On-line Supplementary Material An agent-based model of child sugar-sweetened beverage consumption: implications for policies and practices Authors: Matt Kasman1, Ross A. Hammond1,2,3, Rob Purcell1, Benjamin Heuberger1, Travis R. Moore4, Anna H. Grummon6,7, Allison J. Wu6,8, Jason Block6, Marie-France Hivert6,9, Emily Oken6,7, Ken Kleinman10 1 Center on Social Dynamics and Policy, Brookings Institution, Washington, DC, USA (MK, RAH, RP, BH) 2 Brown School at Washington University in St. Louis, MO, USA (RAH) 3 The Santa Fe Institute, Santa Fe, NM (RAH) 4 ChildObesity180, Gerald J. and Dorothy R. Friedman School of Nutrition Science and Policy, Tufts University, Boston, MA, USA. (TRM) 5 Department of Community Health, School of Arts and Sciences, Tufts University, Medford, MA, USA. (TRM) 6 Division of Chronic Disease Research Across the Lifecourse, Department of Population Medicine, Harvard Medical School and Harvard Pilgrim Health Care Institute, Boston, MA, USA (AHG, AJW, JB, MH, EO) 7 Department of Nutrition, Harvard TH Chan School of Public Health, Boston, MA, USA (AHG, EO) 8 Division of Gastroenterology, Hepatology and Nutrition, Boston Children’s Hospital, Boston, MA, USA (AJW) 9 Diabetes Unit, Massachusetts General Hospital, Boston, MA, USA (MH) 10 Department of Biostatistics and Epidemiology, School of Public Health and Health Sciences, University of Massachusetts—Amherst, MA, USA (KK) Supplementary Materials and Methods Model Description The ABM model simulates young children’s sugar-sweetened beverage (SSB) consumption over time, with consumption being driven by the time that children spend in different settings. Parental behavior plays an active role in this model, making decisions that determine children’s SSB availability in the home. The model uses a weekly time resolution: each model step represents one week of real-life time. One model one represents five years, and such is comprised of 260 model steps. Agent Properties The model is populated by a set of agents πΆπ for π ∈ {1, … , π}, representing a cohort of young children, who age through early childhood and elementary school during the model. Each agent object πΆπ is associated with properties describing the represented child, as well as those describing that child’s parents; parents are not represented by separate object, but by certain properties and behaviors of the child agent. The model population is 1014 children. Each agent has the following properties: Individual Properties - Age, ππ ∈ [2,7]. Each child begins at age 2 and the simulation ends when they are age 7. - BMI z-score, π΅ππΌ(πΆπ ), intended to approximately reflect the contribution to BMI from consumption of sugar-sweetened beverages. The BMI z-score is relative to the CDC growth chart for the current age of the child. Household Properties - Race, ππ ∈ [π€, π, β]. We limit our racial groups to white, Black, and Hispanic, which represent most of the Viva cohort. Several other properties and distributions in the model are correlated with race. - SES (socio-economic status), π ππ π ∈ [low,high]. We define low SES as a family income under $55,000. We use the notation low_ses to denote the Boolean variable π ππ = low. - Maternal employment status. This is used while drawing simulated values for childcare time. Setting Association - Each child is associated with one of each of the three model settings: a home, a school, and a child-care type. At any step of the model, each child has a vector of weekly times β = [ππ» , ππ , ππΎ ]. spent in each setting, π Agents also take several actions which may bring about changes to their and settings’ properties. Here we summarize agents’ actions, and more detail is found below in the ‘Update Rules’ section. Actions - Consume SSBs: each unit of time, the child consumers some amount of SSB, equal to the sum of the amount consumed in each setting. For each setting, this is equal to the hourly consumption rate times the weekly hours spent in that setting. This also leads to changes in BMI z-score. - Update time distribution: at age 5, the child begins attending school. This reduces the amount of time spent in childcare and at home. - Visit pediatrician: once per year, each agent potentially visits a pediatrician, with a probability depending on agent properties. - Update home beverage distribution: each unit of time, each child’s parents update the home SSB offering. Model Environment The model includes three simulated settings. Homes, π»π for π ∈ {1, … , π}, represent each separate agent’s home environment. That is, no two agents share a home environment. We use π΅ π» to denote the beverage offering of an agent’s home environment; all beverage offerings are in units of servings of SSB per hour of time in that setting. Schools, ππ for π ∈ {1, … , π}, represent agent’s school environments, but only abstractly correspond to specific schools. That is, we do not simulate social interaction between agents assigned to the same school. Agents are associated with schools such that segregation by race and SES approximately matches reality. Each school entity has its own beverage offering, which we denote by π΅π . Childcare types, πΎπ for π ∈ {1, … , π}, represent the different settings in which children receive childcare. Childcare types are even further abstracted than schools, and each type of childcare is represented by a single object. The model contains four types: center-based, own home care, outside home-based care, and none. We represent the childcare offering for a given agent by π΅πΎ . Model Initialization Here we describe the rules for initializing agent properties: Agent Properties Agent race is chosen by selecting from the options white, Black, and Hispanic weighted by Viva cohort population. SES is initialized second using probability conditional on race. Age is initialized to 2 years (104 weeks). Child BMI conditional on both race and SES, and is simulated from the coefficients of regression bmiz ~ race + low_ses +race*low_ses. Beverage Offering Distributions Home beverage offering is initially drawn from a gamma distribution whose parameters depend on the agent’s race and SES: π΅π»,π ~ Γ(πΌ(ππ , ππΈππ ), π½(ππ , ππΈππ )). The beverage offering for each school is drawn from a folded (positive, about zero) normal distribution, π΅π ,π ~ |π(ππ , ππ )|. The beverage offerings for each childcare type are fixed. Agent Setting Association Each household is associated by definition with a unique agent. Childcare-types are assigned simply by using probability weighted by cohort incidence of each type. School association is accomplished by an algorithm that uses matrices exposure coefficient matrices to attempt to recreate real world levels of school segregation, described in pseudocode below. Here, “race-ses group” refers to one of the six combinations of race and SES: white-high, white-low, black-high, etc., which we notate w_h, w_l, b_h, etc. Objects: - child_list: list of all child agents - school_list: list of all schools - exposure_matrix: matrix of exposure coefficients, with rows and columns indexed by race-ses groups; i.e., exposure_matrix[h_h,w_l] is the exposure coefficient of a Hispanic, high-SES child to white, low-SES children. Classes: - child: a child agent o child.group: race-ses combination of the child agent - school: a model school o school.children: list of children assigned to this school for school in school_list: initialize school.children as empty list uniformly sample child c1 from child_list add c1 to school.children remove c1 from school_list for i in [2,...,school.size]: randomly select a race-ses combination, weighted by exposure_matrix[c1.group,] (the exposure coefficents to the race-ses combination of child c1) try: uniformly sample child cn from subset of child_list matching the chosen race-ses combination if subset is empty: uniformly sample cn from child_list add cn to school.children remove cn from child_list Setting Time Distribution Agent’s childcare time is initialized is set to zero if their childcare-type is “none”, and otherwise it is initialized by simulation from the coefficients of regression for the formula childcare_time ~ mother_part_time + mother_stay_at_home + childcare_own_home + childcare_other_home, where each of the regressors is a self-explanatory Boolean variable. The weekly childcare time TK is equal to five times the simulation draw. Agent school time is initialized to zero at age 2, and home time is equal to remaining time in the week, ππ» = 24 × 7 − ππΎ . Model Dynamics Each agent’s age increments by one week each time step, and its other properties vary according to several rules. Child SSB Consumption At each time step, the amount of SSB consumed is: SSB(πΆπ , π‘) = ππ (πΆπ , π‘) × π΅π (π(πΆπ )) + ππ» (πΆπ , π‘) × π΅π» (π»π ) + ππΎ (πΆπ , π‘) × π΅πΎ (πΎ(πΆπ )) This is added to the total SSB consumption, and used to calculate instantaneous BMI change for this time step. Child Time Distribution Update At ages 2 through 4, children attend only child-case and not school. At age 5, children begin attending school. Each child Ci begins attending school at a random model step π‘π (πΆπ ) between the ages of 4.5 and 5.5, at which point their weekly time at school is set to 5 times the length of the school day (in the main policy sweep, 6 hours). If a child’s initial childcare time is less than six hours per weekday, their childcare time after starting school is set to zero; otherwise, their school time is subtracted from their childcare time: ππΎ (πΆπ , π‘π (πΆπ )) = max(0, ππΎ (πΆπ , π‘0 ) − ππ (πΆπ , π‘π (πΆπ )) After the child attends school, the new home time is ππ» (πΆπ , π‘π (πΆπ )) = 24 × 7 − ππΎ (πΆπ , π‘π (πΆπ )) − ππ (πΆπ , π‘π (πΆπ )). BMI Update Each agent’s BMIz updates using a simple linear update function, π΅ππΌ(πΆπ , π‘ + 1) = π΅ππΌ(πΆπ , π‘) + πππ΅(πΆπ , π‘) ∗ πΎπ΅ππΌ πππ΅ , where πΎπ΅ππΌ πππ΅ is a parameter drawn at each time step from the BMI update coefficient distribution, πΎπ΅ππΌ πππ΅ ~ π(ππ΅ππΌ πππ΅ , ππ΅ππΌ πππ΅ ). Home SSB Offering Update Each time step, each agent’s home SSB offering updates by the formula: π΅π» (π»π , π‘ + 1) = π΅π» (π»π ) + π(π‘) + πΉ(π΅π» (π»1 , π‘), π΅π» (π»2 , π‘), … π΅π» (π»π , π‘)) The marginal change in SSB offering has two terms. π(π‘)~ π(ππ ππ€ ) is a normally distributed weekly increase term that represents a constant linear increase with random variation. The function F is a “follow the average” function that depends on the current home beverage offerings of the entire model cohort. Parents move their home beverage offering towards the mean offering by a certain portion (a free parameter) of the distance. Parents within a satisficing threshold (another free parameter) do not update their offering. Pediatrician Visit Once a year (at a random calendar point per child) each child may have a simulated pediatrician visit. The visit occurs with a probability that is dependent on each agent’s race and SES; if they miss a probability draw one year, they may still succeed in other years. If the visit occurs in a model year, parents of a child with high BMI receive a signal to reduce home SSB offering. This occurs if the agent’s BMIz is greater than 1, and the amount of reduction is a variable free parameter. The parameter represents the fractional reduction to home SSB offering if the signal is received; i.e., a signal of 0.25 immediately lowers home SSB offering by 25%. Model Usage We test several policy interventions, which are represented as alterations to several model rules. Beverage Multipliers The model includes four multiplier interventions: home beverage multiplier, school beverage multiplier, childcare beverage multiplier, and childcare center beverage multiplier. These can be used to test setting-wise reductions to SSB availability. Home and school multipliers simply multiply all beverages in those settings by a fixed coefficient. For home, this includes both the initialization and yearly update terms, but does not alter social and pediatrician mechanics. The school multiplier alters each school’s fixed beverage amount. The childcare multiplier effects the initial and update terms for all childcare settings. There is also a childcare center multiplier, which if set to a value other than 1 will override the general childcare multiplier and multiply the childcare setting beverage value instead by the childcare center multiplier. Pediatrician Signal Strength The baseline pediatrician signal strength is determined through model calibration, as described in below. The effect of the pediatrician signal can be increased by using larger values for the pediatrician signal parameter. Pediatrician Access Boost Improved pediatrician access is represented in the model by increasing agents’ probability of a yearly pediatrician visit occurring. This is accomplished by a “pediatrician access boost” parameter. Altering the value of this parameter increases pediatrician signal probability such that an access boost of zero yields baseline probabilities, and a boost of one ensures each agent will see the doctor every year. Specifically, the boost decreases the probability that individual agents do not see a pediatrician. For example, a boost of 0.2 reduces the probability of not seeing a doctor by a factor of 0.8 = 1 – 0.2, for each agent each year. Model Parameterization Model parameters are initialized based on analysis of several data sources. The primary data set is the Project Viva cohort data; to characterize the distribution of SSBs across different settings in more detail we use data from NHANES, PSID, and some other literature. Data sources: Viva (Supplemental Figure 1): Between 1999 and 2002, Project Viva recruited 2,128 pregnant women from obstetric offices in Eastern Massachusetts. Mothers who delivered a singleton infant attended in-person study visits during the first and second trimester of pregnancy, in the days following delivery, infancy (median age = 6.6 months), early childhood (median age 3.4 years), and mid-childhood (median age 8.0 years). They also completed mailed questionnaires on the anthropometry, diet, and health behaviors of their children on an annual basis extending through adolescence. Maternal data included daily SSB (including non-diet soda and fruit drinks with added sugar), fruit juice, and diet soda consumption during pregnancy, as well as prepregnancy BMI, pregnancy weight change, education, smoking status, income, and race/ethnicity. Child data included daily physical activity, height, weight, and sugar-sweetened beverage consumption. Supplementary Figure 1: Flow of participant involvement in Project Viva through the mid- childhood visit NHANES: The National Health and Nutrition Examination Survey (NHANES) is a program of studies designed to assess the health and nutritional status of adults and children in the United States. NHANES, a major program of the National Center for Health Statistics (NCHS), combines interviews and physical examinations. The survey examines a nationally representative sample of about 5,000 persons each year. These persons are located in counties across the country, 15 of which are visited each year. The NHANES interview includes demographic, socioeconomic, dietary, and health-related questions. In particular, we use the 24-hour dietary recall portion of NHANES data to help characterize consumption amounts and settings We use the data from the 2003-2004 and 2005-2006 survey cycles, pooled with appropriate adjustment of the survey weights, which approximately corresponds to the time period of Viva data collection for the relevant age of the cohort. PSID: “The PSID began as a longitudinal study that focused on the transfer of capital within families. In 1997, it was expanded to include a host of measures that are pertinent to children aged 12 and younger. These measures ranged from parenting, children’s academic achievement, behavior, and time use. Of the families in the PSID with children younger than 12 years, 2380 participated.” We use the Child Development Supplement (CDS) within the PSID, which has child time use data. NCES: NCES data is used to calculate exposure coefficients for school segregation by race and SES. Hourly SSB Consumption by Setting We use the NHANES 24-hour recall dietary survey data along with the PSID time use data to produce a data set of hourly SSB availability rates by setting, race, and SES. We restrict the analysis to consumption events of SSBs by the definition used by the Viva study: sugar-added soda (soft drinks) and fruit drinks (specifically, soda at USDA food code 924 but removing diet and unsweetened soda, and fruit drinks at code 925 but removing artificially sweetened lowcalorie drinks). We refer to this below as the NHANES-derived dataset. The NHANES data set allows us to classify consumption events for home and school. We then use the daily time use data from PSID to assign child time by setting for each age and surveyed day of the week, and then calculate the expected hourly consumption by setting for each NHANES survey respondent. We use a similar approach to calculate average daily servings at school per school day. Parameter Values from Data Below we list model parameters along with the data sources and analysis techniques used to calculate them. Cohort Demographics Cohort demographics are calculated using the Viva dataset only. Race is simulated using population probability, low SES status using a logistic regression based on race, and BMIz using a linear regression with the interaction terms listed. We use a simple value for the simulation error, equal to one fifth of a z-score SD. Supplementary Table 1: agent attribute initialization parameters Race probabilities probability white 0.78 probability Black 0.14 probability Hispanic 0.08 SES probability coefficients (logistic) low SES -0.89 low SES * Black 1.77 low SES * Hispanic 1.37 BMIz coefficients constant -0.04 Black 0.18 Hispanic -0.02 low SES 0.004 Black * low SES 0.21 Hispanic * low SES 0.52 Error 0.2 Home Beverage Distribution We parameterize the initial hourly home beverage rates using a gamma distribution which captures the long-tailed structure of SSB availability amounts found in both the Viva cohort data and the NHANES 24-hour dietary recall. Using the NHANES-derived hourly consumption data set described above, and for each race-SES grouping we parameterize a gamma distribution using by moment matching. The methodology of the NHANES survey differs significantly, and further it is conducted on a nationally representative population rather than the specific location of the Viva cohort. To compare model output to the Viva cohort data, we introduce a multiplicative “Viva adjustment parameter” to renormalize the model population mean home beverage rate to that observed in the Viva dataset. The model population is initialized by drawing home beverage rates from each agent’s appropriate gamma distribution, and then multiplying by the Viva adjustment. The final value of the Viva adjustment (0.185) is derived by calibrating the resulting model population so that the average model population mean numerically matches the Viva population mean. Supplementary Table 2: Home beverage availability initialization parameters Home SSB hourly servings Race Low Gamma SES? shape white FALSE 0.30543 Gamma rate 9.290275 white Black Black Hispanic Hispanic TRUE FALSE TRUE FALSE TRUE 0.333679 0.915161 1.15037 0.596015 0.588847 17.46528 26.16993 31.74052 17.70099 28.61967 Home Beverage Yearly Increase The yearly update distribution is derived by regressing mean hourly home consumption on year in the NHANES-derived dataset. In the model, a draw from a normal distribution is added to the home beverage availability amount each time step, converting the draw from yearly to weekly (* 1/52), and further multiplying the draw by the Viva adjustment parameter above; the mean for this distribution is 0.00993 and the standard deviation 0.00089. Childcare Type and Beverage Distribution Each agent is matched to a single childcare type which they are associated with throughout the model. The probability of each type association is calculated directly from the Viva data. Our value for childcare hourly consumption is taken from (Lutzkanin (2015)); we apply a 2.0 “informal care multiplier” to the home-based childcare types based on consultation with subject matter experts. Supplementary Table 3: Childcare initialization parameters Childcare Type Center-based 0.241 Home-based, own 0.248 home Home-based, other 0.181 None 0.33 Childcare SSB hourly servings Hourly servings 0.0208 Informal care 2 multiplier Childcare Time Distribution Probability of maternal employment is calculated directly from the Viva data, and the daily hours of early childcare time are simulated using a regression with the below coefficients which depend on maternal employment and childcare type. The error is a naïve estimate equal to one fifth the population SD. Supplementary Table 4: childcare time use parameters Maternal Employment Probability full time Probability part time Probability stay-athome Childcare time coefficients (hours/day) constant mother part-time mother stay-at-home own home childcare other home childcare Error 0.41 0.31 0.29 7.56 -3.48 -6.6 -0.36 0.53 0.1 School Beverage Distribution Each simulated school has its daily serving rate drawn from a folded normal distribution, which is then converted to hourly based on a six-hour school day. The parameters are the population mean (0.0322) and standard deviation (0.155) of the average daily school setting SSB servings from the NHANES-derived school data. School Assignment: Exposure Matrices School assignment is assortative by both SES and race and conducted according to the pseudocode algorithm above. The exposure coefficients are calculated from NCES data for the relevant ages from the 04/05 and 05/06 data sets. These are the separate exposure matrices for race and SES. The race-SES group exposure matrix is calculated at model run as the Kronecker product of these blocks. Supplementary Table 5: School assignment parameters SES high SES low SES Race white Black Hispanic high SES 0.919434 low SES 0.080566 0.89711 white 0.488077 0.107941 0.154866 0.10289 Black 0.218915 0.579593 0.334939 Hispanic 0.293008 0.312466 0.510196 Pediatrician Access We draw yearly probabilities of pediatrician access from (Weinick 2008). Each agent’s yearly probability of access is to the probability of their racial group in the table below, multiplied by the “Low SES factor” if they are a low SES agent. Supplementary Table 6: Pediatrician access parameters Pediatrician access white probability Black probability Hispanic probability Low SES factor 0.82 0.54 0.52 0.5 Model BMIz Update We draw our BMIz update coefficients from the RCT in (deRuyter 2012); these per-serving distribution parameters are renormalized to the US BMI distribution for the relevant age and converted to a weekly serving format (mean of 0.00161, standard deviation 0.000045). Calibrated Parameter Values The model contains several free parameters that are not set by data analysis or taken from literature. We use these free parameters to calibrate the model baseline to achieve a good match with the Viva cohort data in the absence of any policy intervention. We do a simple difference of means of weekly SSB consumption at age 7 between model data and viva data to rank each parameter combination, and thereby selected the below free parameters as achieving the best fit. Supplementary Table 7: calibrated parameters Home beverage social update Social probability 0.15 Social coefficient 0.15 Social satisficing 0.00275 threshold 0.25 Pediatrician signal strength Intervention Parameter Values As described above, intervention conditions are implemented by adjusting baseline parameter values by specified amounts. Supplementary Table 8: intervention condition parameters Parameter Pediatrician signal strength Home beverage multiplier School beverage multiplier Childcare beverage multiplier Childcare center beverage multiplier Pediatrician access boost Baseline value Policy values 0.25 1 1 1 0.3 0 0 0 0.375 0.559 0.822 0.559 1 0 0.39 0.2 0.4 0.5 0.7 0.964 0.7 0.8 0.982 0.8 0.912 0.425 0.45 0.456 0.822 Supplementary Results Supplementary Figure 2: heatmap representing mean cumulative SSB consumption over a large portion of the parameter space. Outer axes (“faceted” subplots, with labels at top and right) are school and childcare beverage reductions, and inner axes (labels left and bottom) are the home beverage reduction and a combined pediatrician boost variable. Pediatrician boost 0.956 incorporates both the access boost and the signal boost. All dimensions have the total number of axis points reduced for ease of display. The changes brought about by the physician policy are small relative to direct SSB availability reduction policies. Supplementary Figure 3: Heatmap plot showing only pediatrician policies. Increasing frequency and effect of pediatrician visits both decrease mean cumulative SSB availability, but do not have population-level effects as large as those that affect availability directly. Supplementary Table 9: Expanded Results Table. Condition impact means and confidence intervals obtained from Welch Two-Sample t-test comparison of distributions of average weekly sugar-sweetened beverage (SSB) consumption across 50 repeated runs under each intervention condition to 50 repeated runs of the baseline condition. Home Beverage Reduction - School Beverage Reduction - Childcare Beverage Reduction - Home Beverage Reduction 100.00% School Beverage Reduction 100.00% - Childcare Beverage Reduction 100% 100% 100.00% - 100.00% - - - 1.8% (10% soda reduction) 3.6% (20% soda reduction) 17.8% (100% soda reduction) - 100.00% 100.00% 100.00% 100.00% 100.00% 100.00% 4.4% (10% soda reduction) 100.00% 100.00% - - - Pediatrician Signal Increase - Mean SSB Consumption Weekly 2.061548 CI lower 2.047698 CI upper 2.075398 Pediatrician Access Increase - Pediatrician Signal Increase - Mean SSB reduction weekly 0.768885 1.807694 CI lower 0.756324 1.796599 CI upper 0.781447 1.818789 - 100% 100% - - 1.533543 0.516005 0.251382 1.27969 0.517284 0.519132 1.523193 0.502363 0.238495 1.268242 0.503659 0.505524 1.543894 0.529647 0.26427 1.291138 0.530909 0.532739 100% 100% - - 0.516811 0.517915 0.503184 0.504305 0.530439 0.531525 100% - - - 0.520568 0.506963 0.534173 100% - - - 0.52509 0.511523 0.538658 100% - - - 0.560961 0.547653 0.574269 - 0.241694 0.246249 0.240776 0.244059 0.314176 0.32104 0.267417 0.271916 0.266498 0.26971 0.339853 0.346741 100% 100.00% 100.00% Pediatrician Access Increase - - - Childcare Center Beverage Reduction Childcare Center Beverage Reduction - - - 20% 50% - 20% 50% 20% 50% 55.00% 60.00% - - 0.254555 0.259082 0.253637 0.256885 0.327015 0.33389 20% 50% 20% 30% - - - 0.353204 0.404567 0.340454 0.391864 0.365954 0.417271 100% - - - 0.579401 0.566022 0.59278 4.4% (10% soda reduction) 8.8% (20% soda reduction) 8.8% (20% soda reduction) 20.00% 20.00% 30.00% 30.00% 44.1% (100% soda reduction) 44.1% (100% soda reduction) 100.00% 100.00% 100.00% 100.00% 100.00% 100.00% 100.00% 100.00% 100.00% 100.00% 100.00% - - 100.00% - - 100.00% - 0.31395 0.301342 0.326558 100% - - - 0.642398 0.62921 0.655586 100% - - - 0.376056 0.803116 0.534792 0.363645 0.79046 0.522882 0.388467 0.815772 0.546702 100% - - - 0.944484 0.674652 0.932257 0.663143 0.956711 0.686161 100% - - - 1.136929 1.125199 1.148658 - - - 0.865355 1.27969 1.27969 1.27969 1.27969 0.854368 1.268242 1.268242 1.268242 1.268242 0.876342 1.291138 1.291138 1.291138 1.291138 - - 100.00% - - 100.00% - - 100.00% - - 1.8% (10% soda reduction) 3.6% (20% soda reduction) 17.8% (100% soda reduction) - - 1.8% (10% soda reduction) - 20% 50% 55.00% 60.00% 20% 50% - - - 1.356641 1.363636 1.385291 1.438091 1.34525 1.352245 1.373984 1.426843 1.368032 1.375028 1.396597 1.449339 - - - - 1.284259 1.27285 1.295668 - - - - 1.288828 1.277458 1.300199 - - - - 1.324876 0.003513 0.00859 0.002653 0.00654 1.313788 -0.0103 -0.00519 -0.01117 -0.00725 1.335963 0.01733 0.022366 0.016479 0.020331 20% 30% 55.00% 60.00% 20% 30% - 20% 50% 20% 50% - - - 0.075318 0.082275 0.101492 0.152678 0.061561 0.068507 0.087772 0.138997 0.089075 0.096043 0.115212 0.16636 - - - 0.004502 -0.00931 0.018316 - 3.6% (20% soda reduction) 17.8% (100% soda reduction) - - - - 0.00906 -0.00472 0.02284 - - - - 0.044703 0.031163 0.058244 - - - - - 0.06239 0.048812 0.075968 - - - - - 0.124183 0.282399 0.421951 0.110771 0.26946 0.40942 0.137594 0.295339 0.434483 - - - - - 0.612264 0.600179 0.624349 - - - - 20% 20% 50% 50% 0.006655 0.012308 0.011272 0.017774 0.078574 0.083087 -0.00713 -0.00143 -0.00246 0.004097 0.064847 0.069403 0.020441 0.026043 0.025002 0.031451 0.0923 0.09677 - - - 20% 50% 0.077726 0.081222 0.085458 0.089972 0.084683 0.088229 0.064 0.067534 0.071724 0.076282 0.070944 0.074522 0.091451 0.094911 0.099191 0.103662 0.098421 0.101936 - - 20% 20% 20% 20% 30% 30% - - 20% 50% 0.104371 0.108449 0.10355 0.106494 0.155281 0.158969 0.090673 0.094787 0.089858 0.092835 0.141626 0.145349 0.11807 0.122112 0.117241 0.120153 0.168936 0.17259 - 1.8% (10% soda reduction) 1.8% (10% soda reduction) 1.8% (10% soda reduction) 30% 30% - - 20% 50% 0.154488 0.157084 0.140834 0.143467 0.168143 0.170702 4.4% (10% soda reduction) 8.8% (20% soda reduction) 20.00% 30.00% 44.1% (100% soda reduction) - - - 20% 50% 20% 50% 20% 50% 55.00% 55.00% 55.00% 55.00% 60.00% 60.00% 60.00% 60.00% - 20% 50% - - 20% 50% 20% 50% 20% 50% - - - - 20% - 0.008025 -0.00575 0.021801 - - 50% - 0.013075 -6.66E-04 0.026817 - - 0.007154 -0.00663 0.020939 - 20% - - - - - - - - - - - - - - - - - 1.8% (10% soda reduction) 1.8% (10% soda reduction) 1.8% (10% soda reduction) 1.8% (10% soda reduction) 1.8% (10% soda reduction) 3.6% (20% soda reduction) 3.6% (20% soda reduction) 3.6% (20% soda reduction) 3.6% (20% soda reduction) 3.6% (20% soda reduction) 3.6% (20% soda reduction) 3.6% (20% soda reduction) 3.6% (20% soda reduction) 17.8% (100% soda reduction) 17.8% (100% soda reduction) 17.8% (100% soda reduction) 17.8% (100% soda reduction) - - - 50% 0.0111 -0.00264 0.024845 - 55.00% - - 0.079825 0.066102 0.093548 - 60.00% - - 0.086795 0.073064 0.100526 20% - - - 0.106011 0.092326 0.119697 30% - - - 0.157176 0.143529 0.170822 - - 20% - 0.012583 -0.00115 0.026321 - - 50% - 0.017644 0.003937 0.031351 - - - 20% 0.01172 -0.00203 0.025473 - - - 50% 0.015621 0.00191 0.029333 - 55.00% - - 0.084422 0.070732 0.098111 - 60.00% - - 0.091338 0.077642 0.105034 20% - - - 0.110571 0.096924 0.124217 30% - - - 0.161739 0.148132 0.175346 - - 20% - 0.048154 0.034648 0.061659 - - 50% - 0.053145 0.039665 0.066626 - - - 20% 0.047299 0.033786 0.060812 - - - 50% 0.051063 0.037591 0.064535 - - - 4.4% (10% soda reduction) 4.4% (10% soda reduction) 4.4% (10% soda reduction) 4.4% (10% soda reduction) 4.4% (10% soda reduction) 4.4% (10% soda reduction) 4.4% (10% soda reduction) 4.4% (10% soda reduction) 4.4% (10% soda reduction) 4.4% (10% soda reduction) 4.4% (10% soda reduction) 8.8% (20% soda reduction) 17.8% (100% soda reduction) 17.8% (100% soda reduction) 17.8% (100% soda reduction) 17.8% (100% soda reduction) - 55.00% - - 0.120137 0.106675 0.1336 - 60.00% - - 0.127025 0.113565 0.140484 20% - - - 0.146189 0.132771 0.159607 30% - - - 0.197451 0.184073 0.210829 - - - 20% - 0.065651 0.0521 0.079202 - - - 50% - 0.070123 0.056597 0.083649 - - - - 20% 0.064751 0.051203 0.0783 - - - - 50% 0.068235 0.054729 0.081741 - - 55.00% - - 0.137915 0.124421 0.15141 - - 60.00% - - 0.144748 0.131235 0.158262 - 20% - - - 0.16401 0.150522 0.177497 1.8% (10% soda reduction) 3.6% (20% soda reduction) 17.8% (100% soda reduction) 30% - - - 0.215342 0.201896 0.228788 - - - - 0.06691 0.053363 0.080456 - - - - 0.071429 0.05792 0.084937 - - - - 0.107244 0.093967 0.120522 - - - 0.127151 0.113772 0.14053 - 20% 8.8% (20% soda reduction) 8.8% (20% soda reduction) 8.8% (20% soda reduction) 8.8% (20% soda reduction) 8.8% (20% soda reduction) 8.8% (20% soda reduction) 8.8% (20% soda reduction) 8.8% (20% soda reduction) 8.8% (20% soda reduction) 8.8% (20% soda reduction) - - - - - - - - - - - - - 55.00% - - - 60.00% 0.117759 0.14448 20% 0.126243 0.112853 0.139632 50% 0.129273 0.115918 0.142628 - 0.199841 0.186495 0.213187 - - 0.206861 0.193519 0.220202 - - - 0.226157 0.212814 0.239499 1.8% (10% soda reduction) 3.6% (20% soda reduction) 17.8% (100% soda reduction) 30% - - - 0.277691 0.264414 0.290969 - - - - 0.128743 0.11537 0.142117 - - - - 0.133336 0.119992 0.146679 - - - - 0.169087 0.155996 0.182178 - - - 0.271633 0.274716 0.271 0.273136 0.345663 0.352613 0.297482 0.30052 0.296854 0.298952 0.371396 0.378326 20.00% 20.00% 1.8% (10% soda reduction) 3.6% (20% soda reduction) 17.8% (100% soda reduction) - 20.00% 30.00% 30.00% 30.00% 0.13112 20% - 20.00% - - 20.00% 20.00% 20.00% 20.00% 20.00% 20.00% 20.00% 50% 20% 50% - - 0.284558 0.287618 0.283927 0.286044 0.35853 0.365469 - - - 0.385167 0.436909 0.372352 0.424127 0.397981 0.449691 - - - - 0.286926 0.274023 0.29983 - - - - 0.291498 0.27863 0.304366 - - - - 0.327344 0.423596 0.425938 0.423098 0.314729 0.411076 0.413432 0.410576 0.339959 0.436116 0.438443 0.435619 55.00% 60.00% 20% 30% 20% 50% 20% 50% - 20% 30.00% - - 30.00% 30.00% 30.00% 30.00% 1.8% (10% soda reduction) 3.6% (20% soda reduction) 17.8% (100% soda reduction) - 30.00% 30.00% 30.00% 44.1% (100% soda reduction) 44.1% (100% soda reduction) 44.1% (100% soda reduction) 44.1% (100% soda reduction) 44.1% (100% soda reduction) 44.1% (100% soda reduction) 44.1% (100% soda reduction) 44.1% (100% soda reduction) 44.1% (100% soda reduction) 44.1% (100% - - 0.424633 0.412124 0.437142 - - - 0.498109 0.50508 0.525325 0.577221 0.485642 0.492617 0.512912 0.564846 0.510577 0.517543 0.537739 0.589596 - - - - 0.426502 0.414006 0.438998 - - - - 0.431038 0.418576 0.443501 - - - - 0.466892 0.454695 0.479088 - - - 20% - 0.613317 0.601239 0.625396 - - - 50% - 0.614864 0.602792 0.626936 - - - - 20% 0.61296 0.600878 0.625042 - - - - 50% 0.613953 0.601875 0.62603 - - 55.00% - - 0.688818 0.676804 0.700831 - - 60.00% - - 0.69581 0.683798 0.707821 55.00% 60.00% 20% 30% 50% - 20% - - - 0.716428 0.704485 0.728372 - 30% - - - 0.768694 0.756809 0.780579 - - - - 0.616812 0.604765 0.628859 - - - - 0.621373 0.609363 0.633383 1.8% (10% soda reduction) 3.6% (20% soda reduction) soda reduction) 44.1% (100% soda reduction) 17.8% (100% soda reduction) - - - - 0.657257 0.645523 0.668991 Sensitivity Analyses We performed sensitivity analysis by computing the model on a reduced sweep of the policy space for several additional “sensitivity conditions”. Conditions of interest include lengthening and shortening the school day, as well as three alternative sets of calibrated free parameter values which also scored well in the procedure outlined above. Supplementary Table 10: Sensitivity analyses parameters Sensitivity Condition School Day Pediatrician Signal Home Social Probability Home Social Coefficient Baseline School day 2 hours School day +2 hours Free params alt. 1 Free params alt. 2 Free params alt. 3 6 4 0.25 0.25 0.15 0.15 0.15 0.15 Home Social Satisficing Thresh 0.00275 0.00275 8 0.25 0.15 0.15 0.00275 6 0.2 0.15 0.1 0.0035 6 0.35 0.05 0.1 0.00275 6 0.3 0.05 0.1 0.00325 Overall, results from these analyses were either as expected a priori (e.g., school-based interventions had greater impact when the school day was long) in the case of school day adjustment or qualitative similar to our main results, in the case of alternative free parameters.