2.1. Data

We use survey data obtained from the Meat Demand Monitor (MDM) project administered by K-State Research and Extension and jointly funded by the U.S. Beef and Pork Checkoff programs. The MDM is a national, online consumer survey that is intended to measure domestic meat demand in both retail and foodservice settings (Tonsor, Reference Tonsor2020). It is distributed every month by the market research company Dynata to a new sample of the population (i.e., the data takes a pooled cross-sectional structure) that is designed to be representative in terms of sex, age, income, educational attainment, race, and geographic region. Survey respondents receive redeemable “points” through the Dynata rewards system upon completion of the survey to incentivize participation.

A prior day consumption recall embedded in the MDM survey is key to this research. Participants are first asked to recall where they consumed breakfast, lunch, and dinner the day before. Available options are “At Home,” “Away from Home,” and “Neither” (i.e., skipped meal). This last option is especially pertinent in our application to GLP-1 use (and more broadly, weight management) and has not been formally considered in prior, related work assessing food outlet choices. If a participant reports consuming a meal at home, they are then asked if the ingredients for that meal were purchased in person from a grocery store or other physical retailer (FAH Retail), purchased online and delivered to the home (FAH Online Retail), or purchased from a meal preparation company or other source (FAH Other). If a meal was consumed away from home, participants report whether it was consumed at a fine dining restaurant (FAFH Fine), a casual dining restaurant (FAFH Casual), a fast casual or quick service restaurant (FAFH Quick), or in some other setting (FAFH Other). Thus, participants have three meal decisions (i.e., breakfast, lunch, and dinner) and eight location alternatives across FAH, FAFH, and skipping the meal.

In a similar fashion, participants are asked which aggregate protein sources were included in each of their prior day (non-skipped) meals with available options being beef, chicken, pork, seafood, alternative proteins (including eggs), and no protein. Importantly, this prior day consumption recall measures the inclusion of protein in a meal but not the volume. Thus, this study does not speak to consumption quantity, but rather frequency.

Last, participants report whether they are currently taking a GLP-1 medication and, if so, how long they have been taking it. Industry reporting suggests that food purchasing behavior may evolve over the course of GLP-1 use (Circana, 2025b; Ulie, Reference Ulie2025). While our data do not allow us to observe within-person changes over time, reported duration of use provides a means of comparing meal choices across GLP-1 users at different stages of use. As such, duration is used to define cross-sectional groups rather than to track individual behavioral changes. Importantly, these questions related to GLP-1 use in the MDM appear toward the end of the survey instrument and after all prior day meal location and protein inclusion questions. Thus, concerns over possible framing effects and bias in the prior day recall related to the presentation of GLP-1 information is mitigated compared to more common stand-alone or one-time survey assessments focused heavily on GLP-1 use.

A total of 51,557 participants completed the MDM survey between January and December 2025. We restrict our analysis to 30,708 participants that are over the age of 18, who pass two embedded attentiveness checks, and who provide complete prior day recalls and sociodemographic information. Table 1 describes the sample and their meal choices. Importantly, 12.1% of our usable sample is currently using a GLP-1 medication, which is consistent with growing prevalence reported by Hristakeva et al. (Reference Hristakeva, Liaukonyte and Feler2026) and a 12.4% usage rate among U.S. adults through October 2025 (Witters and James, Reference Witters and James2025).

Table 1.

MDM participant and meal characteristic relative frequencies

Table 1. Long description

The table presents data on 30,708 participants over the age of 18, detailing their meal choices and sociodemographic information. It includes columns for overall data, non-users, GLP-1 users, and GLP-1 use by duration (1-6 months, 7-12 months, over 12 months). The table is divided into several sections: GLP-1 use, sex, age, annual household income, household size, meal location, and prior day meal protein inclusion. Each section lists relative frequencies for different categories. For example, under GLP-1 use, the table shows 87.9% non-users and 12.1% GLP-1 users. The sex section shows 51.6% females and 48.4% males overall. Age groups range from 18-34 years to 65 years or over, with varying percentages. Annual household income is categorized into less than $40,000, $40,000-$99,999, and $100,000 or greater. Household size includes 1 person, 2 people, and 3 or more people. Meal locations are divided into categories like FAH retail, FAH online retail, FAH other, FAFH fine, FAFH casual, FAFH quick, FAFH other, skipped meal, and prior day meal protein inclusion, which includes beef, chicken, pork, seafood, alternative proteins, and none. The table provides a comprehensive overview of the participants’ characteristics and meal choices.

Note: Asterisks (*) in columns 3 through 6 indicate that the estimate is statistically different from “Non-users” (column 2) at the five percent level or lower using two-proportions z-tests. We use a Bonferroni correction for 26 between-group comparisons. Shares of meals including each protein source will not sum to one because multiple proteins can be consumed in a single meal.

We see stark differences in the sociodemographic composition of the GLP-1 user (column 3) and non-user (column 2) groups. Relative to non-users, GLP-1 users are more frequently male (52%), under the age of 50 (58%), earning at least $100,000 in annual household income (29%), and belonging to households of at least three people (49%). Users’ location of meal consumption is also distinctly different from that of non-users’. Importantly, 41% of meals eaten by GLP-1 users are at home and using ingredients purchased in FAH Retail, which compares to 58% of non-users’ meals. Correspondingly, users’ meal location frequency is higher for FAH Online Retail, FAFH Fine, FAFH Casual, and, to a lesser extent, FAFH Quick, relative to non-users’. Additionally, GLP-1 users report skipping a meal at a rate about six percentage points higher than non-users (roughly 21% versus 14%). These differences in meal location extend to protein consumption, with GLP-1 users reporting higher rates of including each protein source in their meals and a lower rate of not consuming protein in a meal by around 11 percentage points relative to non-users.

We also observe differences in sociodemographic characteristics within the GLP-1 user group itself. For instance, individuals who have taken the medications for more than twelve months are more frequently female (57%), while newer users (one to six months) are more frequently male (56%). Similarly, individuals who have used GLP-1 medications for more than twelve months are more frequently over the age of 50 (approximately 73%) and substantially more likely to have annual household incomes below $40,000 (34%), while those who have taken the medications for between one and six months are more frequently under the age of 50 (about 73%) and with annual household incomes above $100,000 (30%).

These differences in characteristics of shorter- and longer-term GLP-1 users may drive consumption decisions. For example, longer-term users (i.e., more than twelve months) more frequently report consuming their meals at home with ingredients purchased in FAH Retail, while shorter-term users (i.e., one to six months) report higher rates of eating from FAH Online Retail, in FAFH Fine, in FAFH Casual, in FAFH Quick, and skipping meals. Shorter-term users also report consuming each protein source in their meals at a relatively higher rate. For instance, individuals who have used GLP-1 medications for one to six months consume beef in 31% of their meals, on average, while individuals who have used the medications for more than twelve months consume beef in 25% of their meals. Overall, longer-term users tend to more closely resemble non-users across observable characteristics than shorter-term users, with the primary exception being age.

Although the MDM survey is designed to approximate U.S. sociodemographic proportions, the resulting sample (as reflected in Table 1) does not perfectly mirror the national population. Importantly, the observed share and composition of GLP-1 users in the sample should not be interpreted as nationally representative prevalence estimates. However, this study does not seek to estimate population-level GLP-1 usage. Rather, the objective is to obtain reliable econometric estimates of differences in meal location and protein inclusion choices across GLP-1 use groups. Further, for weighting purposes, nationally representative sociodemographic profiles of GLP-1 users have not yet been established in related literature, necessitating a focus on within-sample comparisons rather than population inference.

2.2. Choice model

Consistent with prior food outlet choice and meat demand research (Arnold et al., Reference Arnold, Oum and Tigert1983; Dong and Stewart, Reference Dong and Stewart2012; Kyureghian et al., Reference Kyureghian, Nayga and Bhattacharya2013; Lusk and Tonsor, Reference Lusk and Tonsor2016), we analyze consumer choices using a random utility framework (McFadden, Reference McFadden1974). For individual i, the deterministic portion of utility of consuming meal k purchased at outlet j is expressed as:

(1) $\eqalign{{{V_{ijk}}} =\; & ({\alpha _j} + {\sigma _j}{d_{ij}}) + {\gamma _j}GL{P_i} \cr & + {\sum\limits_{p = 1}^5 {{\pi _{jp}}} Protei{n_{ikp}} + \sum\limits_{p = 1}^5 {{\tau _{jp}}} (Protei{n_{ikp}} \times GL{P_i})} \cr& + {\sum\limits_{m = 1}^2 {{\delta _{jm}}} Mea{l_{ikm}} + \sum\limits_{w = 1}^6 {{\rho _{jw}}} WkDa{y_{iw}} + \sum\limits_{l = 1}^L {{\theta _{jl}}} {Z_{il}};{\mkern 2mu} {d_{ij}}\,\sim N(0{{,}}1),}\cr}$

where α _j and σ _j are the mean and standard deviation, respectively, of the marginal utility of consuming a meal in FAH Retail, FAH Online Retail, FAH Other, FAFH Fine, FAFH Casual, FAFH Quick, and FAFH Other (with skipping a meal normalized to zero for identification purposes), and d _ij is the standard normal deviate. These collectively form alternative-specific constants (ASCs) that capture preferences for each location arising from all factors not explicitly incorporated into the model (e.g., pre-treatment health focus, convenience, etc.). ASCs are normally distributed to allow outlet-specific marginal utilities to be either negative or positive relative to the normalized alternative, similar to Taylor and Villas-Boas (Reference Taylor and Villas-Boas2016).

GLP _i is a binary indicator equal to one if an individual is currently using a GLP-1 medication and zero otherwise. In this application, GLP _i is included as a segmentation variable that allows preferences for meal location to differ between GLP-1 users and non-users through the parameters γ _j .Footnote ¹ Importantly, we do not interpret GLP _i as a causal determinant of food outlet choice. GLP-1 use is a self-selected behavior that is likely correlated with underlying factors such as health concerns, income, and GLP-1-related behavioral adjustments that also influence meal decisions. Accordingly, our objective is not to identify causal effects of GLP-1 use, but rather to document differences in meal location choices between users and non-users that are relevant for industry decision makers. The MDM data also precludes distinguishing GLP-1 use by purpose (i.e., weight management versus diabetes treatment), and thus the estimates reflect overall differences across users irrespective of use motivation.

The subscript p indexes beef, chicken, pork, seafood, and alternative proteins and Protein _ikp is then a vector of binary variables that are equal to one if individual i consumes protein source p in meal k and equal to zero otherwise. This specification allows us to document how preferences for protein types vary both within and across food outlets, though we do not imply that protein inclusion itself determines outlet choice. Further, the MDM grouping of prior day protein inclusion by animal source maintains variation in nutritional and behavioral distinctions as consumers have views on protein quality and healthfulness that vary across animal species (Font-i Furnols and Guerrero, Reference Font-i Furnols and Guerrero2014; Szendrő et al., Reference Szendrő, Dalle Zotte, Fülöp, Garamvölgyi and Tóth2024). Importantly, protein inclusion in a meal impacts utility through the parameters π _jp for non-users and through the parameters π _jp and τ _jp for users (via protein inclusion and GLP-1 use interaction terms), allowing preferences to vary between groups. Further, the vector of binary protein inclusion variables, Protein _ikp , are overlapping in the sense that MDM participants can report consuming multiple protein items in a single meal. Thus, the estimates of π _jp and τ _jp are interpreted relative to not consuming any protein source.

Meal _m is a vector of binary variables indicating if a meal was lunch or dinner (with breakfast being the default, dropped level). Meal occasion is fundamental in determining where to consume a meal (Marshall and Bell, Reference Marshall and Bell2003) and should be controlled for to ensure that outlet marginal utilities do not reflect within-day temporal confounding. These binary indicators influence utility through the parameters δ _jm . Similarly, WkDay _w is a vector of binary variables indicating the weekday on which the survey response was recorded (with Saturday as the omitted base category). These variables control for differences in consumption patterns arising from work schedules and potentially greater flexibility on weekends. Survey participation varies moderately across weekdays, with the lowest share of responses observed on Sunday (11.5%) and highest share on Wednesday (17.8%).

Z _il is then a vector of characteristics of the individual including age, sex, income, and household size that influence utility through the parameters θ _jl . These are correlated with GLP-1 use (see Table 1) and are likely highly predictive of consumers’ meal location choices. Since these consumer traits are observed in the MDM data, we viewed it as appropriate to explicitly control for them.

Equation (1) represents a random parameters logit (RPL) model with outlet-specific constants that are normally and independently distributed. We note that prior assessments of store choice have controlled for a more comprehensive suite of consumer characteristics (e.g., race, vehicle availability). Those efforts have either used reduced form approaches (Dilley et al., Reference Dilley, Adhikari, Silwal, L. Lusk and McFadden2025; Kyureghian et al., Reference Kyureghian, Nayga and Bhattacharya2013), multinomial logit (MNL) modeling (Briesch et al., Reference Briesch, Chintagunta and Fox2009; Dong and Stewart, Reference Dong and Stewart2012; Kyureghian and Nayga, Reference Kyureghian and Nayga2013), or linearized MNL modeling (Taylor and Villas-Boas, Reference Taylor and Villas-Boas2016). The discrete nature of MDM prior day meal reporting and our desire to allow for random taste variation (requiring numerous random coefficients and increased computational demands) make our empirical framework most appropriate but require limiting the included participant characteristic variables to maintain an estimable choice model. Finally, in accordance with our final stated objective, we consider that preferences for meal location and protein may change over the course of GLP-1 use (Circana, 2025b; Ulie, Reference Ulie2025). Using MDM data, we do not observe an individual’s meal choices over an extended period. However, we do observe meal choices made across GLP-1 users that vary in how long they have taken the medications. Thus, we alter Equation (1) by replacing γ _j GLP _i with $\sum _{d=1}^3\gamma _{jd}Duration_{id}$ , where Duration _id is a vector of binary variables that indicate if individual i has used GLP-1 medications for between one and six months, between seven and twelve months, or for over one year (with being a non-user the default, dropped level).Footnote ² This allows the utility obtained from consuming a meal in a specific food outlet to vary across different durations of GLP-1 use. We also replace GLP _i with Duration _id in the protein inclusion and GLP-1 use interaction terms to allow the utility received from the respective protein sources to likewise vary across duration of GLP-1 use.

Estimation of the RPL model is conducted using simulated maximum likelihood (Train, Reference Train2009) and via the Apollo package in R version 4.4.1 (Hess and Palma, Reference Hess and Palma2019). We use 750 individual-specific Modified Latin Hypercube Sampling draws to avoid correlation patterns that arise in Halton sequences in higher-dimension integration [such as this study that considers seven random parameters] (Hess et al., Reference Hess, Train and Polak2006). These draws have been shown to outperform various Halton sequences in up to a 16-dimensional RPL design and at a number of draws as few as 50. Starting values of parameters are obtained from first-stage multinomial logit (MNL) modeling to aid in convergence.

2.3. Economic measures

Because choice probabilities from an RPL model cannot be computed in closed form, we estimate them using simulation following Train (Reference Train2009). The procedure is as follows: i) draw values of β from the estimated parameter distribution f(β∣θ), where β denotes the parameters and θ denotes the estimated distributional moments; ii) calculate the logit probability using each draw [see Train (Reference Train2009) for formulas]; iii) repeat this process many times; and iv) average the resulting probabilities across draws. In this study, we use 1,000 draws from the standard normal distribution for d _ij in Equation (1). For each consumer group, data are held at group means, choice probabilities are simulated separately, and group differences are then computed from those simulated values.

Marginal utilities are constructed in a similar way. We again use 1,000 simulation draws to average over heterogeneity in preferences and obtain the ASCs. In this context, ASC _j represents a non-user’s marginal utility of consuming a meal from outlet j without any protein included, relative to skipping the meal. We then form linear combinations of the ASCs, GLP-1 use parameters, and protein parameters to obtain the marginal utility of meals that include a specific protein source. For example, ASC _j + γ _j + π _jp + τ _jp is the marginal utility of a GLP-1 user consuming a meal from outlet j that includes protein source p, relative to skipping the meal. Higher values of π _jp + τ _jp imply that outlet j becomes more attractive when protein source p is included for a user. If π _j1 + τ _j1 > π _j2 + τ _j2, then protein source 1 contributes more to utility in outlet j than protein source 2, and is therefore more preferred within that outlet for a GLP-1 user.

To construct these utility measures, data are again held at group means. In doing so, all observed covariates (including meal occasion, weekday indicators, and participant characteristics) enter utility through the ASCs as shifters evaluated at their mean values within each consumer group. The protein indicator of interest is then set equal to one while the remaining protein indicators are set to zero. This approach allows us to isolate the contribution of each protein source to utility and to calculate marginal utilities separately by consumer group and protein type.

2.4. Variance estimation

In RPL models, uncertainty arises from both parameter estimation and individual preference heterogeneity. To account for both sources of variation, we follow Hensher and Greene (Reference Hensher and Greene2003), Rigby and Burton (Reference Rigby and Burton2005), and Tonsor et al. (Reference Tonsor, Schroeder, Pennings and Mintert2009) using a two-stage simulation procedure.

First, we draw 1,000 vectors of parameter estimates using a Krinsky and Robb (Reference Krinsky and Robb1986) resampling procedure. Second, for each parameter draw, we compute the economic measures described in the previous subsection using simulation over heterogeneous preferences. Specifically, we take 1,000 draws of the random parameters [based on α _j and σ _j from Equation (1)] and use these to calculate choice probabilities, marginal utilities, and between-group differences.

This procedure yields one set of simulated economic measures for each Krinsky-Robb draw. Collectively, these form an empirical distribution of 1,000 values for each measure. The reported point estimates are the means of these distributions, and the associated 95% confidence intervals are constructed from their empirical percentiles. These results are reported throughout the remainder of the study.

2.5. Estimation considerations and sensitivity analysis

Two important estimation considerations are made: i) omitted outlet-level price data and ii) household and store location endogeneity. First, prices are an important consideration when determining both where to purchase a meal and what type of protein to include in it, if any. Taylor and Villas-Boas (Reference Taylor and Villas-Boas2016) justify missing price data by noting that as long as outlet type j always has higher (or lower) prices than outlet type k, then time-invariant differences in prices will be captured by the outlet-specific constants.

In our application, as long as GLP-1 user and non-user groups experience the same relative prices between outlets (e.g., fine dining prices are always higher than physical retail prices for both groups), then the relative ordering of utilities between groups is unaffected by missing prices. Within FAH decision making, we can reasonably assume that purchasing ingredients in physical retail is less expensive than in online retail (which usually entails shipping costs), and purchasing ingredients in online retail is less expensive than in the “FAH Other” category that includes meal preparation purchases (which entail costs associated with shipping and value-added activities). We can also reasonably assume that FAFH purchases are more expensive than FAH and, within FAFH decision making, fine dining is more expensive than casual dining, and casual dining is more expensive than fast casual and quick service dining.

Further, regarding omitted protein prices, which are unobserved by the researchers given the nature of the MDM meal reporting, to the extent that GLP-1 user and non-user groups experience the same relative prices between protein sources in outlet j (e.g., they both experience higher prices for beef products than pork products in fine dining), then our conclusions regarding relative ordering of utilities between groups likewise do not depend on prices of the respective protein sources. We have no reason to believe that relative prices between protein sources within a specific outlet would differ systematically across GLP-1 users and non-users. However, estimated utility measures should be interpreted as reflecting the joint price and non-price attributes associated with each protein source.

Second, the food store choice literature has widely acknowledged the endogeneity of household and store locations (Cuffey and Beatty, Reference Cuffey and Beatty2022; Currie et al., Reference Currie, DellaVigna, Moretti and Pathania2010; Kyureghian et al., Reference Kyureghian, Nayga and Bhattacharya2013; Taylor and Villas-Boas, Reference Taylor and Villas-Boas2016). Food retailers and restaurants make decisions on where to locate stores based on population characteristics, while households make decisions on where to live based on the available food amenities. Kyureghian and Nayga (Reference Kyureghian and Nayga2013) note the challenge in finding valid instruments and elect instead to use lagged values of the retail environment. Currie et al. (Reference Currie, DellaVigna, Moretti and Pathania2010) take a different approach, leveraging detailed geographic and household data and finding no evidence of endogenous store placement when examining small distances. Subsequent research draws from their efforts by controlling for a large variety of household and food environment characteristics (Taylor and Villas-Boas, Reference Taylor and Villas-Boas2016) or assuming that, within a small geographic space, store placement decisions are due to factors uncorrelated with the characteristics of nearby households (Cuffey and Beatty, Reference Cuffey and Beatty2022).

Unlike data used in prior, related studies, MDM survey data does not include information on participants’ food shopping environments. Additionally, U.S. Census Bureau County Business Patterns data utilized by Kyureghian et al. (Reference Kyureghian, Nayga and Bhattacharya2013) and Kyureghian and Nayga (Reference Kyureghian and Nayga2013) was not available for the 2024–2025 period at the time of writing. To address concerns of household and store location endogeneity, we consider two alternatives. First, we follow related work by including county-level population density as a proxy for physical distance between households and stores in Equation (1) (Dong and Stewart, Reference Dong and Stewart2012; Taylor and Villas-Boas, Reference Taylor and Villas-Boas2016; Volpe et al., Reference Volpe, Okrent and Leibtag2013). MDM participants’ self-reported ZIP codes are mapped to counties using U.S. Department of Housing and Urban Development (2025) crosswalk files. Then, county-level population densities are constructed by dividing U.S. Census Bureau (2025) estimates for population by land area (in square miles).

We also consider that participants may live in either urban or rural areas and have differing access to certain food outlets (Kyureghian et al., Reference Kyureghian, Nayga and Bhattacharya2013; Taylor and Villas-Boas, Reference Taylor and Villas-Boas2016). Specifically, we use Rural-Urban Continuum Codes from the U.S. Department of Agriculture Economic Research Service (2025) to identify which participants live in counties in metropolitan areas of at least 1 million people. We then estimate Equation (1) using only survey data obtained from these participants. Here, we argue that i) individuals living in metropolitan areas have ample access to all food store types (in terms of distance from and density of outlets) and ii) firms have ample access to consumers regardless of store location decision. Thus, the potential for bias due to household and store location endogeneity is minimized.