Predicting curbside food waste generation from responses to Self-Administered Surveys: An Application of Machine learning to improve household estimates

Abstract

Accurate measurement of household food waste is essential for designing and evaluating waste reduction and management policies. However, the mode of waste measurement and the approach to formulating sample populations whose waste is measured can impact accuracy. We find estimates of household food waste differ between measurement approaches applied to the same sample of participants and between the same measurement approach applied to differently recruited samples. The magnitudes of the estimates of per-person once-edible food waste follow the same order as the relative costs of collecting each type of data with the order from low to high being self-administered surveys (SAS) from a national consumer panel (388 g/person/week, $4.42/sample), SAS from locally recruited residents (552 g/person/week, $37.50/sample), and curbside waste composition analysis (WCA) from locally recruited residents (809 g/person/week, $375/sample). This study aims to bridge the gap across measurement and sampling approaches by using machine learning techniques to predict WCA-based edible and inedible food waste from SAS data. Using data from locally recruited respondents with contemporaneous SAS and WCA measurements, we develop least absolute shrinkage and selection operator (LASSO) regression models to predict WCA outcomes with data collected via SAS and compare these predictions against a literature based standard adjustment method. Our results show that LASSO models outperform the root mean square error of the standard adjustment method’s predictions by 9 to 12 percent, hence providing improved ability to predict curbside waste amounts from low-cost survey data. Implications for cost-effective evaluation of policies affecting household food waste levels are discussed.

1. Introduction

Measurement motivates management. However, measurement requires resources with higher quality measurement often costing more. This stimulates short-run questions about how much should be spent on measurement and longer-run questions about how much to invest in research and development to lower the cost of high-quality measurement. In this article, we explore an important tension between the cost and quality of measurement in the context of household food waste generation and present results of research designed to deliver more accurate and granular measurement at a lower cost.

Food waste represents a significant global challenge, contributing to environmental degradation and economic inefficiencies (Babbitt et al. 2022; Bilska et al. 2016; Cattaneo et al. 2021; Kör, Krawczyk and Wakkee 2022). In the United States, food waste accounts for approximately 30–40 % of the food supply, with significant implications for sustainability and resource allocation (Buzby et al., 2014). Households represent the largest single origin of wasted food in the United States (35 % in 2023, ReFED Insights Engine, 2023) with Li et al. (2023) documenting an upward trend in per-person household food waste levels as the U.S. emerged from the COVID lockdown period. To evaluate interventions and policies intended to reduce household food waste generation, structures facilitating measurement and data interoperability are critical (Quested 2019, Babbitt et al. 2022). However, there are multiple options for measuring household food waste generation which vary by cost and accuracy (Quested et al. 2020). While past empirical work has established that various methods can yield strongly correlated responses when applied to the same households (van Herpen et al. 2019, van der Werf et al. 2020), this literature also firmly establishes that certain methods yield systematically smaller estimates, e.g., lower-cost options like retrospective self-administered consumer surveys regularly yield significantly smaller estimates of waste than methods that involve physical collection and measurement of waste (van Herpen et al. 2019, Quested et al. 2020, Shu et al. 2023).

This study involved two key tasks. The first assessed the amount of food wasted by households using two measurement methods and two approaches to recruiting participant samples. A sample of 146 participants provided observations over two consecutive weeks using two measurement methods: a self-administered survey (SAS) and a waste compositional analysis (WCA) of the same household’s curbside waste stream from the week of the survey. The same measurement method (SAS) was also applied to two samples recruited via two approaches: via a bespoke local recruitment effort and via recruitment from a nationally maintained consumer panel.

The second task involved developing machine learning models to predict the WCA amounts with information gathered from the SAS. It is well documented that household characteristics, such as income, education, size, food purchasing patterns, and food management skills, are associated with household food waste generation (González-Santana et al. 2022; Li et al. 2023; Özbük and Coşkun 2020; Shu et al. 2023; Stancu et al. 2016). For example, larger households may generate more total waste (though less on a per-person basis) due to greater food consumption (Li et al., 2023), while higher-income households might waste more due to greater purchasing power (Joardder and Masud, 2019). We hypothesized that some of these same factors would predict the degree of divergence between edible food waste estimates obtained via SAS and WCA and would predict the amount of inedible waste identified via WCA. We employed a Least Absolute Shrinkage and Selection Operator (LASSO) regression and compared the predictions from this machine-learning approach to the standard approach of estimating WCA figures by applying a single under-reporting calibration factor to SAS data.

By examining both self-reported and third-party measures of the same households’ waste, and by assessing the relative costs of implementing the different measurement and sample recruitment approaches, the research aimed to bridge the gap between popular approaches to household waste measurement and understand the trade-offs between cost and accuracy. Self-reported food waste, while convenient and less costly to collect, has regularly been found to result in smaller estimates of household food waste than WCA. However, given that the cost of WCA may be prohibitive, organizations undertaking evaluations of household food waste generation behaviors may have few other options beyond surveys.

This study makes several key contributions. First, it develops a machine learning–based framework that leverages SAS data to predict WCA outcomes, thereby providing a systematic approach to bridging the gap between low-cost and high-accuracy measurement methods. Second, it demonstrates that this approach yields meaningful improvements in predictive accuracy relative to commonly used calibration methods, highlighting the value of incorporating household characteristics and behavioral variables. Third, the study offers a pathway for a practical and scalable solution for estimating household food waste in contexts where direct measurement via WCA is cost-prohibitive, thereby supporting more cost-effective monitoring and evaluation of food waste policies and interventions.

By integrating subjective and objective measures of food waste and employing advanced statistical techniques, this research contributes to a deeper understanding of household waste dynamics. The findings have the potential to guide policymakers and stakeholders in designing effective interventions aimed at minimizing food waste, promoting sustainability, and addressing food insecurity.

2. Materials and methods

2.1. Data collection

2.1.1. Columbus panel – Survey and waste composition analysis

This study relies on two data sources. The first is household-level data collected in Columbus, Ohio. Households were recruited via messages delivered through an opt-in email and text messaging system that provides reminders of waste hauling pickup days and announces schedule changes to Columbus residents. Responding residents who indicated interest in the study provided contact information. Research staff followed up by email or phone explaining that the purpose of the study was to understand household food waste generation and that participants would be required to answer a sequence of two weekly surveys in which they would report their household’s food waste and that during those two survey weeks they would be required to set out their household waste for collection and analysis by research staff. 12 cohorts of 10 to 14 spatially adjacent volunteers were selected for participation in the study, which lasted for two to three weeks. Each cohort was selected from neighborhoods that varied in geography and socio-economic status throughout the city. Participants were drawn from 11 of the city’s 41 zip codes, representing all geographical quadrants of the city. The sampled zip codes have median household income ranging from $51,860 to $92,838 (the city-wide median household income at the zip-code level was $65,327).

The first cohort’s data was collected in August of 2023 while the last cohort’s data was collected in November of 2024 with data collection during each season of the year (3 cohorts in Q1, 2 cohorts in Q2, 4 cohorts in Q3, and 3 cohorts in Q4) to ensure year-round coverage to mitigate potential seasonal biases. Of the 153 participants enrolled, 146 (95 %) provided at least one week of data. Participants provided informed consent and received compensation upon completion of data collection. All study activities were approved by a relevant institutional review board.

Each participant provided at least two forms of data. First, participants answered a sequence of self-administered surveys (SAS) on the Qualtrics platform. The first collected demographic data and prompted respondents to monitor discarded food items in the upcoming weeks. The next two surveys arrived 7 and 14 days later and prompted participants to provide estimates of once-edible (hereafter, edible) food that was discarded by their household over the past 7 days regardless of the method of discard (garbage, sink/disposal, fed to animal, composted). Participants were told not to report items they would not normally eat (e.g., bones, pits). The surveys are similar to those used by Li et al. (2023) and Shu et al. (2023) and prompt participants to record food discards in up to 24 distinct categories (see Table 1; note the 24 categories are not revealed to the respondents during the baseline demographic data collection). For each category, participants mark the appropriate quantity range (e.g., 1–2 cups), which is converted to a gram weight using methods detailed in Shu et al. (2021).

Table 1.

Waste categories by method.

Self-administered Survey (SAS) – edible only	Curbside Waste Composition Analysis (WCA)
Fresh vegetables & salads	Produce
Other vegetables (jar / canned / frozen)
Fresh fruit
Other fruit (jar / canned / dried / frozen)
Potatoes	Potatoes and products
Potato products (fries, hash browns, but report potato chips under ‘salty snacks’)
Pasta	Staples
Rice & other grains (including wraps, couscous, etc.)
Beans, lentils, chickpeas, etc.
Cereals (breakfast cereal, corn meal, oats, etc.)
Bread
Meat (report deli meat under ‘sandwich ingredients’)	Meat and fish
Fish
Sandwich ingredients (deli meats, cheese slices, relishes, etc., but report lettuce & vegetables in ‘fresh vegetables and salads’)
Meat substitutes
Dairy products and cheese (report cheese slices under ‘sandwich ingredients’)	Dairy and eggs
Eggs
Yogurt, custard, etc.
Soups / stews	Soups and sauces
Condiments and sauces (ketchup, mayonnaise, cocktail sauce, etc.)
Candy / cookies / granola bars / chocolate bars	Snacks
Salty snacks (chips / nuts / pretzels, etc.)
Non-alcoholic beverages (milk, juice, soda. Exclude: water, tea, coffee)	Beverages
Alcoholic beverages
	Inedible from any category (bones, peels, stumps, shells, spent fryer grease and oils, coffee grounds, etc.)
	All other trash

Second, and in addition to the surveys, each participant was directed to ensure their normal weekly garbage output was placed in a location that was accessible to research staff for the purposes of curbside waste composition analysis (WCA). Participants were asked to exclude animal excrement, chemicals, diapers, medical waste, and excessively heavy items (e.g., unwanted furniture) from bags set aside for collection and to include food items that would normally be fed to animals or composted. Experienced contractors collected the bags and, for each household, sorted all the material into one of 10 bins (Table 1, column 2) with food packaging removed where feasible. Fewer categories are used in the WCA than the SAS because of cost considerations (labor costs for the WCA increase in the number of categories chosen) and due to practicalities (e.g., it is impossible for a third-party sorter to know if the discarded broccoli was purchased fresh or frozen). The categories for the survey were selected to adhere to those of a household food waste survey that is frequently administered in the Unites States (Li et al. 2023, Shu et al. 2023), and the WCA categories were selected such that they constituted a consolidation of the 24 survey categories to ensure that food category specific correspondence could be assessed between SAS and WCA. The eight categories of once-edible foods (produce, potatoes and products, staples, meat and fish, dairy and eggs, soups and sauces, snacks, beverages), the total of all once-edible foods, and the amount of inedible food scraps (bones, pits, peels) are each the focus of later modeling efforts as each category may lend itself to different intervention efforts. For example, a campaign focused on provision of fresh produce to bolster nutrition may desire accurate prediction of curbside produce waste (Roe et al. 2026), while campaigns focused on composting uptake may seek accurate prediction of inedible food scraps in curbside collections (Shu et al. 2023).

Each bin was weighed and photographed. Best practices were implemented with bin tare weights assessed prior to each household’s sort and with sorters supervised by the same leader for all cohorts to ensure temporal consistency.

Photos of each bin were reviewed by two independent raters. Raters adjusted previously recorded bin weights when they identified (1) items with substantial inedible parts in bins labeled as edible food (e.g., observed a whole watermelon in the edible produce bin and, using USDA standard reference estimates of the weight of a watermelon rind, increased the value of the inedible food bin and decreased the value of the edible produce bin), (2) items with substantial edible parts in the inedible bin, or (3) identified items that were not de-packaged (e.g., identified a can of peaches, estimated the can weight using industry material weight figures, deducted the estimated can weight from the edible produce bin and added it to the trash bin estimate). The average of the two raters’ estimates was used and, if rater estimates diverged by more than 50 g, a third rater reviewed the photographs with the average of the three raters’ adjustments being used. 10.9 % of all bins for edible food categories featured a photo-rater’s adjustment to account for inedible portions while 17.4 % featured adjustment for packaging. Sample wide, these adjustments largely cancelled out: between adjustments for inedible parts and packaging, the average adjustment in bin weight was an increase of 1.2 % while the median adjustment was 0 %.

132 households provided complete data (SAS and WCA) for two weeks, while 14 households provided one week of complete data. Data was pooled from both reporting periods for analysis as paired t-tests revealed no significant differences in average food waste estimates across weeks for either SAS (p = 0.61) or WCA (p = 0.70). The final data set contains 278 observations. A subset of 26 participants also used the FoodImage smart phone app (Roe et al. 2020) to collect food waste generation data for two weeks and completed a third week of surveys, though these additional data will be analyzed elsewhere.

2.1.2. National panel – SAS

To develop a scalable and widely applicable approach for predicting the amount of waste at curbside based upon SAS, we also seek to address sources of bias that might arise from the methods of recruitment used to draw the sample of households from the Columbus, Ohio community. For example, participants in the Columbus, Ohio sample were recruited by using a messaging system used by the local government, while samples for national studies employed for similar surveys are often drawn from commercially maintained consumer panels (e.g., Li et al. 2023, Li and Roe 2024).

Participants recruited via different approaches could feasibly suffer from different levels of selectivity bias and behavioral reactivity, which are known challenges to measurement validity (Quested et al. 2020). In terms of selectivity bias, consumers who are in large consumer panels often answer surveys on a wide variety of topics as a means to augment income, which may lessen the tendency for those with stronger interests in food waste to select into the study and rigorously report food waste quantities compared to those like the Columbus sample who are recruited for a single study opportunity focused on food waste. In terms of behavioral reactivity, those recruited via a local messaging system, who are also aware that researchers will observe their curbside waste, may alter their survey reporting behavior to ensure greater consistency between the two measures, while national panelists have no similar motivation.

To assess whether these two recruitment methods may drive differences in SAS food waste estimates, we recruited a sample of consumers from throughout the United States at four different time points (October 2023, February 2024, April 2024, August 2024) that overlap with the Columbus, Ohio study timeframe. These consumers (hereafter, the National Panel) were drawn from the Prolific consumer panel platform (Douglas et al., 2023; Peer et al., 2021) and, parallel to the Columbus, Ohio sample, completed a sequence of weekly surveys over the course of three weeks including a first survey to capture household information and prompt participants to monitor food discards over the coming weeks, and then three follow up surveys in which food discards over the past 7 days were reported. For this article, we focus on 599 National panel participants who provided one or two weeks of responses to mirror the data gathered from the Columbus panel (542 provided two weeks; 57 provided a single week for 1141 total observations).

2.2. Prediction via LASSO regression

We employed the Least Absolute Shrinkage and Selection Operator (LASSO) regression (Tibshirani 1996) to identify the most influential predictors of the level of food waste obtained from the curbside WCA among the variables available from the SAS while addressing potential multicollinearity and model overfitting. Given that the underlying goal was not only to predict curbside food waste levels but also to identify key predictors from SAS data, LASSO’s shrinkage and variable elimination properties are particularly well-suited for high-dimensional data environments where feature selection is necessary to isolate the most relevant explanatory variables.¹ By applying LASSO, we ensured that the final model retains only the strongest predictors.²

LASSO performs variable selection by introducing an L1 regularization penalty, which shrinks the coefficients of less important variables to zero, effectively eliminating them from the model. This feature is particularly beneficial because many household characteristics and socioeconomic factors are correlated. The objective function for LASSO regression is given by:

$$\underset{{\beta }_0,\beta }{\text{min}}\frac{1}{2n}\sum _{i=1}^n{\left(y_i-{\beta }_0-\sum _{j=1}^p{\beta }_j{x}_{ij}\right)}^2+\lambda \sum _{j=1}^p\left|{\beta }_j\right|$$ (1)

where y_i represents the curbside WCA waste value for household i, and x_ij denotes the j-th predictor for household i. β_j is the coefficient for the predictors, n is the number of households, p is the total number of predictors and λ is the regulation parameter controlling the degree of shrinkage.

We randomly selected 236 (85 %) observations to train the LASSO model and used the remaining 42 observations to validate the model. This method ensures that the model is properly trained while reserving a portion of the data for assessing its predictive accuracy. The LASSO model was trained using Stata’s LASSO linear regression function (version 18), applying a 10-fold cross-validation to determine the optimal regularization parameter λ that minimizes prediction error. We used a 10 % significance level for selecting predictors; only significant predictors are maintained in the final model. Cross-validation supports model robustness by ensuring that the model generalizes well to new data, preventing overfitting while maintaining predictive performance.

This study employs two LASSO models to predict edible food waste: a short model and a long model. The short model includes only household characteristics and socioeconomic variables (e.g., household size, income, education, shopping frequency, and demographic indicators). This specification ensures broad applicability, as general survey datasets typically contain these variables. The long model extends the short model by incorporating additional food-related behavioral variables. The inclusion of these predictors can enhance the precision of curbside food waste estimation while leveraging the improved accuracy of the Columbus survey data. Estimation of both models helps understand the improvement possible from administering more comprehensive surveys.

After fitting both models, we evaluated predictive performance by comparing Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and R-squared values across the test dataset. Additionally, we analyzed the distribution of residuals to ensure that prediction errors were not systematically biased. Finally, we compared the results to a standard approach from the literature in which the estimates from SAS are adjusted by a calibration factor that represents the mean under-reporting rate for SAS versus WCA. We assess the relative improvement from applying the LASSO machine learning model by assessing the percent reduction in the three prediction performance metrics (MSE, RMSE, R²). We applied the under-reporting rate for the 236 observations in the model fitting sample to each SAS estimate from the hold out sample.

The study incorporates an additional LASSO model to predict the amount of inedible food waste collected at the curbside based upon survey data, which prompts participants to record only edible items that were discarded. Predicting inedible food waste is crucial for assessing the survey’s effectiveness in estimating the total organic waste stream, comprising both edible and inedible components, entering landfills. We use both a short and a long model for estimating inedible food waste, consistent with the LASSO models employed to predict edible food waste. By predicting both edible and inedible food waste amounts, the study evaluates the extent to which SAS data can serve as a reliable proxy for overall organic waste generation, thereby improving waste management strategies and policy interventions.

2.3. Calibrating for potential biases of local study recruitment

To gauge the extent to which the mode of survey recruitment may influence the estimated level of food waste obtained via SAS, we pool data from the Columbus and National panels and estimate a censored regression model that explains the variation in weekly per-person food waste as a function of household characteristics, the timeframe of data collection, and the recruitment method (National Panel vs. Columbus):

$$TWP{P}_i=\alpha +N{P}_i\gamma +X_i\theta +{\in }_i$$ (2)

where TWPP_i represents total waste per person, NP_i is an indicator variable (=1 for National Panel observations, =0 otherwise), X_i is a vector of demographic, socioeconomic and survey date factors, and ∊_i is the error term. The estimated National Panel coefficient (γ) captures the systematic difference in self-reported waste levels between the two datasets, which is tested for any difference from zero using a t-test. A censored regression model (the Tobit procedure in Stata) is selected as it accounts for censoring of the dependent variable (i.e., many survey respondents report zero waste³).

Since future researchers may only have access to national panelists rather than locally recruited samples, this result provides a first-stage approach to adjusting estimates from National Panel surveys prior to predicting curbside levels from the LASSO model described previously. By leveraging the self-reported food waste data from both food waste data from Columbus and the national dataset, we create a bridge that links the two datasets, enabling a first attempt of a more generalizable prediction framework for food waste estimation.

3. Results and Discussion

3.1. Summary statistics

Table 2 provides summary statistics of participating households in the Columbus, Ohio sample. While representing a broad cross-section of the city, the sample contains more participants identifying as white, non-Hispanic, college-educated, older than 35 years of age, and higher income than would be suggested by recent Census figures.

Table 2.

Household Characteristics of Columbus Sample.

Variable	Sample	Census
Household Size: Mean	2.70	3.03
Median	2.00
SD	1.52
Race
White	0.75	0.53
Black	0.18	0.29
Other	0.07	0.18
Hispanic	0.04	0.08
Female	0.78
Age
18–34	0.18	0.40
35–44	0.24	0.18
45–54	0.25	0.14
55–64	0.17	0.14
65 and older	0.17	0.14
# Survey Waste Categories Markeda: Mean	4.90
Median	4.00
SD	3.15
Any Unexpected Issues During Waste Reporting Weekb	0.85
Marked Zero Waste on Survey	0.03
Percent of Time at Home: Mean	63.03
Median	70.00
SD	21.60
Employment
Full-time	0.62
Self Employed	0.14
Retired	0.16
Unemployed	0.05
Other	0.03
Annual Household Income
Income < 50 k	0.20	0.39
50 k<=Income < 100 k	0.41	0.31
Income >= 100 k	0.39	0.30
No Special Dietary Restrictions	0.74
Eligible for Supplemental Nutrition Assistance Programc	0.70
Eligible for Women, Infants and Children Nutrition Programc	0.04
House Ownership
Single Family House	0.90
Other	0.10
Living in Urban (vs. Suburban)	0.50
Education
High School or less	0.05	0.35
Some College	0.30	0.25
Bachelor’s Degree	0.30	0.24
Graduate/Professional Degree	0.35	0.16
Frequency of Grocery Shopping
Twice A Week	0.23
Once A Week	0.47
2–3 Times A Month	0.22
Once A Month or Less Often	0.07

Notes: Based on the 146 Columbus households participating in both self-administered surveys (SAS) and curbside waste composition analysis (WCA). Census figures from U.S. Census Bureau for the city of Columbus, OH.

^aNumber of survey categories (out of 24) with positive levels of reported waste.

^bVariable = 1 if respondent indicated no food waste on the survey.

^cVariable = 1 if respondent agreed there was an unexpected event (e.g., unexpected guests, dinner guest no-shows, ate out unexpectedly, obtained low quality food, hosted an event with food, tried new recipe, fewer meals at home, expired items from sale purchases, another issue) that caused an unusual amount of waste during the reporting week; =0 otherwise.

^dVariable = 1 if respondent reported eligibility for this program; =0 otherwise.

The dataset includes food waste data derived from SAS and curbside WCA, which is detailed by method and food category in Table 3. Among all once-edible food items, the mean WCA estimate is 32 % higher than the mean SAS estimate, though the median WCA estimate is only 9 % greater than the median SAS estimate. The variability of the WCA estimates, as measured by the coefficient of variation, is 66 % greater than that of the SAS estimates (1.66 vs. 1.06). Notably, both SAS and WCA have weekly recordings that feature zero.

Table 3.

Waste estimates (g/person/week) by method and category.

	Mean	Median	SD	Min	Max	% of Mean Edible Waste
Produce
SAS	242	173	263	0	1721	43.8
WCA	258	91	444	0	3538	31.9***
Ratio	0.94	1.91	0.59	–-	0.49	–-
Potatoes
SAS	31	0	82	0	845	5.6
WCA	51	0	159	0	1648	6.3
Ratio	0.61	–-	0.52	–-	0.51	–-
Staples
SAS	112	39	180	0	1325	20.3
WCA	163	91	274	0	3084	20.1
Ratio	0.69	0.43	0.66	–-	0.43	–-
Meat & Fish
SAS	45	0	93	0	900	8.2
WCA	70	0	135	0	907	8.7
Ratio	0.64	–-	0.69	–-	0.99	–-
Dairy & Eggs
SAS	40	0	103	0	840	7.2
WCA	40	0	106	0	771	5.0
Ratio	0.99	–-	0.98	–-	1.09	–-
Soups & Sauces
SAS	32	0	117	0	1260	5.8
WCA	115	0	343	0	3402	14.2***
Ratio	0.28	–-	0.34	–-	0.37	–-
Soups & Sauces
SAS	5	0	14	0	120	0.9
WCA	72	0	158	0	1225	8.9***
Ratio	0.07	–-	0.09	–-	0.10	–-
Beverages
SAS	44	0	180	0	2375	8.0
WCA	39	0	137	0	1134	4.8
Ratio	1.14	–-	1.31	–-	2.09	–-
Inediblea	1248	896	1166	0	7439	–-
Other trasha	2508	1814	2604	54	24,404	–-
Total Edible
SAS	552	414	583	0	5115	–-
WCA	809	454	1024	0	6759	–-
Ratio	0.68	0.91	0.57	–-	0.76	–-

Notes: g/person/week unless otherwise noted. N = 278. Columbus, Ohio sample only. SAS is self-administered survey, WCA is curbside waste composition analysis. Ratio is SAS/WCA.

^adenotes the measure only available for the WCA data.

^{*, **, ***}denotes the % Total of Edible Waste in this category differs between SAS and WCA data at the 10 %, 5 % and 1 % level, respectively.

We are not the first to document the difference between the household food waste generation estimates generated by different measurement approaches. Shu et al. (2023), compare estimates from data from 210 households in a single suburb of Columbus, Ohio, and find that SAS estimates are 59 % smaller than WCA estimates. The downward bias from Shu et al. (2023) is nearly two-fold greater than the 32 % downward bias in the current study (552/809 = 0.68). However, in the Shu et al. study, the WCA measurement occurs several weeks after the SAS was collected (as opposed to the same week for the current study) and data collection only occurred during two seasons of the year (late winter and summer) as opposed to during every season in the current study, making comparisons imperfect. Van Herpen et al. (2019) collect food waste generation data in multiple ways from 92 European households and, unlike Shu et al. (2023), different data collections occur during the same time period within each household. They find that the estimates from a SAS are 39 % less than the estimate from a kitchen caddie measurement approach. The kitchen caddie approach is similar to the current study’s WCA approach in that researchers weigh the waste collected by household members. However, the kitchen caddie approach requires a household member to separately collect food waste during the week whereas the WCA approach in the current study merely requires a household member to set aside all household waste from the week without the extra burden of separating food out of the waste stream. Quested et al. (2020) report results from five studies featuring household food waste generation measured by both WCA and food waste diaries. Food waste diaries are similar to SAS in that a household member is required to observe and report food waste quantities, but are distinct in that the reporting typically occurs on a daily basis and, hence, suffers less recall bias. Quested et al. (2020) report downward bias levels that range from 7 % to 36 % from the two studies where the two methods were applied to the same households contemporaneously (sample sizes ranging from 120 to 164). Hence, the current study’s finding of a within-household downward bias of 32 % for SAS versus WCA is within the range of past studies.

However, in the current study, the pattern of survey under-reporting is not uniform across food categories. Mean WCA estimates are much larger than mean SAS estimates for the categories of snacks (14x) and soups/sauces (3.6x) while mean estimates from WCA and SAS are within 10 % of one another for produce and dairy/eggs categories. Furthermore, for the beverages category, the mean SAS estimates exceed mean WCA estimates by 14 %, which is not unexpected given the convenience of discarding unwanted beverages down the sink, which would prevent their measurement via curbside WCA. In terms of the categorical composition of once-edible waste streams as captured by the two methods, the proportion attributable to any one category is similar in five of the eight categories (potatoes, staples, meat/fish, dairy/eggs, beverages) and statistically different in the remaining three (produce, soups/sauces, snacks). We note that previous research has not articulated measurement differences at the food category level.

3.2. Adjusting survey estimates for National vs. Local recruitment

The summary statistics for the national panel are reported in Table A.1 and Table A.2 (supplemental materials). We note that, in this case, the Columbus sample showed higher average food waste across all categories compared to the National sample. For example, the mean total food waste per person reported in Columbus is 552.3 g, compared to 387.7 g nationally (see Fig. S2 for visual comparison across all three methods). The pattern holds across all individual categories such as fruit and vegetables (241.9 g vs. 147.5 g) and meat and fish (45.2 g vs. 32.6 g), suggesting systematic underreporting or behavioral differences in the national sample. Furthermore, Table 2 and Table A.2 highlight demographic contrasts between Columbus and national samples. Columbus households tend to be larger (mean 2.71 vs. 2.15), and more likely to hold graduate and professional degrees (34 % vs. 19.5 %). These differences likely contribute to the observed gap in reported food waste amounts and underscore the importance of calibrating national self-reported data when estimating actual food waste levels.

Table 4 presents the results of a censored regression model analyzing the relationship between self-reported food waste data from Columbus households and the National panel. The analysis is crucial for assessing the relationship between results gathered via different survey recruitment approaches as the regression approach controls for observable differences in household characteristics noted above. The coefficient for recruitment via the National panel is −161.317 (p < 0.05), indicating that recruitment from a national panel resulted in significantly smaller self-reported levels of waste than among those recruited in Columbus.

Table 4.

Cross-Survey Relationship: Columbus vs. National Self-Reported Food Waste.

Variable	Regression Coefficient	Standard Error
National Panel (vs. locally-recruited Panel)	−161.3**	51.6
Household Size	−133.8***	16.0
Any Children (0–5)	99.2	59.9
Any Children (6–17)	19.5	50.7
Household % Female	−3.8	45.6
Education (base: High School or Less)
Some College	−53.1	51.7
Bachelor’s degree	−161.3**	53.4
Grad/Professional Degree	−48.7	57.0
Household Income (base: <$50 k)
$50 k – $99 k	48.8	36.1
$100 k +	112.7**	40.3
Employment Status (base: unemployed)
Full-time employment (≥35 h / week)	−32.7	49.0
Part-time employment (≤35 h / week)	42.6	54.0
Retired	−136.8*	53.3
Student	170.2	203.6
Shop Grocery Frequency (base: Shop > Weekly)
Shop Less than Weekly	−74.0	41.7
Shop Weekly	−60.2	34.9
Special Diet	−52.1	32.9
Shop Online or Use Mealkit	54.1	28.8
Urbanicity (base: Rural)
Suburban	−16.3	40.5
Urban	39.4	44.4
Region (base: Northeast)
Midwest	75.5	44.6
South	−5.8	43.8
West	−5.5	45.2
Year.Quarter (base: 2023.Q3)
2023.Q4	193.0	100.1
2024.Q1	−8.1	40.7
2024.Q2	61.8	43.5
2024.Q3	42.6	41.5
2024.Q4	−63.9	96.2
Constant	858.0***	112.0
Pseudo R Square	0.009	–-
Observations	1417	–-

Notes: The dependent variable is edible food waste (g/person/wk). Data from the Columbus, Ohio, and National panel participants. 162 observations are left-censored at 0.

^*p < 0.10,

^**p < 0.05,

^***p < 0.01.

The finding suggests that national panelists systematically report lower food waste compared to locally recruited Columbus households, potentially due to the differences in self-selection and behavioral reactivity biases across recruitment methods. The discrepancy underscores the need for calibration adjustments to align national self-reported data with more localized measurements, improving the reliability of food waste predictions.

3.3. Predicting curbside edible waste from survey responses

Due to the literature’s focus on WCA as a preferred measure, and due to its greater cost of execution, we treat the WCA measures as the focal variable to be predicted. We treat the SAS measure and other survey variables as the source for prediction variables and rely upon the data collected from households in the Columbus, Ohio sample. Fig. 1 plots food waste estimates from SAS on the y-axis against food waste estimates from curbside WCA on the x-axis. Most data points fall below the 45-degree line, reflecting that total self-reported edible waste is lower than the amount found curbside, reflecting statistics from Table 3. Fig. S1 plots food waste data from SASs on the y-axis against the difference (SAS waste – WCA waste) on the x-axis. The mean difference is −257 g per person with a standard deviation of 832.45 and a range of −6115.89 to 1672.58, reinforcing the patterns discussed in Table 3 as well as most findings in the literature. Fig. 1 and S1 visually confirm the under-reporting and motivation for building a model that can predict WCA estimates based upon self-reported data using machine learning methods.

Fig. 1.

Edible food waste (g/person/week) as estimated by curbside waste composition analysis (WCA, x-axis) vs. self-administered survey (SAS, y-axis, N = 278) among the Columbus, Ohio sample..

Source: Authors’ calculations

Table 5 presents the LASSO regression coefficients for predicting total edible curbside food waste, comparing results from short and long models. Similar models are also estimated individually for each food category (see supplemental materials, Table A.3). These two models incorporate different predictors to evaluate the effectiveness of household characteristics in estimating curbside food waste collected.

Table 5.

LASSO Coefficients for Selected Predictors of Total Edible Curbside Waste.

Predictors	LASSO Estimates
	Short Model	Long Model
Self-report Waste Per Person	0.813	0.815
Household Size >= 3	12.112	54.890
Black	320.519	359.206
Age (35–44)	−25.541	−77.613
Age (45–54)		15.276
Age (55–64)	90.086	197.335
Number of Waste Categories reported		10.483
Any Unexpected Issue		219.365
Hispanic	170.214	285.075
Household Income (below 50 k)		58.561
Home Ownership (single house no attach)		58.711
Education (bachelor or above)	−129.118	−114.249
Grocery Shopping Frequency (once a week)		115.218
Grocery Shopping Frequency (2 to 3 times a month)		−68.622
Constant	489.303	97.254
R2	0.372	0.412

Notes: Estimated based on a randomly selected 85 % of Columbus sample.

The coefficient for self-reported food waste per person remains consistent across both models (0.813 in the short model, 0.815 in the long model), suggesting that self-reported waste is a strong and stable predictor of actual curbside edible food waste. Despite potential biases in self-reporting, the relatively high coefficient confirms that individuals’ estimates contain useful predictive information. However, self-reported data is not the only significant predictor, suggesting predictive improvements are possible from incorporating other information gathered via SAS.

Household size (>=3 members) is a significant predictor in both the short and long models, as are categorical indicators for race, ethnicity, and age. Identifying as Black or Hispanic is a predictor of greater curbside waste generation as is being an older respondent (aged 55–64). Those with a bachelor’s degree or above and those in a younger age group (25–44) are negatively associated with curbside waste generation. Lower-income households (<$50 k) are associated with slightly higher curbside waste generation (58.561 in the long model, absent in the short model), though this association is less than other demographic factors.

Predictors that only appear in the long model draw from a broader array of SAS questions that might not be included in shorter-form surveys. Households that: report unexpected issues during the week, shop for groceries once a week or own their home tend to generate more waste at curbside, while households that shop 2–3 times per month tend to generate less. Other significant predictors in the long model are drawn from variables also available for the short model. This includes having annual incomes less than $50,000, reporting more total waste categories on the survey, and being aged 45–54.

Table A.4 presents the results from the LASSO regression for predicting total inedible curbside food waste, utilizing both short and long models. The variable set used in the LASSO models remains consistent with the model used for edible food waste prediction. In line with the findings for edible food waste, the strongest predictor in both models is self-reported edible waste per person, with coefficients of 0.803 and 0.812 in the short and long models. The finding reinforces that self-reported edible food waste can be a useful predictor of curbside inedible waste, though we note that the goodness of fit for models predicting inedible curbside WCA are considerably lower than those for edible waste (e.g., an R² of 0.41 for edible waste vs. 0.28 for inedible).

3.4. Prediction Accuracy: LASSO versus simple adjustment

In this section we examine the accuracy of predictions of the LASSO models estimated in the previous section against the standard adjustment calibration described in the methods section, which in this case is 0.680 (mean SAS/mean WCA from the estimation sample). Fig. 2 compares the predictive accuracy of three approaches in the hold-out sample (the 15 % of the sample observations not used in model estimation) for the short-form LASSO model, long-form LASSO model, and standard adjustment method, while Fig. S3 contrasts the WCA values for the holdout sample against the predicted values from the standard adjustment and long LASSO model.

Fig. 2.

Predictive Performance: LASSO Models vs. Simple Adjustment Method Notes: Models for edible waste are compared for the Columbus, Ohio participants. MSE – mean squared error (reported in 1000′s); RMSE – root mean square error. Figures at the top of the first two clusters of columns are the percent difference in MSE and RMSE from the Standard Adjustment’s values. Figures at the top of the third cluster of columns are the deviation of the waste composition analysis (WCA) mean predicted by the three approaches from the observed WCA mean. Based on hold-out sample of 15 % of the Columbus sample (N = 42)..

Source: Authors’ calculations

When it comes to predicting the mean of this hold-out sample, the standard adjustment yields an estimate that is 2 % greater than the observed value, whereas the short and long LASSO models yield estimates that are 7 % and 3 % greater than the observed value. When considering core metrics of predictive capacity, the long LASSO model, which incorporates both demographic and behavioral variables, yields the lowest mean squared error (MSE = 386,138.60) and root mean squared error (RMSE = 593.82), outperforming both the short LASSO model and the standard adjustment. The standard adjustment produces the highest prediction error (RMSE = 673.6, 12 % greater than the long LASSO and 9 % greater than short LASSO). These results underscore the benefits of machine learning-based prediction over the standard adjustment as well as the predictive benefits of including food-related behavioral factors that augment the long LASSO model, suggesting that richer survey data can improve curbside WCA waste estimates.

3.5. Implementation costs by method

A critical motivation for developing prediction models is to leverage machine learning to help mitigate the inherent tension between measurement quality and cost. In this section, we briefly outline key costs expended by the research team to undertake the three data collection approaches used in this study (Table 6). The detailed costs cover expenditures on labor or commercial services deployed to collect the sample data plus the costs of incentives provided to participants to encourage cooperation. We omit costs related to study initiation (e.g., obtaining IRB permission, developing research protocol and instruments), general project administration, and data analysis. These were the costs faced by the current research team during 2023–2024 in Columbus, Ohio.

Table 6.

Costs of Data Collection by Method.

Curbside Waste Composition Analysis
Tasks undertaken	Amount	Unit
pre-audit prep	300	$/day
collection of physical samples	2400	$/day
physical sort of collected samples	1100	$/day
consolidate data for transmission	250	$/day
Total/day	4050	$/day
Average samples per day	12	samples
Average cost per sample collected	337.5	$/sample
Participant incentives/sample*	37.5	$/sample
Cost/sample estimate	375	$/sample
Self-administered Surveys – Columbus Panel
Participant incentives/survey*	37.5	$/survey
Cost/survey	37.5	$/survey
As a % of Curbside	10.0	%
Self-administered Surveys — National Panel
Participant Incentives	3.4	$/survey
Recruitment platform service fee	1.02	$/survey
Cost/survey	4.42	$/survey
As a % of Curbside	1.2	%

Notes:

^*Columbus participants were paid $150 for the entire study, which involved 2 weeks of surveys and 2 weeks of setting aside physical waste for pick up. For the purposes of this calculation, this was divided equally across weeks and methods yield 150/4=$37.50.

The costs of curbside WCA are 10 times those of SAS for the Columbus panel and 85 times that for the National panel SAS. The major difference between SAS and WCA costs is driven by the need to rent vehicles and hire people to collect waste samples, transport materials to a sorting location, and sort materials. SAS replaces the hired labor with the effort of the participant, who is compensated with a modest participation fee.

Participation fees for those already part of a consumer panel are more than eight-fold less ($4.42/survey vs. $37.50) than for those who must be recruited for the location-specific study and convinced to participate. While participation fees could feasibly be lowered for local study recruitment, one must worry about attracting adequate participation from a locality; hence higher participation fees are warranted from the point of view of mitigating the risk of the study failing for lack of sufficient participation. In contrast, participation response rates among members of national consumer panels are well understood, and the national scope permits for a broader pool of possible participants and the ability to keep the survey open longer to attract sufficient participation for a given incentive.

We conclude this section by noting some limitations of the study. Because both SAS and WCA involve researcher interactions with study participants, behavioral reactivity may influence the resulting measures. Future studies that validate SAS and WCA estimates against measures insulated from behavioral reactivity (e.g., route-level audits) could calibrate the severity of any bias. While WCA estimates from on-site waste sorters were validated by detailed inspection of waste bin photos by trained raters, further efforts could be exerted to increase the precision of WCA estimates through, for example, more frequent capture of curbside waste from participating households. Another limitation is that estimated LASSO models and their prediction variables are based upon data obtained for a single region over a relatively short time. While the study design featured data collection in a second city (Baton Rouge, LA), inconsistencies in the curbside collection of waste in this location rendered these data unsuitable for a parallel analysis. Hence, the predictive capacity of these models may decline for regions further from Ohio and as time passes and be low for different contexts or for longer monitoring periods. These critiques also apply to the assessment of costs of methods. We also re-emphasize that LASSO is a predictive model, and does not imply any causal relationships between predictors and waste generation, which is left for future work. We also note that, while other machine learning approaches were assessed and deemed inferior, new approaches are emerging in machine learning literature and should be assessed in future work for their relative predictive capacity in this setting.

4. Conclusions and Perspectives

We document that estimates of household food waste generation differ between measurement approaches applied to the same sample of participants (SAS vs. WCA) and between the same measurement approach applied to differently recruited samples (National vs. Columbus SAS responses). The magnitudes of the estimates of per-person once-edible food waste follow the same order as the relative costs of collecting each type of data with the order from low to high being SAS from a national consumer panel (388 g/person/week, $4.42/sample), SAS from locally-recruited residents (552 g/person/week, $37.50/sample), and curbside WCA from locally-recruited residents (809 g/person/week, $375/sample).

Given our maintained assumption (derived from the literature) that surveys under-report WCA, this implies that there is no method among these three that is both less biased and less costly than any other approach. In other words, if these were the only three options, the study coordinator would expect to pay more to decrease bias further.

However, our results suggest several practical insights. For example, if evaluation is to focus on certain subcategories of food, surveys may provide similar performance to more expensive curbside audits. For our Columbus sample featuring 278 weekly observations from 146 households, there was little bias when deploying SAS rather than WCA for certain food subcategories (e.g., produce, dairy and eggs). In fact, for the beverages category, there was a downward bias induced by deploying WCA rather than SAS. Such a bias may not be surprising given that discarded beverages may be disposed of in the sink and never make it to the curbside for collection and analysis or, if discarded beverages are included in a household’s general trash collection, the liquid component may be difficult to quantify using curbside waste measurement protocol.

The current study goes beyond simply documenting the difference in estimated food waste generation captured by different methods. We leverage LASSO, a machine learning approach, to develop models that use information from a household’s SAS to predict the amount of food waste from the same household estimated via WCA of the curbside waste stream collected the same week as covered by the survey. We compare the predictive capacity of these models to the standard adjustment approach used to offset survey under-reporting, i.e., where each value derived via SAS is multiplied by the mean level of under-reporting for a sample where both measures are available. The LASSO models generate predictions for a hold-out sample that are substantially more accurate than the simple adjustment approach.

The improvement in predictive performance is primarily driven by self-reported waste per person, which emerges as the strongest and most consistent predictor across model specifications. This finding highlights that survey-based measures, while subject to underreporting, still contain strong signals about real waste levels. In addition, household characteristics (e.g., household size), behavioral factors (e.g., grocery shopping frequency), and demographic variables further enhance predictive accuracy, indicating that food waste generation is shaped by a combination of consumption patterns and household context.

LASSO models that incorporate more information about the household, as might be possible if longer surveys were conducted, feature stronger prediction than a model with fewer characteristics, though the incremental improvement diminishes. In other words, the biggest gain in predictive capacity comes from switching from the standard adjustment to a short LASSO model rather than from the short to the long LASSO model.

We also document that the estimates of food waste available from SAS differ according to the way participants are recruited. Those recruited locally in Columbus via a bespoke recruitment process and given a higher financial incentive reported 42 % more edible waste than those recruited via a National consumer panel once we control for incidental differences in household characteristics. This suggests a pathway to a scalable solution for estimating household food waste when composition analysis is infeasible due to logistical or cost considerations. A pathway for the prediction of curbside WCA based upon SAS data gathered from national consumer panels would involve (A) adjusting for local conditions before applying the predictive LASSO model using other variables collected on the SAS and (B) ensuring the SAS collects all the predictors in the long LASSO model. The exact level of local adjustment to comply with (A) and the exact LASSO predictors and model deployed in recommendation (B) will require additional study as local conditions are likely to be distinct from those captured in the current study. In terms of local adjustment factors as part of (A), we were able to undertake a similar analysis comparing our national sample’s SAS data to SAS data collected in Baton Rouge, Louisiana (Table A5, supplemental materials) where we found the locally recruited SAS estimates to be 21 % greater than the national SAS estimates, which was not significantly different once other explanatory factors are controlled (see supplemental materials).

Our assessment of the relative costs of measurement approaches also provides practical insights for those planning evaluations. We document that, for the current study, the cost of conducting a curbside WCA is 10 times greater than collecting SAS from the same locally recruited study participants. Cost-constrained teams planning evaluation efforts may simply be limited to surveys in order to be able to gather data from a sufficient number of households for meaningful evaluation. In these settings, an available model to predict curbside WCA levels could be essential and preferable to using standard correction factors from the literature, which don’t account for characteristics of the sample population. Evaluation teams with more accommodating budgets may still prefer to allocate budgets towards SAS rather than WCA. According to this study’s particular cost structure, this would permit the collection of data from 10 times as many households in the same location on the same budget and would provide the opportunity to collect other forms of data only available from a survey (e.g., attitudes and behavioral reactions to interventions). When coupled with a model that can predict curbside WCA based on survey responses, this option may be quite appealing to evaluation teams.

Appendix A. Supplementary data

Supplementary data to this article can be found online at https://doi.org/10.1016/j.wasman.2026.115727.

Data availability

The data that has been used is confidential.