Individual Differences in the Influence of Taste and Health Impact Successful Dietary Self-Control: A Mouse Tracking Food Choice Study in Children

Abstract

In order to improve dietary quality among children, there is a need to understand how they make decisions about what foods to eat. This study used a mouse tracking food choice task to better understand how attributes such as health and taste contribute to food decisions among 70 children aged 7-to-11 years old. Children rated health, taste, and desire to eat for 76 common foods that varied in energy density and then used a computer mouse to select which of two presented foods they would like to eat. The presented food pairs were based on children’s own ratings of taste and health so that some trials required self-control to choose the healthier option (i.e., healthy/not tasty paired with unhealthy/tasty). Children’s body mass index (BMI) percentile was not associated with number of healthy choices. To examine mouse trajectories, we replicated previous analytic techniques and applied a novel technique, time-varying effects modeling (TVEM). Results showed that desire to eat impacted food decision-making sooner than taste and health during trials that required self-control, with TVEM showing that early discounting of desire to eat enabled healthier choices. However, these temporal dynamics varied by age, BMI percentile, and overall self-control performance in the task. When the less healthy food was chosen (i.e., self-control failure), older children and children with better overall self-control were more influenced by taste and desire to eat. However, children with higher BMI percentiles showed stronger discounting (i.e., negative influence) of taste when choosing the healthier food. Together this highlights how the influence of hedonic food attributes on food decision-making varies by individual child-level characteristics. Understanding individual differences in the cognitive processes that support healthy food choices in children may help identify targets for interventions aimed at improving child nutrition

1. Introduction

Dietary quality among children in the United States is poor and tends to worsen with age, and as a result, children’s diets are not meeting national dietary recommendations [1]. This is a concern because poor diet quality in childhood is associated with worse cognitive functioning in adolescents [2,3] and poor metabolic outcomes in adults [4]. Children’s dietary patterns and eating behaviors have been shown to be stable and track into adolescence and adulthood [5,6]. Therefore, childhood is a critical developmental window for establishing healthy dietary patterns that can prevent excess weight gain. In order to develop effective interventions that improve children’s diets, it is essential to understand how children make food choices and how attributes such as health and taste influence their decisions.

Food choices can be viewed as value-based decisions that require the integration of information about internal homeostatic states with attributes related to the food (e.g., taste, health, texture) and environment (e.g., time of day, social context) [7,8]. However, not all attributes are considered in every decision. Evidence shows that food-related decisions are often made using simplified heuristics based only on the information that is most highly valued [9]. Although this may be adaptive considering the number of food-related decisions made every day, it may also lead to choices that are not optimal because both adults and children are more influenced by taste than health [9–11]. Since there is often a trade-off between taste and health, incorporating health into decisions for foods that are highly palatable (i.e., tasty) requires self-control.

In order to understand how taste and health influence dietary self-control, researchers have used computer mouse-tracking food choice tasks. Motor movements elicited when making choices using a mouse reflect on-going cognitive processing [12]. Therefore, mouse trajectories can be used to understand conflict and temporal dynamics of the decision-making process. Mouse tracking studies indicate that both children and adults show greater conflict when deciding whether or not to eat unhealthy than healthy foods, which is characterized by an initial movement toward “Yes” before ultimately engaging self-control and choosing “No” [13,14]. In a more direct assessment of self-control, adults showed more conflict when choosing between two competing foods discordant for taste and health (i.e., tasty/not healthy versus healthy/not tasty) compared to when foods were similar for these attributes [15]. However, the amount of conflict may vary between individuals, as adult women who are higher in restrained eating showed less conflict when choosing the healthier food than adult women lower in restrained eating [16]. In a similar series of studies that explored the temporal dynamics of dietary self-control in adults, health influenced the food decision-making process later than taste [14,17], however, there were individual differences in the decision making process that were related to weight status [14] and dietary self-control [17]. Together, these studies show the utility of mouse-tracking tasks to examine individual differences in the food choice process, but they also highlight the need to conduct additional studies in pediatric samples.

To better understand the temporal dynamics of food choice and dietary self-control in children, the current study examined the influence of taste and health on food decision-making using a mouse-tracking task that was previously used in adults [17]. While prior mouse tracking studies asked children to make decisions about foods presented one at a time [13], the current study extended previous work by having children choose between two competing foods as done in adults [17]. The first aim of this study was to determine the temporal dynamics of taste and health on children’s food decisions when they are faced with a choice between two foods (e.g., apple versus cookie), a situation more relevant to real-world decisions. This was done by replicating the procedure and analytic steps used in Sullivan et. al (2015). The second aim of this study was to characterize mouse trajectories using a novel analytical method called time-varying effects modeling (TVEM) [18], which allowed us to capture multiple windows of time during the decision-making process where taste and health differentially influenced children’s food choice. The final aim of this study was to examine individual differences in children’s performance on trials that required self-control to determine how the decision-making process changes during successful versus unsuccessful efforts. Since previous work in children has shown that both age and body mass index (BMI) percentile influence the decision making process [13], we focused analyses on these characteristics.

2. Materials and methods

2.1. Participants

Children were enrolled in a 4-visit study at the Children’s Eating Behavior Laboratory with the aim of understanding the relationships between food decision-making and weight status. The current paper focused on one visit that examined dietary self-control in 70 children aged 7-to-11 years old (M = 9.49, SD = 1.39). Children were in good health and did not meet criteria for metabolic, psychiatric, or learning disorders, spoke English as their first language, and could read at grade level according to parental report. By design, half of the sample was classified as having overweight or obesity (BMI-for-age ≥ 85^th percentile) and half was classified as having healthy weight (BMI-for-age between 5^th – 85^th percentile). Exclusion criteria included being underweight according to parental report (BMI-for-age < 5^th percentile) or the presence of food allergies or dietary restrictions that would preclude intake of the foods served in the study. Due to the use of functional MRI in one of the study visits (data reported elsewhere [19,20]), children also did not have any common MRI contraindications (e.g., metal in the body, claustrophobia) and were not left-handed. Child assent and parental consent were obtained in accordance with the Institutional Review Board of The Pennsylvania State University. Demographic and family characteristics are presented in Table 1.

Table 1.

Participant Characteristics

Total (Number Males)	70 (34)
Age (Mean [range]), yrs	9.49 [7 – 11]
BMI percentile (Mean [range])	74.5 [11–99]
BMI Z-score (Mean [range])	0.92 [−1.25 − 2.57]
Weight Status, N
Obese	19
Overweight	16
Healthy Weight	35
Ethnicity, N
Hispanic/Latinx	4
Not Hispanic/Latinx	66
Race, N
Black/African American	3
Caucasian	64
Other or Multiple	3
Household Income, N
>$100,000	20
$50,000 - $100,000	32
<$50,000	17
Not Reported	1
Maternal Education, N
Beyond Bachelor's Degree	11
Bachelor's Degree	28
Associate's/Technical Degree	8
High School Diploma	13
Other or Not Reported	10

2.2. Study Design and Procedure

The overall study was a within-subjects, crossover design that took place across four visits to the Children’s Eating Behavior Laboratory. The current study focused on a subset of data pertaining to the Food Choice mouse-tracking task (referred to as the “Food Choice” task throughout). Anthropometrics and fMRI were always performed on the first and fourth visits, respectively. However, the order of the rest of the assessments, which included a range of eating behavior and cognitive tests, were randomly assigned and counterbalanced across visits 1–3 (for more information see [19,20]). Therefore, the Food Choice task could have been administered in any of the first three visits. This study used the same mouse tracking task from a prior study published in adults [17], however, the food images used in the current study were modified to include foods commonly eaten by children that were photographed without brand information (see section 2.4.1). Children were asked to fast for 3 hours prior to each visit to the laboratory in order to approximate a hunger state similar to what is experienced prior to lunch or dinner.

2.3. Anthropometrics

On the first visit, a trained researcher measured height and weight of the child in light clothing and without shoes using a stadiometer (Detecto model 437, Webb City, MO; precision to nearest 0.1 cm) and standard scale (Seca model 202, Chino, CA; precision to nearest 0.1 kg). The average height and weight from two separate measurements was used to compute BMI (kg/m²) and determine the children’s BMI percentile and z-score [21].

2.4. Food Choice Task

During the Food Choice Task, children first rated the food images on health, taste, and desire to eat (see 2.4.1). After the ratings were complete, children read a passage from WebMD.com used in the original task [17] that described the definition of a healthy food in order to encourage them to think about the health of food items when making choices (supplemental Figure S1). Children then rated their hunger on a 4-point Likert scale (“Not At All”, “A Little”, “Moderately”, “Extremely”) and completed the Food Choice task (see 2.4.3). After the Food Choice task, children received a randomly selected snack from the list of foods they chose during the task.

2.4.1. Food Stimuli

A total of 76 food images (347 × 576 pixels, 72 dpi, RBG color) were used during the Food Choice task. Photographed foods were selected from the Continuing Survey of Food intake (CSFII What we eat in America) of foods commonly eaten by children in this age group. The foods ranged in energy density (ED; M = 2.3, SD = 1.9, Range = 0.1 – 6.6 kcal/g) and were roughly balanced in the number of fruits (N = 13), vegetables (N = 16), sweets (N = 13), snacks (N = 16), and meal items/entrees (N = 18). All foods were photographed on plates at an angle that would approximate the perspective a typical sized child would have sitting at a table with the plate in front of them (see Figure 1A for examples).

Figure 1.

A) Prototype images of foods used in the Food Choice Task. B) Displays for a single choice and an example mouse trajectory during

2.4.2. Food Ratings

Before completing the Food Choice Task, children rated each of the 76 food images on Health, Taste, and Desire to Eat on 5-point Likert scales. Each of the images was presented one at a time on a desktop computer screen and children rated the following attributes: 1) Health: “how HEALTHY each food item is to you” (“Very Unhealthy”, “Unhealthy”, “Neutral”, “Healthy”, “Very Healthy”); 2) Taste: “how TASTY each food item is to you” (“Very Bad”, “Bad”, “Neutral”, “Good”, “Very Good”); and 3) Desire to Eat: “how much you would LIKE to eat each food at the end of the experiment” (“Strong Dislike”, “Dislike”, “Neutral”, “Like”, “Strong Like”). Although Sullivan et al. (2015) referred to the third rating scale as “Liking,” we decided to use “Desire to Eat” in order to more closely reflect the wording of the question and avoid confusion with the closely related “Taste” attribute. Prior to making ratings, children were told they had the option to rate images as “Neutral”, but to avoid doing so when possible and to “use their ‘gut’ feelings” about the food if unsure. All food images were rated for a single attribute before moving to the next attribute. The order in which the attributes were rated was counterbalanced across participants and the images were presented in a randomized order for each attribute. During the ratings, colored food images were presented on a black screen along with the attribute rating scale until the child made a selection with a white fixation cross presented for 0.3 s between each food rating (presentation software: Psychtoolbox3 [22,23]).

2.4.3. Food Choice Trials

For each food choice trial, children were presented with two of the previously rated food items and instructed to choose the food they would like to eat right now by moving the computer mouse continuously to that item and clicking on it. They were told that one of the chosen foods would be selected at the end of the task and they would be required to eat it, so to be sure they were choosing foods they would like to eat. However, they were also reminded to factor the health of each food into their decisions. They were instructed to try to make their decision quickly, to move the mouse smoothly, and to not “play” with the mouse. There was a total of six randomly presented blocks of 40 choice trials each, resulting in a total of 240 trials.

Each trial began with a white ‘START’ box centered at the bottom of the black screen. Children were instructed to use the computer mouse to click the ‘START’ box when they were ready for the next choice. After children initiated the trial, a black screen with no mouse cursor was presented for a short duration (M = 350 ms, range: 200–500 ms) before the two food images were presented and the mouse cursor reappeared where the “START: box was. The food images were presented inside white frames at the top left and right of the screen (174 × 131 pixels; see Figure 1B) and the food pairs were selected based on the child’s own ratings of Health and Taste. After excluding foods children rated “Neutral”, all possible combinations of Health and Taste ratings were randomly presented (e.g., “very healthy”/”good” and “healthy”/”good”, “unhealthy”/”good” and “very unhealthy”/”good”, etc.). See Supplemental Materials full description of image pairing and randomization. The Psychophysics Toolbox [23] was used to track the cursor’s position (temporal resolution of 67 Hz), which was recorded in x, y coordinates. A white fixation cross was presented between trials (M = 550 ms, range = 400 – 700 ms).

2.5. Analytic Approach

All data were analyzed using R [24], with some data processing steps conducted using MATLAB 2018a. Since this food choice task had not been used in children before, we repeated the paradigm validations done in adults [17] (see supplemental materials). In addition to methods established by Sullivan et al., (2015) for analyzing mouse tracking during the Food Choice task (referred to as the “standard” approach throughout) [14,15,17], the current study used time-varying effects models (TVEM), which is a novel approach to analyzing mouse tracking data that is able to estimate coefficients as a function of time [25]. All TVEM analyses were conducted using SAS 9.4 and the %TVEM_v311 macro [18]. Both the standard and TVEM approaches are described in detail in sections 2.5.1 – 2.5.3.

In addition to analyzing mouse trajectories across all trials, this study also isolated trials that required self-control in order to analyze mouse trajectories where the child was successful versus unsuccessful. Self-control trials were defined as trials where one food was rated as tastier but less healthy than the other food (e.g., “unhealthy”/”very good” versus “unhealthy”/”bad”) [17] and overall self-control performance was assessed with the self-control success ratio (SCSR), defined as the percent of trials where the healthier, less tasty option was chosen [17]. In previous studies, differences in mouse trajectories have been compared across groups with better or worse SCSR [17], however, differences in the impact of Taste and Health on mouse trajectories during Successful (choose healthier/less tasty food) and Unsuccessful (choose less healthy/tastier food) self-control have not been examined. To account for repeated measurement, mixed-effects general linear models (GLMs) with a random intercept were used.

2.5.1. Mouse trajectory data processing

The mouse trajectories were processed in MATLAB 2018a using the same steps used in previous publications [14,15,17], which included: 1) resampling mouse coordinates so the start is at (0,0) and choice is at (−1,1) or (1,1) for left or right food items, respectively; 2) excluding trials where reaction time exceeds 2 standard deviations of the child’s average (M (SD) = 9.7 (2.7), 5% of trials), where the mouse trajectory crosses the y-axis (vertical center) 3 or more times (M (SD) = 32.1 (19.7), 16% of trials), or the mouse leaves the screen or overshoots the food item (i.e., M (SD) = 2.2 (2.9), 1% of trials); 3) normalizing the mouse trajectory using linear interpolation of the recorded timestamps so that each trial has 100 normalized time spaces (i.e., X_t1, Y_t1 through X_t101, Y_t101). The mouse trajectory angle was then calculated at each normalized timepoint for each trial as the angle of the cursor with respect to the origin (0,0) using the inverse arc tangent. For a more detailed description please see the original paper, Sullivan et al., (2015), or this study’s project on Open Science Framework (https://osf.io/sz4dj/).

2.5.2. Trajectory Analysis

Standard Approach. The processed mouse trajectory data were analyzed in R [24] following the procedure used in other studies [14,15,17]. For each trial there were 100 mouse angle trajectory measurements in addition to the trial-specific differences in the child’s Health, Taste, and Desire to Eat ratings for the pairs of foods presented. Since the ratings for Health and Taste, but not Desire to Eat, were used to determine the food pairs, differences in Desire to Eat ratings were analyzed in separate GLM than Taste and Health (see Equations 1.1 and 1.2, where tp indicates the time point from 1 to 100). For each child, the association between mouse angle trajectories and the differences in attribute ratings (right – left) was determined at each of the 100 normalized time points. This was done so that positive associations between attribute differences and angle trajectory would indicate that the mouse was moving toward the higher rated food image

$$MouseAngl{e}_{tp}\sim {\beta }_{\theta ,tp}+{\beta }_{1,tp}DesireToEat(R-L)\#$$ (1.1)

$$MouseAngl{e}_{tp}\sim {\beta }_{\theta ,tp}+{\beta }_{1,tp}Health(R-L)+{\beta }_{2,tp}Taste(R-L)\#$$ (1.2)

The Significant Time Points (STPs) for each attribute were estimated for both the entire sample and for each child individually. For the entire sample, t-tests were conducted at each normalized time point to determine if the attribute’s β coefficients across children were significantly greater than zero (1-tailed α = 0.05). For individually determined STPs, one-tailed tests of the attribute’s β coefficient were used at each time point (i.e., β > 0; α = 0.05). For both approaches, the STP for each attribute was defined as the time point at which the β coefficient became significant and remained significant across all time points until the end of trial, with a minimum of 10 time points in a row (i.e., time point 90 would be the last possible STP). The definition of STPs requires the association between mouse angle trajectory and the difference in attribute ratings to remain significant through the end of the trial, therefore, some individuals may not show an STP. In order to determine if children with an STP differ in their ratings from those without, we conducted an Attribute (Health, Taste, Desire to Eat) x STP Present (Yes, No) ANOVA. To examine Self-Control trials, these steps were completed separately for trials where children had Successful (i.e., choose healthier/less tasty option) and Unsuccessful (i.e., choose less healthy/tastier option) self-control. To examine individual differences, correlations and GLMs with STPs were conducted.

2.5.3. Trajectory Analysis

Time-Varying Effects Model (TVEM). In addition to the standard approach, a novel analytical approach was applied to the mouse tracking data. TVEM, initially termed the varying-coefficient model, originated over 20 years ago but has only recently been applied to behavioral and psychological sciences [25,26]. With the increased use of intensive longitudinal data and technological resources, TVEM has most commonly been used in ecological momentary assessment studies related to smoking cessation [7–9] and longitudinal studies related to the onset and development of high risk sexual [29] and substance use [30–32] behaviors. It is well suited for studies with intensive longitudinal data because it allows researchers to model how a predictor (e.g., difference in Health ratings) is differentially associated with an outcome (e.g., mouse angle trajectory) continuously over time. TVEM estimates β coefficients as a continuous function over time rather than assuming them to be fixed. That is, TVEM allows for the association between the difference in Health ratings and mouse angle trajectory to change over the course of the decision-making process rather than assuming a static or fixed association. Mouse tracking data can be viewed as a unique type of intensive longitudinal data in that there is frequent and comprehensive measurement of behavior and its associated context over more than 10 repeated instances [25]. Using TVEM, the influence of trial characteristics (e.g., difference in health ratings) on mouse angle trajectory can be modeled dynamically across the decision-making process.

Mixed-effects TVEMs were used, allowing random intercept and slope for time (i.e., the 100 normalized time points). As in the standard analyses, Desire to Eat was modeled separately from Health and Taste due to how the food pairs were chosen and to follow previously established methods [17]. In addition to models with all the trials, the influence of Self-Control (Successful, Unsuccessful) was tested as a moderator of the association between attribute differences and mouse angle trajectory in self-control trials. The impact of SCSR, age, and BMI percentile were tested as moderators of the association between attribute differences and mouse angle trajectory in separate TVEMs for trials with Successful and Unsuccessful Self-Control. All models were estimated with 1 to 5 knots for each time-varying effect starting with the intercept, then attribute differences, and finally any moderators of interest. The best model fit was determined using Bayes Information Criterion and Akaiki information criterion (see Supplemental Materials Tables S1–S5) [26,27,30].

3. Results

3.1. Descriptive Statistics

3.1.1. Demographic Characteristics

Sample characteristics are presented in Table 1 and reflect the general demographics of central Pennsylvania. The majority of children were Caucasian (not Hispanic or Lantinx N = 64), had mothers with education beyond high school (N = 47), and family income of $50,000 a year or more (N = 52). By study design, the sample ranged in BMI with half the sample meeting criteria for either overweight or obesity (overweight: BMI > 85^th percentile, N = 16; obese: BMI ≥ 95^th percentile, N = 19). BMI percentile was not related to child age and did not differ by sex (p’s > 0.459).

3.1.2. Hunger

Prior to the Food Choice task, the majority of children (N =54, 77%) were ‘Moderately’ or ‘Extremely’ hungry with only two children stating they were not hungry. Hunger was negatively associated with BMI percentile (β(se) = −0.01 (0.004), p = 0.006) and girls reported greater hunger than boys (t(59) = −2.50, p = 0.015). Age was not associated with hunger (β(se) = 0.01 (0.07), p = 0.839).

3.1.2. Food Ratings

The mean rating for all foods was above “Neutral” for Taste and Desire to Eat while the mean rating for 76 % (N = 58) of foods was above “Neutral” for Health (supplemental Figure S2). Across individual foods, Health ratings were negatively associated with Taste (β(se) = −0.05 (0.02), p = 0.002) and Desire to Eat (β(se) = −0.05 (0.02), p = 0.002) while Taste and Desire to Eat were positive associated with each other (β(se) = 0.79 (0.01), p < 0.001). When looking at children’s average ratings, those who rated foods as Healthier also tended to rate foods as higher in Taste and Desire to Eat (Table 2). There were no associations between children’s average food ratings and age or BMI percentile (p’s > 0.446; Table 2) and there were no differences by sex (p’s > 0.167).

Table 2.

Correlations Between Children’s Average Attribute Ratings and Individual Characteristics

	1	2	3
1. Health
2. Taste	0.83***
3. Desire to Eat	0.61***	0.44***
4. BMI percentile	−0.02	−0.12	−0.17
5. Age	−0.14	−0.11	0.07

BMI: Body Mass Index

^***p < 0.001

Since the foods were a mix of higher (≥ 1.5 kcal/g) and lower (<1.5 kcal/g) ED, the impact of ED on children’s ratings was tested with an ED (higher, lower) x Attribute (Health, Taste, Desire to Eat) analysis of variance (ANOVA). There were main effects of both Attribute (F(2,325) = 135.00, p < 0.001) and ED F(2,325) = 4.34, p = 0.038), but no ED by Attribute interaction (p = 0.12). The main effect of ED showed greater mean ratings for higher (M=0.55, SE = 0.05) than lower ED foods (M = 0.61, t(325) = 2.08, p = 0.038), regardless of attribute, while the main effect of Attribute showed lower mean ratings for Health (M=0.27, SE = 0.03) compared to Taste (M = 0.79, t(325)=−12.12, p < 0.001) and Desire to Eat (M = 0.67, t(325)=−15.67, p < 0.001) and lower mean ratings for Desire to Eat compared to Taste (t(325)=−3.54, p = 0.001), regardless of food ED. There were no interactions between ED and age, BMI percentile, or sex for ratings on any of the Attributes (p’s > 0.064).

3.1.3. Food Choices

Reaction time (M = 1552 ms, SD = 0.28 ms) was not related to BMI percentile or sex (p’s > 0.248). However, age was negatively associated with reaction time (r = −0.36, p = 0.003). Food choices during self-control trials (i.e., healthier/less tasty food paired with less healthy/tastier food) were examined via the SCSR, which was not related to hunger, BMI percentile, age, sex or reaction time (p’s > 0.101). Additionally, the ED of chosen foods was not associated with BMI percentile when using a mixed effects model to look across trials (β(SE) = −0.001 (0.002), p=0.427) or when testing a correlation with the average energy density of chosen foods (r = −0.04, p = 0.801).

3.3. Temporal Dynamics During Food Choice

Following the standard approach [17], group STPs were first estimated across the entire sample and then for each child individually. To determine if children with and without an individual STPs differed (see Table 3 for N’s), we conducted Attribute (Health, Taste, Desire to Eat) x STP Present (Yes, No) ANOVAs. To match the standard approach, two TVEM models were used to determine how differences in Attributes are associated with mouse angle across time points (see section 2.5.2 for more details): 1) differences in Desire to Eat (referred to as “Desire to Eat” throughout) and 2) differences in Health (referred to as “Health” throughout) plus differences in Taste (referred to as “Taste” throughout). Time points were considered significant if the 95% confidence interval for the β coefficient did not include zero.

Table 3.

Significant Time Points by Attribute

	Children (N)	Mean (SD)	Range
All Trials (n = 240)
Health	25	66.8 (13.1)	45-82
Taste	31	66.5 (17.3)	5-90
Desire to Eat	28	67.4 (17.1)	6-86
Successful Self-Control Trials
Health (mean: 40 trials)	53	70.8 (10.6)	32-85
Taste (mean: 41 trials)	46	71.2 (12.1)	40-90
Desire to Eat (mean: 41 trials)	49	55.7 (17.7)	2-77
Unsuccessful Self-Control Trials
Health (mean: 39 trials)	48	70.0 (13.0)	35-90
Taste (mean: 38 trials)	50	71.7 (11.1)	44-90
Desire to Eat (mean: 39 trials)	51	59.1 (13.3)	2-79

SD: standard deviation

STP: significant time point

The standard approach showed a group STP for Health approximately half way through the decision-making process (STP = 52). There were no group STPs for Taste or Desire to Eat. Less than half the children had individual STPs for each attribute (Health: 36%; Taste: 44%; and Desire to Eat: 40%; see Table 3). There were no differences in SCSR, Age, BMI percentile, or sex between children with versus without an STP (p’s > 0.109). However, there was a significant Attribute (Health, Taste, Desire to Eat) x STP Present (Yes, No) interaction for Hunger (F(2, 189) = 4.06, p = 0.019) such that children with an STP for Taste reported greater hunger than those who did not (p = 0.011). There were no differences in Hunger for those with or without STPs for Health (p = 0.152) or Desire to Eat (p = 0.931). Additionally, individual STPs did not differ by attribute (F(2, 46) = 0.32, p = 0.773; Table 3). Taste, Health, and Desire to Eat were all associated with mouse angle trajectory about two-thirds of the way into the decision-making process (Figure 2A, Table 3). Together, this indicates that children who were hungrier were more likely to show significant influence of taste that lasted through the end of the trial and that, for children with STPs, there was no difference in when Taste, Health, and Desire to Eat influenced choice.

Figure 2.

Associations between mouse angle and differences in attribute ratings (Right – Left) for all trials. Figures A and C show Average β coefficient from individual GLMs with shaded area indicating standard error regions and vertical lines indicating average individual STPs for each Attribute. Figures B and D show Time-Varying Effect Model β coefficient with dotted lines or shading indicating 95% confidence interval. Horizontal bars on the X-axis indicate timepoints where the β coefficient is significant (i.e., 95% confidence region does not include 0). Attributes are distinguished by color and self-control performance by line type. A) standard approach for all trials, B) TVEM approach for all trials, C) standard approach for self-control trials, D) TVEM approach for self-control trials.

3.4. Temporal Dynamics During Self-Control

While general self-control ability is related to the amount of conflict seen during self-control trials in adults [15], is not possible to determine whether children are truly engaging in self-control when selecting the healthier/less tasty option. To be consistent with previous studies we are using the term “Successful” self-control when children chose the healthier/less tasty food and “Unsuccessful” self-control when children chose the less healthy/tastier food. On average, participants had 67 self-control trials (SD = 31). The number of self-control trials did not differ by BMI percentile, age, sex or SCSR (p’s > 0.314). When looking at self-control trials, group and individual STPs were estimated separately for Successful (chose the healthier/less tasty option) and Unsuccessful (chose the less healthy/tastier option) trials. For the standard approach, differences in individual STPs were tested using an Attribute (Health, Taste, Desire to Eat) x Self-Control (Successful, Unsuccessful) ANOVA. Because the TVEM approach uses Attribute Differences directly, as opposed to STPs, two models were run to maintain consistency with the standard approach (see section 2.5.2 for more details): 1) Desire to Eat Difference x Self-Control (Successful, Unsuccessful) and 2) Health Difference x Self-Control (Successful, Unsuccessful) plus Taste Difference x Self-Control (Successful, Unsuccessful).

The standard approach showed group STPs for all attributes around halfway through the decision-making process for Successful (Health STP = 55, Taste STP = 42, Desire to Eat STP = 34) and Unsuccessful (Health STP = 56, Taste STP = 51, and Desire to Eat STP = 39) self-control trials. Approximately two-thirds of children had individual STPs (Successful-Health: 76%, Taste: 66%, Desire to Eat: 70%; Unsuccessful-Health: 69%, Taste: 71%, Desire to Eat: 73%; Table 3). There was a significant main effect of Attribute (F(2, 240) = 47.2, p < 0.001) such that children had earlier individual STPs for Desire to Eat than Health (p < 0.001) and Taste (p < 0.001), but showed no difference between Health and Taste (p = 0.840; Table 3). Individual STPs did not differ by Self-Control (p = 0.646) and there was no interaction between Self-Control and Attribute (p = 0.412). None of the individual characteristics (i.e., SCSR, Age, BMI percentile, sex, or hunger) differed by presence of STP (p’s > 0.163). Together, this indicates that information about Desire to Eat was incorporated into the food decision-making process earlier than information about Taste or Health, regardless of whether children were successful in selecting the healthier food or not (Figure 2C).

When using TVEM, all three attributes had time windows where there were significant interactions between Attribute and Self-Control (Health: time points 44–100, Taste: time points 36–100, and Desire to Eat: time points 1–4 and 35–100; Figure 2D). Health was positively associated with mouse angle for Successful self-control trials but negatively associated with mouse angle for Unsuccessful self-control trials. In contrast, Taste and Desire to Eat were negatively associated with mouse angle for Successful self-control trials but positively associated with mouse angle for Unsuccessful self-control trials. Adjusting for number of self-control trials did not change results (see Supplemental Materials Figures S8–S11). Together, this indicates that as early as a third of the way through the decision-making process, children are moving toward the healthier/less tasty option and away from the less healthy/tastier option when self-control is successful and the opposite when it is not.

In order to further characterize the interaction between Attribute and Self-Control, we examined the influence of each attribute separately during Successful and Unsuccessful trials. Health and Taste became significant at the same time points regardless of self-control performance, with Taste influencing mouse angle earlier (time points 37–100) than Health (time points 51–100). Desire to Eat, however, showed a significant association with mouse angle earlier for Successful (time points 5–19 and 24–100) than Unsuccessful (time points 37–100) self-control trials. This indicates that movement away from the more highly desired food during Successful trials occurred earlier than movement toward the more highly desired food during Unsuccessful self-control trials.

3.5. Individual Differences in Temporal Dynamics During Food Choice

Since less than half the sample had STPs across all trials, only TVEM was used to probe individual differences in temporal dynamics. Individual Characteristic (e.g., Age) x Attribute models were run for each attribute separately. Adjusting models for number of self-control trials did not change results (see Supplemental Materials). In order to visualize how the associations between Attribute and mouse angle changed with individual characteristics, all significant TVEM interactions were plotted by showing the association separately for children in the first quartile, quartiles 2–3, and fourth quartile of the relevant individual characteristic. For example, for a Health x SCSR interaction, children were grouped by quartile for SCSR in order to see how the association between Health and mouse angle differed for those with lowest, middle, and highest SCSR. This approach for categorizing the data was used because the first and fourth quartiles better represented the two tails of the distribution than if the sample was evenly split into thirds.

3.5.1. Associations with Overall Self-Control (SCSR)

There were significant interactions between SCSR and Health (Figure 3A), Taste (Figure 3B), and Desire to Eat (Figure 3C). For Health, there was a short early window (time points 1–6) where those with greater SCSRs had a stronger positive association between Health and mouse angle. This indicates those with better overall self-control were more likely to show an initial movement toward the healthier/less tasty option. In contrast, both Taste (time points 44–100) and Desire to Eat (time points 47–100) showed significant interactions later in the decision-making process. Children with better overall self-control showed more negative associations between both Taste and Desire to Eat and mouse angle while children with worse overall self-control showed more positive associations. This indicates that starting around half-way through the decision-making process, children with better self-control were more likely to show movement away from the option rated higher on Taste and Desire to Eat while those with worse dietary self-control were more likely to be moving towards it.

Figure 3.

Effect of self-control success ratio (SCSR) on temporal dynamics during food choice. Figures depict the Time-Varying Effects Model (TVEM) results for SCSR across all trials. There are 2 Y-axes—Left Y-axis: TVEM β coefficient for the interaction between Attribute and SCSR, plotted in the black line. Shaded areas represent the 95% confidence interval for the interaction β coefficient; Right Y-axis: TVEM β coefficient for the association between Attribute and mouse angle estimated for each SCSR quantile (quantile 1: <42%, quantiles 2 & 3: 42–57%, and quantile 4: >57%), distinguished by line color. Horizontal bars on the X-axis indicate Time Points where the interaction β coefficient is significant with Attribute distinguished by color. A) TVEM for Health x SCSR. B) TVEM for Taste x SCSR. C) TVEM for Desire to Eat x SCSR.

3.5.2. Associations with Age

The TVEM approach did not show any moderation of rating differences by child age (supplemental Figure S3).

3.5.3. Associations with BMI percentile

There was a significant BMI percentile x Desire to Eat interaction (Figure 4A) such that children with lower BMI percentiles showed a stronger negative association between Desire to Eat and mouse angle at the end of the decision-making process (time points 95–100). This indicates that late in the decision-making process, children with lower BMI percentiles were more likely to be moving away from the item they rated as higher in Desire to Eat. BMI percentile did not significantly interact with Health or Taste (supplemental Figure S4).

Figure 4.

Effect of body mass index (BMI) percentile on temporal dynamics during food choice. Figures depict the Time-Varying Effects Model (TVEM) results for Desire to Eat x BMI percentile. There are 2 Y-axes—Left Y-axis: TVEM β coefficient for the interaction between Desire to Eat and BMI percentile, plotted in the black line. Shaded areas represent the 95% confidence interval for the interaction β coefficient; Right Y-axis: TVEM β coefficient for the association between Desire to Eat and mouse angle estimated for each BMI percentile quantile (quantile 1: ≤52^nd, quantiles 2 & 3: 52–94.9^th, and quantile 4: ≥95^th), distinguished by line color. Horizontal bars on the X-axis indicate Time Points where the interaction β coefficient is significant with Attribute distinguished by color. A) TVEM for all trials, B) TVEM for self-control trials.

3.6. Individual Differences in Temporal Dynamics During Self-Control

Since around a third of the children did not have STPs for self-control trials, individual differences in temporal dynamics were only examined using TVEM. Given the significant Self-Control (Successful, Unsuccessful) x Attribute interaction (see Results section 3.4), separate Attribute x Individual Characteristic models were run for Successful and Unsuccessful self-control trials.

3.5.1. Associations with Self-Control Success Ratio

There were significant interactions between SCSR and Desire to Eat for both Successful and Unsuccessful trials. For Successful trials, children with worse overall self-control had a more negative association between Desire to Eat and mouse angle starting about half-way through the decision-making process (time points 53–100; Figure 5A). In contrast, for Unsuccessful trials, children with better overall self-control had more positive associations between Desire to Eat and mouse angle at the end of the decision-making process (time points 83–97; Figure 5B). This indicates that when children were able to engage self-control and select the healthier option, those with better overall self-control were more likely to show movement away from the food they rated higher in Desire to Eat about halfway through the decision-making process. However, when self-control was unsuccessful those with better self-control were more likely to be moving toward the food rated higher in Desire to Eat at the end of the decision-making process.

Figure 5.

Effect of self-control success ratio (SCSR) on temporal dynamics during self-control. All figures depict the Time-Varying Effects Model (TVEM) results for Attribute x SCSR. There are 2 Y-axes—Left Y-axis: TVEM β coefficient for the interaction between Attribute and SCSR, plotted in the black line. Shaded areas represent the 95% confidence interval for the interaction β coefficient; Right Y-axis: TVEM β coefficient for the association between Attribute and mouse angle estimated for each SCSR percentile quantile (quantile 1: <42%, quantiles 2 & 3: 42–57%, and quantile 4: >57%), distinguished by line color. Horizontal bars on the X-axis indicate Time Points where the interaction β coefficient is significant with Attribute distinguished by color. A) Desire to Eat x SCSR for Successful Trials, B) Desire to Eat x SCSR for Unsuccessful trials, C) Health x SCSR for Unsuccessful trials, D) Taste x SCSR for Unsuccessful trials

SCSR moderated the associations between mouse angle and Health and Taste for Unsuccessful but not Successful self-control trials (supplemental Figure S5). During the middle of the decision-making process (Health: time points 33–70; Taste: time points 33–70) for Unsuccessful self-control trials, those with better overall self-control showed a less negative association between Health and mouse angle (Figure 7C), but a more positive association between Taste and mouse angle (Figure 7D). Together, this indicates that during trials where children were unsuccessful at selecting the healthier option, those with better overall self-control (i.e., higher SCSR) were more strongly influenced by Taste than Health.

3.5.2. Associations with Age

For Unsuccessful self-control trials, there were significant interactions between Age and both Taste (Figure 6A) and Desire to Eat (Figure 6B), but not Health (supplemental Figure S6). For both Taste (Time Points 86–100) and Desire to Eat (Time Points 87–100), older children showed more positive associations. This indicates that when self-control was unsuccessful, older children were more likely to show movement toward the option rated higher in Taste or Desire to Eat at the end of the decision-making process. Age did not moderate the association between mouse angle and Attribute for Successful self-control trials (supplemental Figure S6).

Figure 6.

Effect of Age on temporal dynamics during self-control. All figures depict the Time-Varying Effects Model (TVEM) results Attribute x Age for Unsuccessful self-control trials. There are 2 Y-axes—Left Y-axis: TVEM β coefficient for the interaction between Attribute and Age, plotted in the black line. Shaded areas represent the 95% confidence interval for the interaction β coefficient; Right Y-axis: TVEM β coefficient for the association between Attribute and mouse angle estimated for each Age percentile quantile (quantile 1: <8.5 years, quantiles 2 & 3: 8.5–10.5 years, and quantile 4: >10.5 years), distinguished by line color. Horizontal bars on the X-axis indicate Time Points where the interaction β coefficient is significant with Attribute distinguished by color. A) Taste x Age, B) Desire to Eat x Age.

3.5.3. Associations with BMI percentile

There was a significant BMI percentile x Taste interaction for Successful, but not Unsuccessful (supplemental Figure S7), self-control trials. Towards the end of the decision-making process (Time Points 66–87, Figure 9B), children with higher BMI percentiles showed a more negative association between Taste and mouse angle (Figure 4B). This indicates when self-control was successful, children with higher BMI percentiles were more likely to show movement away from the tastier option toward the end of the decision-making process. There were no significant interactions between BMI percentile and Health or Desire to Eat for Successful or Unsuccessful (supplemental Figure S7).

4. Discussion

This study advances the field by demonstrating that mouse tracking can be used to investigate how children make decisions when they are faced with two competing foods [17], a scenario that is more complex and ecologically valid than previous work. Further, by presenting children with two foods that differed by perceived health and hedonic properties, we were able to determine the characteristics of successful versus unsuccessful self-control, which may lead to more successful approaches to improving children’s diets. Overall, both the standard analytical approach used in adults [17] and the novel analytic technique, TVEM, revealed a relatively consistent pattern of results. Across all trials, there were no differences in when attributes impacted decision-making, however, when trials that required self-control were isolated, desire to eat impacted mouse trajectory earlier than health and taste. This suggests that relative to taste and health, desire to eat, as assessed in the current study, is a more salient feature guiding the early decision-making process for healthy foods. Additionally, TVEM allowed us to examine individual differences in when health, taste, and desire to eat influenced children’s food decisions. When the less healthy food was chosen (i.e., self-control failure), older children and those with better overall self-control (i.e., SCSR) more strongly incorporated taste and desire to eat (i.e., positive association) into their decision-making process. In contrast, when self-control was successful, children with higher BMI percentiles discounted taste more strongly (i.e., negative association). Together, this suggests that the influence of hedonic properties on the ability to make healthy food choices varies by individual child-level characteristics, which may help identify non-traditional targets for interventions aimed at improving child nutrition.

Across all trials, there was consistent evidence using both analytic approaches that health and taste influenced food choices about two-thirds of the way through the decision-making process, which differs from previous findings. Adults’ food choices are influenced by taste and desire to eat sooner than health [14,17] and children have been shown to rely more on taste than health when choosing what foods they would eat [10]. The lack of difference in the timing of attribute influences in the current study may be because the children were more responsive to authority and took a more literal interpretation of the researcher’s instructions to consider health when making food decisions compared to adults. Indeed, children chose the healthier/less tasty option around half the time (50.2%), while it was reported that adults only chose the healthier option 25% of the time [17]. Children have been shown to shift their choices and increase their consideration of health information when the importance of health is emphasized [33]. Similarly, adults with overweight shifted their incorporation of health information to be sooner in the decision-making process when calorie labels were added to the food images [14]. Together, this suggests that children may be more influenced by task instructions to consider health than adults, which may influence when taste and health information is incorporated into their food decisions.

In contrast to health and taste, the standard and TVEM approaches differed in their findings related to the influence of desire to eat. While the standard approach showed that all attributes influenced the decision-making process at a similar time point, TVEM showed that there was no effect of desire to eat on food decisions when all trials were considered. One potential reason for this difference is that TVEM included all children while the standard approach only included children with an STP. Overall, less than half of the children had an STP for any attribute, which could indicate that the influence of the attributes could not be captured by STPs or that they were not salient drivers of food decisions for all children. Since TVEM included all children and showed significant influences of health and taste, it is likely that the lack of STPs for these attributes was due to more dynamic patterns of influence that were not capture by STPs. Although more dynamic patterns could indicate greater variability in mouse trajectories due to poorer motor control in children, that does not seem to be the case in this sample because children did not show differences in performance when a keyboard was used to make choices (see supplementary materials). Since desire to eat was not significant in the TVEM approach but showed significant influence for a subset of children in the standard approach, it is possible the influence of desire to eat across all trials is more variable across children. Indeed, since desire to eat was assessed by the question “how much would you like to eat this at the end of the experiment”, the measure may have been less stable than taste and health.

When food choices during self-control trials were isolated, both approaches showed an earlier influence of desire to eat than taste or health, regardless of whether children were successful or unsuccessful in choosing the healthier food. Although this pattern was consistent across approaches, TVEM also showed that desire to eat influenced decision-making even earlier during successful compared with unsuccessful self-control. Desire to eat had a negative influence during successful self-control trials which suggests early discounting of the better liked food was required to select the healthier food. In contrast, desire to eat had a positive influence on mouse trajectory during the unsuccessful self-control trials, which suggests that desire to eat is a more salient driver of food choice when children are making selections for foods that are less healthy. Children’s desire to eat ratings have been found to be more predictive of food choice than health [33], therefore, the earlier influence of desire to eat when the healthier food was chosen suggests this attribute is more salient and needs to be discounted quickly in order for self-control to be exercised successfully.

When the less healthy food was chosen (i.e., unsuccessful self-control), TVEM revealed that children with worse overall self-control (indexed by SCSR) showed greater discounting of health while those with better overall self-control showed a stronger positive influence of taste and desire to eat. A negative influence of health during the middle of decision-making indicates that children were moving away from the healthier option, which when combined with the weak influence of taste suggests that rather than being drawn toward the tastier option, it is the food’s perceived health attributes driving the decision-making process. Since prior studies have found that children who perceive themselves to have lower self-control rate healthier foods as lower in taste [34] and we told the children they would have to eat one of the foods they selected, our findings may reflect a tendency to avoid selecting healthy foods they would not want to eat after the task. In contrast, the strong positive influence of a taste and desire to eat in those with better overall self-control suggests that when self-control failures did occur, they were more related to the hedonic value of foods. Together, our results indicate that children with poor overall self-control show a pattern of avoiding healthier food choices, while children with better overall self-control show individual self-control failures that are driven by the hedonic value of food.

A critical advance of the current study is that by using TVEM we were able to distinguish factors that were associated with children’s ability to both successfully and unsuccessfully choose the healthier food. For example, while BMI percentile was related to the influence of taste when selecting the healthier/less tasty food, age was related to the influence of taste and desire to eat when selecting the less healthy/tastier food. When choosing the healthier food, children with higher BMI percentiles showed a stronger negative influence of taste toward the end of the decision-making process which reflects greater discounting or movement away from the tastier food. The stronger negative influence of taste indicates that this attribute was more salient for children with higher BMI percentiles, which may be related to findings showing children with overweight and obesity have heightened sensitivity to food cues [35,36]. If children with higher BMI percentiles are more sensitive to perceived taste qualities of the food images, they may require stronger discounting of this attribute to choose the healthier/less tasty option. Conversely, when choosing the less healthy food, older children showed stronger positive influences of taste and desire to eat late in the decision-making process. A previous study in 8-to-13 year old children showed that relative to younger children, older children required greater cognitive effort to reject healthy foods, possibly due to greater attempts to engage self-control before ultimately making their decision [13]. Neither BMI percentile or age was related to SCSR in this study, which indicates that temporal dynamics of food choice may be more sensitive to individual differences in self-control than overall choice behavior. Self-control abilities develop through childhood and adolescence [37] and play an important role in the regulation of eating behavior and food choice [38]. Together, these findings highlight the potential utility of developing interventions to help children resist the influence of taste when trying to improve dietary choices. For example, inhibitory control training has shown short-term utility for helping both children and adults resist the selection of highly energy dense foods [39–42]. Additionally, these findings highlight the importance of child-level factors on the influence of attributes related to the hedonic rather than health value food.

Although this study advanced our understanding of food choice and self-control in children, there are limitations to the generalizability of the findings. First, due to the demographic characteristics of rural central Pennsylvania, this sample of children did not show substantial economic, racial, or ethnic diversity. As food preferences are influenced by early exposures and experiences [8], the temporal dynamics of self-control during food choice may vary by economic, racial, or cultural differences. Indeed, a qualitative study showed that while health and desire to eat are drivers of food choices, children see their parents as a source of information about their food decisions [43]. Additionally, overall self-control ability may also differ by these sample characteristics. Second, the specific protocol used in this study may limit the generalizability of findings. Children were instructed to fast three hours prior to participating in this study. While this creates an ecologically relevant state of hunger similar to that before a meal, findings may not reflect decision-making processes when children are sated or fasted for a greater amount of time. Additionally, this study assessed desire to eat by asking how much children would “like to eat” a food at the end of the session, which could be interpreted as an indicator of how much children “want” the food. Although this phrasing may help reduce the overlap in the measurement of desire to eat and taste, these hedonic attributes were still highly correlated. It is possible that different measurement approaches for desire to eat and taste would show different patterns of influence on children’s food choices. Additionally, it is possible that having the children rate the images prior to the choice task influenced their performance by changing their hunger state or the saliency of the images. Future studies are needed to determine if the temporal dynamics of food choice are influenced by the novelty food images. This study also specifically prompted children to consider health when making food choices. It is therefore unclear whether these findings generalize to all food choices or only when children are explicitly told to consider health. It is also unclear the extent to which children followed the prompt to consider health throughout the task, so we cannot be sure that children were actually exerting self-control during discordant food choices. Future work is needed to understand the influence of explicit direction on the temporal dynamics of food choice and dietary self-control in children. Third, although performance did not appear to differ when making choices using a keyboard rather than a computer mouse (see supplementary materials), it is possible that individual differences in motor control could impact mouse speed and trajectories. It is also possible that fatigue over the course of the task could result in changes in strategy or reduced focus on the health prompt. Lastly, this study did not assess whether overall self-control performance was related to actual food choices or intake in the lab. Testing whether individual differences in the temporal dynamics of successful and unsuccessful self-control relate to food choice and/or intake is an important step to understanding how to translate these findings to improving dietary quality among children.

5. Conclusion

This study showed that mouse tracking can be used to dynamically assess the influence of health, taste and desire to eat during food decision-making in children. TVEM produced consistent findings with the standard approach, but offered advantages in its ability to make use of the full sample of children to examine child-level factors that impacted both successful and unsuccessful self-control. In particular, discounting of hedonic attributes was critical for children to successfully choose the healthier food. This is important because it may provide insight into the underlying cognitive processes required for children to make healthy dietary choices. Future studies are needed to determine the impact of interventions such as inhibitory control training on the temporal dynamics of hedonic and health attributes during food choice. Mouse tracking tasks have the potential to provide insight into the cognitive mechanisms that support changes in food choices, which may help improve interventions to increase healthy food-choices and improve dietary patterns in children.