The effects of food advertising and cognitive load on food choices

Participants were students at UCLA, recruited through posters, ads in the campus newsletter, and a campus-wide student participant pool maintained by the Anderson Behavioral Lab, a part of the UCLA Anderson School of Management. All willing students aged 18 or older and without any major self-reported health problems (such as asthma, diabetes, heart disease or depression) were eligible to participate. Participants were told that they would be participating in a study of “television viewing and short-term memory”. Those who completed the study were given a $10 gift certificate to on-campus stores and restaurants.

Participants who met the above inclusion criteria were randomly assigned to one of four groups:

Participants were invited in groups of 20 to view prerecorded movie segments interspersed with advertising. Each session included, in this order, a brief introduction to the study, 45 minutes of viewing, a brief break for snacks (including water and soda options), and the completion of a survey of demographic and other information. The entire session typically lasted just under an hour. Eligible enrollees were asked to enroll for a particular study session via an online scheduling system. The study slots were then randomly assigned to one of the four experimental arms, ensuring only that approximately equal number of sessions were conducted in morning, noon and afternoons.

The viewing consisted of a 3 blocks of content and each block included three 30-second commercials. Each block began with an introductory or transition screen displayed for 15 seconds, an introductory announcement such as that seen in movie theaters requesting that people silence their cell phones, and one, 30-second commercial. This introductory material— 45 seconds total—was followed by a 6-to-7-minute segment excerpted from a movie or television show, a second 30-second commercial break, a second movie or TV segment, and finally one more 30-second commercials. For those participants assigned to the food-advertising arm, the 2 of the 3 commercials were for an obesigenic food product—potato chips, chocolate candy and sugary soda. The order in which food commercial was introduced within each block varied. Each participant assigned to the food-advertising arm accordingly was exposed to 6, 30-second food commercials. In the intervention arm, 1 of the 3 commercials was for irrelevant products (such as cars, sneakers or cell phones). Those assigned to the control-advertising arm saw the same irrelevant commercials as in the food-advertising arm, and in addition saw additional non-food commercial in the place of food commercial in each block. Participants in both arms saw the same number of commercials and the same TV and movie programming. The movie and TV excerpts were chosen to be entertaining, but not highly stimulating, and to avoid mention of food, eating, or obesity-related topics. The same TV and movie excerpts were used in both arms. Additional file 1 reports the full schedule of viewing in both arms.

There were two parts to the cognitive task, a task involving remembering a number and a task involving tracking information on screen.

Immediately before the introductory screen of the third block, participants were shown a number for 7 seconds and asked to memorize the number. Participants were asked not to write the number down and were told that they would be asked to record the number on their final survey. Those assigned to the high-cognitive-load condition were asked to remember a 7-digit number. Those assigned to the low-cognitive-load condition were asked to remember a 2-digit number. These cognitive tasks were chosen because of their similarity to a previous experiment involving cognitive load and food choice [30]. The specific numbers are reported in Additional file 1. The information task demanding high cognitive load required the participants to mentally keep track of the number of times a particular word was uttered in a movie segment. (For example, in the sector showing ‘Duck Dynasty’, the participants were asked to count the number of time the word duck is uttered by any of the actors.) At the end of that segment, they were required to write down the total count on the task answer sheet given to them.

In addition to memorizing a 2-digit number, the low cognitive load task was to answer a simple question per segment. (For example, in the ‘Duck Dynasty’ segment, the question asked about the show’s location, which was mentioned multiple times during the segment.)

At the beginning of the study, participants were informed of the study purpose and protocol and provided their verbal consent to participate. The study protocol was approved by the UCLA Institutional Review Board, approval #12-000323.

A variety of snack and drink items were made freely available to the participants during a break that took place after the viewing and before the survey. Participants were told that there were snacks on the table on one side of the room, and that they were invited to help themselves. These items included water, small bags of sliced apples, small bags of trail mix, granola bars, Coca Cola, small bags of M&M’s, Reese’s Peanut Butter Cups, Hershey’s Kisses, and Lay’s Potato Chips. Ads for Coca Cola, Hershey’s Kisses, M&M’s and Lay’s Potato Chips were included as part of the experimental manipulation in the food-advertising arm. Because no ads were presented for water, apples, trail mix or granola, these items were deemed healthy, with the candy, soda, and potato chips deemed unhealthy. These labels are intended as convenient descriptors only, as it is true that excessive consumption of, say, trail mix, would not be healthy.

Two main outcome variables were assessed: the number of snack items chosen and the total count of calories of food that was chosen. These outcomes were chosen to reflect each of the two distinct dimensions of food-related choices: the type of food chosen and the quantity chosen. Actual consumption of food was not a behavioral target of the experiment and was not observed in the study. Within each of these outcomes, the analysis separately tracks the number of calories from healthy and unhealthy items and the number of healthy and unhealthy items chosen.

The number and types of snack items (including drinks) were observed and recorded by one of the coauthors (SS). Discrete video recording of the snacks area permitted accurate assessment of the items taken by each study participant. Calorie counts were available for each of the healthy and unhealthy items.

The final questionnaire included questions on age, gender, year in school, major, exercise and sleeping habits, fast food consumption, soda consumption, and. television viewing habits. Following previous work on economic disparities in obesity, students were asked to provide the zip code of their parents’ address as a proxy for socioeconomic status [38]. Data from the 2011 American Community Survey, collected by the US Census, were used to determine the average income for each zip code. Participants were dichotomized into high vs low socioeconomic status according to whether the average income in their home zip code is above or below the within-sample median. Foreign students (N = 48) were dropped from these analyses.

The number of unhealthy snack items chosen is count data, with a Poisson distribution. A likelihood ratio test failed to reject the assumption of equidispersion (i.e., that the conditional mean and conditional variance of the outcome are equal; p-value = 0.38), suggesting that poisson is preferred to a negative-binomial regression. The Vuong test revealed no evidence of zero-inflation. Accordingly, the assumptions of Poisson regression could not be rejected and hence, it was the preferred model.

The Poisson regression was first conducted in the whole sample to test the main effects of advertising. To gain some purchase on the statistical meaning of the differences in the effects of advertising between high-cognitive-load and low-cognitive-load conditions two tests were conducted. First, an advertising-cognitive-load interaction term was added to the regressions and its significance was tested. If the coefficient on this term were significant, it would indicate that the analysis could reject the null hypothesis of no effect-modification of advertising by cognitive load.

Second, an equivalence test was conducted [39], using a two one-sided test (TOST) with a delta of 50 kcal for the total calories and 25% for the number of unhealthy snacks. The purpose of an equivalence test is to determine whether the observed point estimate, along with its entire confidence interval, is contained within a specified margin around some anchor, often either zero or some other known quantity. Unlike a statistical significance test, the purpose of an equivalence test is to test the magnitude of difference between an estimate and some other quantity. In this analysis, the question is whether the effects of advertising can be said to be of similar magnitude in a high-cognitive-load and a low-cognitive-load condition. Note that significant differences and equivalences are conceptually distinct: estimates in these two conditions could be statistically significantly different and yet equivalent; not statistically significantly different and yet not equivalent; or any other combination. The equivalence used a one-sided test of whether the interaction of cognitive load and advertising was associated with a change in either total calories or the number of unhealthy snacks of less than 50 calories or less than 25%, respectively.

The sample was then split into sub-samples of high-cognitive-load and low-cognitive-load, and the Poisson regression was conducted in each sample separately to test the effects of advertising under these distinct conditions.

Finally, as a secondary analysis, the samples were further stratified within cognitive-load arms by socioeconomic status, divided at the sample median (excluding the foreign-born participants). Again, the Poisson regression was conducted, this time in 4 distinct sub-samples of the data.

In each regression, the participant’s status in the food-advertising or non-food advertising arm is the only regressor.

The number of calories is a normally distributed variable, but truncated on the left at zero. With this distribution for the dependent variable, Tobit regression is indicated. As for the first outcome, the number of calories chosen was analyzed first in a Tobit regression of the whole sample, with tests for effect modification and equivalence (with a delta of 50 kcal), and then in a stratified regression by cognitive load and finally in a sub-analysis in which the sample was further stratified by socioeconomic status.

All analyses were carried out using Stata 10.1.

A small number (N = 3; <1%) of the participants were observed either to have written their number down when they were asked to remember it, or recalled a number that was substantially different than the number they had been given. The results reported here were analyzed without correcting for this protocol violation. However, an analysis in which these participants were dropped (not reported here) produced highly similar results.

Several additional analyses were conducted to test the robustness of the results to alternative specifications. These analyses included ordinary least squares regression instead of Poisson or Tobit, and analyses that were adjusted for the gender, parental SES, year in college, foreign citizenship and past food habits of the participant. All analyses produced results that were highly similar to those reported here.

Participants in this study were all students at a top-ranked university, which may limit the external generalizability of the study. UCLA is one of the most ethnically and economically diverse universities in the country [53] and has the highest proportion of students receiving Pell Grants of any major university, an important indicator of economic diversity [54]. All the same, many of the results on food choice obtained to date have been conducted among college students, and research in the community would enhance confidence in the generalizability of the results.

Socioeconomic status in this study was measured by a proxy of parental zip code, which is clearly an imperfect measure. In the US, a zip code includes approximately 7,000-10,000 people. Because housing costs in the US tend to follow geographic patterns, zip codes tend to have some degree of economic homogeneity. Yet this homogeneity is not absolute, and there can be variations of income within zip code. In this data set the standard deviation of parental SES as measured by zip code proxy was 38% of the mean. This measure was used because the socioeconomic status of college students is hard to operationalize. An advantage of replicating these results in the community would be the ability to capture more reliable measures of socioeconomic status.

This research is motivated by the possibility that chronic cognitive load enhances the effect of chronic exposure to food advertising. Yet in the confines of this experiment neither chronic cognitive load nor chronic exposure to food advertising could be experimentally manipulated. It could be that the effects of chronic exposures are either greater or lesser than the very brief and relatively small doses manipulated in this experiment. Given how pervasive and profound both cognitive load and food advertising are in American society, other methods besides experimental manipulation will be necessary to tease out the causal roles and interactions of these two factors on eating behaviors.