Understanding influencing attributes of adolescent snack choices: Evidence from a discrete choice experiment

The quality of diet is crucial to the health and well-being of America ’ s adolescents. While a critical step in the goal toward healthier eating is facilitated by school nutrition guidelines that address the food environment, additional information on snack preferences can contribute to this goal. The purpose of this study was to determine which snack nutrient and snack characteristic attributes affect adolescent snack choices and quantify their relative importance. The method used in this study was a discrete choice experiment (DCE) designed with a unique approach of block fractional factorial designs. The study participants were middle school students, aged 11 to 13 years old (n = 166) from an underrepresented, minority-serving middle school in Orange County, California. A mixed logit model was used to analyze the data from DCEs that examined five snack nutrients and five snack characteristics. The results indicated whole grains had the highest relative importance of snack nutrients followed by salt, protein, calories and sugar. Students were more attentive to negative nutrients when two attributes interacted. Price had the highest relative importance of snack characteristics followed by nutritiousness, social, taste, and convenience. Students were more likely to choose a snack in which their family eats even if the snack was low in nutrient content. Understanding students ’ preferences can potentially enhance healthy eating practices on school campuses. This provides implications for stakeholders working to implement and comply with federal school nutrition guidelines.


Introduction
Currently in the United States, nearly one in five school-aged children and adolescents have obesity (CDC, 2018;Hales, Carroll, & Fryar et al., 2017).As childhood obesity remains a critical public health issue, researchers are taking a more focused examination of diet and eating behaviors.One area of interest is snacking patterns.From 1977 to 1978, adolescents consumed about 300 calories per day from snacks; this number increased to an average of 526 calories from 2005 to 2006 (Sebastian et al., 2010).Recent research reveals close to one-third of children's daily caloric intake comes from snacking (Piernas & Popkin, 2010a,b;USDA ARS, 2014).Many of the snacks consumed by adolescents are high in energy density yet low in nutrient density, a pattern often associated with increased risk of obesity and cardiovascular disease (Piernas & Popkin, 2010a,b).
One-third to one-half of adolescents' meals, which account for more than 35% of their daily calories (Mozaffarian et al., 2012), are consumed at school (Glickman, Parker, Sim, Cook, & Miller, 2012).Thus, the quality of school diet is crucial to the health and wellbeing of America's adolescents.In July 2014, the U.S. Department of Agriculture implemented Smart Snacks in Schools Standards (USDA FNS, 2016).This is a set of national nutrition standards for any foods and beverages sold outside of the federal reimbursable school meal program during school hours, such as items from vending machines, school snack bars, or student stores.The implementation of this science-based nutrition program creates a positive change in the school food environment to help make "the healthy choice, the easy choice" (USDA FNS, 2016).While policy change that addresses the food environment is a critical step in the goal toward healthier eating, additional information on snack choice can contribute to the program's efficacy.
Food preferences predict food choices and consumption (Birch & Fisher, 1998;Drewnowski, 1997).A recent study among adolescents showed a strong relationship between snack preferences and actual snack choices (Mielby, Nørgaard, Edelenbos, & Thybo, 2010).An analysis of snack purchasing and consumption behaviors found that snack nutrients such as protein, whole grains and fiber, and snack characteristics such as taste, price, convenience, and nutritiousness were key attributes for snack choices (Forbes, Kahiya, & Balderstone, 2016).It is noted that nutrients in food or snack products are usually categorized into negative nutrients and positive nutrients.A negative nutrient is a nutritional characteristic that should be decreased e.g.salt, sugar, while a positive nutrient is a nutritional characteristic that should be increased e.g.protein, whole grain (Balasubramanian & Cole, 2002).Several studies have found that consumers are more likely to focus on the absence of negative nutrients than on the presence of positive nutrients (Balasubramanian & Cole, 2002;Burton, Garretson, & Velliquette, 1999;Russo et al., 1986).The characteristic attributes of food and snack products are often focused on intrinsic attributes and extrinsic attributes.The intrinsic attributes are related to the physical aspects of a product (e.g.taste, nutrition) while the extrinsic attributes are related to the product, but not in the physical part (e.g.price, convenience) (Bernués, Olaizola, & Corcoran, 2003).Consumers' perceptions of food quality are influenced both by product's intrinsic attributes and extrinsic attributes (Grolleau & Caswell, 2005).According to the literature, consumers are compelled to rely on extrinsic food attributes to make rapid choices when intrinsic attributes are missing (Symmank, 2019;Deng & Srinivasan, 2013;Irmak et al., 2011).
In previous studies, methods such as Likert scales and forced ranking were utilized to assess snack choices.These methods were only able to evaluate an individual effect of each attribute and unable to evaluate the interaction effect between two attributes.To assess snack choices among adolescents, this study utilized a discrete choice experiment (DCE), which is a survey method to facilitate the understanding of subjects' preferences.Since DCEs consider multi-level attributes at the time of the decision, they can be used to identify interactions between attributes and to quantify the relative importance of attributes (de Bekker-Grob, Ryan, & Gerard, 2012).Furthermore, DCEs better represent the tradeoffs between attributes that subjects must consider when making a decision (Bolt, Mahlich, Nakamura, & Nakayama, 2018;Marshall et al., 2016).DCEs can generate a rich source of data and resemble real-life situations in which subjects make decisions based on several choices offered (Veldwijk, Lambooij, De Bekker-Grob, Smit & De Wiz, 2014), each of which features a combination of multi-level attributes (Lambooij et al., 2015;Mangham, Hanson, & McPake, 2009).
Snack preferences vary substantially among adults and children.Adolescence is an important developmental transition from childhood into adulthood in which independence is gained (Smith et al., 2016).Given that snack foods are a large part of the adolescent diet and impact health outcomes, it is critical for adolescents to exercise their independence to establish healthy snacking habits before reaching adulthood (Svisco, Shanks, Ahmed, & Bark, 2019).

Hypotheses and research questions
To continue to expand the literature on adolescents' snack choices, this study presents two DCEs for adolescents, aged 11 to 13 years old.The first DCE (DCE1) focused on snack nutrient attributes, which were selected based upon Smart Snacks in Schools Standards (USDA ARS, 2016) and are categorized into negative nutrients and positive nutrients.This leads to the first hypothesis (H1) and the two research questions (RQ1-RQ2): H1: It is hypothesized that negative nutrients will have higher relative importance compared to positive nutrients.
RQ1: Is there an interaction between negative nutrients and positive nutrients?
RQ2: If interactions exist, are negative nutrients or positive nutrients preferred?
The second DCE (DCE2) focused on a subset of snack characteristic attributes, broken up into extrinsic attributes and intrinsic attributes.This leads to the second hypothesis (H2) and the third research question (RQ3): H2: It is hypothesized that extrinsic snack characteristics will have higher relative importance compared to intrinsic snack characteristics.
RQ3: Is there an interaction between extrinsic and intrinsic snack characteristic attributes?
While the existing literature explores the individual snack attributes, this study contributes to the literature with the ability to quantify the interactions between attributes allowing us to address the hypotheses that could not be considered with existing methods.For example, based on this we are able to explore the interactions between positive and negative nutrients, as well as extrinsic and intrinsic attributes.

Attribute selection
Ideally, all attributes that influence consumers' choices should be considered in a DCE; however, some attributes may need to be excluded so that the choice sets are reasonable for the subjects of the study (Bridges et al., 2011).As the number of attributes increases so does the cognitive difficulty of the DCE, therefore subjects may base their choice on a single or subset of attributes (Mangham et al., 2009).In addition to the number of attributes, the number of levels for each attribute is also a critical consideration.As the number of levels of an attribute increases so does the possibility of the significance of that attribute (Kløjgaard, Bech, & Søgaard, 2012).This is due to the fact that subjects may be more likely to place a higher value on an attribute with more levels (Ratcliffe & Longworth, 2002).
Given the age range of the adolescents, this study considered a limited number of attributes each with two levels to keep the complexity of the DCEs simple.To determine these levels, age of the subjects and ease of cognition were considered (Kløjgaard et al., 2012).Qualitative and quantitative literature was used to determine a baseline level that reflected current snack conditions for adolescents.A second level for each attribute was then determined to represent a change from the baseline.The following section is dedicated to the justification of the selection of the attributes and their levels.

Snack nutrient attributes
The Smart Snacks in Schools Standards (USDA ARS, 2016) require snacks that are sold at school meet at least one of three nutrition criteria: 1) be a grain product that contains 50% or more whole grains by weight (have whole grains as the first ingredient); or 2) have the first ingredient a fruit, a vegetable, a dairy product or a protein food; or 3) be a combination food that contains at least ¼ cup of fruit and/or vegetable.In addition, the snack must meet the following nutrient standards: 200 calories or less, 200 mg of sodium or less, 35% of calories from fat or less, <10% of calories from saturated fat, 0 g of trans fat, and for sugar, 35% by weight or less.These standards intend to promote snacks at school with a high content of whole grains, fruits, vegetables, and protein while limiting snacks with a high content of calories, salt, sugar, and fat.However, there is limited literature on whether students perceive specific snack nutrients as equally important or whether positive nutrients (e.g., whole grains, fruits, vegetables, and protein) and negative P. Rusmevichientong et al. nutrients (e.g.calories, salt, and sugar) are different in terms of importance for their snack choices.Thus, five snack nutrient attributes for this study were selected based on the Smart Snacks in Schools Standards (USDA ARS, 2016) that include three negative nutrient attributes: sugar, salt, and calories and two positive nutrient attributes: protein and whole grains.Each snack nutrient attribute was considered at a corresponding low and high level.To reduce the cognitive burden on the subjects in our study, quantitative values were not considered for the levels of these attributes, as students may have a hard time quantifying the exact amounts of the nutrient.The five snack nutrient attributes, their levels, and symbols used in DCE1 are presented in Table 2.
The negative nutrient attributes selected were sugar, salt, and calories.According to the Survey of Snacking with a Conscience from Nielsen (2014), respondents worldwide reported they care more about the absence of nutrients than the addition of them.In particular, among negative nutrients, roughly one-third of global respondents find it is very important to have snacks low in sugar, salt, fat, and calories.Similarly, Forbes et al. (2016) found sugar, total fat, salt, and calories to be the four most important nutritional factors consumers consider when purchasing a snack.With regard to dietary fat, studies show many consumers are unaware of the different types of fat and the range of nutrition implications for each (Diekman & Malcolm, 2009;Lin & Yen, 2010).One study that exemplifies this is Diekman & Malcolm (2009).They surveyed over 6,000 people from 16 countries and found that approximately half of the consumers were unaware of the different types of fats and were confused about whether fats are good or bad and which are healthier.To avoid confusion, fat was not included as one of the negative nutrients in this study.
The positive nutrient attributes selected were protein and whole grain.The positive nutrients included in the Smart Snacks in Schools Standard are protein, whole grains, fruits, and vegetables.Classifying fruits and vegetables is complicated as they themselves are not nutrients, but rather products.While fruits and vegetables can be used as ingredients in snack foods, when they are, the snack products are often significantly higher in negative nutrients such as calories, saturated fat, sugar, and salt (Wellard et al., 2015).Thus, the focus of this study was placed on protein and whole grains as the positive nutrients.
Five snack characteristic attributes were selected for this study based on existing snack literature (Forbes, Kahiya, & Balderstone, 2016;Symmank, 2019;Rusmevichientong, Jaynes, & Kazemi, 2020;Story, Neumark-Sztainer, & French, 2002).The three extrinsic characteristic attributes were: price, convenience, and social influence, and the two intrinsic characteristic attributes were: taste and nutritiousness.The five snack characteristic attributes, their levels and symbols used in DCE2 are presented in Table 3.Note the levels of these attributes were qualitative.When describing qualitative levels, it is essential to define clearly what is meant by the levels since they must be interpreted by the subjects in the way intended by those who designed the survey (Ryan et al., 2012).
The extrinsic characteristic attributes selected were price, convenience, and social influence.Studies in marketing and business literature showed that consumer judgments are based upon heuristic conclusions about the product's extrinsic characteristic attributes, particularly in the absence of intrinsic characteristic properties (Irmak et al., 2011;Deng & Srinivasan, 2013).When consumers do not have enough information about the intrinsic values of the product, the price attribute plays an essential cue (Acebrón & Dopico, 2000).For this study, the two levels of the price attribute ($3 or less and more than $3) were selected based on the $3 calculated average price of snack items from the US Department of Agriculture, Food and Nutrition Service (USDA FNS, 2019).
Convenience was found to be among one of the top characteristics for consumers and was particularly applicable when purchasing a snack (Forbes et al., 2016).For this study, the two levels of convenience were waiting in line, and no waiting in line.Defining convenience is a complex task as people have varying interpretations of food-related convenience (Jaeger, 2006).Jaeger (2006) defines the elements of convenience as the time and effort to obtain the food.Given that this study was conducted within the school environment and students are provided limited time periods at school for eating, therefore the levels of the convenience attribute were based on how quickly students could obtain a snack.
The final extrinsic attribute was a social influence.Social environment exhibits an substantial role in determining unhealthy food intake among adolescents (Wouters et al., 2010).Interpersonal processes within the family and among friends have a substantial impact on adolescents' food choices and eating behaviors through mechanisms such as modeling, reinforcement, social support, and perceived norms (Story et al., 2002).Therefore, the levels of the social influence attribute for this study were considered at the peer and family level: friends eat and family eats.
The intrinsic characteristic attributes selected were taste and nutritiousness.Several studies found that taste or flavor is the most important factor when consumers make snack purchase decisions (Zbib et al., 2010;Chaplin & Smith, 2011;Forbes et al., 2016).The two levels of the taste attribute (just okay taste and my favorite taste) were selected based upon Kamphuis et al., 2015 in which their reference level of taste attribute was defined as sufficient (just okay) taste.This level is realistic and relevant because consumers will not choose food or snacks that have a bad or unacceptable taste.To determine the second level for the taste attribute, Dong et al. (2016) suggested selecting a more extreme level to best identify the level of the attribute at which a subject would switch choices.Therefore, the second level was defined as favorite taste.
The study of Babicz et al. (1994) reported that when consumers make a snack purchase, one of the highest rated attributes was nutritiousness.The nutritiousness of a snack is deemed relevant for promoting healthy eating among children (Melanson, 2008).Adolescence is a time when young people start to focus more on particular product attributes such as nutritiousness (Bahn, 1989) and are becoming more interested in how healthy eating can affect the body (Contendo, 1981;Ton, MacLeod & Barthelemy, 1996).A study on 11 to 15-year-olds found that the older adolescents in this group had more knowledge about healthy product alternatives, and they tended to choose healthier product options (Berg et al., 2000).Therefore, the two levels of the nutritiousness attribute were selected to be low nutrient and high nutrient.

Experimental design 2.2.1. Discrete choice experiment design
In DCEs, subjects are presented with several questions, known as choice sets.Each choice set is made up of several options, and these options are made of different attributes and their levels (Veldwijk et al., 2014).Subjects are then asked to select one option they prefer.A critical component in the use of a DCE is how the DCE is constructed.The construction of a DCE refers to the selection of the different combinations of the attributes and their levels to form the corresponding options and choice sets.The construction of a DCE is a pivotal step to determine which attributes and their interactions can be estimated.
A well-designed experiment should obtain maximum information with minimum cost and time and is an imperative precursor for any analysis.As the number of attributes increases, so does the number of possible choice sets, which can be overwhelming for the subjects (Dhar P. Rusmevichientong et al. & Simonson, 2003;Iyengar & Kamenica, 2010).To reduce the number of choice sets, this study used an experimental design technique known as a Blocked Fractional Factorial Design (BFFD) based on the foundational work of Jaynes, Wong and Xu (2016), to construct the DCEs.

Blocked fractional factorial design
There are two components in a BFFD: the number of blocks and the size of the blocks.The relationship between the two components of a BFFD and a DCE is as follows.First, let the number of blocks in a BFFD be the number of choice sets.Second, let the size of the blocks in a BFFD be the number of options within each choice set.A BFFD is constructed by first selecting a fraction of a full factorial design through what is called the design generator, and a block is determined by what is called the block generator (Wu & Hamada, 2009).The design and block generators of the BFFD provide the entire structure of the BFFD and thus identify which attributes, and their interactions can be unbiasedly estimated.Two advantages of using a BFFD for a DCE are 1) potentially reducing the number of choice sets while still being able to answer the research question, and 2) estimating all individual effects and some interactions between two attributes.In this study, two DCEs were designed using BFFDs to quantify the relative importance of adolescent snack choices pertaining to snack nutrient and snack characteristic attributes.
For this study, to illustrate how to construct a DCE using a BFFD, five attributes were considered, which are denoted with letters: A, B, C, D, and E, each with two levels.A full factorial design of these five attributes, each with two levels, consists of 32 (2 5 ) possible combinations.To reduce the number of combinations, a one-half fraction of a full factorial design can be considered, which has 16 (2 5-1 ) different combinations.This process to select an optimal subset of the full factorial design is known as a fractional factorial design.
The 16 different combinations from the one-half fraction of the full factorial design are determined by the design generator, which is denoted as: E = ABCD.This fraction is then divided into blocks based on what is called the block generators, which are defined as: b 1 = AB, b 2 = AC.These blocks divide the 16 different combinations into four blocks (2 2 ) each of size four (2 5-1-2 ) (Wu & Hamada, 2009).Thus, using this BFFD to construct a DCE with five attributes (A, B, C, D, and E), each with two levels, results in four choice sets with four options.With this design, all five attributes (A, B, C, D, E) plus seven interactions between two attributes (AD, AE, BD, BE, CD, CE, DE) can be estimated unbiasedly.
Table 1 presents the BFFD for the DCEs used in this study.Both DCE1 and DCE2 consist of four choice sets, each with four options.It is worth noting that in addition to the snack options in each choice set, a none option was also included.This allows subjects to opt-out of selecting an option, which better reflects real-life choices as one may decide not to have a snack if not satisfied with their available options.Forcing subjects to make a choice induces bias, as they would not always make that same choice in real life (Dhar & Simonson, 2003).

Examples of realistic snacks
DCEs in this study did not include actual snack products because individual taste and related brand preference may bias subjects' choice decisions.Rather, the focus was on snack nutrient attributes and snack characteristic attributes.In DCE1, to ensure that realistic snacks with specific nutrient attributes and levels exist in the market, each run in Table 1 was examined to verify that the attribute level combinations represent actual snack products.The Smart Snack Calculator (Retrieved from https://foodplanner.healthiergeneration.org/calculator/) was utilized to confirm either high or low levels for each of the nutrient attributes in selected snack products.For example, in Table 1, run 4 would have the following attribute level combinations: high sugar, low salt, low calorie, high protein, high whole grains.An illustration of such a snack could be a Nabisco Newton's Whole Grain Note: 1 Using BFFD, sixteen different combinations (runs) were divided into four choice sets (blocks) each has four options (block size).Option 5 in each choice set was an opt-out or none option which was later added into the design.

Snacking behavior variables
In addition to the choice sets regarding snack attributes, students were asked three additional questions related to hungriness, snacking frequency, and vending frequency.According to the literature, being hungry can affect consumers' judgment, and people often make different choices when hungry than when they are full (Hunter, 2013).Moreover, snacking and vending frequency are associated with students' food choices and dietary behaviors (Hartmann, Siegrist, & van  The hungriness variable was coded to take a value of 1 if the student reported that they were very hungry or hungry; otherwise a value of 0 was assigned.The snacking frequency variable was coded to take a value of 1 if students reported they always or often snack during a day; otherwise a value of 0 was assigned.The vending frequency variable takes a value of 1 if the student reported never snacked from a vending machine; otherwise it takes a value of 0 if the student indicated they had ever snacked from a vending machine.The responses from these questions were incorporated into the statistical analysis as snacking behavior variables.

Data collection 2.3.1. Sample
On October 26, 2018, the survey was administered during physical education (PE) classes.Data collection was conducted from a convenience sample of students, aged 11-13 years old from a middle school within La Habra City School District in Orange County, California.The City of La Habra is among one of the cities with the highest rates of childhood obesity in the county (Orange County's Healthier Together, 2017).This middle school is designated as an underrepresented, minority-serving school with over 85% of the students of Hispanic or Latino descent.About 70% of the students are eligible for free or reduced meals, and over 40% of the students speak Spanish at home (California Department of Education, 2017).In the U.S., Hispanic and Latino populations are the largest minority group; they experience numerous health disparities compared to white populations, including obesity, diabetes, mental health, chronic obstructive pulmonary disease, and others (Raymond-Flesch, 2017).
A total of 166 students completed the survey and were entered into an opportunity drawing for 36, $50 gift cards to a local grocery venue.The obtained sample size is about 25% larger than the minimum sample size required based on the sample size calculation proposed by Johnson and Orme (2003), which was also verified through computer simulation and prior studies in the field.Approval for this study was granted by the   California State University Fullerton Institutional Review Board.

Procedure
Before implementation, parent permission was granted for student participation, only one student did not provide parent permission.A team of three researchers and six trained research assistants were involved with administering the survey.The team was broken into subgroups of two-three to visit 12 classrooms.Each group utilized a script and a power-point presentation to introduce the study, explain the survey layout, and review the survey procedures.Time was also allocated to provide various examples as well as answer any questions, thus increasing the potential for students' understanding of the survey layout and the symbols.In addition, student assent was obtained, and students were informed their participation was entirely voluntary and that they could quit the survey at any time with no adverse consequences.Students used school-issued iPads to access and complete the survey via Qualtrics, an online survey platform.
Based on previous experience and DCE literature, the survey was developed by a team of researchers and tested using computer simulation.In addition, during this development phase, numerous meetings with school personnel were conducted to better understand student literacy levels, school philosophy, and day-of logistics.A lengthy exchange with the school principal and counselors also provided insight and specific feedback regarding the appropriateness of the language, symbols, and layout of the survey.

Pilot testing
The survey was then pilot tested with over 50 students, aged 11-13 years old, from a different middle school, to examine cognitive burden on the subjects as well as comprehension and content validity.In each DCE, symbols were used to complement the text to represent attribute levels in each choice set.This dual approach increases accessibility to students.Furthermore, symbols and pictures used in DCEs help to aid in comprehension and can be useful in low-literacy populations (Mangham et al., 2009).During the pilot process, a verbal assessment was also conducted with the subjects to determine the understanding of the symbols utilized for each DCE.The feedback provided by the students facilitated the fine-tuning of symbols to be more reflective of the text descriptions for each attribute.

Multinomial logit model
In a DCE, subjects must choose an option from a set of options in each choice set.The outcome variable represents the option selected from each of the choice sets.The benefit that a subject experiences by selecting a particular option is known as the utility.The utility can be viewed as two additively separable parts: 1) a systematic observable component specified as a function of the attributes of the options, and 2) an unobservable random component representing unmeasured variation in the preferences i.e. other attributes not specified in the current DCE.Let there be S choice sets, each containing a finite number of J options.The utility for each individual from choice set s associated with option j is presented in the equation ( 1): where is the utility of an option j in choice set s (as perceived by each individual); is the observable utility component function defined by a vector of attribute levels in option j in choice set s; is the vector of estimated attribute level coefficients, and is an unobservable random component.It is assumed that each individual will choose option j if that option maximizes their utility among all options in the choice set s.The observed utility can be broken into a linear combination of attribute levels and interactions between attributes.Different assumptions on the random component produce different DCEs models.
One of the most commonly used models for analyzing DCEs is the multinomial logit (MNL) model.The MNL model assumes the error terms to be independent and identically distributed (IID) type I extreme value.This assumption leads to the independence of irrelevant alternatives (IIA), which implies the choice between one option over another is unaffected by the inclusion of additional options (McFadden, 1974).To overcome the limitations of the MNL model, a mixed logit (MXL) Fig. 2.An example of a choice set in a discret choice experiment 2 (DCE 2).
P. Rusmevichientong et al. model is considered.A MXL model allows for preference heterogeneity to vary across the subjects (Hauber et al., 2016) and relaxes the IIA property.With a MXL model, the utility function is modified to allow correlations among the error components for different choice alternatives (Train, 2003).
Separate MXL models were used to analyze the data from the two DCEs, where random effects were assumed for the individual snack attributes in the models to allow heterogeneity in subject responses.The model for DCE1 assumed random coefficients for all five individual snack nutrient attributes: sugar, salt, calories, protein, and whole grains.Additionally, as dictated by the BFFD the following interactions between two attributes were included in the model: sugar × protein, sugar × whole grains, salt × protein, salt × whole grains, calories × protein, calories × whole grains, protein × whole grains.Two snacking behavior variables were included in the model: hungriness and snacking frequency, as well as the interaction of each of these variables with the five individual snack nutrient attributes.
The model for DCE2 assumed random coefficients for all five individual snack characteristic attributes: price, convenience, taste, social, nutritiousness.Similar to the model for DCE1, seven interactions between two attributes were included: price × social, price × nutritiousness, convenience × social, convenience × nutritiousness, taste × social, taste × nutritiousness, social × nutritiousness.The hungriness and vending frequency snacking behavior variables as well as their interactions between all five individual attributes were included in the model for DCE2.

Effect coding
The attributes in the DCEs were coded using effects coding.Effects coding for qualitative attributes ensures that the systematic observed utility effects are uncorrelated with the intercept (Louviere, Hensher, & Swait, 2000).The reference level is defined as the negative sum of the estimated coefficients, which means it is internalized in the coefficient estimates and cannot be carried over onto the intercept (Bech & Gyrd-Hansen, 2005).An alternative specific constant (ASC) variable for the none option was created and included as a fixed effect in the model to examine the opt-out effect.The ASC variable was assigned a value of 1 for the four options within a choice set and a value of 0 for the none option.Data was analyzed using the software package mlogit in R® (Croissant, 2012) (Version 3.6.2 R Foundation for Statistical Computing Inc., Vienna, Austria, 2019).Significance was measured at P < .05;however, there were a couple variables that were marginally significant with P slightly greater than 0.05 (0.05 < P < .1)that are worth mentioning.

Relative importance
Additionally, the relative importance of each attribute was estimated to compare its relative effect on adolescents' snack choices.To estimate the relative importance, it requires commensurable measurement units because attributes' parameter size cannot be directly compared due to the underlying subjective scale of the utility size (Lancsar, Louviere, & Flynn, 2007).The partial log-likelihood method was used in this study.This method investigates the explanatory power of each attribute (or attribute level).Attributes that are considered more important in explaining choice decisions will contribute more to the total loglikelihood in the choice model (Crouch & Louviere, 2004).Thus, the relative importance was estimated based upon how much each attribute contributes to the overall log-likelihood.First, the reduced model loglikelihood values were calculated when an individual effect and any of its estimable interactions were excluded from the full model.The partial effect of each attribute was calculated as the change in the reduced model log-likelihood from the full model log-likelihood.The relative effect was calculated as the percent change in the log-likelihood.Attributes which contributed more to the total log-likelihood have a larger relative effect and have a higher ranking of relative importance.

Discrete choice experiment 1: Snack nutrient attributes
The demographics of the study participants are presented in Table 4, with a fairly even split between gender and student grade level (7th and 8th grade).Relating to H1 and R1-RQ3, the results of DCE1 are as follows.The relative importance of the five snack nutrient attributes are presented in Table 5. Whole grains and salt accounted for almost 50% of the log-likelihood, 25%, and 24%, respectively.Protein, calories, and sugar collectively added to about 50% of the log-likelihood, 21%, 15%, and 15%, respectively (H1).The results from the mixed logit model for DCE1 are presented in Table 6.The reference level is the low level for all five snack nutrient attributes.Students were less likely to choose a snack high in sugar, salt, or calories; however, they were more likely to choose a high protein or whole-grain snack.Two interactions between snack nutrient attributes in the model had a significant negative impact on students' snack choices (RQ1).Students were less likely to choose a high-sugar snack even if the snack was considered whole grains (RQ2).Also, despite their selection of a high-protein snack, they were less likely to choose a snack if it was also high in salt (RQ2).
Of the two snacking behavior variables, students who were very hungry or hungry were less likely to choose a snack, and students who reported always or often snacking were more likely to choose a snack.Looking at the interaction between these two snacking behavior variables and the snack nutrient attributes, students were more likely to choose a high protein snack if they were very hungry or hungry.Students who always or often snacked were more likely to choose a snack high in sugar, high in calories, and less likely to choose a snack with whole grains.Lastly, an alternative specific constant (ASC) variable for the none option was added in the model to examine the opt-out effect.The estimate of the ASC variable for the none option indicated students were more likely to choose a snack option than the none option.

Discrete choice experiment 2: Snack characteristic attributes
Relating to H2 and RQ3, the results of DCE2 are as follows.The relative importance of the five snack characteristic attributes are presented in Table 5.Price accounted for 58% of the log-likelihood; nutritiousness accounted for another 18%, social influence 12%, taste 9%, and convenience 3% (H2).The results from the mixed logit model for DCE2 are presented in Table 7. Five snack characteristic attributes were examined (price, convenience, taste, social influence, and nutritiousness).The reference level for each attribute was the low level.Two of the snack characteristic attributes in the model had a significant negative impact on students' snack choices: students were less likely to choose a snack that was more than $3, and students were less likely to The reduced model log-likelihood values were calculated when one attribute and its interaction was excluded from the full model.The partial effect of each attribute was calculated as the change in the reduced model log-likelihood from the full model log-likelihood.The relative effect was calculated as the percent change in the log-likelihood.Attributes which have a larger relative effect have a higher ranking of relative importance.choose a snack with just okay taste.Two of the snack characteristic attributes in the model had a significant positive impact on students' snack choices: students were more likely to select waiting in line to obtain a snack, and students were more likely to choose a snack that was high in nutrients.One interaction between snack characteristic attributes in the model had a significant negative impact on students' snack choices; students were less likely to choose a snack which their friends eat when the snack was high in nutrients (RQ3).
With regard to the two snacking behavior variables (hungriness and vending frequency) in DCE2, the results showed that students who were very hungry or hungry at the time of the survey were less likely to choose a snack.Looking at the interaction between these two snacking behavior variables and the snack characteristic attributes.Students who indicated very hungry or hungry, were more likely to choose a snack in which their friends eat; however, they were less likely to choose a high-nutrient snack.Students who indicated they never snacked from a vending machine were less likely to choose a snack that had waiting in line; while they were more likely to choose a high-nutrient snack.Lastly, the estimate of the ASC variable for the none option indicated students were more likely to choose a snack than opt out.

Discussion
This study provides insights regarding the influencing attributes of snack choices among the sampled adolescents.The use of DCEs quantified the relative importance of snack nutrient and snack characteristic attributes that affect adolescent snack choices.Moreover, the unique construction of the DCEs identified the interactions between attributes.The whole grains nutrient attribute (positive nutrient) had the highest relative importance of the snack nutrients.This may be related to recent efforts by the food industry to prominently utilize the term whole grains in their packaging and messaging as a marketing strategy for revitalizing their products and brands (Future Market Insights, 2018).When analyzing the interaction between two snack nutrient attributes, students were more attentive to negative nutrient attributes e.g., sugar and salt, than positive nutrient attributes e.g., whole grains and protein.This is supported by previous research that found information on negative nutrients was more salient than positive nutrients (Sadler, 1999;Shannon, 1994).
Furthermore, students who reported always or often snacking were more likely to choose high-sugar or high-calorie snacks and were less likely to choose whole grain snacks.This is supported by food science, which states eating foods high in sugar and low in whole grains/fiber typically will cause a spike in blood sugar and then a subsequent crash, leaving the individual feeling unsatiated (Grosvenor & Smolin, 2018).Additionally, students who reported being very hungry or hungry selected a high-protein snack.This is not surprising as subjective measurements of satiety find protein to be more satiating than fats or carbohydrates (Latner & Schwartz, 1999;Porrini et al., 1997).
Among the snack characteristic attributes in DCE2, price (extrinsic attribute) was the most important attribute for students' snack choices.Although adolescents' choices are often dependent on food purchased by their parents at home, they have a certain degree of autonomy in purchasing and consumption outside of home using their disposable income (pocket money) (Li et al., 2017;Roberts, Blinkhorn, & Duxbury, 2003;van Ansem, Schrijvers, Rodenburg, & van de Mheen, 2015).This amount of pocket money is found to be positively correlated with family income (Shah, Syeda, & Bhatti, 2014).Given that the majority of students participating in this study were from low-income families, they likely have limited pocket money.Therefore, the price attribute was significantly more important for their snack choices.
In addition, the snack characteristic attribute that was the least important was convenience.The significant individual effect of convenience suggests adolescents preferred to wait in line.Thus, convenience was not of high priority for the adolescents in this study as they indicated a willingness to wait in line for a snack.While this is surprising, it may be explained by the high social nature of adolescents and their affinity for being with their peers while in line.
The significant interaction between family eats (extrinsic attribute) and low nutrients (intrinsic attribute) underscores the strength of family influence on snack choice.Even though in the individual effects, students were more likely to choose high nutrient snacks, once the extrinsic attribute of social influence was introduced, it outweighed the intrinsic attribute of nutritiousness.Socialization into, and cultural norms around eating habits significantly influence family diet and parental decisions establishing food choices and attitudes towards food consumption within the family (Hardcastle & Blake, 2016).
The findings revealed students chose whole-grain snacks and snacks low in salt and high in protein.This is a promising finding as it renders some validation for programming founded on basic nutrition fundamentals.However, students did not value the role of calories or sugar in snack options relative to the other snack nutrient attributes.As the prevalence of childhood obesity in low-income and Hispanic and Latino communities continues to increase, further examination on calorie and sugar attributes may be needed to ascertain adolescents' understanding of the role these two attributes play in health.

Limitations
There are some limitations to this study.It is noteworthy that the significant findings are limited in scope to the sample demographics.In addition, a larger sample size could have provided a more representative sample of the population.
Furthermore, DCEs present some limitations of their own.There are some limitations in the DCE method that need to be addressed.Despite a less cognitively taxing survey design, the choice task of the DCE may become too challenging for subjects if many attributes are included in the model (Buttorff, Trujillo, Diez-Canseco, Bernabe-Ortiz, & Miranda, 2015).While it may have been feasible to include both snack nutrient attributes and snack characteristic attributes in a single DCE, such a study would require a large number of attributes and thus make choice tasks very complex, particularly for students aged 11 to 13.Other potential snack characteristic attributes such as brand or packaging, and other nutrient attributes such as fat or specific fruits and vegetables could have been incorporated into the study; however, they were omitted to make the study more manageable for subjects.As with any survey, the endogeneity bias may occur as a consequence of omitted variables.
Although DCEs are frequently used for eliciting individual preferences, they can only elicit stated preferences because subjects make a decision based on the hypothetical choices available, not the choices in a real-life situation.As a result, hypothetical bias may arise when choice tasks do not fully reflect reality in the characteristics of choices (Quaife, Terris-Prestholt, Di Tanna, & Vickerman, 2018).In addition, in an experimental setting subjects know their decisions are being thoroughly investigated and examined, thus their decision-making is likely to be influenced by controlled, deliberate, and effortful processes even though subjects are assured of anonymity of their actions (Levitt & List, 2007).For example, subjects may know that whole grains are good for their health, and therefore they may feel they should choose a snack containing whole grains.However, in real-life, subjects may be more likely influenced by impulsive, automatic, and effortless processes.Furthermore, future research in this area could also consider the use of labeled experiments to increase the validity of the DCE (Jaeger & Rose, 2008).

Conclusions and implications
While policy guided by dietetics is an essential component to public health, it can be strengthened when complemented with the understanding of subject preferences, as with the use of DCEs.Thereby, Smart Snacks in Schools Standards, can be reinforced in the school setting by offering snacks that are in line with students' preferences.This approach can potentially enhance healthy eating practices on school campuses.Overall, this study sheds light on how adolescents, aged 11-13 years old, in this underrepresented community consider snack nutrients and characteristics associated with snacks.Thus, providing implications for stakeholders working to implement and comply with federal school nutrition guidelines.

Ethics of human subject participation
This study protocol was approved by the California State University Fullerton Institutional Review Board.All participants provided electronic written informed consent.

Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
2 DCE1 -A: Sugar, B: Salt, C: Calories, D: Protein, E: Whole grains. 3DCE2 -A: Price, B: Convenience, C: Taste, D: Social, E: Nutritiousness. 4The attribute level of − 1 indicates the first level of the attribute.The attribute level of 1 indicates the second level of the attribute.The opt-out option was coded as 0.

Fig. 1 .
Fig. 1.An example of a choice set in a discret choice experiment 1 (DCE1).

Table 1
Block Fractional Factorial Design.

Table 4
Participant Demographics of Students from a Middle School in La Habra City School District in Orange County, California.

Table 5
Relative Importance of Snack Attributes.