HOW ADDICTIVE IS SODA? ANALYZING THE COMPENSATORY EFFECT OF RESTRICTIONS ON SODA CONSUMPTION FOR SCHOOL-AGED ADOLESCENTS A Thesis Presented to the faculty of the Department of Economics California State University, Sacramento Submitted in partial satisfaction of the requirements for the degree of MASTER OF ARTS in Economics by Chantel Crane SUMMER 2012 HOW ADDICTIVE IS SODA? ANALYZING THE COMPENSATORY EFFECT OF RESTRICTIONS ON SODA CONSUMPTION FOR SCHOOL-AGED ADOLESCENTS A Thesis by Chantel Crane Approved by: __________________________________, Committee Chair Kristin Kiesel __________________________________, Second Reader Terri Sexton ____________________________ Date ii Student: Chantel Crane I certify that this student has met the requirements for format contained in the University format manual, and that this thesis is suitable for shelving in the Library and credit is to be awarded for the thesis. __________________________, Graduate Coordinator Kristin Kiesel, Ph.D Department of Economics iii ___________________ Date Abstract of HOW ADDICTIVE IS SODA? ANALYZING THE COMPENSATORY EFFECT OF RESTRICTIONS ON SODA CONSUMPTION FOR SCHOOL-AGED ADOLESCENTS by Chantel Crane Sugar-sweetened beverages have been pinpointed as a major contributor to adolescent obesity; the rates of obesity have dramatically increased over the past twenty years. This study investigates the effects of California’s SB 965, which banned certain beverages from being sold in high schools, including sodas. Using store-level scanner data, I estimate the effect of the ban on out-of-school purchases. Using a triple difference specification, I find that soda purchases post-ban decrease when stores are close to a school. However, when looking to stores within proximity to an open campus school, there is a significant increase in sales, although this finding is plagued by major deficits in the data. _______________________, Committee Chair Kristin Kiesel _______________________ Date iv ACKNOWLEDGEMENTS I would like to honor and appreciate Kristin Kiesel for her tireless commitment to putting my knowledge into practice. Your honest, consistent feedback and suggestions gave me a command on the material that I could not master alone. My future would be far different without your guidance, support, and expertise. I would also like to acknowledge and thank Terri Sexton for encouragement that that led me to the Master’s program in Economics. Your kind words, encouragement, patience and dedication to my education are unmatched; I am grateful every day for having met you. Finally, to my partner Matthew Jones, Esq.: I am grateful for your endurance, understanding, unwavering support and steadfast belief. You never questioned my study time and waited many late nights for me to complete my day’s work; you made my goals possible through your endless support and strength. v TABLE OF CONTENTS Page Acknowledgements .................................................................................................................... v List of Tables .......................................................................................................................... vii List of Figures ........................................................................................................................ viii INTRODUCTION ……………..………………………………………………………….... 1 BACKGROUND OF THE STUDY ......................................................................................... 6 Regulatory Background .................................................................................................... 6 Literature Review ............................................................................................................ 11 EMPIRICAL MODEL AND DATA ...................................................................................... 20 Economic Framework ...................................................................................................... 20 Data Description ............................................................................................................. 23 Empirical Model ............................................................................................................. 29 RESULTS AND ROBUSTNESS CHECKS .......................................................................... 35 Regression Results .......................................................................................................... 38 Robustness Check: Open Campus Identification ............................................................ 42 Robustness Check: Semester Effects .............................................................................. 46 CONCLUSION. ...................................................................................................................... 59 Appendix 1. Survey Administered .......................................................................................... 64 Appendix 2. Variable Descriptions ......................................................................................... 66 References .............................................................................................................................. 70 vi LIST OF TABLES Tables Page 1. Summary Statistics for All Stores .................. .………………………………. 28 2. The Effect of a Ban on Soda Sales in Schools on Out-of-School Purchases….41 3. The Effect of a Ban on Soda Sales in Schools on Out-of-School Purchases Utilizing Store Fixed Effects ..... ………….…………………………………. 45 4. The Effect of a Ban on Soda Sales in Schools on Out-of-School Purchases Utilizing Semester Indicators .................................. …………………………. 50 5. The Effect of a Ban on Soda Sales in Schools on Out-of-School Purchases Utilizing Hourly Data .................................... .………………………………. 57 6. Summary of Final Results for All Models ........... …………………………….58 vii LIST OF FIGURES Figures 1. Page Market Segment for Sodas Purchased in Grocery Stores Near Schools: With Compensation .................................................... ..…………………………….22 2. Market Segment for Sodas Purchased in Grocery Stores Near Schools: Without Compensation ........................................... …………………………. 23 3. Stores and Number of Proximate Schools….………………………………….27 4. Stores and Number of Schools Within Two Miles . …………………………. 27 5. Soda vs. Other Beverage Sales for All Stores in the Data- Not Within Half Mile .…………………………………………………………………………. 36 6. Soda vs. Other Beverage Sales for Stores Within Half Mile of a School……. 36 7. Soda vs. Other Beverage Sales for Stores Within Half Mile of an Open Campus………………………………………………………………………. 37 viii 1 INTRODUCTION Childhood obesity has more than doubled over the past two decades and is considered an epidemic (James and Kerr 2005; IOM 2007). Many factors contribute to this epidemic; however, sugar-sweetened beverages (sometimes referred to as caloricallysweetened beverages, or CSBs) have been identified as major contributors to an adolescent’s increased risk of becoming obese (Frieden et al. 2010). Consumption of food obtained at schools and other sources away from the home are labeled Food Away From Home (FAFH) and have become increasingly important within the context of the childhood obesity epidemic especially with their relationship to beverages consumed in tandem with these items. FAFH not only consists of food that is often higher in fat and calories, but also provides additional opportunities for adolescents to obtain CSBs outside the control of parents. A better understanding of the role of schools in the context of consumption of CSBs throughout the day may shed light on the effectiveness, and the potential success, of policies aimed at decreasing the obesity rate among children and adolescents. This paper investigates the extent to which students in Sacramento high schools, which like all California high schools experienced a ban on soda sales, substituted their usual consumption of sodas on campus with out-of-school purchases of CSBs (which will be referred to as compensation). This research utilizes store-level scanner data from a major grocery chain, which is aggregated to weekly observations for each store, extended by a survey of high schools in the Sacramento region. It employs difference-indifferences (DD) and triple difference (DDD) estimators to compare soda purchases in 2 stores within a close distance of a school to soda purchases in stores without nearby schools. I further compare soda purchases to alternative beverage purchases, as well as sales made before and after the implementation of state regulations which limit student access to CSBs in California high schools. Finally, I analyze differences in open versus closed campus policies and their potential effect on store purchases. The results suggest a decrease in soda sales post-ban for stores close to schools. Specifically, I find a 4.5% decrease in soda purchases in the treatment stores. This decrease is identified in a triple difference specification that compares soda purchases in stores close to schools as compared to other stores, in the post-ban compared to the preban period, and when isolating sodas versus other beverage sales of non-soda purchases for stores within half a mile to a school. Furthering the results with a robustness check, this effect is significant with a store-level fixed effect employed as well. For stores close to an open campus, I find that there is an increase in soda purchases; however, these results are characterized by many caveats that plague their validity. Additionally, robustness checks for seasonal characteristics are employed to control for the potential increases due to high or low volume periods caused by seasonal fluctuations at stores proximately located to an open campus, as well as differences in response over time. The findings suggest an initial compensation in spring 2007 after the regulation was passed. Further, when inspecting the transaction-level data, aggregated by hour instead of week, the compensation effect persists in open campus schools, and suggests increased sales before and after school, but not during school as would be expected when comparing open versus closed campuses. 3 Potential compensation effects may severely decrease the putative success of policies that restrict access to CBSs at school, because high school students, like adults may employ adverse behavioral responses to these policies that may undercut the desired outcomes. Many different compensatory effects can occur: high school students may bring beverages from home, obtain them on the way to and from school, or depending upon the school policies, leave campus during lunch. Policies aimed toward reducing obesity in adolescents focus on limiting consumption; however, these can only be effective if they reduce excess calorie consumption throughout the day; including out-ofschool caloric consumption. Thus, this study investigates potential compensation by students at easily accessible grocery stores in order to gauge out-of-school consumption changes as a response to school-based regulations, specifically in conjunction with the school-level open campus policy. To understand the impact of school regulations on the prevention of childhood obesity, it is important to recognize the consumption patterns that shape adolescent diets. Nielsen et al. (2002) discuss sources where adolescents consume calories and determine that although schools are main hubs for caloric consumption, children are eating less at school and home than in previous years and instead increasing their patronage of other FAFH venues. This leaves many questions about compensatory behavior and motivates this research to answer some of them, and focuses primarily on in-school reductions in soda consumption to determine the effect on out-of-school purchases. While many studies are designed to engage this topic, the different strategies employed do not individually paint a clear picture of behavioral responses to these school-based policies. 4 For instance, Briefel et al. (2009) look to on-campus behavior of adolescents attending schools where regulations on CSBs have been implemented to quantify the change in consumption patterns of students. Others, like Fletcher et al. (2004) observe the effect that changes in taxes have on the consumption of CSBs by all individuals as opposed to students. Huang and Kiesel (2012) study the effectiveness of a ban on CSBs on consumption outside of the school using household-level data and DD by household type and by state. In contrast, this paper will look to out-of-school consumption by using store-level data and proximity to a school to quantify changes in sales that may be due to the regulation. While limited in scope, it attempts to take a more detailed look at potential compensation effects of banning sodas from being sold in high schools. While I do not find evidence for compensation in general, my results suggest that open campus policies may result in a pattern of initial compensation at stores within proximate distance to a school after the regulation. However, when isolating semesters to map the compensation over time, the pattern does not persist. Additional robustness checks question the extent of the compensation, and these results suggest that schoollevel policies may be the best indicator of whether overall reduction in CSBs may occur in response to supply reductions at schools. This paper is outlined as follows: The section called Regulatory Background presents the background of policies regarding adolescents and soda consumption. The Literature Review describes previous studies aimed at determining the relationship between soda and caloric consumption in adolescents. Economic Framework describes the econometric tools used in this research. The Data Description outlines and explains 5 the data used in this paper. The Empirical Model presents the model and regression equation used for the regressions. The Results and Robustness Checks report and discuss regression results and robustness checks employed in this study. The final section, Conclusion, summarizes the research and outlines some possible goals and opportunities for further study. 6 BACKGROUND OF THE STUDY This chapter will describe many public policy strategies restricting CSB consumption. Although not all regulations pertain to adolescents, and particularly high school students, it is important to understand the expansion of control over the past decade made by California and its localities. Regulatory Background This paper focuses on SB 965, a California regulation which first took effect in 2007, and limits the sales of CSBs in high schools. In 2004, California was the first to pass a state-level regulation banning soft drinks from elementary, middle, and junior high schools (except at special events). Initially, high schools were not subjected to the effects of this regulation, and limitations were made by very few schools as either a school-level or district-level policy. SB 965 was passed to correct the omission of high schools in previous legislation and therefore expanding the ban on CSBs to high schools as well. There were two checkpoints for the regulation: July 1, 2007 when the schools were required to meet 50% or more of the requirements outlined in the regulation, and July 1, 2009 when the schools were required to be 100% compliant with the new law. Many states have followed in California’s path, implementing policies aimed at curbing obesity; a common strategy is limiting student access to CSBs with the projected outcome of decreasing access and consumption of these items. This paper aims to explore the extent to which adolescents compensate outside of school in reaction to regulations in schools that limit student access to CSBs. 7 Competitive foods and beverages are defined as items that are not offered in accordance with the United States Department of Agriculture’s National School Lunch Program (NSLP). Some states, like California, have outright banned CSBs from being offered on school campuses in the cafeteria and alternative sources such as snack clubs or vending machines. Although California’s regulations have been among the most stringent, 17 other states have also implemented limitations on portion size and specific nutritional requirements for competitive foods within the school (IOM 2007). At the local level, limitations and taxes of carbonated beverages have been passed by many cities: primarily targeting adolescents because of their particular risk for long lasting obesity. There are various policy tools advanced in California to reduce obesity at the community level: Richmond, CA proposed to target sodas and has placed on the November 2012 ballot a one cent tax per ounce of carbonated beverages in an attempt to reduce demand with the goal of reducing the extreme obesity epidemic in their city (Kenyon 2011). In San Francisco, Happy Meal toys can no longer be offered in meals that did not meet city nutritional standards of child-approved portions, with the intention of weakening the meals’ marketing appeal to young children (Bernstein 2011). These examples are intended to be illustrative, rather than exhaustive, of the policy tools employed in California to combat childhood obesity at FAFH locations. At the school level, some approaches target schools with voluntary recommendations, rather than mandatory restrictions; for example, regulators have attempted to work with schools and major players in the soft drink industry to create healthier environments for adolescents, aimed at creating healthy habits for the future. 8 Through the Healthy Schools program, launched in 2006 by former President Bill Clinton’s Alliance for a Healthier Generation, schools are offered support through presentations and grant opportunities supporting ideas aimed at making schools healthy and active places for students (Robert Wood Johnson Foundation 2011). Through the Healthy Schools program, voluntary guidelines were agreed upon by soft-drink manufacturers and others designed to lower the calories and increase the nutritional content intended for adolescent consumption. Despite the voluntary augmentations to the school environment, lawmakers continue to seek legislative solutions for application at the school level. Schools have been targeted for limitations of CSBs by legislative and regulatory action, despite reports of schools’ heavy financial dependence on contracts with soda manufacturers that fund extracurricular and other activities. In 2002, over 240 school districts in the United States entered into exclusive contracts with providers of these products often in exchange for upfront or yearly payments if schools met sales quotas (Kolb 2004; Shin 2007). These contracts are threatened (and in the case of California almost completely eliminated) by regulations targeting CSBs in schools; opponents cite the elimination of these contracts as a high price to pay for little to no impact on adolescent health. To combat obesity and institute a well-rounded nutritional environment, many schools and school districts reached beyond programs designed to encourage positive behavior but have instead created their own policies toward competitive foods and beverages. A 2005 report found that the majority of schools had policies of their own limiting or eliminating competitive foods and beverages in their schools: many of which 9 started at the school level as opposed to a district mandate (Story et al. 2006). Story et al. (2006) also report that the Los Angeles Unified School District banned all soft drinks, and other major public school districts throughout the country have drastically limited the offering of competitive foods and drinks in their schools. The local districts took action before the states in many cases, in response to parental requests for better food environments. These policies are not the focus of this study because they vary greatly by school. This study focuses on the effect within the Sacramento region of the statewide ban, and does not take into account policies that were in place before the ban’s enactment. Despite state, local, and school level policies preventing soda sales on campuses, schools still were the main sources for competitive beverages obtained by adolescents. Prior to the California regulation, adolescents had easy access to CSBs at school through vending machines, snack clubs, and even school cafeterias (Briefel et al. 2009; Bowman et al. 2004). According to Fox et al. (2009), the 2004-2005 school year was characterized by 100% of U.S. high schools offering competitive food. In order to reduce the availability of competitive foods in schools, and with the aim of decreasing overall caloric intake, California passed SB 965, a regulation of competitive beverages in high schools: prohibiting high fat, high sugar content beverages and foods on all high school campuses (IOM 2007). Although SB 965 institutes strict standards on competitive foods and beverages, there is no formal measure of compliance outlined in the bill. Samuels et al. (2009) aim to measure school compliance by analyzing the quantitative measure of soft drink availability in high schools and the consumption of the product overall in high schools. 10 Since every school is different, it is important to distinguish their ability to implement the state-mandated changes in soft drink availability at schools. During their study (and still through today), there was no account for the actual number of high schools that completed the transition in applying the new law in California; therefore the authors created their own process to determine soft drink availability and compliance rate. The compliance rate is determined by the amount of competitive foods (and drink) that meet the fat, calorie and sugar limitations outlined in the regulation. Samuels et al. (2009) implement a study that investigates 56 schools and records the amount of competitive foods (and beverages) being offered in these California high schools. They gather data by sending trained representatives to observe the school first-hand and record their findings. The process of their study overcomes many of the problems of self-reported data, which in this case may be a significant issue, especially because the question being investigated is a matter of legal compliance (even if no formal measurement has been instituted by the state). Samuels et al. (2009) find a high level of compliance and conclude that California has completed the process of implementing the CSB regulation at school sites. Additionally, they found that after the first checkpoint in the implementation of SB 965 (50% compliance by July 1, 2007), 86% of the schools in their sample were meeting the requirements (Samuels et al. 2009). They conclude that compliance varies greatly depending upon the school, but overall, the regulation seems to be effective in reducing the offerings of competitive foods and beverages. Since California schools are reducing their offerings of soda and should not be offering any of these products on high school campuses, further research with a focus on 11 out-of-school consumption should illuminate the actual effect of SB 965. A triple difference specification is used that examines soda sales in stores within close distance of a school to those further away, comparing soda to non-soda beverage purchases, before and after the soda ban. These markers provide an unique quasi-natural experiment where a triple difference framework can be used to estimate a treatment effect defined as the effect of banning soft drinks in schools on supermarket purchases. Investigating stores near open campus schools as an additional treatment group will indicate the effect of policies with taking into consideration school-level policies that affect the Sacramento region. Literature Review The recent literature on soda bans in schools suggests that it is effective in limiting access to CSBs, and in changing student behavior patterns while students are on a high school campus. For instance, Samuels et al. (2009) went to schools to determine the level of compliance post-implementation of the ban in high schools; they found significant progress in implementing SB 965 and consequently a successful transition to its limitations. Briefel et al. (2009) conclude that school regulations change adolescent behavior and can greatly reduce the amount of calories high school students consume; however, they also found that food at home and particularly other FAFH greatly contribute to the level of adolescent consumption of CSBs and therefore leave further room for analysis on the access and level of student intake of such products. 12 Differing from the previously outlined research, my results illustrate out-of-school consumption as a proxy to determine the effectiveness of the policy. It follows Huang and Kiesel (2012) more closely. They focus on Connecticut’s regulation of CSBs compared with states not employing a similar regulation. In addition, they use households without school age children in Connecticut as their control, and they compare the purchasing patterns before and after the regulation to households with school age children. Their study also attempts to control for advertising expenditures as a possible indicator for increased efforts to reach adolescents outside of the school environment; they find that there was a decrease in advertising expenditures after the ban. However, they conclude that the decrease is most likely due to seasonal changes, as the advertising expenditures followed a similar pattern the following year. Most notably, their regression results at the aggregated level do not support the notion that soda consumption outside of schools increased in any statistically significant way, in response to decreased accessibility of CSBs within schools. Like Huang and Kiesel (2012), this paper examines the effect of the soda ban on adolescent consumption, but digresses by focusing on Sacramento, CA instead of multiple states which provides a local and more in depth perspective on the results. Moreover, this paper will employ a robustness check to account for the possible seasonal effects identified by the Huang and Kiesel (2012) study. By looking to semester consumption patterns, I aim to isolate the effect of the policy on adolescents during the time in which the regulation is effective (during the school year). Although the present study has a similar goal and technique, the purchase data is actual store-level transactions 13 as opposed to the AC Nielsen Homescan data used by Huang and Kiesel (2012). I distinguish stores that are close to a school and those that are not by utilizing geocoded information about stores and school proximity. In addition, the survey data collected attempts to ensure compliance, and incorporate school characteristics, including open campus policies and school instruction (and lunch break) times. If compensation were to occur, then students would purchase CSBs potentially from many sources. Although there may be many potential locations for student purchases, this study focuses on grocery stores close to a school. It is expected that if consumption patterns change, some of these changes in CSB transactions post-regulation would occur at nearby grocery stores and should be reflected through this strategy. Understanding the externalities associated with the removal of CSBs from schools specifically with regard to out-of-school purchases is essential to support current policies and creating future policies that may successfully reduce obesity among adolescents, (relative to consumption of CSBs). A number of studies suggest that removing unhealthy foods from schools reduces obesity by making it less convenient to obtain competitive foods and beverages that add excessive empty calories to a child's diet (Frieden et al. 2010; Fox et al. 2009; Mancino et al. 2010). Children consume 26% of their daily energy from school sources; even when taking into account other forms of FAFH, it is clear that schools are still a viable forum for combating obesity through limitations on competitive foods and beverages (Briefel et al. 2009). Briefel et al. (2009) determined that children consume less soda at school when soda is no longer offered at schools and reduces overall calorie consumption at schools. Although this is true for schools without snack 14 clubs or CSBs offered through the school lunch program, that study was unable to find a similar statistical significance relating the lack of vending machines to lower calories overall consumed from CSBs. This inconclusive result suggests compensation may have occurred and thus explaining why an in-school reduction did not lead to an overall reduction in calorie consumption by students in response to a vending machine ban. This indicates that further research at the school level and elsewhere may be necessary (Briefel et al. 2009). This paper aims at extending the findings outlined by the authors in order to determine whether there is an overall decrease in consumption by investigating potential out-of-school consumption as well. Although Briefel et al. (2009) found that students are more likely to obtain calories at school than any other single source, their paper does not relate consumption of CSBs to patterns of consumption outside the hours of the school day. Analyzing the difference in sales from before and after the regulation, this research will determine the existence of a compensation effect representing an increase in adolescent consumption of calories off-campus in response to the regulation. The study will use geocoding techniques to pinpoint the nearest school to proximately located grocery stores in order to measure any purchasing change pre- and post-implementation of the policy. Finally, a survey conducted of schools within the Sacramento region will provide additionally relevant information, such as whether the school has an open or closed campus policy. Furthering these studies to out-of-school consumption, Mancino et al. (2010) attempt to find a correlation between FAFH and dietary choices. If there is a positive 15 correlation between FAFH and caloric intake, then the schools may play a larger role in weight gain than the home; thus demonstrating that school-centered legislation on nutritional standards may be effective toward the ultimate goal of curbing childhood obesity and one of the few places that legislative policy is applicable. Similar to other studies, Mancino et al. (2010) used recall data, which asks participants to recall and record their food and beverage intake on a daily basis, obtained from the 2003-2004 National Health and Nutrition Examination Survey, and the 1994-1996 Continuing Survey of Food Intakes by Individuals. Their study analyzes children between the ages of 6 and 18, and identifies change in total daily caloric intake as the dependent variable. The recall data classified meals by location that the individual consumed: the meal’s location was categorized by where the majority of the calories were obtained. For instance, if a lunch of 500 calories is purchased at school and a soda of 100 calories is brought from home, the items are not separated to different locations but instead all calories are considered food at school. Mancino et al. (2010) narrowly focused the study to hone in on individual preferences in order to purge the endogeneity problem in previous research. To isolate individual preferences they use recall data from two non-consecutive days which accounts for individual trends that are consistent between both days. The data level describes the consumption of each individual. The researchers compared that information and the difference between the two days to create a dependent variable of the average difference among all children. They find that caloric intake from FAFH increased the number of calories consumed overall by adolescents between the ages of 6 16 and 18. This makes intuitive sense because fast-food and other restaurants are typically more greasy and high in sugar content. Furthermore, consumption of carbonated beverages seems more likely when eating away from home. The study concedes that it is difficult to ascertain the nutrition level of FAFH and therefore it only looks to number of calories instead of nutrition level. Nielsen et al. (2002) emphasize the necessity of discovering the effects of FAFH as well, and more specifically the role of school food choice on the potential transition to increasing consumption of calories away from school and away from home. In the context of food choice, the potential compensation effect of restrictive policies in schools is critical to understanding the purchases of CSBs at other locations and subsequently the changes in overall CSB consumption. Adolescents increasingly seek nourishment at fastfood establishments, restaurants, and other FAFH, which are traditionally characterized by greater portion sizes and higher caloric value. This illustrates the fact that schoolcentered regulations alone may not reach far enough to have the intended decrease in obesity (Mancino et al. 2010; Bowman et al. 2004). Through a survey of 12-29 year olds, Nielsen et al. (2002) found that school-aged adolescents were consuming less food at school than in previous years; this coupled with the fact that they are consuming more FAFH paints a disturbing picture of the potentially poor nutritional decisions adolescents today are making, including increasing their caloric intake with beverages and failing to decrease consumption of other calories. Although potentially of decreasing importance, schools still play a large role in adolescent calorie consumption by providing significant opportunities to consume competitive foods throughout the day. 17 Fox et al. (2009) found that 55% of students attending high school consumed competitive foods while at school in the time period before any regulations were implemented. With this market substantially altered, as SB 965 does, adolescents are likely to react and look to other markets for substitutes. Whether these students do in fact access the alternative markets will play an important role in the success of schoolcentered policies aimed at preventing obesity among adolescents; if adolescent’ compensatory behavior is not only prevalent but also unaccounted for in studies, then the goals of these regulations may not be reached. I combined research on the compliance level, FAFH, and compensatory practices to examine soda purchases made by adolescents in response to SB 965 with the overall goal of accounting for these alternative markets in context with other research. While the previously discussed studies are the most closely related to this paper, it is important to note other studies that provide additional context to the topic. For instance, often cited as a potential policy option is taxation on sugary beverages. These studies do not analyze supply limitations as a framework, but instead curbing the demand through price changes as an additional strategy to reach the same goal; an overall reduction in daily calories consumed. Fletcher et al. (2011) analyze the relationship between obesity and soft drink consumption. They evaluate incentives designed to stop adolescents from drinking soda and the effectiveness of reducing caloric consumption in adolescents. The Fletcher et al. study is pivotal because it illustrates the link between soft drink consumption, caloric intake and Body Mass Index (BMI) of children. Fletcher et al. extend the literature by 18 utilizing state-fixed effects as well as demand and cross-price elasticities with the intent of determining if public policies such as an optimal tax rate can reduce soda consumption and overall caloric intake. Following the literature on cigarette taxes, they consider individual state tax rates and their changes over time as provided through the Council of State Governments and add a BMI calculator to measure the fixed effects of soda consumption on weight. The study also uses the National Center for Health Statistics data; this is a representative sample of about 5000 persons each year, which is a questionnaire that includes a physical examination and demographic information. The survey details the food and beverages consumed in the last 24 hours and a representative sub-sample is also used to determine vitamin levels in the blood, which serve to determine nutrition levels. This research analyzes a full day of calorie intake and includes long-term trends and tax changes within each state. However, it does not look to weeklong eating habits, and therefore may incorrectly categorize an individual’s actual behavior (caloric consumption changes on a day-to-day basis). In addition, there are many problems associated with self-reported nutritional data: people are not always completely honest or may simply forget their previous day’s consumption. They conclude that although taxes on sodas did not have an effect on whether or not a youth will consume soda, it did affect how much soda a youth consumed. Fletcher et al. (2011) also state that the presence of close substitutes may greatly reduce the effectiveness of a tax on CSBs. Their study finds a compensatory effect of soda taxation in added milk consumption, suggesting that they may be substitutes. Although Fletcher et al. (2011) finds that taxation lowers the consumption of soda in adolescents, it is not 19 significant in application to overall calorie consumption, and they conclude that the policy is not effective in reducing the number of calories consumed each day. But because behaviors changed with an increase in price, their study does suggest that adolescents are rational actors that react to regulations as expected; this assumption will be relevant in the current study. Finally, there are additional studies on CSB consumption which warrant mention. Bonnet and Requillart (2011) analyze changes to prices of the components of CSBs; they find that decreasing the price of sugar increases the number of CSBs purchased each year by about one-liter per person. However, their study does not describe a limitation of access, but instead a change in price of an input which, like tax studies, is not the purview of the current study. Taber et al. (2012) find that limiting access to sodas is not singularly effective at keeping adolescents from consuming sugary drinks, but instead more comprehensive strategies including excluding sales of sports drinks and other sugary drinks were more effective in changing adolescent behavior. Runge et al. (2011) validate and further the nutritional implications of Taber et al. (2012) citing the positive aspects to milk and juice substitution from sodas. Their paper suggests that adolescents can be incentivized to increase consumption of milk and juices and consequently reduce their risk of becoming obese by 5% with an increase in price on CSBs of 20% (Runge et al. 2011). Also important to acknowledge is the body of studies regarding increased vigorous activity as an option to decrease obesity rates; but that is well beyond the parameters of this study and does not directly speak to the issues of SB 965. 20 EMPIRICAL MODEL AND DATA This chapter describes the data used in this study. The econometric specification and variables employed are explained in the context of this study and econometric theory. Further, the outcomes will be outlined with a description of expected results. Economic Framework Schools play an important role in forming adolescent habits, specifically with regard to nutrition. Health classes are mandatory and are designed to bring awareness and understanding to nutritional issues. Their lack of reliance on parents for meals provides high school students new found flexibility in nutritional decisions and consequently more opportunities to make poor decisions regarding nutrition such as over consuming fat and calories while under consuming healthy dietary components. Since students are assumed to be rational actors (the assumption follows with previous studies as outlined in the pervious chapter), they are expected to respond to policies by taking into account opportunity cost, and accessibility of the items into their cost analysis. In order to pursue alternative markets to that of the school in response to the removal of a specific item, substitute markets will need to be reliable, and overcome transportation costs and time constraints that may limit high school students’ access. In Sacramento, schools are often positioned near many businesses and large neighborhoods. If a school is located near fast-food restaurants and convenience stores, there is low opportunity cost to accessing sodas and other competitive foods that are restricted on campus. However, adolescent’ accessibility to these substitute markets varies by school; for instance, open campus policies permit students to leave campus 21 during school hours including during lunch time. If students are assumed to follow the applicable rules, then the lack of an open campus policy should prevent students from purchasing competitive items during school hours. Such is the presumed framework of a simple competitive market for substitute goods, which I have been referring to as compensation. Here, although the goods are the same, the times and locations of the purchases will play a key role in the results. Since students face numerous constraints which may prevent them from accessing these items, compensation may not be possible and an overall reduction in the consumption of CSBs may occur. For instance, school-level barriers such as a closed campus policy should prevent students from leaving campus to purchase outside food or drinks during the school day. Additionally, curfews requiring students to be home immediately following the conclusion of a school day, lack of transportation to alternative sources of CSBs, and many other factors may make the opportunity cost of pursuing substitute markets substantial enough to deter compensatory behavior, unless parents assist in the behavior. There are two potential effects of the ban with regard to increased beverage consumption: within-school substitution of alternative beverages and out-of-school compensation of soda (or utilizing an alternative market). The first is not the focus of this study; to determine level of compensation, the regressions will show the level at which students are consuming the same regulated products in a non-school location. The second effect can be modeled by a simple shift in the demand curve in out-of-school markets: convenience stores, fast-food restaurants, grocery stores, etc. When students are prevented from purchasing sodas on campus, they can either bring sodas to school, obtain 22 them after school, or purchase them during school hours off-campus (if there is an open campus policy in place). With the ban on CSB sales in schools, supply of CSBs by schools decreases to a quantity of zero after full implementation. If adolescents compensate fully, there should be a one-to-one shift in outside markets. However, this study focuses on grocery store data only, and would not measure fully the extent of any compensation. It is expected, however, that if students look for other alternatives, a proximately located store should capture some of the non-school purchases. Therefore, it is expected that an increase should reflect only a portion of the shift, as determined by the increase in transactions of one of the alternative markets; this is illustrated in Figure 1. Note: This is a simple illustrative model of the potential effect of the regulation on the soda markets. The supply curve is flat and assumed to be a long run supply; students are presumed to be a small enough population that they should not play a substantial role in the market, and thus should not affect the underlying cost structure of the supply curve. Therefore the quantity increases from Q1 to Q2 in the grocery store market. 23 Alternatively, Figure 2 depicts a change in preferences created by the ban. If students do not compensate, but instead embrace their new alternatives, it is reasonable to expect that students would increase consumption of soda alternatives. This may be seen in increased purchases of juices, sports drinks, milk, or other alternatives to soda at school and outside of school. This preference change, if significant, may lead to a decrease in soda purchases at stores within a reasonable distance from a school. Shifts in preferences seen in the data as a negative coefficient for the DDD. The DDD tests for the presence of either one of these potential effects illustrated by Figures 1 and 2. Note: This is a simple illustrative model of the potential effect of the regulation on the soda markets. The supply curve is flat and assumed to be a long run supply; students are presumed to be a small enough population that they should not play a substantial role in the market, and thus should not affect the underlying cost structure of the supply curve. Therefore the quantity decreases from Q 2 to Q1 in the grocery store market. 3.2 Data Description The data used is proprietary data from a major supermarket chain and aggregates transaction-level beverage purchase records into weekly store-level sales at the individual 24 product level from week 48 in 2006 to week thirteen in 2011 for all stores in the Sacramento region.1 The weekly sales data combines the Universal Product Code (UPC) identification matched with the UPC description, which was then used to create a dummy variable for a soda category. This will be used to separate sales of soda beverages from other beverage sales. Additionally, the purchases can be further categorized by a sale dummy that indicates whether the item was purchased while it was under promotional pricing. Finally, I looked separately at both quantity and net sales of beverages purchased from each store. The weekly data are supplemented by data from a survey conducted by the author of high schools in the Sacramento region, which may be referenced in Appendix 1. The schools were asked to respond to a number of questions including whether or not they have an open campus policy that allows students to leave campus at lunchtime. Additionally, they were asked what types of food programs they participate in to determine whether they were utilizing the National School Lunch Program (NSLP); if they participate they are required to follow the regulation and therefore this question helps distinguish which schools should be compliant. All schools in the region were sent a request and the response rate was 41%; there are 54 schools in the final set. This final set includes four charter schools, 39 public schools and eleven private schools, that responded to the survey. Of the survey respondents, there were four with open campus policies, (two were 1 The data used in this study is a subsample of a larger data set received through the SIEPR-GIANNINI Data Center (http://are.berkeley.edu/SGDC/ ) for an ongoing research project titled: Mandated Soda Bans— Adherence at School and Compensation at Home. 25 non-continuation high schools where the majority of students attended full time). Only one school was included in the regressions as an open campus school because it was the only school considered in direct proximity to a grocery store. In the survey, respondents were asked whether they offered soda on campus and the only schools that replied in the affirmative were all private institutions. This is further evidence that the regulation has been fully implemented in public schools. The results of Samuels et al. (2009), an independent study taken by trained professionals who did not rely on response data, suggest that schools were ahead of state regulatory timelines on their way to completing the transition to a soda-free school environment; in conjunction with the survey, the assumption of full compliance seems reasonable (Samuels et al. 2009). In a next step, the store addresses from this major grocery store chain were geocoded to find the distance to the schools with a maximum of 10 miles, and all 129 schools (including non-respondents) were included in the geocode set which is designed to determine if a store has a school near regardless of their response to the survey (six charter, 93 public, and 30 private schools). There are 33 schools whose distance is between one and two miles of a store, 14 schools between half a mile and one mile, and three within half a mile or less of a store, one of which is an open campus. A count variable was then created that indicates the number of schools within varying ranges of the stores (within half of a mile to two miles) to describe the effect on beverage transactions of having schools near the store. The store-to-school distances are summarized in Figure 3. Figure 4 provides a different illustration: all stores were quantified by the number of schools within two mile proximity. 26 Combining these data sources results in a final data set with 3,008,915 weekly observations, with a total of 35 stores, 16 of which have a school within two miles. The summary statistics for the final data set are reported in Table 1 and additional variable descriptions are provided in Appendix 1. The products are identified using their UPC; the product_n variable assigns an unique indicator that counts the products instead of using the long UPC identification. There are 9747 different beverage products identified and these are further truncated to categories of sodas and non-soda purchases. The soda variable indicates soda purchases, and according to the data, 50% of the purchases were identified as soda products. 41% of the products purchased were under promotional pricing. Additionally, 42% of the products were sold during the treatment 2 period, or after July 1, 2009 which makes the treatment period large enough to be adequately compared to the control period. 27 Figure 3: Stores and Number of Proximate Schools 18 16 14 12 10 8 6 4 2 0 Stores with a school within Stores with a school within Stores with a school within half mile one mile two miles Note: The Y-axis represents the number of stores. Stores within two miles include all schools within two miles (including half and one mile). Stores within one mile include all schools within one mile (including half mile). Stores within half mile include schools less than half mile. Figure 4: Stores and Number of Schools Within Two Miles Number 20 of stores 18 16 14 12 10 8 6 4 2 0 0 schools 1 School 2 3 School 4 5 School 6 7 School Schools Schools Schools Stores with a school within 2 miles Note: The Y-axis represents the number of stores. Stores within two miles only include schools between one to two miles. Stores within one mile only include schools between half mile and one mile. Stores within half mile include schools less than half mile. 28 Table 1. Summary Statistics for all Stores Variable Observations Mean Std. Dev. Min Max Weekly Quantity 3000819 19.464 50.775 -74 4415 Weekly Net 3000819 40.708 90.440 -1267 8658 Price 2997042 2.826 3.117 -24 996 Week ID (yyyyww) 3000819 200887.400 126.588 200648 201113 Product_n 3008915 4205.133 2795.822 1 9747 Store_n 3000819 18.01261 9.598595 1 35 Sale 3000819 0.417 0.493 0 1 Distance- Half Mile 3000819 0.094 0.291 0 1 Distance- One Mile 3000819 0.337 0.473 0 1 Soda 3008915 0.501 0.500 0 1 Open Campus 3008915 0.034 0.181 0 1 Treatment 1 (July 1, 2007 and on) 3008915 0.867 0.339 0 1 Treatment 2 (July 1, 2009 and on) 3008915 0.416 0.493 0 1 DDD- Open Campus 3008915 0.007 0.083 0 1 DDD- Treat 2 (Half Mile) 3000819 0.019 0.138 0 1 DDD- Treat 2 (One Mile) 3000819 0.072 0.259 0 1 Fall06 2994250 0.004 0.064 0 1 Spring07 2994384 0.071 0.257 0 1 Fall07 2995224 0.061 0.239 0 1 Spring08 2994770 0.069 0.253 0 1 Fall08 2994358 0.058 0.234 0 1 Spring09 2993933 0.080 0.271 0 1 Fall09 2994300 0.060 0.238 0 1 Spring10 2993555 0.074 0.261 0 1 Fall10 2993804 0.062 0.241 0 1 Spring11 3008915 0.046 0.210 0 1 DDD- Fall 06 (Half Mile) 2986154 0.000 0.014 0 1 DDD- Spring 07 (Half Mile) 2986288 0.003 0.058 0 1 DDD- Fall 07 (Half Mile) 2987128 0.003 0.055 0 1 DDD- Spring 08 (Half Mile) 2986674 0.003 0.058 0 1 DDD- Fall 08 (Half Mile) 2986262 0.003 0.052 0 1 DDD- Spring 09 (Half Mile) 2985837 0.004 0.061 0 1 DDD- Fall 09 (Half Mile) 2986204 0.003 0.054 0 1 DDD- Spring 10 (Half Mile) 2985459 0.003 0.058 0 1 DDD- Fall 10 (Half Mile) 2985708 0.003 0.053 0 1 DDD- Spring 11 (Half Mile) 3000819 0.002 0.046 0 1 Note: No negative values were included in the data when the regressions were run. These were removed to ensure accurate results. 29 Empirical Model To curb the obesity epidemic, SB 965 and policies like it aim to reduce the availability of CSBs at schools, with the goal of reducing overall CSB consumption throughout the day for school age adolescents. I do not attempt to track individual consumption, but instead observe the change in weekly purchases of sodas in grocery stores located within proximity of a high school compared to stores that are not within proximity to a school and should therefore not see a change in sales from adolescents post-implementation of the policy. If the policy obtains the results initially expected by lawmakers, then the data would not show an increase in purchases of CSBs at grocery stores close to schools as compared to grocery stores without schools close by after implementation; one could conclude that the total consumption of CSBs by high school students might have been measurably reduced. Although this study does not take into account fast-food outlets and other restaurants, or alternative grocery stores, which are also possible locations at which adolescents may obtain CSBs, the compensation effect should at least partially be observed in the purchasing patterns at grocery stores before and after the policy went into effect. The present study should, therefore be able to test for potential compensation outside of schools in response to new regulations. To determine the change in soda purchases in response to the California regulation, I created a treatment variable that compares the time periods before and after the ban. The treatment 1 dummy variable takes the value of 1 for the time period beyond the regulation’s first checkpoint on July 1, 2007; this indicates the period in which schools were required to meet the standards at a 50% compliance rate. The treatment 2 30 dummy variable takes the value of 1 for the time period beyond the regulation’s second checkpoint on July 1, 2009 when the schools were required to be fully compliant with the new regulations. They are interacted with other variables in a triple difference specification to quantify the treatment effect of the policy change by comparing weekly purchases of sodas to other types of beverages pre- and post-implementation of the regulation. Compared to a simple comparison of means in stores located close to schools versus stores that do not have a school close by, this specification allows differences across stores that are close to a school and that are not, and across soda sales versus other beverage sales that are unaffected by the policy at the school-level. The single variables compare the average effect of that variable, whereas the interaction terms account for the trends consistent among both variables. The DDD is a stronger specification because it relaxes the identification assumptions by controlling for time, school-level policies, and soda versus other beverage sales, gaining a measured treatment effect. This is critical because when determining the effect of a policy, it is important to isolate the effect that occurs in response to the policy, instead of outside factors (such as the economy, changes in population size or demographic information). The base model (Model one) is specified as follows: Yijt= 0+β1Xijt+β2Ddistance+β3Dsoda+β4Dtreatment+β5Wijt + β6Zi + 1Dsoda*treatment +2Dsoda*distance +3Dtreatment*distance +1DDD + ijt 31 Yijt is the dependent variable: weekly beverage sales at the product level that is indexed by i, indicating the store; j, the product; and t, the week. The dependent variable is in log form and measures the percentage change in weekly purchases; this variable is preferred over an average of the sales because it is easier to interpret and better understood for all stores regardless of size. A log-linear specification serves to tighten the data to account for many of the differences in the stores. For instance, larger stores may have the same percentage increase in sales, but their weekly quantity will be much higher than that of a smaller store. Thus, logging the dependent variable accounts for the difficulties in interpreting level changes in sales. Additionally, the log-linear specification yields a significantly higher R2, and it is used rather than the other weekly quantity sales dependent variable (although the results not shown support the findings of the reported results). The intercept common to all stores and products is 0. The price variable, Xijt, represents the weekly average price of each beverage purchase at a given week and store. Multiple dummy variables are utilized in the model: Ddistance,i represents stores within a particular distance of a school which should control for the amount of students that may frequent the store for beverage purchases; Dsoda,j is a dummy that is valued at 1 if the item was a soda product; Dtreatment,t is a dummy valued at 1 if the purchase was during the treatment period. The dummy Wijt is a control variable for sale, defined as a 1 for a product at that has been marked down at a specific store (which is store and product specific by UPC), and Zi is the open campus dummy which is 1 for a store that has at 32 least one open campus school within a half mile, regardless if additional schools are proximate but do not have open campus policies. To create the triple difference variable, interaction terms are created: Dsoda*treatment, ijt is the interaction between soda and treatment period which should isolate the DD of a change in soda purchases during the treatment period; Dsoda*distance, ijt is the interaction between the soda and distance dummy variables which controls for a directional change in soda purchases that are due to the proximity of a school regardless of when the items were purchased (within or without the treatment period); Dtreatment*distance, it represents the interaction between treatment period and distance from a store to a school to isolate the effect of the policy with regard to proximity of a school. Finally, DDDijt is the triple difference effect comparing treatment period to before, soda to non-soda purchases, and the distance variable that isolates stores within proximate distance to schools; ijt is the error term. The DDD combines the dummy variables in a triple interaction, and 1 estimates the average change in soda purchases as the treatment effect. A more detailed description of the variables may be found in Appendix 2): Since there are likely to be unobserved effects that are not accounted for in the OLS model, it may be beneficial to employ a regression that accounts for these effects. The random effects model assumes that there is no correlation between the stores and the other explanatory variables, but instead there are some unexplained random differences among the variables included in the regression. It does not create new coefficients, but instead alters the error term, which will in turn adjust the standard errors of the coefficients. Alternatively, the fixed effects model utilizes the available information, 33 such as stores identifiers, and semesters to account for potential differences that are constant over time, or over stores. The difference between the two models is basically whether the time or entity effects reduce the variance in the data or not. To find which specification is appropriate, I turn to a commonly used test: the Hausman Test. This test should determine which model best fits the data but may only be beneficial if differences between stores (such as size, sales, items offered, etc) are significant, and when looking at the data this seems to be the case (Hahn et al. 2011). The fixed effects regression is easier to justify: although some changes may be controlled by additional time fixed effects (which will be dealt with in a later model), the store-level fixed effect can additionally control for the differences across stores that are time invariant. Each store is likely to sell the same number of soda products over time, have a predominantly consistent size and have a similarly concrete sales level; because stores do not change dramatically, they are relatively stable focal points for entity fixed effects. Ultimately, I was able to reject the null hypothesis that the random effects model was appropriate. An additional robustness check utilizes a DDD specification to find the effect of school proximately on the store’s transactions of sodas compared to other types of beverages within different semesters. The seasonal effects represent a control for characteristics that are the same across stores, but potentially vary across time. For instance, in 2008 the economy was reported to be in decline and purchases across all stores should have reacted similarly making a potential decrease in sales consistent in each store (assuming there is a correlation between the economy and food consumption). Controlling for potential seasonal change, especially those that happen each year (specific 34 months: for instance holiday months) potentially further reduces the differences in the sales. It also allows for differentiation between short term and long-term effects of the policy. The semester model removes the treatment variable and instead adds fall and spring semester indicators to track changes over time. Dsemester, t is a dummy variable valued at 1 if the transaction took place within the semester; there is a fall semester and spring semester beginning in fall 2006 and ending in spring 2011. To create the triple difference variable, interaction terms are created in addition to the base model: Dsemester*distance, it is the interaction between each semester and distance dummy variables; Dsemester*soda, ijt represents the interaction between the soda dummy and each semester. Finally, 1…10 represent the effect of the purchases in a semester compared to summer months, soda purchases versus non-soda purchases, and stores within proximity of a school to those not within proximity. Particular differences in purchases during summer and school year may occur since during the summer adolescents are at home for a larger portion of their day and may be more likely to consume beverages in family size packaging rather than individual (the difference between a half gallon of orange juice, or a 12 ounce for instance). Therefore, an additional robustness check for summer is included. I remove the summer months to compare the semesters to a base semester of fall 2006 to exclude periods where the effect should not take place at all. 35 RESULTS AND ROBUSTNESS CHECKS Based upon the econometric specification outlined in the previous chapter, the regressions in this chapter will outline the results of the study. Multiple robustness checks are employed in order to ascertain the effectiveness of all the models congruently. I employ a graphical analysis that may reveal differences between stores. Figures 5-7 show stores further than half mile than a Sacramento high school, within half mile of any campus, and within half mile of an open campus, respectively. The reference lines delineate the two treatment periods to isolate the potential compensation shocks. The graphs indicate similar sales data for the time periods: slight downward trend, but no significant change after July 1, 2007 or July 1, 2009. Additionally, the fluctuations are common among the different store types, and no visible difference between soda or other beverage sales as reflected through changes in transactions. Based upon the figures alone, I would expect no compensation effect because there are no significant spikes in purchases that indicate sharp changes in behaviors. The regression analysis is intended to control for potentially relevant factors such as prices, sales, or store size; all of which may paint a clearer picture of the effect of the regulation. 36 0 0 10 10 20 20 30 30 40 40 Figure 5: Soda vs. Other Beverage Sales for All Stores in the Data- Not Within Half Mile 2007w1 2007w1 Week Week 2008w1 2009w1 2008w1 Treatment 1 2009w1 2010w1 2010w1 Treatment 2 2011w1 2011w1 (mean) sodaweeklyqty Average Weekly Quantity- Soda Average (mean) Weeklyweeklyupcqty Quantity- All Beverages NOTE: the top line is All Beverages; the bottom line is soda. 0 10 20 30 40 Figure 6: Soda vs. Other Beverage Sales for Stores Within Half Mile of a School 2007w1 Week 2008w1 Treatment 1 2009w1 2010w1 Treatment 2 Average Weekly Quantity- Soda Average Weekly Quantity- All Beverages NOTE: the top line is All Beverages; the bottom line is soda. 2011w1 37 0 10 20 30 40 Figure 7: Soda vs. Other Beverage Sales for Stores Within Half Mile of an Open Campus 2007w1 Week 2008w1 Treat 1 Begins 2009w1 2010w1 2011w1 Treat 2 Begins Average Weekly Quantity- Soda Average Weekly Quantity- All Beverages Note: the top line is All Beverages; the bottom line is soda. Before I discuss the results of the regression analysis, it is worth mentioning that there are common issues that accompany DD and DDD regressions. To prevent potential problems often associated with these specifications, I employ techniques as suggested by Bertrand et al. (2003), such as robust standard errors to ensure proper specification and valid test-statistics. Finally, serial correlation in the variables could lead to t-statistics that reject the null because they are close to |2|. However, I analyze each coefficient estimate for potential bias; I find that the majority of the t-statistics are well above that threshold thus strengthening the findings and providing evidence against serial correlation in my variables. 38 Regression Results The first model (Table 2, Column (1)) reports the results for the base model with no fixed effects and no seasonal dummy. The second treatment variable (treatment 2) is characterized by a requirement for full compliance whereas the first treatment required only 50% compliance; thus the first treatment is distinguished by varying compliance rates determined at the school level. Although the first treatment period is not reported, the coefficients were similar to that of the second treatment, which is utilized because it reflects full compliance and yields significant results (whereas the first treatment does not have significance in the pertinent variables). In this model, the dependent variable is weekly net sales, which is the net dollar amount of sales of each product in a particular store for a given week. Stores within one mile to a school are compared to those without. Having a school within one mile is correlated with an increase in the net sales of beverages sold at those stores close to schools. Although negative, suggesting no compensation but instead a change in preference, the DDD variable is not significant. The next model utilizes a half mile indicator, which follows more closely to previous research.2 The second model (Table 2, Column (2)) compares stores within a half mile to those without. The triple difference is negative and significant; having a school within half mile decreases the sales of soda compared to other beverages at the proximately located store by 4.1% in the treatment 2 period. Since the DDD is statistically significant and isolates the effect of the ban alone, this specification might be preferred: the half mile 2 The shorter distance of a quarter mile was specified in the Currie (2010) study, but was not feasible in this study because of too few data points associated with that specification. 39 distance is better accessible to students and more likely to see a change in sales in reaction to a school policy change. The dependent variable of quantity instead of net sales is used in the next models and provides a much better fit for the data. However, in order to utilize this specification, one caveat must be emphasized: the quantity measure indicates a 1 for a single serve purchase as well as a 1 for a twelve-pack purchase. Due to this data issue, the assumption that the pattern of purchases before and after the regulation remain constant, meaning students and parents purchase similar beverage sizes before and after implementation and is in interpreting the results of this regression. This seems to be a reasonable assumption because students are likely purchasing soda directly before or after school and more likely to pick up a single-serve option. Per unit, twelve-packs are usually cheaper, so a student who takes advantage of this cheaper per unit cost likely did so before the regulation as well and therefore their purchases of particular beverage sizes should be unaltered before and after the treatment begins. The next three models (Table 2, Column (3,4,5)) all utilize the dependent variable of ln weekly quantity sales, and are consistent with each other, though they undertake different strategies. Column (3) uses the one mile specification; the triple difference estimator is slightly negative, but with no significance, which confirms that the half mile specification has more information about the relationship between schools and grocery transactions and will from this point be used as the proximate distance between a school and store. Additionally, using this dependent variable, the price coefficient is negative and is now consistent with basic supply and demand theory. Oddly, the sale coefficient is 40 negative indicating that an item being on sale will decrease the quantity sold. This may be due to the caveat in the former paragraph: number of units in the package may play a large role, encouraging people to buy packages with more units as opposed to more separate packages of the same product (example: one twelve pack over two six-packs). Models four and five are specified identically except that five has a control for time invariant characteristics that should better isolate the effect of the school on store transactions as opposed to store characteristics that do not vary over time. These storelevel fixed effects are employed to account for differences in store-level characteristics such as size or demographics. In both models four and five, the treatment period is consistent with the previous models outlined in Table 2 and shows fewer sales were made during that time period, with a corresponding increase in soda sales of 12%. The triple difference estimators for both the base model (Table 2, Column (4)) and the store-level fixed effect model (Table 2, Column (5)) interact stores within half mile to a school, comparing soda to non-soda beverage purchases, further differencing the before and after the treatment 2 time period. The coefficient on the DDD directly contradicts the theory that children compensate outside of schools, and actually finds a decrease in consumption of the regulated beverages by 4.5% (which is significant at 99% level of confidence). 41 Table 2. The Effect of a Ban on Sodas in Schools on Out-of-School Purchases Variables Withinhalf Within1 Price Sale Treat2 Soda Sodatreat2 (1) Dependent Variable: ln (weekly net sales) 0.090*** (0.003) 0.057*** (0.003) -0.280*** (0.002) -0.075*** (0.003) 0.156*** (0.002) 0.086*** (0.003) Sodahalfmile Soda1mile DDD Treat 2 (One Mile) (3) Dependent Variable: ln (weekly qty sales) 0.078*** (0.003) -0.085*** (0.005) -0.331*** (0.002) -0.060*** (-0.003) 0.114*** (0.002) 0.075*** (0.004) -0.025*** (0.004) -0.020*** (0.004) -0.004 (0.006) -0.009 (0.006) -0.041*** (0.010) NO 2989661 (4) Dependent Variable: ln (weekly qty sales) -0.032*** (0.005) (5) Dependent Variable: ln (weekly qty sales) -0.157*** (-0.007) -0.085*** (0.005) -0.331*** (0.002) -0.069*** (0.002) 0.121*** (0.002) 0.076*** (0.003) 0.005 (0.006) -0.085*** (0.005) -0.310*** (0.002) -0.056*** (0.002) 0.121*** (0.002) 0.074*** (0.003) 0.008 (0.006) 0.032*** (0.007) 0.022*** (0.007) -0.045*** (0.010) -0.045*** (0.010) NO 2996037 YES 2996037 0.022*** (0.004) 0.025*** (0.007) DDD Treat 2 (Half Mile) Store Fixed Effect Observations 0.057*** (0.003) -0.281*** (0.002) -0.085*** (0.002) 0.160*** (0.002) 0.089*** (0.003) -0.001 (0.006) 0.013*** (0.004) Withinhalftreat2 Within1treat2 (2) Dependent Variable: ln (weekly net sales) -0.018*** (0.005) NO 2989661 NO 2996037 R2 0.041 0.040 0.067 0.066 0.091 NOTE: Soda is a dummy indicating a soda purchase, Treat 2 is a 1 for sales made post July 1, 2009; within half and within1 are half mile and one mile distances (between a store and school) respectively. Price is the cost at which the item was purchased and sale indicated if it is purchased on sale. The DDD estimator is the soda dummy interacted with treatment 2 and distance. This coefficient quantifies a change in soda sales compared non-soda sales in the treatment period compared to the control for stores within a proximate distance to a school to those without. Robust standard errors are utilized in every regression to ensure that the problems of DD may be limited and the standard errors do not overestimate the actual t-value and thereby reduce the chance of type I error (Bertrand et al. 2003). Standard errors are reported in parenthesis and levels of significance are denoted as follows: ***(1%), ** (5%), * (10%). 42 Robustness Check: Open Campus Identification Sacramento has starkly different neighborhood types and different communities will likely have different purchasing patterns that will be reflected at the store within proximity to that community. In the final regression of the basic model (Table 2, Column (5)) the store-level fixed effect seemed to better explain the effect of the policy by controlling for differences across stores that are time invariant. Next, using a half mile proximity for school-to-store distance, I add a control for open campus that may further the results and dichotomize the regulation’s impact by different school-level policies. The indicator for an open campus is added to this model to isolate the compensation when students are provided opportunities to leave campus during school hours. The first model (Table 3, Column (1)) includes both the store dummies for half mile to any school and half mile to an open campus, whereas the second model (Table 3, Column (2)) takes out the half mile specification since the open campus variable already accounts for half mile, and the DDD describes sales at a store with an open campus policy within half a mile. For both models, there is about 12% more soda sold compared to other beverages. The Treat 2 time period is characterized by a decrease in overall beverage sales. Strikingly, a store with an open campus within half mile, looking only to a mean comparison, experiences 44% more soda sales than stores without an open campus near. This seems excessive and may be due to other external factors such as strategic locations of stores, schools, and housing opportunities as opposed to student purchasing patterns. It is also possible that open campus policies are often at schools where students are close to home; therefore the campus is close to a high concentration of 43 homes, which are also within half mile to a store. For the first model, that includes the control for half mile to any school, the triple difference estimator comparing a store within half mile to any school is significant at 1% and illustrates a 4.1% decrease in consumption of sodas compared to other beverages in the second treatment period, which is strikingly similar to that of the second model of Table 2 and provides additional evidence of non-compensation. However, the open campus DDD illuminates the regulation’s different impact based upon school policy: there is a 9.8% increase in soda sales compared to other beverages in the second treatment if the store was within half mile to an open campus. This potentially suggests a compensation effect when students are able to leave the school campus during school hours. This is confirmed by the second model where the DDD estimator finds that 10% more sodas were sold compared to non-sodas during the treatment period for the treatment group (this variable is significant at 1%). This is slightly higher than the previous specification, which is probably because some of the effect is due to stores being within half mile of a school, and that effect is negative. Therefore, stores close to an open campus will see a higher magnitude change compared to all other stores. The store fixed effects specification controls for store-level characteristics that do not vary over the time of our data; by holding these characteristics constant, I can isolate the change that is specific to the ban as opposed to differences between stores. The store fixed effect specification yields significant results and intuitively seems necessary; additionally, the open campus policy should identify a difference in behavior among adolescents and consequently explain some of the change in sales over time. It appears 44 that open campus policies may be the driver for much of the difference in sales because students have a low barrier to entry into alternative markets (transportation costs, time constraints, etc.) and this phenomenon will be investigated further in an additional robustness check that investigate the store transactions by time and proximity to the open campus school. There is one major limitation in using the open campus within half mile of a store: this school is a small private school. Upon further review of this private school, I find that it is small and have not changed their policies with regard to soda. They offer soda on campus and still do to this day which seriously undermines the findings of compensation. However, this was the only school within a reasonable distance to a store that would provide students with the opportunity to engage in compensatory behavior. Additionally, the treatment store is also close to other schools with open campus policies which may be driving some of this change in sales during the treatment period, though the other open campus schools are continuation high schools that do not require students to be on campus all day. Although these results are tenuous at best, they were included to provide some context for the differences in school policies. When I use the transactionlevel data, I will expand the distance to allow for the other open campus schools to be included, and differentiate between hours before and after school is in session. 45 Table 3. The Effect of a Ban on Sodas in Schools on Out-of-School Purchases Utilizing Store Fixed Effects Dependent variable is ln(weekly quantity sales) (1) (2) Variables D.V. : ln (weekly qty sales) D.V. : ln (weekly qty sales) Withinhalf -0.154*** (0.007) Price -0.085*** -0.085*** (0.005) (0.005) Sale -0.310*** -0.310*** (0.002) (0.002) Treat2 -0.050*** -0.048*** (0.002) (0.002) Soda 0.123*** 0.124*** (0.002) (0.002) Open Campus 0.435*** 0.435*** (0.008) (0.008) Opensoda -0.051*** -0.051*** (0.010) (0.010) Opentreat2 -0.154*** -0.156*** (0.012) (0.011) Sodatreat2 0.070*** 0.066*** (0.003) (0.003) Sodahalfmile 0.006 (0.006) Withinhalftreat2 0.016** (0.007) DDD- Treat 2 -0.041*** (0.010) DDD- Treat 2 (Open) 0.098*** 0.102*** (0.016) (0.016) Store Fixed Effects YES YES Observations 2996037 2996037 R2 0.091 0.091 NOTE: Soda is a dummy indicating a soda purchase, Treat 2 is a 1 for sales made post July 1, 2009; withinhalf is the half mile distance between a store and school. Price is the cost at which the item was purchased and sale indicated if it is purchased on sale. The DDD estimator is the soda dummy interacted with treatment 2 and an Open Campus within half mile. This coefficient quantifies a change in soda sales compared non-soda sales in the treatment period compared to the control for stores within a half mile of a school with an open campus policy to those without. Robust standard errors are utilized in every regression to ensure that the problems of DD may be limited and the standard errors do not overestimate the actual t-value and thereby reduce the chance of type I error (Bertrand et al. 2003). Standard errors are reported in parenthesis and levels of significance are denoted as follows: ***(1%), ** (5%), * (10%). 46 Robustness Check: Semester Effects Since students are only in school for certain months in the year, it is therefore important to control for the seasonal changes in consumption of CSBs. Furthermore, it will be interesting to differentiate between long and short term responses to these policy changes. In order to evaluate the differences in soda purchases in response to the California regulation while taking into account the seasonal characteristics of transactions, I created additional dummy variables that split the school year into fall and spring semesters. I define the spring semesters as starting in February and going through the end of April. The beginning and end of the semester varies by school, which is why January and June were excluded from the spring semester. Additionally, only three months are observed to ensure comparability with the fall semester, which runs from September through November so as to exclude August when some schools are in session for only a short portion and December (in which student attendance varies greatly due to the holidays) for the same reason. They are interacted with the half mile distance variable and the soda indicator to create a DDD estimator that quantifies the changes in purchases of sodas at local grocery stores specific to that particular semester. The DDD estimators for stores with any school within half mile suggest preference change for the semesters in the treatment 2 period, although there is only significance for the semester immediately following the implementation of the regulation. However, the store with the open campus near sees significant changes for multiple semesters following the regulation. Starting in the fall of 2007, there is what appears to be an initial preference change as indicated by the negative coefficients that continue 47 through fall 2008. Yet, in spring of 2009, the indicator switches signs and shows over 10% compensation that persists through spring of 2010. The next three models use the fall 2006 semester as a stable base that will allow a mapping of the change in sales relative to that semester. The first model (Table 4 Column (1)) illustrates the semester effects without taking into consideration open campus policies. The coefficients on the DDD estimators were predominantly negative, but not significant. Recall that July 1, 2009 is the date of full implementation of the regulation, meaning that schools were to be 100% compliant by that date. Since most schools are not in session in July, the next semester should be the most substantial in terms of compensation. However, the fall 2009 DDD is negative and significant at 5%, although these results are isolated to that one semester, and there is no significance thereafter. This illustrates a 5% decline in sodas purchased compared to other beverages in grocery stores near the school during the same time periods previously. This strongly supports the results found in all the models. This DDD is the estimator for all schools regardless of their decision to participate in an open campus policy and although the results are not significant, they are consistent with previous models. The second seasonal model (Table 4, Column (2)) includes the open campus effect but does not include the within half distance because it is already included in the open campus variable. This specification compares the open campus store to all other stores. The results are almost identical to that with the control for half mile school-tostore distance. Similar to the fixed effects model, having an open campus near the store increases beverage sales by nearly 38%. The DDD however, identifies an increase of 48 13.5%, a slight increase from the previous model, indicating that there is an increase in soda sales after implementation at stores proximate to open campus schools immediately following implementation of the regulation. Based upon these results, the open campus specification supports the hypothesis of compensatory behavior by students. The data suggests that open campus policies provide the necessary flexibility for students to overcome the opportunity costs of going elsewhere to purchase beverages. Predominantly negative before the policy was fully implemented, there does seem to be some compensation beginning in the spring of 2009 as indicated by the positive coefficient (right before full implementation was required). At that time, there was a compensatory effect for the store with an open campus within half a mile, increasing the quantity sold by 12%. Since the results are only significant two semesters after the regulation and dissolve by the Fall 2010 semester, they are temporary at best and are only indicative of compensatory behavior in students who attend schools with an open campus policy. Although the previous models found no compensation, and in fact a decline, this model suggests potential short-term compensation. However, out of 54 responses from schools, only four schools had open campus policies. This represents less than 8% of schools. The other regressions reflect a possible change in preferences by students who do not have the flexibility to leave campus during school. It is worth pointing out once more, however, that upon further investigation, it is unclear if students in the open campus school did experience limited availability of sodas. This result therefore warrants further investigation into open campus policy schools. 49 The third model removes the within half variable and the corresponding permutations utilizing this variable. The results are almost identical: this model illustrates the potential explanatory strength of an open campus. With or without taking into account schools within proximity, an open campus seems to be the best indicator of an increase in sales at grocery stores near schools post-implementation of the regulation. The three previous models included all available data for the time period. The DDD variables were therefore comparing the sales in the particular semester to all other times, including summer and holidays. As a final robustness check, I remove all the summer and other time periods. The base semester for the next two regressions will be Fall 2006, the first semester of data available. The fourth model is similar to the first semester regression, but with the data removed as indicated previously. This model shows no significance in the DDD. The fifth model, although similar direction and magnitude as model 2 of the semester regressions, sees no significance in the DDD variables. The first three models include data for the entire year, but isolate the effects during the summer. Using the summers and non-school year as a base, there does seem to be some effect on outside purchases of a ban on soda sales in schools. However, you would expect these results to persist when using a control month as the base, omitting all other time periods not in question. Looking to models 4 and 5, the DDD estimators are no longer significant. These results are probably due to unexplained fluctuations not considered in the data; but the results of the last two models strongly question the conclusion that there is an increase in soda sales post implementation of the regulation. 50 Table 4. The Effect of a Ban on Sodas in Schools on Out-of-School Purchases Utilizing Semester Indicators Variables Withinhalf Price Sale Soda Dependent variable is ln(weekly quantity sales) (1) (2) (3) 0.352*** (0.008) -0.085*** (0.005) -0.306*** (0.002) 0.146*** (0.002) opencampus opensoda Fall06 Spring07 Fall07 Spring08 Fall08 Spring09 Fall09 Spring10 Fall10 Spring11 DDD Fall 06- (Open) DDD Spring 07- (Open) -0.600*** (0.015) -0.014*** (0.004) 0.055*** (0.005) 0.013*** (0.004) 0.004 (0.005) -0.016*** (0.004) -0.041*** (0.005) -0.041*** (0.004) -0.051*** (0.005) -0.088*** (0.005) 0.353*** (0.008) -0.085*** (0.005) -0.306*** (0.002) 0.147*** (0.002) 0.379*** (0.010) -0.014 (0.013) -0.599*** (0.016) -0.016*** (0.005) 0.052*** (0.005) 0.008* (0.005) -0.001 (0.005) -0.011*** (0.004) -0.034*** (0.005) -0.037*** (0.004) -0.049*** (0.005) -0.083*** (0.005) 0.006 (0.116) -0.010 (0.033) -0.085*** (0.005) -0.306*** (0.002) 0.146*** (0.002) 0.378*** (0.010) -0.013 (0.013) -0.597*** (0.015) -0.016*** (0.004) 0.049*** (0.005) 0.008* (0.004) -0.001 (0.005) -0.014*** (0.004) -0.033*** (0.005) -0.037*** (0.004) -0.047*** (0.005) -0.082*** (0.005) 0.011 (0.116) -0.013 (0.033) (4) (5) -0.126** (0.050) -0.083*** (0.007) -0.288*** (0.002) -0.597*** (0.020) (Omitted) -0.125** (0.050) -.0083*** (0.007) -0.288*** (0.002) -0.060*** (0.021) 0.379*** (0.089) -0.007 (0.115) (Omitted) 0.590*** (0.016) 0.659*** (0.016) 0.617*** (0.016) 0.607*** (0.016) 0.587*** (0.016) 0.564*** (0.016) 0.564*** (0.016) 0.554*** (0.016) 0.515*** (0.016) (Omitted) 0.589*** (0.016) 0.657*** (0.016) 0.612*** (0.016) 0.605*** (0.016) 0.592*** (0.016) 0.571*** (0.016) 0.568*** (0.016) 0.556*** (0.016) 0.520*** (0.016) (Omitted) -0.016 (0.119) NOTE: Robust standard errors are utilized in every regression to ensure that the problems of DD may be limited and the standard errors do not overestimate the actual t-value and thereby reduce the chance of type I error (Bertrand et al. 2003). Standard errors are reported in parenthesis and levels of significance are denoted as follows: ***(1%), ** (5%), * (10%). 51 Table 4. (Continued) Variables DDD Fall 07- (Open) DDD Spring 08- (Open) DDD Fall 08- (Open) DDD Spring 09- (Open) DDD Fall 09- (Open) DDD Spring 10- (Open) DDD Fall 10- (Open) DDD Spring 11- (Open) (1) (2) -0.072** (0.034) -0.116** (0.032) (3) -0.074** (0.034) -0.115*** (0.032) -0.092*** (0.035) 0.122*** (0.030) 0.135*** (0.033) 0.097*** (0.030) -0.036 (0.033) 0.005 (0.039) -0.093*** (0.035) 0.121*** (0.030) 0.139*** (0.033) 0.099*** (0.030) -0.035 (0.033) 0.006 (0.039) (4) (5) -0.079 (.0.119) -0.123 (.119) -0.099*** (0.012) 0.115 (0.118) 0.128 (0.119) 0.090 (0.118) -0.044 (0.119) -0.002 (0.121) (Omitted) (Omitted) -0.049 (0.067) 0.084 0.084 DDD Spring 07 0.035* (0.069) (0.069) (0.019) 0.068 0.065 DDD Fall 07 0.019 (0.069) (0.070) (0.020) 0.042 0.038 DDD Spring 08 -0.006 (0.069) (0.069) (0.019) 0.071 0.067 DDD Fall 08 0.022 (0.070) (0.070) (0.021) 0.067 0.071 DDD Spring 09 0.018 (0.069) (0.069) (0.019) -0.003 0.003 DDD Fall 09 -0.051** (0.069) (0.070) (0.020) 0.024 0.028 DDD Spring 10 -0.024 (0.069) (0.069) (0.019) 0.030 0.028 DDD Fall 10 -0.019 (0.070) (0.070) (0.021) 0.045 0.045 DDD Spring 11 -0.005 (0.070) (0.071) (0.024) YES YES Store Fixed Effects YES YES YES 1748060 1748060 Observations 2864606 2864606 2864606 .090 .090 R2 0.091 0.091 0.091 NOTE: Robust standard errors are utilized in every regression to ensure that the problems of DD may be limited and the standard errors do not overestimate the actual t-value and thereby reduce the chance of type I error (Bertrand et al. 2003). Standard errors are reported in parenthesis and levels of significance are denoted as follows: ***(1%), ** (5%), * (10%). DDD Fall 06 52 Robustness Check: Transaction-Level The results reported previously suggest potential compensation depending on school policy. As an additional robustness check, I returned to the transaction-level data to ascertain the hours in which students were compensating. The survey provides information regarding start and end times of the lunch period, and school hours. I combined these details with purchase data and aggregated by the hour rather than the week. The results found previously suggest that open campus policies seem to be the driver for most of the possible compensation. Therefore, a robustness check isolating open campus schools as the treatment group was utilized in this analysis. Starting with all four open campus schools, I returned to the geocode information and determine the grocery stores within my sample that are closest. However, for these regressions, I expanded the distance to 5.14 miles, in order to include all stores that are close to an open campus school. Additionally, it is important to reiterate that the open campus school that was under half mile from a store was very small and a private school. Additionally, the school continues to serve soda on the school grounds after the implementation of the regulation and to this day. However, as stated previously: two other open campus schools are also located close to that store, and might have affected the results reported. Therefore, that store was included in the robustness check, along with a second store that is close to the fourth and final open campus school. An additional control store was selected at random from all stores that did not have a Sacramento high school within 5.14. 53 The dependent variable is log of hourly sales by quantity of hourly sales which is determined by store, product, and week. The log was utilized once again to obtain a consistent measure for each store regardless of size; therefore the results can be expressed as percentage increases in the dependent variable instead of a quantity of hourly sales which varies significantly across stores. Because of varying school-times, a one-hour specification could not be made for the before3, after4 or lunch5 times and two hours is used instead. To obtain the number of units, and a hands-on account for the soda category, I went to the store with the three open campus schools near in order to find unit information and number of ounces per container, which was not included in the previous regressions. As expected, when accounting for number of units in each package, the coefficient for the price variable as indicated by per unit price is negative. This follows with basic economic theory and is consistent through all the models at the transactionlevel. Consistent among all the models employed in this robustness check, when a product is on sale, hourly purchases increased by 28-29%. Additionally, in treatment 2, the quantity sold was about 3% lower than the previous time period. Models 4 through 6 investigate the comparison of school when they are in session by removing summers and weekends. In the first model, the differences estimator is a quadruple difference (D4) that compares purchases before school to those made at all other times (not including the two hour period after school and during lunch), soda to 3 The schools had the following start times: 8:30 am, 7 am, 8:20 am, and 8:30 am. Therefore, the beforeschool dummy was a 1 if the purchases were made between 6 and 7:59 am inclusive. 4 The schools had the following end times: 3:30 pm, 4 pm, 3:06 pm, and 3:30 pm. Therefore, the afterschool dummy was a 1 if the purchases were made between 3 pm and 4:59 pm inclusive 5 The schools had the following lunch times: 12:15-12:45 pm, 12-12:30 pm, 12:26-1:02 pm, 11:30 am12:30 pm. Therefore, the lunchtime dummy was a 1 if the purchases were made between 11 am and 12:59 pm inclusive. 54 non-soda purchases, stores with an open campus proximate to a store with no schools near, and finally purchases made in the treatment 2 period to those made before the regulation took effect. The D4 estimator finds a weak effect of the before school period at 4.9% in model 1 and 6% in model 4 with 90% and 95% confidence respectively. The model without the summers or weekends provides stronger evidence of compensation; according to these results, students may be purchasing sodas in the hour before school, and the stores are seeing a 6% increase in their hourly sales during this time period. This is slightly lower than previous results from Table 3 that found about a 10% increase in sales for stores within a half mile to an open campus. This is likely due to the base specification change; the base in the previous models included summers and weekends where soda sales might be lower due to traveling, or alternative beverages purchased that are more refreshing like lemonade or juices. However, it is likely that more purchases would occur either at lunch time or after school when the students are already out of the house. While an open campus policy should affect the school lunch time period, after school is a reasonable time for students to make purchases at local grocery stores on their way home from school. The results of the after school regressions, Models 3 and 6, are striking: the stores experience between 14% and 18% increases in their hourly sales of sodas. The 18% figure compares the school time periods to a base of summers and weekends which were excluded from the model. These results persist with the base of all times when students should not have access to sodas (while in school, well after school, etc.). These results suggest that there may be some compensation by students in response 55 to the regulation. Finally, open campus policies which allow students to leave campus during lunch should provide students the opportunity to compensate at lunch time as well. However, the results of the lunch models do not support a compensation theory. In fact the models displays decreases in the hourly sales during this time period of about 2.5%. This follows with both the base period of non-school days and the model that includes every day of the year. However, these results are only significant at the 90% level of confidence. It is possible that students whom attend open campuses experienced changes in their preferences which they carried out to their off-campus consumption locations. However, if they were able to leave campus to purchase lunch, it begs the question of why they would change their preferences with no functional removal of the options. However, the D4 results are weak and this theory is not supported by any research- only conjecture. Couple these results with the other models in this robustness strategy, preference change seems very unlikely at schools that have open campus policies. Overall, students do seem to compensate; although the store-to-school distance is much further than is supported by research, students would still have access to these stores, and at least these results show a portion of the compensation. The results suggest some students do in fact consume at grocery stores, and this sub-population may be enough to drive up hourly sales at these proximate stores by almost 20%. Additionally, it is not clear how many of these students take advantage of the open campus policy and whether they go to a grocery store during that time. Another possibility is that since it is likely that adults take their lunch in the middle of the day, that there are unaccounted decreases in soda consumption during the same period; information effects of media 56 representation of the issue, new healthy eating habits that were legislated for vending machines and businesses, and calorie postings in California businesses were all potential contributors to adult decreases in consumption of CSBs through limited offerings or potential preference changes. However, buses and other forms of transportation after school may explain why the five mile distance is not deterring students: the location of a student’s home may be closer to the store than the school. These results are strong, especially for the after school time period and suggest an even larger compensation than did the previous results. 57 Table 5. The Effect of a Ban on Sodas in Schools on Out-of-School Purchases Utilizing Hourly Data Variable perunitprice singleserve sale treat2 soda afterschool beforeschool lunchtime opensoda opentreat2 sodatreat2 afteropen beforeopen lunchopen D4 (After) D4 (Before) D4 (Lunch) Dependent variable is ln(hourly quantity sales) (1) (2) (3) (4) (5) -0.048*** -0.048*** -0.048*** -0.053*** -0.053*** (0.001) (0.001) (0.001) (0.002) (0.002) -0.119*** -0.118*** -0.118*** -0.113*** -0.113*** (0.002) (0.002) (0.002) (0.003) (0.003) 0.292*** 0.293*** 0.293*** 0.284*** 0.284*** (0.001) (0.001) (0.001) (0.001) (0.001) -0.031*** -0.031*** -0.034*** -0.030*** -0.030*** (0.003) (0.003) (0.003) (0.003) (0.003) -0.084*** -0.083*** -0.084*** -0.086*** -0.086*** (0.003) (0.003) (0.003) (0.003) (0.003) -0.255*** -0.251*** (0.016) (0.017) -0.009 0.004 (0.006) (0.007) 0.010** (0.005) 0.091*** 0.091*** 0.087*** 0.089*** 0.089*** (0.004) (0.004) (0.004) (0.004) (0.004) 0.006* 0.006* 0.006* 0.006* 0.006* (0.003) (0.003) (0.003) (0.004) (0.004) -0.074*** -0.075*** -0.075*** -0.072*** -0.073*** (0.005) (0.005) (0.005) (0.005) (0.005) 0.075*** 0.073*** (0.016) (0.018) -0.079*** -0.083*** (0.007) (0.007) 0.002 (0.005) 0.149*** 0.182*** (0.055) (0.059) 0.049* 0.060** (0.029) (0.030) -0.024* -0.026* (0.014) (0.015) (6) -0.054*** (0.002) -0.113*** (0.003) 0.284*** (0.001) -0.033*** (0.003) -0.087*** (0.004) 0.006 (0.005) 0.086*** (0.004) 0.007* (0.004) -0.073*** (0.005) 0.005 (0.005) Store Fixed Effects YES YES YES YES YES YES Observations 827996 827996 827996 715207 715207 715207 R2 .101 .101 .101 .099 .099 .099 NOTE: Robust standard errors are utilized in every regression to ensure that the problems of DD may be limited and the standard errors do not overestimate the actual t-value and thereby reduce the chance of type I error (Bertrand et al. 2003). Standard errors are reported in parenthesis and levels of significance are denoted as follows: ***(1%), ** (5%), * (10%). 58 Table 6 outlines the results of all the models employed in this study. Stores within half mile of a school did not see an increase in soda sales compared to other beverages during the treatment 2 time period. However, if the store was near an open campus specifically (as compared to any other school) there was almost a 10% increase as indicated by the DDD estimator. Finally, these results are not consistent nor conclusive when controlling for seasonal effects. Table 6. Summary of Final Results for All Models Variable: DDD (Half Mile) Model Log-Linear (Base) With Store-level Fixed Effect DDD (Open Campus) (- 4.1% to - 4.4%) N/A (- 4.1% to - 4.5%) (9.8% to 10.2%) D4 (Open Campus After school: (14.9% to 18.2%) Before School: (6%) Lunch-time : Weak results With Semester Effects Spring07: 3.5% Fall09: - 5.1% Fall07 to Fall08 (- 7.2% to - 11.6%) Negative thereafter (not significant) Spring09 to Spring10 (9.7% to 13.9%) Varies thereafter with no significance 59 CONCLUSION The school environment is the most important location for students in terms of where they obtain their calories. Public policies regarding regulation of CSBs have already been implemented at every level of adolescent’ schooling; opponents state that the effects on school budgets outweigh any benefits that come from these obesity-cutting measures. This study investigates the effect of a statewide ban on selling CSBs in California high schools and the extent to which this leads to compensation by adolescents. This paper takes a local focus, isolating the urban city of Sacramento as local indicator of the extent of the success of the ban at reducing adolescent consumption of soda in multiple markets. I investigate potential change in preferences or market compensation of students in response to a statewide ban on soda (SB965) sales in Sacramento high schools. I take weekly beverage sales at grocery stores within half mile to a local high school, and compare the sales in the treatment period and the control, while also comparing sodas versus other beverage sales in addition to open campus policy of the school within close range of the store. Including store-fixed effects to account for inherent differences among the stores, I find significant changes in weekly sales at stores within proximity to any school type although the school-level policy changes the magnitude and direction of the variable. I employ a difference-differences (DD) and triple difference specification (DDD). The base model utilizes a DDD estimator to determine the level of compensation, if any, of the ban on soft drinks in high schools. While these results will not determine the effect of SB 965 on obesity, it will instead highlight the compensation made by students 60 through grocery store transactions in response to the regulation. In addition, the seasonal effects clearly distinguish open and closed campuses by their purchasing patterns. The semester immediately following implementation is correlated with a corresponding and persistent decrease in sales for stores within half mile of a closed campus and over a 13% increase in sales for stores close to an open campus. For both groups, however, the coefficients are negative and insignificant by fall 2010. This effects found in stores near open campus schools only confirm a short term compensation effect that may be reflective of immediately responsive students that had access to CSBs the year prior. With the dissipation of the effect over time, it is possible that the students that begin attending high school, who did not have access to CSBs at their middle school, do not respond to the regulation because there was no de facto change in availability for these students: they never had access to them. So, although having an open campus within proximity to a store might play a role in determining compensatory behavior, whether this compensation is long-term will need to be further investigated to validate or renounce regulation aimed at reducing caloric consumption by adolescents. The final robustness check investigates the data at the hourly transaction-level. This robustness check was designed to determine if the data shows that students at open campus schools are in fact compensating by estimating a quadruple difference estimator for times where students would be on their way to or from the campus. The D4 estimator estimates an almost 20% increase in sales for the two hour period after the school day ends. Although not as large, there is a 6% increase in sales in the two hour period before schools begin. Previous results suggested about a 10% increase in sales from students in 61 response to SB 965 that were driven by an open campus policy that allows students to leave campus during the school day, the hourly transaction results do not support the theory that these students are leaving during lunch to obtain beverages that they do not have access to on campus, however. The D4 estimator for lunch times was negative and may have been biased by other patrons that have lunch hour at the same time who possibly changed their preferences for some reason that was not controlled for by this study. However, there is strong evidence to suggest that students are in fact compensating before and after school. It appears that during the school day, the regulation is successful in reducing CSB consumption; whether this results in an overall decrease in calories consumed by these students is unclear. My results do not provide consistent support for the arguments made by the opponents of the policy. Although there is some compensation, these results are found only in stores with open campus schools within proximity. Because open campuses represent less than 8% of the sample and the Sacramento open campus is not a good treatment entity, the vast majority of the schools compensatory effect should be predominantly determined by the models without an open campus DDD. The DDD is negative for standard campuses, and indicates that student preferences have changed due to the policy and this preference spills over to their off-campus opportunities. Fernandes (2008) finds that a ban on soft drinks does reduce the number of calories consumed while at school; coupling his findings with the results of this research suggests an overall decrease in calories because the majority of high school students do not compensate for sodas outside of the school. Although, the DDD estimator in this case isolates sodas and 62 not other CSBs; if there was a systematic shift to sports or energy drinks which were not categorized as soda, then the validity of the former statement is in question because it is possible (although unlikely) that these sports drinks would counteract any increases in soda. This strongly suggests that policies aimed at removing high calorie items may be successful in reducing the number of overall calories consumed by adolescents in and out of the school environment. There is a possibility that, given the opportunity, students will obtain sodas from alternative sources as indicated by the open campus results. However, when constrained (according to our sample, over 92% of schools have such constraints), it appears that adolescent behavior outside of school may actually be augmented through public policy. Where stores with an open campus near are seeing a 10-14% increase in soda sales compared to other beverages during the post-regulation period, the schools within half a mile without that policy saw a 5% decline in that same factor. In addition, this research may be enhanced by comparing public versus private schools to a store instead of simply using location to any school. Since private and charter schools were only subjected to the regulation if they participate in the National Lunch Program, an additional variable controlling for this factor may potentially alter the results reported here. Including demographic information may also improve the outcome of the study and better describe the impact of the regulation on populations traditionally threatened by a lack of nutritional education and the obesity epidemic. Though these results adequately display the direction of the effect of the regulation, additional controls may be necessary to understand more extensively the magnitude of the effect. A control 63 for fast-food restaurants in the area around schools may do well to increase the accuracy of the specified models. Further research may include an in depth analysis of the store with an open campus within half a mile. In summary, my results suggest that public policy may play an important role in altering eating habits in adolescents that may be resulting in obesity. While these results do not describe any calorie changes, previous research suggests that reducing consumption at school alone, regardless of preference changes that spillover to alternative choices, may reduce overall calories consumed in a day (Briefel et al. 2009; Fox et al. 2010; Frieden et al. 2010; Mancino et al. 2010). If these results are only a portion of the overall response, then students may be changing their consumption patterns at alternative locations as well. 64 Appendix 1: Survey Administered School Beverage Survey 1/2 1. Does your school offer grades other than 9th-12th? Yes No a. If yes, what grades? 2. What are your hours? a. School: From __________ to ____________ b. Lunch: From __________ to ____________ 3. Does your school participate in the following programs? National School Lunch Program National Breakfast Program Neither 4. What alternative food services does your campus offer within the time period of one half hour before and one half hour after school is in session? Snack clubs Vending machines Neither 5. What beverage choices do you offer as part of your lunch service? Please check all that apply: Diet Soda Flavored Water Non Diet Soda Milk Sweetened Sports Beverages Flavored Milk Energy Drinks 50-100% Juice Water (No Sweetener Added Other Juice 65 School Beverage Survey 2/2 6. What beverage choices do you offer as part of your alternative food choices? Please check all that apply: Diet Soda Flavored Water Non Diet Soda Milk Sweetened Sports Beverages Flavored Milk Energy Drinks 50-100% Juice Water (No Sweetener Added Other Juice 7. Do you have an open-campus policy? Yes No Don’t Know a. If yes, are the above beverages available off-campus during school hours (e.g. fast-food outlets, grocery stores, drugstores, corner stores) within two miles? Yes No Don’t Know b. If yes, where? Grocery Store Concession Truck Corner Store Fast-Food Restaurant Other, please specify __________________ 8. Are the majority of your students required to attend school during set hours throughout the week (or are a majority continuation or independent study students)? Yes No Don’t Know 9. How confident are you about the correctness of your answers? Extremely Confident Confident Not Confident Not sure at all 66 Appendix 2. Variable Descriptions Variable Weeklyupcqty: Description The weekly sum of the quantity of sales by the store and beverage category Weeklyupcnet: The weekly sum of the net sales by the store and beverage category. Week ID: Listed as Year, week number; (going from Wednesday ending Tuesday (yyyyww). Product_n Arbitrary value assigned by UPC- identifies an unique product. Store_n Arbitrary value assigned to each store. Sale: Binary variable; 1 for a UPC that corresponds to a product that is on sale, 0 if not on sale. Soda: Binary variable; 1 for a UPC that corresponds to a soda product, 0 for other beverage. Opencampus: Binary variable; 1 for an open campus within a half mile to a store, 0 for none. Within1: Binary variable; 1 for a store with a school within one mile, 0 if none. Withinhalf: Binary variable; 1 for a store with a school within half mile, 0 if none. Price: Price paid for the beverage product. Treat1: Records all beverage sales made during the first treatment (after July 1, 2007). Treat2: Records all beverage sales made during the second treatment (after July 1, 2009). Fall (Year): Records all beverage sales made during the fall semester of that year (September trough the end of November). Spring (Year): Records all beverage sales made during the spring semester of that year (February through the end of April). Opensoda: Indicates soda purchases purchased at a store within half mile to an opencampus (opencampus*soda). Opentreat2: Indicates purchases purchased at a store within half mile to an opencampus during the treatment 2 period (opencampus*treat2). Soda1mile: Interaction between soda dummy and schools within one mile (soda*within1mile). Sodahalfmile: Indicates soda purchases purchased at a store within one mile to a school (withinhalfmile*soda). 67 Variable Sodatreat2: Description Records all soda sales made during the second treatment (treat2*soda). Within1treat2: Records all sales made during the second treatment at a store within one mile of a school (treat2*within1mile). Withinhalftreat2: Records all sales made during the second treatment at a store within half mile of a school (treat2*withinhalfmile). Soda(s for spring, f for fall)(Year): Records all soda sales made during the semester indicated; semesters determined by: fall is September through the end of November, and spring is February through end of April (semester*soda). Within1(s for spring, f for fall)(Year): Records all sales made at a store with a school located within one ile during the semester indicated (within1*semester). Soda1mile(s for spring, f for fall)(Year): Difference- in-differences estimator for the change in soda sales compared to non-soda beverage sales from the given semester relative to before the treatment in stores with a school within one mile compared to those with no school within one mile. opensoda (semester): Indicates soda purchases purchased at a store within half mile to an opencampus during a particular semester(opencampus*soda*semester). DDD Treat2 (One Mile): Difference- in-differences estimator for the change in soda sales compared to non-soda beverage sales from the second treatment relative to before in stores with a school within one mile compared to those with no school within proximity. DDD Treat2 (Half Mile): Difference- in-differences estimator for the change in soda sales compared to non-soda beverage sales from the second treatment relative to before in stores with a school within half mile compared to those with no school within proximity. DDD Treat2- (Open): Difference- in-differences estimator for the change in soda sales compared to non-soda beverage sales from the second treatment relative to before in stores with an open campus school within half mile compared to those with no school within proximity. DDD Semester- (Half Mile): Difference- in-differences estimator for the change in soda sales compared to non-soda beverage sales from the semester relative to before in stores with a school within half mile compared to those with no school within proximity. DDD Semester(Open): Difference- in-differences estimator for the change in soda sales compared to non-soda beverage sales from the semester relative to before in stores with an open campus school within half mile compared to those with no school within proximity. 68 Variable hourlyqty Description The hourly sum of the quantity of sales by the store and beverage category perunitprice This variable is the price per unit of individual containers of sodas. For example, a twelve-pack would have 12 units, and the cost of the twelve-pack would be divided by price to obtain this variable. singleserve Binary variable; 1 for a UPC that corresponds to a product that is an individual package with less than 32oz, 0 if greater than 32oz or multiple units. afterschool Binary variable; 1 the time period after school (3 pm and 4:59 inclusive), 0 otherwise. beforeschool Binary variable; 1 the time period before school (6 am and 7:59 inclusive), 0 otherwise. lunchtime Binary variable; 1 the time period during designated lunch times (11 am and 12:59 inclusive), 0 otherwise. afteropen Indicates purchases purchased during the after school time period for stores within 5.14 miles of an opencampus (afterschool*opencampus). beforeopen Indicates purchases purchased during the before school time period for stores within 5.14 miles of an opencampus (beforeschool*opencampus). lunchopen Indicates purchases purchased during the designated lunchtime for stores within 5.14 miles of an opencampus (lunchtime*opencampus). beforesoda Indicates soda purchases purchased during the before school time period (beforeschool*soda). aftersoda Indicates soda purchases purchased during the after school time period (afterschool*soda). lunchsoda Indicates soda purchases purchased during the designated lunchtime (lunchtime*soda). beforetreat2 Indicates purchases during the before school time period as well as the treatment 2 time period (beforeschool*treat2). aftertreat2 Indicates purchases during the after school time period as well as the treatment 2 time period (afterschool*treat2). lunchtreat2 Indicates purchases during the designated lunchtime as well as the treatment 2 time period (lunchtime*treat2). beforetreat2soda Indicates soda purchases during the before school time period as well as the treatment 2 time period (beforeschool*treat2*soda). 69 Variable aftertreat2soda Description Indicates soda purchases during the after school time period as well as the treatment 2 time period (afterschool*treat2*soda). lunchtreat2soda Indicates soda purchases during the designated lunchtime as well as the treatment 2 time period (lunchtime*treat2*soda). beforeopensoda Indicates soda purchases during the before school time period at a store within 5.14 miles to an open campus (beforeschool*opencampus*soda). afteropensoda Indicates soda purchases during the after school time period at a store within 5.14 miles to an open campus (afterschool*opencampus*soda). lunchopensoda Indicates soda purchases during the designated lunchtime at a store within 5.14 miles to an open campus (lunchtime*opencampus*soda). D4 (After) Indicates soda purchases during the before school time period at a store within 5.14 miles to an open campus as well as during the treatment 2 time period (beforeschool*opencampus*soda*treat2). D4 (Before) Indicates soda purchases during the after school time period at a store within 5.14 miles to an open campus as well as during the treatment 2 time period (afterschool*opencampus*soda*treat2). D4 (Lunch) Indicates soda purchases during the designated lunchtime at a store within 5.14 miles to an open campus as well as during the treatment 2 time period (lunchtime*opencampus*soda*treat2). 70 REFERENCES Bernstein, Sharon. 2010. "San Francisco bans Happy Meals." Los Angeles Times, November 2. http://articles.latimes.com/2010/nov/02/business/la-fi-happy-meals20101103 Bertrand, Marianne, Esther Duflo, and Sendhil Mullainathan. 2004. “How much should we trust differencesin-differences estimates? Quarterly Journal of Economics 119: 249-275. Binkley, James K. 2006. “The Effect of Demographic, Economic, and Nutrition Factors on the Frequency of Food Away from Home.” Journal of Consumer Affairs 40: 372-391. Bonnet, Celine, and Vincent Requillart. 2011. "Does The EU Sugar Policy Reform Increase Added Sugar Consumption? An Empirical Evidence On The Soft Drink Market." Health Economics 20 (9): 1012-1024. Bowman, Shanthy, Steven L. Gortmaker, Cara B. Ebbeling, Mark A. Pereira, and David S. Ludwig. 2004. "Effects Of Fast-Food Consumption On Energy Intake And Diet Quality Among Children In A National Household Survey." Pediatrics 113 (1): 112-118. Briefel, Ronette, Ander Wilson, and Philip M. Gleason. 2009. "Consumption of LowNutrient, Energy-Dense Foods and Beverages at School, Home, and Other Locations among School Lunch Participants and Nonparticipants." Journal of the American Dietetic Association 109 (2): S79-S90. 71 Bruce, Stephanie. “SB965 Fact Sheet,” CALSNA, Accessed on November 22, 2011. http://www.calsna.org/documents/Legislation/SB965-FactSheet.pdf. Currie, Janet, Stefano Della Vigna, Enrico Moretti, and Vikram Pathania. 2010. "The Effect of Fast Food Restaurants on Obesity and Weight Gain." American Economic Journal: Economic Policy, American Economic Association 2 (3): 3263. National Bureau of Economic Research, Working Papers. Fernandes, Meenakshi M. 2008. “The effect of soft drink availability in elementary schools on consumption.” Journal of the American Diet Association 108 (9): 1445-52. Fletcher, J. M., D. Frisvold, and N. Tefft. 2010. “Can Soft Drink Taxes Reduce Population Weight?” Contemporary Economic Policy 28: 23-35. Fletcher, J. M., D. E. Frisvold and N. Tefft. 2010. “The Effects of Soft Drink Taxes on Child and Adolescent Consumption and Weight Outcomes.” Journal of Public Economics 94: 967-974. Fox, Mary Kay, Anne Gordon, Renee Nogales, and Ander Wilson. 2009. "Availability and Consumption of Competitive Foods in Us Public Schools." Journal of the American Dietetic Association 109 (S): S57-S66. Frieden, T., W. Dietz, and J. Collins. 2010. "Reducing Childhood Obesity Through Policy Change: Acting Now To Prevent Obesity." Health Affairs 29 (3): 357-363. 72 IOM Committee on Nutrition Standards for Foods in Schools. 2007. “Leading the Way Towards Healthier Youth.” Washington DC: National Academy Press; Accessed October 1, 2011. http://www.iom.edu/Reports/2007/Nutrition-Standards-forFoods-in-Schools-Leading-the-Way-toward-Healthier-Youth.aspx James, Janet, and D. Kerr. 2005. “Prevention of Childhood Obesity by Reducing Soft Drinks.” International Journal of Obesity 29: S54-S57. Kenyon, Alexis. 2011. "City Council moves forward with soda tax." Richmond Confidential. December 12. http://richmondconfidential.org/2011/12/12/citycouncil-moves-forward-with-soda-tax/. Kolb, Carol, and Carol Medlin. April 2004. "Soda Ban in Schools." Health Policy Monitor (April). <http://www.hpm.org/en/Surveys/IGH__USA/03/Soda_Ban_in_Schools.html>. Mancino, Lisa, Jessica Todd, and Biing-Hwan Lin. 2009. “Separating What We Eat from Where: Measuring the Effect of Food Away from Home on Diet Quality.” Food Policy 34: 557-562. Nielsen Samara Joy, Anna Maria Siega-Riz, and Barry M. Popkin. 2002. “Trends in food locations and sources among adolescents and young adults.” Preventative Medicine. 35: 107–113. Robert Wood Johnson Foundation, 2011. "Alliance for a Healthier Generation's Healthy Schools Program Now Reaches More Than 10,000 Schools." Childhood Obesity. Robert Wood Johnson Foundation, January 13. Accessed January 10, 2012. <http://www.rwjf.org/childhoodobesity/product.jsp?id=71750>. 73 Runge, Carlisle Ford, Justin Johnson, and Carlisle Piehl Runge. 2011. "Better Milk than Cola: Soft Drink Taxes and Substitution Effects." Choices: The Magazine of Food, Farm and Resource Issues. 26 (3). Accessed December 12, 2011. http://farmdoc.illinois.edu/policy/choices/20113/2011317/2011317.html. Samuels, Sarah, Sally Lawrence Bullock, Gail Woodward-Lopez, Sarah E. Clark, Janice Kao, Lisa Craypo, Jay Barry, and Patricia B. Crawford. 2009. “To What Extent Have High Schools in California Been Able to Implement State-Mandated Nutrition Standards.” Journal of Adolescent Health 45: S38-S44. Shin, Annys. 2007. “Removing Schools’ Soda is Sticky Point.” Washington Post. March 22. Accessed December 12, 2011. http://www.washingtonpost.com/wpdyn/content/article/2007/03/21/AR2007032101966.html. Story, Mary, Karen M. Kaphingst, and Simone French. 2006. "The Role of Schools in Obesity Prevention." Princeton University. 16 (1): 109-42. Taber, Daniel R., Jamie F. Chriqui, Lisa M. Powell, and Frank J. Chaloupka. 2012. “Banning All Sugar-Sweetened Beverages in Middle Schools: Reduction of InSchool Access and Purchasing but Not Overall Consumption” Archives of Pediatrics and Adolescent Medicine 166 (3): 256-262. DOI: 10.1001/archpediatrics.2011.200 Todd, Jessica E. and Chen Zhen. 2010. “Can Taxes on Calorically Sweetened Beverages Reduce Obesity?” Choices 25 (3)