Open Collections

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Health implications of food away from home consumption : a marketing and policy analysis Ghotbi, Sina 2015

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata

Download

Media
24-ubc_2016_february_ghotbi_sina.pdf [ 3.57MB ]
Metadata
JSON: 24-1.0221500.json
JSON-LD: 24-1.0221500-ld.json
RDF/XML (Pretty): 24-1.0221500-rdf.xml
RDF/JSON: 24-1.0221500-rdf.json
Turtle: 24-1.0221500-turtle.txt
N-Triples: 24-1.0221500-rdf-ntriples.txt
Original Record: 24-1.0221500-source.json
Full Text
24-1.0221500-fulltext.txt
Citation
24-1.0221500.ris

Full Text

  HEALTH IMPLICATIONS OF FOOD AWAY FROM HOME CONSUMPTION: A MARKETING AND POLICY ANALYSIS  by  Sina Ghotbi BS, University of Tehran, 2005 MS, University of Illinois at Chicago, 2008    A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY in THE FACULTY OF GRADUATE AND POSTDOCTORAL STUDIES (Business Administration)   THE UNIVERSITY OF BRITISH COLUMBIA (Vancouver) December 2015  © Sina Ghotbi, 2015  ii  Abstract  This thesis provides two independent empirical quantitative studies on consumer decision-making in the context of food-away-from-home consumption, and examines important public policy and marketing issues. We first study whether meals with diet carbonated soft drinks (CSD) have more calories than meals with regular CSD by investigating an individual-level meal consumption panel dataset between 2000 and 2007 at a major fast-food chain in Canada. This study is motivated by the debate on the effectiveness of diet CSD on weight loss and is the first to provide field-based evidence, supported by within subject variation, on the harm and benefit of diet drinks in terms of caloric consumption within a meal. We find no increase in total calories consumed during a meal, and most diet CSD drinkers also exhibit a reduction in total calories. Looking more narrowly at the food calories in a meal we find that young males are the only group who have more food calories when they choose a diet drink, yet the extra food calories do not exceed the calories saved by consuming diet rather than regular CSD. In the second study, using an individual-level panel dataset on food consumed in all restaurants in Canada, we examine the effect of media coverage of multiple instances of Bovine Spongiform Encephalopathy (BSE – mad cow disease) on the decision to dine out and, once in a restaurant, whether to order beef. We find that when media coverage of BSE crises is high, people eat out less often, decrease the number of beef orders when eating out, and increase the ordering of other products. Interestingly, we find that these effects only hold for well-known, national chain restaurants and not for local restaurants. We develop and implement a modified double-hurdle model that explicitly takes into account participation (i.e., dine out) and consumption (e.g., number of beef orders) information in estimation of the model. This work  iii  extends the product harm crises literature to the case of industry level crises and the context of restaurants. The thesis concludes with a brief chapter on limitations and extensions of the research.       iv  Preface  I was the primary author of the work presented in this PhD thesis. I was responsible for identifying the research questions, conducting the literature review, collecting and managing supplementary data, analyzing the data, modeling and coding the estimation procedures, and preparing the manuscript. Specific contributions for each chapter are described below.  1 Introduction  I was the author of this chapter with intellectual contributions from Charles Weinberg.   2 Essay 1: Do You Diet by Drinking Diet Drinks? An Empirical Study of Food and Drink Choice at Fast-food Restaurants  I was primarily responsible for identifying the research question, preparing the literature, collecting the nutrition data, implementing data clean up, carrying out the estimations, and preparing the manuscript. Charles Weinberg contributed by editing of the manuscript, and identifying and positioning of the research question. Tirtha Dhar provided the primary dataset, contributed to writing and estimations of an earlier version of the paper, and edited the current manuscript. Ting Zhu and Kevin Milligan provided intellectual contributions.  3 Essay 2: Restaurant Diners’ Reaction to Incidents of Mad Cow Disease: Stay Home, Eat Less Beef, or Life as Usual?  v   I identified the research question, reviewed the literature, collected the local media and weather data, cleaned and rearranged the data, built the econometric model, coded and carried out the estimations, and prepared the manuscript. Charles Weinberg assisted with the editing of the manuscript and positioning of the paper. Tirtha Dhar provided the primary dataset, offered guidance on the estimation procedure, and assisted with editing the manuscript. Ting Zhu contributed to the model development and editing of the manuscript. Kevin Milligan provided intellectual contributions  4 Limitations and Concluding Remarks  I was the author of this chapter with intellectual contributions from Charles Weinberg.      vi  Table of Contents  Abstract .......................................................................................................................................... ii	  Preface ........................................................................................................................................... iv	  Table of Contents ......................................................................................................................... vi	  List of Tables ................................................................................................................................ ix	  List of Figures ............................................................................................................................... xi	  Acknowledgements ..................................................................................................................... xii	  Dedication ................................................................................................................................... xiv	  Chapter 1: Introduction ............................................................................................................... 1	  Chapter 2: Do You Diet by Drinking Diet Drinks? An Empirical Study of Food and Drink Choice at Fast-food Restaurants .................................................................................................. 6	  2.1	   Introduction ........................................................................................................................ 6	  2.2	   Background ...................................................................................................................... 10	  2.3	   Data and Variable Description ......................................................................................... 15	  2.3.1	   Variable Description ................................................................................................. 17	  2.3.2	   Descriptive Statistics ................................................................................................. 20	  2.4	   Empirical Approach and Findings ................................................................................... 22	  2.4.1	   Total Calorie Analysis .............................................................................................. 24	  2.4.2	   Food Calorie Analysis ............................................................................................... 28	  2.4.3	   Identification Discussion .......................................................................................... 31	  2.5	   Discussion ........................................................................................................................ 34	  2.5.1	   Public Policy and Marketing Implications ................................................................ 35	   vii  2.5.2	   Limitations and Concluding Remarks ...................................................................... 38	  2.6	   Tables and Figures ........................................................................................................... 39	  Chapter 3: Restaurant Diners’ Reaction to Incidents of Mad Cow Disease: Stay Home, Eat Less Beef, or Life as Usual? ....................................................................................................... 53	  3.1	   Introduction ...................................................................................................................... 53	  3.2	   Data .................................................................................................................................. 58	  3.2.1	   Sample of Respondents ............................................................................................. 58	  3.2.2	   Measure of Media Coverage of the BSE Outbreaks ................................................. 62	  3.2.3	   Local Weather Information ....................................................................................... 63	  3.3	   Empirical Strategy ........................................................................................................... 64	  3.3.1	   Econometric Modeling: Double Hurdle with Explicit Participation Model ............. 66	  3.3.2	   Predictor Variables and Identification ...................................................................... 69	  3.3.3	   Model Comparison and Discussion .......................................................................... 71	  3.4	   Findings ............................................................................................................................ 75	  3.4.1	   Dine-Out and Meal-Order Analysis .......................................................................... 75	  3.4.2	   National Chain versus Local Restaurants ................................................................. 77	  3.4.3	   Magnitude of BSE Effect .......................................................................................... 79	  3.5	   Discussion ........................................................................................................................ 80	  3.6	   Tables and Figures ........................................................................................................... 85	  Chapter 4: Limitations and Concluding Remarks .................................................................. 94	  References .................................................................................................................................... 97	  Appendices ................................................................................................................................. 109	  Appendix A –  Chapter 2: Supplemental Analyses ................................................................ 109	   viii  Appendix B –  Chapter 3: Newspapers Used in Constructing the Media Index ..................... 117	  Appendix C –  Chapter 3: Constructing the Local Weather Variables ................................... 119	  Appendix D –  Chapter 3: Double-hurdle and DHEP Methods Derivations .......................... 123	  Appendix E –  Chapter 3: Investigating Supply Side Reaction to Food Scare Outbreaks ..... 131	  Appendix F –  Chapter 3: Counterfactual Analysis Formulations .......................................... 135	  Appendix G –  Chapter 3: Supplemental Analysis ................................................................. 139	    ix  List of Tables  Table  2-1. Descriptive Statistics of Demographic and Meal Occasion Variables ........................ 39	  Table  2-2. Total Calories Estimations – Main Effect ................................................................... 43	  Table  2-3. Total Calories Estimations – Matching Methods ........................................................ 43	  Table  2-4. Total Calories Estimations – Full Model ..................................................................... 44	  Table  2-5. Food Calories Estimations – Main Effect ................................................................... 48	  Table  2-6. Food Calories Estimations – Matching Methods ........................................................ 48	  Table  2-7. Food Calories Estimations – Full Model ..................................................................... 49	  Table  3-1. Average of Response Variables ................................................................................... 85	  Table  3-2. Summary Demographic Statistics for Sample and for Canadian Household Heads ... 85	  Table  3-3. Summary Statistics for BSE Media Index ................................................................... 86	  Table  3-4. Summary Statistics for Weather Variables .................................................................. 87	  Table  3-5. Models Comparison for Number of Beef Orders and Dine Out Probability ............... 88	  Table  3-6. DHEP Estimation for Beef- and Chicken-oriented Restaurants .................................. 89	  Table  3-7. DHEP Estimation for Meal Orders and Dine Out in All Restaurants ......................... 90	  Table  3-8. DHEP Estimation for Beef Orders and Dine Out in National Chain and Local Restaurants .................................................................................................................................... 91	  Table  3-9. Marginal Effect of Crises on Meal Orders and Dining Out in the Peak Months ........ 93	  Table  A-1. Total Calories Model Construction Steps (with Interaction) .................................... 110	  Table  A-2. Total Calories Marginal Effects for Model Construction Steps ............................... 111	  Table  A-3. Food Calories Model Construction Steps (with Interaction) .................................... 112	  Table  A-4. Food Calories Marginal Effects for Model Construction Steps ............................... 113	   x  Table  A-5. Percentage of CSD and Diet CSD in Restaurants .................................................... 114	  Table  A-6. Total Calories – Full Model Coefficient Estimations with Three-way Interactions 115	  Table  A-7. Food Calories – Full Model Coefficient Estimations with Three-way Interactions . 116	  Table  C-1. Province-City Size Matrix of Canadian Cities ......................................................... 120	  Table  E-1. Beef Price Analysis ................................................................................................... 132	  Table  E-2. Promotion Analysis – Coupon Models ..................................................................... 133	  Table  E-3. Promotion Analysis – Special Offer Models ............................................................ 134	  Table  G-1. Beef-Oriented Brand, Local, and Local Fast-food Restaurants ................................ 140	  Table  G-2. Weather Link to Food Orders ................................................................................... 141	  Table  G-3. Mad Cow Effect in Ontario and Alberta .................................................................. 142	  Table  G-4. Media Index Analysis ............................................................................................... 143	  Table  G-5. Avian Influenza Estimation for All Meal Items in All Restaurants ......................... 144	  Table  G-6. Chicken-Oriented Restaurants – BSE and Avian Influenza ..................................... 145	    xi  List of Figures  Figure 2-1. Diet and Regular Drink Orders by CSD Size ............................................................. 40	  Figure 2-2. Diet and Regular Drink Orders by Gender and Age .................................................. 40	  Figure 2-3. Average Total Calories by CSD Type and Size ......................................................... 41	  Figure 2-4. Average Food Calories by CSD Type, Size, and Gender .......................................... 41	  Figure 2-5. Average Food Calories by CSD Type, Gender, and Age Groups .............................. 42	  Figure 2-6. Total Calories – Estimated Effect by Gender ............................................................ 45	  Figure 2-7. Total Calories – Estimated Effect by CSD Size and Gender ..................................... 46	  Figure 2-8. Total Calories – Estimated Effect by CSD Sizes, Gender, and Age .......................... 47	  Figure 2-9. Food Calories – Estimated Effect by Gender ............................................................. 50	  Figure 2-10. Food Calories – Estimated Effect by CSD Size and Gender ................................... 51	  Figure 2-11. Food Calories – Estimated Effect by CSD Sizes, Gender, and Age ........................ 52	  Figure 3-1. Canada-Wide (normalized) BSE Media Index by Month .......................................... 87	  Figure 3-2. Actual and Counterfactual Beef Order and Dine Out Probability ............................. 92	  Figure 3-3. Marginal Effect of Crises on Meal Orders and Dining Out in the Peak Months ....... 93	   xii  Acknowledgements  I am not sure if I can fully express the gratitude I feel for the people who have helped me through this journey. I never believed in using the word ‘I’ even in single authored documents. If I can generate anything meaningful, it is only because “we” have brought it into existence.  I am truly grateful to my advisors Charles Weinberg and Tirtha Dhar, and committee members Ting Zhu and Kevin Milligan for everything they have done for me. Chuck had been a caring and guiding father along the way. I am honored to have been his student and am delighted that I had the chance to know his pure soul. Tirtha had been a kind older brother who was there for me in my difficult moments. Even though he was overcoming ongoing challenges is his own career he was never anything short of a great mentor and a supportive friend. Ting is one of the greatest minds I have ever met, and definitely much more than that. She is compassionate and humble, and she deeply cared about my growth. Kevin has a wonderful personality and was tremendously supportive along the way.  I would like to extend my gratitude to everyone in the Marketing division. Joey was the most supportive and affable person with whom I consulted about any problem. Dale, Darren, Tim, Chunhua, Ann, Kate, Jack, Yi, Dave, and all other faculty have been very generous in providing support and guidance whenever I needed them. I am also delighted that I had such an incredible team of PhD students, or better to say friends, with whom I shared this experience. And I want to thank Elaine Cho for her extraordinary responsiveness and approachability.    xiii  I thank Professor Ellen Goddard at the University of Alberta for providing us access to the data. The research presented in this dissertation was supported by Research Grants from the Social Sciences and Humanities Research Council to Charles Weinberg and Tirtha Dhar.   I would like to say a big thank you to Michael Scott, my MS supervisor at the University of Illinois at Chicago, who generously supported me in early days of graduate studies in a new continent, and kindly helped me to find my path into where I am today.   Finally, I am deeply grateful to my family, relatives, and friends without whom any achievement is meaningless. My parents, Soheila and Babak, are the ones who made this decade-long journey of graduate study possible, and my beloved aunts and uncles—Soraya, Eskandar, Sofia, Esfandiar, and Soosan—and Abbas and Leyla were my family when I was away from my family.  xiv  Dedication  This thesis is dedicated to my parents, Soheila and Babak, my sister, Dena, and my partner in the years to finish it, Ramineh. 1  Chapter 1: Introduction  Among the major changes to occur in the modern lifestyle is the shift toward the purchase and consumption of food away from home. Food away from home, defined as food prepared outside home as opposed to food prepared in a consumer’s kitchen, constitutes an increasingly larger portion of consumer expenditure on food and drink. In Canada, expenditure on food away from home (FAFH) constitutes more than 28% of total food expenditure (Statistics Canada – 2013), and this number is more than 50% in the USA (US Department of Agriculture’s Economic Research Services). This increase in FAFH consumption is paralleled by growing health issues, including high body mass index (BMI) and a high number of publicized food safety concerns. Concerns over such issues have led to responses by restaurants (e.g., introducing healthier food and drink items), public policy restrictions (e.g., mandatory posting of nutrition information and stricter food safety regulations) and a large body of research addressing the ongoing multifaceted debate on the matter. The current thesis is broadly motivated by such concerns on the healthfulness of food away from home consumption. It aims to provide a better understanding of consumer decision making in this domain, and to generate insight for marketers and policy makers.   This thesis adopts a quantitative approach to answer two sets of empirical questions in two independent essays. The primary data source is NPD Group’s Consumer Reporting of Eating Share Trends (CREST), a self-reported individual level consumer panel on meals when eating away from home. The data are from Canada and cover the years 2000 to 2007. The first essay studies the link between the choice of diet versus non-diet carbonated soft drink (CSD) and the  2  total calorie consumption in a meal. The core question it addresses is whether choosing diet CSD is an effective calorie cutting strategy in the context of a meal eaten away from home. The second essay is an empirical investigation of consumers’ reaction to major food scares, in particular BSE (or mad cow disease). The central research question is, do consumers change their decisions of whether or not to dine out and what to order during news of food scare outbreaks? If yes, to what extent and in which restaurants? Findings of both essays contribute to the relevant literatures, have implications for consumer health, and generate practical insight for managers and policymakers.   The first essay provides insight into the debate on the harm and benefit of diet CSD.  Some nutritionists and policy makers have claimed that choosing diet CSDs as compared to regular CSDs leads consumers to overcompensate for the saved calories in the drink by increasing consumption of other meal items. The empirical evidence on this notion, which is supported by behavioral theories such as multiple goal pursuit and calorie underestimation, is mixed. We merged the NPD Group’s CREST panel dataset with nutritional information from McDonald’s to estimate our models parameters. We exploit the within-subject variations in our estimations to control for individuals’ time-invariant unobservables, a key source of bias in other field studies. In addition, we discuss the assumptions under which we can identify a causal link between drink choice and consumed food calories.   Our analysis provides novel evidence in support of diet CSD as a calorie cutting option for consumers who are not willing to forego soft drinks. In brief, diners who choose diet CSD consume no more total calories in a meal than those who consume regular CSD. Additionally,  3  we find that although young males consume more calories from food accompanied with diet drinks than regular drinks, they do not increase consumption to the extent that it overcompensates for the calorie saved on the drink. Interestingly, middle-age females show the opposite effect—consuming fewer food calories when choosing a diet drink. This work suggests that promoting diet CSD, by providing more flavors of non-sugary drinks for example, can be beneficial to consumers in terms of calorie cutting and to restaurants in providing a health-oriented image. This study contributes to the marketing literature by testing the implications of behavioral theories such as multiple goal pursuit in the field setting, and identifying sources of heterogeneity in the effect of one decision on other decisions within the same consumption episode.  The second essay quantifies the reaction of consumers to multiple outbreaks of mad cow disease (BSE), a major case of industry level food scare. Food scares are typically referred to as food safety concern incidents that lead to spiraling public anxiety and escalating media attention. The first incident of Canadian BSE occurred in May 2003 and resulted in extensive media coverage; multiple subsequent cases were discovered in the following years. Some other examples of food scares affecting markets are Salmonella, E-coli, swine influenza, and bird influenza. We merge the NPD Group’s CREST panel dataset on all Canadian restaurants with the Canadian local press news index on BSE and local weather information to implement our analysis. Developing a two-stage decision-making model tailored to our problem, we estimate the probability of dining out, and condition on dining out, the number of different meal items to order. We model the interrelatedness of the dine out decision and the food order decision by the use of a flexible error structure for the two stages.   4   We find that the likelihood of dining out and the orders for beef decreased at the time of the BSE outbreak. We also find some evidence of substitution of pork and seafood for beef. Interestingly, the reported changes only happen for national chain restaurants, and we do not find any evidence of local restaurants being affected by the crises. According to our analysis, national chain restaurants have witnessed a 1.7% reduction in likelihood of being visited and a 3.4% reduction in beef orders at the peak month of the crisis. This study indicates that consumers tend to react rationally to changes of perceived risk, and policy makers should expect congruent reactions from consumers if a health threat is effectively communicated. Managers can use the magnitude of the reported effects to have a more accurate cost and benefit evaluation to design their prevention and reaction strategies when facing food scares. Our work contributes to the product harm crises literature by providing evidence on a crisis at an industry level, as opposed to at a brand level, and in restaurants (FAFH industry)—an area rarely explored in this literature.   The chapters proceed as follows. First, we discuss the “Do You Diet by Drinking Diet Drinks? An Empirical Study of Food and Drink Choice at Fast-food Restaurants” essay (Chapter 2) in six sections. We provide the research problem and overview of the project in Section 2.1, and discuss the literature and roots of the debate on diet drinks in the background section (Section 2.2). We then describe the dataset construction and the variables included in the analysis (Section 2.3). We provide the empirical strategy, robustness checks, and identification discussions in Section 2.4, and close the chapter with limitations, suggestions for future work, and managerial and policy implications (Section 2.5). Tables and figures are provided at the end of chapter 2 (Section 2.6), but the appendix (Appendix A) is located at the end of the thesis.  5  Second, we structure the “Restaurant Diners’ Reaction to Incidents of Mad Cow Disease: Stay Home, Eat Less Beef, or Life as Usual?” essay (Chapter 3) in six sections. Section 3.1 introduces the research questions and provides a brief background on food scares and the related literature. It also discusses an overview of the project and its contribution. Next we describe the dataset construction process and the descriptive statistics (Section 3.2). We detail the model development, provide a comparison to existing models, and discuss the estimation procedure in Section 3.3. Section 3.4 presents the model estimations and results for the entire FAFH market as well as for national chains and local restaurants. Discussions on managerial and policy implications are provided in Section 3.5. Tables and figures are provided at the end of chapter 3 (Section 3.6), but appendices (Appendix B to G) are located at the end of the thesis. Finally, Chapter 4 provides limitations of our studies, a brief overview of the work, and opportunities for future research.  6  Chapter 2: Do You Diet by Drinking Diet Drinks? An Empirical Study of Food and Drink Choice at Fast-food Restaurants  2.1 Introduction  Diet soft drinks1 were first introduced in 1952 by Kirsch Bottling in Brooklyn, New York, in the form of a sugar-free ginger ale called No-Cal. Later, in 1962, Royal Crown Cola introduced Diet Rite as the first mass-produced diet cola drink in the North American market. Subsequently, Coca-Cola Company introduced Tab, a precursor to their market-leading Diet Coke brand, and Pepsi Co. introduced Diet Pepsi. From the beginning, these diet beverages were marketed as products with the ability to deliver a particular health benefit based on almost zero caloric intake as compared to regular carbonated soft drinks (CSDs), without sacrificing taste and other desirable characteristics of such drinks. From this perspective, diet drinks can be defined as a functional food with the ability to deliver specific health benefits to consumers, allowing former Coca-Cola CEO Neville Isdell to argue, “Diet and light brands are actually health and wellness brands” (Martin, 2007).   However, the claimed benefits of diet drinks have been constantly challenged in the news headlines citing studies that link weight gain and other health problems to diet-drink                                                 1 We use drink, soft drink, and CSD interchangeably to refer to carbonated soft drinks. 2 For example: “Diet drink linked to weight gain,” Huffington Post, May 27, 2013; “Study: drinking diet soda actually causes weight gain, blood sugar spikes,” Natural News, July 24, 2011. 3 For example, Bleich, Wolfson, Vine, and Wang (2014) report that “overweight and obese adults drink more diet  7  consumption.2 One popular argument against diet drinks is that they harm consumers by encouraging a “Big Mac and Diet Coke” mentality. In other words, diet-drink consumption leads people to over-consume in other parts of the meal. In addition to this psychological argument, a body of nutrition literature has explored, with mixed results, the link between intake of artificial sweeteners (such as aspartame in diet CSDs) and satiation from a meal. Given the mixed findings, advocates on both sides of the debate over the consequences of diet-drink consumption have been able to support their position.   In the present study, by focusing on actual consumption data in a field setting—food and drinks ordered at a major fast-food restaurant chain—and employing heterogeneous treatment effect methodology, we provide new insight into the debate on diet drinks. Specifically, we address two research questions by studying the linkage between choice of diet drink and the caloric intake within a meal. First, when people order a diet drink as part of a meal, is the total calorie count from the meal higher than when they order a regular drink? Second, are the calories from food affected by the choice of diet drink compared with a regular drink?  We are the first study to measure the net caloric consumption difference between meals with diet versus regular soft drinks in a field setting where consumers know the type of drink they are choosing and make the choice voluntarily. In addition, the panel structure of the data provides within subject variations that helps us control for the time-invariant unobserved differences across individuals on caloric intake, and thereby overcome one of the key                                                 2 For example: “Diet drink linked to weight gain,” Huffington Post, May 27, 2013; “Study: drinking diet soda actually causes weight gain, blood sugar spikes,” Natural News, July 24, 2011.  8  shortcomings of existing correlational studies (Fowler et al., 2008) on the link between diet CSD choice and caloric intake.3 Finally, by examining consumption data as compared to purchase data, we can directly measure each person’s actual consumption of calories without the need to infer it from household-level purchase data, as was done in past studies (Binkley & Golub, 2007). On the other hand, we recognize that our data relate only to calorie consumption within a meal, and the study’s scope is limited to the short-term effects of a diet soft drink, as compared to a regular soft drink—and not other drink options such as water. Therefore, our study does not measure arguable long-term effects of diet-drink consumption on, for example, weight gain (Swithers & Davidson, 2008; perhaps through permanent metabolism change), cardiovascular problems (Gardener et al., 2012), and diabetic or blood sugar level issues (Suez et al., 2014). Still, studying short-term effects is of particular value since behavioral theories suggest that the proximity of decisions is an important factor in explaining the possible impact of making a healthy (or, at least, calorie-cutting) choice on the healthfulness of related decisions.   Understanding the impact of diet-drink choice on consumers’ caloric intake is important for the following reasons. First, this knowledge tests the validity of the industry’s focus on highlighting the advantages of diet drinks.4 The results will help develop a more informed public-policy debate on the regulation of diet and regular drinks and other functional foods in general. This is particularly relevant today because regulators have imposed sugar taxes (e.g., in                                                 3 For example, Bleich, Wolfson, Vine, and Wang (2014) report that “overweight and obese adults drink more diet beverages than healthy-weight adults and consume significantly more solid-food calories and comparable total calories than overweight and obese adults who drink sugar-sweetened beverages.” 4 Major beverage industry companies, through affiliations such as the Healthy Weight Commitment Foundation, have voluntarily pledged to reduce calories through a combination of product innovation, smaller portion sizes, and marketing of low-calorie options (Slining, Ng, & Popkin, 2013).   9  France and Mexico) that raise the price of regular but not diet drinks, and they have sought (e.g., in New York City) to limit the maximum size of regular but not diet drinks sold in restaurants (Leon, 2011; Comlay, 2013). Second, the marketing literature has developed a rich body of theories and a set of supporting laboratory studies for understanding how the choice of a presumably healthy (virtue) product like diet drinks can lead to overindulgence or increased consumption of other items in a meal (Chernev, 2011; Chernev & Gal, 2010; Dhar & Simonson, 1999). However, implications of these theories are largely untested in a field setting. Third, if, in line with prediction of such behavioral theories, consumption of diet drinks is associated with higher caloric intake in the rest of the meal, then it is important to learn the heterogeneity of this effect across the population. This can help policy makers better target or assist more vulnerable segments of the population.  We find, in response to the first question above, that total calories in a meal (food plus drink) is significantly lower when consuming diet as compared to regular drinks. However, in response to the second question, we find that some segments of the population—primarily younger males—order more food calories when choosing a diet as compared to a regular drink. These findings indicate that, despite some significant caloric consumption variation across foods with diet and regular drinks, diet soft drinks have calorie-cutting benefits.  In the following sections, we discuss the background, dataset, and the variables used in analysis. We then move on to the empirical approach and findings to discuss analysis, results (for both total calorie and food calorie models), and identification. We end with discussions on the  10  implications and limitations of our work. The tables and figures are presented at the end of the chapter.  2.2 Background  At a broad societal level, researchers and social critics have pointed out that despite the widespread usage of “diet” beverages and other functional food in North America, the obesity rate is alarmingly high (Government of Canada, 2009).5 The correlation between diet-drink consumption and weight gain is highlighted in the news based on long-term correlational field studies such as Fowler et al. (2008) and Stellman and Garfinkel (1986). Fowler et al. (2008), for example, followed more than 5,000 adults for seven years and reported that although people who drank both sugar-sweetened (regular) and diet sodas gained weight, diet soda drinkers were more likely to become obese. Also, the more diet sodas the participants drank, the greater their weight gain. However, these findings sometimes are contradicted by other long-term studies in which researchers find evidence to link diet-drink consumption to weight loss (e.g., Schulze, 2004). Such mixed findings and the prevalence of CSD consumption have created a debate on harm and benefit of soft drinks and especially their influence on weight gain. Numerous academic studies have tried to understand this behavior, policy makers have proposed practices such as a sugar tax, and industry players have voluntarily extended the lines of diet products and devised smaller packaging to address concerns about consumption of CSD.                                                  5 For further details on obesity crisis please refer to the Centers for Disease Control site at http://healthypeople.gov/2020/LHI/nutrition.aspx?tab=data.   11    The puzzle of gaining weight by consuming a lower-calorie product, diet drinks, is typically explained by the extra calories contributed from other food items—the behavior we refer to as “balancing.” Based on the existing literature on this topic, we identify two general paths for the effect of diet drink on food consumption: physiological and psychological. In the physiological path, consumption of artificial sweeteners (used in all diet CSD) may influence food consumption through short-term satiation level or may transform metabolic cycles, resulting in higher food intake in the long run. The second path attributes the attendant weight gain to psychological factors involved in choosing a diet (or healthier, “virtue”) option and its influence on adjacent decisions. This path is supported by implications of behavioral theories such as multiple goal pursuit and calorie underestimation, and a number of lab experiments on the effect of low-fat or low-cal labels on consumption behavior.  The evidence on physiological path, mostly provided by the nutrition literature, is mixed and overall suggests that the physiological short-term effect of diet drinks does not lead to excessive calorie consumption (see meta-analysis by Miller & Perez, 2014). Nutritional scientists specifically tested the consumption impact of diet drinks (i.e., artificially sweetened drinks) on satiation levels (Rogers, Carlyle, Hill, & Blundell, 1988) and food intake, primarily conducting lab studies in which the subjects were unaware of (blind to) whether or not the products they consumed were artificially sweetened. Blind experiment design was intended to rule out psychological factors. While some studies reported no link between extra calorie consumption and artificially sweetened drinks (DellaValle, Roe, & Rolls, 2005; Parham & Parham, 1980; Rolls, Kim, & Fedoroff, 1990; Schulze, 2004), others found that some extra food calories are consumed with an artificially sweetened drink but normally not to the extent that it surpasses the  12  calories saved by the diet drink (Holt, Sandona, & Brand-Miller, 2000; Lavin, French, & Read, 1997; Rogers et al., 1988). The evidence for long-term effects of artificial sweeteners on metabolic reflex and calorie-regulating mechanism is scant and inconclusive. While there are no detailed studies of such an effect in humans, Swithers and Davidson (2008) found that artificial sweeteners interfere with the body’s natural ability to judge calorie content and therefore resulted in higher caloric intake among rats with an artificially sweetened diet.  On the psychological path, primarily investigated in the area of consumer behavior, researchers aim to better understand whether and why awareness (via labeling) of choosing a healthy (virtue) product may affect consumption in the short term. Implications of the two mentioned psychological theories and related laboratory evidence can explain the relationship between decisions (e.g., food order and drink order) made within a consumption episode. First, the theory of goal pursuit addresses why people may make decisions congruent with a single goal, or may try to serve multiple and even opposing goals. Fishbach and Dhar (2005) argued that people make decisions (e.g., eating healthy and indulgence) either by focusing on the pursuit of a single goal or by “balancing” potentially contradictory goals (Dhar & Simonson, 1999). When pursuing a single goal, consumers choose actions consistent with that goal; however, when individuals pursue multiple goals, they may choose different actions to balance among conflicting goals. More specifically, when a single goal (e.g., losing weight) is pursued, consumers should consistently choose lower-calorie options. Alternatively, in choice episodes involving a trade-off between the two goals of vice (e.g., pleasure) and virtue (e.g., losing weight), consumers tend to balance between the two goals. In our context, this implies that if the  13  choice of diet drink fulfills the losing-weight goal, then consumers will allow themselves to pursue an indulgence goal with their subsequent food choice.  Second, Chernev and Gal (2010) found that consumers tend to systematically underestimate the combined calorie content when evaluating bundles of healthy (virtue) and indulgent (vice) options in a meal, such that they end up averaging rather than adding the calories contained in the vice and virtue.6 In the present context, such averaging also implies consumption of more high-calorie food conditioned on the choice of diet drink. More broadly, behavioral theories demonstrate that a current choice (Fishbach & Dhar, 2005), past choice (Mukhopadhyay, Sengupta, & Ramanathan, 2008), or the opportunity to choose a healthy as compared to hedonic food (Wilcox, Vallen, Block, & Fitzsimons, 2009) may fulfill healthy eating goals and trigger a rebound effect, leading to a less healthy choice now or in the near future (Bublitz, Peracchio, & Block, 2010).  In addition to the discussed theories, some lab studies found evidence for the effect of a low-fat or low-cal label on eating level. In one notable study where subjects were served yogurt labeled low-fat or high-fat, Shide and Rolls (1995) showed that participants exposed to different fat-content labels consumed significantly different quantities of calories in a meal. Controlling for the yogurt’s calorie level, subjects in the low-fat-labeled yogurt consumed more total calories                                                 6 Because people rely on their evaluations of a meal's healthiness to determine its calorie content, they consequently conclude that a meal combining a healthy and an unhealthy item has fewer calories than the unhealthy item alone. Such a habit, according to the author, is due to the qualitative nature of people’s information processing, which stems from their tendency to categorize food items according to a good/bad dichotomy into virtues and vices. In other words, when presented with a meal that combines both a virtue and vice, people form an overall impression of this meal's healthiness in a way that the vice/virtue combination is perceived to be healthier than the vice alone.  14  as compared to subjects having a meal with high-fat-labeled yogurt. By contrast, when the fat-content labels were removed, the finding reversed and subjects with low-fat yogurt consumed fewer calories than did subjects with high-fat yogurt. Moreover, in the studies of Wansink and Chandon (2006), subjects were given an identical snack, and in one condition the snack was labeled as low-fat versus no label in the control condition. In three studies these authors showed that low-fat labels led all consumers—particularly those who are overweight—to overeat snack foods. These experiments indicate that low-fat and low-cal labels can nudge participants to consume more. Therefore, from a policy perspective, we cannot solely rely on blind studies to infer the effect of diet drinks on caloric intake. However, these behavioral experiments do not clarify the net effect of these labels on calorie consumption, a measure that is necessary to determine the effectiveness of the labels. For instance, we need to know if the calories consumed from the extra amount of “low-fat” snack is more than the calories saved by switching from the regular snack to the low-fat version.7  Based on the discussed body of theories and evidence, we clarify the position of our study as informing the debate on diet drinks’ link to calorie consumption. The current work is the first to measure the short-term effect of drink choice on caloric consumption, where consumers are aware of the drink type. Its field-level rich panel data ensure the external validity of findings and allow for controlling for observables and some key unobservables (individual fixed effects).8                                                  7 We clarify this point with a hypothetical numeric example. Suppose that a normal bar of chocolate has 100 calories and the low-fat version of the same product has 40 calories. In this case even if consumers switch from consuming a single bar of regular chocolate to two bars of the low-fat chocolate, they still save 20 calories.   8 A conceptually related study measured the impact of purchase of diet drinks on other food purchases at supermarkets by consumers. In  particular, Binkley and Golub (2007) used consumer shopping data at grocery stores and found that in a subsample of people who live alone, panel members who buy diet drinks also purchase a higher  15  In addition, we comment on the heterogeneity of response across multiple covariates (i.e., gender, age, income, and party size) to provide a deeper understanding of consumers’ decision-making process.   2.3 Data and Variable Description  We use data from the NPD Group’s CREST (Consumer Reports on Eating Share Trends) panel for the years 2000–2007 to estimate our model. NPD collects the data using diary-based surveys from rolling panels of representative samples of Canadian consumers. The unit of observation in the present study is the individual’s away-from-home meal consumption occasion (i.e., breakfast, lunch, dinner, or snacks). For each consumption occasion by a panel member, the database provides information on restaurant visited and food and drinks ordered. It also provides detailed demographic information on each head of household.9 The database in hand does not provide calorie count or nutritional information for meal items ordered. Thus, to supplement this database we merged it with data collected on calorie and other nutritional information from the McDonald’s Nutrition Facts and United States Department of Agriculture databases.10                                                                                                                                                         share of lower-calorie and healthier products (fruits, vegetables, low-fat diary). The study is limited to examining the share of different grocery items purchased and does not provide any link between the shopping behavior and consumption. 9 In this database, each household has at least one head and typically two heads, one female and one male. 10 Specifically, McDonald’s Canada Nutrition Fact data can be accessed here: http://www1.mcdonalds.ca/NutritionCalculator/NutritionFactsEN.pdf and USDA National Nutrient Database for Standard Reference can be accessed at: http://www.ars.usda.gov/Services/docs.htm?docid=8964.   16   For our analysis we use data on consumption occasions at McDonald’s for the following reasons. First, the standardized product offerings at McDonald’s enable us to measure calorie content of meals with a high degree of accuracy relative to any other restaurants or any combination of restaurants. Second, comparing the patterns of CSD order in McDonald’s to other fast-food restaurants (Subway, Wendy’s, Burger King, A&W, Arby’s, KFC, and Harvey’s), we find that the McDonald’s sample descriptive statistics are close to the average. That is, the average percentage of meals with CSD and the average percentage of meals with a diet drink conditioned on any CSD order across all fast-food chains is 35% and 23%, respectively; the corresponding percentages for McDonald’s are 37% and 25% (details in Appendix A). These descriptives suggest that behavior with regard to drink choice in McDonald’s is representative of that in other fast-food places. Third, as one of the largest fast-food chains in Canada, McDonald’s offers a sizable consumer base that is a demographically representative sample. Finally, by studying the impact of diet-drink choice on caloric intake in McDonald’s, we shed light on consumption behavior at a typical quick-service restaurant (QSR), one of the fastest-growing segments of food-away-from-home (FAFH) consumption.11  For the purpose of our study, we filter the original dataset in multiple steps. Our NPD Group’s CREST database contains 682,830 observations of individual meal occasions, 59,730 of which are meals at McDonald’s. Within McDonald’s data, only observations with a CSD and one other item are relevant for this study. Therefore, we retained 22,075 observations of meals with CSD and at least one other meal item. Of these, we excluded 2,178 cases for which we                                                 11 In 2013, quick-service restaurants (QSR) constituted 64% of total FAFH incidence (NPD Group, 2013).  17  could not ascertain the calories from food because of the discrepancies between the menu items and CREST food item coding. In the database only household heads can be identified across multiple eating occasions, so we drop the observations for other individuals (e.g., guests or children); this step of the data-selection process drops 8,044 observations. In our exploratory data analysis we found very few incidents of ordering a diet drink over the age of 80, so we focus on individuals between the ages of 19 (the minimum age for household heads observed in the sample) and 80. This selection criteria leaves us with 11,777 observations. We also implemented three systematic outlier detection analyses and excluded 39 observations that were flagged as outlier by all methods.12 The final sample of analysis has 11,738 observations.  2.3.1 Variable Description   Next we discuss the variables in our analysis. The two dependent variables we use in analysis are the calories from food and drink and calories from food alone within a meal. To estimate the effect of drink choice on caloric intake, we conceptualize a quasi-experimental design whereby choice of caloric intake can be explained by the treatment conditions of drink type (i.e., diet and regular) while controlling for drink size (small, medium, and large) and other mean-centered covariates including gender (male dummy), age (both linear and quadratic terms to capture any nonlinear pattern),13 income (ordinal variable with ten levels: under $15,000; $15,000–19,999; $20,000–24,999; $25,000–29,999; $30,000–34,999; $35,000–44,999; $45,000–                                                12 The three outlier detection methodologies we used are studentized residual, Hosmer and Lemeshow leverage (on coefficients) measure (Hosmer & Lemeshow, 2004), and Cook’s distance measure (Cook, 1977).   13 Our approach is consistent with findings in the literature that report a non-linear relationship between age and caloric intake (Chesher, 1997).  18  54,999; $55,000–59,999; $60,000–69,999; over $70,000), party size (the number of family members and guests present at the eating occasion), meal occasions (breakfast, lunch, and dinner dummies with snacks as the base), and their interaction with drink choice variable. Next we justify inclusion of these variables in the calorie estimation model.  Drink size is a key variable to fully characterize the choice of CSD. This variable clearly affects the total calorie consumption, and the asymmetric caloric content of drink size across diet and regular soft drinks requires us to include the interaction of drink size and drink type (diet vs. regular) in the model. Consumer decision on the size of a diet drink is potentially driven by factors other than those used in decisions about the size of a regular drink. Given the zero-calorie level of diet drinks regardless of their size, the choice of size for diet drinks is primarily driven by price and thirst. In choosing the size of a regular drink, however, one may expect the calorie concerns to be an additional driver. As shown in Figure 2-1, the distribution of drink size across diet and regular is almost identical in our sample (the null hypothesis of independence of size and type of CSD is not rejected, p > .1), suggesting that size decision is probably driven by similar underlying forces for both types of drinks. This indicates that size decision is not the primary calorie-cutting strategy. In addition, the experiments of Flood, Roe, and Rolls (2006) and Rolls et al. (1990) did not show any change in food intake due to change in the size of drink. Therefore, while controlling for drink size we keep the focus of our study on switching patterns between diet and regular drink.  Gender is a common demographic classification for marketers and policy makers (e.g. Diet Coke is targeted to females), and numerous studies indicated that females were better in  19  terms of their consumption within dietary constraints and attention to health (Wardle et al., 2004), using diet/low-fat labels in their decision making (Braun, Gaeth, & Levin, 1997), intention to buy functional food (Bhaskaran & Hardley, 2002; Verbeke, 2006), and following recommended dietary guidelines (Kiefer, Rathmanner, & Kunze, 2005; Roos, Lahelma, Virtanen, Prättälä, & Pietinen, 1998; Turrell, 1997). Therefore, we expect to find a different behavior across genders with males having fewer diet drinks and making less healthy eating decisions.  Age is also a key targeting dimension for both policy makers and marketers, and therefore the heterogeneity of response on this dimension has substantial and practical value. In addition, Kim, Nayga, and Capps (2000, 2001) found that older people follow healthier diets—lower intake of calories from total fat and cholesterol and higher score on the Healthy Eating Index. Also, studies on intentions to purchase functional foods showed that older consumers were more willing to purchase functional food (Bhaskaran & Hardley, 2002; Verbeke, 2006). Based on this body of studies, we expect older people to be more likely to order a diet drink, and also to be more careful in their food choices.  Numerous studies, especially in the obesity literature (e.g., Garn et al., 1977), documented the link between income (sociodemographic class) and calorie consumption or body size. In addition, consumers with higher education (and presumably with higher income level) had lower (higher) intake of fat and sodium (fiber) (Kim at al., 2000). Kim et al. (2001) found that income is positively linked to the Healthy Eating Index, reflecting higher-quality diets and stronger health concerns. Therefore, we might expect the more affluent individuals to make  20  healthier decisions and be less likely to overcompensate for the calories saved when choosing a diet drink. We kept only income in the analysis since income and education are highly correlated in our sample.  Party size is reported to have a direct positive effect on calorie consumption, especially when eating at restaurants (Ariely & Levav, 2000; de Castro & de Castro, 1989). Also, it may differentially influence the decisions across meals with diet and regular drinks, for instance, through an impression-management mechanism. Previous research showed that impression-management motives prompt consumers to strategically alter their behaviors to present themselves positively (Ashworth, Darke, & Schaller, 2005; White & Dahl, 2007). For example, when eating in a group one might switch to diet soft drink and replace the fries with salad to provide a healthy image to companions.  Finally, the meal occasion variables mostly control for the systematic differences across different eating occasions. For instance, breakfast and snacks—meals in between the three main meal occasions—typically have lower calorie levels. We include the interaction of these variables to capture any moderating effect for the meal occasion.   2.3.2 Descriptive Statistics  Table 2-1 displays the descriptive statistics of the discussed variables. For the purpose of comparison, we provide descriptive statistics for all eating occasions (Full Sample), all occasions at McDonald’s (All McDs), and for the sample we use in our analysis (Analysis Sample). On  21  most characteristics, the three samples are largely similar. The percentage of breakfast observations is lower in the analysis sample because we retained the meals with CSD, a less popular drink for breakfast.  We present the distribution of meals with diet versus regular soft drink in our analysis sample. Figure 2-1exhibits the percentage of diet-drink orders across small, medium, and large drink sizes. Consistent with our earlier discussions, the share of diet drink is virtually the same across drink sizes. In Figure 2-2, we present the percentage of diet-drink choice across gender and age groups, illustrating that, as expected, younger and male consumers have proportionally less diet-drink orders.  Next, we generate a set of figures to provide further insights into the relationships between our dependent variables and key covariates—drink size, gender, and age. Figure 2-3, the only figure exhibiting the total calories (not food calories), demonstrates that meals with a diet drink have lower calories—199, 261, and 291 for small, medium, and large CSD, respectively. A large portion of the difference is driven by the fixed calorie differences across diet and regular drinks (150, 220, and 320 calories for a small, medium, and large drink, respectively), so we focus on food calories in the following figures. In Figure 2-4, we present the average caloric intake from food by drink type, size, and gender. On average and across all drink sizes, individuals consume fewer food calories when drinking a diet CSD or small CSD. Males have higher food caloric intake. To explore further, we then break down average caloric intake from food by drink type, size, gender, and age in Figure 2-5. To simplify our exposition, we graph averages by six age groups (ages 19–30, 31–40, 41–50, 51–60, 61–70, and 71–80). In the case of  22  males and females in the 19–30 age group and males in the 31–40 age group we find greater consumption of food calories with diet drinks—81, 4, and 15 calories, respectively. For the rest we find lower consumption of food calories with a diet drink. These variations, in addition to theoretical and substantial reasons explained above, suggest that both age and gender can play a critical role in explaining food consumption behavior by drink choice.   2.4 Empirical Approach and Findings  The first goal of this empirical work is to establish the effect of a diet soft drink order on total caloric consumption within the same meal. We are interested in measuring the relative forces of two opposing theories in terms of calorie consumption: balancing between the choice of diet drink and indulgence in food and direct effect of cutting calories by switching to diet drink. The second objective is to establish the effect of choice of diet drink on the food accompanied by the drink. In addition to multiple robustness checks we address the identification assumptions for the effect on food calories, and we explore the heterogeneity in response and the mechanism by use of interaction terms.  In line with the treatment effect framework, our main effect model (equation 1 below) is a linear regression of the dependent variable (total calories or food calories) on the diet drink dummy (  if individual i at occasion t orders a diet drink and  otherwise). We bring in drink size dummies (similarly defined), demographics (male dummy, age, age squared, and income), and eating occasion variables (party size and meal occasion dummies) to observe dietit = 1 dietit = 0 23  the influence of controlling for those factors on the coefficient (coefficient subscripts are matching in equation 2-1 and equation 2-2).          2-1  Our full model is, in line with the heterogeneous treatment effect framework, the regression of the dependent variable (total calories or food calories) on the diet drink dummy, the discussed variables, the full set of two-way interactions of diet by drink size and diet by covariates, plus the full set of three-way interactions of diet by drink size by covariates (equation 2-2). We include the full set of interactions to avoid arbitrary selection of interaction terms, but for both the main effect model and the full model we provide estimates of models with a smaller set of variables to demonstrate the robustness of the findings.    1 2 3 45 620 1 2 3where1, , , , , , , ,, , , ,it it it it it it it itit it it it it it itit i it it it it it it itn n n n nDV X diet X small X large Xdiet small X diet large XX male age age income partysize breakfast lunch dinnerεβ β β β β=Β + Β × +Β × +Β ×+ Β × × + Β × × +⎡ ⎤= ⎣ ⎦Β = [ ]4 5 6 7 8, , , , for  1, ,6n n n n n nβ β β β = K 2-2  We primarily use Fixed Effect, i.e., within (subject), estimator to estimate the model’s coefficients. In addition, we implement OLS, Random Effect, and matching estimators to examine the robustness of the analysis. Fixed Effect is our preferred estimation procedure because it controls for individual time-invariant unobservables. The Fixed Effect estimation procedure is based on the standard assumption that the error term can be decomposed into a dietit10 20 30 40211 12 13 1415 16 17 18it it it iti it it itit it it it itDV diet small largemale age age incomepartysize breakfast lunch dinnerβ β β ββ β β ββ β β β ε= + + ++ + + ++ + + + + 24  linear combination of individual fixed effect, , and the remaining unobservables:  (e.g., Wooldridge, 2010). However, we also need to consider the additional assumptions required to address the endogeneity due to correlation of and the variable. In our observational setting, the assignment of treatment (choice of diet drink) is not random, and therefore we need to make additional assumptions to support the causality for some of the reported findings. In the final part of this section we use the framework of direction of bias due to omitted factors to clarify the conditions under which part of our findings has a causal implication.  2.4.1 Total Calorie Analysis  In this subsection we pursue the first objective of the analysis—investigating the link between choosing a diet drink and total calorie consumption in a meal. We present estimation of the main effect model with total calories as the dependent variable in Table 2-2. We report that based on both between- and within-subject variations, using OLS (column 6), meals with a diet drink have 241 calories fewer than ones with a regular drink. When we control for individual fixed effect, in other words, confining our analysis to within-subject variations (column 3), we still observe 227 fewer calories for meals with a diet drink. This indicates that diet drinks have strong overall calorie-cutting benefits.  We take multiple steps to demonstrate the robustness of the reported main effect of diet drink on total calorie consumption. First, in columns 1, 2, 4, and 5 of Table 2-2, we present the Fixed Effect and OLS estimations of model with a smaller set of covariates. We also provide the Random Effect estimation of the model (Table 2-2, column 7) and implement three matching iθ it i iteε θ= +ite itdiet 25  methods—propensity score, nearest neighbor, and regression adjustment—to estimate average treatment effect of diet drink on total calories (Table 2-3). We observe that the negative (calorie-cutting) effect of diet drinks persists across these specifications and estimation procedures.    In the next step, using the framework of Altonji, Elder, and Taber (2005), we provide evidence on the bounds of the unobservables’ effect on our findings. Under assumptions discussed by Altonji et al. (2005), the robust estimated effects of diet drinks across presented specifications provide a relative measure for the potential bias from unobservables. In this exercise, the estimation of model with complete set of covariates (restricted model) is compared to the estimation of model with only the independent variable of interest (unrestricted model). We examine how much inclusion of the covariates influences the coefficient of the independent variable of interest. Based on this change in the estimation of this coefficient, we can comment on the proportional explanatory power of unobservables required to eliminate all the effect for the key independent variable. Explicitly, let 620R columndietβ β=  be the coefficient estimate of the diet variable in a restrictive model with controls (Table 2-2, column 6, row 1), and 420UR columndietβ β=  be the diet coefficient estimate in an unrestrictive model (Table 2-2, column 4, row 1). Altonji et al. (2005) procedure indicates that the ratio is a measure of how much larger the selection on unobservables needs to be relative to the selection on observables in order to fully account for the estimated effect of diet drink on the total calorie count. For the above ratio we plug in the least restrictive estimate of the diet coefficient from a regression with no controls (Table 2-2, column 4) and its most restrictive estimate from the regression with the highest explanatory power (Table 2-2, column 6). We find the ratio to be equal to five. Therefore, the bias due to !dietR!dietUR " !dietR 26  unobservables would have to be more than five times the bias due to omitting observables to account for the entire effect of the diet drink (this ratio would have been higher if we used Fixed Effect coefficients). This does not indicate that the selection on unobservables may not introduce a serious bias, but that it is unlikely to explain all of the estimated effect. In the Identification subsection, we further expand the discussion on unobservables in a different framework to clarify the assumptions required for making a causal claim.  So far we have established a strong link between diet drink and total calorie reduction. Next, we move on to the full model estimations (Table 2-4) to investigate the heterogeneity in response. As discussed earlier, among the three estimations for the full model presented in Table 2-4, we focus on Fixed Effect estimations to control for individual time-invariant factor (the Hausman test rejects Random Effect in favor of Fixed Effect, p > .05).14   We explore the heterogeneity in effect of diet drinks by calculating the marginal effect of diet drink selection across drink size, gender, age, and other covariates. For example, the difference in caloric intake between two drink type conditions (diet subtracted by regular) for a medium-size CSD (base case) can be expressed as (other variables are fixed at 0) the coefficient of the diet dummy variable:   2-3                                                  14 In Appendix A, Table A-1 and Table A-2, we provide estimation and marginal effects for the simplified interaction model to exhibit the model-building process and show the robustness of the effects with smaller sets of covariates and interactions. Table A-1 and Table A-2 also show that the effects are qualitatively similar across smaller models estimated by OLS and Fixed Effect. E cal diet = 1!" #$ % E cal diet = 0!" #$ = &20 27  Similarly, the difference in caloric intake for a male ordering a small-size drink can be expressed as a linear combination of coefficients of diet, diet interacted with mean-centered male, diet interacted with small (drink size), and diet interacted with small interacted with mean-centered male dummy variables’ (presented in order in equation 2-4):  20 22 50 521, 1, 10, 1, 1 (1 ) (1 )where  is the mean of male dummymale malemaleE cal diet small maleE cal diet small male β µ β β µ βµ⎡ ⎤= = = −⎣ ⎦⎡ ⎤= = = = + − + + −⎣ ⎦     2-4  As exhibited in Figure 2-6, we report -210 calories for overall marginal effect of diet drink on total calories with smaller (larger) in magnitude effect for males (females). To further illustrate the drink choice, we calculate the marginal effect of diet by gender for small, medium, and large drinks in Figure 2-7. We see that the marginal effect of diet drink becomes insignificant only for males in the small drink condition. To probe this further, we bring in the age variable and generate Figure 2-8. Here we observe that males under 40 and above 63 years old and in the small drink condition are the largest group for whom we cannot report any calorie reduction. Moreover, income and party size interactions turn out to be insignificant, so we do not observe heterogeneity in response on these dimensions. Overall we observe that no demographic group consumes extra total calories when choosing a diet drink (no significant positive effect on any of the graphs), and most people have significantly fewer total calories in meals with a diet drink.     28  2.4.2 Food Calorie Analysis  The second objective of this analysis is to investigate the effect of choice of diet drink on food consumed with that drink. The total calorie count, which includes both food and drink calories, directly addresses the debate on harm and benefit of diet drinks. However, to provide a deeper understanding of the consumer decision-making process, we isolate the food decision and shift our dependent variable to the food calories eaten with the soft drink.   Similar to our explication in the previous subsection, we start by the main effect model presented in Table 2-5, columns 3 and 6. This time we do not detect any main effect on food calories by choosing a diet drink. Even in the OLS estimations, the main effect present in the smaller model (columns 4 and 5) vanishes by including all the covariates. Similarly, we do not detect any significant average treatment effect based on the matching estimations in Table 2-6. This means that we do not have any overall main effect for diet drink on food; or in other words we do not detect any average treatment effect for diet drink across the whole population.    Next, we discuss the full model presented in Table 2-7. OLS estimation is driven by both between- and within-subject variations, and between-subjects variation can be driven by the inherent difference between individuals (e.g., BMI, diabetic problems). As discussed earlier, we are interested in controlling for individuals’ time-invariant unobservables, so we generate the marginal effects based on the Fixed Effect estimations.15 In this analysis, we focus on the                                                 15 Also, the Hausman test rejects the Random Effect model in favor of the Fixed Effect model in this case too, so we used fixed effect for further analysis.  29  interaction of covariates with diet drink to explore the heterogeneity and the underlying mechanism for the effect of diet drink on food calories.16  We start by explaining the marginal effect of selecting diet drinks across gender using full-model Fixed Effect estimations. Based on the discussed literature in the Data and Variable Description section, we expect females to have healthier food choices and more salient dieting concerns. One can speculate that they remain committed to one goal, i.e., cutting calories, when they are on a constrained diet, and therefore they should make decisions congruent with their single goal (see Background section for discussion). On the other hand, male consumer decisions are more hedonic, and they may be balancing between competing goals when they choose diet drinks. Interestingly, as exhibited in Figure 2-9, we find a positive effect for diet drink on food calories (+51 calorie) among males (balancing effect).17 As expected based on the main effect model findings (Table 2-5), we do not detect an overall effect, because the balancing effect (more food with a diet drink) among males is masked by the opposite-in-direction (less food with a diet drink) but non-significant effect for females.  Similar to the total calorie analysis, we incorporate the information on drink choice into marginal effects by showing the effect across drink size (Figure 2-10). We see that the balancing                                                 16 As a further robustness check, we illustrate the model-building process by providing the estimations and marginal effect for smaller interaction models based on Fixed Effect and OLS (Appendix A and Table A-3 and Table A-4). The coefficient estimation patterns for gender and age in the full model emerge even in simpler interaction models. The model with gender interaction only has a significant interaction coefficient (column 2), and age patterns are similar based on the example age marginal effects (columns 3 and 4). Also, the model without income, party size, and meal occasion variables provides qualitatively similar findings. 17 Note that the male dummy variable is dropped in the Fixed Effect estimations since it does not have any within-subject variation.  30  among males holds for medium and to a greater extent small CSDs; and the negative effect (fewer food calories when choosing a diet drink), even though insignificant, holds for all drink sizes among females.  In the next step, we probe further by bringing age variables into marginal effects (Figure 2-11). As discussed in an earlier section, we expect younger consumers who switch to diet drink to have hedonic goals competing with dieting goals. The estimation of marginal effects in this step clarifies that the male’s balancing effect is driven by, as expected, the younger demographics—19 to 42 and 19 to 35 years old for small and medium-size drinks, respectively. Additionally, and in line with our speculations, we find that middle-aged females—in our results, 38 to 57 years old for a medium-size drink—consume significantly fewer calories (max 50 cal) when they switch to diet drink. For young females, the expected effect of age and gender on having competing goals (hedonic and dieting) works in the opposite direction, possibly resulting in no effect. The reported null effect for older females is worthy of further study.   The next key dimensions to explore are income, party size, and meal occasion dummies. We are unable to detect any significant interaction between the choice of a diet drink and income, meaning that we do not find different patterns of food calorie change with choice of drink across individuals with different income levels. In addition, our estimations find a direct positive effect for party size on food and total calorie consumption (i.e., higher calorie count for larger party size), but do not provide any evidence on the interaction of diet and party size, meaning that we cannot find any evidence that presence of other people at a meal would affect the relative calorie of the food eaten with diet versus regular drinks. Finally, the meal occasion  31  variables, despite their intuitive direct effect on calories, discussed in Data and Variable Description section, do not show any interaction with diet-drink variable in the Fixed Effect estimations  2.4.3 Identification Discussion  So far we have established a relationship between the choice of a diet drink and change in the level of food calorie intake relative to the level of calories accompanying a regular soft drink.18 We found that young males consume more calories from food when they switch to diet drink, and middle-age females do the opposite. However, the assignment of treatment, choice of diet drink, is not random in our observational setting, thus we need to clarify the assumptions under which our findings have causal interpretation.19  We adopt the omitted-factors framework to identify and discuss the sources of endogeneity. Along with the multiple factors controlled for by the covariates included, omitted factors, captured by the error term, can potentially affect choice of both drink and food. Such “third” factors are the source of endogeneity or systematic (not ignorable) treatment assignment (or treatment selection on unobservables). We classify these unobservables into three mutually                                                 18 For the identification discussion we focus only on the food calorie model, because total calories already includes drink calories and it is trivial that choice of drink affects total calories. In other words, our discussion will be clearer when we focus on effect of drink choice on food calories rather than on calories derived from food plus drink choice. 19 In our case, the notion of random treatment assignment is not easy to conceptualize even in a hypothetical laboratory experiment. Given that the treatment of interest to address the policy debate is the deliberate choice of diet drink, it is not clear if the forced choice of diet drink in laboratory setting will generate externally valid insight for the field case where consumers freely choose between diet and regular drinks. Once we remove the forced drink choice feature from the experiment design, the randomness and therefore exogeneity of the treatment is under question, and we end up with a set-up similar to the current study.   32  exclusive categories and explain our strategy in addressing them or the assumptions required to make causal claims despite their potential influence: 1. Individual-fixed time-invariant factors (e.g., chronic disease, BMI, eating habits) 2. Time-variant factors that affect the choice of drink but are independent from the food choice (e.g., trying a new flavor available for diet drinks, being nudged by a promotion for diet drinks, or experimentally sampling the taste of a diet drink)  3. Time-variant factors that affect both the choice of the drink and food (adopting a calorie-cutting diet because summer is ahead, a marriage ceremony is expected, or a sudden weight gain is noticed)  The first group of unobservables is already controlled for in the Fixed Effect estimation procedure. These factors are the potential drivers for the findings in other field studies (e.g., Fowler et al., 2008). The second group of unobservables, by definition, is not a threat to exogeneity of the choice of diet drink. In other words, the ignorability of treatment is preserved if the switch to the diet drink is due solely to such factors, and our reported findings will have causal implications. The third group of unobservables is of primary interest in the identification discussion. We cannot make any causal claim without imposing assumptions on existence or direction of their effect. Neglecting these factors requires a strong assumption that we avoid, but we show that by making a reasonable assumption on the direction of the effect for these unobservables, we can claim causality for a key portion of our findings.  We refer to the third group of unobservables as calorie-cutting-goals. The assumption that we impose is a rather intuitive one: The effect of calorie-cutting-goals, if any, will be  33  negative on the food calorie consumption (and obviously positive on the likelihood of choosing a diet drink). Therefore, any change in the food calorie count (food calories with diet CSD subtracted by food calories with regular CSD) observed in the presence of calorie-cutting-goals is smaller than in the counterfactual scenario where calorie-cutting-goals are controlled for. In the case of our research, any significant positive effect on food calorie consumption (i.e., more calories with diet drink) is an underestimate of the effect in a hypothetical experimental scenario where calorie-cutting-goals are controlled for. Revisiting the reported findings, when we take the effect of unobservables into account, the reported positive effect of diet drink on food calories for young males is an underestimation. Therefore, we identify a causal positive effect on food calorie count by diet drink among young males at least as large as the reported magnitude of the balancing effect. To complete the discussion, we can say that under the mentioned assumption on unobservables, the negative and null effects remain only correlational in nature.  To summarize our empirical results, we find strong evidence for a calorie-cutting benefit of diet drinks in terms of total meal (food plus drink) calories. Moreover, we find that young males increase their caloric consumption from food when they choose a diet drink, and middle-age females do the opposite, but none of the changes in the food calorie count are great enough to surpass or cancel out the calories saved with choice of a diet drink. Further, we show that under rather intuitive assumptions on unobservable factors, the reported effect of a diet drink choice for young males has a causal interpretation, and other reported effects remain correlational.   34  2.5 Discussion  In this paper, we offer a novel insight addressing the debate on the harm and benefit of diet soft drinks. We use food and drink consumption data at the individual level from people eating at a major fast-food restaurant to examine whether having a diet drink leads to over-consumption of calories in other parts of the same meal. To control for possible confounding factors in the field setting, we use within-subject variation and recent developments in econometrics of treatment effects to address some of the concerns regarding biases in estimation. In terms of outcome measures, we focus on the differences in caloric intake from total meal (food plus drink) and food consumed with diet CSD as compared to regular CSD. The large size of our dataset helps us to estimate these differences across various demographic groups.  We find robust evidence for lower total calorie level in meals with a diet drink as compared to meals with a regular drink. This result generally holds for both males and females, and across age groups. Then we focus on the influence of choice of diet drink on the food’s calorie count and report that young males (19 to 35) have more food when they select a diet drink, but the extra calories in the food are compensated for by the calories saved with the choice of a diet drink. In addition, we interestingly find that on average females consume fewer calories from food with diet CSDs than with regular CSDs, and this difference is mainly driven by females in the age range of 36 to 58. Overall, we find strong support for calorie cutting benefit of diet drink, and no evidence for overcompensating the calorie saved by choosing the diet drink, in a meal eaten away from home.   35  Our findings also provide field evidence to test implications of behavioral theories—such as multiple goal pursuit (e.g., Fishbach & Dhar, 2005)—that seek to explain the relationship of decisions within a consumption episode. While the majority of our sample did not change their eating behavior based on drink choice, we find effects in line with theoretical scenarios where an individual pursues a single goal (middle-age females remaining committed to calorie-cutting goal) as well as the pursuit of multiple goals (young males balancing opposing goals—calorie-cutting and indulgence). Our observational non-experimental methodology cannot directly test the validity of behavioral theories, but our findings encourage special attention to measuring the behavioral effects in terms of quantities that are relevant to the research issue (for example, instead of reporting probability of choosing an indulging food, report the expected calorie or fat intake) and including wide demographic groups in testing the discussed theories.     2.5.1 Public Policy and Marketing Implications  In the present debate on the link between diet drinks and weight gain, one of the common arguments against diet soda is the notion of substituting the drink’s calories with more calories from other items. Our findings may help to alleviate such concerns since they show that in terms of total caloric intake, consumers significantly benefit from consuming diet CSDs. Therefore, our findings generally support the policies or practices that nudge consumers away from sugary  36  drinks and to diet drinks. Still, our evidence is based on the decisions within a meal at a restaurant, and we should be careful not to overgeneralize the implications.20  For example, policy makers have suggested that a sugar tax (e.g., in New York City) be added to the price of regular drinks. This practice could nudge consumers to choose diet drinks, and our study suggests that we should not expect backfiring due to choice of diet drink in that meal. However, one should note that our study does not take price difference into account, and in this case consumers may decide to spend the money (not the calories) saved by choosing a diet drink on other food items. Therefore, both dollar cost and calories should be considered in an ideal study of this policy. This is an interesting avenue to add to the knowledge in this area—see Schroeter, Lusk, and Tyner (2008) for an example of related work.  Another policy option, similar to the route taken by the tobacco industry, is to put warning labels on sugary drinks reminding consumers about their sugar content and adverse health consequences. Our findings suggest that this practice will not be harmful in terms of extra calories consumed in other parts of the meal. However, the body of literature on nutrition labeling suggests that providing information is not as effective in changing behavior as altering convenience and cost. According to George Loewenstein, an economist at Carnegie Mellon University, “There are very few cases where social scientists have documented that giving people information has changed their behavior very much. Changing prices and changing                                                 20 As mentioned earlier, diet drinks are arguably linked to other adverse health consequences such as cardiovascular problems and diabetic disease. The validity of these claims and the relative effect of diet drinks compared to regular drinks in these issues are still open questions and are beyond the scope of the current work.  37  convenience have a big impact. Providing information doesn’t” (Tavernise, 2014). Thus, effectiveness of this practice in the field is an empirical question and a good area to explore.  In addition to policy practices, major industry players have reacted to criticism of CSDs by introducing more low-calorie or diet-drink options, smaller packages, and promotion of diet drinks. Our findings support these actions and indicate that we should not expect an immediate rebound in calorie consumption. For instance, the introduction of Coke Zero, which unlike Diet Coke targets males (Tungate, 2008), can benefit consumers and in particular may facilitate the choice of a low-calorie option for males.  Furthermore, we suggest some practices for the context of the current analysis, restaurants and fast-food outlets, to encourage customers to order diet drinks as compared to regular ones. These practices can help consumers cut calories and will enhance the health-oriented image of restaurants. First, restaurants can increase the number of flavors offered for diet drinks. Many restaurants offer greater variety for regular drinks, and this practice can be changed by providing more types of non-sugary drinks. Second, fast-food restaurants can make diet drink the default choice in their combo/meal order (e.g., burger, fries, and drink). A strong default effect documented in other areas (Johnson & Goldstein, 2003) can work to the benefit of consumers by lowering their calorie consumption. Third, restaurants can charge for refill of  38  sugary drinks and have free refill for diet drinks.21 This intriguing possibility should be tested for potential price effects and restaurant-switching behavior of consumers.   2.5.2 Limitations and Concluding Remarks  We recognize, of course, that our findings are based on individual meals and not on long-term physiological or behavioral changes in consumption patterns. Most of the studies we cite have a similar focus on short-term effects, mainly because conducting long-term experiments on human subjects is extremely challenging. We also cannot comment on inter-meal substitutions since we do not observe the meals at home, and the incidents of eating out are too sparse in our dataset. Inter-meal substitution provides the full picture of calorie substitution patterns and is a critical extension to this work—see Khare and Inman (2009) for an example of research on inter-meal dynamics.  Finally, our results are based on the consumption of food and drink at McDonald’s, the largest fast-food chain in the world. We expect that our results hold for similar fast-food chains, and likely for other away-from-home settings. With fast-food restaurants accounting for a substantial portion of all away-from-home food consumption occasions (approximately 64% in Canada), this is an important area to understand. Examining behavior in other away-from-home settings would require detailed data on caloric content of menu items, and would be a great extension.                                                  21 Or make refills of sugary (regular) drinks less convenient by requiring customers to ask for refills of sugary drinks from the counter.  39  2.6 Tables and Figures  Table 2-1. Descriptive Statistics of Demographic and Meal Occasion Variables Variables Full Sample  All McDonald’s Analysis Sample % Male 43.58 45.05 45.45 Age (average) 45.71 37.84 41.18 % Income Level 1-3 ($24,999 and under) 13.41 14.68 13.74 % Income Level 4-6 ($25,000–$54,999) 23.95 26.14 23.07 % Income Level 7-9 ($55,000–$69,999) 28.68 29.41 30.88 % Income Level 10 ($70,000 and above) 33.96 29.77 32.31 Size of the Party (average) 2.25 2.41 2.30 % Breakfast 9.72 13.23 1.56 % Lunch 30.64 37.18 52.21 % Dinner 35.47 28.64 38.78 % Diet CSD 24.88 23.37 27.72 % Small Size CSD 21.55 15.16 14.21 % Large Size CSD 24.59 23.52 26.57 Total Calories (average) NA NA 816.96 Sample Size 682,830 59,730 11,738     40   Figure 2-1. Diet and Regular Drink Orders by CSD Size   Figure 2-2. Diet and Regular Drink Orders by Gender and Age   020406080100percentLarge CSD Medium CSD Small CSDMeal with Diet CSD Meal With Regular CSD020406080100percentFemale Male19 to 29 30 to 39 40 to 49 50 to 59 60 to 69 70 to 80 19 to 29 30 to 39 40 to 49 50 to 59 60 to 69 70 to 80Meal with Diet CSD Meal with Regular CSD 41  Figure 2-3. Average Total Calories by CSD Type and Size    Figure 2-4. Average Food Calories by CSD Type, Size, and Gender   650.2 849   801.4 1062  809.5 1201  05001,0001,500Food + Drink CalorieSmall CSD Medium CSD Large CSDDiet Regular Diet Regular Diet Regularby CSD Type and SizeAverage Total Meal Calories624  671  777  812  774  820  753  758  859  876  861  919  02004006008001,000Food CalorieFemale MaleSmall CSDMedium CSDLarge CSDSmall CSDMedium CSDLarge CSDby CSD Type, CSD Size, and GenderAverage Food CaloriesDiet Regular 42  Figure 2-5. Average Food Calories by CSD Type, Gender, and Age Groups  832.5 828.9 798.8 813.4 767.4 776.5 711.3 736   684.4 720.9 514.6 639.8 1019  937.6 928.1 913.4 870.6 881.3 795.8 815.8 757.2 770.9 688.1 708.2 02004006008001,000Food CalorieFemale Male19 to 2930 to 3940 to 4950 to 5960 to 6970 to 8019 to 2930 to 3940 to 4950 to 5960 to 6970 to 80by Age Group, and GenderAverage Food CaloriesDiet Regular 43   Table 2-2. Total Calories Estimations – Main Effect   DV: Total Calories  FE FE FE OLS OLS OLS RE Diet -229.95*** -227.58*** -227.46*** -289.91*** -283.36*** -241.48*** -239.22*** Small  -134.78*** -128.68***  -192.26*** -154.56*** -145.37*** Large  74.69*** 71.68***  102.69*** 86.72*** 87.18*** Male   NA   88.48*** 81.74*** Age   2.34   2.64 1.48 Age2   -0.06   -0.07*** -0.05*** Income   -2.32   -4.27*** -4.02*** Party Size   13.80***   3.72* 5.34** Breakfast   -93.26***   -198.74*** -120.47*** Lunch   121.96***   109.98*** 113.78*** Dinner   145.52***   147.33*** 147.63*** Intercept 1052.16*** 1051.40*** 1052.63*** 1069.81*** 1068.77*** 1059.45*** 1055.92*** N 11738 11738 11738 11738 11738 11738 11738 R2 0.053 0.110 0.148 0.211 0.307 0.394  Note: * p < 0.1, ** p < 0.05, *** p < 0.01    Table 2-3. Total Calories Estimations – Matching Methods  Matching Method DV: Total Calories  Average Treatment Effect Propensity Score -230.24*** Regression Adjustment -237.9*** Nearest Neighbor -235.66*** N 11738 Note: * p < 0.1, ** p < 0.05, *** p < 0.01; these estimations include Small, Large, Male, Age, Age2, Income, Party Size, Breakfast, Lunch, and Dinner variables as covariates. 44  Table 2-4. Total Calories Estimations – Full Model  OLS Random Effect Fixed Effect Diet -223.30*** -223.85*** -208.14*** Small -166.90*** -164.29*** -156.78*** Large 116.86*** 112.30*** 99.86*** Male 71.83*** 65.93*** N.A. Age 2.07 3.54* 4.52 Age2 -0.06*** -0.07*** -0.10 Income -3.29** -4.76*** -2.70 Party Size 4.70 6.02 15.10* Breakfast -172.93*** -162.28*** -153.15*** Lunch 60.59*** 76.10*** 83.47*** Dinner 90.95*** 103.19*** 100.30*** Diet X Small  65.33*** 78.97*** 98.92*** Diet X Large -111.68*** -102.43*** -103.69*** Diet X Male 29.59 30.49* 75.00*** Diet X Age 8.75* -0.87 -12.04* Diet X Age2 -0.10* 0.00 0.12 Diet X Income -2.19 -1.66 -3.05 Diet X Party Size -3.69 -3.85 -6.27 Diet X Breakfast -34.38 -25.81 -3.75 Diet X Lunch 74.71** 37.26 16.28 Diet X Dinner 68.59* 28.59 15.99 Small X Male 27.15 23.42 8.51 Small X Age -7.12* -9.66** -13.30*** Small X Age2 0.07* 0.10*** 0.15*** Small X Income 3.44 5.91** 8.63** Small X Party Size 3.17 3.48 5.75 Small X Breakfast 164.17*** 145.95** 89.76 Small X Lunch 75.38** 57.69** 52.53 Small X Dinner 103.81*** 89.39*** 108.00*** Large X Male 32.26** 23.47* 34.09* Large X Age -1.65 -0.00 -0.08 Large X Age2 0.03 0.01 0.02 Large X Income -2.68 -0.83 -1.33 Large X Party Size  -2.99 -0.70 -5.06 Large X Breakfast 36.39 30.90 23.25 Large X Lunch 56.54* 45.76 39.69 Large X Dinner 71.54** 61.03* 50.32 Small X Diet X Covariates Included Included Included Large X Diet X Covariates Included Included Included Intercept 1055.54*** 1053.12*** 1053.51*** Sample Size 11738 11738 11738 R2 0.409 0.167 0.172 Note: * p < 0.1, ** p < 0.05, *** p < 0.01 - Three way interactions coefficient are not significant (p > .05) and are shown in Appendix A for the sake of exposition     45   Figure 2-6. Total Calories – Estimated Effect by Gender  Note: The marginal effects are based on the coefficient estimates presented in Table 2-4, Fixed Effect column    -300-250-200-150-100(Diet - Regular) Total Meal CaloriesBy GenderMale Overall Femaleand 95% CIEstimated Diet Drink Effect on Total Meal Calorie 46  Figure 2-7. Total Calories – Estimated Effect by CSD Size and Gender  Note: The marginal effects are based on the coefficient estimates presented in Table 2-4, Fixed Effect column-400-300-200-1000(Diet - Regular) Total Meal Calories By Gender and Drink SizeMale, Small CSDMale, Medium CSDMale, Large CSDFemale, Small CSDFemale, Medium CSDFemale, Large CSDand 95% CIEstimated Diet Drink Effect on Total Meal Calorie 47  Figure 2-8. Total Calories – Estimated Effect by CSD Sizes, Gender, and Age  Note: The marginal effects are based on the coefficient estimates presented in Table 2-4, Fixed Effect column -600-400-2000200Total Meal Calories20 40 60 80AgeMale, Small CSD-600-400-2000200Total Meal Calories20 40 60 80AgeMale, Medium CSD-600-400-2000200Total Meal Calories20 40 60 80AgeMale, Large CSD-600-400-2000200Total Meal Calories20 40 60 80AgeFemale, Small CSD-600-400-2000200Total Meal Calories20 40 60 80AgeFemale, Medium CSD-600-400-2000200Total Meal Calories20 40 60 80AgeFemale, Large CSD 48    Table 2-5. Food Calories Estimations – Main Effect  DV: Food Calories  FE FE FE OLS OLS OLS RE Diet 2.57 2.95 3.23 -52.97*** -49.36*** -7.90 -6.49 Small  -92.01*** -86.31***  -145.17*** -107.64*** -99.37*** Large  4.73 1.61  30.19*** 14.06** 14.38** Male   NA   87.55*** 80.42*** Age   2.16   2.80 1.44 Age2   -0.06   -0.07*** -0.05*** Income   -2.33   -4.20*** -3.89*** Party Size   13.45**   3.11 5.06** Breakfast   -88.31**   -185.04*** -115.87*** Lunch   120.44***   108.34*** 111.66*** Dinner   142.91***   146.63*** 145.42*** Intercept 816.52*** 828.21*** 830.14*** 832.86*** 844.60*** 835.47*** 832.47*** N 11738 11738 11738 11738 11738 11738 11738 R2 0.000 0.019 0.059 0.010 0.060 0.177  Note: * p < 0.1, ** p < 0.05, *** p < 0.01   Table 2-6. Food Calories Estimations – Matching Methods  Matching Method DV: Food Calories  Average Treatment Effect Propensity Score 5.66 Regression Adjustment -9.29* Nearest Neighbor -0.89 N 11738 Note: * p < 0.1, ** p < 0.05, *** p < 0.01; these estimations include Small, Large, Male, Age, Age2, Income, Party Size, Breakfast, Lunch, and Dinner variables as covariates.   49  Table 2-7. Food Calories Estimations – Full Model   OLS Random Effect Fixed Effect Diet -3.30 -3.85 11.86 Small -96.90*** -94.29*** -86.78*** Large 16.86** 12.30* -0.14 Male 71.83*** 65.93*** N.A. Age 2.07 3.54* 4.52 Age2 -0.06*** -0.07*** -0.10 Income -3.29** -4.76*** -2.70 Party Size 4.70 6.02 15.10* Breakfast -172.93*** -162.28*** -153.15*** Lunch 60.59*** 76.10*** 83.47*** Dinner 90.95*** 103.19*** 100.30*** Diet X Small  -4.67 8.97 28.92 Diet X Large -11.68 -2.43 -3.69 Diet X Male 29.59 30.49* 75.00*** Diet X Age 8.75* -0.87 -12.04* Diet X Age2 -0.10* 0.00 0.12 Diet X Income -2.19 -1.66 -3.05 Diet X Party Size -3.69 -3.85 -6.27 Diet X Breakfast -34.38 -25.81 -3.75 Diet X Lunch 74.71** 37.26 16.28 Diet X Dinner 68.59* 28.59 15.99 Small X Male 27.15 23.42 8.51 Small X Age -7.12* -9.66** -13.30*** Small X Age2 0.07* 0.10*** 0.15*** Small X Income 3.44 5.91** 8.63** Small X Party Size 3.17 3.48 5.75 Small X Breakfast 164.17*** 145.95** 89.76 Small X Lunch 75.38** 57.69** 52.53 Small X Dinner 103.81*** 89.39*** 108.00*** Large X Male 32.26** 23.47* 34.09* Large X Age -1.65 -0.00 -0.08 Large X Age2 0.03 0.01 0.02 Large X Income -2.68 -0.83 -1.33 Large X Party Size  -2.99 -0.70 -5.06 Large X Breakfast 36.39 30.90 23.25 Large X Lunch 56.54* 45.76 39.69 Large X Dinner 71.54** 61.03* 50.32 Small X Diet X Covariates Included Included Included Large X Diet X Covariates Included Included Included Intercept 835.54*** 833.12*** 833.51*** Sample Size 11738 11738 11738 R2 0.187 0.068 0.074 Note: * p < 0.1, ** p < 0.05, *** p < 0.01; Three way interactions coefficient are not significant (p > .05) and are shown in Appendix A for the sake of exposition      50   Figure 2-9. Food Calories – Estimated Effect by Gender  Note: The marginal effects are based on the coefficient estimates presented in Table 2-7, Fixed Effect column      -50050100(Diet - Regular) Food CaloriesBy Drink SizeMale Overall Femaleand 95% CIEstimated Diet Drink Effect on Food Calorie 51    Figure 2-10. Food Calories – Estimated Effect by CSD Size and Gender  Note: The marginal effects are based on the coefficient estimates presented in Table 2-7, Fixed Effect column     -50050100150200(Diet - Regular) Food CaloriesBy Gender and Drink SizeMale, Small CSDMale, Medium CSDMale, Large CSDFemale, Small CSDFemale, Medium CSDFemale, Large CSDand 95% CIEstimated Diet Drink Effect on Food Calorie 52  Figure 2-11. Food Calories – Estimated Effect by CSD Sizes, Gender, and Age  Note: The marginal effects are based on the coefficient estimates presented in Table 2-7, Fixed Effect column  -2000200400Food Calories20 40 60 80AgeMale, Small CSD-2000200400Food Calories20 40 60 80AgeMale, Medium CSD-2000200400Food Calories20 40 60 80AgeMale, Large CSD-2000200400Food Calories20 40 60 80AgeFemale, Small CSD-2000200400Food Calories20 40 60 80AgeFemale, Medium CSD-2000200400Food Calories20 40 60 80AgeFemale, Large CSD 53  Chapter 3: Restaurant Diners’ Reaction to Incidents of Mad Cow Disease: Stay Home, Eat Less Beef, or Life as Usual?  3.1 Introduction   How do consumers react to product-related crises, such as those involving unsafe food or product design flaws? To what extent do they change their consumption behavior in different markets? While some crises—for instance, those involving Tylenol (Murray & Shoehn, 1992) and Firestone (Advertising Age, 2000)—have led to sales drops and recalls for individual companies, others have the potential to affect categories of products and can compromise an entire industry. In particular, incidents tarnishing the safety of basic food products (e.g. beef) can result in spiraling public anxiety and escalating media attention, a situation sometimes labeled a “food scare.” Since the early 1990s, many countries have experienced one or more significant food scares such as those triggered by outbreaks of bird influenza, salmonella, listeria, and E. coli (Knowles, Moody, & McEachern, 2007). The focus of our study, bovine spongiform encephalopathy (BSE; also known as mad cow disease), has been a prime case of a food scare that captured public attention and has significant health and economic consequences across multiple countries.22                                                  22 The United Kingdom’s outbreak of BSE in the 1990s, for example, which led to the tragic death of 176 people, dropped beef retail sales from 617 Kilo-tons in 1988 to 390 Kilo-tons in 1996 (Yeung and Morris, 2001). Moreover, the first North American discovery of BSE disease in the Canadian province of Alberta in May 2003 resulted in severe international trade restrictions on beef and cattle and the destruction of thousands of cattle.    54  The unique features of incidents of BSE make it an appropriate example to study the effect of food scares on consumers’ decision making. It offers a case of a crisis that strongly captured public awareness and increased food-safety concerns. This is due to BSE’s extreme health consequences (once infected there is no cure) and the recurrence of BSE crises.23 More importantly for our purposes, it provides a clean set-up to investigate the change of consumers’ behavior due to crises. In contrast to swine influenza and bird influenza, BSE is not contagious. Therefore, BSE should not change people’s behavior such that they avoid public places, and any change in consumer behavior can be attributed to concerns regarding safety of the product. In addition, it affects only one type of product (unlike salmonella), and contaminated product cannot be made edible by any processes such as cooking, as in the case of bird influenza.   At the macro and retail levels, BSE outbreaks have been shown to negatively influence beef consumption around the world—for example, by studies in the UK (Burton & Young, 1996), the Netherlands (Mangen & Burrell, 2001), Italy (Mazzocchi & Lobb, 2005), Japan (Ishida, Ishikawa, & Fukushige, 2010; Peterson & Chen, 2005), and the USA (Schlenker & Villas-Boas, 2009). In Canada, the country of our study, Ding, Veeman, and Adamowicz (2009) reported a reduction in total consumption of beef followed by a recovery in less than four months after the first three Canadian BSE outbreaks. Additionally, Peng, McCann-Hiltz, and Goddard                                                 23 BSE has affected multiple countries over the past thirty years. BSE’s human variant, called variant Creutzfeldt Jakob Disease (vCJD), was first diagnosed in the UK in 1996. vCJD is transmitted by eating contaminated beef, is not contagious, and has no cure. By October 2010, vCJD cases had been reported in 12 countries, resulting in 227 deaths globally. BSE was the first food scare to affect food safety on a European scale and the one that triggered reform of existing legislation and established new regulatory institutions across Europe (Knowles, Moody, & McEachern, 2007).    55  (2004) found that newspaper articles related to the first BSE case in Canada had a significant, but small, negative effect on retail beef purchases.24  As discussed below, we found only two studies examining the effect of food scares on food away from home (restaurants), a central and vulnerable sector at the time of crises.25 Compared to eating at home, restaurants provide lower control over food preparation and less information on ingredients of food—features that could intensify the perception of risk at the time of crises (Yeung & Morris, 2001). Further, if consumers are concerned about restaurant food safety, they can simply stay home and avoid the food away from home (FAFH) market altogether, a less viable course of action for grocery stores. Therefore, restaurants are an important sector to study at the time of crises, and findings from other contexts are not generalizable to this area.26  We found only two other studies on the effect of food scares in the restaurant context. First, in a small-scale study, Reynolds and Balinbin (2003) surveyed 86 London restaurant                                                 24 Another stream of literature has tried to explain the variations in the reaction to crises by risk framework and nature of media coverage (Maynard & Wang, 2010). For instance, using Pennings, Wansink, & Meulenberg (2002) risk perception/risk aversion framework, Yang and Goddard (2011) segmented Canadian consumers according to their stated level of risk aversion and risk perception and showed that segments with higher risk aversion or risk perception react more strongly than others to the crises.  25 In Canada, the Food Away From Home (FAFH) expenditure share of the total food expenditure was 27% in 2013 (Statistics Canada). FAFH constitutes about half of households’ expenditure on food in the US (US Department of Agriculture’s Economic Research Services). Expenditure on FAFH in the USA reached $565 billion in 2008 (48.5% of total food expenditure). This percentage was 45.5% in 1988 and 33.1% in 1968 (Roosen, Lusk, & Shogren, 2011).  26 We have some anecdotal evidence on food scares raising concern for restaurants or affecting their sales. In Japan, sales in stores open a year or more were reported to drop by about 10% in October because of fear of mad cow disease  (see: http://atimes.com/atimes/Japan/DG11Dh05.html). In addition, “Fast-food chains [such as McDonald’s and Burger King] took pains in May 2003 [after the first Canadian case of BSE] to emphasize their beef safety standards.” (Day, 2003)   56  owners on changes in their restaurant performance following an outbreak of BSE. The restaurant owners did not report an overall decrease in sales, but they indicated a reduction in beef orders. Moreover, restaurants featuring few or no beef items reported stronger overall sales than the ones with a strong beef orientation (places where beef items constitute a large portion of consumers’ orders). Second, Maynard, Goddard, and Wang (2008) analyzed household beef purchase at fast-food restaurants in Alberta and Ontario at the time of the BSE crises. Using a self-reported household-level dataset, they implemented a double-hurdle model to capture the change in consumers’ beef purchase decision—whether or not to buy any beef, and how many beef items to order. They found a reduction in beef purchase likelihood of Ontario consumers but not Albertans, and stable levels of beef consumption for those who did buy beef in both provinces after the increase in BSE media coverage.27   In this work, we analyze a consumer-level FAFH consumption dataset merged with local news index based on print media and local weather information to shed light on consumers’ decisions made at the time of BSE outbreaks. This is the first paper in the current context to model the dine-out and meal-order decisions in all types of restaurant while allowing for interrelatedness of the dine-out and food-order decisions, perhaps due to a forward-looking attitude of consumers when deciding whether to dine out, as discussed later.28 We are also the                                                 27 We replicated their analysis (Appendix G) for all types of restaurants in Alberta and Ontario. We did not detect any reaction in Alberta, but we found beef order reduction in the second stage in Ontario. These findings are in line with those of Maynard et al. (2008). 28 Unlike our modeling procedure, Maynard, Goddard, and Conley (2008) do not model the dine out decision and the interdependence of the participation and consumption decisions. This limits their ability to control for common unobserved shocks to both stages (dine out and meal order), and also, their study cannot provide insight for changes in restaurant traffic.    57  first to look into variation of consumers’ reaction to a food scare across national chain restaurants and local restaurants. Furthermore, our proposed estimation procedure incorporates the participation variable (dining out in our case) into a modified double-hurdle model. This helps us to explicitly distinguish two types of zero consumption in our dataset and eventually in our estimations: zero consumption due to not participating (not dining out) and zero due to no consumption (no beef orders) conditioned on participation. The existing double-hurdle estimation procedures typically combine these two types of zeroes and therefore cannot use the participation data in estimation. The proposed methodology matches the nature of decision-making and the data-generating process in the FAFH context as well as any other two-stage decision-making scenario in which the decision in each stage is observable and decisions of the two stages are interrelated.  During a BSE crisis, we find a marginally significant decrease (up to 1.2%) for the probability of dining out in all types of restaurants and a significant negative effect for beef-oriented restaurants. We also report a significant decrease (up to 1.7%) for conditional beef orders at the time of BSE news coverage, a positive effect on seafood and pork, and no effect for chicken and vegetarian orders. Focusing more narrowly on beef-oriented restaurants, we find a significant and larger drop (up to 3.4%) in visits to these restaurants and, conditional on this behavior, a further decrease (up to 1.7%) in beef orders at these restaurants. We then look into beef-oriented restaurants to examine the ability of brand equity to soften the effect of food crises on national chain restaurant sales. Surprisingly, despite the substantial expenditures by nationally branded fast-food chains to build consumer loyalty and trust, in times of BSE crises, consumers shy away from national chains and, once there, order fewer meals containing beef. By contrast,  58  for locally owned beef-oriented restaurants, we find no significant decrease in monthly visits to these restaurants or number of beef meals ordered during the time of a BSE crisis.  In the rest of the chapter, we first describe the main dataset and additional data collection processes. Next we elaborate on empirical strategy, and then discuss the findings. In the last section, we discuss the managerial and public policy implications of our work and suggest directions for future research in this area.  3.2 Data  The central dataset in our analysis is the National Panel Division Group’s Consumer Reports on Eating Share Trends (NPD Group’s CREST). It is composed of self-reported food-away-from-home consumption data drawn from a wave panel of Canadian households. We merged these data with the print media coverage of BSE crises and the local weather data, extracted from the Canadian Newsstand Complete and Environment Canada, respectively.29  3.2.1 Sample of Respondents    The NPD Group’s CREST data panel includes 20,364 Canadian adults (19 years old and above) from 12,651 households, and it spans the time period from September 2000 to December 2006. The database contains detailed demographic characteristics such as age, education,                                                 29 We also used the Factiva database to access French-language press.  59  income, marital status, children, province, city size, and preferred language (English/French). For each meal occasion, the database contains information on specific items ordered by each individual in the panel. Each restaurant has just one food orientation (e.g. hamburger, steak, beef/roast beef, or chicken orientation) and type (fast-food/drive-in, family type, informal/casual, and formal dining) in the dataset. For example, McDonald’s is coded as hamburger fast-food/drive-in restaurant. The database contains information on more than 1,262 restaurants and coffee shops in Canada and codes 253 different food and drink items.   The coding of the meal items in the NPD Group’s CREST dataset for the most part clearly distinguishes the type of meat ordered. We code such items as hamburger, steak, and roast beef as “beef” meals, and fried chicken, chicken wings, chicken nuggets, chicken strips, and chicken sandwiches as “chicken” meals. Similarly, pork and seafood items are separately coded, and we combine the orders of salads, vegetarian subs, and vegetarian burgers in a “vegetarian” group. However, certain food items such as pizza and submarine sandwiches do not have a clear indicator of their meat type in the dataset, and thus they are excluded from the above-mentioned categories.30 Orders of each meal item within a month for each individual are captured by a separate count variable. In addition, we generate a binary participation variable that is equal to one when a panelist reports consuming at least one meal away from home in that month and is equal to zero otherwise.                                                  30  The data do not provide clear information on pizza toppings, and subs are categorized in chicken, turkey, meat, and other subs. That is, beef subs are not distinctively coded, and given the prevalence of ham sub sandwiches we chose not to code “meat subs” as a meal with beef; this way we have a more reliable count of beef items. Also, note that exclusion of these items from the mentioned categories would not result in dropping those observations, because the mentioned categorization is used to construct the dependent variables, not the sample of analysis.  60  Table 3-1 provides descriptive statistics for the key variables in our analysis. On average, 29% of the sample dine out once or more in a month. The average numbers of monthly beef, chicken, seafood, pork, and vegetarian orders in the sample are .15, .13, .06, .09, and .16, respectively. Conditioned on dining out in the period, the average count of monthly beef, chicken, seafood, pork, and vegetarian orders per month rises to .52, .46, .20, .29, and .53. In addition, among all the meals ordered, 16% are beef meals, and conditioned on ordering some type of meat the percentage goes up to 26%. Based on the presented statistics, most of the participants are not highly active in the FAFH market; so to alleviate the extreme sparsity of the dataset we aggregate the data to the monthly level, a common unit of observation in the FAFH literature (e.g., Maynard et al., 2008; Pritchett, Johnson, Thilmany, & Hahn, 2007; Saghaian & Reed, 2007).31  In Table 3-2, we summarize the sample’s demographic variables included in the analysis and compare them to the corresponding national measure obtained from Statistics Canada for the year 2003. The table reveals that 57% of adults in the sample are married and 36% have at least one child. Average age is 49 and 29% of them have a university degree. Income variable has ten categories, and its weighted average is $63,700. Demographic profile of our sample (Nielsen) is                                                 31 We implemented a supplemental analysis to check whether the monthly-level aggregation of our sparse meal occasion data would influence our findings. The concern arising from this aggregation is that the binary participation variable may not fully capture the change in the store traffic. Participation variable captures the changes of restaurant visits from any positive number to zero, but it does not capture, for instance, the reduction from four visits to one. To address this issue, we ran a two-stage model with the number of restaurant visits as the dependent variable for the second stage while controlling for participation in the first stage. The conditional marginal effect of the food-scare crises on the number of visits was not significant, implying that the binary participation variable has already captured the effect of interest, and therefore, the aggregation of data is highly unlikely to perturb the implications on restaurant visit.    61  similar to that of the Canadian population,32 but the Nielsen sample is slightly more educated than the population average. While the average age of our sample is approximately equal to the national average, most of our sample is between 45 and 55 years of age, thus accounting for the higher percentage of married individuals in our sample compared to the Statistics Canada percentage over a younger population. Comparison of statistics indicates that our sample is representative of the Canadian adult population.  Next we explain the categorization process of restaurants; these categories are used to capture the heterogeneity of a food scare’s effect across restaurants. We use two categorizations for restaurants in our analysis. First, we categorize restaurants based on their beef or chicken orientation. The food specialty of restaurants is explicitly coded in the NPD Group’s CREST dataset. We categorize restaurants with “hamburger,” “steak,” and “beef and roast beef”33 food specialties into a beef-oriented restaurant set (constituting 19.7% of visits). Similarly, restaurants coded with “chicken” specialty food make up the chicken-oriented restaurant set (constituting 5.22% of visits). Second, we categorize beef-oriented restaurants in terms of their brand status, national chain or local. We define national chain restaurants as those that are located in all ten Canadian provinces and have at least 3,000 meals served in the dataset. These conditions filter out all beef-oriented restaurants except for McDonald’s, Burger King, Wendy’s, A&W, Harvey’s, and Arby’s—a set labeled as national chain restaurants (constituting 16.36% of visits).                                                 32 Note that marital status data obtained from Statistics Canada is reported for all ages. Once we filter for 45- to 55-year-olds (the average age range in our sample), the percentage of married people goes up to 59%, matching our sample percentage.  33 In addition to the “food specialty code” variable discussed above, the “restaurant code” variable also classifies some of the restaurants as beef and chicken places. Specifically, we categorize restaurant codes “A/O beef/steak” and “A/O hamburger” (A/O chicken) as beef (chicken)-oriented places too.  62  The rest of the beef-oriented restaurants are classified as a local restaurant set (constituting 3.33% of visits).   3.2.2 Measure of Media Coverage of the BSE Outbreaks    Next we operationalize media coverage intensity using the frequency of local press articles containing food scare-related keywords in each province. Even though newspapers are not the only source of exposure to the news, we use their content as proxy for the relative media and public attention shifts.34 We used the Canadian Newsstand Complete and Factiva databases to search for the following keywords in the title, abstract, or subject of news items: “mad cow disease,” “BSE,” and “bovine spongiform encephalopathy.” Furthermore, to address the bilingualism in Canada, we calculated the frequency of the keywords translated to French appearing in French newspapers. Consequently, we generated ten provincial English media indices and one French media index. However, we note that more populated provinces publish more newspapers, and hence they have higher numbers of articles in the database.35 So one unit of article in a larger province (e.g. British Columbia) is expected to have a smaller share of effect on the shift of public attention compared to a province with less press (e.g. Nova Scotia). Therefore, the absolute number of articles is not a clean indicator of public attention shift in cross-province analysis. For that reason, we normalize the provincial media variable by dividing                                                 34 Moreover, we use online resources such as the Canadian Food Inspection Agency and United States Department of Agriculture websites to find the dates for the food scare's major incidents, and to ensure consistency between those data and the constructed media variables.  35 For the British Columbia province, for example, we included the Vancouver Sun, The Province, Victoria Times Colonist, Nanaimo Daily News, Daily News (from Prince Rupert), and Abbotsford Times in generating the province-specific media index. In Alberta, on the other hand, we included the Calgary Herald and Edmonton Journal newspapers.  63  the monthly food-scare article counts by the total number of food-scare articles in that province for the whole six-year period of our analysis (and we multiply this ratio by 100). That is, we measure the articles on BSE in each month for each province as a percentage of the total number of articles on BSE between 2000 and 2006 for the same province. In Table 3-3, we provide average and maximum numbers for the normalized BSE media variables and pure article counts indices. As a robustness check, we estimated our model using the article count media variable (not normalized) and media variable normalized by the number of newspapers covered in each province (see Appendix G). We found that the parameter estimates are qualitatively similar to the estimations based on the media index used in the paper. This gives us further confidence in use of the normalized media variable.   Once media indices were generated, we matched each person to the media index of her province. She was assigned to the French news index if her preferred language was French in the dataset. Figure 3-1 exhibits the longitudinal variation of average BSE media variables over the 11 media indices. We observe the peaks in the media attention to the crises after the major incidents of mad cow disease in May 2003 and January 2004. Details on the process and a complete list of the covered newspapers are in Appendix B.   3.2.3 Local Weather Information   We collect information on local weather because it helps us explain consumers’ decision to dine out. We constructed five local weather variables using the historical weather data  64  recorded in the Government of Canada (Environment Canada) database.36 We extracted average temperature, extreme maximum and minimum temperature, total rain, and total snow for each location in each month. The location-identifying variables in the NPD Group’s CREST data are the province and city size, a categorical variable reflecting a city’s population. We constructed a matrix with ten provinces as the rows and six levels of city size as the columns. We allocated 140 Canadian cities with more than 10,000 population to the matrix cell matching their population and province. Thirty-seven out of 60 province–city-size combinations had at least one city allocated to them.37 In each province–city-size cell with multiple cities, we recorded the weather information for up to the three most populated cities with complete weather data between 2000 and 2006. For major cities with multiple weather stations, we used the data from the airport weather station. Table 3-4 provides summary statistics for the weather variables (with unit of observation as city size–province) used in the model. The cities matrix and details of this process are exhibited in Appendix C.   3.3 Empirical Strategy  Food scares can influence a consumer’s probability of participation (i.e., to dine out) in the FAFH market (first/participation stage) and can change the food-order patterns conditioned on participation in the market (second/consumption stage). The objective of our analysis is to investigate the changes in consumers’ dine-out and food-order decisions due to BSE media                                                 36 http://climate.weather.gc.ca/ is an open governmental database with comprehensive data on Canada’s weather. 37 Seven cities were uniquely identified by the province variable (with ten categories) and city-size variable (with five categories). For example, the only city in BC with more than 1 million population is Vancouver, but there are two cities, Kelowna and Abbotsford, with population between 100K and 250K.   65  discussion. Note that making inferences about each stage of the decision has strategic managerial implications. For example, if the reduction in beef order is due only to lower dining-out incidents, then the crises would affect the restaurants’ traffic; and the restaurants’ marketing strategy should primarily focus on bringing customers in. However, if we observe a lower level of beef orders among participating consumers, then managers should focus on providing non-beef options that would be attractive to consumers who would otherwise order beef. We discuss these points further in the final section.   The two decisions, participation and consumption, are interrelated, because the first-stage decision (to dine out) is affected by the questionability of the desired food’s safety in the second stage. In other words, the probability for participating in the FAFH market depends on the expected desirability of the food consumed in the second stage. We model this interdependence of the two decisions by allowing for common unobserved shocks to both stages. Empirically, this translates to correlated error structure for the two stages of estimation, and it can be modeled using double-hurdle methods with correlated error terms. As we will explain, however, none of the existing methods fully match our data-generating process in which we observe the participation in the (FAFH) market for any level of (beef) consumption, including zero. Accordingly, we developed a modified version of the double-hurdle method—what we call the “Double Hurdle with Explicit Participation model” (DHEP)—to estimate the parameters of interest. This method is more appropriate than the existing double-hurdle methods for the case when the participation decision is observed and can be incorporated in the estimation.    66  In this section, we first discuss our econometric model and its functional specification and identification. We then compare our proposed model estimates to benchmark models and finally discuss and rule out some confounding factors.  3.3.1 Econometric Modeling: Double Hurdle with Explicit Participation Model  We develop an econometric model, Double Hurdle with Explicit Participation (DHEP), to explain the probability of dining out (participation) and number of meal orders conditioned on dining out (consumption). This two-stage model simultaneously estimates the parameters of both stages and the errors structure. In addition, the model allows for corner solutions (zero values) of consumption conditional on participation in the market. Therefore, our model is categorized as a double-hurdle approach, yet it is modified to incorporate two explicit decision states—decision to dine out and conditioned on dining out decision to eat beef. This modification results in new formulations for the likelihood function that require a new estimation procedure using likelihood maximization. Details of the likelihood function derivation and likelihood formulations for other types of two-stage models are presented in Appendix D.   We now explain our double-hurdle method and clarify the proposed modification at the end of this subsection. In double-hurdle models, the decision maker has to pass two “hurdles” before being observed with a positive level of consumption ( ). The first hurdle models the decision to participate in the market. Conditioned on passing the first hurdle, each individual decides on the level of consumption in the second hurdle. Participation and consumption decisions are normally driven by partially overlapping behavioral forces. For example, in the y 67  case of cigarette smoking, people do not purchase cigarettes either because they are non-smokers (not passing the first hurdle) or because they happen not to need cigarettes in that period (not passing the second hurdle) (Jones, 1989). The binary participation decision ( ) for each individual in each occasion is governed by the latent variable , which is modeled as a linear function of vector of covariates ( ) and an unobserved term ( ). In particular, we define (subscripts are dropped for simpler exposition), 1 if  0 0 if  0 where  wdww Zα ν>⎧= ⎨ ≤⎩= + 3-1 and for the consumption decision we define, **if   00 if   0.X u X uyX uy d yβ ββ+ + ≥⎧= ⎨ + <⎩=         3-2 where  is a latent variable that is a non-negative linear function of a set of consumption regressor ( ) and an unobserved term ( ). The consumption variable ( ) conditions the non-zero consumption values on participation ( ) in the market. Following conventions in double-hurdle models (Jones, 1992), the unobserved terms are allowed to have a joint bi-variate normal mean zero distribution with correlation coefficient and variance parameters  and , respectively, that are estimated along with the coefficient vectors of the two stage decisions () through likelihood maximization: ( ) ( )2, 0,1where  u Nνσρσρ σΣ⎡ ⎤Σ = ⎢ ⎥⎣ ⎦:        3-3  dwZ !y*X u y1d =! ! 2! ," 68  Unlike the existing double-hurdle procedures, our methodology incorporates the information on participation in the estimation. In our case, the first hurdle (dining out or not) is an observed key decision, and hence, its information should be incorporated in the estimation process. This allows our proposed procedure to distinguish between two mechanisms that result in zero beef consumption—low interest to eat beef in the period and/or not participating in the market (not dining out). However, based on an extensive literature search, we determined that existing double-hurdle methods are designed for datasets in which we do not observe the participation decision, and therefore, we cannot empirically distinguish between the mentioned two mechanisms that generate zero values for consumption. These methods posit that a positive consumption is observed only if both hurdles are passed, and zero consumption level is due to not passing one or both of the hurdles, but we do not know which hurdle is not passed. That is, in many applications, the first hurdle (participation) is implicit and explains the behavior of different latent types of decision makers. For example, Maynard et al. (2008) model beef orders in restaurants and note that vegetarians may never order beef (resulting in zero orders due to not participating in the market) while others happen to choose a zero quantity. Thus, consumers should pass both a not-being-a-vegetarian hurdle and a hurdle of having a high utility for beef at the current period to be observed with a positive beef order. Note that vegetarian status is not observed in the dataset and is implicit in their modeling process.   Hence, the proposed DHEP method better matches the underlying data-generating process in our context. When participation decision is observed, DHEP is also potentially more efficient than the existing double-hurdle models since it exploits the participation as well as  69  consumption information to estimate parameters. As noted above, a complete development of this model and its likelihood function formulation is presented in Appendix D.   3.3.2 Predictor Variables and Identification  Next we describe the independent variables included in the and vectors, explaining the dine-out and food-order decisions, respectively. The key independent variable in the participation decision is the BSE media index variable38 ( ), and the included covariates are age, having children, marital status, having a university degree, income, region, linear trend, city size, and local weather information—including maximum, minimum, and mean temperatures (in Centigrade), total rain (mm), and total snow (cm) in a month. Specifically,    3-4  In the second stage, we model the conditional food-order decision as a function of BSE media coverage ( ), linear trend, and the demographics variables including age, having children, marital status, income, having a university degree, and region.   3-5                                                 38 Various studies (Maynard & Wang, 2010; Peng, McCann-Hiltz, & Goddard, 2004; Pritchett, Johnson, Thilmany, & Hahn, 2007) have established the link between the public reactions to the crises and the media coverage. Z Xp[ ]: BSE media :Weather variables {mean/max/min temp, total rain/snow}, city size: Demographics {age, child, married, income, education, region}, linear trendZ p W DpWD=p[ ]: BSE media : Demographics {age, child, married, income, education, region}, linear trendX p DpD= 70   Age, having children, and income level (Soberon-Ferrer & Dardis, 1991; Yen, 1993), education (Michael & Becker, 1973), and region (Jensen & Yen, 1996) variables are known to affect the decision to dine out and to order meals, and they are generally included in previous studies. Moreover, based on a collection of surveys on risk-related attitudes (Muringai & Goddard, 2011), the demographic variables included in the present study are among the factors influencing confidence in food products including meat. Weather condition is expected to affect the utility of dining out by changing the desirability of engaging in outdoor activities or shifting transportation costs. City size links to the geographical density of restaurants and their accessibility for customers, thus affecting the dine-out decision. Note that weather variables and city size are excluded from the vector of covariates in the food-order stage—a key notion for identification, discussed next.  It is well-established in the literature (Jones, 1992; Wooldridge, 2010) that we require defendable exclusion restrictions to improve identification of two-stage models such as double hurdle and Heckman (Heckman, 1979). Otherwise, the identification of the coefficient is based only on the functional form. We propose that while the weather variables help us explain the decision to dine out, they can be excluded from the consumption stage. This is because while weather condition affects the decision to dine out by shifting the transportation cost and availability of other outdoor activities, it is unlikely to shift food orders once a consumer is at the restaurant. For example, snowy weather may affect the quality of roads, making it less attractive to dine out at a restaurant, but it should not shift preference for beef once customers make it to the restaurant. Also, food orders are not driven by weather variables in our sample prior to  71  incidents of BSE.39  The city-size variable is also excluded from the meal-order decision. Size of city affects restaurant density and accessibility, and thus, the decision to dine out. On the other hand, we argue that variation in this variable does not change the preference for different food items. We do not have any reason or evidence to believe that adults from larger cities have a greater preference for beef, for example.  3.3.3 Model Comparison and Discussion  In our modeling procedure, as discussed above, we are interested in disentangling the effects of food-scare news on participation and consumption decisions while allowing for some common unobservable factors affecting both decisions. In this part, we compare our proposed model estimates to those of four other estimation procedures and find that there is an overall consistency across procedures. Further, we provide extra support for identification of the food-scare effect by ruling out two alternative explanations: the supply-side reaction and concurrent macroeconomic trends at the time of BSE outbreaks.  We estimate four models along with our proposed methodology. First, we ignore the participation decision and estimate an OLS model of beef consumption. Second, we consider a two-part estimation procedure that models the participation decision by probit and consumption decision by an OLS on participating subsample. It assumes that the two decisions and therefore                                                 39 In a supplementary analysis presented in Appendix G, we find that food orders prior to BSE incidents are not driven by the weather variables included in the model. We regress the food orders on weather variables in the participating sample and, controlling for the common covariates, the coefficient for weather variables is insignificant.   72  their error terms in the underlying latent utilities for dine out and beef consumption are independent. Third, we examine the Heckman selection model where zero consumption does not arise from a standard corner solution but instead represents a separate discrete choice (not participating). This method uses only the first hurdle to determine the participation decision. Next we estimate a double-hurdle (DH) model with correlated error terms in which positive consumption is observed only when both hurdles are passed, but the participation decision variable is not incorporated. Finally our DHEP model allows for correlated error terms and zero levels of consumption conditional on participation. It also incorporates the participation variable in estimation. The following condensed formulations show the conceptual set-up for these five models: { } { }{ }{ }{ }if  0  & 0 &0 otherwiseif  00 otherwiseif  0 & 00 otherwiseif  0 & 00 if  0 & 00 if  0where OLSTwo partHeckmanDHDHEPy X uX u X u w uyX u wyX u X u wyX u X u wy X u www Zββ β νββ ββ ββα ν−= +⎧ + + ≥ > ⊥= ⎨⎩+ >⎧= ⎨⎩⎧ + + > >= ⎨⎩⎧ + + > >⎪= + ≤ >⎨⎪ ≤⎩= +   3-6    Parameter estimates for the five models are shown in Table 3-5. Across the tested models, we find that the key results for the effect of the BSE media index are largely consistent. In terms of dining out, the BSE media index is marginally significant (except for the DH model). In terms of number of beef orders, the BSE media index has a significant (p < .05) effect for the DHEP, an effect that is directionally consistent with the results for the other models. The results  73  for the other coefficients are also directionally equivalent across the models. The correlation coefficient between the error terms of the two stages is significant, supporting the use of two-stage models.   Next we discuss two confounding factors that may bias our estimation results: supply-side reaction and concurrent macroeconomic trends. First, the strategic reaction of supply side is a potential source of endogeneity whenever we model the demand in isolation. In the specific case of food scare crises, restaurants may strategically react by lowering prices and offering promotions. In a supplementary analysis, documented in Appendix E, we investigated the relationship between BSE media and two measures of supply reaction: beef price index and promotional efforts indices (i.e., coupons received and special offers taken by panelists). We do not find any evidence for supply-side reaction at the time of crises. Beef prices and the level of promotions in our sample did not change systematically in reaction to the BSE outbreaks. Additionally, anecdotal evidence based on interviews with two restaurant owners suggests that the high cost of menu changes and the risk of sending negative signals to customers hinders restaurants from changing prices in short time windows and upon occasions such as a BSE outbreak.  Second, the issue of endogeneity can potentially arise because of concurrent macroeconomic factors, seasonality, or secular trends in consumption such as gradual decrease of beef consumption. Even though one cannot fully rule out all such factors, we took several steps to alleviate the concerns. First, we included a linear trend variable in all the models to control for any secular upward or downward linear trend in consumption or participation.  74  Second, we conducted an analysis to measure the change of participation for two distinct sets of restaurants, beef-oriented restaurants and chicken-oriented restaurants. If the dine-out decision is driven by macroeconomic factors, both sets of restaurants should be affected similarly, but our evidence suggests otherwise.40 Table 3-6 reports the first-stage estimations with the DHEP method for beef- and chicken-oriented sets of restaurants. We observe that beef-oriented restaurants are significantly less likely to be visited at the time of a BSE outbreak (p < .05); interestingly, however, chicken-oriented restaurants are more likely to be visited (p < .05) at the time of BSE crises. This finding supports our proposition regarding dine-out decisions being influenced by BSE crises and not the macroeconomic trends affecting all restaurants.   Third, the weather variables proxy for seasonal variations in dine-out decision, and their use rules out seasonality explanations. Finally, the continuous (rather than step function) and multimodal nature of our media index provides a more reliable platform, compared to before-after dummies for BSE outbreaks, for separating the effect of BSE crises from potential macroeconomic shocks, because it is unlikely that a macroeconomic factor’s trend mimics the multimodal variations of the BSE media index.   In summary, a range of empirical tests and arguments confirms the robustness of our findings to supply-side reaction and concurrent macroeconomic trends. Also, consistency of the estimated effect across multiple methods further enhances our confidence in the estimation                                                 40 Note that the estimation included a linear trend that captures the secular changes in the beef and chicken consumption.  75  procedure. Next we present the findings for different meal orders and national chain versus local restaurants.  3.4 Findings  We present our findings in this section based on the DHEP estimation. To provide a fuller understanding of consumers’ behavior in response to BSE crises, we estimate the effect of the crises on monthly number of orders of beef, chicken, seafood, pork, and vegetarian items conditional on dining out, along with the probability of dining out in restaurants. Furthermore, we measure the extent of consumer reaction to BSE crises in national chain and local beef-oriented restaurants to compare the performance of different types of restaurants in food-scare outbreaks.  3.4.1 Dine-Out and Meal-Order Analysis  Looking at the top part of Table 3-7, we see that dine-out estimates are essentially similar across different food orders since they are primarily driven by variations in the same dependent variable, dine-out decision. The slight differences between coefficients are due to correlated error structure in the joint likelihood estimation. We pick the beef-order estimate, which is of primary interest in our context, to explain consumers’ dine-out decision. As in Table 3-5, we observe a marginal negative effect for BSE crises news index on the probability of dining out (p < .1). Younger, more affluent, more educated adults without children are more likely to dine out. Additionally, the excluded factors partially explain the decision to dine out. The colder the  76  lowest temperature and the higher the highest temperature in a month, the less likely people are to visit restaurants. Adults in smaller cities tend to dine out more often.   Table 3-7 also presents the DHEP parameter estimates for beef, chicken, seafood, pork, and vegetarian item orders conditional on dining out. BSE crises have decreased the number of beef orders (p < .05) and increased seafood orders and pork orders (p < .05), but have not affected chicken or vegetarian orders. This provides evidence for substitution of beef with seafood and pork. Among demographic factors, having a child and being married (a higher income) decreases (increases) number of beef orders per month. Interestingly, holding a university degree works only in favour of ordering vegetarian items, and aging shifts consumer orders from beef and chicken to other protein and vegetarian orders.  We can further clarify the big picture of findings by looking at the presented findings beside the reported effect for beef-oriented and chicken-oriented restaurants in the previous section. At the times when BSE was in the news, consumers dined out marginally fewer times across all restaurants, but they significantly reduced their visits to beef-oriented restaurants, perhaps by switching to chicken-oriented restaurants. Once in the restaurant, they reduced their beef orders, and more strongly so in beef-oriented restaurants. Next we focus on beef-oriented restaurants, where we see the strongest reaction to the crises, to explore the effect of brand equity on consumer behavior at the time of food-scare crises.    77  3.4.2 National Chain versus Local Restaurants  In addition to our investigation of the entire FAFH market, we study consumers’ behavior in national chain versus local restaurants. National chain restaurants invest substantial sums of money in developing strong brand images that are designed to signal consistent quality and reliability across all locations. On the other hand, the prominence of national chains sometimes makes them a focal point for social issues, including health scares. Moreover, anecdotal evidence indicates that major brands were concerned about a possible sales drop after the first outbreak of Canadian BSE (Day, 2003). By contrast, local restaurants have lower brand equity, and they typically focus on serving people in their neighborhoods and building a more organic relationship with customers. Therefore, it is interesting to empirically compare the performance of national chain to that of local restaurants in the face of an industry-level crisis, as this can provide some evidence for the level of customers’ trust in different restaurants, which is linked to loyalty (Hyun, 2010),41 and further hone our managerial implications.  The estimation procedure of the restaurant-level analysis is identical to the main analysis except that we replace the dine-out and food-order variables by dependent variables defined for each restaurant set. For example, in the national chain restaurants analysis, we define the dependent variables based on visits and food orders in any of the national chain restaurants:                                                 41 In his extensive review of customer relationship quality literature in the restaurant context, (Hyun, 2010) links loyalty to satisfaction and trust (e.g., Crosby, Evans, & Cowles, 1990). Trust is defined as “one party’s belief that its need will be fulfilled in the future by actions undertaken by the other party” (Anderson & Weitz, 1989). Hence, we expect trust to be the dimension with the primary influence on consumer reaction at the time of food safety crises, since they should believe that their future food safety is provided by the trusted restaurant.    78  McDonald’s, Burger King, Wendy’s, A&W, Harvey’s, and Arby’s. Therefore, our procedure involves estimation of separate DHEP models for national chain and local restaurant sets defined in the data section.  Table 3-8 reports the DHEP model estimates for beef orders in national chain and local sets of restaurants both of which are a subset of beef-oriented restaurants.   The results for local and national restaurants are strikingly different. Based on the BSE media coefficient estimates, national chain restaurants are negatively affected (p-value < .05) both in terms of visit probability and conditional beef orders, but we do not find any significant BSE effect for local restaurants. Moreover, the magnitude of the coefficients is larger than those reported for all beef-oriented restaurants and all restaurants (Table 3-8). This interesting finding suggests that national chain restaurants are vulnerable to food-scare crises, whereas local restaurants seem to be largely unaffected by these crises.42 Even though surprising, the reported effect is in line with findings of Liu and Shankar (2015) in the context of automobile recalls. They found that brand preference and demand decline by a greater amount when the perceived quality of the recalled brand is higher.                                                    42 The challenge in comparing national chain versus local restaurants is finding a set of local restaurants matching the national chain restaurants. Our national chain restaurant selection criteria, choosing the most frequently visited restaurants present in ten provinces, has left us with fast-food restaurants only; therefore, we tried to control for fast-food status in a supplementary analysis documented in Appendix G. Hence, we formed a new set of local restaurants that are coded as fast-food, and repeat the DHEP analysis. Our findings remain the same: no effect for the local fast-food restaurants.   79  3.4.3 Magnitude of BSE Effect  In this section, we estimate the magnitude of the BSE effect and project it to the industry-level scale. Using our model parameter estimates, we calculate predicted percentage change in the number of food orders/dine out probability had there been no crises (BSE Media=0). In calculating the marginal effect on food order we incorporate the indirect effect of the food scares through the inverse Mills ratio term (detail shown in Appendix F). We implement these calculations for the two peak months of crises, May 2003 and January 2004, to find the counterfactual level of dine out probability and beef orders—Figure 3-2 juxtaposes these numbers with the actual levels. As reported in Table 3-9 and illustrated in Figure 3-3, the conditional order of beef would have been 1.7% and 1.4% higher if there were no crises in May 2003 and January 2004, respectively; and similarly, dine-out probability would have been 1.2% and .9% higher, respectively. As shown in the table, the effects are stronger for national chain restaurants: participation likelihood and beef orders would have been 3.4% and 1.7% higher in May 2003 if there had been no crises in the peak months.  The estimated percentage changes translate into sizable numbers once they are projected to industry scales. The wide coverage of more than 1,250 restaurants in this study provides a representative sample of the Canadian restaurant market; so we can cautiously project the findings to the national level. According to Statistics Canada,43 the estimated total Canada-wide dollar value of FAFH receipts for May 2003 is $3.1 billion. Therefore, ceteris paribus, a 1.2%                                                 43 Table 355-0001 of CANSIM data on Restaurant, Caterer and Tavern Statistics  80  increase in going out translates to a roughly $37 million sales increase across all Canadian restaurants under the counterfactual scenario of no food crisis in that month.  3.5 Discussion  Food scares and more generally product-category crises frequently affect markets. In this work, we investigated the effect of food scares, specifically mad cow disease (BSE), on consumer decisions in the food-away-from-home (FAFH) market—an understudied sector that absorbs 27% of Canadian households’ expenditure on food, with important managerial and policy implications. Overall, we found that media coverage of BSE outbreaks marginally decreased the likelihood of dining out in the FAFH market and significantly reduced ordering of beef products. Once we focus on beef-oriented restaurants, both effects amplify and become significant. We also find evidence for increase in dining at chicken-oriented restaurants, and substitution of beef by pork and seafood. Moreover, the negative effects for both dining out and food choice were significant for beef-oriented national chain restaurants, but not significant for beef-oriented locally owned restaurants. As we discuss below, management of national chain restaurants needs to be particularly vigilant in monitoring food scares and forceful in designing quality assurance systems in order to minimize the impact on their sales and their relationships with customers. For consumers who still choose to dine out, alternative food choices should be provided.  We base our results on a detailed analysis of multiple outbreaks of BSE in Canada in the years 2000–2006 using individual-level NPD Group’s CREST panel data. We operationalize the  81  perceived intensity of the crises by counting the local press articles on the topic in each month for each province. To conduct our analysis and to take advantage of the data on dining-out decision (participation in market), we developed a two-stage double-hurdle model—Double Hurdle with Explicit Participation. Our model, while providing results consistent with other approaches, matches our data structure and allows us to make potentially more efficient estimates.   From the public policy perspective, it is important to recognize that consumers’ reaction is consistent with a model of risk reduction. That is, broadly speaking, consumers dine out less, choose beef less often when they do dine out, and substitute other foods for beef when there is a concern for its safety. To the extent that the media continue to report incidents of BSE and give it relatively high coverage, policy makers can expect people to respond appropriately. However, as product harm crises become more commonplace, media coverage may decay or consumers may be less likely to respond to repeated crises for which they do not experience such ill effects. Our paper examines multiple cases of BSE in Canada where the later crisis received less news coverage. Exploring such dynamics are beyond the scope of the present research but are worthy of future study.   Beyond the immediate health and safety of their customers, management of restaurants need to be concerned about the incidents of food crises since our study shows that impact of the BSE crisis was to decrease the number of people entering the beef-oriented stores and to change their food orders away from beef. The situation can be more complicated for the restaurant industry as compared to the packaged-food industry. The packaged-food provider can replace its  82  infected products by shipping new items to the retail store and publicizing these efforts, and consumers are unlikely to skip the retail market altogether at the time of crises; however, consumers may stop going to some restaurants, or a segment of them, at the time of crises. Thus, as the crisis itself is outside the control of the restaurants, they can take short-term and long-term actions to limit the effects of crises on their sales. Restaurants need to have systems in place to monitor food-scare threats and be prepared to react. In an international setting, the company may be able to obtain early indicators from areas outside its main market. In the face of a crisis, managers need to provide strategies to continue to induce people to dine out, assuming it is safe to do so, by focusing on promoting other foods that consumers are predisposed to order. However, this is a challenging strategy for single-product restaurants like KFC. Anecdotal evidence suggests that such restaurants tend to be hit strongly by food crises.44   Our results also suggest that managers of national chain restaurants should be particularly sensitive to consumers’ prior expectations and their potential distrust of safety procedures. The focal event of large-scale food-scare crises (e.g. discovery of BSE) is external to all restaurants and not under their control, so the crises result in similar levels of publicity and blame attribution for all restaurants. Therefore, any variation in reaction to different sets of restaurants should probably be attributed to relationship of customers to restaurants and consumer level of trust in different restaurants. For example, it is worthwhile to empirically investigate the influence of service quality on consumers’ trust, since Hyun (2010) found that among key restaurant                                                 44 Yum! Brands Inc., owner of KFC and Pizza Hut, sales fell 29% in China in April 2013 due to concern about the safety of its chicken and the spread of avian flu.  83  attributes (including food quality, price, service, location, and clean environment), service was the factor with the strongest effect on consumer trust. Moreover, consumers may have different prior expectations—defined as their belief about the behavior of a firm in a given situation based on their experience and knowledge (Dawar & Pillutla, 2000)—about brand versus local restaurants, which would be another potential area of study. In addition, it would be interesting to note whether a focus on local sourcing of products would have beneficial effects on consumers’ perception of the store’s safety during a food-scare crisis.   In considering the generalization of our results for managers, we should note that BSE is a dreadful scare, but one with a very low incidence level in the population and no threat of contagion. In apparently contagious disease crises, such as avian influenza and swine influenza, the impact may be considerably greater on the likelihood of dining out. In a supplementary analysis (Appendix G), we run DHEP analysis on avian influenza cases occurring between 2000 and 2006 in Canada. Interestingly, we find marginal effect for the decision to dine out, but no effect on food orders. Focusing on chicken-oriented restaurants, we find significant reduction in dining out at the time of avian influenza yet still no effect on chicken orders conditioned on dining out. This highlights the influence of the nature of crises, e.g. being contagious or not, on consumer reaction and response strategies. In the case of contagious disease, therefore, different managerial strategies need to be considered first to heighten the cleanliness level of restaurants and employees and to ensure that customers are accurately informed about the hygiene of the restaurant and are encouraged to visit. This is another potentially interesting addition to this work.   84  One other interesting extension would be to develop a discrete choice model of selecting restaurants and food items. This approach requires a dataset that provides clear information on the choice set of customers, a feature lacking for our dataset. Finally, even though a minor factor at the time of our study, nowadays social media may serve to magnify the crises, and it might be a useful gauge for measuring public sentiment and reaction.    85   3.6 Tables and Figures  Table 3-1. Average of Response Variables  Across All Observations Cond. on Participation in Market Food Orders per Month Average Std. Dev. Average Std. Dev. Beef .15 .54 .52 .91 Chicken .13 .48 .46 .79 Seafood .06 .31 .20 .55 Pork .09 .42 .29 .74 Veggie .16 .56 .53 .94 % average participating in FAFH in a month   29%        Table 3-2. Summary Demographic Statistics for Sample and for Canadian Household Heads  Sample     Canadian Population Variable Mean  Age  48.8 4745 Income (in CAD$): 63,700 68,50046  Percentage Married (Y/N) 57.2 4647 Household with child (Y/N) 36.5 3548 University degree (Y/N) 29 2149                                                    45 Source: Statistics Canada, 2011 Census of Population, Statistics Canada Catalogue no. 98-312-XCB2011028. 46 Source: Statistics Canada, Table 202-0403 (year 2003). Average total income, by economic family type, annual (dollars), CANSIM (accessed: 2015-05-20). 47 Source: Statistics Canada, 2011 Census of Population, Statistics Canada Catalogue no. 98-312-XCB2011042.  48 Source: Statistics Canada, 2011 Census of Population, Statistics Canada Catalogue no. 98-312-XCB2011032. 49 Source: Population 15 years and over by highest degree, certificate, or diploma (1986 to 2006 Census), http://www.statcan.gc.ca/tables-tableaux/sum-som/l01/cst01/educ42-eng.htm  86   Table 3-3. Summary Statistics for BSE Media Index   Article count Normalized value  Mean Max.  Mean Max Average Canada 97.38 715 1.25 7.13  British Columbia 18.65 164 1.25 10.99 Alberta 20.97 189 1.25 11.26 Saskatchewan  12.96 89 1.25 8.58 Manitoba 0.25 8 1.25 10 Ontario 20.38 145 1.25 8.89 Quebec 3.88 27 1.25 6.68 New Brunswick 2.57 16 1.25 7.76 Prince Edward Island 5.26 50 1.25 11.87 Nova Scotia 2.18 12 1.25 6.85 Newfoundland 3.8 30 1.25 9.86 French Press Index  6.45 49 1.25 9.49 Note: Mean and maximum of the normalized media index and article count. Since the total number of articles is normalized to 100, the overall averages for all indices will be 100 divided by 80 periods (1.25). The maximums for the normalized media index are the highest percentage of BSE articles in a month per index.     87   Figure 3-1. Canada-Wide (normalized) BSE Media Index by Month  Note: The first North American discovery of BSE occurred in Alberta, Canada, in May 2003. The second BSE case was reported in Canada in January of 2005 (Jan 2, 2005, and Jan 11, 2005), with additional outbreaks in April, July, August 2006 and January 2007 (Maynard & Wang, 2010).    Table 3-4. Summary Statistics for Weather Variables Monthly Local Weather Variable Mean Std. dev. Mean Temperature (Cº) 6.55 10.39 Max Temperature (Cº) 21.42 9.81 Min Temperature (Cº) -7.37 12.76 Total Rain (mm) 56.27 46.21 Total Snow (cm) 12.79 19.64     88   Table 3-5. Models Comparison for Number of Beef Orders and Dine Out Probability  Overall OLS Two-part Heckman DH DHEP           Dine Out BSE Media  -0.0009* -0.0009* -0.0001 -0.0009* HH With Child  -0.0420*** -0.0420*** -0.0658*** -0.0425*** Uni. Degree  0.0203*** 0.0203*** -0.0069 0.0202*** HH Income  0.0107*** 0.0107*** 0.0107*** 0.0106*** Age  -0.0027*** -0.0027*** -0.0071*** -0.0028*** Married  0.0069 0.0069 0.0516*** 0.0063 Ontario  -0.0038 -0.0038 -0.1101*** -0.0046 Quebec  -0.0417*** -0.0417*** -0.2309*** -0.0429*** Atlantic Provinces  -0.0255*** -0.0255*** -0.1318*** -0.0260*** City Size  -0.0054*** -0.0054*** 0.00076 -0.0049*** Mean Temperature  -0.0019 -0.0019 -0.0076*** -0.0017 Max Temperature  -0.0020*** -0.0020*** -0.0023** -0.0023*** Min Temperature  0.0029*** 0.0029*** 0.0084*** 0.0029*** Total Rainfall  -0.0001 -0.0001 0.0002** -0.0000 Total Snowfall  0.0002 0.0002 0.00037 0.0001 Linear Trend  0.0003** 0.0003** -0.0008*** 0.0003** Constant  -0.4919*** -0.4919*** -0.6170*** -0.4831***               Beef Order BSE Media -0.0004** -0.0003* -0.0013* -0.0006 -0.0035** HH With Child -0.0207*** -0.0512*** -0.0614*** -0.0941*** -0.1412*** Uni. Degree -0.0018 -0.0308*** -0.0115 -0.0103 -0.0355** HH Income 0.0026*** 0.0025** 0.0067*** 0.0818*** 0.0209*** Age -0.0019*** -0.0046*** -0.0058*** -0.0082*** -0.0142*** Married 0.0137*** 0.0314*** 0.0443*** 0.0729*** 0.1600*** Ontario -0.0343*** -0.1080*** -0.1211*** -0.1570*** -0.3344*** Quebec -0.0691*** -0.1698*** -0.2332*** -0.3288*** -0.6429*** Atlantic Provinces -0.0429*** -0.1330*** -0.1432*** -0.1879*** -0.3952*** City Size -0.0014**     Mean Temperature -0.0006     Max Temperature 0.0004 Excluded Excluded Excluded Excluded Min Temperature 0.0002     Total Rainfall -0.0000     Total Snowfall 0.0000     Linear Trend -0.0002*** -0.0006*** -0.0006*** -0.0010*** -0.0018*** Constant 0.3523*** 1.1285*** 0.5914** 0.1220 -0.4631** Lambda   0.5022** NA NA Rho   NA 0.71*** 0.54*** Sigma   NA 4.15*** 2.26*** Sample Size 380004 380004 380004 380004 380004 Log Likelihood  -177362.16 NA -277737.7 -352949.15 Note: * p < 0.1, ** p < 0.05, *** p < 0.01. The Heckman model is estimated with a two-stage procedure. Lambda is the coefficient of selection term (inverse Mills ratio) in the two-stage Heckman estimation method.      89   Table 3-6. DHEP Estimation for Beef- and Chicken-oriented Restaurants   Beef-oriented Restaurants Chicken-oriented Restaurants                                                    Dine Out BSE Media -0.0018*** 0.0023*** HH With Child 0.0226*** -0.0277*** Uni. Degree -0.0143** -0.0541*** HH Income -0.0045*** 0.0012 Age -0.0084*** -0.0014*** Married 0.0436*** 0.0148 Ontario -0.1128*** 0.2301*** Quebec -0.2559*** 0.3361*** Atlantic Provinces -0.0950*** 0.1674*** City Size -0.0113*** 0.0137*** Mean Temperature -0.0013 0.0069*** Max Temperature 0.0005 -0.0000 Min Temperature 0.0029*** -0.0068*** Total Rain Fall -0.0002** 0.0002** Total Snow Fall 0.0004** 0.0001 Linear Trend -0.0001 -0.0009*** Constant -0.5932*** -1.4777***                                                         Beef Order BSE Media -0.0043** 0.0129 HH With Child 0.0103 0.0941 Uni. Degree -0.0775*** 0.1333 HH Income 0.0045 0.0008 Age -0.0264*** 0.0143*** Married 0.1220*** 0.1733 Ontario -0.1726*** -0.4565* Quebec -0.6769*** 0.1120 Atlantic Provinces -0.1558*** -0.0862 Linear Trend -0.0009** -0.0013 Constant -0.7891*** -7.0615*** Rho 0.84*** 0.27 Sigma 2.12*** 2.7*** Sample Size 380004 380004 Log Likelihood -184919.58 -70056.19 Note: * p < 0.1, ** p < 0.05, *** p < 0.01       90  Table 3-7. DHEP Estimation for Meal Orders and Dine Out in All Restaurants  Dine-Out Estimations BSE Media -0.0009* -0.0009* -0.0009* -0.0010* -0.0008* HH With Child -0.0425*** -0.0422*** -0.0378*** -0.0398*** -0.0439*** Uni. Degree 0.0202*** 0.0205*** 0.0184*** 0.0213*** 0.0205*** HH Income 0.0106*** 0.0107*** 0.0092*** 0.0105*** 0.0108*** Age -0.0028*** -0.0027*** -0.0030*** -0.0028*** -0.0028*** Married 0.0063 0.0067 0.0103 0.0050 0.0040 Ontario -0.0046 -0.0042 -0.0061 0.0060 -0.0050 Quebec -0.0429*** -0.0421*** -0.0412*** -0.0338*** -0.0441*** Atlantic Provinces -0.0260*** -0.0265*** -0.0218*** -0.0141* -0.0277*** City Size -0.0049*** -0.0056*** -0.0002 -0.0077*** -0.0043*** Mean Temperature -0.0017 -0.0021 -0.0006 0.0012 -0.0030** Max Temperature -0.0023*** -0.0021*** -0.0003 -0.0004 -0.0019*** Min Temperature 0.0029*** 0.0030*** 0.0011*** -0.0006 0.0033*** Total Rain Fall -0.0000 -0.0001 -0.0001** -0.0000 0.0000 Total Snow Fall 0.0001 0.0002 -0.0000 -0.0001 0.0001 Linear Trend 0.0003** 0.0003** 0.0002** 0.0002** 0.0003** Constant -0.4831*** -0.4865*** -0.5487*** -0.5415*** -0.4919*** Meal Order beef chicken seafood pork vegetarian BSE Media -0.0035** 0.0015 0.0137** 0.0109*** -0.0006 HH With Child -0.1412*** -0.1318*** 0.0196 -0.4308*** -0.3810*** Uni. Degree -0.0355** -0.0945*** -0.1657*** -0.2193*** 0.1010*** HH Income 0.0209*** 0.0192*** -0.0208* -0.0370*** 0.0852*** Age -0.0142*** -0.0097*** 0.0522*** 0.0320*** 0.0048*** Married 0.1600*** -0.0294 -0.0504 0.2130*** -0.0913*** Ontario -0.3344*** 0.2112*** 0.1797*** -0.0942** 0.1311*** Quebec -0.6429*** 0.2101*** 0.3573*** 0.6429*** 0.3673*** Atlantic Provinces -0.3952*** 0.0453* 0.5697*** -0.1195* -0.0174 Linear Trend -0.0018*** 0.0001 -0.0014 0.0001 -0.0058*** Constant -0.4631** -0.7905*** 6.1137*** 4.0272*** -0.0735 Rho 0.55*** 0.18*** -0.9*** -0.9*** 0.56*** Sigma 2.27*** 1.88*** 12.02*** 7.97*** 2.37*** Sample Size 380004 380004 380004 380004 380004 Log Likelihood -352949 -346426 -294807 -314004 -353398 Note: * p < 0.1, ** p < 0.05, *** p < 0.01     91   Table 3-8. DHEP Estimation for Beef Orders and Dine Out in National Chain and Local Restaurants    National Chain Restaurants Local Restaurants Beef-oriented Rest. All Restaurants Dine Out BSE Media -0.0019*** -0.0003 -0.0018*** -0.0009* HH With Child 0.0536*** -0.2372*** 0.0226*** -0.0425*** Uni. Degree -0.0249*** 0.0556*** -0.0143** 0.0202*** HH Income -0.0102*** 0.0301*** -0.0045*** 0.0106*** Age -0.0092*** 0.0006*** -0.0084*** -0.0028*** Married 0.0411*** 0.0306*** 0.0436*** 0.0063 Ontario -0.1078*** -0.0225*** -0.1128*** -0.0046 Quebec -0.2811*** 0.0470*** -0.2559*** -0.0429*** Atlantic Provinces -0.0789*** -0.0818*** -0.0950*** -0.0260*** City Size -0.0113*** -0.0194*** -0.0113*** -0.0049*** Mean Temperature -0.0002 -0.0010 -0.0013 -0.0017 Max Temperature 0.0005 -0.0029*** 0.0005 -0.0023*** Min Temperature 0.0019** 0.0046*** 0.0029*** 0.0029*** Total Rain Fall -0.0002*** -0.0001* -0.0002** -0.0000 Total Snow Fall 0.0003 0.0000 0.0004** 0.0001 Linear Trend -0.0001 0.0002 -0.0001 0.0003** Constant -0.6178*** -1.2089*** -0.5932*** -0.4831*** Beef Order BSE Media -0.0053** 0.0011 -0.0043** -0.0035** HH With Child 0.0564** -0.0725 0.0103 -0.1412*** Uni. Degree -0.1062*** -0.0256 -0.0775*** -0.0355** HH Income -0.0078** 0.0373*** 0.0045 0.0209*** Age -0.0288*** 0.0062*** -0.0264*** -0.0142*** Married 0.0938*** 0.0976*** 0.1220*** 0.1600*** Ontario -0.1910*** -0.1209*** -0.1726*** -0.3344*** Quebec -0.6824*** -0.2671*** -0.6769*** -0.6429*** Atlantic Provinces -0.1312*** -0.2797*** -0.1558*** -0.3952*** Linear Trend -0.0010** -0.0005 -0.0009** -0.0018*** Constant -0.7369*** -1.4660*** -0.7891*** -0.4631** Rho 0.85*** 0.08 0.85*** 0.55*** Sigma 2.17*** 1.71*** 2.12*** 2.27*** Sample Size 380004 380004 380004 380004 Log Likelihood -165368.49 -189727.39 -184919.58 -352949 Note: * p < 0.1, ** p < 0.05, *** p < 0.01     92  Figure 3-2. Actual and Counterfactual Beef Order and Dine Out Probability  Note: Counterfactual dine out probability and number of beef ordered are measured by assigning zero (no crises) value to the BSE media in the mentioned months. See Appendix F for more detail. .288 .291 .294 .297.095 .098 .087 .0890.1.2.3All Restaurants National ChainsMay 2003 Jan 2004 May 2003 Jan 20041st Stage: Dine Out Probability2.14 2.18 2.16 2.192.79 2.83 2.78 2.820123All Restaurants National ChainsMay 2003 Jan 2004 May 2003 Jan 20042nd Stage: Conditional Beef Order CountActual Level Counterfactual LevelGraphs bytgeFor Two Peak Months of CrisesActual and Counterfactual Beef Order & Dine Out Prob. 93  Figure 3-3. Marginal Effect of Crises on Meal Orders and Dining Out in the Peak Months  Note: Reported numbers are % change in average beef order and probability of dine out were there no food scares in each of the months. See Appendix F for more detail.   Table 3-9. Marginal Effect of Crises on Meal Orders and Dining Out in the Peak Months  Marginal Change (%) in  All Restaurants All Rest National Chains Brand Rest  May 2003 Jan 2004 May 2003 Jan 2004 Average Dine-Out Prob. -1.2 -0.9 -3.4 -2.7  Std. Error .6 .5 1.2 1.0     Average Conditional Beef Order -1.7 -1.4 -1.7 -1.5  Std. Error .8 .7 .7 .6  Note: Reported numbers are % change in beef order and probability of dine out were there no food scares in each of the months. See Appendix F for more detail.          -1.16 -.897-3.36-2.71-4-3-2-10All Restaurants National ChainsMay 2003 Jan 2004 May 2003 Jan 20041st Stage: Dine Out-1.71-1.39-1.73 -1.55-2-1.5-1-.50All Restaurants National ChainsMay 2003 Jan 2004 May 2003 Jan 20042nd Stage: Conditional Beef OrderPercentage Change Due to CrisesGraphs bytgeFor Two Peak Months of CrisesMarginal Effect of Crises on Meal Orders and Dining Out 94  Chapter 4: Limitations and Concluding Remarks  As is the case with any empirical study, the current research does have limitations that must be noted.  The first essay’s data are limited to meals away from home. This restricts the scope of study to within-meal behaviors. Ideally, we would want to explore all eating occasions and also the substitution patterns between away-from-home and at-home meals. On a similar note, upon availability of further nutrition information, expanding the analysis to other restaurants could be of interest. For example, one may want to compare the decisions made at Subway Restaurants to those made at McDonald’s. Unfortunately, the size of the Subway sandwich is not coded in the NPD Group’s CREST dataset, and thus, we could not measure the calorie content of meals at this chain.  Even more broadly, it would be interesting to see if these results held in different types of restaurants, such as those providing table service.  Another limitation of the first essay is the exclusion of water and no-drink from the analysis. Switching to water and no-drink is conceptually related to a calorie cutting decision and is a natural extension of this work. Moreover, price implications when drinking water also need to be considered. People who choose tap water are typically having a free drink, whereas people buying bottled water may be paying more than those who choose CSDs. The unavailability of a clean meal-item price variable in the CREST dataset limits our ability to address this question.   95  In terms of study design, an external shock to consumption of diet drinks can provide a quasi-experimental framework that can further alleviate endogeneity concerns. For example, the introduction of Coke Zero to restaurants would be particularly interesting since marketing studies suggest it is primarily directed to males (Tungate, 2008). More generally, our data do not cover any such incidents, but this would be worthwhile to explore with new data sets.  In terms of limitations of the second essay, the food scare research project does not incorporate the choice of specific restaurant brands at the time of the food scare outbreak into the analysis. This would require implementing a discrete choice model and information on the choice set of restaurants for customers. This is a fruitful avenue for future research.  The reported findings on the difference in reaction to local versus national chain restaurants are worthy of further investigations as well. Upon availability of survey data, one could shed light on the underlying mechanism when choosing a restaurant in a perceivably unsafe environment, and explore the role of restaurant brand in this choice.  Additionally, we do not model the potential change in response, or, the possibility of increasing  “numbness,” to multiple incidents of food scares over time. Such a study requires a direct measure of consumers’ risk perception and cannot rely on news indices only. Because, news coverage intensity is itself a function of public sensitivity to the matter, we would need to disentangle the effect of lower coverage of news from the lower reaction level of consumers to the crises. This would be an interesting extension of our work, but one with an ambitious data requirement.   96    In sum, this thesis in two essays provides insight into consumer decision making when having food away from home and informs policy debate and managerial decisions. The first essay contributes to the marketing literature by field testing the implications of behavioral theories such as multiple goal pursuit. Our empirical findings reported in the first essay also add to the nutrition and obesity literature and inform the debate on the effectiveness of low-calorie products by providing unique evidence on the calorie cutting benefits of diet carbonated soft drinks in a fast-food meal. The second essay extends the product harm crises literature to the case of industry level crises, as opposed to brand level crises, and to the restaurants context. Our results in the second essay suggest that consumers adopt the rational perceived risk reduction strategy, and indicate that Canadian restaurants, and especially national chains, are affected by the incidents of BSE (mad cow disease). Our empirical findings in the second essay, based on the econometric model we developed for this problem, can provide managers with a basis for more informed cost benefit analysis to design their strategy in encountering future food scares. Finally, we hope that the current work and its future extensions will result in policies and industry practices that align firm objectives with the protection of consumers’ health and well being.  97  References  Altonji, J. G., Elder, T. E., & Taber, C. R. (2005). Selection on Observed and Unobserved Variables: Assessing the Effectiveness of Catholic Schools. Journal of Political Economy, 113(1), 151–184. Anderson, E., & Weitz, B. (1989). Determinants of Continuity in Conventional Industrial Channel Dyads. Marketing Science, 8(4), 310–323. http://doi.org/10.1287/mksc.8.4.310 Ariely, D., & Levav, J. (2000). Sequential Choice in Group Settings: Taking the Road Less Traveled and Less Enjoyed. Journal of Consumer Research, 27(3), 279–290. Ashworth, L., Darke, P. R., & Schaller, M. (2005). No One Wants to Look Cheap: Trade-Offs Between Social Disincentives and the Economic and Psychological Incentives to Redeem Coupons. Journal of Consumer Psychology, 15(4), 295–306. http://doi.org/10.1207/s15327663jcp1504_4 Bhaskaran, S., & Hardley, F. (2002). Buyer Beliefs, Attitudes and Behaviour: Foods with Therapeutic Claims Null. Journal of Consumer Marketing, 19(7), 591–606. http://doi.org/10.1108/07363760210451410 Binkley, J., & Golub, A. (2007). Comparison of Grocery Purchase Patterns of Diet Soda Buyers to Those of Regular Soda Buyers. Appetite, 49(3), 561–571. http://doi.org/10.1016/j.appet.2007.03.225 Bleich, S. N., Wolfson, J. A., Vine, S., & Wang, Y. C. (2014). Diet-Beverage Consumption and Caloric Intake Among US Adults, Overall and by Body Weight. American Journal of Public Health, 104(3), e72–e78. http://doi.org/10.2105/AJPH.2013.301556  98  Braun, K. A., Gaeth, G. J., & Levin, I. P. (1997). Framing Effects with Differential Impact: The Role of Attribute Salience. Advances in Consumer Research, 24(1), 405–411. Bublitz, M. G., Peracchio, L. A., & Block, L. G. (2010). Why Did I Eat That? Perspectives on Food Decision Making and Dietary Restraint. Journal of Consumer Psychology, 20(3), 239–258. http://doi.org/10.1016/j.jcps.2010.06.008 Burton, M., & Young, T. (1996). The Impact of BSE on the Demand for Beef and Other Meats in Great Britain. Applied Economics, 28(6), 687–693. http://doi.org/10.1080/000368496328434 Chernev, A. (2011). The dieter’s paradox. Journal of Consumer Psychology, 21(2), 178. Chernev, A., & Gal, D. (2010). Categorization Effects in Value Judgments: Averaging Bias in Evaluating Combinations of Vices and Virtues. Journal of Marketing Research, 47(4), 738–747. http://doi.org/10.1509/jmkr.47.4.738 Chesher, A. (1997). Diet Revealed? Semiparametric Estimation of Nutrient Intake–Age Relationships. Journal of the Royal Statistical Society: Series A (Statistics in Society), 160(3), 389–428. http://doi.org/10.1111/j.1467-985X.1997.00073.x Comlay, E. (2013, October 31). Coke Femsa Shares Fall as Mexico Passes Food, Drink Taxes. Reuters. Mexico City. Retrieved from http://www.reuters.com/article/2013/10/31/us-mexico-sodatax-idUSBRE99U16120131031 Cook, R. D. (1977). Detection of Influential Observation in Linear Regression. Technometrics, 19(1), 15–18. http://doi.org/10.2307/1268249 Crosby, L. A., Evans, K. R., & Cowles, D. (1990). Relationship Quality in Services Selling: An Interpersonal Influence Perspective. Journal of Marketing, 54(3), 68–81. http://doi.org/10.2307/1251817  99  Dawar, N., & Pillutla, M. M. (2000). Impact of Product-Harm Crises on Brand Equity: The Moderating Role of Consumer Expectations. Journal of Marketing Research, 37(2), 215–226. http://doi.org/10.1509/jmkr.37.2.215.18729 Day, S. (2003, May 21). Mad Cow Case Raises Worries in United States. The New York Times. Retrieved from http://www.nytimes.com/2003/05/21/world/mad-cow-case-raises-worries-in-united-states.html de Castro, J., & de Castro, E. (1989). Spontaneous Meal Patterns of Humans: Influence of the Presence of Other People. The American Journal of Clinical Nutrition, 50(2), 237 –247. DellaValle, D. M., Roe, L. S., & Rolls, B. J. (2005). Does the Consumption of Caloric and Non-caloric Beverages with a Meal Affect Energy Intake? Appetite, 44(2), 187–193. http://doi.org/doi: DOI: 10.1016/j.appet.2004.11.003 Dhar, R., & Simonson, I. (1999). Making Complementary Choices in Consumption Episodes: Highlighting versus Balancing. Journal of Marketing Research, 36(1), 29–44. http://doi.org/10.2307/3151913 Ding, Y., Veeman, M. M., & Adamowicz, W. (2009). BSE and the Dynamics of Beef Consumption: Influences of Habit and Trust (2009 Annual Meeting, July 26-28, 2009, Milwaukee, Wisconsin No. 49284). Agricultural and Applied Economics Association. Retrieved from http://econpapers.repec.org/paper/agsaaea09/49284.htm Fishbach, A., & Dhar, R. (2005). Goals as Excuses or Guides: The Liberating Effect of Perceived Goal Progress on Choice. Journal of Consumer Research, 32(3), 370–377. Flood, J. E., Roe, L. S., & Rolls, B. J. (2006). The Effect of Increased Beverage Portion Size on Energy Intake at a Meal. Journal of the American Dietetic Association, 106(12), 1984–1990. http://doi.org/10.1016/j.jada.2006.09.005  100  Fowler, S. P., Williams, K., Resendez, R. G., Hunt, K. J., Hazuda, H. P., & Stern, M. P. (2008). Fueling the Obesity Epidemic? Artificially Sweetened Beverage Use and Long-term Weight Gain. Obesity, 16(8), 1894–1900. http://doi.org/10.1038/oby.2008.284 Gardener, H., Rundek, T., Markert, M., Wright, C. B., Elkind, M. S. V., & Sacco, R. L. (2012). Diet Soft Drink Consumption is Associated with an Increased Risk of Vascular Events in the Northern Manhattan Study. Journal of General Internal Medicine, 27(9), 1120–1126. http://doi.org/10.1007/s11606-011-1968-2 Garn, S. M., Bailey, S. M., Cole, P. E., & Higgins, I. T. (1977). Level of Education, Level of Income, and Level of Fatness in Adults. The American Journal of Clinical Nutrition, 30(5), 721–725. Government of Canada, P. H. A. of C. (2009, June 17). Obesity in Canada: snapshot - Public Health Agency of Canada. Retrieved March 27, 2014, from http://www.phac-aspc.gc.ca/publicat/2009/oc/index-eng.php#tot Heckman, J. J. (1979). Sample Selection Bias as a Specification Error. Econometrica, 47(1), 153–161. http://doi.org/10.2307/1912352 Holt, S. H., Sandona, N., & Brand-Miller, J. C. (2000). The Effects of Sugar-free Versus Sugar-rich Beverages on Feelings of Fullness and Subsequent Food Intake. International Journal of Food Sciences and Nutrition, 51(1), 59–71. http://doi.org/10.1080/096374800100912 Hosmer, D. W., & Lemeshow, S. (2004). Applied Logistic Regression. John Wiley & Sons. Hyun, S. S. (2010). Predictors of Relationship Quality and Loyalty in the Chain Restaurant Industry. Cornell Hospitality Quarterly, 51(2), 251–267. http://doi.org/10.1177/1938965510363264  101  Ishida, T., Ishikawa, N., & Fukushige, M. (2010). Impact of BSE and Bird Flu on Consumers’ Meat Demand in Japan. Applied Economics, 42(1), 49–56. http://doi.org/10.1080/00036840701564392 Jensen, H. H., & Yen, S. T. (1996). Food Expenditures Away From Home by Type of Meal. Canadian Journal of Agricultural Economics/Revue Canadienne D’agroeconomie, 44(1), 67–80. http://doi.org/10.1111/j.1744-7976.1996.tb00143.x Johnson, E. J., & Goldstein, D. (2003). Do Defaults Save Lives? Science, 302(5649), 1338–1339. http://doi.org/10.1126/science.1091721 Jones, A. M. (1989). A Double-hurdle Model of Cigarette Consumption. Journal of Applied Econometrics, 4(1), 23–39. http://doi.org/10.1002/jae.3950040103 Jones, A. M. (1992). A Note on Computation of the Double-Hurdle Model with Dependence with an Application to Tobacco Expenditure. Bulletin of Economic Research, 44(1), 67–74. http://doi.org/10.1111/j.1467-8586.1992.tb00507.x Khare, A., & Inman, J. J. (2009). Daily, Week-Part, and Holiday Patterns in Consumers’ Caloric Intake. Journal of Public Policy & Marketing, 28(2), 234–252. http://doi.org/10.1509/jppm.28.2.234 Kiefer, I., Rathmanner, T., & Kunze, M. (2005). Eating and Dieting Differences in Men and Women. Journal of Men’s Health and Gender, 2(2), 194–201. http://doi.org/10.1016/j.jmhg.2005.04.010 Kim, S.-Y., Nayga, R. M., & Capps, O. (2000). The Effect of Food Label Use on Nutrient Intakes: An Endogenous Switching Regression Analysis. Journal of Agricultural and Resource Economics, 25(1), 215–231.  102  Kim, S.-Y., Nayga, R. M., Jr., & Capps, O., Jr. (2001). Food Label Use, Self-Selectivity, and Diet Quality. The Journal of Consumer Affairs, 35(2), 346–363. Knowles, T., Moody, R., & McEachern, M. G. (2007). European Food Scares and Their Impact on EU Food Policy. British Food Journal, 109(1), 43–67. http://doi.org/10.1108/00070700710718507 Lavin, J. H., French, S. J., & Read, N. W. (1997). The Effect of Sucrose- and Aspartame-Sweetened Drinks on Energy Intake, Hunger and Food Choice of Female, Moderately Restrained Eaters. Published Online: 06 January 1997; doi:10.1038/sj.ijo.0800360, 21(1). http://doi.org/10.1038/sj.ijo.0800360 Leon, W. (2011, December 29). France Approves Fat Tax on Sugary Drinks Such as Coca-Cola and Fanta. Retrieved March 27, 2014, from http://www.dailymail.co.uk/news/article-2079796/France-approves-fat-tax-sugary-drinks-Coca-Cola-Fanta.html Liu, Y., & Shankar, V. (2015). The Dynamic Impact of Product-Harm Crises on Brand Preference and Advertising Effectiveness: An Empirical Analysis of the Automobile Industry. Management Science, 61(10), 2514–2535. http://doi.org/10.1287/mnsc.2014.2095 Mangen, M.-J. j., & Burrell, A. m. (2001). Decomposing Preference Shifts for Meat and Fish in the Netherlands. Journal of Agricultural Economics, 52(2), 16–28. http://doi.org/10.1111/j.1477-9552.2001.tb00922.x Martin, A. (2007, March 7). Makers of Sodas Try a New Pitch: They’re Healthy. The New York Times. Retrieved from http://www.nytimes.com/2007/03/07/business/07soda.html Maynard, L., Goddard, E., & Conley, J. (2008). Impact of BSE on Beef Purchases in Alberta and Ontario Quick-Serve Restaurants. Canadian Journal of Agricultural Economics/Revue  103  Canadienne D’agroeconomie, 56(3), 337–351. http://doi.org/10.1111/j.1744-7976.2008.00132.x Maynard, L., & Wang, X. (2010). Context-Dependent BSE Impacts on Canadian Fresh Beef Purchases. Journal of International Food & Agribusiness Marketing, 23(1), 32–55. http://doi.org/10.1080/08974438.2011.534027 Mazzocchi, M., & Lobb, A. E. (2005). A Latent-Variable Approach to Modelling Multiple and Resurgent Meat Scares in Italy (2005 International Congress, August 23-27, 2005, Copenhagen, Denmark No. 24509). European Association of Agricultural Economists. Retrieved from http://econpapers.repec.org/paper/agseaae05/24509.htm Michael, R. T., & Becker, G. S. (1973). On the New Theory of Consumer Behavior. The Swedish Journal of Economics, 75(4), 378–396. http://doi.org/10.2307/3439147 Miller, P. E., & Perez, V. (2014). Low-calorie Sweeteners and Body Weight and Composition: A Meta-analysis of Randomized Controlled Trials and Prospective Cohort Studies. The American Journal of Clinical Nutrition, 100(3), 765–777. http://doi.org/10.3945/ajcn.113.082826 Mukhopadhyay, A., Sengupta, J., & Ramanathan, S. (2008). Recalling Past Temptations: An Information-­‐Processing Perspective on the Dynamics of Self-­‐Control. Journal of Consumer Research, 35(4), 586–599. http://doi.org/10.1086/591105 Muringai, V., & Goddard, E. (2011). Bovine Spongiform Encephalopathy, Risk Perceptions, and Beef Consumption: Differences Between Canada and Japan. Journal of Toxicology and Environmental Health, Part A, 74(2-4), 167–190. http://doi.org/10.1080/15287394.2011.529327  104  Murray, E., & Shoehn, S. (1992). Lessons from the Tylenol Tragedy on Surviving a Corporate Crisis. Medical Marketing and Media, 14–19. NPD Group. (2013). Fast Food Still King In Canada. Retrieved February 1, 2014, from https://www.npdgroup.ca/wps/portal/npd/ca/news/press-releases/fast-food-still-king-in-canada/ Parham, E. S., & Parham, A. R. J. (1980). Saccharin Use and Sugar Intake by College Students. Journal of the American Dietetic Association, 76(6), 560–563. Patton, L. (n.d.). Yum’s China Sales Slump 29% in April on Chicken Safety Concern. Retrieved June 23, 2015, from http://www.bloomberg.com/news/articles/2013-05-10/yum-s-china-sales-drop-29-in-april-amid-chicken-safety-concerns Peng, Y., McCann-Hiltz, D., & Goddard, E. (2004). Consumer Demand for Meat in Alberta, Canada: Impact of BSE (2004 Annual Meeting, August 1-4, Denver, CO No. 20331). American Agricultural Economics Association (New Name 2008: Agricultural and Applied Economics Association). Retrieved from http://econpapers.repec.org/paper/agsaaea04/20331.htm Pennings, J. M. E., Wansink, B., & Meulenberg, M. T. G. (2002). A Note on Modeling Consumer Reactions to A Crisis: The Case of the Mad Cow Disease. International Journal of Research in Marketing, 19(1), 91–100. http://doi.org/10.1016/S0167-8116(02)00050-2 Peterson, H. H., & Chen, Y.-J. (Kelly). (2005). The Impact of BSE on Japanese Retail Meat Demand. Agribusiness, 21(3), 313–327. http://doi.org/10.1002/agr.20050  105  Pritchett, J., Johnson, K. K., Thilmany, D., & Hahn, W. (2007). Consumer Responses to Recent BSE Events. Journal of Food Distribution Research, 38(2). Retrieved from http://econpapers.repec.org/article/agsjlofdr/43498.htm Reynolds, D., & Balinbin, W. M. (2003). Mad Cow Disease: An Empirical Investigation of Restaurant Strategies and Consumer Response. Journal of Hospitality & Tourism Research, 27(3), 358–368. http://doi.org/10.1177/1096348003251394 Rogers, P. J., Carlyle, J.-A., Hill, A. J., & Blundell, J. E. (1988). Uncoupling Sweet Taste and Calories: Comparison of the Effects of Glucose and Three Intense Sweeteners on Hunger and Food Intake. Physiology & Behavior, 43(5), 547–552. http://doi.org/10.1016/0031-9384(88)90207-7 Rolls, B. J., Kim, S., & Fedoroff, I. C. (1990). Effects of Drinks Sweetened with Sucrose or Aspartame on Hunger, Thirst and Food Intake in Men. Physiology & Behavior, 48(1), 19–26. http://doi.org/10.1016/0031-9384(90)90254-2 Roos, E., Lahelma, E., Virtanen, M., Prättälä, R., & Pietinen, P. (1998). Gender, Socioeconomic Status and Family Status as Determinants of Food Behaviour. Social Science & Medicine, 46(12), 1519–1529. http://doi.org/10.1016/S0277-9536(98)00032-X Roosen, J., Lusk, J. L., & Shogren, J. F. (Eds.). (2011). The Oxford Handbook of the Economics of Food Consumption and Policy (1st ed.). Oxford University Press. Retrieved from http://www.oxfordhandbooks.com/view/10.1093/oxfordhb/9780199569441.001.0001/oxfordhb-9780199569441 Saghaian, S. H., & Reed, M. R. (2007). Consumer Reaction to Beef Safety Scares. International Food and Agribusiness Management Review, 10(01). Retrieved from http://econpapers.repec.org/article/agsifaamr/8173.htm  106  Schlenker, W., & Villas-Boas, S. B. (2009). Consumer and Market Responses to Mad Cow Disease. American Journal of Agricultural Economics, 91(4), 1140–1152. http://doi.org/10.1111/j.1467-8276.2009.01315.x Schroeter, C., Lusk, J., & Tyner, W. (2008). Determining the Impact of Food Price and Income Changes on Body Weight. Journal of Health Economics, 27(1), 45–68. http://doi.org/10.1016/j.jhealeco.2007.04.001 Schulze, M. (2004). Sugar-sweetened Beverages, Weight Gain, and Incidence of Type 2 Diabetes in Young and Middle-aged Women. JAMA, 292(8), 927–934. http://doi.org/10.1001/jama.292.8.927 Shide, D. J., & Rolls, B. J. (1995). Information About the Fat Content of Preloads Influences Energy Intake in Healthy Women. Journal of the American Dietetic Association, 95(9), 993–998. http://doi.org/10.1016/S0002-8223(95)00273-1 Slining, M. M., Ng, S. W., & Popkin, B. M. (2013). Food Companies’ Calorie-Reduction Pledges to Improve U.S. Diet. American Journal of Preventive Medicine, 44(2), 174–184. http://doi.org/10.1016/j.amepre.2012.09.064 Soberon-Ferrer, H., & Dardis, R. (1991). Determinants of Household Expenditures for Services. Journal of Consumer Research, 17(4), 385–397. Stellman, S. D., & Garfinkel, L. (1986). Artificial Sweetener Use and One-year Weight Change Among Women. Preventive Medicine, 15(2), 195–202. http://doi.org/10.1016/0091-7435(86)90089-7 Suez, J., Korem, T., Zeevi, D., Zilberman-Schapira, G., Thaiss, C. A., Maza, O., Elinav, E. (2014). Artificial Sweeteners Induce Glucose Intolerance by Altering the Gut Microbiota. Nature, 514(7521), 181–186. http://doi.org/10.1038/nature13793  107  Swithers, S. E., & Davidson, T. L. (2008). A Role for Sweet Taste: Calorie Predictive Relations in Energy Regulation by Rats. Behavioral Neuroscience, 122(1), 161–173. Tavernise, S. (2014, November 26). Calories on Menus: Nationwide Experiment Into Human Behavior. The New York Times. Retrieved from http://www.nytimes.com/2014/11/27/upshot/calories-on-menus-a-nationwide-experiment-into-human-behavior.html Tungate, M. (2008). Branded Male: Marketing to Men. Kogan Page Publishers. Turrell, G. (1997). Determinants of Gender Differences in Dietary Behavior. Nutrition Research, 17(7), 1105–1120. http://doi.org/10.1016/S0271-5317(97)00082-1 Verbeke, W. (2006). Functional Foods: Consumer Willingness to Compromise on Taste for Health? Food Quality and Preference, 17(1–2), 126–131. http://doi.org/10.1016/j.foodqual.2005.03.003 Wansink, B., & Chandon, P. (2006). Can “Low-Fat” Nutrition Labels Lead to Obesity? Journal of Marketing Research, 43(4), 605–617. http://doi.org/10.1509/jmkr.43.4.605 Wardle, J., Haase, A. M., Steptoe, A., Nillapun, M., Jonwutiwes, K., & Bellisie, F. (2004). Gender Differences in Food Choice: The Contribution of Health Beliefs and Dieting. Annals of Behavioral Medicine, 27(2), 107–116. http://doi.org/10.1207/s15324796abm2702_5 White, K., & Dahl, D. W. (2007). Are All Out-Groups Created Equal? Consumer Identity and Dissociative Influence. Journal of Consumer Research, 34(4), 525–536. http://doi.org/10.1086/520077  108  Wilcox, K., Vallen, B., Block, L., & Fitzsimons, G. J. (2009). Vicarious Goal Fulfillment: When the Mere Presence of a Healthy Option Leads to an Ironically Indulgent Decision. Journal of Consumer Research, 36(3), 380–393. http://doi.org/10.1086/599219 Wooldridge, J. M. (2010). Econometric Analysis of Cross Section and Panel Data. MIT Press. Yang, J., & Goddard, E. (2011). Canadian Consumer Responses to BSE with Heterogeneous Risk Perceptions and Risk Attitudes. Canadian Journal of Agricultural Economics/Revue Canadienne D’agroeconomie, 59(4), 493–518. http://doi.org/10.1111/j.1744-7976.2011.01225.x Yen, S. T. (1993). Working Wives and Food Away from Home: The Box-Cox Double Hurdle Model. American Journal of Agricultural Economics, 75(4), 884–895. http://doi.org/10.2307/1243976 Yeung, R. M. W., & Morris, J. (2001). Food Safety Risk: Consumer Perception and Purchase Behaviour. British Food Journal, 103(3), 170–187. http://doi.org/10.1108/00070700110386728     109  Appendices  Appendix A  –  Chapter 2: Supplemental Analyses  This appendix presents the Fixed Effect and OLS estimations of the model with interactions with smaller sets of covariates for both dependent variables, total calories (Table A-1) and food calories (Table A-3). It also contains the key marginal effects for both cases in Table A-2 for total calorie estimations and in Table A-4 for food calorie estimations.  In Table A-5, we provide descriptive statistics on share of meals with CSD and share of meals with diet CSD conditioned on order of any CSD. The complete tables of the full models (with three-way interactions) are presented in this appendix (Table A-6 and Table A-7).             110  Table A-1. Total Calories Model Construction Steps (with Interaction)   Fixed Effect  OLS  Specification: (1)  (2)  (3)  (4)  (1)  (2)  (3)  (4)  Diet -223.5***  -226.3***  -210.1***  -209.7***  -260.6***  -271.4***  -219.3***  -225.8***  Small -176***    -178.6***  -178.4***  -212.6***    -184.6***  -180.9***  Large 104.3***    104.2***  104.2***  138.7***    120.8***  120.8***  Male   N.A.  N.A.  N.A.    110.3***  82.2***  71***  Age     7.1  7.1      -0.1  1.4  Age2     -0.1  -0.1      0**  -0.1***  Diet X Small  107.7***      120.9***  61.5***    62.1***  80.9***  Diet X Large -98.9***      -101.3***  -130.7***    -138.3***  -121.3***  Diet X Male   42  66.9**  61.4**    -10.8  21.2  24.6  Diet X Age     -9.4  -9.9      3  10.1*  Diet X Age2     0.1  0.1      0  -0.1*  Small X Male       11.3        27.6  Small X Age       -9.3*        -4.2  Small X Age2       0.1**        0  Large X Male       38**        30.3*  Large X Age       -3.9        -3.6  Large X Age2       0.1        0  Small X Diet X Male       41.4        24.3  Small X Diet X Age       -9.4        -18.6**  Small X Diet X Age2       0.1        0.2**  Large X Diet X Male       3        -26.3  Large X Diet X Age       1.8        -16.6  Large X Diet X Age2       0        0.1  Intercept 1050.4***  1052.9***  1056***  1054.9***  1062***  1066***  1059.3***  1059***  Notes: * p < 0.1, ** p < 0.05, *** p < 0.01; N.A.: Male dummy is dropped in FE analysis (no variation).        111  Table A-2. Total Calories Marginal Effects for Model Construction Steps  E(cal|diet) - E(cal|regular) Fixed Effect  OLS  at  (1)  (2)  (3)  (4)  (1)  (2)  (3)  (4)  Overall Overall   -226.3***        -271.4***       Small CSD -115.8***    -100.6***  -88.7***  -199.1***    -157.1***  -144.9***   Medium CSD -223.5***    -210.1***  -209.7***  -260.6***    -219.3***  -225.8***   Large CSD -322.4***    -315.1***  -311***  -391.3***    -357.5***  -347***   Age = 20     -127.9**        -216.8***     Age = 45     -228.5***        -214.3***     Age = 70     -226.2***        -266.9***    Male Overall   -202.6***        -277.5***       Small CSD     -62.8**  -30.7      -145.2***  -117.3***   Medium CSD     -172.4***  -175***      -207.3***  -211.9***   Large CSD     -277.3***  -274.7***      -345.6***  -348***   Age = 20     -195.1***(L)  -186.7**(L)      -142.7***(S),-204.8***(M),-343.1***(L) -265.1***(M),-248.2**(L)  Age = 45     -81.2**(S),-190.7***(M),-295.7***(L) -68.1*(S),-194***(M),-291***(L)     -140.2***(S),-202.3***(M),-340.6***(L) -133.8***(S),-193.1***(M),-362***(L)  Age = 70     -188.5***(M),-293.4***(L) -167.9***(M),-320.1***(L)     -192.8***(S),-254.9***(M),-393.2***(L) -122.7***(S),-260***(M),-438.3***(L) Female Overall   -244.6***        -266.7***       Small CSD     -129.7***  -133.6***      -166.4***  -166.3***   Medium CSD     -239.3***  -236.5***      -228.5***  -236.5***   Large CSD     -344.3***  -339.1***      -366.8***  -346.3***   Age = 20     -157.1***(M), -262***(L) -162.6***(M),-251.1***(L)     -163.9***(S),-226***(M),-364.3***(L) -289.7***(M),-246.5***(L)  Age = 45     -148.1***(S),-257.7***(M),-362.6***(L) -171***(S),-255.5***(M),-355.4***(L)   -161.4***(S),-223.5***(M),-361.8***(L) -182.8***(S),-217.8***(M),-360.3***(L)  Age = 70     -145.8***(S),-255.4***(M),-360.3***(L) -142***(S),-229.3***(M),-384.5***(L)   -214***(S),-276.1***(M),-414.4***(L) -171.7***(S),-284.7***(M),-436.6***(L) Notes: * p < 0.1, ** p < 0.05, *** p < 0.01; (S) and (M) is marginal at small and medium-size CSD. N.S: Not significant across any of the CSD sizes. "Overall" marginals are evaluated at mean value of the rest of the covariates.     112   Table A-3. Food Calories Model Construction Steps (with Interaction)         Fixed Effect  OLS  Specification: (1)  (2)  (3)  (4)  (1)  (2)  (3)  (4)  Diet -3.5  7.3  9.9  10.3  -40.6***  -35.2***  0.7  -5.8  Small -106***    -108.6***  -108.4***  -142.6***    -114.6***  -110.9***  Large 4.3    4.2  4.2  38.7***    20.8***  20.8***  Male   N.A.  N.A.  N.A.    91.3***  82.2***  71***  Age     7.1  7.1      -0.1  1.4  Age2     -0.1  -0.1      0**  -0.1***  Diet X Small  37.7    39.6  50.9**  -8.5    -7.9  10.9  Diet X Large 1.1    -4.9  -1.3  -30.7    -38.3  -21.3  Diet X Male   55.1**  66.9**  61.4**    8.2  21.2  24.6  Diet X Age     -9.4  -9.9      3  10.1*  Diet X Age2     0.1  0.1      0  -0.1*  Small X Male       11.3        27.6  Small X Age       -9.3*        -4.2  Small X Age2       0.1**        0  Large X Male       38**        30.3*  Large X Age       -3.9        -3.6  Large X Age2       0.1        0  Small X Diet X Male       41.4        24.3  Small X Diet X Age       -9.4        -18.6**  Small X Diet X Age2       0.1        0.2**  Large X Diet X Male       3        -26.3  Large X Diet X Age       1.8        -16.6  Large X Diet X Age2       0        0.1  Intercept 830.4***  817.5***  836***  834.9***  842***  829.7***  839.3***  839***  Notes: * p < 0.1, ** p < 0.05, *** p < 0.01; N.A.: Male dummy is dropped in FE analysis (no variation).        113   Table A-4. Food Calories Marginal Effects for Model Construction Steps        E(cal|diet) - E(cal|regular) Fixed Effect  OLS  at  (1)  (2)  (3)  (4)  (1)  (2)  -3.0  -4.0  Overall Overall   7.3        -35.2***       Small CSD 34.2    49.4**  61.3**  -49.1**    -7.1  5.1   Medium CSD -3.5    9.9  10.3  -40.6***    0.7  -5.8   Large CSD -2.4    4.9  9  -71.3**    -37.5  -27   Age = 20     87.2        2.6     Age = 45     -13.4        5.1     Age = 70     -11.1        -47.4*    Male Overall   37.3*        -30.7*       Small CSD     85.9***  117.3***      4.4  31.8   Medium CSD     46.3**  43.8*      12.3  7.7   Large CSD     41.4  44.1      -26  -28   Age = 20     123.6**  264.2**(S) 113.4*(M)     14.2  N.S   Age = 45     56.7  70.8*(S)      -7  N.S   Age = 70     23.1  N.S.      16.7  -124.9***(L)  Female Overall   -17.7        -38.9**       Small CSD     19  14.4      -16.8  -17.2   Medium CSD     -20.6  -17.6      -8.9  -17   Large CSD     -25.5  -20.3      -47.2  -26.3   Age = 20     -43.8**  161.4*(S)      -4.5  -67.6*(M)   Age = 45     25.4  -40.8*(M)      -35.9  N.S.   Age = 70     -41.6  N.S.      -57.1*  -62.5*(M) -123.2*(L) Notes: * p < 0.1, ** p < 0.05, *** p < 0.01; (S) and (M) is marginal at small and medium-size CSD. N.S: Not significant across any of the CSD sizes.  "Overall" marginals are evaluated at mean value of the rest of the covariates.             114   Table A-5. Percentage of CSD and Diet CSD in Restaurants  % order with CSD  % diet CSD conditional on CSD order Subway 34  27 Wendy’s 45  28 McDonald’s (full sample) 37  25 KFC 31  20 Harvey's 15  27 Arby's 39  29 Burger King 44  18 A &W 39  16 All Above Fast-food Outlets 35  23 All Restaurants 18  24         115  Table A-6. Total Calories – Full Model Coefficient Estimations with Three-way Interactions   OLS Random Effect Fixed Effect Diet -223.30*** -223.85*** -208.14*** Small -166.90*** -164.29*** -156.78*** Large 116.86*** 112.30*** 99.86*** Male 71.83*** 65.93*** N.A. Age 2.07 3.54* 4.52 Age2 -0.06*** -0.07*** -0.10 Income -3.29** -4.76*** -2.70 Party Size 4.70 6.02 15.10* Breakfast -172.93*** -162.28*** -153.15*** Lunch 60.59*** 76.10*** 83.47*** Dinner 90.95*** 103.19*** 100.30*** Diet X Small  65.33*** 78.97*** 98.92*** Diet X Large -111.68*** -102.43*** -103.69*** Diet X Male 29.59 30.49* 75.00*** Diet X Age 8.75* -0.87 -12.04* Diet X Age2 -0.10* 0.00 0.12 Diet X Income -2.19 -1.66 -3.05 Diet X Party Size -3.69 -3.85 -6.27 Diet X Breakfast -34.38 -25.81 -3.75 Diet X Lunch 74.71** 37.26 16.28 Diet X Dinner 68.59* 28.59 15.99 Small X Male 27.15 23.42 8.51 Small X Age -7.12* -9.66** -13.30*** Small X Age2 0.07* 0.10*** 0.15*** Small X Income 3.44 5.91** 8.63** Small X Party Size 3.17 3.48 5.75 Small X Breakfast 164.17*** 145.95** 89.76 Small X Lunch 75.38** 57.69** 52.53 Small X Dinner 103.81*** 89.39*** 108.00*** Large X Male 32.26** 23.47* 34.09* Large X Age -1.65 -0.00 -0.08 Large X Age2 0.03 0.01 0.02 Large X Income -2.68 -0.83 -1.33 Large X Party Size  -2.99 -0.70 -5.06 Large X Breakfast 36.39 30.90 23.25 Large X Lunch 56.54* 45.76 39.69 Large X Dinner 71.54** 61.03* 50.32 Small X Diet X Male 2.50 -6.61 27.66 Small X Diet X Age -12.79 -7.77 -5.72 Small X Diet X Age2 0.14* 0.08 0.05 Small X Diet X Income -2.20 -0.54 -3.23 Small X Diet X Party Size -10.40 -5.53 0.90 Small X Diet X Breakfast 41.23 155.24 267.98* Small X Diet X Lunch 10.81 37.00 83.40 Small X Diet X Dinner 21.39 41.08 53.98 Large X Diet X Male -53.67 -26.72 -20.35 Large X Diet X Age -1.52 1.85 4.99 Large X Diet X Age2 0.01 -0.03 -0.06 Large X Diet X Income 5.60 7.62 7.48 Large X Diet X Party Size 2.27 0.33 19.80 Large X Diet X Breakfast -93.52 70.27 164.92 Large X Diet X Lunch -57.68 -0.55 66.79 Large X Diet X Dinner -10.67 22.02 57.01 Intercept 1055.54*** 1053.12*** 1053.51*** Sample Size 11738 11738 11738 R2 0.409 0.167(0.404) 0.172 Note: * p < 0.1, ** p < 0.05, *** p < 0.01    116  Table A-7. Food Calories – Full Model Coefficient Estimations with Three-way Interactions   OLS Random Effect Fixed Effect Diet -3.30 -3.85 11.86 Small -96.90*** -94.29*** -86.78*** Large 16.86** 12.30* -0.14 Male 71.83*** 65.93*** N.A. Age 2.07 3.54* 4.52 Age2 -0.06*** -0.07*** -0.10 Income -3.29** -4.76*** -2.70 Party Size 4.70 6.02 15.10* Breakfast -172.93*** -162.28*** -153.15*** Lunch 60.59*** 76.10*** 83.47*** Dinner 90.95*** 103.19*** 100.30*** Diet X Small  -4.67 8.97 28.92 Diet X Large -11.68 -2.43 -3.69 Diet X Male 29.59 30.49* 75.00*** Diet X Age 8.75* -0.87 -12.04* Diet X Age2 -0.10* 0.00 0.12 Diet X Income -2.19 -1.66 -3.05 Diet X Party Size -3.69 -3.85 -6.27 Diet X Breakfast -34.38 -25.81 -3.75 Diet X Lunch 74.71** 37.26 16.28 Diet X Dinner 68.59* 28.59 15.99 Small X Male 27.15 23.42 8.51 Small X Age -7.12* -9.66** -13.30*** Small X Age2 0.07* 0.10*** 0.15*** Small X Income 3.44 5.91** 8.63** Small X Party Size 3.17 3.48 5.75 Small X Breakfast 164.17*** 145.95** 89.76 Small X Lunch 75.38** 57.69** 52.53 Small X Dinner 103.81*** 89.39*** 108.00*** Large X Male 32.26** 23.47* 34.09* Large X Age -1.65 -0.00 -0.08 Large X Age2 0.03 0.01 0.02 Large X Income -2.68 -0.83 -1.33 Large X Party Size  -2.99 -0.70 -5.06 Large X Breakfast 36.39 30.90 23.25 Large X Lunch 56.54* 45.76 39.69 Large X Dinner 71.54** 61.03* 50.32 Small X Diet X Male 2.50 -6.61 27.66 Small X Diet X Age -12.79 -7.77 -5.72 Small X Diet X Age2 0.14* 0.08 0.05 Small X Diet X Income -2.20 -0.54 -3.23 Small X Diet X Party Size -10.40 -5.53 0.90 Small X Diet X Breakfast 41.23 155.24 267.98* Small X Diet X Lunch 10.81 37.00 83.40 Small X Diet X Dinner 21.39 41.08 53.98 Large X Diet X Male -53.67 -26.72 -20.35 Large X Diet X Age -1.52 1.85 4.99 Large X Diet X Age2 0.01 -0.03 -0.06 Large X Diet X Income 5.60 7.62 7.48 Large X Diet X Party Size 2.27 0.33 19.80 Large X Diet X Breakfast -93.52 70.27 164.92 Large X Diet X Lunch -57.68 -0.55 66.79 Large X Diet X Dinner -10.67 22.02 57.01 Intercept 835.54*** 833.12*** 833.51*** Sample Size 11738 11738 11738 R2 0.187 0.068 0.074 Note: * p < 0.1, ** p < 0.05, *** p < 0.01  117   Appendix B  –  Chapter 3: Newspapers Used in Constructing the Media Index  We generated a monthly media index across Canada using Canadian Newsstand Complete and FACTIVA databases to get the local coverage of the news. We also generated a separate index for French newspapers by searching for French translation of keywords in FACTIVA in Canada. We present the list of newspapers for each region in this appendix. To select the newspapers we picked up to five newspapers (depending on availability of newspapers in the database) in cities with more than 100,000 population. Here is the list of newspapers considered in each province.   Newfoundland and Labrador St. John's Telegram (Telegram)  Nova Scotia Sydney Cape Breton Post  The Chronicle Herald (Halifax); Metro Halifax – not available in the database Prince Edward Island  Charlottetown Guardian (Guardian) – No city above 100,000, most circulated New Brunswick Times-Transcript (Moncton); Telegraph-Journal (New Brunswick) Gleaner (New Brunswick) – Fredericton town has less than 100,000, picked the highest circulating  Quebec  118  The Gazette; Sherbrooke Record  Ontario (city cut-off 300K+) Toronto Star; The Windsor Star; The Ottawa Citizen; Hamilton Spectator (The Spectator); Ottawa-Gatineau Le Droit; St. Catharines Standard (Standard); Welland Tribune (Welland Tribune); Waterloo Region Record (The Record) – The Record in the database has two cases in BC but nothing in Ontario; Toronto Sun; Metro Toronto; Toronto 24 Hours; Ottawa Sun; London Free Press – not available in the database Manitoba Winnipeg Free Press Winnipeg Sun; Metro Winnipeg – not available in the database Saskatchewan Regina Leader-Post; Saskatoon Star Phoenix Alberta Calgary Herald; Edmonton Journal  Metro Calgary; Edmonton Sun; Metro Edmonton; Calgary Sun – not available in the database British Columbia Vancouver Sun; Vancouver Province; Victoria Times Colonist; Nanaimo Daily News (city is less than 100K); Daily News (Prince Rupert, B.C.); Abbotsford Times; Abbotsford News Daily; Kelowna Daily Courier; Vancouver 24 Hours; Metro Vancouver; The Record (New Westminster, B.C.); The Journal (Ashcroft, B.C.); The Prince George Citizen – not available in the database   119   Appendix C  –  Chapter 3: Constructing the Local Weather Variables   We collected the monthly weather data for each province between the years 2000 and 2006 from the website http://climate.weather.gc.ca/ on the following variables: average temperature, extreme maximum and minimum temperature, total rain, and total snow.  The location-identifying variables in the NPD Group’s CREST data are the province and city size, a categorical variable reflecting city population. We constructed a matrix of ten provinces and six levels of city size in which we allocated 140 Canadian cities and towns with more than 10,000 population. Some cities have multiple weather stations (we used the airport weather station), and some minor cities do not have any report (we used other cities in the same cell to represent their weather). In this appendix, we present the matrix that classifies Canadian cities based on their population (city size) and province (region). In each province–city-size combination, we recorded the weather information for (up to) the three largest cities with complete weather data between 2000 and 2006 in the Environment Canada database. Next we present the cities matrix: 120  Table C-1. Province-City Size Matrix of Canadian Cities Province/city size 10K-30K 30K-100K  100K-250K 250K-500K 500K-1M Above 1M BC Parksville (incomp), Fort St. John (collected), Port Alberni (incomp downloaded the Cox Lake station of Port Alberni),  Cranbrook (collected),  Quesnel (incomp),  Williams Lake (collected)  (the rest ignored) Salmon Arm, Squamish,  Powell River ,  Terrace,  Prince Rupert,  Dawson Creek Kamloops, Nanaimo, Chilliwack (the rest ignored) Prince George,  Wood Buffalo,  Vernon (Coldstream),  Courtenay,  Duncan,  Penticton,  Campbell River Kelowna (West Kelowna), Abbotsford-Mission Victoria NONE Metropolitan Vancouver AB Lloydminster (collected), Okotoks (collected), Brooks (collected), (the rest ignored) Camrose,  Cold Lake,  High River,  Sylvan Lake,  Wetaskiwin,  Strathmore,  Canmore,  Lacombe Lethbridge, Red Deer, Medicine Hat (ignored), Grande Prairie NONE NONE Calgary, Edmonton (the position in the dataset) NONE SK Lloydminster (collected), North Battleford (Battleford) (collected from two stations over time- merged), Yorkton (collected from two stations over time- merged),  Swift Current (collected), Estevan (ignored) Prince Albert, Moose Jaw (incomp) Saskatoon, Regina NONE NONE NONE  121  Province/city size 10K-30K 30K-100K  100K-250K 250K-500K 500K-1M Above 1M MB Steinbach (incomp), Portage la Prairie (collected), Thompson (collected) Brandon NONE NONE Winnipeg NONE ON Centre Wellington (incomp- not found), Pembroke (incomp), Collingwood (collected no 2007),  Cobourg (collected no 2007),  Port Hope (incomp),  Petawawa (collected no 2007),  (the rest ignored) Kenora, Tillsonburg,Temiskaming Shores,  Ingersoll,  Hawkesbury (Grenville),  Elliot Lake Belleville (Quinte West) (missing 2007 but collected), Sault Ste. Marie (collected), Kawartha Lakes (incomp), North Bay (collected),  Sarnia (St. Clair) (missing 2007 but collected),  Norfolk (incomp),  Cornwall (ignored) (South Stormont) (incomp),  Leamington (incomp),  Timmins (ignored),  Orillia,  Brockville (incomp),  (the rest ignored) Woodstock,  Midland,  Owen Sound,  Stratford Barrie (used Lake Barrie station and Barrie MPC station files), Greater Sudbury (used Sudbury A station), Kingston, Guelph (incomp), Brantford (Brant) (collected missing 2007),   (the rest incomp)  Thunder Bay, Peterborough (Selwyn) Kitchener-Cambridge-Waterloo (incomplete), London (incomplete), St. Catharines-Niagara (incomp), Oshawa (incomp), Windsor Hamilton  Metropolitan Toronto, Ottawa QC Baie-Comeau (incomp), Sept-Îles (incomp),  Thetford Mines (collected),  Rivière-du-Loup (Saint-Antonin) (incomp),  Matane (incomp),  Amos (incomp),  Dolbeau-Mistassini (incomp),  Lachute (collected),  Cowansville (not found),  Hawkesbury (Grenville) (ignored) Saint-Jean-sur-Richelieu (incomp), Drummondville  (collected), Granby (collected),  Saint-Hyacinthe (collected), (the rest ignored)  Shawinigan,  Rimouski,  Sorel-Tracy,  Joliette,  Victoriaville,  Rouyn-Noranda,  Salaberry-de-Valleyfield,  Saint-Georges,  Val-d'Or,  Alma Sherbrook (incomp), Saguenay (incom), Trois-Rivières  NONE Quebec Montreal  122  Province/city size 10K-30K 30K-100K  100K-250K 250K-500K 500K-1M Above 1M NB Miramichi (North Esk)   (two stations merged), Edmundston (two stations merged),  Campbellton (incomplete), New Brunswick (Addington, Dalhousie, Listuguj) (not found)    Fredericton (collected), Bathurst (collected) Moncton (Dieppe, Riverview), Saint John (Quispamsis), Chatham-Kent (incomp) NONE NONE NONE PE Summerside (two files should be merged 2003 is missing) Charlottetown (collected) (Stratford) NONE NONE NONE NONE NS Kentville (collected)  Truro (incomp), New Glasgow (incomp) Cape Breton (incomp) Halifax NONE NONE NF Corner Brook (collected), Grand Falls-Windsor (incomp),  Bay Roberts (Spaniard's Bay, Upper Island Cove)  NONE St. John's (Mount Pearl, Conception Bay South) NONE NONE NONE  123  Appendix D  –  Chapter 3: Double-hurdle and DHEP Methods Derivations   Our data structure does not fit the existing two-stage models, thus we proposed an extended version of the double-hurdle model. In this appendix, we discuss double-hurdle models and categorize them based on two assumptions detailed below (Madden, 2008). We present the likelihood function for various models, including the extended double-hurdle method that we propose. We elaborate on the steps to calculate the likelihood function for our proposed procedure.  The underlying idea for the basic double-hurdle model is that individuals need to pass two hurdles before being observed with a positive level of consumption in the sample. The first hurdle in our study is the participation decision (e.g. dine out in a restaurant), and the second hurdle is the consumption decision (e.g. ordering beef in the restaurant). Two crucial assumptions will determine the form of the double-hurdle model (Madden, 2008): 1. Independence between the unobserved terms in the participation and consumption equations (we argue that in the food-scare case, errors are dependent) 2. Dominance of the participation decision over the consumption decision. We will define this condition formally below (our data structure suggests that we do not have a dominance condition since we can have a corner solution when participating).   We formulate the problem as follows. The binary participation decision ( ) for each individual in each occasion is governed by the latent variable , which is modeled as a linear dw 124  function of participation regressor vector ( ) and an unobserved term ( ). In particular, we define (subscripts are dropped for the sake of exposition), 1 if  0 0 if  0 where  wdww Zα ν>⎧= ⎨ ≤⎩= +  D-1 and for the consumption decision we define, **if   00 if   0.X u X uyX uy d yβ ββ+ + ≥⎧= ⎨ + <⎩=        D-2 where  is the latent variable that is a non-negative linear function of a set of consumption regressor ( ) and an unobserved term ( ). The consumption variable ( ) conditions the non-zero consumption values on participation ( ) in the market.  In the classic double-hurdle model we do not observe the participation variable, so the formulations can be summarized as:         D-3  Assumptions on the correlation of the error terms and dominance of the participation decision result in different likelihood functions. Participation dominance means that the zero consumption does not arise from a corner solution to the consumption equation. Once the first hurdle, participation decision, is crossed we have positive consumption value. First-hurdle dominance indicates (g is the conditional probability density function of the consumption variable): Z !y*X u y1d =0 & 00 otherwiseX u X u Zyβ β α ν+ + > + >⎧= ⎨⎩ 125         D-4  We start with the case of a general double-hurdle model with correlated error terms and not dominant participation. When participation is not dominant, we can have two types of zero consumption. The first type is due to not participating, and the second type is due to zero-level consumption conditioned on participating in the market. Even though we conceptually allow for two types of zero we may not be able to distinguish them in the dataset. The likelihood function for the double-hurdle model can be divided into zero consumption outcomes and positive consumption outcomes, and we will have: D-5 where g is the conditional probability density function of the consumption variable. The unobserved terms are allowed to have a joint normal mean zero distribution with correlation coefficient and variance parameters  and , respectively, that are estimated along with the coefficient vectors of the two stage decisions ( ) through likelihood maximization:          D-6   p y* > 0 w > 0( ) = 1 and g y* y* > 0,w > 0( ) = g y* w > 0( )( ) ( )( ) ( ) ( )( ) ( )( ) ( ) ( )*0* *0001 0 0 00 0 0 0, 01,DHyyDHyyL p w p y wp w p y w g y y wL p Z p u X Zp Z p u X Z g y u X Zν α β ν αν α β ν α β ν α=>=>⎡ ⎤= − > > >⎣ ⎦⎡ ⎤> > > > >⎣ ⎦⎡ ⎤= − > − > − > −⎣ ⎦⎡ ⎤> − > − > − > − > −⎣ ⎦∏∏∏∏! ! 2! ," u,!( ) ! N 0,"( )" =1 #$#$ # 2%&''()** 126  The second specification holds when we keep the condition of no dominance for the participation decision, but restrict the error terms to be independent ( ). This will generate the Cragg model (Cragg, 1971) with the following likelihood function:   D-7  The third specification is generated when error terms remain dependent but the participation decision dominates the consumption decision. This specification is equivalent to the Heckman sample selection model with the following likelihood function:   D-8  Finally, the fourth specification results from assuming both independence of the error terms and dominance of the participation decision. The method, sometimes called a two-part model, reduces to a probit for participation and OLS for consumption on the sample with positive consumption. Simplifying the likelihood function even further we have:   D-9  Based on the above classifications, our problem most closely matches the double-hurdle specification. We allow for correlated error terms, and participation should not dominate consumption. However, the proposed likelihood formulation for the double-hurdle model does not take the participation variable (d)—which is unobserved in many applications—into the ! = 0 Lcragg = 1! p " > !Z#( ) p u > !X$( )%& '(y=0)p " > !Z#( ) p u > !X$ " > !Z#( )g y u > !X$( )%& '(y>0) LHeckman = 1! p " > Z#( )$% &'y=0( p " > Z#( )g y " > Z#( )$% &'y>0(( ) ( ) ( )0 01two party yL p Z p Z g yν α ν α−= >= − > >⎡ ⎤ ⎡ ⎤⎣ ⎦ ⎣ ⎦∏ ∏ 127  estimation. In other words, the proposed double-hurdle model does not distinguish between the two types of zero consumption: zero consumption due to not participating, and zero consumption conditioned on participation. Given the observed participation variable in our dataset, however, we can incorporate the information of the d variable in the likelihood function and discern the two types of zero. The conditional probability density function for this new data structure will be:  D-10  We can write the likelihood function for three segments of sample: not participating, participating and zero consumption, and positive consumption:  D-11  After substitution of covariates and error terms we have:  D-12  To calculate the contribution of each observation to the likelihood function we can write: ( ) ( ){ } [ ] ( ) ( ){ } [ ] [ ]( ) ( ){ } [ ] [ ]1 1 1 01 0 *1 1 1 0*0 1 0 11 0 1d ydd yf y X p d p d p y dp d p y d= === >= = = ≤ == > =( ) ( ) ( )( ) ( ) ( )*_0 1, 0* *1, 00 1 0 11 0 1 0, 1double hurdled d yd yL p d p d p y dp d p y d g y y d= = == >⎡ ⎤= = = ≤ =⎡ ⎤⎣ ⎦ ⎣ ⎦⎡ ⎤= > = > =⎣ ⎦∏ ∏∏( ) ( ) ( )( ) ( ) ( )_0 1, 01, 01,double hurdled d yd yL p Z p Z p u X Zp Z p u X Z g y u X Zν α ν α β ν αν α β ν α β ν α= = == >⎡ ⎤= − > − > − < − > −⎡ ⎤⎣ ⎦ ⎣ ⎦⎡ ⎤> − > − > − > − > −⎣ ⎦∏ ∏∏ 128   D-13 where g is conditional joint pdf for y, which is practically pinned down by u value and therefore is based on the joint distribution of . We need to evaluate the three brackets to form the likelihood function. To make the problem tractable we make the joint normality assumption for the error terms—a commonplace assumption in the literature (Greene, 2011). For identification purposes, we will assume  in implementations, but for the sake of exposition we stick to the general case in this note.    D-14  Next we evaluate the three terms in brackets. Note that  and  indicate univariate or joint (depending on subscripts) cumulative distribution function (CDF) and pdf, respectively. The derivation for the first term is straightforward:   D-15  To evaluate the second term we use the following characteristic for joint probability intervals (Hogg, McKean, & Craig, 2005, equation 2.1.2):   D-16   Li = 1! p " > !Zi#( )$% &'1$% &'! "### $###1!di p " > !Zi#( ) p u < !Xi( " > !Zi#( ))1$% &'2$% &'! "####### $#######di1 yi=1$% &'p " > !Zi#( ) p u > !Xi( " > !Zi#( )g yi u > !Xi( ," > !Zi#( )$% &'3$% &'! "############ $############di 1 yi>1$% &' andu ν1νσ = ! ,u( ) ! N 0, "!2 #" u"!#" u"! " u2$%&&'())*+,,-.//Φ φ[ ]1 iZννασ⎛ ⎞−=Φ ⎜ ⎟⎝ ⎠ p a1 < X1 ! b1,a2 < X2 ! b2"# $% = FX1,X2 b1,b2( )& FX1,X2 a1,b2( )& FX1,X2 b1,a2( ) + FX1,X2 a1,a2( )where F is joint CDF in the general form (with correlation) 129   Using the previous equation we can write for the second bracket:  D-17  Finally, the third bracket is evaluated in two references (Amemiya, 1984; Jones, 1992)  by use of conditional joint normal distribution. We follow Jones (1992):   D-18  All the terms used (standard bivariate CDF and pdfs) can be evaluated in standard statistical software. We coded the likelihood function using CDF and pdf functions and maximized it with the “ml” routine in Stata (and “optim” routine in R for some cases). For coding convenience, we transfer the function to log form over all the observations:  D-19  [ ] ( ) ( )( )( )( ) ( ) ( ) ( ),,2,,, , , ,0 , 0,i i iu u vu ui i iu u vu up Z p u X Zp u X Zp u X ZF X F F X Z F ZX X ZX X Zννν α β ν αβ ν αβ α νβ β α αβ β ασ σ σβ β ασ σ σ= > − < − > −= < − > −= −∞ < < − − < < ∞= − ∞ − −∞ ∞ − − − + −∞ −⎛ ⎞ ⎛ ⎞− − −=Φ − −Φ +⎜ ⎟ ⎜ ⎟⎝ ⎠ ⎝ ⎠⎛ ⎞ ⎛ ⎞− − −=Φ −Φ⎜ ⎟ ⎜ ⎟⎝ ⎠ ⎝ ⎠[ ]( ), 2131ii iu i iu v uu uZ y Xy Xν να ρ βσ σ σ βφσ σρ⎡ ⎤+ −⎢ ⎥ ⎛ ⎞−⎢ ⎥=Φ ⎜ ⎟⎢ ⎥− ⎝ ⎠⎢ ⎥⎣ ⎦( ) [ ]( ) [ ]( ) [ ]( ) [ ]( ) [ ]( )log 1 log 1 .1 0 log 2 .1 0 log 3i i i i i i iil L d d y d y⎡ ⎤= = − + = + >⎣ ⎦∑ 130  On a separate note, according to Jones (1992), the appearance of the Z and X term in [3] helps to emphasize the role of exclusion restrictions in identifying the model.   To use the “ml” procedure in Stata to maximize the likelihood function we need to express it as a function of theta parameters—see Gould, Pitblado, and Poi (2010) for further detail. We assign . We also transfer the and  to  and , respectively, to confine the correlation coefficient between -1 and 1 and standard deviation to positive values. The Stata and R estimation codes are available upon request.  D-20       !" = 1 !3 j !4 j ! = tanh("3 j ) ! u = exp("4 j )( ) ( )( )[ ]( )[ ]( )( ),2121 loglog .1 0 log , ,1.1 0 log1i ii ii i i i u v ii u ui i iu i ii i uu uj ij id ZX Xl L d y ZZ y Xy Xd yZXν αβ β α ρσ σρα βσ βφσ σρθ αθ β⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥− Φ −⎢ ⎥⎢ ⎥⎛ ⎞⎛ ⎞ ⎛ ⎞− −⎢ ⎥= = + = Φ −Φ −⎜ ⎟⎜ ⎟ ⎜ ⎟⎜ ⎟⎢ ⎥⎝ ⎠ ⎝ ⎠⎝ ⎠⎢ ⎥⎛ ⎞⎡ ⎤⎢ ⎥+ −⎜ ⎟⎢ ⎥⎢ ⎥⎛ ⎞−⎜ ⎟⎢ ⎥+ > Φ⎢ ⎥⎜ ⎟⎜ ⎟⎢ ⎥− ⎝ ⎠⎢ ⎥⎜ ⎟⎢ ⎥⎢ ⎥⎣ ⎦⎝ ⎠⎣ ⎦==∑34jj uθ ρθ σ== 131   Appendix E  –  Chapter 3: Investigating Supply Side Reaction to Food Scare Outbreaks  In this appendix, we provide supplementary analysis to investigate the reaction of restaurants in our sample at the time of mad cow disease (BSE) outbreaks. Using OLS, we regress the price paid for meals with only a single beef item on the BSE media index, searching for evidence for systematic price change at peak times of BSE. In another set of analyses, we regress the number of coupons and special offers received by households in each month on the BSE media index and its first and second lagged terms (with month as the time series unit). Our goal in these estimations is to see if restaurants have changed their pricing, coupon distribution or special offer levels at the time of crises. In all analyses, we include covariate, explained further, to control for other factors. As shown below, we do not find any significant effect for the BSE media index on the beef item price or promotions. These findings weaken the alternative supply side reaction explanations for the reaction of consumers to BSE crises documented in the body of this paper.   Beef Price Analysis: We cleaned the data for single-person meals with only one beef order and no other meat order. Note that the price of food items is not individually coded in the dataset and we have to infer it from the total bill variable. We kept only the participating sample, since we cannot infer any price from the non-participating subsample. The following tables are the regression of bill for the whole meal (not each item in the meal, as mentioned the data only report the total bill for each eating occasion) on BSE monthly media index—one analysis with normalized index used in main analysis and one analysis with non-normalized media index as  132  robustness check—and a vector of covariates controlling for regional variations, linear trend (inflation), if tip was given, coupon, special offer, number of items ordered in the meal (one of which is beef), meal occasion (breakfast etc.), and type of restaurant (formal, casual, family, fast-food):  Table E-1. Beef Price Analysis  Normalized Media Index Non-Normalized Media Index  DV: Meal Price DV: Meal Price BSE Media -0.8374 0.3806 Alberta -0.1164 0.4455 Saskatchewan  -77.1093*** -77.8122*** Manitoba -26.4911** -26.2121** Ontario 35.0829*** 34.8627*** Quebec -90.0131*** -94.6664*** New Brunswick -36.9252*** -35.6510*** Prince Edward Island -46.8254 -48.1931 Nova Scotia -71.0492*** -77.4822*** Newfoundland 59.2869*** 52.0721** Tip Given Dummy 325.2723*** 325.3969*** Linear Trend 2.0815*** 2.0714*** Coupon 110.4956** 111.1756** Special Offer -216.7611*** -217.4855*** Number of items ordered 125.8686*** 125.9613*** Breakfast Dummy -36.6826 -36.6872 Snack Dummy -90.3049*** -91.1146*** Family Type Place 159.9696*** 160.7000*** Informal Casual Place 448.0695*** 447.8355*** Formal Dining Place 1664.3105*** 1664.9703*** Constant -798.1747*** -795.5624*** Sample Size 20588 20588 Note: * p < 0.1, ** p < 0.05, *** p < 0.01 Sample of single meals with one beef item and no other meat  Media index coefficients are insignificant, so we do not find evidence in our dataset for the price of beef items to change with BSE outbreaks.   Coupon and special offer analysis: We used coupon and special offer variables as a measure of promotional efforts at the time of crises. We regress (OLS) these variables on BSE  133  media index using data aggregated on a monthly level for each province. This structure makes it possible to probe the effect of the lagged media index variable on total coupons or special offers received in the sample for each month. We use region, linear trend, and a dummy for restaurant chain as covariates in this model. The first column in each table is the model with only the current BSE media effect. The next two columns report the model with one and two lagged BSE media variables.  Table E-2. Promotion Analysis – Coupon Models  (1) (2) (3) BSE Media (Normalized) 0.0002 0.0004 0.0004 Lag BSE Media  -0.0003 -0.0007 2 Lags BSE Media   0.0008 Alberta 0.0214*** 0.0212** 0.0214** Saskatchewan  0.0402*** 0.0402*** 0.0399*** Manitoba 0.0330*** 0.0331*** 0.0336*** Ontario -0.0194*** -0.0195*** -0.0196*** Quebec 0.0118* 0.0128** 0.0124* New Brunswick -0.0243*** -0.0246*** -0.0252*** Prince Edward Island 0.1181*** 0.1134*** 0.1124*** Nova Scotia -0.0314*** -0.0313*** -0.0305*** Newfoundland 0.0778*** 0.0773*** 0.0771*** Linear Trend -0.0002* -0.0002* -0.0003* Restaurant Chain (Y/N) 0.0831 0.0910 0.0938 Constant 0.3093*** 0.3121*** 0.3190*** Sample Size  799 788 777 Note: * p < 0.1, ** p < 0.05, *** p < 0.01- Analysis on aggregated monthly provincial sample      134   Table E-3. Promotion Analysis – Special Offer Models  (1) (2) (3) BSE Media (Normalized) 0.0002 0.0004 0.0003 Lag BSE Media  -0.0004 -0.0009 2 Lags BSE Media   0.0009 Alberta 0.0190** 0.0189** 0.0190** Saskatchewan  0.0391*** 0.0394*** 0.0391*** Manitoba 0.0283*** 0.0285*** 0.0286*** Ontario -0.0184*** -0.0184*** -0.0187*** Quebec 0.0065 0.0079 0.0070 New Brunswick -0.0256*** -0.0259*** -0.0266*** Prince Edward Island 0.1237*** 0.1199*** 0.1186*** Nova Scotia -0.0317*** -0.0316*** -0.0309*** Newfoundland 0.0801*** 0.0799*** 0.0797*** Linear Trend -0.0000 -0.0000 -0.0001 Restaurant Chain Dummy 0.0837 0.0910 0.0942 Constant 0.1893** 0.1945** 0.2000** Sample Size 799 788 777 Note: * p < 0.1, ** p < 0.05, *** p < 0.01- Analysis on aggregated monthly provincial sample   The insignificant BSE media (and its lags) coefficients indicate that we cannot detect any systematic reaction to BSE crises by change in the level of coupon distribution or special offers received by consumers.   In summary, the supplementary analysis in this appendix rules out the reaction of the restaurants to BSE outbreaks by changing the beef price or promotional efforts in our dataset of analysis. One may want to further investigate these effects in other datasets with a focus on supply side, a resource unavailable to the authors.     135  Appendix F  –  Chapter 3: Counterfactual Analysis Formulations  In this appendix, we formulate the counterfactual analysis based on the Double Hurdle with Explicit Participation variable method. Our goal is to comment on the percentage change of dine-out probability and meals ordered conditioned on participation for counterfactual values of the BSE media index. Specifically, we pick two peak months of BSE crises and show the counterfactual change in the level of participation and meals ordered had there been no crises during those months. To implement the counterfactual analysis we change the value of the media variable to zero in the peak months of crises and measure the change in the conditional expected number of beef orders and the probability of going out. We calculate the counterfactual changes for the observations belonging to the months of interest and report their average.   When estimating the conditional meal orders in models with correlated error terms, we need to incorporate both direct effect of crises on meal order and indirect effect of the crises through the first stage. The first-stage effect on the second-stage decision is captured by the inverse Mills ratio (IMR) term (Greene, 2011).   ( ) ( )1 1, , 0where: food choice (second stage) regressors: dine out (first stage) regressors:  excluding food scare media variable ( ):  excluding food scare media variable ( )R RE y X Z w p R p RXZR X pR Z pλβ β β λ α α ′′> = + + +′ F-1   136  Therefore, to find the counterfactual differences we need to find the selection correction term ( ) as a function of media index. We plug the covariates and coefficients of the first stage into the inverse Mills ratio formulation:    F-2  The counterfactual difference is the difference between the two expected values based on the actual ( ) and counterfactual ( ) values of the media variable:   F-3  Calculating this counterfactual difference is similar to marginal effect estimations in non-linear models. The two common practices are implementing the computations at the mean value of covariates, or averaging the effects evaluated at the observation level (called average marginal effect). The two practices normally provide similar estimates, and we use the average marginal effect procedure in this work. That is, we evaluate each observation separately then average the effects across all observations. For example, to compute the counterfactual change in we write:  F-4  λ( ) ( )( )( )( )11=where  is Normal PDF is Normal CDFRRZ p RZZ p Rφ α φ α αλ αα α αφ′′′+=′Φ Φ +Φp cp( ) ( )( ) ( ) ( )1 1 1| , Z , 0 | ,Z, 0c cc cR RE y X w E y X wE p p E p R p Rλβ β λ α α λ α α′ ′> − >⎡ ⎤′ ′= − + + − +⎣ ⎦λ( ) ( )( )( )( )( )1 11 11 11 1 11cR RcN Nn R n Rcn n n Rn Rp R p Rp R p RN Np Rp Rλ α α λ α αφ α α φ α αα αα α′ ′′ ′− −= = ′′′ ′+ − + =′+ ′+−′Φ +′Φ +∑ ∑ 137  On the other hand, value is determined by the counterfactual scenario that we would like to consider. We fix the counterfactual value for the media variable at zero ( ). Conceptually, this would correspond to the case of no food-scare occurrence. Doing a simple derivation for the discussed counterfactual scenarios and averaging over observations we have:   F-5  Monthly counterfactual analysis: The counterfactual scenarios discussed provide the general framework to measure the market reaction in different scenarios. Media coverage of food scares peaks for one or two months and fades away. Therefore, for most of the periods in the data we observe low media coverage of crises and do not expect to see a major change in consumers’ behavior for the counterfactual scenario. Consequently, the average counterfactual effect over all the periods does not showcase the effects of interest, which is consumer reaction to peaks of crises. This encourages us to formulate a monthly counterfactual analysis for the peak months of media coverage. Thus, we condition the expectations on the month of interest ( ), and the average marginal effect is calculated on the observations that belong to the month of interest:   F-6   pc pc = 0( ) ( )( ) ( ) ( )( ) ( )( )( )( )11 111111 1 1| , Z , 0 | , Z, 000c cNn R RnN Nn R n Rn nn R n RE y X w E y X wN p R p RR p Rp NR p Rλλβ β λ α λ δ αφ α φ α αβ βα α α−′ ′=′ ′−= =′ ′> − >⎡ ⎤′ ′= − + − +⎡ ⎤⎣ ⎦⎣ ⎦⎡ ⎤′ ′+= − + −⎢ ⎥′ ′Φ Φ +⎣ ⎦∑∑ ∑ t = !( ) ( )| t , ,Z , 0 | t , ,Z, 0c cCFD E y X w E y X wτ τ= = > − = > 138  Note that to provide a more straightforward interpretation we report our findings based on the percentage change. We generate the percentage numbers by dividing CFD by the base expectation value:  F-7  Finally, we calculate the standard errors in each case with delta method. We use the calculated standard errors to indicate the level of significance for the reported counterfactuals.   ( ) ( )( )| t , , Z , 0 | t , , Z, 0Counterfactual Change%| t , , Z, 0c cE y X w E y X wE y X wτ ττ= > − = >== > 139  Appendix G  –  Chapter 3: Supplemental Analysis   This appendix contains the supplemental analysis used in Chapter 3. Specifically,  1. Analysis on local fast-food restaurants 2. Analysis showing that weather variables are not linked to any specific food order 3. Analysis on Alberta and Ontario samples to replicate the findings of Maynard et al. (2008) 4. Analysis with different BSE media indices to show the robustness of findings to normalization of the media index 5. Analysis on the effect of avian influenza on all restaurants and on chicken-oriented restaurants     140   Table G-1. Beef-Oriented Brand, Local, and Local Fast-food Restaurants    National Brand Local Local – Fast-food Dine Out BSE Media -0.0019*** -0.0003 -0.0016 HH With Child 0.0536*** -0.2372*** -0.1155*** Uni. Degree -0.0249*** 0.0556*** -0.0167 HH Income -0.0102*** 0.0301*** 0.0132*** Age -0.0092*** 0.0006*** -0.0082*** Married 0.0411*** 0.0306*** 0.0122 Ontario -0.1078*** -0.0225*** 0.0171 Quebec -0.2811*** 0.0470*** 0.1618*** Atlantic Provinces -0.0789*** -0.0818*** -0.3084*** City Size -0.0113*** -0.0194*** 0.0152*** Mean Temperature -0.0002 -0.0010 0.0021 Max Temperature 0.0005 -0.0029*** 0.0010 Min Temperature 0.0019** 0.0046*** -0.0023 Total Rain Fall -0.0002*** -0.0001* -0.0002** Total Snow Fall 0.0003 0.0000 -0.0007** Linear Trend -0.0001 0.0002 -0.0008** Constant -0.6178*** -1.2089*** -1.7677*** Beef Orders BSE Media -0.0053** 0.0011 -0.0003 HH With Child 0.0564** -0.0725 0.3848*** Uni. Degree -0.1062*** -0.0256 0.1428* HH Income -0.0078** 0.0373*** -0.0484*** Age -0.0288*** 0.0062*** 0.0272*** Married 0.0938*** 0.0976*** 0.0701 Ontario -0.1910*** -0.1209*** 0.0036 Quebec -0.6824*** -0.2671*** -0.9875*** Atlantic Provinces -0.1312*** -0.2797*** 1.1894*** Linear Trend -0.0010** -0.0005 0.0056*** Constant -0.7369*** -1.4660*** 7.3208*** Rho 1.2709*** 0.0793 -2.5668*** Sigma 0.7763*** 0.5375*** 1.4465*** Sample Size 380004 380004 380004 Log Likelihood -165368.49 -189727.39 -21746.18 Note: * p < 0.1, ** p < 0.05, *** p < 0.01; Estimations based on DHEP method.     141  Table G-2. Weather Link to Food Orders  Beef Orders Chicken Orders Seafood Orders Pork Orders Vegetarian Orders Max Temperature -0.0273 0.0076 0.0007 -0.0058 0.0270 Min Temperature -0.0229 0.0006 -0.0012 -0.0170 0.0056 Mean Temperature 0.0508 -0.0081 0.0017 0.0230 -0.0312 Total Rain Fall 0.0000 0.0000 -0.0001** 0.0001 -0.0001 Total Snow Fall 0.0001 -0.0002 -0.0002 -0.0000 0.0002 Linear Trend -0.0010*** -0.0002 0.0004*** 0.0005*** -0.0019*** HH With Child 0.0308*** 0.0026 -0.1085*** -0.1628*** -0.1596*** Uni. Degree -0.0186*** -0.0336*** 0.0077** -0.0203*** 0.0315*** Married 0.0297*** -0.0270*** 0.0141*** 0.0573*** -0.0436*** Income 0.0059*** 0.0067*** 0.0110*** 0.0041*** 0.0243*** Alberta 0.0742*** 0.0760*** -0.0255*** 0.0070 -0.0368*** Saskatchewan  0.0731*** 0.0199 -0.0420*** 0.0356*** 0.0129 Manitoba 0.0878*** 0.0640*** -0.0247*** 0.0290*** 0.0730*** Ontario -0.0743*** 0.1187*** 0.0084 0.0054 0.0755*** Quebec -0.1616*** 0.1276*** -0.0318*** 0.0958*** 0.1612*** New Brunswick -0.0821*** 0.0706*** 0.0448*** -0.0170 0.0109 Prince Edward Island -0.0927** -0.1003*** 0.0176 0.2151*** 0.0988** Nova Scotia -0.1066*** 0.0691*** 0.0518*** -0.0659*** 0.0103 Newfoundland -0.0297 0.1689*** 0.0740*** 0.0369 0.0395 Constant 1.0478*** 0.4108*** -0.0893** -0.0749 1.2447*** Sample Size 112997 112997 112997 112997 112997 Log Likelihood -148582.17 -133272.25 -91118.54 -125246.54 -152218.68 Note: * p < 0.1, ** p < 0.05, *** p < 0.01; OLS Estimations 142   Table G-3. Mad Cow Effect in Ontario and Alberta  Ontario Alberta  Dine Out BSE Media -0.0018 -0.0020 HH With Child 0.0540*** 0.0979*** Uni. Degree 0.0182* -0.0069 HH Income -0.0060*** -0.0049* Age -0.0091*** -0.0082*** Married 0.0584*** 0.0949*** City Size -0.0156*** -0.0057 Mean Temperature 0.0005 -0.0065 Max Temperature -0.0010 0.0049* Min Temperature 0.0035* 0.0044** Total Rain Fall -0.0001 0.0001 Total Snow Fall 0.0008** 0.0009 Linear Trend -0.0002 0.0001 Constant -0.6887*** -0.9392***  Beef Orders BSE Media -0.0182*** -0.0107 HH With Child 0.0850** 0.2085*** Uni. Degree 0.0584 -0.0300 HH Income -0.0253*** -0.0136 Age -0.0285*** -0.0281*** Married 0.2358*** 0.2585*** Linear Trend 0.0010 -0.0032*** Constant -2.2454*** 0.0252 Rho 1.3495*** 1.3393*** Sigma 0.8182*** 0.8278*** Sample Size 137204 59812 Log Likelihood -57113.27 -30193.85 Note: * p < 0.1, ** p < 0.05, *** p < 0.01; Estimations based on DHEP method.      143    Table G-4. Media Index Analysis   Normalized Media by Total Number of Articles in Province (used in paper) Article Count Media Index Article Count Normalized by Number of Newspapers  Dine Out BSE Media -0.0009* 0.0003 0.0001 HH With Child -0.0425*** -0.0458*** -0.0456*** Uni. Degree 0.0202*** 0.0189*** 0.0190*** HH Income 0.0106*** 0.0106*** 0.0106*** Age -0.0028*** -0.0026*** -0.0026*** Married 0.0063 0.0048 0.0048 Ontario -0.0046 0.0033 0.0037 Quebec -0.0429*** -0.0407*** -0.0400*** Atlantic Provinces -0.0260*** -0.0145* -0.0137* City Size -0.0049*** -0.0024*** -0.0024*** Mean Temperature -0.0017 -0.0017** -0.0016** Max Temperature -0.0023*** 0.0009** 0.0009* Min Temperature 0.0029*** 0.0003 0.0003 Total Rain Fall -0.0000 -0.0000 -0.0000 Total Snow Fall 0.0001 0.0000 0.0000 Linear Trend 0.0003** 0.0003** 0.0003** Constant -0.4831*** -0.6004*** -0.6011***  Beef Orders BSE Media -0.0035* -0.0062*** 0.4439*** HH With Child -0.1412*** 0.4462*** -0.2631*** Uni. Degree -0.0355** -0.2618*** -0.0983*** HH Income 0.0209*** -0.0981*** -0.0181*** Age -0.0142*** -0.0181*** 0.0945* Married 0.1600*** 0.0955* -0.3024*** Ontario -0.3344*** -0.2928*** -0.6921*** Quebec -0.6429*** -0.6836*** -0.1022* Atlantic Provinces -0.3952*** -0.1041* -0.0027*** Linear Trend -0.0018*** -0.0027*** -0.0045* Constant -0.4631** 7.4717*** 7.4648*** Rho 0.6139*** -2.0765*** -2.0782*** Sigma 0.8178*** 1.8871*** 1.8883*** Sample Size 380004 380004 380004 Log Likelihood -352949.15 -308240.49 -308245.79 Note: * p < 0.1, ** p < 0.05, *** p < 0.01; Estimations based on DHEP method.     144  Table G-5. Avian Influenza Estimation for All Meal Items in All Restaurants    Probit DHEP DHEP                                               Dine Out AI Media -0.0019* -0.0017* -0.0013 HH With Child -0.0419*** -0.0424*** -0.0425*** Uni. Degree 0.0203*** 0.0202*** 0.0202*** HH Income 0.0107*** 0.0106*** 0.0096*** Age -0.0027*** -0.0028*** -0.0027*** Married 0.0068 0.0062 0.0057 Ontario -0.0033 -0.0042 -0.0034 Quebec -0.0453*** -0.0465*** -0.0485*** Atlantic Provinces -0.0247*** -0.0252*** -0.0121 City Size -0.0053*** -0.0048*** -0.0014* Mean Temperature -0.0021 -0.0019 0.0009 Max Temperature -0.0019*** -0.0022*** -0.0003 Min Temperature 0.0029*** 0.0029*** -0.0006 Total Rain Fall -0.0001 -0.0000 -0.0000 Total Snow Fall 0.0001 0.0001 -0.0001 Linear Trend 0.0004*** 0.0003*** 0.0003*** Constant -0.5454*** -0.5324*** -0.6209***   Beef Order Chicken Order AI Media  0.0006 0.0057 HH With Child  -0.1414*** 0.0580* Uni. Degree  -0.0355** -0.1898*** HH Income  0.0208*** -0.0343*** Age  -0.0142*** 0.0037*** Married  0.1599*** -0.0888** Ontario  -0.3348*** 0.2218*** Quebec  -0.6589*** 0.4418*** Atlantic Provinces  -0.3949*** 0.1105** Linear Trend  -0.0018*** -0.0014** Constant  -0.4333** 5.7489*** Rho  0.6135*** -2.3672*** Sigma  0.8176*** 1.7520*** Sample Size 380004 380004 380004 Log Likelihood -229966.31 -352950.13 -345343.45 Note: * p < 0.1, ** p < 0.05, *** p < 0.01     145  Table G-6. Chicken-Oriented Restaurants – BSE and Avian Influenza                                                                    Dine Out  BSE Media 0.0024***  AI Media  -0.0033** HH With Child -0.0300*** -0.0297*** Uni. Degree -0.0522*** -0.0521*** HH Income 0.0017 0.0017 Age -0.0014*** -0.0014*** Married 0.0093 0.0093 Ontario 0.2330*** 0.2346*** Quebec 0.3458*** 0.3584*** Atlantic Provinces 0.1689*** 0.1695*** City Size 0.0100*** 0.0098*** Mean Temperature 0.0055*** 0.0052*** Max Temperature 0.0002 0.0004 Min Temperature -0.0058*** -0.0058*** Total Rain Fall 0.0001 0.0001 Total Snow Fall -0.0001 -0.0001 Linear Trend -0.0009*** -0.0007*** Constant -1.4586*** -1.5726***                                                                   Chicken Orders  BSE Media -0.0040**  AI Media  -0.0009 HH With Child -0.0212 -0.0220 Uni. Degree 0.0365*** 0.0362** HH Income 0.0009 0.0008 Age -0.0015*** -0.0015*** Married 0.0130 0.0127 Ontario -0.1732*** -0.1739*** Quebec -0.3614*** -0.3812*** Atlantic Provinces -0.1487*** -0.1479*** Linear Trend 0.0004 0.0004 Constant 3.0914*** 3.0948*** Rho -1.3739*** -1.3747*** Sigma 0.1476*** 0.1481*** Sample Size 380004 380004 Log Likelihood -85134.28 -85138.11 Note: * p < 0.1, ** p < 0.05, *** p < 0.01; Estimations based on DHEP method.    

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
https://iiif.library.ubc.ca/presentation/dsp.24.1-0221500/manifest

Comment

Related Items