Robust standard errors in parentheses. Statistical significance of the two-sided t-tests (63 df) that regression coefficients differ from zero are indicated as:

^***p<0.01

^**p<0.05

^*p<0.1.

Attendant Behaviors: Waste Destination

In addition to wasting less food, participants in the treatment group were also coached on how to save food items for later consumption and how to divert uneaten food items and inedible waste from the landfill or sewer by using the items to feed pets or through composting or sharing surplus food. While there was no significant change in the amount saved for later consumption (Figure 1), there was a nearly 20 percentage point decline in the percent of disposed items that went to landfill or via sink to the sewer, which was a significant reduction compared to the control group (Figure 5).

Figure 5. % of Plate Waste Items Sent to Landfill or Sewer.

Error bars depict 95% confidence intervals. N = 647. Difference in difference estimate from a linear probability regression model is significantly different than zero (−0.223, p < 0.001 from a two-sided t-test with 643 df), yielding an effect size of −24.6% (−0.223 divided by a baseline fraction of 0.907 expressed as a percent). See Table SI-7 for detailed results.

Discussion

In this study, an individually tailored food waste reduction intervention yielded large and statistically significant reductions in plate waste. The tailored intervention’s effect size is relatively large compared to other rigorously tested food waste interventions and aligns with extant literature documenting that technology-tailed interventions in the health and wellness space can deliver large effect sizes.²⁴ Specifically, in the realm of food waste reduction, previous studies have demonstrated reductions of 24%,³⁷ 29%³⁸ or did not report an effect size due to insignificant differences between treatment and control groups during post-intervention measurements.³⁹ The results of this study reveal a 79% reduction for the behavior that became the focus of the coaching interventions, and this was statistically significant. Given that participants in the treatment group averaged about a half a kilogram of avoidable plate waste per week (495.76 g/week), this suggests that if the gains could be sustained for a year, more than 20 kg of wasted food would be avoided from reduced plate waste alone. When aggregating across all sources of avoidable waste (preparation, plate waste and food storage clean outs), the effect size remains large (a 46% reduction compared to controls, see Figure 2, panel (b)), but the heterogeneity inherent in infrequent food storage clean out events (see Figure 3) and the short duration of this project decreased the precision of this overall effect and drove it to statistical insignificance (p = 0.26).

We note that for several measures, the treatment and control groups feature significantly different weekly values during the pre-treatment phase. Hence, one must be concerned that any significant difference-in-difference regression coefficients may simply be capturing a regression to the mean effect that occurs incidentally during the post-treatment phase rather than the effect of the treatment itself. To explore this alternative explanation, we plot daily averages during the seven individual days of the pre-treatment measurement for the key category of avoidable plate waste (see Figure SI-1). If regression to the mean were the driver of this result, we would expect daily trends in plate waste during this period to be negative and distinct from daily trends in the control group. We find both treatment and control groups feature negative daily trend rates with no statistically significant difference between treatment and control rates (see notes of Figure SI-1). Hence, the data reveal parallel trends between the treatment and control groups during the pre-treatment phase, which meets the third requirement detailed in the methods section for this approach to provide internal validity.⁵⁷

The data monitoring and measurement infrastructure deployed in the study lent itself to providing highly granular insights (e.g., differences between reductions observed during different meals occasions) and to identifying spillover effects across different parts of the participant’s in-home food system. The finding that greater plate waste reductions occurred during lunch and dinner meals is not too surprising given that these meals typically feature greater variety and amounts of Calorie consumption than breakfast or snack occasions. Several such spillovers were identified, including a marginally significant reduction in the concurrent behavior of food consumption, and a significant reduction in portion size selection. In this project, portion size selection is not a spillover, however, since some participants were instructed to reduce the amount of food that they served to themselves as a means to reduce plate waste. Furthermore, the attendant behavior of diverting waste from landfill was also significantly bolstered. However, no significant change was measured for waste created during less frequently occurring events such as food storage clean out events or meal preparation (not all meals involve preparation due to heat-and-serve items, ready-to-eat items, take-out meals, etc.). This pattern may occur for several reasons. First, the intervention lasted about one week, which makes setting goals and revising actions in response to less frequent events more difficult. With food selection, portion sizing and plate waste disposition occurring multiple times per day, these actions may have more easily lent themselves to coaching feedback than other activities (shopping tactics, produce storage and trimming advice, etc.). For example, only about 64% of U.S. consumers report cleaning out their refrigerator often.⁴⁵ However, given the current design, we cannot rule out that, even with a longer intervention time frame, these events may simply prove more difficult for participants to change, but we do not feel this would be the case.

We find that the tailored intervention did not change antecedents such as the purchasing and consumption of fresh fruits and vegetables, which may affect nutrition and health outcomes. Given that fruits and vegetables are among the most wasted of all food items in U.S. homes,⁵⁸ one could imagine undirected admonitions to reduce waste leading to reductions in the consumption of these health sustaining foods. While there was a significant reduction in overall food consumption for the treatment group, the only subsets of food featuring marginally significant reductions in consumption were carbohydrates in the form of cereals and grains, while the consumption of fruits and vegetables did not differ significantly between treatment and control.

While the above results show promise, we acknowledge several limitations of the study that must be addressed in subsequent work to explore the robustness and efficacy of technology aided tailored interventions. First, the sample of participants is small, largely female, and from a single location, which limits generalizability. Also, the sample consists only of those willing participate. While this is not unusual for pro-environmental behavioral interventions, the intensive nature of intervention (e.g., technology training and multiple coaching sessions) may drive additional selectivity beyond that observed in less intensive interventions. In addition, the coaching and plate waste measurements focused on a single individual within a household. While food preparation and food storage cleanout measurements likely encompassed the activities relevant to all household members, additional measurements across household members would be needed to understand the intra-household spillover effects of the coaching intervention on plate waste activities and, hence, the full effect of the intervention on waste. Finally, the study time frame was limited to one week of baseline measurement and one week of intervention with no medium or long-term follow up measurements to assess intervention dynamics. Future studies would benefit from longer durations and follow up measurement periods after coaching has ceased, though we recognize doing so may further exacerbate the selectivity issues previously highlighted.

Our results suggest that individual tailoring holds promise for increasing the effect sizes associated with sustainable consumer behavior interventions without inducing or exacerbating unwanted antecedent, concurrent or attendant behaviors. However, additional research is required to confirm these promising initial results and to ensure that any effect size enhancement can be generated at scale. The intensiveness of the intervention, and the need for individuals to opt into such treatments, poses challenges for intervention scaling. Not only is the intervention intensive for participants, it is also resource intensive for those enacting the intervention, as recruitment, participant training and individual coaching entail substantial staff costs while the technology requirements involve up front development costs as well as back-end data processing costs to provide feedback to enhance coaching efforts. Like in the realm of health and wellness, we foresee the need for further research to introduce more passive measurement techniques (e.g., waste bin cameras and digital scale readings automatically activated during discard) to reduce the burden borne by participants in their daily routines as well as to integrate artificial intelligence into the tailoring process in order to ease resource constraints and leverage scarce and expensive human coaching effort.^59-60 One can imagine pairing non-tailored interventions with follow up tailored interventions in pursuit of the urgent goals of altering individual behavior to reduce emissions.