Literature review

The term convenience has been used in the literature with different connotations, but there is consensus that convenience is associated with any aspect of the food that would enable saving physical and mental energy, as well as time, in grocery planning and shopping, meal preparation, consumption, and post-meal activities such as clean up (Scholderer and Grunert Reference Scholderer and Grunert2005; Scholliers Reference Scholliers2015). Of all those activities, meal preparation is the most time-intensive. Okrent and Kumcu (Reference Okrent and Kumcu2016) report that between 2003 and 2011, women in the United States spent on average 48 minutes on meal preparation (men spent 18 minutes on average); whereas, in 1920, rural women in the United States spent on average 122 minutes cooking and 68 minutes on meal clearing and clean up.

Several studies have analyzed consumer’s demand for ready meals. Capps, Tedford, and Havlicek (Reference Capps, Tedford and Havlicek1985) found that manufactured convenience foods (e.g., foods with no home-prepared counterparts) were more responsive to own-price changes compared to the nonconvenience foods (e.g., the fresh, unprocessed, or home-produced foods) and that basic convenience foods (e.g., food where processing was more related to the preservation method rather than ease of preparation) were more sensitive to own-price changes than complex convenience foods (e.g., multi-ingredient foods with high levels of time-saving and energy inputs). Verlegh and Candel (Reference Verlegh and Candel1999) found that the working status of the person responsible for preparing meals at home had a significant and positive relationship with the consumption of convenience meals. De Boer et al. (Reference De Boer, McCarthy, Cowan and Ryan2004) found that the importance of freshness negatively affected the purchase of ready meals and that increases in the perceived time pressure positively contributed to the purchase of ready meals. Harris and Shiptsova (Reference Harris and Shiptsova2007) found that households with increased disposable incomes, for whom the opportunity cost of time was higher, were positive to expenditures on ready meals. To sum, these studies concur that disposable time and income are positively associated with the purchases of convenience meals, whereas the notion of freshness negatively impacted its purchase.

Consumers’ preferences for nutritional content are typically associated with preferences for health-related aspects of consuming a food product. Specific to convenience foods, Binkley (Reference Binkley2006) and Burton, Howlett, and Tangari (Reference Burton, Howlett and Tangari2009) proved that nutritional content has little impact on the consumption of food away from home. This contrasts with general grocery store food products, like bread, for which nutritional content has a positive impact on consumers’ preferences (Ginon et al. Reference Ginon, Lohéac, Martin, Combris and Issanchou2009). There are no conclusive findings for ready meals. Geeroms, Verbeke, and Van Kenhove (Reference Geeroms, Verbeke and Van Kenhove2008) found that health-related statements do not have an impact on consumers’ preferences for ready meals. Remnant and Adams (Reference Remnant and Adams2015) found that ready meals exhibited high contents of saturated fat and salt and low sugars, compared to the nationally recommended UK front-of-pack labeling. Interestingly, they found that the cost of the meals was associated with higher contents of energy, fat, saturated fat, protein, and fiber, and not to healthier ingredients. Kanzler et al. (Reference Kanzler, Manschein, Lammer and Wagner2015) found similar results to those in Remnant and Adams (Reference Remnant and Adams2015); in that ready meals were nutritionally imbalanced, being high in fat content and low in carbohydrate levels, with protein content being above dietary recommendations. We extend the previous literature by analyzing the effects of nutritional content, specifically including those ingredients that have a negative health connotation, but that guarantee an appealable flavor on households’ demand for ready-to-heat convenience meals sold at grocery stores in the United States.

Data

This study uses the IRI InfoScan retail scanner data. IRI has agreements with retail outlets across the United States, to provide weekly retail sales data including prices and quantities for products with a Universal Product Code (UPC) and perishable products or random weights. Specifically, the data fields include the following: UPC, store ID for store-level or geography key for retailer marketing area, week, number of units sold, and total revenue in dollars and cents. The retail outlets included in the InfoScan data set include grocery stores (>33,000 stores), drug stores (both chain and independent with >42,000 stores), convenience stores with scanning capacity (chain and independent with > 150,000 stores), mass merchandisers, supercenters, traditional neighborhood markets, club stores, dollar stores, defense commissary stores, and exchanges (Muth et al. Reference Muth, Sweitzer, Brown, Capogrossi, Karns, Levin, Okrent, Siegel and Zhen2016).

In this study, the InfoScan data are linked to product attributes such as nutrition facts, brands, flavor, meal preparation time, product description, and net weight using the UPC. IRI obtains the scanned image of the nutrition fact label and then codes the information from the package adding it to the database. An example of a nutrition fact label is presented in Figure 1. Note that in some cases, the nutrition data are not complete. IRI codes nutrition data for products with significant sales volume, as the intention is to cover a large portion of the sales and not a large portion of UPCs (Muth et al. Reference Muth, Sweitzer, Brown, Capogrossi, Karns, Levin, Okrent, Siegel and Zhen2016).

Figure 1. Nutrition fact label for pepperoni pizza.

The nutrition information available for ready-to-heat meals includes the macronutrients sugar content, calories, total fat, cholesterol, fiber, and sodium. The reason for containing sugar and fiber rather than carbohydrates is to prevent perfect collinearity with the variable calories in the regression, which is an indicator of total energy. We acknowledge that sugar and fiber are listed under the carbohydrates in the nutritional label and that these two components are not the only source of carbohydrates. For fats, we also acknowledge that there are different types of fats (i.e., trans fats, monounsaturated and polyunsaturated fats), and each of them have different impacts on consumers’ health. As explained in the preceding paragraph, not all the information in the nutritional label is available for all UPC products, and the nutrition variables included in the study are the ones that are consistently present in the data set, for most ready-to-heat meals. This study does not include micronutrients among the variables used, because the amounts in which micronutrients are present in ready-to-heat meals is negligible.

About the units, the nutrient information is presented on a per-serving-size basis, and we convert these observations to a per ounce unit. The preparation time is available in the data set in minutes.

These data were collected between 2008 and 2017 across all 50 states (plus Washington DC and Puerto Rico) in the country.

Table 1 describes the data used in this study, that is, twenty ready-to-heat meal products with the largest market shares in the entire sample. Also, for each product, the table includes the meal preparation time in minutes, net weight in ounces, and nutritional content including sugar, calories, total fat, cholesterol, fiber, and sodium. The products include pasta, salads, side dishes, and pizza. We found enough variability in terms of the different attributes included, for example time preparation ranges from 2 to 10 minutes, net weight from 4.3 to 24 oz, sugar content from 0.12 to 2.67 g/oz, calories from 22.17 to 80 per oz, total fat 0.10 to 3.90 g/oz, cholesterol 1.33 to 25.80 mg/oz, fiber 0.14 to 1.26 g/oz, and sodium content 72.09 to 326.06 mg/oz.

Table 1. Summary of characteristics for each ready-to-heat meal

Source: InfoScan data, IRI.

Table 2 reports the average price in cents per ounce, the standard deviation of the prices, and the average market share of each product in the sample. The prices range from 12.69 to 38.89 cent/oz; the standard deviation for the product bundle prices ranges from 1.53 (pepperoni pizza) to 4.06 cent/oz (salad and pasta box); the average market share ranges from 0.60% (pepperoni pizza) to 2.94% (salad and pasta box). In sum, the products used in this study represent 34% of all ready-to-heat meal sales in the data set.

Table 2. Market share and average prices for each ready-to-heat meal

Source: InfoScan data, IRI.

The observations on the average price and total sale percentage (market share) of each product are included for 52 markets (i.e., each state in the United States plus DC and Puerto Rico) and 522 weeks (i.e., 10 years). In total, the regression sample includes a total of 542,880 observations for 20 ready-to-heat meal products most frequently bought with the largest market shares during 2008–2017 in the IRI grocery scanner data set.

This study uses the InfoScan data set for the empirical modeling and does not use the household-based scanner data, mainly because the ready-to-heat meals are not purchased on a frequent basis by a critical mass of households. Thus, there are insufficient observations to provide a stable market share for each time-market combination. Nonetheless, to provide complete information, this study includes a summary of the household-based scanner data set, which is different from the InfoScan data set, to compare the sociodemographic characteristics between households who purchase ready-to-heat meals, at least once during the period of analysis, and those households in the entire data set. The description of the household-based scanner data set is presented in Table 3. One observes that those who purchase ready-to-heat meals exhibit a larger percentage of white, higher educated individuals, are less likely to have both male and female household heads compared to the entire sample, have smaller households in terms of number of individuals, and report higher annual incomes.

Table 3. Description of household sociodemographic characteristics

Source: IRI household scanner data.

Empirical method

This study applies the BLP random coefficients logit model (Berry, Levinsohn, and Pakes Reference Berry, Levinsohn and Pakes1995; Nevo Reference Nevo2001; Zhang and Palma Reference Zhang and Palma2021) to estimate the demand for ready-to-heat meals. Consumer i’s utility of consuming product j on period t is given by,

(1) $${U_{ijt}} = {\alpha _i}\left( {{y_i} - {p_{jt}}} \right) + {\boldsymbol{x_{jt}}}{\boldsymbol{\beta _i}} + {\xi _{jt}} + {\varepsilon _{ijt}}$$

where ${y_i}$ is consumer i’s income, ${p_{jt}}$ is the observed price of product j in time-market combinationFootnote ² t. ${x_{jt}}$ , is a 1 $\times$ K dimensional vector and depicts the attributes of product j and includes the convenience variable expressed as preparation time and the nutrition variables: sugar content, calories, total fat, cholesterol, fiber, and sodium content. ${\alpha _i}$ is the consumer i’s marginal utility of income, ${\boldsymbol{\beta _i}}$ is a K $\times$ 1 dimensional vector and represents the consumer i’s marginal utility of product attributes, ${\xi _{jt}}$ captures the unobserved product-specific shock in each market, and ${\varepsilon _{ijt}}$ is the error term.

Equation 1 assumes both ${\alpha _i}$ and ${\beta _i}$ have a constant and a random component. Randomness stems from standard normal distributions, which are used to represent the heterogeneity of parameters. The parameters are composed by a mean and a variance-covariance matrix, following,

(2) $$\left[ {\matrix{ {{\alpha _i}} \cr {{\beta _i}} \cr } } \right] = \left[ {\matrix{ \alpha \cr \beta \cr } } \right] + {\boldsymbol{v_i}},\;{\boldsymbol{v_i}} \sim {P_v}\left( v \right)$$

where ${\boldsymbol{v_i}}$ is the K by 1 dimensional vector of taste parameters for consumer i. The distribution of ${\boldsymbol{v_i}}$ is denoted by ${P_v}\left( v \right)$ .

${\boldsymbol{x_{jt}}}\;$ denotes the attributes capturing heterogeneity and has two parts coefficients: random ${\boldsymbol{v_i}}$ and non-random ${\boldsymbol{\beta _0}}$ ; in other words, ${\boldsymbol{\beta _i}}$ consists of two vectors: ${\boldsymbol{\beta _0}}$ and ${\boldsymbol{v_i}}$ , which yields,

(3) \begin{equation}\matrix{ {{u_{ijt}}} & = \,\, {{\alpha _i}\left( {{y_i} - {p_{jt}}} \right) + {{\boldsymbol{x}}_{{\boldsymbol{jt}}}}{{\boldsymbol{\beta }}_{\boldsymbol{i}}} + {\xi _{jt}} + {\varepsilon _{ijt}}} \hfill \cr {} \,\,\,\,\,\,\,\,\, & = \,\, {{\alpha _i}{y_i} - {\alpha _i}{p_{jt}} + {{\boldsymbol{x}}_{{\boldsymbol{jt}}}}{{\boldsymbol{\beta }}_{\boldsymbol{0}}} + {{\boldsymbol{x}}_{{\boldsymbol{jt}}}}{{\boldsymbol{v}}_{\boldsymbol{i}}} + {\xi _{jt}} + {\varepsilon _{ijt}}} \hfill \cr {} \,\,\,\,\,\,\,\,\, & = \,\, {{\alpha _i}{y_i} + \left( { - {\alpha _0}{p_{jt}} + {{\boldsymbol{x}}_{{\boldsymbol{jt}}}}{{\boldsymbol{\beta }}_{\boldsymbol{0}}} + {\xi _{jt}}} \right) + {{\boldsymbol{x}}_{{\boldsymbol{jt}}}}{{\boldsymbol{v}}_{\boldsymbol{i}}} + {\varepsilon _{ijt}}} \hfill \cr } \end{equation}

The expression above is composed of the utility from income, the mean utility from the product attributes, consumer heterogeneity, and the independently and identically distributed (iid) error term. The mean utility from the product attributes and consumer heterogeneity is defined by,

(4) $${\delta _{jt}} \equiv - {\alpha _0}{p_{jt}} + {\boldsymbol{x_{jt}}}{\boldsymbol{\beta _0}} + {\xi _{jt}}$$

(5) $${\mu _{ijt}} \equiv {\boldsymbol{x_{jt}}}{\boldsymbol{v_i}}$$

Then, the utility function can be expressed as,

(6) $${u_{ijt}} = {\alpha _i}{y_i} + {\delta _{jt}} + {\mu _{ijt}} + {\varepsilon _{ijt}}$$

where ${\delta _{jt}}$ is the mean utility from a consumer’s choice of product $j$ that is homogeneous across all consumers, ${\mu _{ijt}}$ is the heteroskedastic disturbance term that shows consumer heterogeneity, and ${\varepsilon _{ijt}}$ is the homoscedastic iid error term.

Each consumer purchases one product unit at a time that gives the highest utility compared to all others, including the outside product. Conditional on the product attributes $\left( {\boldsymbol{x,\;\xi}} \right)$ and market prices $p$ , a consumer i chooses product $j$ if,

(7) $${u_{ijt}} \ge {u_{ikt}}{\rm{\;for\;all\;}}j,k \in \left\{ {0,,1,2, \ldots ,{\rm{J}}} \right\}$$

Further, if ${q_{jt}}$ represents the quantity of the product $j$ sold in market $t$ , then the observed probability of a consumer i choosing the product $j$ over other products is given by,

(8) \begin{align} {\rm{Pr}}\left( {{u_{ijt}} \gt {u_{ikt}}} \right) & = {\rm{Pr}}\left( {{\alpha _i}{y_i} + {\delta _{ijt}} + {\mu _{ijt}} + {\varepsilon _{ijt}} \gt {\alpha _i}{y_i} + {\delta _{ikt}} + {\mu _{ikt}} + {\varepsilon _{ikt}}} \right)\\ & = {\rm{Pr}}\left( {{\varepsilon _{ikt}} - {\varepsilon _{ijt}} \lt {\delta _{ijt}} - {\delta _{ikt}} + {\mu _{ijt}} - {\mu _{ikt}}} \right)\\ & \matrix{{= \mathop \int \nolimits_\varepsilon I\left( {{\delta _{ijt}} - {\delta _{ikt}} + {\mu _{ijt}} - {\mu _{ikt}}} \right)f\left( {\varepsilon |{v_i}} \right)d\varepsilon \;} \cr { = {s_{ijt}}\;\;\;\;\forall \;j \ne k\# } \cr } \end{align}

Equation 8 is integrated over the density of unobserved preference to obtain the theoretical share of product $j$ in market $t$ , resulting in Equation 9,

(9) $$\matrix{ {{s_{jt}}\left( {p,\boldsymbol{x}} \right) = \mathop \int \nolimits_\nu {s_{ijt}}dP_v^*\left( v \right)\# \;} \cr } $$

The study uses the STATA BLP package to analytically estimate the coefficients using Monte Carlo integration. The package uses 200 draws for the Monte Carlo simulation, the tolerance level used to define the convergence of the contraction mapping algorithm is 10^e-15, and the starting value for the standard deviations of the random coefficients is 0.5 (Vincent Reference Vincent2015). Considering that price is endogenous to the market share, this study uses as an instrumental variable, the average price of the same product in other marketsFootnote ³ (Hausman Reference Hausman, Bresnahan and Gordon1996). The instrument was tested, and results of the first-stage F-test larger showed this was not a weak instrumental variable (Nevo Reference Nevo2001).

To capture the effect of convenience, expressed in terms of preparation time, and nutritional content, as the demand for ready-to-heat meals, product j is depicted by 20 ready-to-heat meal products j (j = 1,2…,20). The time-market combinationFootnote ⁴ (t) includes 50 states plus Washington DC and Puerto Rico, over 522 weeks (from 2008 to 2017 or 10 years): t = 1, 2, …, 27144. We include the net weight of each ready meal to control for different weights per package that could result in different prices per unit as control variables. The net weight is an important control variable for two reasons, as discussed by Cohen (Reference Cohen2008). First, with respect to sales strategy, large packages could be used for the strategy bundling selling for price discrimination. Second, the net weight of a product is related to the packaging cost (although it is not a large cost). Thus, weights per package could result in different prices per unit.

We also include binary primary ingredient indicators to control product heterogeneity: pizza, pasta, potato, and salad. For example, the binary variable pizza equals one if the observation is a pizza product, zero otherwise. This information is given by the data set. State fixed effect (FE) and week FE variables are also included to control for state heterogeneity and time seasonality. That is, these variables control variations over place (states within the United States) and time. For example, households might have greater supply of salads in lower latitude states such as California or Florida, compared to states such as Minnesota. Also, there might be a larger supply of salads during the summer season weeks. A White test for heteroscedasticity shows evidence of heteroscedastic error terms. Therefore, the model uses the robust standard error to control for heteroscedasticity, given that the time and geographic range is broad (White Reference White1980; Vincent Reference Vincent2015).We also apply the variance inflation factor (VIF) method to test for multicollinearity. Results indicate no multicollinearity (VIF > 10) in the set of independent variables included in the model. Further, when presenting the results, we include the proportion of households who have a positive (negative) marginal utility parameter for each nutritional attributes. Because we assume a normal distribution for the marginal utility parameters, and we can observe the mean and the standard deviation for each marginal utility parameters, we calculate the share of households with a positive/negative marginal utility for a given nutritional variable.

A limitation of this approach is that when using grocery store scanner data, the researcher observes choices, or households’ actual purchases, leaving out the opt-outs. That is, the researcher does not observe the entire choice setting, as this information is not available in the data set. This study offers one mitigation strategy. Because not all the 20 convenience food options were available to every household in every time-market combination, this study uses the national average price for the ready-to-heat meal, when it was not available in a specific time-market combination. This caveat is discussed in Nevo (Reference Nevo2001), who argues that not having the entire choice set results in an overestimated unobservable taste heterogeneity. For future research, we hope for improvements in the data collection, by including observations of those products not purchased or the entire choice set faced by households.