Fear, Peer Pressure, or Encouragement: Identifying Levers for Nudging Towards Healthier Food Choices in Multi-Cultural Singapore

: With roots beyond behavioural economics to psychology, nudges can be applied for influencing healthy behaviours such as food choice and portions to decrease obesity for better public health outcomes. However, the effectiveness of the type of nudges are contentious with conflicting literature. In this pilot study, we conducted a 23-day study surveying the food choices that included portion, locus of control, demographic data, and psychological measures of personality, perceived stress, narcissism, regulatory focus, food choice motive and dietary restraint, with the participants given four intervention conditions of 12 instant messaging sent every two days through WhatsApp. The messages were either factual (control), focused on consequences, through social comparison, or persuasive. Running over the COVID19 pandemic, 17 participants completed the full surveys show-ing significant effects between the experimental conditions with the psychological parameters ex-cept for diet confidence and extraversion and conscientiousness, as well as cognitive restraint. We found BMI and waistline measurements to be suitable measurements, with promising results from the fear and social comparison nudges for food-related behaviours and exercise. Our pilot findings have implications to the use of nudges upon which future studies investigating psychological factors can build on. participants. The study was conducted in compliance of prevailing ethical guidelines.


Introduction
The 2017 Nobel Prize in Economic Sciences was awarded to Richard H. Thaler for his contributions to behavioural economics. Thaler's research, built upon work that were subjects of the 1978 and 2002 prizes [1], highlighted the role of nudges. This was followed by decades of research and interest by governments all over the world, not surprisingly notably in the area of tax collection [2]. The notion of nudges has roots beyond behavioural economics, including psychology where Pavlovian conditioning [3], biopsychological habituation of sea slugs [4], and conscious and unconscious behavioural priming [5] have been continually pursued and incorporated into modern applications. These include the study of subliminal priming in social engineering effects, or lack thereof [6]. While controversial, new ideas for influencing behaviour such as the use of sonic devices to prevent teen loitering [7], or blue lights at Japanese train stations to deter suicide attempts [8] continue to emerge. Regardless of their aim or methodology, the science underlying nudging

Hypotheses Formation
Singapore is a good locale to study the effects of behavioural nudging on food health behaviour given her multicultural backdrop that can deeply influence dietary choices [50], and that nudges are already applied in areas such as taxation [51] and corporate processes [52]. In fact, Singapore is reportedly one of the most compliant countries to COVID-19 pandemic measures [53]. In the same spirit as these recent studies where behavioural nudges have shown some efficacy, the present study was conceived with the goal of shedding light on the prospects of applying behavioural nudge strategies on food choice, such as messages based on fear (of negative health consequences), constructive encouragement, and social pressure (such as social comparison), with these being delivered digitally over a smartphone. The hypotheses include: Hypothesis 1: Individuals receiving nudge messages (all types) would have lower body measurements after intervention than before while those receiving control messages would have no difference.
Hypothesis 2: Individuals receiving nudge messages (all types) would have improved food consumption behaviour after intervention than before while those receiving control messages would have no difference.
Hypothesis 3: Individuals receiving nudge messages based on constructive encouragement would have higher diet confidence and diet persistence after intervention than before while those receiving messages based on fear, social pressure, and those in the control condition would experience negligible differences.

Research Design
The study adopts a single-blind randomized controlled trial design with a control condition ("control") and three experimental conditions ("health consequences", "behaviour substitution", and "social comparison") corresponding to the three techniques (Figure 1). The outcomes were measured at three time points: pre-intervention (baseline), midintervention, and post-intervention ( Figure 2A). This design allowed us to detect significant changes in outcomes across the intervening time points and between the conditions. Ethics approval was granted by the Agency for Science, Technology, and Research (A*STAR), Singapore Institutional Review Board under number 2019-007 and the protocols and procedure adhered to for the whole study, with informed consent from the participants. The study was conducted in compliance of prevailing ethical guidelines.

Participants
Volunteers for the study were recruited between June 2019 and June 2020 through convenience sampling in various settings that included social media platforms. Festive periods (e.g., Chinese New Year, Ramadan) were avoided to prevent anomalous effects due to feasting or fasting. A priori power analysis of mixed-design ANOVA composing of 4 between-conditions and minimally 2 within-measurements with power (1-β) = 0.80, α = 0.05, and effect size = 0.25 was conducted using G*Power, suggesting a minimum sample size of 136. Accounting for a 50% participant dropout rate, the study aimed to recruit at least 250 participants with the following criteria: (a) ages 21 years and above, (b) healthy with no known medical condition, (c) familiar with and actively using WhatsApp as an instant messaging service, (e) currently working or residing in Singapore, (f) proficient in English language, and (g) not undergoing any existing medical-directed food or physical activity programmes. Participation was strictly voluntary, and subjects did not receive monetary reimbursements for their participation. This would help reduce the impact of selection biases.  Table S3 for the detailed list.

Nudge-Message Development
The concept behind intervention design utilizes the smartphone as the means for a direct message-based behavioural change intervention, with messages designed based on the Behavioural Change Wheel framework [54], and delivered at particular time points to participants. The goal of behaviour nudging in this case, was to promote a reduction of food portion sizes, as well as intake of sugary foods and drinks. Given its prevalent use in Singapore, the WhatsApp instant messaging service ("WhatsApp") was chosen as the medium for delivering these messages. Four sets of behavioural change messages comprising images or illustrations and text captions (Table S3) were designed, corresponding to three different behaviour change techniques [55] and a control condition. With a total of four study conditions, a total of 12 messages were developed ( Figure 2B). Each pair of image illustrations with accompanying text captions would present a different health or nutritional information snippet, contextualised according to the nudge techniques involved in the conditions they were based on, including neutral facts ("control"); scare tactics ("health consequences"); constructive encouragement ("behavioural substitution"); and peer pressure ("social comparison"). Specifically, messages in the "control" condition described the nutritional information of various food items as neutral facts method. Messages in the "health consequences" condition highlighted the health risks and negative consequences of not consuming smaller food portions, restricting sugar intake, or not exercising, as scare tactics. In contrast, message in the "behaviour substitution" condition applied constructive encouragement by suggesting actions or alternatives that would prompt individuals to consume smaller food portions, restrict sugar intake, or exercise more. Lastly, messages in the "social comparison" condition provided exemplars of physically fit and healthy individuals to prompt these social comparisons including comparison with peers.

Survey Measurements
Participants were asked to complete an online survey at pre-, mid-, and post-intervention time points after the initial informed consent (see Supplementary Table S1). The pre-intervention survey assessed participants' body measurements, food consumption behaviours, preferred food portion, sugary food and drinks intake, diet confidence and diet persistence, as well as internal locus of control with respect to health. It also gathered socio-demographic features and assessed various psychological constructs including personality, perceived stress, narcissism, regulatory focus, food choice motive, and dietary restraint. Both the mid-and post-intervention surveys contained similar sets of questions, but excluded the socio-demographic questions and psychological measures. These two surveys further gathered participants' perception of message validity. Body measurements were assessed through self-reported height, weight, and waistline measurements. The body mass index (BMI) was then calculated from the height and weight. Food consumption behaviours and preferences were assessed in two ways. Preferred portion sizes were assessed through a locally adapted food portion selection task [56], while a specific subscale of the dietary practice questionnaire [35] gathered self-reports on the frequency that sweetened drinks (e.g., tea and coffee), as well as desserts and snacks, were consumed in the past week. Diet confidence and diet persistence were assessed respectively through single-item questions, namely "If you are to go on a diet now, how confident are you that you will succeed?" and "If you are to go on a diet now, how long do you think you can sustain it?". Health internal locus of control was assessed with the corresponding subscale of the multidimensional health locus of control scale [57]. Lastly, validity for the nudge messages were assessed through single-item questions, namely "Do you agree that the messages sent to you during the study help you to manage your diet or eating habits" and "Do you agree that you feel healthier after the intervention?" respectively.
Socio-demographic characteristics surveyed include gender, ethnicity, age group, and household income level. Personality was assessed based on five factor model of personality with the Mini-IPIP scale [58]. Perceived stress was assessed with the Perceived Stress Scale [59]. Narcissism was assessed with the Single-Item Narcissism Scale [60]. Promotion and prevention regulatory focus was assessed with the Regulatory Focus Scale [61]. Food choice motive was assessed with the subscales for the health and weight factors from the Food Choice Questionnaire [62]. Lastly, dietary restraint was assessed with the uncontrolled eating, cognitive restraint, and emotional eating factors of the Thee Factor Eating Questionnaire [63].
Participants were also asked to take a photo of every meal that they had in a day and submit them through WhatsApp at the pre-and post-intervention time points (Supplementary Table S2). Their meal photos would be processed using the APD Areametric App [64] which would estimate the average food portion sizes of the meals that they had consumed in a day at pre-and post-intervention timepoints. This was used for the analysis of their actual food portion consumption.

Procedures
Recruited participants were asked to visit the online survey [46] for the research information sheet and to acknowledge it and the provide their informed consent before completing the pre-intervention survey. Mobile numbers were also collected for the sending of nudge messages via WhatsApp after the participants were assigned to the one of the four study conditions on a rolling basis.
Two days after a participant had signed up, a WhatsApp message would be sent to remind them to take a top-down photo of their meals in a day (with a Singapore $1 coin placed next to it for automatic area calibration by the APD Areametric app). For the next three days, an automated script would send messages to participants over WhatsApp at 7 a.m. local time daily, to remind them to submit photos of their meals.
Following the photo request, intervention messages (across the various conditions) were sent over 23 days in a similar manner. Every two days at 7 a.m., an automated script would send a nudge message over WhatsApp to the study participants based on their assigned study conditions. In this way, each participant would receive 12 nudge messages by the end of the intervention period. On Day 16, study participants were prompted to complete the mid-intervention survey online hosted on the same platform.
A day after the end of the intervention period, WhatsApp messages would be sent to prompt study participants to complete the post-intervention survey online, and to take the final meal photos of the meals the participants had, for that day. As with the earlier prompts, reminder messages for the photos would be sent for the next three days at 7 a.m. local time. On the last day, a WhatsApp message would be sent to the participants to announce the end of the study and thank them for their participation. No further messages would be sent beyond that point.  Table 1 summarizes the socio-demographic characteristics of the participants based on gender, ethnicity, age group and household income level by study conditions. 33 (47%) participants were female while 37 (53%) were male. Ethnically, 55 (79%) participants were Chinese while 8 (11%), 4 (6%) and 3 (4%) were Indian, Malay and others respectively. 17 (24%) participants were in the 21 -24 age group while 35 (50%), 12 (17%), 5 (7%) and 1 (1%) were in the 25 -34, 35 -44, 45 -54 and 55 -64 age group respectively. Lastly, 8 (11%) participants indicated a household income of ≤ $2,000 while 12 (17%), 12 (17%), 10 (14%), 10 (14%), 6 (9%) and 12 (17%) indicated a household income of $2,001 -$4,000, $4,001 -$6,000, $6,001 -$8,000, $8,001 -$10,000, $10,001 -$15,000 and >$15,000.  Table 2 presents the descriptive statistics of the outcome measures by study conditions and intervention time points. Additionally, the Cronbach alpha scores for the preferred food portion for consumption were 0.87, 0.89 and 0.94 for pre-, mid-, and postintervention respectively, while for the health internal locus of control they were 0.72, 0.86 and 0.83 for pre-, mid-, and post-intervention respectively. Table 3 presents the descriptive statistics of the psychological measures including their Cronbach alpha scores.

Hypothesis 1: Individuals receiving nudge messages (all types) would have lower body measurements after intervention than before while those receiving control messages would have no difference.
In terms of body measurements (Table 4), the mean BMI and waistline were lower after the intervention period, for both "control" and "social comparison" conditions than before the intervention period. However, the mean BMI was higher for the "health consequences" condition after the intervention, even though the mean waistline was lower after intervention. Curiously, for the "behaviour substitution" condition, both the mean BMI and waistlines were higher after intervention. A two-way MANOVA did not reveal any significant differences in BMI and waistline across intervention time points (Pillais' Trace = .02, F(4, 200) = 0.47, p > .05), and between the four study conditions (Pillais' Trace = .08, F(6, 200) = 1.34, p > .05) with no significant interaction effect, Pillais' Trace = .07, F(12, 200) = 0.86, p > .05.

Hypothesis 2: Individuals receiving nudge messages (all types) would have improved food consumption behaviour after intervention than before while those receiving control messages would have no difference.
It was observed ( Table 4) that, for the "control" condition, the mean preferred food portion for consumption, tea and coffee consumption, and desserts and snacks consumption were lower whereas the mean sweetened drinks consumption was higher after intervention. For the "health consequences" condition, the mean preferred food portion for consumption, sweetened drinks consumption, and tea and coffee consumption were lower whereas the mean desserts and snacks consumption remained unchanged after intervention. For the "behaviour substitution" condition, the mean preferred food portion, sweetened drinks consumption, and tea and coffee consumption were lower whereas the mean desserts and snacks consumption was higher after intervention. Lastly, for the "social comparison" condition, the mean preferred food portion and sweetened drinks consumption were both higher whereas the mean tea and coffee consumption, and desserts and snacks consumption were lower after intervention. A two-way MANOVA did not show significant main effect for study conditions, Pillais' Trace = .13, F(12, 282) = 1.06, p > .05, and intervention time points, Pillais' Trace = .02, F(8, 186) = 0.19, p > .05, while also did not show significant interaction effect, Pillais' Trace = .19, F(24, 380) = 0.80, p > .05.

Hypothesis 3: Individuals receiving nudge messages based on constructive encouragement would have higher diet confidence and diet persistence after intervention than before while those receiving messages based on fear, social pressure and those in the control condition would experience negligible differences.
For diet confidence and diet persistence (Table 4), the "control" condition participants had lower mean diet confidence and diet persistence after the intervention period than before, while in the case of the "health consequences", "behaviour substitution", and "social comparison" conditions, the mean diet confidence was higher whereas the mean diet persistence was lower (Table 4). A two-way MANOVA confirmed the significance of the main effect for the study conditions, with Pillais' Trace = .17, F(6, 206) = 3.23, p < .05, but there was no significant main effect between intervention time points, Pillais' Trace = .05, F(4, 206) = 1.26, p > .05, nor any significant interaction effect, Pillais' Trace = .09, F(12, 206) = 0.81, p > .05.
Participants in the "behaviour substitution" condition (i.e., constructive encouragement) reported the highest mean message helpfulness and health perception scores, followed by those in the "health consequences", "social comparison", and "control" conditions. A one-way MANOVA did not reveal significant differences between the study conditions for the two measures, Pillais' Trace = .39, F(6, 36) = 1.43, p < .05.

Control
Health consequences

Relationship between Specific Outcome and Psychological Measures
Pearson correlations were derived for specific pair of outcome and psychological measures at the pre-intervention time point (Table 5). Notably, for outcome measures, BMI had significant moderate positive correlation with the uncontrolled eating and emotional eating factors (Table 5). Preferred food portion had a significant but weak negative correlation between extraversion and uncontrolled eating and there was significant moderate negative correlation with agreeableness. Diet confidence had a significant, though weak, positive correlation with extraversion and conscientiousness as well as with diet persistence and had significant moderate positive correlation with health food choice motive. Lastly, the health internal locus of control had a significant but weak positive correlation with health and weight control food choice motive. For psychological measures, extraversion had a significant but weak positive relationship with uncontrolled eating. Health and weight control food choice motive have significant moderate positive relationship with cognitive restraint. Note: * p < .05, ** p < .01

Discussion
We sought to examine the effects of administering various types of nudges ("health consequences", "behaviour substitution", "social comparison") through WhatsApp on various diet and health related measurements in this ambitious study that included investigating the psychological factors and perception of the participants. While the G*Power calculation required a minimum of 250 participants, we were only able to recruit 70 over the late Oct 2019 to early 2021 period. There was a significant number of incomplete submissions, leading to only 17 completed participants. Many factors contributed to this from the numerous items in the surveys and the high demand of participant attention to the ongoing COVID19 pandemic measures interrupting lives and the change in WhatsApp policy in January 2021 that resulted in many users ending their use of WhatsApp for other platforms. This change of events impact our initial utility of the WhatsApp app for its widespread use and successful utility in education [65,66] and in healthcare [67]. From the limited 17 participants, we did not find any significant effects of nudge messages nor within the various types on body measurements of BMI nor waistlines, food consumption behaviours, diet confidence and diet persistence, or perceived message helpfulness. Nonetheless, an in-depth investigation of the collected data provided insights to the use of nudges for food-related behaviours that may guide future studies.
We found an interesting decrease of BMI for the control and social comparison conditions, but it was reversed for health consequences and behaviour substitution conditions, and there was a waistline decrease across all measures with few exceptions ( Table  2). This finding suggests that being made self-conscious of the food eaten alone was able to elicit some effort of control by the people with the exceptions. These exceptions include the waistline for the behavioural substitution and the preferred food portions, sweetened drinks, and health internal locus of the social comparison condition. Taking together the BMI increase for health consequences and behaviour substitution and the decrease of waistline for the health consequences, the possibility of weight gain could be due to muscle mass gain (assumed from decreased waistline) from increased exercise. What is surprising is that the overall increase of snacks, and decreased diet persistence of behaviour substitution suggested that this form of intervention may not yield the desired effect and that with the assumption of muscle mass, fear in health consequences may have encouraged exercise and muscle gain better than the control and social comparison condition.
What was interesting in the social comparison condition was that despite it having the best decrease in BMI of ~2.2, its waistline decrease of ~2.5 cm was less than the control of ~4 cm (Table 4), yet its increase in preferred food consumption, sweetened drinks and decreased diet persistence may suggest practices of starvation that caused some compensation given the decrease in snacking frequency and second largest decrease in food portion at ~-78.5 after behaviour substation at ~-87. Social comparison methods may, at least in the short term, be the most effective intervention (in terms of highest increase in internal locus of control as well) for getting people to eat less and decrease their BMI.
Despite perceived stress shown to be associated with unhealthy dietary habits [68,69] and weight gain [70], we did not find any correlation with BMI for the possible reason that the 21 days of this study was not likely to provide sufficient time for notable weight changes, as well as the small sample size. Even so, this was in agreement with some past studies [68,71] that did not find such correlations. In addition, the study showed uncontrolled and emotional eating to have significant positive correlations with BMI affirming that diet plays the major role in increased BMI. We further found a weak negative correlation with preferred food portion regardless of whether it was healthy or unhealthy food that was contrary to a past study reporting a positive correlation between uncontrolled eating and food portion size [72]. This discrepancy may be explained by a compensation of more frequent consumption despite choosing smaller portions by the those with uncontrolled eating.
On psychological parameters, we found extraversion to have a weak negative correlation with preferred food portion which may support the above observation on uncontrolled eating given that a past study [73] reported that high extraversion individuals may consume more sweet and savoury food and sugary beverages. Thus, extraverts may in fact, consume smaller portions of food but in higher frequency.

Limitations and Future Studies
Due to the many tasks in this study, there was a 68.6% drop-out rate impacting the significance of the findings. Only ~31.4% of the initial 70 participants completed all aspects of the study. Certainly, further investigations need to be utilized less parameters than what we have attempted here. Notably, given the lack of obvious associations of sociodemographic factors with the interventions, it may be possible to exclude these parameters as well as leaving out the need mid-point intervention given the general consistency with the end of intervention. Given that the control condition also saw a decrease, a better control that did not involve food at all may perhaps decrease awareness and reminders for a better baseline analysis. Given the problems with the instant messaging app, it may be possible that better convenience of data collection also in-built app nudges that also include steps taken and exercise logging e.g. APD Health Nudge app [74].

Conclusion
Our findings in this pilot study suggest the mere awareness of food intake can have an effect in eating habits with potential effectiveness in the use of fear and social comparison nudges for food-related behaviours and exercise across the multi-cultural background while reaffirming the suitability of BMI together with waistlines in more objective measurements, even within the 23-day experiment. Certainly, the use of apps is feasible, although further adaptions for the ease of use would reduce participation drop-off.
Supplementary Materials: The following are available online at www.mdpi.com/xxx/s1, Table S1: Survey Instructions and Question Items (Data Collection Form), Table S2: Study Procedure Instructions, Table S3  Informed Consent Statement: Informed consent was obtained from all subjects involved in the study.

Data Availability Statement:
The data presented in this study are available on request from the corresponding author.