Social norms and obesity prevalence: From cohort to system dynamics models

Summary Group‐level obesity can be seen as an emergent property of a complex system, consisting of feedback loops between individual body weight perception, individual weight‐related behaviour and group‐level social norms (a product of group‐level ‘normal' body mass index (BMI) and sociocultural ‘ideal' BMI). As overweight becomes normal, the norm might be counteracting health awareness in shaping individual weight‐related behaviour. System dynamics modelling facilitates understanding and simulating this system's emergent behaviour. We constructed six system dynamics models (SDMs) based on an expert‐informed causal loop diagram and data from six sociocultural groups (Dutch, Moroccan and South‐Asian Surinamese men and women). The SDMs served to explore the effect of three scenarios on group‐level BMI: ‘what if' weight‐related behaviour were driven by (1) health awareness, (2) norms or (3) a combination of the two. Median BMI decreased approximately 50% and 30% less in scenarios 2 and 3, respectively, than in 1. In men, the drop in BMI was approximately two times larger in scenario 1 versus 3, whereas in women, the drop was approximately equal in these scenarios. This study indicates that the overweight norm in men holds group‐level BMI close to overweight despite health awareness. Since norms are counteracting health awareness less strongly in women, other drivers of obesity must be more relevant.

from its parts and is more than a direct outcome of their aggregation. 4 In addition, a complex system shows non-linear behaviour, with a large disproportionality between its inputs and outputs, 5 which often results from the presence of feedback loops.
Previous research has indicated that norms affect individual dietary intake and physical activity. It has been shown that it is problematic for individuals to change their weight-related behaviour if others in their environment do not. 6 In fact, it has consistently been shown that weight loss resulting from interventions targeting individual weight-related behaviour is likely to not be maintained. 7 One possible explanation is that norms may counteract health awareness in shaping individual health-related behaviour. In that regard, the relevance of addressing norms to change individual health-related behaviour as a policy strategy has received attention with respect to smoking. 8 That is, 'the shifts in social norms (…) that occur as a result of major tobacco control interventions and campaigns can strengthen smokers' motivation to quit and commitment to staying quit'. 9 Understanding the effect of the feedback loops between individual characteristics and group-level processes over time can guide similar policy strategies aiming to lower obesity prevalence, in which we do justice to influences at both levels.
System dynamics models (SDMs) can aid in understanding and simulating this system's emergent behaviour. 10 SDMs can reveal how variables interact by expressing the causal links between them using difference equations. Specifically, system dynamics modelling is used to understand and simulate a complex system's non-linear behaviour in different scenarios. 10 This is valuable when evaluating the effect of 'what if' scenarios is unfeasible using conventional empirical methods, as holds for comparing the effect of health awareness versus norms on group-level body mass index (BMI).
In this study, we model the system of social norms regarding body weight perception and obesity prevalence using SDMs. These SDMs are designed to test the hypothesis that as overweight becomes normal, the norm might be counteracting health awareness in shaping individual weight-related behaviour. Hereto, we design an expertinformed causal loop diagram (CLD) to conceptualize this system, which we subsequently use as a template for functioning SDMs. We consider this system as consisting of feedback loops between individual body weight perception, individual weight-related behaviour and group-level norms towards body weight. These norms can be conceptualized as a product of social comparison and cultural preferences: 3 what is considered normal, i.e. group-level BMI, and ideal, i.e. sociocultural ideal BMI. The SDMs simulate this system's behaviour for a multi-ethnic cohort in Amsterdam (registered in the HELIUS study) 11 -for whom body weight perception was measured and whose data we use to tune the SDMs.
We present this-as a proof-of-concept of studying feedback loops between individual characteristics and group-level processes using this methodology-with the aim to simulate the effect of three scenarios on group-level BMI. These reflect the question 'what if' weight-related behaviour were driven (1) only by health awareness, (2) only by norms and (3) by their interaction, i.e. health awareness and norms.

| METHODS
In this section, we describe (1) the formulation of the expertinformed CLD, (2) the study population that we select from the Amsterdam-based cohort whose data can be used to quantify this CLD and (3) body weight perception as it was measured for this study population. We then explain (4) the conversion of the CLD to six SDMs using stocks, flows, auxiliaries and constants and (5) the variables and equations that we use in these SDMs. Lastly, we describe (6) the use of cross-sectional data for the quantification of the SDMs, (7) the use of validation statements to validate this quantification and (8) the scenarios we test using the validated SDMs.

| Causal loop diagram
We start by formulating an expert-informed CLD of the system, The CLD shows that an increase in individual BMI increases group-level BMI (CL1), 3 that is, what is normal, which in turn drives up the norm (CL2). 3 Sociocultural ideal BMI, that is, what is ideal, also affects the norm (CL3) 3 and represents the relatively stable sociocultural perception an individual has of the ideal BMI in their group, which induces an individual variation in how the norm is regarded. Individual ideal BMI, representing the BMI an individual strives for, is then driven both by the norm (CL4) 12 and by knowledge of what a healthy BMI is (CL5), 12 creating a possible conflict between these two influences. This sequence leading to individual ideal BMI drives two feedback loops in parallel (R1 and R2). In R1, an increase in individual ideal BMI increases food intake (total daily energy intake (TDEI)) (CL6), [13][14][15][16] which then increases BMI (CL7). 17 In R2, this increase in individual ideal BMI causes a decrease in physical activity level (PAL) (CL8) 14,18,19 and consequently in total daily energy expenditure (TDEE) (CL9), 20 which similarly increases BMI (CL10). 17

| Study population
We then convert the CLD to SDMs, which requires data that can quantitatively represent the CLD's variables. Therefore, we turn to the multi-ethnic HELIUS study, 11 in which data on characteristics reflecting these variables were collected.
We construct six SDMs for three ethnic groups from this cohort, as we hypothesize that norms influence individuals to a different extent in cultures that contrast concerning the degree of individualism/collectivism. We define the positions of these cultures on the dimension of individualism/collectivism as previously identified based on the Hofstede model, a well-established paradigm for comparing cultures. 22 Here, collectivism, versus its opposite individualism, is 'the degree to which people in a society are integrated into groups' . 22 We accordingly construct SDMs for three groups that differ in the degree of individualism/collectivism, which we then stratify by gender-as we expect the influence of norms on women to be greater 23 -resulting in six groups. These are Dutch ('individualistic'), 24 Moroccan ('collectivistic') 24 and South-Asian Surinamese (Suriname-'slightly collectivistic' 24 ; South-Asian ancestry from India 25 -'both collectivistic and individualistic traits') 24 men and women. We confirm the diversity in body weight perception that underlies this stratification with statistical analyses (Appendix S1.1). Constructing an SDM for each of these groups implies that we assume that, specifically for social comparisons of body weight, a member of one group is unaffected by the individuals in other groups. This assumption is based on the similarity hypothesis, which is the idea that 'comparison with targets that are close on a variety of dimensions, such as age, gender, or family ties, have a greater affective impact' . 3 Thus, we expect that individuals are most likely to compare themselves to others from their own sociocultural group.
The data include characteristics concerning sex (which we use as a proxy for gender), age, ethnicity, education, BMI and body weight perception for each individual. Details on their collection are given in Appendix S1. 1

| Body weight perception
For this cohort, perceived BMI was determined via questionnaires, asking each individual to indicate which image they most looked like on a randomly ordered version of the body image scale developed by Pulvers et al. 26 (Figure 2).
For all groups we map each image to a corresponding measured BMI, where each image is represented by the average BMI for all individuals that selected that image as their perceived BMI (Table 1).
In addition, each individual was asked which image they would prefer to look like and which image they thought others in their environment would find most attractive. We used the answers to these two questions to quantitatively represent individual ideal BMI and sociocultural ideal BMI, respectively, based on the abovementioned calibration (visual representation in Appendix S1.2).

| Stocks, flows, auxiliaries and constants
Using these data, we then convert the CLD into six SDMs. The structure of the SDMs is considered representative for each of the groups, but quantification of the strengths of the causal links for each SDM is dependent on each group's data. From here onwards, the conceptual variables in the CLD are defined as quantifiable variables based on the cohort data (see Section 2.5).
The SDMs' structure ( Figure 3) is designed to mirror the CLD but contains a number of additional variables (e.g. 'Age' and 'Height') that are necessary to meet the mathematical and structural SDM requirements. These requirements have been described in detail elsewhere. [27][28][29][30][31] To facilitate the conversion of the CLD into six SDMs, we define which of the CLD's variables will be stocks, flows, auxiliaries and constants. A stock represents a 'concrete aspect […] of the system that can be seen and measured' ('Weight' in Figure 3), whereas a flow is connected to a stock and determines how it changes over time-it can be perceived as a rate (e.g. 'Weight loss rate'). 31 An auxiliary is 'any dynamic variable that is computed from other variables at a given time' (e.g. 'BMI'), whereas a constant is a variable that does not change during the simulation (e.g., 'Height'). 32 F I G U R E 1 Causal loop diagram (CLD) of the system of social norms regarding body weight perception and obesity prevalence. Variables are connected by arrows indicating causal links, where a plus indicates that an increase in the variable at the tail of the arrow constitutes an increase in the variable at the head of the arrow whereas a minus indicates that an increase in the variable at the tail constitutes a decrease in the variable at the head. Reinforcing feedback loops R1 and R2 and balancing feedback loop B1 are indicated with loop symbols Perceived body mass index (BMI) (body image scale) in kg/m 2 corresponding to average measured BMI per group

| Variables and equations
The variables are formulated based on the cohort data ( Table 2) and are displayed in italics in the remainder of Section 2.
We use a list of equations (Table 3), representing the SDMs' structure, that takes the values corresponding to all variables in Table 2 for a given individual as inputs and generates their values for the next time point as outputs. These equations should be solved in the order as defined in Figure 3 (starting from IIB (individual ideal BMI) and following the causal links).
These equations ensure that dimensional consistency is preserved within the SDMs, that is, the units between the left-and the righthand side of all equations are consistent. [27][28][29][30][31] The body weight perception variables and their updating procedures and the fitted parameter (Equations 1, 2 and 10 as listed in Note: Gender and ethnicity are not technically variables included in the SDMs but are shown as we use them in the development of the SDMs. The values for MedBMI, the optimized parameters and the fitted parameter are the same for all individuals in a given group (as explained in this section and in Section 2.6). We treat Age as a constant. PAL has no unit: It is expressed as TDEE/BMR (see Appendix S1.3). 21 Impact HBonIIB also has no unit, it is expressed as BMI/BMI (as explained in this section).
hypothesize that an individual might be inclined to alter their BMI based on the frequency of individuals in their group with a different BMI. The average is considered unsuitable as it would also be affected by a relatively small fraction of the group with a different BMI, which we do not suppose would change the behaviour of most individuals.
We use the initial value for SCIB as a constant, which via the calibration explained in Section 2.3 could be derived from the cohort data.
We use this initial value for SCIB as a constant to introduce individual variation to the norm, as it represents the sociocultural perception an individual has of the ideal BMI in their group, which we consider to be relatively stable. Note that we choose to let both MedBMI and SCIB exert an equal influence on Norm, that is, the most conservative ratio of 50/50, as we are unaware of research that has specifically addressed this issue.
We form Equation 1 by rewriting the output of regression through the origin between distance Norm and HB (the independent variable, x), represented by and distance Norm and IIB (the dependent variable, y), that is, which we conduct for each group. This provides us with the value for Impact HBonIIB for each group, given by which we can rewrite to get Equation 1. We thus determine the value for Impact HBonIIB directly from the cohort data, where the interpretation of Equation 1 is that IIB is a weighted average of HB and Norm where Impact HBonIIB determines the relative weight of each. A larger Impact HBonIIB implies a larger effect of HB on IIB and, as a consequence, a lower effect of Norm on IIB. This is based on the assumptions that (1) if Norm equals HB, then IIB equals HB; implying individuals prefer to be healthy and (2) HB and Norm work against each other. The values for Impact HBonIIB for each group are given in Table 4.
Impact HBonIIB can also be regarded as uniquely determining an individual's behaviour with respect to HB and Norm, which we show in Appendix S1.3.

T A B L E 3
The list of equations that represents the system dynamics models (SDMs) Women 52 (24) 39 (18) 46 (23) 51 (20) 33 (18) 48 (18) BMI (kg/m 2 ), median (IQR) 25 We provide both the median and the mean for IIB and SCIB as providing the median only for these variables makes them appear identical. b Here, discrepancy is calculated between perceived BMI (obtained by calibration of measured BMI to the body image scale) and IIB so that both are estimated according to the body image scale to allow for direct comparison. Therefore, if discrepancy < 0, individuals think they should gain weight; if discrepancy = 0, individuals are satisfied with their weight; if discrepancy > 0, individuals think they should lose weight.
Descriptive characteristics of the study population at t = 0 corresponding to the SDMs' variables are presented in Table 5.

| Cross-sectional data
The data we have are cross-sectional; however, to construct operational SDMs, we require data from multiple time points. We therefore develop a new way to generate pseudo-time series data from the available cross-sectional data by generating a set of qualitative 'data-generating assumptions'. A summary of alternative methods and a detailed explanation of our method are given in Appendix S1.4. These assumptions are based on the system's temporal behaviour that is expected to exist across all groups, for example, that, on average, an individual can lose 2 kg/month. 35,36 Here, we assume linear dynamics for the system's short-term behaviour. We fit the time steps in the SDMs as months in this study to identify an approximate timescale of the system's behaviour. However, this timescale is not exact, and therefore, the relative trends are more important to interpret than the exact timescale they occur on.
We optimize the values for the parameters Intent PAB and Intent EB such that the SDMs reproduce the data-generating assumptions and fit the cohort data. Intent PAB and Intent EB represent the change that individuals make in physical activity behaviour (changing their PAL) and eating behaviour (changing their TDEI), respectively, based on how much their BMI differs from their IIB (captured in Discrepancy BMIandIIB ). We use the optimization method basin-hopping to find the optimal values for these parameters. 37 individually and comparing it with available knowledge about the real system' . 39 We address direct structure by basing the CLD on literature, using equations that were previously defined and validated elsewhere where possible and maintaining dimensional consistency in the SDMs.
As structure-oriented behaviour tests, which 'assess the validity of the structure indirectly, by applying certain behaviour tests on model-generated behavior patterns' using simulation, 39 we generate 32 qualitative validation statements derived from expert knowledge and literature (see Appendix S1.5 for references and further details). The majority of the literature behind the statements is based on previous empirical research that addresses the Dutch population in particular, which adds to the internal validity of the statements. These statements contain comparisons we make concerning how the SDMs' behaviours are expected to differ among the groups. We hypothesize for instance that norms tend to have a larger effect in the Moroccan than the Dutch, as Moroccan culture is regarded as being more collectivistic than Dutch culture. 24 We expect a representative set of SDMs to reproduce a considerable number of these validation statements. A set can obtain a validation score out of 32 points (one for each statement), representing its quality.

| Scenarios
The IIB t = Norm t : For driven by health awareness and norms combined, we leave Equation 1 as it is, with Impact HBonIIB as estimated from the cohort data using regression through the origin (Table 4), where IIB is dependent on both HB and Norm: IIB t = Impact HBonIIB × HB + 1− Impact HBonIIB ð Þ × Norm t :

| RESULTS
We divide the results in three sections corresponding to the results of the (1) optimization, (2) validation and (3) scenarios.

| Validation
We obtain a validation score of 20/32 points (62.5% correct) for the set of SDMs. This implies that the set of equations that we use is capable of capturing the expected behaviour to a large extent. Table 7 shows for each behaviour whether it is exhibited by the respective SDMs.

| Scenarios
For all groups (Figure 4), we see that if weight-related behaviour were driven by a combination of health awareness and norms (green), then group-level median BMI would be lower than if driven only by norms (blue) and higher than if driven only by health awareness (red). If driven only by health awareness (red), there would only be minor differences between the male and female groups: median BMI would drop 11% (3 BMI points) among groups of both genders, which is trivial given the definition of this scenario.
For the other scenarios, that is, if driven only by norms (blue) and by health awareness and norms combined (green), we see different emergent behaviours between the male versus female groups, while they demonstrate only slight differences between groups of the same gender.
These small quantitative differences (given in Appendix S2.2) are not as reliable as the large qualitative differences that our simulations show, given the cross-sectional data that we use. We therefore mostly focus on the qualitative differences, that is, those between genders. There are no significant physiological differences between Dutch and South-Asian Surinamese men regarding the effect of physical activity on weight loss.
Yes: In the male groups, the effect of PAL on BMI is of the same order of magnitude in the Dutch as the South-Asian Surinamese group.

2
There are no significant physiological differences between Dutch and Moroccan men regarding the effect of physical activity on weight loss.
Yes: In the male groups, the effect of PAL on BMI is of the same order of magnitude in the Dutch as the Moroccan group.

3
There are no significant physiological differences between South-Asian Surinamese and Moroccan men regarding the effect of physical activity on weight loss.
Yes: In the male groups, the effect of PAL on BMI is of the same order of magnitude in the South-Asian Surinamese as the Moroccan group.

4
There are no significant physiological differences between Dutch and South-Asian Surinamese women regarding the effect of physical activity on weight loss.
Yes: In the female groups, the effect of PAL on BMI is of the same order of magnitude in the Dutch as the South-Asian Surinamese group.

5
There are no significant physiological differences between Dutch and Moroccan women regarding the effect of physical activity on weight loss.
Yes: In the female groups, the effect of PAL on BMI is of the same order of magnitude in the Dutch as the Moroccan group. 6 There are no significant physiological differences between South-Asian Surinamese and Moroccan women regarding the effect of physical activity on weight loss.
Yes: In the female groups, the effect of PAL on BMI is of the same order of magnitude in the South-Asian Surinamese as the Moroccan group.

7
There are no significant physiological differences between Dutch men and women regarding the effect of physical activity on weight loss.
Yes: In the Dutch group, the effect of PAL on BMI is of the same order of magnitude in the male as in the female group.

8
There are no significant physiological differences between South-Asian Surinamese men and women regarding the effect of physical activity on weight loss.
Yes: In the South-Asian Surinamese group, the effect of PAL on BMI is of the same order of magnitude in the male as in the female group.

9
There are no significant physiological differences between Moroccan men and women regarding the effect of physical activity on weight loss.
Yes: In the Moroccan group, the effect of PAL on BMI is of the same order of magnitude in the male as in the female group.
10 Norms tend to have a larger effect on Moroccan than on Dutch men, as Moroccan culture is regarded as being more collectivistic than Dutch culture.
Yes: In the male groups, (1-Impact HBonIIB ), representing the impact of Norm on IIB, is higher in the Moroccan than in the Dutch group.

11
Norms tend to have a larger effect on South-Asian Surinamese than on Dutch men, as South-Asian Surinamese culture is regarded as being more collectivistic than Dutch culture.
No: In the male groups, (1-Impact HBonIIB ), representing the impact of Norm on IIB, is not higher in the South-Asian Surinamese than in the Dutch group.

12
Norms tend to have a larger effect on Moroccan than on South-Asian Surinamese men, as Moroccan culture is regarded as being more collectivistic than South-Asian Surinamese culture.
Yes: In the male groups, (1-Impact HBonIIB ), representing the impact of Norm on IIB, is higher in the Moroccan than in the South-Asian Surinamese group. 14 Norms tend to have a larger effect on South-Asian Surinamese than on Dutch women, as South-Asian Surinamese culture is regarded as being more collectivistic than Dutch culture.
No: In the female groups, (1-Impact HBonIIB ), representing the impact of Norm on IIB, is not higher in the South-Asian Surinamese than in the Dutch group.

15
Norms tend to have a larger effect on Moroccan than on South-Asian Surinamese women, as Moroccan culture is regarded as being more collectivistic than South-Asian Surinamese culture.
Yes: In the female groups, (1-Impact HBonIIB ), representing the impact of Norm on IIB, is higher in the Moroccan than in the South-Asian Surinamese group. 16 Dutch women tend to be more prone to be influenced by norms than Dutch men.
No: In the Dutch group, (1-Impact HBonIIB ), representing the impact of Norm on IIB, is not higher in the female than in the male group.

17
South-Asian Surinamese women tend to be more prone to be influenced by norms than South-Asian Surinamese men.
No: In the south-Asian Surinamese group, (1-Impact HBonIIB ), representing the impact of Norm on IIB, is not higher in the female than in male group.

18
Moroccan women tend to be more prone to be influenced by norms than Moroccan men.
No: In the Moroccan group, (1-Impact HBonIIB ), representing the impact of Norm on IIB, is not higher in the female than in the male group.

19
The intent to change physical activity behaviour tends to be greater in Dutch than in Moroccan men, as physical activity is more embedded in Dutch than in Moroccan culture.
No: In the male groups, Intent PAB is not higher in the Dutch than in the Moroccan group.

20
The intent to change physical activity behaviour tends to be greater in Dutch than in South-Asian Surinamese men, as physical activity is more embedded in Dutch than in South-Asian Surinamese culture.
Yes: In the male groups, Intent PAB is higher in the Dutch than in the South-Asian Surinamese group.

21
The intent to change physical activity behaviour tends to be greater in Dutch than in Moroccan women, as physical activity is more embedded in Dutch than in Moroccan culture.
No: In the female groups, Intent PAB is not higher in the Dutch than in the Moroccan group.

22
The intent to change physical activity behaviour tends to be greater in Dutch than in South-Asian Surinamese women, as physical activity is more embedded in Dutch than in South-Asian Surinamese culture.
Yes: In the female groups, Intent PAB is higher in the Dutch than in the South-Asian Surinamese group.

23
The intent to change physical activity behaviour tends to be greater in Dutch women than in Dutch men.
Yes: In the Dutch group, Intent PAB is higher in the female than in the male group.

24
The intent to change physical activity behaviour tends to be greater in South-Asian Surinamese men than in South-Asian Surinamese women.
No: In the South-Asian Surinamese group, Intent PAB is not higher in the male than in the female group.
are counteracting health awareness in shaping this behaviour. Particularly in groups where overweight is the norm, that is, all male groups, the norm holds group-level BMI close to overweight despite of health awareness.
These results should be seen in light of a number of methodological considerations. We opted for SDMs as they allow us to assume that within a group all individuals compare themselves to that group equivalently, that is, their interactions are homogeneous. We did not have the data necessary to model possible heterogeneity in behaviour nor in social connectivity, as would be required to make, for example, an agent-based modelling approach 40 -with diverse interactions of unique agents with others and with their environment-valuable. We also assumed that there is no interaction between the sociocultural groups with respect to this system, implying that individuals are most likely to compare themselves to others from their own sociocultural group. Although we cannot rule out that there is interaction affecting this system, the similarity hypothesis that we based our assumption on is well-established. 3 In our SDMs, individuals compare themselves to their entire sociocultural group, whereas this comparison group might actually be more confined. Stratifying by other characteristics that also affect body weight perception and the influence of norms in addition to gender and ethnicity, for example, age, 41 would result in smaller comparison groups. Zooming in on these subgroups could reveal additional emergent behaviours, which we did not detect in this study as we focused on the current groups' average dynamics.
In addition, we assumed that median BMI and sociocultural ideal BMI exert an equal influence on the norm, as the most conservative ratio, which might not be accurate. We also assumed that individuals adjust their weight-related behaviour to narrow the gap between their actual and ideal BMI and that their behaviour is driven by health T A B L E 7 (Continued)

Validation statement
Operationalization: Behaviour exhibited by the SDMs?

25
The intent to change physical activity behaviour tends to be greater in Moroccan men than in Moroccan women.
No: In the Moroccan group, Intent PAB is not higher in the male than in the female group.

26
The intent to change eating behaviour tends to be greater in Dutch than in Moroccan men, as eating behaviour is more important in Moroccan than in Dutch culture.
No: In the male groups, Intent EB is not higher in the Dutch than in the Moroccan group.

27
The intent to change eating behaviour tends to be greater in Dutch than in South-Asian Surinamese men, as eating behaviour is more important in South-Asian Surinamese than in Dutch culture.
Yes: In the male groups, Intent EB is higher in the Dutch than in the South-Asian Surinamese group.

28
The intent to change eating behaviour tends to be greater in Dutch than in Moroccan men, as eating behaviour is more important in Moroccan than in Dutch culture.
No: In the female groups, Intent EB is not higher in the Dutch than in the Moroccan group.

29
The intent to change eating behaviour tends to be greater in Dutch than in South-Asian Surinamese women, as eating behaviour is more important in South-Asian Surinamese than in Dutch culture.
Yes: In the female groups, Intent EB is higher in the Dutch than in the South-Asian Surinamese group.

30
The intent to change eating behaviour tends to be greater in Dutch women than in Dutch men.
Yes: In the Dutch group, Intent EB is higher in the female than in the male group.

31
The intent to change eating behaviour tends to be greater in Moroccan women than in Moroccan men.
Yes: In the Moroccan group, Intent EB is higher in the female than in the male group.

32
The intent to change eating behaviour tends to be greater in South-Asian Surinamese women than in South-Asian Surinamese men.
Yes: In the South-Asian Surinamese group, Intent EB is higher in the female than in the male group.
Note: Each validation statement is supported by 1-8 references, as we show in Appendix S1.5. In the right column, we show whether the behaviour is exhibited by the respective pair of SDMs based on the computational operationalization of the validation statement (see Appendix S1.5). Abbreviation: SDM, system dynamics model. awareness and norms. Our assumption implies that these are the only determinants of weight change, which is a simplification of reality. We deliberately kept the scope of the SDMs limited, omitting other drivers of weight-related behaviour. This is related to the proof-ofconcept function of this study, where we intended to demonstrate the methods required to design functioning SDMs with the aim to study feedback loops relevant to the field of public health. The SDMs therefore include only those factors that we deem necessary and sufficient to describe this system as the aim is to understand the implications of its feedback loops on the long term rather than to propose interventions or make predictions. In reality, broader environmental variables as well as individual characteristics such as education and income could affect whether individuals are willing and able to adjust their behaviour. Given that the aim of our SDMs was to unravel a specific causal mechanism and its long-term effects at the group level, we did not deem it essential to take individual physiological factors or genetic predispositions into account. Nevertheless, we acknowledge that these factors have a role to play in weight loss and weight gain at the individual level. We argue, however, that norms and their impact Our SDMs may serve as seed models for iterative extensions to include additional variables, which would make weight-related behaviour dependent on more determinants, possibly introducing more divergence in emergent behaviour. We however considered it important to understand the impact of norms in isolation before compounding additional causal mechanisms into the same SDMs.
As the data were cross-sectional, we used expert-informed datagenerating assumptions to optimize the values for the parameters for intent to change physical activity and eating behaviour, respectively.
This means that in each group, we based our simulations on the same hypothesized fundamental behaviours, as supported by literature, with respect to weight gain/loss, for example, concerning physiological constraints (rate of weight change). These fundamental behaviours however do not markedly influence the differences we observe between the groups over time. These differences are primarily based would. This implies that norms are counteracting health awareness less strongly in women, suggesting that they must be subject to additional drivers of obesity. The major driver of this divergence in behaviour is the norm, which is closer to the healthy weight range in women. This is in line with previous research showing that women are generally more engaged in health and weight-related behaviour. 46,47 Previous research has also shown 'that appearance norms encountered by women in daily life are more rigid, homogeneous and pervasive than those for men, and that more messages implying the attainability of the ideal appearance are directed at women' . 48 For women, having this pressure to conform to the sociocultural ideal BMI might add to the discrepancy between the results of the scenarios and reality. That is, despite of being aware of the sociocultural ideal BMI and the sociocultural ideal BMI being healthy, which under our assumptions should lead to weight loss in women, we see in reality that, at the group level, women are not losing weight. In this regard, it has for instance been shown that weight stigma increased the inclination to avoid exercise, independent of BMI and body satisfaction, among college-aged females. 49 In addition, under our assumptions, the cohort data underlying our SDMs suggest that the relative impact of norms on individual ideal BMI is larger among men. This is of interest, as we usually consider the effect of norms on weight change to be indicative of their strength. That is, norms towards body weight are generally regarded as being stronger in women, as women more likely to want to adjust their weight based on norms. 50 Our results however show that norms are stronger in men. This suggests that even though for men norms do not contribute to weight loss, they do establish the reinforcing feedback loops of group-level obesity. As the cohort data also show that the norm is approximately equal to group-level BMI in men, this might indicate that what is normal may be more important than what is ideal in determining what body weight men strive for. This is supported by findings from previous research showing that men are less likely to recognize their own overweight status and are more satisfied with their body weight than women. 51,52 Secondly, this study adds evidence that efforts to address the obesity epidemic need to consider norms, as an example of grouplevel processes. Our results indicate that in populations where overweight is the norm, the potential impact of policies is greater if they address these norms, as compared with policies that use individual- Studying 'what if' scenarios using computational modelling approaches allows us to test policy strategies under the exact same conditions. Although the SDMs do not include potential solutions for influencing norms, they show that we should do justice to influences on both the individual as well as the group level. Our results suggest that norms limit the effectiveness of interventions targeting individual weight-related behaviour, especially in men. A restricted focus on individual weight-related behaviour, which places the responsibility for body weight on the individual in a context where overweight is the norm, might even have adverse effects. 53 The results of this study indicate that shifting the focus to group-level interventions aiming to change the norm can contribute to the prevention of group-level obesity. A starting point for this can be considering how the effects of policy strategies may diverge among groups based on the prevailing norm towards body weight. Given that in our study population overweight prevalence is high in both men and women, the results of our scenarios imply that failing to consider the norm and its effect over time might result in a similar policy strategy for both genders, which would be misguided. Our SDMs can be instrumental in considering the norms of groups that, besides gender and ethnicity, differ with respect to characteristics such as socio-economic status. This can be explored in future research using these SDMs and would allow us to differentiate policy based on the group-level norm towards body weight, in addition to obesity prevalence.

| CONCLUSIONS
We showed that norms regarding body weight perception withhold individuals from losing as much weight as they could if driven by health awareness alone. When overweight is the norm, that is, in all male groups, the norm holds group-level BMI close to overweight despite of health awareness. Our results thus suggest that norms limit the effectiveness of interventions targeting individual weight-related behaviour, especially in men. Since norms are counteracting health awareness less strongly in the female groups, there must be additional drivers of obesity in women.