This web-based study used a rigorous 2×2 mixed-factorial experimental design with 2 independent variables: the ethnicity of the participant (Latinx or White) and whether the person in the advertisement was a celebrity or noncelebrity. Latinx and White adolescent participants (aged 13-17 years) each viewed 8 advertisements for novel foods and beverages, including 4 advertisements that featured Latinx celebrities and the same 4 advertisements that featured Latinx noncelebrities (matched on all other attributes), in addition to 2 neutral advertisements. Novel food advertisements feature brands that participants are likely to be unfamiliar with, either created by researchers or brands from other countries. Using novel brands minimizes the likelihood that participants’ ratings will be driven by previous brand preferences and experiences. For example, using a well-known brand such as Coca Cola would drown out our ability to examine the effects of Latinx celebrities’ endorsements due to preexisting attitudes and preferences by consumers. After viewing each advertisement, participants provided attitude measures (for someone like me, advertisement likeability, positive and negative affect, and brand perceptions) and behavioral intentions (consumption, “liking,” following, commenting, and tagging a friend), before viewing the next advertisement. These outcome measures have been used in other food marketing research studies as primary and secondary outcomes [39,41,52].

In January 2019, we recruited participants through Dynata, a survey firm that maintains panels of thousands of adolescents and adults in the United States who are interested in completing surveys in exchange for incentives such as cash, prize drawings, or donations to charity. Dynata’s recruitment involves a rigorous process that uses multiple steps to recruit and screen potential participants. First, randomly selected participants from members of Dynata’s panels are combined with a pool of potential participants who are joining Dynata for the first time after responding to web-based recruitment materials. Those recruitment materials include banner advertisements on social media platforms and other commonly visited websites where participants might notice and respond to the flyers. This combined group of potential participants then receives an invitation to “take a survey,” but are not provided with additional details about surveys that might introduce selection bias (eg, “take a survey about food” might attract people with an interest in food). Second, to be eligible for a study, potential participants complete quality control questions and are randomly matched with surveys for which they likely qualify. Finally, among the randomly selected participants who expressed an interest in and have been randomly matched with a survey, a subset is randomized to complete a specific survey and then receives a link to the survey.

Pretest (N=50) and study participants (N=903) were adolescents, aged 13-17 years, who self-identified as Latinx, or White, had daily internet access, and could read and write in English. Of the 1509 eligible participants who responded to the survey, 992 participants (65.7%) completed all questions. An additional 89 participants did not answer 8 of 10 attention-check questions correctly and were excluded from the analyses. These attention-check questions asked, “What does this advertisement say?” with 3 multiple-choice answers. The final sample included 903 adolescents (436 (48.3%) Latinx adolescents and 467 (51.7%) White adolescents).

The Institutional Review Board for the New York University School of Medicine approved the study (protocol #115-0087), in accordance with all applicable regulations. Parents voluntarily provided consent and adolescent participants provided assent after reading general information about the study, explaining the nature and possible consequences of the study, and the length of the survey. Study data were deidentified to protect participants’ privacy and confidentiality. Participants were compensated by our recruitment firm, Dynata, with incentives such as cash, prize drawings, or donations to charity.

To develop the 8 food and beverage advertisements, we first created pairs of advertisements for 8 novel foods and beverages using Photoshop (Adobe). Each pair included 1 advertisement that featured a Latinx celebrity and a matched advertisement with a Latinx noncelebrity. Because brand familiarity for well-known products (eg, Coca Cola) heavily influences advertisement ratings and the difficulty of changing preexisting attitudes about known brands [67], we selected brands that were only available in countries outside the United States. To increase the likelihood that the advertisements were perceived as Latinx-targeted and that participants would recognize the celebrities, the advertisements included a slogan with at least 1 Spanish word or the name of the celebrity in the slogan or as a signature.

The 16 advertisements (8 pairs) for pretesting included 4 male and 4 female celebrity or noncelebrity pairs and 4 foods and 4 beverages. We pretested these advertisements among 20 Latinx and White adolescents to match the advertisement pairs on the attractiveness of the person in the advertisement, perceived target audience, and celebrity recognition.

Pretest participants first rated the attractiveness of the person in the advertisement, indicated the perceived race or ethnicity of the person, and identified whether the person in the advertisement was a celebrity. They also indicated how much they thought the advertisement was intended for people like them and for people in their age group.

We excluded 1 pair of advertisements because only 10% of participants (n=2) perceived the celebrity to be Latinx and another pair because the majority did not recognize the celebrity. We excluded 2 additional pairs because they scored the lowest on both celebrity recognition and attractiveness measures. The final set of 4 advertisement pairs (Figure 1) included 3 advertisement pairs that featured Latina women alongside a sugar-sweetened beverage (n=2 pairs) or unhealthy food product (n=1 pair) and one that featured a Latinx male alongside a sugar-sweetened beverage. Similar to previous research, we blurred the faces and names of the celebrities in the advertisements in Figure 1. The participants saw unblurred images [35,68].

We also used Photoshop to create 2 additional advertisements to serve as neutral nontargeted stimuli. These 2 advertisements featured bland, unsavory products (ie, water and saltine crackers) and did not show people (Figure 2).

They were used to compare advertisement ratings by Latinx and White adolescents to enable us to control for overall differences in advertisement ratings in our final models.

The web-based survey was designed using Silver Lake and CPP Investment’s Qualtrics software, and the functionality of the survey was tested by the investigators before launching the study. Participants were required to complete the survey on a computer (instead of a mobile device) to ensure the visibility of advertisements. Participating on a mobile device might have made it more difficult to view the features of the advertisements in the survey due to the smaller screen size compared to desktops. Median survey completion time was 23 minutes. The study setting was anywhere a participant could access a computer to complete the experimental survey.

After providing assent and answering eligibility questions, participants were told the following cover story to conceal the true purpose of the study: “We are working with MTV to select commercials that their viewers might like and are interested in your attitudes and opinions on products featured in these advertisements. Participation in this study will involve viewing several different food and beverage advertisements and answering questions about them.” All participants viewed and rated each of the 10 advertisements in a specific order as follows: they first viewed the neutral advertisement for water; then 2 celebrity advertisements and 2 noncelebrity advertisements for different brands presented in random order; then viewed the neutral advertisement for crackers; and finally, the remaining celebrity and noncelebrity advertisements presented in random order. This order ensured that participants did not see advertisements for 2 brands in the same group of 4 advertisements.

The primary outcome measures and dependent variables are listed in Table 1.

Participants provided their attitudes about advertisements and brands immediately after viewing each advertisement, before viewing the next advertisement, using a sliding scale (0-100) for all questions (Table 1). We used 0-100 response scales to minimize bias and lack of reliability versus Likert scales [73], and, among some participant populations (eg, under-resourced communities), Likert scales contradict other response formats [74]. Other studies have shown that 0-100 scales offer stronger psychometric properties than Likert scales [75].

The question, “How much did you like this ad?” (0=“strongly disliked it” to 100=“liked it a lot”) assessed advertisement likeability. A commonly used and validated marketing survey was used to evaluate affective advertisement responses [69]. Participants responded to the following 8 prompts: “Did the ad for [Coke] make you feel…cheerful; good; pleased; stimulated (positive affect); soothed; insulted; irritated; repulsed (negative affect)?” Affective advertisement responses included positive (cheerful, good, pleased, stimulated, and soothed) and negative (insulted, irritated, and repulsed) attributes, rated as 0=“not at all” to 100=“very much so.” Positive and negative attributes were averaged into separate composite positive and negative affect scores. Taste preference was measured with 1 question: ”How much do you think you would like the taste of this product” (0=“Not at all” to 100=“A lot”) [69]. Advertisement efficacy was measured with an adapted version of a commonly used marketing scale that asked participants the following 5 questions: “I like the brand in this ad,” “I react favorably to the brand in this ad,” “The brand in this ad is appealing,” “The brand in this ad is good,” and “The brand in this ad is pleasant,” and rated from 0=“disagree” to 100=“agree” [71]. Brand liking statements were averaged for a combined brand perception rating [70].

Behavioral intention measures included the likelihood to consume the advertised brand [71] and willingness to engage with the brand or advertisement on social media, including the following questions: “What is the likelihood you would Follow or Like this brand on Facebook, Instagram, Twitter, Snapchat, or Tumblr?”; How likely would you be to engage with this brand on social media? (“I would comment on the brand’s post” [0=“very unlikely” to 100=“very likely”]; “I would “Like” this brand’s post” [0=“very unlikely” to 100=“very likely”]; “I would tag a friend” [0=“very unlikely” to 100=“very likely”]). The willingness to engage measures have not been validated but have been used before in previous research [16,36]. All responses were assessed on a sliding scale from 0=“very unlikely” to 100=“very likely” (Figure 3).

Participants rated the likelihood they were the intended audience for the post across all 8 advertisements that featured people, using a scale developed by Aaker et al [72]. The scale asked three questions: (1) “I feel the advertisement was intended for people like me,” (2) “I do not believe the company was targeting consumers like me,” and 3) “The advertiser made this advertisement for people like me.” We assessed these responses using a sliding scale from 0=“disagree” to 100=“agree.” We reverse-coded question 2 and then averaged the scores.

Participants provided demographic characteristics, including age, race, ethnicity, grade level, parental education, and height and weight. They then indicated how often they checked social media if they had access to a smartphone, and which social media platforms they used. Participants who identified as Latinx also completed the Multigroup Ethnic Identity Measure-Revised, a 6-item measure that assesses exploration of and commitment to ethnic identity on a 5-point Likert scale: 1=strongly disagree to 5=strongly agree. The overall composite Multigroup Ethnic Identity Measure-Revised score is calculated by averaging the scores for each of the 6 items. The scale has been validated among adolescents to measure ethnic identity (among Latinx adolescents, α=.930) [76-78].

We used 2-tailed t tests (continuous variables) and chi-square tests (categorical variables) to compare the demographic characteristics of Latinx and White participants. We also compared Latinx and White participants’ ratings of the 2 neutral advertisements on all outcomes to determine whether the 2 groups systematically rated advertisements differently on any of the measures.

We used multilevel linear regression to account for multiple ratings by each participant. Dependent variables included all attitude and behavioral measures. Independent variables included binary terms for participant ethnicity and presence of a celebrity in the advertisement, an interaction term between participant ethnicity and celebrity, random effect of participant, and categorical brand indicator. There were no differences in Latinx and White adolescents’ ratings of the neutral advertisements that did not feature people, so we did not control for that variable in the model. Across all the outcomes except targetedness, Latinx participants did not rate the neutral advertisements any differently than the White participants. They differed on targetedness—Latinx participants said they thought the neutral advertisements were more targeted to them. Due to differences in age by participant ethnicity (P=.04), we included a categorical indicator of age as a covariate in all models. All analyses were conducted in R (R Foundation for Statistical Computing) in February 2020.

More than half of participants (n=540, 59.8%) were aged 13-15 years and 40.2% (n=363) were aged 16-17 years (Table 2). Approximately one-half identified as Latinx (n=436, 48.3%). There were no significant differences in parental education (P=.21) or participant sex (P=.13) between Latinx and White participants. The majority of participants (n=870, 96.3%) reported having one or more social media accounts, and 41.1% reported using a smartphone “almost constantly” to access social media.

On average, participants reported liking the 8 advertisements that featured people (mean 53.58, SD 28.70; Table 3).

Positive affective responses were rated neutrally (mean 49.44, SD 28.85), while participants reported overall low negative affective responses to advertisements (mean 18.77, SD 25.79). Participants rated their perception of the brand on average as 51.63 (SD 29.25) across all 8 advertisements that featured people.

For all attitude measures except negative affect, all participants rated advertisements significantly more positively for advertisements with celebrities than with noncelebrities; there were no main effects of participant ethnicity nor interactions between celebrity and participant ethnicity. However, Latinx participants reported significantly more negative affect toward all advertisements (P=.01).

For brand perception, we found a main effect of celebrity but did not find an interaction, indicating that Latinx and White adolescents rated the brands with celebrities significantly higher than the brands without celebrities.

Participants reported that they would likely consume the product (mean 56.92, SD 31.22), and there was no main effect of celebrity or participant ethnicity. However, we found a significant interaction between participant ethnicity and celebrity (Table 3). Only Latinx participants indicated significantly higher consumption intention when the advertisement featured a celebrity compared to advertisements featuring a noncelebrity (P<.001).

On measures of social media engagement, participants rated their likelihood of following the brand on average 39.31 (SD 33.51) across all 8 advertisements. They indicated a similar likelihood to comment on the social media post (mean 37.09; SD 33.34), “like” the post (mean 39.65, SD 33.81), and “tag” a friend in the post (mean 35.55, SD 33.88).

The main effect of celebrity was significant for two of the four social media engagement measures: (1) participants were more likely to comment on or “Like” posts that featured celebrities in their advertisements compared to posts that featured noncelebrities (Table 3); (2) Latinx participants were more likely to “Follow” or “Like” brands or posts than White participants. There was a main effect of celebrity on participant’s likelihood of commenting on the post such that participants reported a higher likelihood of commenting on posts that featured celebrities (mean 38.1, SE 1.04 compared to those without (mean 36.5, SE 1.04; β=1.54; t₆₃₂₀=4.25; P<.001). We also found the main effects of celebrity (β=1.28; t₆₃₁₉=2.38; P=.02) and participant ethnicity (β=4.27; t₉₆₈=2.09; P=.04) for willingness to “like” the brand. The main effects were qualified by an interaction (β=1.98; t₆₃₁₉=2.57; P=.01). While White participants reported higher willingness to “like” advertisements with celebrities compared to noncelebrities (mean 36.7, SE 1.42 vs mean 38.0, SE 1.42; z=2.38; P=.02); the effect of celebrity was significantly greater among Latinx participants (mean 41.0, SE 1.49 vs mean 44.3, SE 1.49; z=5.87; P<.001).

For willingness to “tag” a brand, we found no main effects of celebrity (t₆₃₁₉=1.55; P=.12) or participant ethnicity (t₉₆₃=1.91; P=.06). However, we did find a significant interaction (t₆₃₁₉=2.46; P=.01). While White participants were equally likely to “tag” an advertisement with a celebrity (mean 33.9, SE 1.43) compared with an advertisement without a celebrity (mean 33.1, SE 1.43; z=1.55; P=.12), Latinx participants were more likely to “tag” an advertisement with a celebrity (mean 39.7, SE 1.50) compared to an advertisement without a celebrity (mean 37.0, SE 1.50; z=4.91; P<.001).

Finally, for willingness to follow a brand on social media, we found no main effect of celebrity (t₆₃₁₉=1.76; P=.08), but did find evidence of a main effect of participant ethnicity (β=4.01; t₉₆₄=1.98; P=.05), as well as a significant interaction (β=1.94; t₆₃₁₉=2.59; P=.01). White adolescents reported the same willingness to follow a brand on social media regardless of whether the advertisement featured a celebrity (mean 37.6, SE 1.42) or not (mean 36.7, SE 1.42; P=.08). However, Latinx participants reported higher likelihood of following the brand on social media when the advertisement featured a celebrity (mean 43.6, SE 1.48 vs mean 40.7, SE 1.48; z=5.30; P<.001).

Our hypothesis for research question 1, is that both Latinx and White adolescents would report higher preferences, including more positive attitudes (advertisement likeability; positive affect; and brand perceptions) and behavioral intentions (product consumption; social media engagement), when advertisements featured Latinx celebrities versus Latinx noncelebrities was mostly supported. Specifically, both Latinx and White participants in our sample reported significantly higher positive attitudes and brand perceptions for brands endorsed by Latinx celebrities compared to Latinx noncelebrities. However, we found significant main effects of celebrity advertisements for just 2 behavioral intentions: likelihood to comment and like brand posts. The findings that celebrities increased the appeal of the advertisements and products reinforces previous research that showed celebrity-endorsed products are more appealing to adolescents [41] and preadolescents [79] compared to products endorsed by noncelebrities.

Our hypothesis for research question 2, that Latinx participants would report more positive attitudes and behavioral intentions for both Latinx celebrity and Latinx noncelebrity advertisements compared to White participants was supported for just 2 outcomes: likelihood to follow brands on social media and like a brand post. We found no significant main effects of participant ethnicity on positive attitudes, brand perceptions, or intent to consume. However, we did find the hypothesized interaction between celebrity and participant ethnicity for most behavioral intentions. Relative to White adolescents, Latinx adolescents reported a higher willingness to “like” and follow brands that featured celebrities compared to brands that featured noncelebrities. In addition, Latinx adolescents reported higher consumption intentions and likelihood to tag a post when advertisements featured celebrities compared to noncelebrities.

These results reinforce other studies on the effects of celebrity endorsements and targeted food marketing. Our findings that both Latinx and White adolescents reported a higher likelihood to comment, “like,” or tag a friend in the post on social media when the advertisements featured celebrities versus noncelebrities is similar to previous research showing that celebrity endorsements increase willingness to engage with brands [80]. Like other studies demonstrating that celebrity endorsements made adolescents more likely to choose the endorsed product [79], we identified a main effect of celebrity on adolescents’ willingness to consume the product—but only for Latinx adolescents.

Latinx celebrities may disproportionately influence Latinx adolescents’ behavioral intentions, including intent to consume and follow brands on social media, compared to White adolescents and noncelebrity advertisements. One potential explanation for the disproportionate influence of Latinx celebrities on Latinx youth is that Latinx communities are underrepresented in popular media, and positive portrayals in the media may heighten Latinx adolescents’ responses to targeted advertisements. Future studies should examine the extent to which exposure to Latinx-targeted advertisements increases Latinx adolescents’ actual food purchases and caloric consumption.

This study makes several contributions to the literature on the effects of food advertising and exposure to racially targeted food advertising. First, this is the first experimental study to examine the intersection of food advertising and celebrity endorsements among Latinx youth. The results indicate a direct causal effect of celebrities on willingness to consume the product and follow or “like” the brand on social media for both Latinx and White youth, even though—after controlling for other possible explanations—Latinx youth were more likely to believe that the advertisements with Latinx celebrities targeted them. These findings demonstrate the power of celebrity endorsements on adolescents’ preferences. Second, this study demonstrates that the intersection between celebrity endorsements and Latinx ethnicity is complex and warrants further research to disentangle these effects. Celebrities influenced adolescents’ attitudes about the advertisements, but it appears that Latinx celebrities may have a disproportionate influence on Latinx adolescents’ behavioral intentions.

This study has several limitations. First, although we only included brands from outside the United States to avoid brand familiarity potentially influencing outcomes, our pretest did not assess brand familiarity among participants. Some adolescents may have been familiar with the international brands in our advertisements, thereby influencing their ratings. An additional limitation is that we only included Latinx celebrities and Latinx noncelebrities. The lack of White people in advertisements limits our ability to discern if our outcomes were shaped by ethnic congruity or some other factor. Another limitation is that in using Dynata for participant recruitment, our sample was not nationally representative. Participants in Dynata’s sample are more likely to use the internet than the general US population, but since our research studies social media, this sample may better reflect our population of interest. Another limitation of opt-in panels is the risk of response bias (eg, providing socially desirable answers). However, we believe this risk is greatly reduced because the survey responses are anonymous, and research involving Dynata shows that adolescents express a variety of preferences (ie, they do not give socially desirable answers) [52-57]. Another limitation is that within-subjects study designs introduce possible “order effects” bias. Because our participants saw each advertisement (eg, the popcorn advertisement featuring a celebrity) and later saw its counterpart (eg, popcorn advertisement featuring a noncelebrity), order effects bias would suggest that the ratings of the second advertisement in a given pair could be affected by their reaction to the first advertisement in that pair. We aimed to reduce such bias, however, by (1) separating the presentation order of advertisements so that participants never saw two similar advertisements in a row; and (2) randomizing the order (eg, for each advertisement pair, some participants saw the celebrity advertisement first, whereas others saw the noncelebrity advertisement first). Although we would have preferred to avoid any possibility of order effects by using randomization via a between-subjects design, within-subjects is valuable when large sample sizes are not feasible—when designing the study, we knew we would not be able to recruit a large enough sample of Latinx adolescents to warrant a between-subjects design with multiple study conditions. Within-subjects’ designs are commonly used in experimental research studies, including studies focused on food-related behaviors [81-83]. Finally, we did not assess participants’ familiarity with the people featured in the advertisements, nor did we ask how much they liked the person in the advertisement. Both of those ratings should have been controlled in our analyses because they could have affected participants’ ratings. Our study has major strengths, though, including the tightly controlled manipulation of images and the pretesting of those matched celebrities and noncelebrities on perceived attractiveness and target audience. Research on the effects of monoethnic and celebrity advertising effectiveness on the Generation Z cohort should also be explored.

Our data advance the theoretical understanding of targeting marketing by showing that viewing ethnicity-congruent food advertisements is particularly powerful for Latinx youth, which is concerning given Latinx communities experience high rates of diet-related diseases. These findings can help support policies that address unhealthy food marketing that targets communities of color, similar to how tobacco research on targeted marketing helped shed light on tobacco companies’ targeted marketing practices [84-87]. Furthermore, in response to concerns about targeted marketing, Facebook now prohibits any advertisement from targeting users based on race [88]. New York State’s proposed Predatory Marketing Prevention Act would aim to prevent advertisements from misleading consumers. These findings can be communicated to policymakers to consider including racial/ethnic-targeted marketing in such bills [89]. These findings suggest an urgent need to reduce celebrity endorsements in ethnically targeted advertisements that promote unhealthy food products to communities disproportionately affected by obesity and diabetes. Celebrities should consider choosing endorsements that support the health of their young fans, and powerful celebrities may be able to encourage companies to allow them to endorse healthier products. The food industry limits food advertising to children aged 12 years and younger, but our findings indicate that industry self-regulatory efforts should expand to include adolescents and address disproportionate advertising of unhealthy food to Latinx youth.