MIDWEST FARMERS DECISION - MAKING IN CONSERVATION AGRICULTURE ADOPTION By Qi Tian A DISSERTATION Submitted to Michigan State University in partial fulfillment of the requirements for the degree of Agricultural, Food, and Resource Economics Doctor of Philosophy 202 1 ABSTRACT MIDWEST FARMERS DECISION - MAKING IN CONSERVATION AGRICULTURE ADOPTION By Qi Tian Conservation Agriculture (CA) adoption can alleviate the environmental consequences of conventional agricultural producti on while maintaining yields . A decision - making in CA adoption is needed to inform policy design that encourages adoption . In the absence of the CA adoption market, experimental methods provide an essential alternative to in vestigate decision - maker Experiment (DCE) to analyz - making to shed light on policy desi gn as well as to inform methodological issues associated with DCE approach . The first chapter Willingness - to - Accept (WTA) CA practices and assesses the factors affecting the WTA. In addition to the payment to compensate the e xpenses or efforts of taking a CA practice, a substantial payment is needed to incentivize farmers leaving the status quo and commit ting to a CA program. Internal factors, such as f experience with CA practices , as well as external factors, i.e., policy design in terms of information framing and the decision time windo w, both have impacts on the WTA. These findings provide a practical guide for cost - efficient policy design. The traditional DCE approach for stated preference evaluation builds on a n essential assumption that decision - making is reference independent , i.e., independent of irrelevant alternatives . The second chapter develops a new framework to relax and test this assumption by incorporating b ehavioral realism into modeling. I found that decision - maker s use behavioral strategies, i.e., reference dependence, in decision making, and that different sources of information are evaluated differently as reference points. These findings , on the one han d , set caveats for modeling DCE data based on independence of irrelevance assumption , and on the other hand , indicate a more cost - efficient policy design tool that nudges desired behaviors through shaping the reference point. Th ree decision - making strategi es could describe the decision making in a DCE: reference independence, reference dependence, and attributes non - attend ance. This last chapter explicitly discusses which strategy is adopted and how such strategies evolve in repeated choice tasks. I found t hat decision - maker s use behavioral strategies to make decisions. As decision - maker s collect information over the repeat ed choice scenarios, they are shifting from the current choice set to the path as the reference point. Failing to account for the referen ce dependence behavior in choice modeling could misidentify the attended attributes as non - attended. This finding sugge sts that the reference dependence model can be a guiding choice for DCE modeling. Again, this chapter implies that discrete choice modeli ng without accounting for behavioral realism will fail to reveal the true preference. Copyright by QI TIAN 202 1 v ACKNOWLEDGEMENTS First and foremost, a huge THANK YOU to my adviser, Dr. Jinhua Zhao, for encouraging me to explore my curiosities, find ways to improve, conquer the challenges, and explore career possibilities. His keen insights about research, lea rning, and communication continuously inspire my work and life way beyond the research. I am deeply grateful for my wonderful committee members: Dr. Frank Lupi, for providing advice and critical opportunities to grow my stated preference research capabili ty; Dr. Vincenzina Caput o, for sharing her expertise and passion in choice experiment research; Dr. Joseph Herriges and Dr. Jeffrey Wooldridge for offering critical advice on econometrics for my dissertation and beyond. I am also thankful to the AFRE depa rtment for giving me thi s Ph.D. research opportunity and providing the best resource with the fabulous research talents, fellow graduate students, and departmental staff. I also thank my NSF CNH collaborators, Dr. Adam Reimer, Dr. Bruno Basso, Dr. Diana St uart, Dr. G. Philip Robe rtson, and Dr. Sandy Marquart - Pyatt, for working together in the interdisciplinary team to support my research dataset and make a better planet through environmental research. In addition, I would like to thank my friends for all t he time we spent togethe r. I could never go so far without you all being there for me. Finally, t o my mom and dad, for your love, understanding, and sacrifices, for modeling me living with no regret, being kind, optimistic, and fearless. I developed my dis sertation on the topic o f regret minimization in decision - making, and I love that topic a lot because that is also the motto my mom and dad taught me many years as I was growing up. vi TABLE OF CONTENTS LIST OF TABLES ................................ ................................ ................................ ....................... viii LIST OF FIGURES ................................ ................................ ................................ ....................... ix CHAPTER 1. INTRODUCTION ................................ ................................ ................................ ... 1 INFORMATION TREATMENT EFFECT AND COMMITMENT COST THEORY ................. 5 2.1. Introduction ................................ ................................ ................................ .......................... 5 2 .2. Survey and Data ................................ ................................ ................................ ................... 8 2.2.1. Survey and Experimental Design ................................ ................................ .................. 8 2.2.2. Hypothesis Testing ................................ ................................ ................................ ...... 17 2.2.3. Data ................................ ................................ ................................ .............................. 17 2.3. Estimation Procedures ................................ ................................ ................................ ........ 20 2.4. Results ................................ ................................ ................................ ................................ 24 2.4.1. WTA Estimation ................................ ................................ ................................ .......... 24 2.4.2. Treatment Effect Testing ................................ ................................ ............................. 28 ................................ ................................ ................ 31 2.5. Conclusions ................................ ................................ ................................ ........................ 33 REFERENCES ................................ ................................ ................................ ............................. 51 CHAPTER 3. REGRET MINIMIZATION IN DECISION - MAKING: IMPLICATIONS FOR CHOICE MODELING AND POLICY DESIGN ................................ ................................ ........ 58 3.1. Introduction ................................ ................................ ................................ ........................ 58 3.2. Methodology Foundations ................................ ................................ ................................ .. 60 3.3. Models ................................ ................................ ................................ ................................ 64 3.3.1. Traditional DCE Modeling ................................ ................................ .......................... 65 3.3.2. DCE Modeling in the Presence o f SQ Alternative ................................ ...................... 69 3 .3.3. DCE Modeling in the Presence of Path Dependence Behavior ................................ ... 71 3.3.4. Hypothesis Testing ................................ ................................ ................................ ...... 72 3.4. Survey and Data ................................ ................................ ................................ ................. 74 3.5. Res ults ................................ ................................ ................................ ................................ 78 3.6. Implications ................................ ................................ ................................ ........................ 86 3.6.1. The Implications for Policy Design ................................ ................................ ............. 86 3.6.2. The Impli cations for Discrete Choice Experiment ................................ ...................... 91 3.7. Conclusions ................................ ................................ ................................ ........................ 93 REFERENCES ................................ ................................ ................................ ............................. 96 CHAPTER 4. REGRET MINIMIZATION, PATH DEPENDENCE, AND ATTRIBUTE NON - ATTENDANCE IN DISCRETE CHOICE EXPERIMENTS ................................ .................... 102 4.1. Introduction ................................ ................................ ................................ ...................... 102 4.2. Econometric Frameworks and Hypothesis Testing ................................ .......................... 107 vii 4.2.1. Random Utility Maximization (RUM) ................................ ................................ ...... 108 4.2.2. Random Regret Minimization (RRM) ................................ ................................ ....... 109 4.2.2.1. Rand om Regret Minimization (RRM) ................................ ................................ 109 4.2.2.2. Random - RRM) .......... 111 4.2.2.3. Random Regr et Minimization in the Presence of Path Dependence (P - RRM) .. 112 4.2.3. Attributes Non - Attendance (ANA) ................................ ................................ ........... 113 4.2.4. Decision Rule Testi ng ................................ ................................ ............................... 114 4.3. Survey and Data ................................ ................................ ................................ ............... 115 4.4. Empirical Estimation Results ................................ ................................ ........................... 117 - RRM and P - RRM Estimations ................................ ................................ 118 - RRM, and P - RRM Estimations Accounting for the ANA Behavior ....... 124 4.5. Co nclusions ................................ ................................ ................................ ...................... 133 REFERENCES ................................ ................................ ................................ ........................... 136 CHAPTER 5. CONCLUSIONS ................................ ................................ ................................ . 141 APPENDICES ................................ ................................ ................................ ............................ 144 APPENDIX A. G - RRM REDUCES TO RUM WHEN ................................ ............... 145 - RRM REDUCES TO RUM WHEN ................................ .............. 146 APPENDIX C. P - RRM REDUCES TO RUM WHEN ................................ ................ 147 viii LIST OF TABLES Table 2.1. Treatment Groups ................................ ................................ ................................ ........ 14 Table 2.2. Expected Nitrogen Savings ................................ ................................ .......................... 16 Table 2.3. Attributes and Levels of the Choice Design ................................ ................................ 16 Table 2.4. Sample Characteristics in Percentage (%) by Treatment Sample ................................ 19 Table 2.5. WTA Space Estimations ................................ ................................ .............................. 27 Table 2.6. WTA Space Estimations by Treatment Sample ................................ .......................... 30 Table 2.7. Hypo thesis Tests ................................ ................................ ................................ .......... 31 Table ................................ 32 Table 3.1. Sample Characteristics by Stat e ................................ ................................ ................... 77 Table 3.2. Estimation Results ................................ ................................ ................................ ....... 79 Table 3.3. Hypothesis Test Results ................................ ................................ ............................... 80 Table 3.4. Factors Influencing Regret Profundity ................................ ................................ ........ 85 Table 3.5. Candidate Nudge Programs ................................ ................................ ......................... 88 Table 3.6. WTA ($) Indifferent Betw een Target and SQ Programs ................................ ............. 93 Table 4.1. RUM' Estimations ................................ ................................ ................................ ...... 121 Table 4.2. G' - RRM Estimations ................................ ................................ ................................ .. 122 Table 4.3. P - RRM Estimations ................................ ................................ ................................ ... 123 Table 4.4. RUM' Estimations with Inferred ANA ................................ ................................ ...... 127 - RRM Estim ations with Inferred ANA ................................ ................................ . 129 Table 4.6. P - RRM Estimations with Inferred ANA ................................ ................................ ... 131 ix LIST OF FIGURES Figure 2.1. Positive Information Treatment ................................ ................................ .................. 11 Figure 2.2. Negative Information Treatment ................................ ................................ ................ 12 Figure 2.3. Dynamic Decision Context ................................ ................................ ......................... 13 Figure 2.4. Static Decision Context ................................ ................................ .............................. 13 Figure 2.5. Survey Sample ................................ ................................ ................................ ............ 35 F igure 3.1. RUM and RRM Comparison ................................ ................................ ...................... 67 Figure 3.2. Regret Function Conditional on Regret Weight (gamma) ................................ ......... 69 Figure 3.3. Nesting Stuctur e of Models ................................ ................................ ........................ 73 ......... 88 doption Rate by State with Program 6 as the Nudge .................. 91 1 CHAPTER 1. INTRODUCTION Nitrogen leakage from farming systems brings significant ecological consequences, such as aquatic and marine eutrophication and greenhouse gas. C onservation Agriculture (CA) practices and tools have been developed to reduce nitrogen leakage while sustainin g yields , but the adoption rates remain strikingly low. To encourage CA adoption, a policy with a payment vehicle is needed to fill the gap betwe sion - making on CA adoption to facilitate policy design. - making, it is essential to understand farmers adoption decision - making analy sis is the lack of an existing market implementing a similar ernative for eliciting individual preference confronting such a challenge. However, a key concern when employin g experimental methods, such as Discrete Choice Experiments (DCE), is the incentive compatibility of the experiment. An experimental mechanism is considered incentive - compatible behavior in real life. Failing to incorporate behavioral realism in DCE modeling will conclude with preference evaluation that departures from th e real preference and affect the policy effectiveness. Therefore, this dissertation discusses how to design and model the DCE to reveal the real preference better. Besides preference evaluation, understanding the factors that affect the preference is criti characteristics and wi llingness to adopt a CA program will help identify the farmers who are 2 more willing to adopt a program given the same payment incentive. Learning the external way that incentivizes adoption better. There have been numerous literature justifying the internal vior, however, the studies that examine the relationship in the CA adoption scenario are too limited to guide p olicy design. To inform CA policy design, Chapter 2 evaluates two external factors of a policy design the framing of a policy and a decision - maki ng window. Through designing a two - by - two treatment DCE and examining the interaction effects of the two extern al factors, my analysis - making. Negative framing and a decision - making w Willingness - to - Accept (WTA) of the CA program. In addit ion to external factors, my analysis information on CA adoption. The results in this chapter provide empirical support for better designing of a CA policy. Both C hapter 3 and Chapter 4 build on the same experiment that evaluates Midwest he decision - making strategies in a DCE through developing and examining models that relax the independence of i rrelevant alternative assumption imposed on the DCE modeling framework . The assumption is alternative is independent of the other alternatives in the choice set. This assumption validates modeling DCE behavior through a Random Utility Maximization (RUM) framework. Chapter 3 develops a Random Regret Minimization (RRM) model that relaxes the assu mption imposed on the RUM framework. It allows behavioral decision - making strategies where decision - makers base their choice on a wish 3 to avoid the situation where the non - chosen alternative(s) perform(s) better than the chosen one . The analysis suggests t hat (1) decision - making is within as well as across choice sets dependent; (2) different sources of information are evaluated differently as reference points; and (3) the extent of individuals relying on the behavioral decision strategy is associated with their characteristics. These findings challenge basing on RUM framework to model DCE. In the meantime, policyma kers can make use of the reference dependence behavior to improve policy cost - efficiency by setting a reference point to nudge desired behavior. Chapter 4 extends from Chapter 3 to further assess the decision - making strategies and the changing patterns of these strategies over repeated choice scenarios in a DCE. Through examining the decisions in a DCE survey of four repeated choice scenarios, thi s study verifies that the reference dependence strategy is used and reveals an adaptive pattern of the dependen ce strategy. Decision - makers shift from the current choice set - dependent to path - dependent as they gradually collect information through repeate d choices. The attributes that are otherwise identified as non - attended under an attribute non - attendance frame work are attended in a path - dependent manner. These findings suggest that incorporating reference dependence behavior and correctly identifying t he reference points in DCE modeling are important for preference evaluation. In the meantime, designing an expe riment that perfectly matches the real - world decision scenario from the perspective of reference points will enable the DCE approach to evaluate real preference better. In summary, this dissertation provides empirical supports on CA policy design and discusses the modeling issues of the DCE approach. By properly framing the intention of a policy and setting the decision window, the policy adoption rate can be increased. T argeting the 4 thesis models and justifies behavioral decision strategies, such as reference dependence, loss aversion, and changing patt erns of decision strategi es in a DCE. These findings set caveats on using DCE to reveal preference as well as provide modeling solutions. These behavioral strategies can also be carefully used to facilitate policy design. 5 CHAPTER 2 . FARMERS DECISION ON NITROGEN APPLICATION: TES TING INFORMATION TREATMENT EFFECT AND COMMITMENT COST THEORY 2 .1. Introduction Nitrogen (N) leakage from farming systems brings significant ecological consequences, including water pollution and greenhouse gas emissions (e.g., Fe nn et al., 1998; Syswerda, 2009; EPA, 2011). Global N fertilizer use has increased by approximately ten folds since 1950 (Robertson et al., 2000). However, most crops only take up to 50% of the nitrogen applied, with the other 50% leaking to the environmen t (Syswerda, 2009; Smil, 1999). Therefore, Conservation Agriculture (CA) 1 practices and tools have been developed to reduce N appli cation and leakage, but their adoption rates remain strikingly low (Ribaudo et al., 2011; Osmond et al., 2015). The main reas on for the low adoption of CA is the divergence between social benefits and individual interests: the incremental costs associated with adopting CA accrue at the farm level, while the beneficial environmental effects are captured by the society (Ma et al., 2012). A policy that motivates the CA adoption with production regulation, taxes, or payment incentives are needed to fill the gap . A large body of empirical literature on CA adoption has been developed to facilitate policy design, but there is still no u niversal conclusion on how sociological, economic, and w, 2007; Prokopy et al., 2008; Baumgart - Getz et al., 2012). Among these factors, the financial incentive has been found to be the most crucial motive for CA adoption, though it is not the single factor (e.g., Cary and Wikinson, 1997; Chouinard et al., 2007 ; Stuart et al., 2012; Honlonkou, 2004; Claassen and 1 C onservation Agriculture (CA) is a term defined by the Food and Agricultural Organization of the United Nations - saving agricultural crop production that strives to achieve acceptable profits together with high and sust ai 6 Horan 2000). Internal factors associated with farmers and farms such as levels (e.g., Okoye, 1998), farm size and profitability (e.g., Lambert et al., 2007), household income (e.g., So environmental issues (e.g., Mase et al., 2015; Napier and Camboni, 1993; Traoré et al., 1998; Reimer et al., 2012; Stuart et al. 2012) all have impacts on CA adoption. Besides these int ernal factors, external factors associated with the proposed programs such as information communication (Opdam et al., 2015; Hong and Zinkhan, 1995; Chernev, 20 04; Florack and Scarabis, 2006 ) can also shape CA adoption. External factors study can directly contribute to improving polic y cost - efficiency, but the related research regarding CA adoption is far from sufficient to inform policy design. To shed light on external factors the framing of a policy and the deci sion window, and investigates how these two factors jointly affect the decision - making. I perform a discrete choice experiment (DCE ) with two - by - U.S. Tversky and Kahneman (1981) has shown that framing can affect decision - making and lead to violations of classic axioms of rational choice. In environmental ec onomics, some works suggest that disclosing environment externalities to the consumers can be effective at shifting conservat ion preferences (Thaler and Sunstein, 2008), while other research demonstrates that disclosing potential environmental benefits can encourage conservation behavior among the general public (Opdama et al., 2015; Maibach et al., 2008; Myers et al. 2012; Kran tz and Monroe 2016; Stevenson et al., 2018). However, to my best knowledge, none of this previous literature has studied the effect 7 though environmental service provider is the basis of managing environmental issues. practit ioners, and policymakers design a more cost - effective policy to encourage CA adoption, and in parallel, help elucidate the mo tives for shifting farm management. This paper studies how the framing of the goal of policy affects environmental service provider versions of framing are discussed and compared in this study: positive policy framing emphasizes the positi ve environmental contribution as well as the N efficiency of CA adoption, and negative policy framing focuses on the damaging envir onmental externalities of no - action. The decision window is another essential factor in policy design. In stated preference e xperiments, it is commonly assumed that decision - making is in a static setting without uncertainty. Real - world choices, however, ar e more often made in dynamic settings with uncertainty. Under a dynamic scenario, individuals can delay decisions to obtain f urther knowledge or even reverse the choice (Arrow and Fisher 1974 ; Dixit and Pindyck, 1994 ). Zhao and Kling (2001, 2 004) investigate the quasi - Willingness - to - Pays (WTP) divergence between static and dynamic welfare measures. They found that committing to a decision at the moment give up the opportunity to learn more about the value of an option if a deci sion is made today. Decision - making in dynamic settings has been investigated fr om the WTP perspective in environmental economics (e.g., Arrow and Fisher, 1974; Corrigan et al., 2008), finance (e.g., Dixit and Pindyck, 1994), and consumer behavior (e.g., Castaño et al., 2008; Bazzani et al., 2017). However, this concept has rarely bee n investigated from the Willingness - to - Accept - 8 making can manifest in different ways for consumers and producers. The uncertainty for consumers arises as they do not have complete information about the product, whereas producers have more knowledge of what to expect. Meanwhile, once a decision is made, consumers can immediately explor e all the attributes of the product they have purchased, while producers, on the As such, this paper makes the first try to investigate the learning effect on farmer decisions by providing two decision - making windows. The static dec ision window only gives decision - makers one time to enroll in the program. In contrast, the dynamic decision window offers a second chance to vote and thus allows for informat ion collection and learning. The paper is organized as follows: section 2 introd uces the survey design, survey procedure, and the survey sample characteristics; model specification and estimation procedure are presented in section 3; empirical application results are reported in section 4; I close the discussion and point out directio ns for future research in section 5. 2 .2. Survey and Data 2 .2.1. Survey and Experimental Design The target population of the research is corn growers in the Midwestern U.S., specifically Michigan, Iowa, and Indiana. I chose corn growers because corn is t he most widely planted crop in the U.S. and the largest user of nitrogen fertilizer in terms of application rates per acre, total acres treated, and overall applications 2 . The Midwest was chosen because corn production is most centered in this region, wit h Iowa ranking first in sales of corn (McCurry, 2014). The survey distribution was through mailing, and the survey participants and mailing addresses were 2 According to ERS 2012, around 50% of N fertilizer in the US is applied to corn, 11% to wheat, 10% to turf, and 3% to cotton. 9 drawn from farmers g rowing corn in 2013 from the Farm Service Agency (FSA) 3 address book. Two focus groups with a total of 20 farmers were conducted in 2015 to derive an understanding he questionnaire. A pilot study with a sample of 300 was conducted in 2015 to e stimate the range of the payment as well as to test the design of the questionnaire. For the final survey, 1600 corn growers were randomly drawn from each state with 4 , 800 corn growers in total. I used a modified Tailored Design Method survey design (Dillm an et al., 2014), with two questionnaire mailings and a reminder mailing. An initial contact containing a cover letter, survey questionnaire, and $2 cash incentive was sent out in March 2016, followed by a reminder card one week later. A second questionnai re mailing without incentives was sent out in April 2016 to those who had not responded. To test the effects of framing and the decision window on decision - making, I designed a two - by - two treatment survey and used a between - subject approach following Lusk and Schoeder (2004) , in which each survey respondent will be assigned to one of the four treatment designs . I am interested in how these two factors jointly affect decision - mak ing because in real choice - making these two factors may interact with each othe r . For testing the decision window effect, the respondents from the no - delay option group were notified that whether the program will be implemented depends on their one - time vo te. In contrast, the respondents from the delay option group were informed that they would have a second chance to vote one year later if the program does not pass this year so that they can collect more information to make a decision. This information was highlighted right before each choice question to ensure the information is eff 3 FSA is the payment services agency within USDA. FSA has r ecords for every farmer who receives any form of payment (direct payments, crop insurance subsidies, disaster payments, conservation payments, etc) through USDA. 10 ntrast, the oducing the nitrogen efficiency and environmental benefit of CA adoption. The details of the two treatments are listed in Figure 2. 1 Figure 2. 4. The static/dynamic decision ti me treatments and two information treatments combine into four treatment groups shown in Table 2. 1. I randomly assigned the respondents to one of the four treatment groups. 11 Figure 2. 1. Positive Information Treatment 12 Figure 2. 2. Negative Information Trea tment 13 Figure 2. 3. Dynamic Decision Context + Figure 2. 4. Static Decision Context 14 Table 2. 1. Treatment Groups Positive Information Dynamic Decision Window Treatment 1 (PD) Treatment 2 (PS) Treatment 3 (ND) Treatment 4 (NS) The main body of the mail survey consists of four parts and a preamble. Before the start of the choice experiment, the preamble provides definitions of the terms referred to in the choice setting and as ks whether the survey respondents have heard of these terms to ensure they clearly understand the proposed practice s. Part I contains four choice experiment tasks with two - - ch oice (Batsell perceived s opinions about farming and the environment. Part IV gathers socio - demo graphic information. - adoption. An example survey is attached in Figure 2. 5. Within the choice set, each hypothetical alternative is described by three CA practices, an expected nitrogen saving, and a payment vehicle. The three CA practices are avoidance of fall nitrogen appl ication, i.e., Fall , side - dressing N fertilizer, i.e., Si de , and covering crops in winter, i.e., Winter . I chose these three because they are the most efficient methods of improving N use efficiency without adding operation costs, as compared to other alte rnatives (Osmond et al., 2015; Christianson et al., 2014) Systems Approach to Land Use Sustainability (SALUS) model (Basso et al. 2012) and the 15 specific formula is reported in Table 2. the top of each decision scenario. Note that the survey respondents may not perceive hypothetical a nd SQ alternatives in the same way because SQ is the endo wment alternative and the practice levels are not saliently presented in the choice scenario. The range of the payment is based on Natural Resources Conservation Service (NRCS) payment levels for N m anagement, and is adjusted based on focus group study and pilot survey estimation. Attributes and attributes levels are defined in Table 2. 3. The three proposed CA practices will potentially reduce N usage by 5% to 50% without affecting yields if adequate ly handled. This point is noted before each choice set in the questionnaire. In the meantime, such a program carries benefits as well as costs and risks to the farmer. The benefits include program participation payment, nitrogen fertilizer saving, and soil protection. According to Schnitkey (2015), the gross rev enue of corn growing per acreage is $804, with a net return of $194, and the associated cost of fertilizer is $161. Therefore, even without the payment program, there is still an incentive for CA ado ption in a hope to save fertilizer cost. On the other ha nd, side - dressing fertilizer and covering crops in winter add to farm management costs. Prohibiting fertilizer application in fall and covering crops in winter are perceived to be risky by many farme rs due to challenges in managing the timing of spring pla nting operations (Arbuckle and Roesch - McNally, 2015). 16 Table 2. 2. Expected Nitrogen Savings Winter Cover Crops Required Fall Application Prohibited Sidedress Application Required Nitrogen Savings Yes Yes Yes 50% No Yes Yes 25% Yes No Yes 40% Yes Yes No 25% Yes No No 10% No Yes No 10% No No Yes 25% Table 2. 3. Attributes and Levels of the Choice Design Attributes Levels Winter Cover Crops Required Yes, No Fall Application Prohibited Yes , No Sidedress Application Required Yes, No Expected Nitrogen Savings 0%, 10%, 25% 40% 50% Annual Payment/Acre $0, $5, $20, $40, $100, $180 A Bayesian design that minimizes D - error based on priors from the pilot survey was used to create the choice s ets with variation in attributes levels. First, an orthogonal fractional factorial design generated by SAS was used for the pilot survey. Next, a conditional logit model was applied to analyze the pilot survey data, and the estimates served as the priors f or the fi nal design. Last ly , a Bayesian efficiency design that minimizes D - error based on priors from the pilot survey and contains 24 choice sets , is generated using Ngene software (Choice Metrics, 2012) . I used a block design with six blocks containing f our choic e sets for each to avoid fatigue effects. Respondents were randomly assigned to one of these blocks. The order of presentation and allocation to respondents of the various choice sets is randomized. Examples of the choice tasks are attached in Fig ure 2. 3 a nd Figure 2. 4. 17 2 .2.2. Hypothesis Testing Hypothesis testing is conducted to investigate if there is a significant difference between treatments for WTA. The two - by - two treatments combine four hypotheses tests. Hypothesis 1 and hypothesis 2 are in troduced to test the effect of decision window cond itional on a specific framing approach ; hypothesis 3 and hypothesis 4 are added to test information treatment conditional on decision window setting . The subscripts of WTA, i.e., P, N, D, S, represent , respect ively. Hypothesis 1: Hypothesis 2: Hypothesis 3: Hypothesis 4: 2 .2.3. Data The survey experiment achieved a response rate of 31%. After removing returns with more than 10% of the questions incomplete, I have 1,140 usable retu rns of the survey, with response rates as 30%, 26%, 22%, respectively, for Michigan, Iowa, and Indiana. Table 2. 4 reports the farmer characteristics by the four treatments. 97% of respondents are male. 94% survey respondents are above 35 years old, with ab out 30% for each age group (35 - 54, 55 - 64, 65 an d above). 96% of farmers have above high school or higher education (96%). 18 Half respondents have more than 36 years of experience with farming. About 63% of the respondents are full - time farmers working off - fa rm less than 49 days a year. 59% of respondents operate small to medium farms (less than 500 acres), with 64% of farmers obtained product values more than $100,000 in the year 2015. About 61% of respondents have ever enrolled in a conservation program. I e treatment group to generate an idea about the randomization of the sample. A chi - square test, as reported in Table 2. 3, was conducted to see if there are any unbalanced characteristics associated with the tr eatment group. The results suggest that the nul l hypothesis of equality between the socio - demographic characteristics across treatment samples cannot be rejected at the 5% significance level for all variables except for conservation group enrollment. This result confirms that the randomization is succe ssful in equalizing the characteristics of respondents across treatments. 19 Table 2. 4. Sample Characteristics in Percentage (%) by Treatment Sample Sample PD (n=286) PS (n=279) ND (n=284) NS (n=291) All (n=1 ,140) State IA 36 32 34 32 33 IN 26 30 27 32 29 MI 38 38 40 36 38 Pearson chi2(6) = 3.6579 Pr = 0.723 Gender Female 4 3 3 3 3 Male 96 97 97 97 97 Pearson chi2(3) = 0.6667 Pr = 0.881 Age Betwe en 18 - 34 7 4 6 5 6 Between 35 - 54 27 29 29 28 28 Between 55 - 64 32 32 30 31 31 Above 64 34 35 36 36 35 Pearson chi2(9) = 3.5161 Pr = 0.940 Education Elementary School 5 4 5 2 4 High School or Some College 48 50 48 53 50 University 47 46 48 45 47 Pearson chi2(6) = 4.5376 Pr = 0.604 Years of farming experience Less than 37 years 50 45 50 47 48 Longer than 36 years 50 55 50 53 52 Pearson chi2(3) = 1.6459 Pr = 0.649 Days off - far m Great than 50 days/year 37 39 36 37 37 Less than 49 days/year 63 61 64 63 63 Pearson chi2(3) = 0.4826 Pr = 0.923 Acres of farm operated in 2015 Less than 100 acres 42 44 39 40 41 Between 100 - 500 acres 32 33 36 35 34 Greater than 500 acres 27 23 25 26 25 Pearson chi2(6) = 2.5865 Pr = 0.859 Total values of products sold in 2015 Less than $100,000 44 32 34 36 36 $100,000 - $499,999 35 42 42 43 41 $500,000 - $999,999 13 14 14 12 13 Greater than $1,000,000 8 11 10 9 10 Pearson chi2(9) = 9.7982 Pr = 0.367 Annual Household Income Low (up to $25,000) 12 13 10 11 12 20 Medium ($25,000 - $100,000) 66 62 62 68 65 High (above $100,000) 22 25 28 21 24 Pearson chi2(6) = 5.4099 Pr = 0.492 Conservation program enrollment Never enrolled 54 62 67 60 61 Ever enrolled 46 38 33 40 39 Pearson chi2(3) = 9.7944 Pr = 0.020 2 .3. Estimation Procedures The choice experiment a pproach was initially developed by Louviere and Hensher (1982) and Louviere and Woodworth (1983) and constructed upon the Random Utility Theory (Lancaster 1966; McFadden 1974). In a random utility framework, the researcher observes s ome attributes of the a lternatives, but other components of the individual utility are unobservable and treated as composed of the attributes that describ e the alternative, as w ell as an alternative specific constant that represents the opt - out, i.e. SQ alternative, and a stochastic term. The utility for individual i from alternative j within choice task t can be expressed as: where is the systematic portion of the utility function which depends on the experimentally designed attributes for alternative j , and represents the unobserved random/stochastic term. Assumptions regarding the functional form of and the distribution of are required in order to transform the random utility model into a choice model. To capture unobserved preference hete rogeneity and correlation across repeated choices, this paper uses a Random Parameter Logit Model with Error Component (RPL - EC) and the utility is specified in WTA space for the estimation (Scarpa and Alberini 2005; Scarpa, Ferrini, 21 and Willis, 2005; Hess and Rose, 2009; Thiene and Scarpa, 2009; Scarpa et al., 2008). To that of Pay , follow random normal distributions. Fixing the payment coefficient ensures that the estimated WTAs are normally distributed, and all re spondents have a positive coefficient for s parameter (Train, 1999). It is worthwhile to mention that the SQ alternative is experienced by respondents, while the hypothetical alternatives can only be conjectured. As such, the utilities of hypothetical alternatives are more correlated within themselves than with the SQ option. Besides, as the SQ alternative is familiar to respondents, it is likely to have a smaller individual valuation error. With th ese considerations, an error component stochastic term was included in the model to account for the systematic effects associated with the SQ and the hypothetical alternatives. The advantage of specifying the model in WTA space is that the coefficients can directly be interpreted as marginal WTA values. Beside s, WTA space estimation is a more feasible approach when comparisons across treatments are made than that based on marginal utilities, i.e., preference space estimation (Caputo, Scarpa, and Nayga, 2017 ). The utility of RPL - EC model in WTA space for a resp ondent i choosing alternative j at choice task t is specified as: where represents the payment farmers receive from the proposed program; , , are three proposed practices with 1 indicating that this practice is 22 required, and 0, otherwise; is the expected nitrogen saving in percentage; is a dummy variable with value 1 indicating the SQ alternative and 0, otherwise; is the payment parameter; are the coefficients of the estimated WTA values; error component is normally distributed with zero mean for the hypothetical alternatives, and for the SQ alternative; and is the unobserved error term which follows a Gumbel distribution. Note that here I estimated the model with panel structure, assuming that the error components are the same for all choice sets by the same individual following the suggestion by Scarpa et al., 2007. Also note that, in addition to setting the SQ constant term dummy to contribution to uti lity, here I allow the preferences for the same attribute to be different between hypothetical and SQ alternatives by separately specifying and . A formal testing of the assumption tha t = is reported in the later empirical analysis. Moreover, I allowed the taste parameters to be interdependent by assuming the coefficients of attributes follow a multivariate normal distribution. With this setting, I allowed the prefer ences for the proposed practices to be correlated. For example, I might expect some correlation between preferences for Winter and Side. The multivariate normal distribution has vector mean , and variance - covariance matrix , where C is the Cholesk y matrix. The significa nce of any element of Cholesky matrix C will support the dependence across tastes. It is worth to mention that an SQ constant term was included in equation (2) to explore the potential divergence between hypothetical alternatives an d SQ alternative. Leavi ng the status quo situation has been justified to decrease utility in many DCE applications (e.g., Lehtonen et al. 2003; Hanley, Wright, and Alvarez - Farizo 2006). However, a crucial question for these findings is whether the SQ const ant term is capturing t he average effect on utility of all factors not included in the model, or whether the SQ constant term is associated with a behavioral decision 23 strategy, such as misperceived sunk costs or regret aversion (e.g., Meyerhoff and Liebe, 2009). Therefore, this et al., 2008. In this way, the SQ constant term captures the pure effect of staying at status quo because the individual specific status quo situations have been controlled in the data. function including a set of dummy va riables following de - Magistris 2013, and Bazzani 2017. The data was pooled based on one of the two treat ments so that the dummy variables can test the specified as follo ws: where is a dummy variable indicating the tre atment group, i.e., information or decision window; represent the respe ctive treatment effect on the specific attribute. Note that dummy variables are included within the brackets and minus signs are specified such that the coefficients 24 can be directly interpreted as the difference in WTAs. A total of four extended utility fu nctions are estimated based on the four different dimensions of pooled data. an an Ordinary Least Square model by specifying the individual posterior estimates of WTA as the dependent var iables and the respondent characteristics as the independent variables following Train, 2009: where represents the individual WTA for attribute A estimated from equation (2); represents social - demographic factors, opinions, and information sources. 2 .4. Results This section discusses the estimation res ults of the RPL - EC models in WTA - space and the extended utility RPL - EC models to test the treatm WTA and their social demographic characteristics, attitude, and source of information are also examined. 2 .4.1. WT A Estimation I started with investigating the assumption that = through running two specifications of RUM model. The RPL - EC model estimation results in WTA space are reported in Table 2. 5 with RUM assuming = are as expected except that of Nit rogen , and the estimations of Fall, Side, Nitrogen are not Winter, Side and SQ being significant. The unexpected signs and insignificant estimations of RUM imply that restricting 25 = might fail to capture the true underlying decision criteria. Furthermore, the likelihood ic being 122. Th is rejection again suggests that the preferences over the same attribute in the hypothetical and SQ alternative are practices are un favored, payment and nitrogen saving are favored, and leaving the status quo is unfavored. The statistical significance of the Moreover, the signifi cance of the err or component for the alternative specific constants justifies the hypothesis of correlation across the hypothetical alternatives. The WTA estimations for are most significant among the three CA practices in both hypothetical a nd SQ alternativ es, i.e., $87/acre and $91/acre. This finding is consistent with my expectation as covering winter crops is most costly and time - consuming. WTA estimations for Fall are small and insignificant. With respect to WTA of Side , the estimation is small and insig nificant in hypothetical alternative, however, is significant and relatively high in SQ alternative, i.e., $151/acre. N saving has a positive impact on preferring an alternative : a 10% expected nitrogen saving decrease s the WTA of a hypoth etical alternative by $13/acre and the WTA of an SQ alternative by $29/acre . Although the estimations are not significant. Referring to the costs of fertilizer for corn, which is approximate ly $1 61/acre (Schnitkey, 2015), the scale of estimation in the bon us gained from expected fertilizer saving is reasonable. However, we should be cautious in interpreting the parameters of Here is the expected nitr ogen saving calculated by assuming the farmer currently does not take any of the CA practices and applies 26 170lb/acre fertilizer on the land. It is not the actual nitrogen saving that will occur after the proposed practices are adopted. We should cautiously interpret the coefficient of to be the effect of the information related with expected nitrogen saving, rather than the effect of nitrogen saving. In addition to that, a crucial point to note here is as is hig hly correlated with the three CA practices, be significantly identified if sample size is not sufficiently large. In light of these two points, we did not find shows as significant impacts as we might expect. Shifting farmers away from the status quo significantly brings u p the WTA by about m. Program acceptance aversion is a significant issue in conservation program promotion. In addition to the SQ dummy term, SQ and hypothetical alternatives also dive preference/WTA. Generally, the scales of WTAs for the SQ alternativ es are larger and more significant than for the hypothetical alternatives. That means that if a farmer has adopted a certain practice in the SQ, he/she needs higher incentive to stay at SQ, or lower incentive to switch to the CA program because he/she will gain policy payment by simply committing to the proposed program without changing the practices he/she has adopted. It is worth mentioning that sample size can be a limitation from which this research suffers. On one hand, lifting the assumption of = increases the required sample size to identify the model parameters as the number of parameters doubled. On the other hand, the correlated wi th Winter , F all , and Side . This high correlation raises the required sample size to power the identification of parameters. 27 Table 2. 5 . WTA Space Estimations RUM Mean Values 77*** a 87*** (14.3 ) b (17.3) 1 1 ( 8.7) (9.9) 13 2 (20.6) (23.7) 39 - 128 (84.2) (98.6) 91*** (30.4) 28 (20.9) 151*** (52.9) - 292 (208) - 90*** - 145*** (4.9) (10.5) Standard Deviations 117*** 93*** (37.7) (25.2) 6 3 (16.2) (4.9) 32 4 (45.8) (5.3) 69 151 (122) (119) 107*** (25.5) 42 (39.2) 147*** (23.8) 301* (174) 114*** 108*** (13.3) (11.7) Model Statistics Log - Likelihood - 4495 - 4434 AIC/N 2.09 2.07 BIC/N 2.10 2.08 N 4300 4300 a. p - value with *** 1% , ** 5%, * 10%. b. Standard error is reported in the bracket. 28 2 .4.2. Treatment Effect Testing To examine the treatment effect, I started with rerunning the model based on each treatment group . The results in Table 2. 6 show that the individual treatme are consistent with the findings based on the pooled sample in the previous section: the WTAs for the same attributes in an SQ alternative are higher than that in a hypothetical alternative. In PS an d NS groups, the parameters of Sid e and Nitrogen are extremely high and trading off with each other, which again brings attention to the high correlation between Nitrogen and the three CA practices . Comparing the WTAs among the treatment groups, I found th at the WTA estimations are relativ ely lower for the ND group, and relatively higher for the PS group. To continue, I formally tested the difference in WTAs by running the extended utility equation (3) to make pairwise comparison s. The results are reported in Table 2. 7. To test the decisio , I pooled the positive information observations to run equation (3) . As the extended part for the treatment dummy parameters are also included in the WTA space, the estimated pa rameters can be directly int erpreted as the difference in dollar amount of WTA. Likewise, I separately pooled negative information treated observations, delayed option available observations, and delayed option unavailable observations to test , and respectively . With respect to the three CA practices, ND group generates the lowest WTA estimations and PS group generates the highest WTA estimations. With respect to N saving and staying at SQ, ND group produces the highest WTA estimations and P S group produces the lowest WTA estimations. The scale of the difference in WTA between treatment groups is substantial referring to the scale of WTAs; however, the estimation of the difference is not significant. The statistical insignificance holds us fr om claiming a winner among the treatment groups. However, it is worth mentioning again that 29 sample size is a limitation of this work to identify the parameters significantly. On that note, this treatment testing suggests that negative information framing c ombined with a dynami c decision window has the potential to reduce program costs as compared with the other three treatments. In summary, information framing treatment and decision window setting may have impacts on decision - making. Negative information f raming combined with a dynamic decision combined with a static decision scenario may prevent people from adoption. That being said, people are more effectively nudged by the negative consequ ence of taking no action to conduct good deeds as compared with the positive contributions they could have made. Meanwhile, people feel more confident to commit to a program immediately if they are provided with the option to delay dec isions to collection information rather than being pushed to make a decision immediately. These findings are merely based on raw WTAs comparison and are not statistically justifiable due to the sample size limitation. 30 Table 2. 6. WTA Space Estimations by Treatment Sample Sa mple PD PS ND NS Mean Values 45 a 90.4*** 6 2 .7** 77.1** (72.2) b (30.8) ( 28.0 ) (32.7) - 3.49 14.2 - 14.7 - 0.902 (22.2) (17.5) (20.8) (19) 0.66 13.7 - 18.5 4.63 (52.8) (42.3) (49.7) (44.7) - 153 - 132 - 113 - 95 (220) (175) (208) (186) 106*** 134** 73.6** 144** (38.9) (54.8) (36.1) (66.2) - 69 57.2 9.29 79.3* (49.7) (37.1) (38.5) (46.2) 71.2 250*** 64.7 308*** (124) (95.1) ( 96) (117) 579 - 745** 103 - 884** (501) (369) (379) (458) - 153*** - 123*** - 161*** - 143*** (24.7) (17.6) (22.8) (19.4) Standard Deviations 61 116*** 7 3 * 91** (77.8) (42.1) (37.1) (40.1) 7. 9 32 14 5.3 (9.1) (29.7) (100) (4.7) 2.7 25 24 7.0 (7.5) (37.1) (106) (9.5) 179* 154 116 106 (107) (137) (89.5) (95.7) 121*** 131*** 84*** 160** (10.7) (21.2) (23.3) (63.2) 82 74 13.9 82.2** (92.2) (57.5) (12.8) (33.4) 85 185*** 78 327*** (39.6) (57.5) (69.3) (118) 741 866*** 123 895** (637) (289) (95.4) (394) 102*** 94*** 93*** 108*** (17.1) (15.2) (16.5) ( 13.7) Model Statistics Log - Likelihood - 1094 - 1101 - 1121 - 1100 AIC/N 2.08 2.07 2.07 2.07 BIC/N 2.12 2.12 2.12 2.12 N 1064 1072 1092 1072 a. p - value with *** 1%, ** 5%, * 10%. b. Standard error is reported in the bracket. 31 Table 2. 7. Hypothesis Tests Wint er Fall Side Nitrogen SQ - 3 0 . 2 a - 19.4 - 36.1 136 12.1 s.d. (41.3) b (24.2) (59.1) (243) (14.2) - 20.2 - 17.1 - 49.2 140 13.8 s.d. (40. 5) (24.4) (58.5) (240) (40.5) 39.2* 6.31 15.9 - 51.8 - 1.63 s.d. (22.1) (15.8) (43.3) (260) (15.3) 11.3 8.38 0.722 - 33.5 2.10 s.d. (38.1) (22.8) (54.4) (224) (13.1) a. p - value with *** 1%, ** 5%, * 10%. b. Standard error is reported in the bracket. 2 - making as discussed above , I further tested - making. As the WTA estimations are larger and more significant for the SQ alternative, I ran the posterior estimation s of equation (4) separately on the WTA s for Winter , Fall , Side , Nitrogen , and SQ . As reported in Table 2. 8 , , and , b ut do not have significant impacts on , and . Looking further into the factors, I found that states, gender, education, days off - farm, experience with conservation program s , and trust in neighbors and media as farming information so urces all have impacts on , , and . Generally, Iowa farmers need higher compensation than Michigan farmers; male, higher educated, full - time, conservation program experienced, trusting neighbors more an d media less farmers need more compensation to enro ll in the proposed program. 32 Table 2. 8 s on WTA of the SQ Alternative Factors Winter Fall Side Nitrogen SQ State IA - 3 a 4 38 *** 34 6 ( 10 .1) b (3. 7 ) ( 10.6 ) ( 32.3 ) ( 7.2 ) IN 22 2 24 25 * 17 ( 18.4 ) ( 5.7 ) ( 13.6 ) ( 14.5 ) ( 10.1 ) Gender Male 28 *** 10 22 5 7 * 39 (1 1.2 ) ( 9.2 ) ( 17.8 ) ( 31.1 ) ( 25.6 ) Age Senior (>59 years) 9 3 - 6 7 6 ( 8.8 ) ( 3.1 ) ( 4.9 ) ( 9.2 ) ( 5.7 ) Education College Degree or Above - 8 - 3 2 - 26 * 1 4 ** ( 8.1 ) ( 2.0 ) ( 3.8 ) ( 14.1 ) ( 6.2 2 ) Days off Farm Full - time (less than 50 days off farm) - 8 - 6 29 ** 15 3 ( 6.3 ) ( 5.4 ) ( 12.3 ) ( 11.6 ) ( 4.2 ) Farm Value High Product Value (>$499,999) 3 7 5 - 11 2 ( 3.8 ) ( 6.1 ) ( 3.9 ) ( 7.3 ) ( 4.2 ) Conservation Program Has Experience wi th Conservation Program 25 * 31 - 20 5 - 12 ( 13.3 ) ( 22.6 ) ( 15.2 ) ( 6.2 ) ( 7.5 ) Information Source Trust in Neighbor > 2.5 c 14 * - 1 21* - 38 - 1 0 ( 7.4 ) ( 1.5 ) ( 10.9 ) ( 27.0 ) ( 7.2 ) Trust in Extension > 2 6 1 2 3 - 2 ( 8.2 ) ( 2.0 ) ( 3.4 ) ( 5.2 ) ( 3.2 ) Tr ust in Private Sector > 3 1 3 4 2 4 ( 3.2 ) ( 4.8 ) ( 2.8 ) ( 3.3 ) (4. 1 ) Trust in Media > 2 - 27 *** - 5 7 - 8 1 ( 7.2 ) ( 5.3 ) ( 5.7 ) ( 6.2 ) ( 1.5 ) Trust in Online Calculator > 1 - 2 - 9 - 4 - 7 6 ( 3.4 ) ( 7.8 ) ( 3.4 ) ( 6.3 ) ( 5.7 ) Constant 79 *** 1 9 138 * - 304 - 152 *** ( 31.6 ) ( 12.8 ) ( 81.7 ) (259) ( 35.2 ) N 686 686 686 686 686 F 3.1 3 1. 4 4 4. 52 3. 74 1. 25 Prob > F 0.0002 0. 1312 0.0000 0.0000 0. 2298 R - squared 0.0 41 8 0.0 32 9 0.0 702 0.0 548 0.02 14 a. p - value with *** 1%, ** 5%, * 10%. b. Standard error is reported in the bracket. c. The median value for a 1 - 5 liker scare question is used to generate the binary variable. 33 2 .5. Conclusions This chapter investigated - making for C A adoption through conducting a survey experiment analysis . I estimated the Incentives are effective at encouraging farmers to adopt CA practi ces . The costs come in two parts: one is the direct practice costs that com pensate the added expenses or efforts of taking a particular practice, the other is the program enrollment cost that compensates for leaving the status quo. Among the three CA practices, covering crop s in winter is relatively significant and expensive. Thi s is consistent with the difficulty level of adopting each practice. Besides that, the unwilling ness to change from the status quo also plays a critical part in the compensatio n. This unwilling ness to leave the status quo can be due to 1) the endowment eff ect wherein people ascribe more values to the status quo merely because they have been endowed with it ; 2) concern of the incurred transaction cost ; 3) aversion or distrust of a regulated/government program ; 4) aversion of commitment ; 5) risk aversion wher e people avoid making changes to take any risk ; or 6) protest of the survey. Understanding the reasons for unwillingness to leave the status quo will inform policymakers taking actions to remove the associated concerns and thus improving policy cost effici ency. These tasks are out of the scope of this dissertation and are suggested for future research directions. Two factors effectively reduce the necessary compensations of CA adoption . One is the one hand, taking CA practices in the status quo will naturally lower the utility level of staying at status quo, making people more likely to commit to a CA program. On the ot her hand, the WTAs for the same CA attributes are generally higher and more sign ificant in the SQ alternative than in the hypothetical alternatives. 34 This greater WTAs for the SQ alternative further lowers the CA - taken - staying at S Q. This finding raises the importance of building CA experience among farmers. T he other factor that reduces necessary compensation is the expected N savings, even though this expected N saving might depart from the actual N saving depending on the survey SQ fertilizer application. This indicates the importance of includi ng the expected nitrogen saving as part of the program design to incentivize adoption. Furthermore, this study tested the interaction effect of the policy design from two dime nsions , i.e., the information framing and a decision time window . The two - by - two treatment test suggests that the negative information framing combined with a dynamic decision scenario is t efficiency. However, further research with sufficient sample size should be co nducted to provide statistical evidence for these findings. also showed impacts on program enrollment . T argeting the program among factors wi th program - favor features can potentially increase the adoption rate without inc reasing the policy budget. Understanding the reasons and causal effects of why each factor contributes to their decision will also generate insights on how to encourage partici pation. 35 Figure 2. 5. Survey Sample 36 Figure 37 Figure 38 Figure 39 Figure 40 Figure 41 Figure 42 Figure 43 Figure 44 Figure 45 Figure 46 Figure 47 Figure 48 Figure 49 Figur e 50 Figure 51 REFERENCES 52 REFERENCES Arbuckle Jr, J.G., Morton, L.W. and Hobbs, J., 2015. Understanding farmer perspectives on climate change adaptation and mitigation: The roles of trust in sources of climate infor mation, climate change beliefs, and perceived risk. Environment and behavior, 47(2), pp.205 - 234. Arbuckle, J.G. and Roesch - McNally, G., 2015. Cover crop adoption in Iowa: The role of perceived practice characteristics. Journal of Soil and Water Conservatio n, 70(6), pp.418 - 429. Arrow, K.J. and Fisher, A.C., 1974. Environmental preservation, uncertainty, and irreversibility. In Classic papers in natural res ource economics (pp. 76 - 84). Palgrave Macmillan, London. Artell, J., Ahtiainen, H. and Pouta, E., 2013. Subjective vs. objective measures in the valuation of water quality. Journal of environmental management, 130, pp.288 - 296. Banzhaf, M.R., Johnson, F.R. and Mathews, K.E., 2001. Opt - preferences. The choice modelling appr oach to environmental valuation. Edward Elgar, London, pp.157 - 177. Barton, D.N. and Bergland , O., 2010. Valuing irrigation water using a choice experiment: an Environment and Development Eco nomics, 15(3), pp.321 - 340. Basso B, Sartori L, Cammarano D, Grace P, Sorensen C, Fountas S. 2012. Environmental and economic evaluation of N fertilizer rates in a maize crop in Italy: a spatial and temporal analysis using crop models. Biosystems Engineerin g 113: 103 - 111. Batsell, R.R. and Louviere, J.J., 1991. Experimental analysis of choice. Mar keting letters, 2(3), pp.199 - 214. Baumgart - Getz, A., Prokopy, L.S. and Floress, K., 2012. Why farmers adopt best management practice in the United States: A meta - an alysis of the adoption literature. Journal of environmental management, 96(1), pp.17 - 25. Baz zani, C., Caputo, V., Nayga Jr, R.M. and Canavari, M., 2017. Testing commitment cost theory in choice experiments. Economic Inquiry, 55(1), pp.383 - 396. Caputo, V., Scarpa, R. and Nayga Jr, R.M., 2017. Cue versus independent food attributes: the effect of a dding attributes in choice experiments. European Review of Agricultural Economics, 44(2), pp.211 - 230. 53 Carson, R.T., Louviere, J.J., Anderson, D.A., Arabie, P., Bunc h, D.S., Hensher, D.A., Johnson, R.M., Kuhfeld, W.F., Steinberg, D., Swait, J. and Timmerman s, H., 1994. Experimental analysis of choice. Marketing letters, 5(4), pp.351 - 367. Cary, J.W. and Wilkinson, R.L. . 1997 . vation Behaviour . Journal of Agricultural Economics 48 ( 1 3 ): 13 21 . Castaño, R., Sujan, M., Kacker, M. and Sujan, H., 2008. Managing consumer uncertainty in the adoption of new products: Temporal distance and mental simulation. Journal of Marketing Researc h, 45(3), pp.320 - 336. Chernev, A., 2004. Goal - attribute compatibility in consumer choice. Ch ouinard, H.H., Davis, D.E., LaFrance, J.T. and Perloff, J.M., 2007, June. Fat taxes: big money for small change. In Forum for Health Economics & Policy (Vol. 10, No . 2). De Gruyter. Christianson, L., Knoot, T., Larsen, D., Tyndall, J., and Helmers. 2014. A doption potential of nitrate mitigation practices: an ecosystem services approach. International Journal of Agricultural Sustainability 12(4): 407 - 424. Claassen, R. and R.D. Horan. 2000. Environmental Payments to Farmers: Issues of Program Design. Agricult ural Outlook, June - July, pp. 15 - 18. Corrigan, J.R., Kling, C.L. and Zhao, J., 2008. Willingness to pay and the cost of commitment: an empirical specification and t est. Environmental and Resource Economics, 40(2), pp.285 - 298. Corrigan, J.R., Kling, C.L. and Zhao, J., 2008. Willingness to pay and the cost of commitment: an empirical specification and test. Environmental and Resource Economics, 40(2), pp.285 - 298. de - Ma gistris, Tiziana, Azucena Gracia, and Rodolfo M. Nayga. 2013. On the use of honesty p riming tasks to mitigate hypothetical bias in choice experiments. American Journal of Agricultural Economics, 95(5): 1136 - 1154 Dillman, D.A., Smyth, J.D., and Christian, L .M. 2014. Internet, Phone, Mail, and Mixed - Mode Surveys: The Tailored Design Method. Wiley: Hoboken, N.J. Dixit, A.K., Dixit, R.K., Pindyck, R.S. and Pindyck, R., 1994. Investment under uncertainty. Princeton university press. Domínguez - Torreiro, M. and So liño, M., 2011. Provided and perceived status quo in choice experiments: Implications for valuing the outputs of multifunctional rural areas. Ecological Economics, 70(12), pp.2523 - 2531. 54 EPA (U.S. Environmental Protection Agency). 2011. Reactive nitrogen in the United States: an analysis of inputs, flows, consequence, and management options . Report EPA - SAB - 11 - 013. Washington, D.C., USA. Fenn ME, Poth MA, Aber JD, Baron JS, Bormann BT, Johnson DW, Lemly AD, McNulty SG, Ryan DF, Stottlemyer R. 1998. Nitrogen e xcess in North American ecosystems: Predisposing factors, ecosystem responses, and ma nagement strategies. Ecological Applications 8: 706 733. Florack, A. and Scarabis, M., 2006. How advertising claims affect brand preferences and category brand association s: The role of regulatory fit. Psychology & Marketing, 23(9), pp.741 - 755. Hanley, N., Wright, R.E. and Alvarez - Farizo, B., 2007. Estimating the economic value of improvements in river ecology using choice experiments: an application to the water framework directive. In Environmental value transfer: Issues and methods (pp. 111 - 130). Springe r, Dordrecht. Hess, S. and Rose, J.M., 2009. Should reference alternatives in pivot design SC surveys be treated differently?. Environmental and Resource Economics, 42(3), pp.297 - 317. veness: The influence of congruency, conspicuousness, and response mode. Psychology & Marketing, 12(1), pp.53 - 77. Honlonkou, A.N., 2004. Modelling adoption of natural resources managem ent technologies: the case of fallow systems. Environment and Developmen t Economics, 9(3), pp.289 - 314. and synthesis of recent research. Food policy, 32(1), pp.25 - 4 8. Krantz, S.A. and Monroe, M.C., 2016. Message framing matters: Communicating climate change with forest landowners. Journal of Forestry, 114(2), pp.108 - 115. Lambert, D.M., Sullivan, P., Claassen, R. and Foreman, L., 2007. Profiles of US farm households a dopting conservation - compatible practices. Land Use Policy, 24(1) , pp.72 - 88. Lancaster, K.J., 1966. A new approach to consumer theory. Journal of political economy, 74(2), pp.132 - 157. Lehtonen, E., Kuuluvainen, J., Pouta, E., Rekola, M. and Li, C.Z., 2003. Non - market benefits of forest conservation in southern Finland. Environmental Science & Policy, 6(3), pp.195 - 204. 55 Louviere, J.J. and Hensher, D.A., 1982. On the design and analysis of simulated choice or allocation experiments in travel choice modelling. Transportation research record, 890(1982), pp.11 - 17. Louviere, J. J. and Woodworth, G., 1983. Design and analysis of simulated consumer choice or allocation experiments: an approach based on aggregate data. Journal of marketing research, pp.350 - 367. Lusk, J .L. and Schroeder, T.C., 2004. Are choice experiments incentive c ompatible? A test with quality differentiated beef steaks. American Journal of Agricultural Economics, 86(2), pp.467 - 482. ces programmes. Journal of Agricultural Economics, 63(3), pp.604 - 626. Maibach, E.W., Roser - Renouf, C. and Leiserowitz, A., 2008. Communication and marketing as climate change intervention assets: A public health perspective. American journal of preventive medicine, 35(5), pp.488 - 500. Mase, A.S., Cho, H. and Prokopy, L.S., 2015. Enhancing the Social Amplification of Risk Frame belief in climate change. Journal of Environmen tal Psychology, 41, pp.166 - 176. McCurry, M., 2014. Census of Agriculture 2012. McFadden, D., 1974. The measurement of urba n travel demand. Journal of public economics, 3(4), pp.303 - 328. Metrics, C., 2012. Ngene 1.1. 1 user manual & reference guide. Sydney, Australia: Choice Metrics. Meyerhoff, J. and Liebe, U., 2009. Status quo effect in choice experiments: Empirical evidence on attitudes and choice task complexity. Land Economics, 85(3), pp.515 - 528. Myers, T.A., Nisbet, M.C., Maibach, E.W. and Leiserowitz, A.A., 2012. A public health frame arouses hopeful emotions about climate change. Climatic change, 113(3 - 4), pp.1105 - 1112. Napier, T.L. and Camboni, S.M., 1993. Use of conventional and conservation practices among farmers in the Scioto River basin of Ohio. Journal of Soil and Water Conservation, 48(3), pp.231 - 237. Okoye, C.U., 1998. Comparative analysis of factors in the adop tion of traditional and recommended soil erosion control practices in Nigeria. Soil and Tillage Research, 45(3 - 4), pp.251 - 263. 56 Opdam, P., Coninx, I., Dewulf, A., Steingröver, E., Vos, C. and van der Wal, M., 2015. Framing ecosystem services: Affecting beha viour of actors in collaborative landscape planning?. Land use policy, 46, pp.223 - 231. Osmond, D., D.L.K. Hoag, A.E. Luloff, D.W. Meal nutrient management: lessons from watershed case studies. Journal of Environmental Q uality 44: 382 - 390. Prokopy, L.S., Floress, K., Klotthor - Weinkauf, D. and Baumgart - Getz, A., 2008. Determinants of agricultural best m anagement practice adoption: Evidence from the literature. Journal of Soil and Water Conservation, 63(5), pp.300 - 311. Reim er, A.P., Thompson, A.W. and Prokopy, L.S., 2012. The multi - dimensional nature of environmental attitudes among farmers in Indiana: im plications for conservation adoption. Agriculture and human values, 29(1), pp.29 - 40. Ribaudo, M., Hansen, L., Livingston, M., Mosheim, R., Williamson, J. and Delgado, J., 2011. Nitrogen in agricultural systems: Implications for conservation policy. Roberts on, G.P., Paul, E.A. and Harwood, R.R., 2000. Greenhouse gases in intensive agriculture: contributions of individual gases to the radiative forcing of the atmosphere. Science, 289(5486), pp.1922 - 1925. Rose, J.M., Bliemer, M.C., Hensher, D.A. and Collins, A .T., 2008. Designing efficient stated choice experiments in the presence of reference alternatives. Transportation Researc h Part B: Methodological, 42(4), pp.395 - 406. Scarpa, R. and Alberini, A. eds., 2005. Applications of simulation methods in environment al and resource economics(Vol. 6). Springer Science & Business Media. Scarpa, R., Ferrini, S. and Willis, K., 2005. Perfor mance of error component models for status - quo effects in choice experiments. In Applications of simulation methods in environmental a nd resource economics (pp. 247 - 273). Springer, Dordrecht. Scarpa, R., Thiene, M. and Train, K., 2008. Utility in willingne ss to pay space: a tool to address confounding random scale effects in destination choice to the Alps. American Journal of Agricultural Economics, 90(4), pp.994 - 1010. Scarpa, R., Willis, K.G. and Acutt, M., 2007. Valuing externalities from water supply: St atus quo, choice complexity and individual random effects in panel kernel logit analysis of choice experiments. Journal of Environmental Planning and Management, 50(4), pp.449 - 466. Schnitkey, G., 2015. Revenue and costs for corn, soybeans, wheat, and doubl e - crop soybeans, actual for 2009 through 2014, projected 2015 and 2016. Department of Agriculture a nd Consumer Economy, University of Illinois, Champaign, IL, USA. 57 Smil, V., 1999. Nitrogen in crop production: An account of global flows. Global biogeochemic al cycles, 13(2), pp.647 - 662. Somda, J., Nianogo, A.J., Nassa, S. and Sanou, S., 2002. Soil fertili ty management and socio - economic factors in crop - livestock systems in Burkina Faso: a case study of composting technology. Ecological economics, 43(2 - 3), pp. 175 - 183. Stevenson, K.T., King, T.L., Selm, K.R., Peterson, M.N. and Monroe, M.C., 2018. Framing cl imate change communication to prompt individual and collective action among adolescents from agricultural communities. Environmental Education Research, 24(3 ), pp.365 - 377. Stuart, D., Schewe, R.L. and McDermott, M., 2012. Responding to climate change: Barr iers to reflexive modernization in US agriculture. Organization & Environment, 25(3), pp.308 - 327. Syswerda, S.P., 2009. Ecosystem services from agriculture a cross a management intensity gradient in southwest Michigan. Michigan State University. Thaler, R. and Sunstein, C., 2008. Nudge: The gentle power of choice architecture. New Haven, Conn.: Yale. Thiene, M. and Scarpa, R., 2009. Deriving and testing efficie nt estimates of WTP distributions in destination choice models. Environmental and Resource Economic s, 44(3), p.379. Train, K.E., 2009. Discrete choice methods with simulation. Cambridge university press, p. 262 - 285. Traoré, N., Landry, R. and Amara, N., 19 98. On - farm adoption of conservation practices: the role of farm and farmer characteristics, percep tions, and health hazards. Land economics, pp.114 - 127. Tversky, A. and Kahneman, D., 1981. The framing of decisions and the psychology of choice. science, 21 1(4481), pp.453 - 458. Zhao, J. and Kling, C.L., 2001. A new explanation for the WTP/WTA disparity. E conomics Letters, 73(3), pp.293 - 300. Zhao, J. and Kling, C.L., 2004. Willingness to pay, compensating variation, and the cost of commitment. Economic Inquiry , 42(3), pp.503 - 517. 58 CHAPTER 3 . REGRET MINIMIZATION IN DECISION - MAKING: IMPLICATIONS FOR CHOICE M ODELING AND POLICY DESIGN 3 .1. Introduction Discrete Choice Experiment (DCE) is a survey - based economic approach (developed by Louviere and Hensher, 1982, a nd Louviere and Woodworth, 1983) for eliciting individual preferences. It is increasingly used in n on - market valuation to elicit environmental preferences (Louiviere et al., 2000; Kanninen, 2007; Carson and Groves, 2007). Respondents are provided with hypo thetical choice scenarios, and each choice scenario contains multiple alternatives, usually more th an two, including a Status Quo (SQ) or an opting - out alternative. Each alternative is described by a combination of attributes. The respondents are asked to choose one alternative from each choice scenario in order to elicit their preferences of the goods to be valued. Compared with the binary - alternative (referendum) contingent valuation method, DCE provides more information from a single choice due to its mu ltiple - alternatives setting. Besides, the DCE allows value examination of individual attributes, in addition to the value of the whole package estimated with the contingent valuation method. Disregarding the advantages of the DCE method, an essential is sue with this stated preference method is incentive compatibility. A mechanism is incentive compati ble if truth revelation is best for all participants (Myerson, 1979). DCE applications have long implicitly assumed this approach is truth revelation (e.g., Louviere and Woodworth, 1983). However, increasing studies investigate the hypothetical bias of the experimental approach, where the hypothetical scenarios may fail to generate the same responses as the real scenarios do, and discuss experimental design so lutions to reduce the hypothetical bias (e.g., Carlsson and Martinsson, 2001; Lusk and Schroeder, 2 004; Carson and Groves, 2007; Taylor et al., 2010; 59 Rakotonarivo et al., 2016). However, there has been limited research studying how the multiple - alternative from the true pr eference regardless how cautiously the experiment is designed to incentivize decision - making compatible with real - world behavior. Vossler et al. (2012) sugge sts that a single binary DCE combined with a consequentiality condition is incentive compatible. An derson et al. (2007) finds that a multiple price list auction provides simple incentives for truthful revelation, and this auction mechanism collapses to a b inary choice under certain conditions. Carson and Groves (2007) suggests that expending the choice set to multiple alternatives and/or repeated choice tasks would violate incentive compatibility property of DCE. Besides incentivizing decision - makers to be have consistently in experimental and real - world settings, a more fundamental issue is how to model the decision behaviors to reveal the decision - n experimental setting or a real - world sett ing. The Random Utility Maximization (RUM) framework, where the DCE approach has been built on, may fail to depict the true decision behavior because it makes a naïve assumption that decision - making is rational an d ignores the critical alternative behavior al decision realism. New models have been developed to relax these assumptions (e.g., Swait, 2001; Arentze and Timmermans, 2007; Kivetz et al., 2004; Zhang et al., 2004), but these models are mostly less interpret able than the RUM framework. An exception i s Random Regret Minimization (RRM), which has been recently proposed by Chorus (2008, 2010) to incorporate the behavioral features while still inheriting the RUM interpretable estimation framework. This paper dev elops a new Random Regret Minimization mode l, i.e., Path Dependent Random Regret Minimization (P - RRM) model, that relaxes the assumptions imposed on traditional RUM and RRM. Thereby, we can understand the decision rules through hypothesis 60 testing. Similar to the existing RRM models, this P - RRM mode l allows for reference dependence and loss aversion behavior. Different from the existing RRM models, this P - RRM allows for different impacts from the status quo and hypothetical alternatives as Reference Points ( RPs). Meanwhile, this P - RRM model allows fo r reference dependence behavior not only within but also across choice sets. Finally, this paper uses a Willingness - To - Accept (WTA) context DCE data on Conservation Agriculture (CA) practice adoption among corn g rowers in the Midwest U.S. to examine the m odel empirically 4 . These findings also have applications in analyzing decision making outside of the environmental economics area. This study rejects the assumption that decision making is choice set independent, supporting reference - dependent behavior. Be sides, I found that hypothetical alternatives, individual status quo, and previous choice sets can all affect decision making as RP. This finding implies that mimicking the real choice scenario, in te rms of composition of alternatives, can be a guiding app roach way to support the incentive compatibility of DCE. Lastly, due to reference - dependent behavior, policymakers can nudge their desired choice by strategically altering the choice set composition. 3 .2. Methodology Foundations RUM is the dominant estima tion strategy in the context of DCE. The fundamental assumption of RUM is that the utility derived from an alternative is a function of its attributes with the objective of utility maximization. This RUM framework presumes that decision - makers evaluate eac h alternative independently to maximize utility. However, there has been rich evidence of reference - dependent decision making, where people make choices based on pairwise comparison 4 Conservation Agriculture (CA) is a term defined by the Food and Agricultural Organization o f the United Nations - saving agricultural crop production that strives to achieve acceptable profits 61 and classify the c omparison outcomes as losses or gains. The loss aversion emotion leaves individuals more sensitive to losses than to gains generated from the bilateral comparison (see Prospect Theory, Kahneman & Tverskey, 1979; Regret Theory, Bell, 1982; Fishburn, 1982; L oomes and Sugden, 1982). Therefore, the utility (or regr et, as more often named in regret minimization literature) of an alternative relies on attributes of its own as well as the referred alternatives. Whether the reference - dependent behavior disturbs th e incentive compatibility property of DCE depends on if the RP is endogenous of the choice experiment. If the RP is endogenous of the choice experiment, the decision making will be contingent on the experiment design. In such a case, DCE will be incentive - compatible only if the experiment provides the exactly s ame choice is affected by the prices experienced outside the DCE, such as previous shopping (Caputo et al., 20 18; Tonsor, 2018). An example of endogenous RP is an ind specific alternative is contingent on the whole choice sets design, which is the case I will discuss in the paper. RRM (Chorus, 2008, 2010) is developed to incorporate this endogenous choice set reference behavior. Decision makin g is endogenous of the experiment because the choice sets composed of the experiment are the RPs. RRM inherits two critical points from Regret Theory and Prospect Theory. One is reference - dependent, w hich indicates that decision making is based on a binary comparison of the chosen alternative and forgone alternatives. The other is loss aversion, which implies that decision - makers do not want their foregone alternative to perform better than their chose n alternative. Under the RRM framework, losses and gains are generated from the binary comparison, and a new utility function form is developed to measure the 62 asymmetry of weighting in those losses and gains. Besides that, RRM models are similar to their R UM counterparts and can be estimated with the existing e conometric models for RUM. Due to the advantages in capturing behavior features with no extra estimation requirement by following the framework of RUM, RRM has gained wide attention from literature i n fields such as transportation, marketing, and environm ental economics (e.g., Hensher et al., 2013; Thiene et al., 2012; Chorus and Bierlaire, 2013; Chorus et al., 2014; Boeri et al., 2014; Adamowicz, Glenk, and Meyerhoff 2014). Comparison studies between RRM and RUM have been conducted, and model performance difference between the two models, in terms of statistical criteria, is small and dataset dependent (see Chorus, 2014 for a review of comparison between the two models). However, in terms of choice pr obability or market share prediction, the difference can be large and lead to a substantial difference in policy implications (Chorus et al., 2014). Therefore, choosing between the models has a notably practical impact. Besides model comparison and select ion, another research direction is to incorporate differ ent decision paradigms within a single model. Hess et al. (2012) and Boeri et al. (2014) use a Latent Class Model (LCM) to allow individuals to apply heterogeneous information processing strategies. C horus et al. (2013) uses a hybrid model to allow attribu tes to be processed by different decision rules. Van Cranenburgh et al. (2015 ) and Chorus (2014) generali ze the classical RRM and RUM models by allowing a parame ter to decide between RRM and RUM. Other RRM related works include model adaptation to choice scenarios including SQ alternative, i.e., opt - out, (Thiene et al., 2012; Hess et al., 2014), welfare analy sis (e.g., Dekker & Chorus, 2018), and choice set effici ency design (Van Cranenburgh et al., 2018). Despite the fruitful development of RRM in the past decade, some fundamental issues are still unsolved. To begin, even though RRM was developed to fill the gap between RUM 63 assumptions and decision behavior reali sm, there is not a framework designed to investigate the realism of decision behavior. This study extends the RRM framework to examine the realism of decision behavior by constructing a framework that nests the RUM and RRM assumptions in a single model. Sp ecifically, I relaxed the assumptions that decision making is both within and across choice sets dependent. I discussed how the relaxation of the assumptions affects the incentive property of DCE and choice modeling analysis. To continue, the existing RRM literature does not discuss how to handle SQ alternative in choice modeling. A direct approach is to replace the SQ alternative with individual perceived values or homogeneous no - action if perceived values are not available. This approach does not treat S Q alternatives differently than other alternatives. However, RRM literature with an SQ setting based on this approach results in weaker performance in terms of statistical power (Thiene et al., 2012; Hess et al., 2014; Chorus, 2012). This phenomenon limits the application of the RRM framework in more expansive research fields. This paper, for the first time, develops a framework to handle SQ alternative by distinguishing SQ alternative from hypothetica l alternatives in regret generation. This new framework will not only improve model performance but also help to understand the decision mechanism in the presence of SQ choice. Finally, RRM does not provide direct welfare analysis implications as RUM does (e.g., pecification. This is the first paper to quantify the impacts of accounting for behavioral factors on welfare analysis and policy implementation through a simulation approach. I found that relaxing th e decision - making assumptions leads to a remarkable diff s chosen probability. - To - Accept (WTA) context as opposed to a Willingness - To - Pay (WTP) co ntext on which almost all the other RRM 64 literature is ba sed. To my knowledge, the only relevant work examining reference dependence behavior in the context of WTA is Tonsor (2018), which examines producer decisions within a DCE. This study completes the li terature in understanding the roles behavioral strategie s (e.g., reference - dependent, regret aversion) play in a WTA scenario. The remainder of the paper is organized as follows. Section 3 constructs models to account for different decision rules and dis cusses how to use hypothesis tests to examine the underl ying assumptions. The survey and data used for empirically examining the models are introduced in section 4. Empirical estimation results and hypothesis tests are reported in section 5. Section 6 disc usses the implications of incorporating behavioral facto rs for policy design and discrete choice experiments. The paper closes with a summary and discussion of the findings and future directions in section 7. 3 .3. Models This section discusses how to mo del discrete choices under different decision rule assum ptions. There are three components of DCE: choice scenario and sets of alternatives, a function that describes the observed utility, and an error term that describes the unobserved utility and the ass ociated distribution. In this paper, I am interested in adding regret minimization behavior to the choice modeling by modifying the function that describes the observed utility or regret. To start, I introduced the existing models, which are Random Utilit y Maximization (RUM), Random Regret Minimization (RRM), and Generalized Random Regret Minimization (G - RRM). Next, I discussed how to model the regret minimization behavior in a DCE with the SQ setting. I am interested in understanding whether the SQ altern ative has an equal contribution as the hypothetical alte rnative does in serving as an RP. Third, I developed a new RRM model 65 that relaxes all the assumptions imposed by the existing DCE modeling and discussed how to investigate the underlying behavioral ru les with this new model. 3 .3.1. Traditional DCE Modeli ng Presume a regular choice scenario: a decision - maker, i , faces a choice scenario, s , with J alternatives, each being described in terms of M attributes . RUM (McFadden, 1974) postulates that utility from alternative j is independent of alter natives k , i.e., independent of choice set composition. A decision - maker will choose the alterna tive with the highest utility from the given choice set. The random utility of each alternative is described by a linear combination of the observable attribute s plus a random error term . The random error term represents the inability of researche rs to observe all the factors determining a decision - utility, i.e., the unobserved heterogeneity among decision - makers. An individual i choosing alternative j wit h taste parameters can be described as follows: Or An SQ constant term is added to capture the status quo effect. is a dummy variable: when j is the SQ option and otherwise. preferences ov er attributes for hypothetical alternatives and SQ alternative are differe nt, and therefore I separately specify the corresponding preference parameters as and . I name 66 = conventional RUM . Under the assumption that error term follows independent and identically distributed, i.i.d . Extreme Value Type I with variance equaling , multinomial logit (MNL) model can be used for est imation (McFadden, 1974). The choice probability for al ternative j is: . For the same choice scenario, RRM framework (Chorus, 2008, 2010) postulates that a decision - maker i will choose the alte rnative with the lowest regret from the given choice set, and the regret is composed of a systematic regret described by the ob served attribute and an i.i.d random error . Regret is generated when the considered alterna tive is outperformed by the competing alternatives within the choice set with respect to any attribute. Note that this setting presumes that an i aversion, regardless of whether a tangible c Following the work of Quiggin (1994), Chorus (2010) defines the regret for alternative j att ribute m by bilaterally comparing with alternative k as follows: 5 . Figure 3. 1 illustrates the regret as a function of loss X = as compared with RUM. This function of regret posits that individuals respond more to loss (X > 0) than to gain (X < 0) due to the conv exity of the log function. This setting also presumes that - level regrets, i.e., m , with all RPs, i.e., k ive j : . An individual i 5 An alternative r eg ret minimization model is defined in Chorus (2008) as follows: , where is the preference parameter of attribute m. This formulation implies that gain generates zero weight in formulating regret: when a consid ered alternativ e outperforms its competing alternative, i.e., , the regret is zero. 67 alternative j from choice scenario s will be the observable regre t plus a random error defined as follows: Figure 3. 1. RUM and RRM Comparison Note that the minimization of regret is mathematically equivalent to maximizing the negative of th e regret defined in equation (2). As a result, the SQ constant term parameter has the opposite sign of that from the conventional RUM model. Assuming that follows i.i.d Extreme Value Type I, multinomial logit can be used for model estim ation with the probability of choosing alternative j over other alternatives defined as follows: . Note that in the case of a single binary choice where a choice set contains two alternative s, RRM reduces to linear RUM (see Chorus, 2010, for a formal proof). 68 RRM framework posits a non - substitution behavior. That is, the ratios of preference parameters can no longer measure the marginal rates of substitution. As such, a decrease in one attribu te may not be compensated by an equal increase in another attribute. In addition, an alternative with in - between attribute values generates lower regret than those with mo re extreme attribute values, in which some attributes have very high values while oth ers have very low values. With two frameworks, i.e., RUM and RRM, available for choice modeling, the question is how to choose between the two frameworks. Therefore, Gener alized Random Regret Minimization (G - RRM) model (Chorus, 2014) is constructed to nest RRM and RUM as special cases and allow the model itself to test the underlying decision rule(s). The G - 1 6 . The G - RRM is defined as follows: ( is a regret weight parameter that depicts the curvatu re of the regret function. Figure 3. 2 describes how the regret function responds to the change of regret weight parameter . When = 1 , equation (3) will be the conventional RRM model defined in equation (2); when = 0 , the G - RRM model generates the s ame prediction as a RUM model does (see Appendix A and Chorus, 2014 for a formal proof). As approaches zero, the asymmetry on loss and gain vanishes; as increases, so does the asymmetry. Note that is arbitrarily s et to be between 0 and 1 because the curvature of the regret line is getting less sensitive to the value of as increases. Removing the upper bound of is likely to confound the estimation of with since is 6 By replacing with , I can assume different curvatures for the attributes as proposed in Chorus (2014). In this paper, I assume a single curvature for the reg ret function. 69 trading off with with respective to estimation (Chorus, 2014). La stly, let , where is an individual characteristic, I can measure the relationship between regret weight and individual factors (Chorus, 2014). Figure 3. 2. Regret Function Conditional on Regret Weight (gamma) 3 .3.2. DCE Modelin g in the Presence of SQ Alternative An essential setting of DCE is the existence of a baseline alternative SQ, i.e., opt - out alternative. This setting avoids forced choice of the proposed alternatives and thus guarantees proper welfare measures (Hanley et al., 2001). There has been substanti al literature discussing the importance of including an SQ alternative specific constant term in DCE modeling to account for the endowment effect the fact that people demand more to give up an object they process than th ey would be willing to acquire it (T haler, 1980; Samuelson and Zeckhauser, 1988; Adamowicz et al., 1998). However, there is no literature investigating the role of an SQ alternative in DCE when there exists reference dependence behavior. This paper will di scuss this issue. 70 As has been discu ssed in the RRM section, giving up a hypothetical alternative that outperforms the considered alternative for a particular attribute can generate a loss feeling. Similarly, foregoing the SQ that performs better than the considered alternative for a particu lar attribute can generate loss as well. Prior studies have recognized that decision - makers may have their SQ as a critical RP (Kahneman & Tverskey, 1979; Tversky & Kahneman, 1981, 1986, 1991). The question is how to mea sure the impact of SQ as RP in DCE. To examine the role of SQ option as RP in DCE, this paper sets SQ as an RP and allows the SQ to generate an impact that is different from the hypothetical alternatives. The rationale of allowing the difference is SQ is endowed by the decision - makers and t hus serves as an internal reference, whilst hypothetical alternatives are imposed by choice set design and therefore serves as an external reference. Furthermore, different from the studies that examine reference effects by asking the respondents of their reference prices before decisions are made (Mazumdar et al., consistent with real - world settings since, in most cases, the re is no chance of explicitly remind ing decision - makers of their SQ before they make a choice. Lastly, different from the existing works that focus on the cost variable as the single reference element (e.g., Caputo, et al., 2018), this paper investigates t he contributions of all attributes t hat describe the SQ as the reference of decision making. ndividual i from choosing alternative j is specified as follows. I named the model - RRM to indicate its modificati on from the previous G - RRM model. 71 Assuming that follows i.i.d Extreme Value Type I, an MNL model can be used for model estimation. - RRM model is, preference parameters and based on comparison with different RPs can be different. That is, for a considered hypothetical alt ernative, the exact same difference in one attribute generating from comparing with hypothetical alternative and SQ alternative does not generate the same amount o f regret. Again, when , - RRM reduces to RUM. A formal proof is provided in Appendix B . Similar to G - RRM framework, , where is individual information, I can measure the relationship between regret weight and personal factors. 3 .3.3. DCE Mod eling in the Presence of Path Dependence Behavior As decision making is referenc e - dependent, a natural question is whether decision making is path - dependent. That is, the information delivered from the previous choice sets affects the decision when survey participants are making repeated choices in a single survey . When a decision - maker faces a choice scenario s , where , the previous choice scenarios - 1 including the chosen alternative in the prior choice scenarios, i.e., , constructs the path. To test the path dependence behavior, I chose the chosen alternative in the previous cho ice scenario s - 1, , as the RP from the path. If there exists any path - dependent behavior, the chosen alternative would be the most crit ical RP. Therefore, a Path Dependent Random Regret Minimization (P - RRM) model is defined as follows: 72 Assuming that follows i.i.d Extreme Value Type I, a multinomial logit can be used for model estimation. Similar to the previous RRM models, the P - RRM model presumes that losses and g in loss and gain. Again, when , P - RRM will reduce to RUM. A formal proof is provided in Appendix C . 3 .3.4. Hypothesis Testing With models (1) (5), we ar e bac k to the question: what is the underlying decision rule of decision making in DCE? Specifically, among models (1) (5), which model best describes the decision making? The relationship map of the models is shown in Figure 3. 3. defin ed in equation (1) when . RUM, RRM, G - - RRM defined in equations (1) (4) are nested in P - RRM defined in equation (5) as special cases. If in equation (5), P - - RRM. If - RRM will reduc e to G - RRM. If in equation (3), (4) or (5), the corresponding models will reduce to RUM. Therefore, through hypothesis testing on the parameters of P - RRM, we can understand whether decision making is regret minimiza tion or utility maximization, and wh at is(are) the RP(s). The hypotheses are listed below: Hypothesis 1: Survey respondents have same preferences over attributes for SQ alternative and hypothetical alternative s in RUM framework. 73 Hypothesis 2: Survey res pondents have same preferences over attributes for SQ alternative and hypothetical alternative s in G - RRM framework. Hypothesis 3 : decision making is utility maximization. To test this hypothesis, I need to test whether P - - RRM and G - RMM will reduce to RUM through testing whether in equation (3), (4) and (5). Hypothesis 4 : SQ alternative and hypothetical alternative have the same impacts as RP. To test this hypothesis, I need to test if in equation (4). Hypothesis 5 : decision making is path independent. To te st this hypothesis, I need to test if Figure 3. 3. Nesting Stucture of Models 74 3 .4. Survey and Data As an empirical illustration of the approach, I used data from a choice experiment that elicits objective of the CA program is to incentivize farmers to adopt CA practices to reduce nitrogen fertilizer leakage int o the environment. To incentives was conducted amo ngst corn growers in the Midwestern U.S., specifically in Michigan, Iowa, and Indiana in 2016. Mailing addresses for t he survey are randomly drawn from the Farm Service Agency (FSA). 7 With a response rate of 27%, I have 1,294 completed surveys. The survey contains four repeated choice experiment tasks, with each described by two hypothetical alternatives and an SQ altern ative. Each alternative is described by a payment vehicle and three CA practices plus an expected nitrogen saving. The payment level is sug gested by a focus group study among farmers and adjusted after a pilot study of this survey, which was conducted in 2 015. Attributes and attribute levels are defined in Table 2. 2. The first three irement imposed. Expected nitrogen saving is decided by the combination of the three CA practices calculated by agron omy and environmental experts. Besides the choice tasks, the survey contains questions about the can be linked with individual stated SQ values. The to avoid the questions affecting decision making. The individual status quo CA adoption levels as well as associated expected Nitrogen savi ng levels will be incorporated into the dataset for the later empirical estimation. A Bayesian efficiency design that minimizes D - error based on priors from the pilot survey and contains 24 choice sets is generated using Ngene software (Choice Metrics, 201 2). I used a block 7 FSA is the payment services agency within USDA. FSA has records for every farmer who received any form of payment (direct payments, crop insurance subsidies, disaster payments, conservation payments, etc) throu gh USDA. This FAS address boo k covers over 90% of farmers that the CA program is targeted at. 75 design with six blocks containing four choice sets for each to avoid fatigue effects. A respondent was randomly assigned to one of the blocks. The order of presentation and allocation to respondents of the various choice sets is randomiz ed. A sample of the survey is attached in Figure 2. 5. Beyond discrete choice questions, this dataset also social demographic status, attitude toward the environment policy, different resources for information, as well as their relationship with the regret weight parameter that describes decision - weighting on loss and gain. Details of this survey design and administration can be found in Chapter 2 of this dissertation. The sample characteristics by the state are summarized in Table 3. 2. After excluding the incomplete response, with which more than 10% of the questions incomplete, the response rate is highest in Michigan, i.e., 27 %, and lowest in Indiana, i.e., 21 %. Opting out rate, i.e., the percentage of farmers choosing SQ alternat ive among the three alternatives, reaches its highest level in Iowa, i.e., 43 . Through examining the follow - up questions of the survey 8 , I found that the Conservation tillage rate is significantly lower in Michigan than that in the other two states, while the reduced tillage rate is highest in Michigan. Conservation program enrollment is relatively higher in Iowa. The distributions of age, gender, farmi ng experience, and days off farms are generally consistent across the three states. Iowa has a higher per centage of farmers who completed a ssociate or higher - level degrees. Both farm product values and household incomes are higher in Iowa. Iowa and Indiana have more large farm owners in the sample . The status quo CA adoptions are diverging across states: Iowa has the lowest rates of covering crops 8 For further details of the questions, see Figure 2.5 for a survey sample. 76 in the winter whilst having the highest rate of fertilizer application in the fall ; Indian has the highest rate of fertilizer side - dressing. 77 Table 3. 1. Sample Characteristics by State Variable names and codes IA IN MI All states Survey response rate (%) 24 21 27 24 Alternative chosen (%) Alternative 1 29 33 30 31 Alternative 2 28 30 31 30 Stat us quo 43 37 39 40 Tillage type (%) Conventional tillage: less than 15% residue remaining on surface 5 2 12 7 Reduced tillage: 15 - 30% residue remaining on surface 50 53 65 56 Conservation tillage: more than 30% residue remaining on surface 45 45 23 37 Conservation program enrollment (%) Ever participated in the past 43 37 38 39 Age (%) Between 18 - 34 4 5 5 5 Between 35 - 54 28 25 28 27 Between 55 - 64 34 32 34 33 Above 64 34 38 33 35 Farming experience (%) Average or above expe rience 35 37 38 37 Gender (%) Male 97 95 98 97 Education (%) Some college, no degree, or lower 53 58 64 58 Associate degree, or higher 47 42 36 42 Days off farm per year (%) Less than 100 days 67 68 68 68 Greater than 100 days 33 3 2 32 32 Total values of products sold in 2015 (%) Less than $100,000 30 33 44 36 $100,000 - $499,999 44 39 38 41 $500,000 - $999,999 15 14 12 13 Greater than $1,000,000 11 14 6 10 Annual Household Income (%) Low income (up to $25,000) 11 8 14 12 Medium income ($25,000 - $100,000) 62 66 66 64 High income (above $100,000) 27 26 20 24 Acres of farm operated in 2015 (%) Less than 100 acres 17 17 17 17 Between 100 - 500 acres 43 38 49 44 78 Table 3.1 Greater than 500 acres 40 45 34 39 Certainty about the decision (%) Uncertain 5 8 5 6 Somewhat certain 28 29 29 29 Certain 67 63 66 65 Winter crops covered (%) Adopted 18 23 24 22 Fall application (%) Adopted 31 6 9 15 Side - dress application applied (%) Adopted 32 53 47 44 3 .5. Results Models based on equation (1) (5) are estimated with MNL using Python Biogeme. With a nesting structure, I tested the assumptions imposed on different models. Estimation results of RUM, G - - RRM based o n the whole sample that includes the first choice set are reported in columns (1) ( 4 ) in Table 3. 2 - RRM and P - RRM based on the subsample, which excludes the first choice set, are reported in columns ( 5 ) and ( 6 ) in Table 3. 2 9 . Her e I excluded the first choice set be cause there is no previous choice set to refer to for path - dependence model. Hypothesis test results are summarized in Table 3. 3 . Note that I have incorporated SQ alternatives with individual stated SQ values. 9 - RRM with the subsample data for comparison purpose. 79 Table 3. 2 . Estimation Results (1) ( 2 ) ( 3 ) ( 4 ) ( 5 ) ( 6 ) S ample Set = 1, 2, 3, 4 Set = 1, 2, 3, 4 Set = 1, 2, 3, 4 Set = 1, 2, 3, 4 Set = 2, 3, 4 Set = 2, 3, 4 RUM RUM G - RRM - RRM - RRM P - RRM - 0.643 *** a - 0.718 *** - 0.234* 0.845 ** 0.646 *** 0.2 41 * ( 0.120 ) b (0 .145 ) ( 0.135 ) ( 0.407 ) ( 0.148 ) ( 0.14 5 ) - 0.0118 0.00343 - 0.00669 0.0253 - 0.0689 - 0.91 *** ( 0.0733 ) ( 0.0821 ) ( 0.0368 ) ( 0.0463 ) ( 0.0479 ) ( 0.1 39 ) - 0.108 - 0.0133 - 0.0455 - 0.864 0.734 *** 0.5 6 9 ( 0.173 ) ( 0.196 ) ( 0.102 ) ( 0.549 ) ( 0.129 ) ( 0 . 827 ) 0.324 1.06 0.145 - 0.158 0.505* 1.6 9 ** ( 0.707 ) ( 0.813 ) ( 0.411 ) ( 0.338 ) ( 0.278 ) ( 0 .881 ) 0.00840 *** 0.00827 *** 0.003 00*** 0.00178 0.00174 0.00162*** ( 0.000455 ) ( 0.000472 ) ( 0.00111 ) ( 0.00816 ) ( 0.00132 ) ( 0.000251 ) - 0.752 *** - 2.18 ** - 2.14 *** - 2.20 *** ( 0.434 ) ( 1.02 ) ( 0.236 ) ( 0.315 ) - 0.231 *** - 0.00678 - 0.000682 0.00422 ( 0.173) ( 0.0159 ) ( 0.0195 ) ( 0.185 ) - 1.26 2.02 ** - 2.07 *** - 2.16 *** ( 0.434) ( 0.945 ) ( 0.210 ) ( 0.241 ) 2.44 - 0.304 0.703* 1.25 ** ( 1.72) ( 0.639 ) ( 0.385 ) ( 0.544 ) 0 c 0.0247 0.0319 *** 0.0195 *** - ( 0.0512 ) ( 0.0128 ) ( 0.00247 ) - 3.44 *** ( 0.197 ) - 2.73 *** ( 0.203 ) - 2 . 2 4 *** ( 0.210 ) 3.2 9 ** ( 1.78 ) 0.001 81*** ( 0.000 22 6 ) 0.751 *** 1.2*** - 0.753 *** - 0.83 - 0.946 *** - 1. 28 *** ( 0.0499 ) ( 0.0759 ( 0.0497 ) ( 0.616 ) ( 0.1 55 ) ( 0. 291 ) 0.0699 0.369 0. 350 ** 0.11 4 *** 80 Table 3.2 ( 0.379 ) ( 0.932 ) (0.11 7 ) ( 0.02 24 ) - 4495 - 44 34 - 44 95 - 4414 - 32 78 - 29 65 AIC /N 2.09 2.07 2.09 2.06 2.04 1.85 BIC /N 2.10 2.09 2.10 2.08 2.06 1.88 N 4300 4300 4300 4300 3225 3225 a. *** , ** , * . b. Robust standard error is reported in the bracket. c. is not identifiable sinc e the status quo payment is constant at zero. Table 3. 3 . Hypothesis Test Results Equ Null hypothesis Alternative hypothesis Unrestricted Restricted LR test t - value d.f. p - value (1) - 4495 - 4434 122 5 0 (3) 0.18 4292 0.85 (4) 0.40 4287 0.69 (4) 2.99 3212 0 (5) 5.10 3207 0 (4) - 4414 - 4495 162 5 0 (5) - 2965 - 3278 626 5 0 81 First of all , the signs of the preference parameters for RUM and as reported in columns (1) and (2) of Table 3. 2 are generally as expected. Covering crops in winter , i.e., Winter , avoiding fertilizer application in fall, i.e., Fall , and side - dressing fertilizer, i.e., Side , all decrease utility, while saving nitrogen, i.e., Nitrogen , and payment, i.e., Pay , both increase utility. The SQ constant term is positive , indicating that maintaining SQ is preferred. These findings are also - up questions in the survey, 22 %, 85%, and 44 % of the respondents have e ver met the requirements of cover ing crops in winter, avoiding fertilizer application in fall, and side - dressing fertilizer in the past. The null hypothesis that individuals have same preferences over attributes for hypothetical alternatives and SQ alterna tive is rejected in Table 3. 3. Pe ople do not take hypothetical and SQ alternatives with the same weights either because their preferences over the two types of not salient in the choice scenar ios and cannot be processed similarly. To continue, I investigated whether decision making is utility maximization or alternatively regret minimization through hypothesis testing. Specifically, I investigated whether G - RRM - RRM, and P - RRM reduce to RUM through testing if or for the corresponding models. The regret weight parameter s are respectively 0. 0699 and 0. 369 for G - - RRM as reported in columns ( 3 ) and ( 4 ), and 0.224 and 0. 112 - RRM and P - RRM as reported in columns ( 5 ) an d ( 6 ). The null hypo thesis that is rejected in - RRM and P - RRM which run on the sub - sample at 1% significance level, however, is not rejected in G - - RRM which run on the whole sample. These findings imply that regret minimization is a reason able assumption for decision making modeling when people start to make repeated choices. 82 As decision making is reference - dependent and regret minimizing in repeated choice scenario, I examined what are the RPs and the relative contributions of these RPs. I started with testin g whether SQ and hypothetical alternatives have the same impacts as R Ps. Specifically, I tested the hypothesis that . G - - RRM with . The likelihood ratio test, as reported in row six of Table 3. 3 rejects this hypothesis at the 1% significance level, implying that hypothe tical alternatives do not share the same impacts with SQ alternatives as RP. Therefore, it is reasonable to estimate and separately in DCE that includes an SQ alternative. Besides, the Akaike information criterion (AIC) and Bayesian information criterion (BIC) suggest that - minimizing information loss. - RRM model - RRM running on the subsample because reference dependence b ehavior is only significant in repeated choice scenario of our sample and the results are easy to be compared with that from the P - RRM model. As reported in column (5) of Table 3. 2 . , have all signs, except that of Fall , as expected and generally sta tistically significant. only have signs significant and as expected for attributes Nitrogen . This raises the concern of identifying parameters for hypothetical alternatives - RRM. I will further check this issue after path dependence behavior is incorporated. Generally, the - RRM model implies that both SQ and hypothetical alternatives can pl ay as RP, but hypothetical alternatives might have weaker impacts. Lastly, the P - RRM estimation is reported in column (6) of Table 3. 2. Note that - RRM is the restricted model of P - RRM with . The likelihood ratio test of the nesting structure of G - RRM and P - RRM is reported in the last row of Table 3. 3. The likelihood ratio test rejects - RRM in favor of P - RRM, implying the existence of path - dependent behavior. Meantime , P - 83 RRM sign ificantly - RRM in terms of likelihood value , AIC and BIC. These findings justify that incorporating the last round chosen alternative is vital for correctly modeling the choice behavior. Taking a closer look at the parameters of the P - R RM model , I found that the signs and scales of that repre sent the last round reference - dependent behavior are consistent with those from the previous models for all attributes . Besides, the estimations for are significant with expected signs for all attributes except for Fall, and the estimations for are significant with expected signs for all attributes except for Winter and Side . The scales of parameter for the same attribute among three RPs are highest for the last round cho sen alternative and lowest for the hypothetical alternative, except that Pay has higher weight when SQ alternative is referred. It is worth mentioning this w be significantly identified i f the associated preferences are significant. One reason is due to the nature of the research, correlations between Nitrogen saving and adopting CA practices are relatively high, i.e., 0.4 - 0.8. This high correlation among attributes significantly raises the are also not significantly and correctly identified. Meantim e, P - RRM further raises the required sample size as the number of parameters triple s as compared wi th the RUM model. Last, if the alternative with this round over choice sets. This again increases the required sample size for parameter identification. Taking all the above findings and challenges together, the P - RRM model reveals a solid dec ision pattern. That is, when there is no previous choice experience, decision - the mos t critical RP for decision making. When there is a path, decision making is both current 84 choice set and path - dependent: decision - ternative , SQ alternative, and hypothetical alternatives can all play as RP to influence the decisi on making, though the same attribute might have different weights as different types of RP. For the same attribute, the last round chosen alternative has the highest weight as a RP and the hypothetical alternative has the lowest weight. The limitation of t his work is we need a larger sample to comprehensively understand the roles of SQ and hypothetical alternatives in a path - dependence choice scenario. It is worth mentioning that I have related the individual - specific status quo with SQ alternative. This a pproach, as compared with treating the SQ alternative as homogeneous among survey respondents, is recommended in DCE literature because it addresses the prob lem of heterogeneity of SQ and can increase model explanatory power (Kataria et al., 2012; Glenk, 2 011; Artell et al., 2013; Barton and Bergland, 2010; Banzhaf et al., 2001). On the one hand, this explains why SQ plays as a more important RP as compared wi th the hypothetical alternative. On the other hand, there raises the concern of endogeneity since t he individual specific SQ might be correlated with the error term of the utility function. The potential endogeneity of individual specific SQ is a general p roblem if it is a problem of the DCE literature in the presence of the SQ setting. This is out of t Finally, I examined the relationship between individual characteristics and regret we ight by defining for P - RRM using the subsample of choice sets (2) (4) , where is individual information with dummy coding . Table 3. 4 reports the factors that influence the regret weight by running separately on each factor, i.e., X. Besides e xperience with conservation tillage or conservation programs , and gender, all factors are associated with regret weight. People who have higher education, longer days off - farm, higher farm product value, higher household 85 income , or larger farm are less reg ret minimization oriented. Farmers who are older, more certain about their decisions, or have more farming experience, are more regret minimiza tion oriented. Table 3. 4. Factors Influencing Regret Profundity Par Est Tillage type 1. 0 7*** a (0.037) b Conservation tillage - 0.126 (0.184) Conservation program enrollment 0.78*** (0.183) Ever participated in the past - 0.193 (0.208) Age 0.286*** (0.0819) 55 or older 0.0312* ** (0.00709) Farming experience 0.321*** (0.238) Average or above 0.0672*** (0.00279) Gender 1.02*** (0.312) Male - 0.202 (0.923) Education 1.25*** (0.422) Associated degree or higher - 0.326*** (0.0721) Days off farm per year (%) 0.82*** (0.281) Greater than 100 days - 0.132*** (0.0298) Total values of products sold in 2015 0.92*** (0.363) $500,000 or above - 0.37*** (0.0675) Annual Household Income 0.96*** (0.395) High income - 0.32*** (0.0286) Acres of farm operated in 2015 (acres) 0.79*** (0.274) Greater than 500 acres - 0.293*** (0.0712) Certainty about the decision (%) 0.424* ** (0.106) Certain 0.175*** (0.0547) a. *** , ** , * . b. Robust standard error is reported in the bracket. 86 3 .6. Implications The discussion above rejects the choice set independent assumption imposed on conventional RUM models. This section investigates how such reference - dependent behavior affects welfare analysis and policy design based on DCE and discusses the implications fo r DCE. 3 .6.1. The Implications for Policy Design As decision making is reference - dependent, p olicymakers can influence decisions by manipulating the RP of a policy. Below I used a simulation approach to show how differently pation rate. This will inform policymakers t o increase policy efficiency. Presume a policy scenario which is exactly the same as the scenario in the survey design. There are four choice task s with two hypothetical alternatives plus an SQ alternative for e ach . One alternative is the target program t hat the policymaker aims at maximizing the adoption rate, and the other alternative is the nudge program that the policymaker adds to the choice set to give the decision - maker an alternative option and potentiall y nudge the desired behavior. The decision - m akers do not know which is the target program and which is the nudge program, and they make decisions to minimize their regret. I will show how to choose the nudge program to n. et program is to adopt all three CA practices with an expected nitrogen saving of 50%. Excluding the target program (Yes, Yes, Yes) and no - action program (No, No, No), I have six different nudge programs as candi dates. The first step is to decide the payme nt level for each of the candidate nudge programs. The objective is to set the payment at the level that is reasonable to decision - makers but still low enough such that the 87 nudge program will not be chosen even w ow as $60 10 SQ will be chosen. I refer to the WTA estimation based on RUM models and run a few simulations to decide the lts are listed in Table 3. 5 . Next, I predicted 10,000 simulated decisions based on estimations with RUM models and P - RRM models conditional on different nudge programs using software R. It is important to note values of SQ and the previous chosen alternative based on values 11 . With simulation, the pro - RRM represents the true adoption rate. The program adoption rate estimated with RUM represents the pred iction that fails to account for reference dependence behavior . , is depicted in Figure 3. 4. As the nudge program is designed at the payment level such that it will not be chosen, the staying at SQ rate equals to . 10 I chose $60 by referring to the range of WTA estimations based on RUM model. 11 As there are 1,249 respondents, I randomly replicated their stated status quo to generate 10,000 decisions. 88 Table 3. 5 . Candidate Nudge Programs Program No. Target 1 2 3 4 5 6 winter Yes Yes No No Yes No Yes fall Yes No Yes No Yes Yes No side Yes No No Yes No Yes Yes nitrogen 50% 10% 10% 25% 25% 25% 40% Pay ($) X 85 5 15 6 0 20 65 Figure 3. Figure 3. e increases as payment increases. However, the increasing patterns differentiate under different model assum ptions. In the meantime, the adoption curvature does not follow a sigmoid pattern, which is due to the o. Under the simulation with P - RRM assumption, the adoption is most sensitive when payment is either below $ 100 or above $160; under the simulation with RUM assumption, the adoption is most sensitive when payment is between $100 and $170. The different pa tterns reflect the difference in decision behavior assumptions. 89 The reason is, when the payment is low, the incre ases in payment can be effective in encouraging adoption because their SQs are not favorable enough to compe nsate for the regret of losing the small payment from enrolling in the target program. As the payment increases above $100, the target program begin s to attract group of growers, the regret generated from comparing SQ with the target program, for the three CA practices, is so large that a large increase in p ayment is needed to compensate for the regret om SQ to the target program. Lastly, compensated by paym ent, and payment regains its ability to promote adoption. Payment not working well in the mediocre payment r assumption: alternatives with in - between attribute values generate lower regret than those with more ex treme attribute values because attributes are not linearly substitutable as assumed in the utility maximizat ion framework. For the adoption rate, the program adoption can reach 100% as long as payment is high enough, i.e., $180, according to the RUM model . However, the same amount of payment cannot reach the same adoption rates based on the P - RRM prediction. Similarly, the adoption rates predicted with P - RRM, when payment is low, are not as low as those predicted with RUM. The difference again explai ns the difference in decision behavior assumptions: with the RUM model, the utility will increase as long as any attribute gets better; with the P - RRM model, decision making depends on the bundle of attributes that do not have short slabs rather than on a singl e 90 bounded after payment no longer generates regret. nudge programs and h ow to select the nudge program. As shown in Figure 3. 4, program 6 is most effective in encouraging adoption, while program 2 works the least effective. Referring to the design of nudge programs in Table 3. 5 , I found that the nudge programs with CA attribut e leve that for the target program, the gains are from bilateral comparison with respect to payment, and the losses are from the bilateral comparison with respect to th e three CA practices. We want the nudge program to be designed at the levels that the CA practices are similar to that of the target program, such that regret generated from comparing the target program with the nudge program is minimized. In this wa y, the Finally, I examined the program adoption rate by the state for policy reference purposes. Table 3. 1 shows that the SQ levels are diverging among states: the rates of covering crops in winter and side - dres sing f ertilizer are lower in Iowa, and the rate of fall application is higher in given the same payment as shown in Figure 3. 5. 91 Figure 3. s Adop tion Rate by State with Program 6 as the Nudge In summary, as decision - makers use a behavioral strategy for decision making, a small amount of payment is most effective in attracting people who can easily meet the policy requirement. Otherwise, a large amount of payment is needed to shift the decision - makers from a feeling of loss to a feeling of gain due to committing the program requirements. On the other hand, policymakers can make use of the behavioral strategy to promote program adoption by ca refull y designing the nudge program. 3 .6.2. The Implications for Discrete Choice Experiment The existence of choice set dependent behavior set caveats on how to choose discrete choice modeling framework to correctly describe the true decision behavior. In a c hoice set with multiple alternatives, i.e., greater than two alternatives, the preference evaluation will be contingent on the choice set composition . The problem with a multiple - choice setting is that the evaluation of a considered alternative will not on ly depend on its attributes but also depend on 92 the other hypothetical alternatives provided in the choice set as RP . As a result, DCE will no longer reveal the preferences for another decision scenario unless the experiment design perfectly imitates the re Furthermore, i f the real - world decision is a single binary choice, a single binary choice setting will be the necessary condition for accurate preference estimation of the stated preference approach. Si milarl y, as preference evaluation is contingent on the choice set composition, WTP/WTA will also be contingent on the hypothetical alternatives provided in the choice set. As we know, due to the linear additive form of the utility function of RUM, the RUM model can give direct WTA estimation by taking the ratio of marginal utility of an attribute to the marginal utility of the cost attribute. However, due to the feature of asymmetry in loss and gain of RRM, WTA cannot be directly calculated. To give a direc t comp arison to welfare analysis between the decision - makers indifferent between the proposed plan and their status quo. Specifically, I used the preference estimates from RUM a nd P - RRM to simulate 10,000 decision s with their values of SQ and previously chosen alternatives randomly drawn from the respondents of the survey. I used program 2 and program 6 as the nudge programs for P - RRM model prediction. Lastly, I gradually increas to incur a 50% adopt based on different models are reported in Table 3. 6 . 93 Table 3. 6 . WTA ($) Indifferent B etween Target and SQ Programs Winter Fall Side - dress Nitrogen Saving RUM P - RRM - Nudge 2 P - RRM - Nudge 6 No No No 0% 1 01 143 1 59 Yes No No 10% 1 91 268 242 No Yes No 10% 88 68 88 No No Yes 25% 92 71 58 Yes Yes No 25% 1 59 162 14 2 No Yes Yes 25% 9 9 78 7 5 Yes No Yes 40% 14 1 154 1 31 Yes Yes Yes 50% 1 35 1 3 2 107 I found that WTAs of models. Given that RUM fails to take account of choice set dependent decision behavior, using RUM for welfa re analysis would give a biased estimation. Besides, WTAs estimation depends on t he design of the nudge program. Consistent with previous findings, program 6 is generally more efficient than program 2 in nudging desired behaviors concerning reducing WTAs. Note that nitrogen saving itself can reduce farming costs, and nitrogen saving is decided by the combination of practices that have been taken as such, taking all three practices incurs lower WTA than taking a single or two practices. 3 .7. Conclusions B sts the decision behavior assumptions imposed on RUM and RRM frameworks. This study is among the first few literature investigating regret behavior using the RRM framework in environmental economics (e.g., Boeri et al., 2012; Thiene et al., 2012). Meanwhil e, the findings have general implications in fields outside environmental economics. This is also the first paper investigating the RRM framework in a WTA scenario. 94 First, de cision making is choice set composition as well as path - dependent. The hypothetic al alternatives, the decision - s chosen alternative is the most important if the current choice scenario is not the first choice scenario. When the decision - makers are first exposed to the choice set where there is no last round of information delivered, SQ is the most critical RP. As s urvey respondents gradually collect information over repeated choices, the decisi on making shows a path - dependent pattern. That is, decision making evolves from the current choice set dependent to across choice sets dependent. Moreover, a ttribute depend on their positions of being hypothetical, SQ, or . The existence of the path dependence decision raises the issue of a proper interpretation of DCE analysis results. The gap between the survey and real - world decision scenarios c an produce significant bias in decision prediction and WTP/WTA estimation. This work reveals a discussed in DCE literature . The policy implication of these finding s is that, by manipulating the nudge program proposed together with the target can increase. Setting the nudge program close to the target program is most efficient in nudging desired choices. A future research a venue is to study which approach works best in boosting a target target program with a nudge program or proposing a single binary choice. There are several issues left with future discussions. First and foremost, we need to be cautious that these conclusions are based on an experimental study , whether the decision - mak er s 95 adopt similar decision strategies in the real choice scenario remains further examin ed . Next, how the path dependence behavior evolves over repeated choice remains further investigation . So far, ent the information delivered from the previous choices. However, how the other information collected from the earlier choices wor ks together to influence decision making remains to be discussed. To continue, due to the sample size limitation, it is diffic ult to distinguish the effects of different attributes that work as different kinds of RPs as the model complexity increases. Futu re work is necessary to understand each les. Furthermore, relating the individual - specific status quo with SQ alternative solves the problem of of the SQs as well as raises the issue of endogeneity of individual SQs in discrete choice modeling. An avenue for further re search would be discussing the endogeneity of the SQ setting in DCE. Lastly, a nudge program has been justified to be effective in promoting should be enacted t o investigate whether a single alternative policy or a target alternative combined with a nudge alternative will work better in ac hieving the desired results. 96 REFERENCES 97 REFERENCES Adamowicz, W., Boxall, P., Williams, M., Louviere, J., 1998. Stated preference approaches for measuring passive use values: choice experiment and contingent valuation. American Journal of Agricultural Economics 80, 64 75. Adamowicz, W.L., Glenk, K. and Meyerhoff, J., 2014. Choice modelling research in environmental and resource economics. Chapters , pp.661 - 674. Arentze, T. and Timmermans, H., 2007. Parametric action decision trees: Incorpora ting continuous attribute variables into rule - based models of discrete choice. T ransportation Research Part B: Methodological , 41 (7), pp.772 - 783. Artell, J., Ahtiainen, H. and Pouta, E., 2013. Subjective vs. objective measures in the valuation of water qua lity. Journal of environmental management, 130, pp.288 - 296. Banzhaf, M.R., Johns on, F.R. and Mathews, K.E., 2001. Opt - preferences. The choice modelling approach to environmental valuation. Edward Elgar, London, pp.157 - 1 77. Barton, D.N. and Bergland, O., 2010. Valuing irrigation water using a choice experiment: an Environment and Development Economics, 15(3), pp.321 - 340. Bell, D.E., 1982. Regret in decision making under uncertainty. Operations Research 30 (5), 961 981. Boeri, M., Longo , A., Doherty, E. and Hynes, S., 2012. Site choices in recreational demand: a matter of utility maximization or regret minimization?. Journal of Environmental Economics and Pol icy , 1 (1), pp.32 - 47. Boeri, M., Scarpa, R., & Chorus, C. G., 2014. Stated choice s and benefit estimates in the context of traffic calming schemes: Utility maximization, regret minimization, or both? Transportation Research A, 61, 121 135. 2014.01.003 . Capu to, V., Lusk, J.L. and Nayga, R.M., 2018. Choice experiments are not conducted in a vacuum: The effects of external price information on choice behavior. Journal of Economic Behavior & Organization , 145 , pp.335 - 351. Carlsson, F. and Martinsson, P., 2001. D o hypothetical and actual marginal willingness to pay differ in choic e experiments?: Application to the valuation of the environment. Journal of Environmental Economics and Management , 41 (2), pp.179 - 192. 98 Carson, R.T. and Groves, T., 2007. Incentive and inf ormational properties of preference questions. Environmental and reso urce economics, 37(1), pp.181 - 210. Chorus, C., 2014. A Generalized Random Regret Minimization model. Trans. Res. Part B, Pages 224 - 238. Chorus, C., Arentze, T., and Timmermans, H., 2008. A random regret minimization model of travel choice. Trans. Res. Par t B, 42(1), 1 18. Chorus, C., van Cranenburgh, S. and Dekker, T., 2014. Random regret minimization for consumer choice modeling: Assessment of empirical evidence. Journal of Business Rese arch , 67 (11), pp.2428 - 2436. Chorus, C.G. and Bierlaire, M., 2013. An empirical comparison of travel choice models that capture preferences for compromise alternatives. Transportation , 40 (3), pp.549 - 562. Chorus, C.G., 2010. A new model of random regret mini mization. EJTIR, 10 (2), 2010 . Chorus, C.G., 2012. Random Regret - based discrete choice modeling: A tutorial . Springer Science & Business Media. Chorus, C.G., Rose, J.M. and Hensher, D.A., 2013. Regret minimization or utility maximization: it depends on the attribute. Environment and Planning B: Planning and Design , 40 (1), pp.154 - 169. Crawford, M.T., McConnell, A.R., Lewis, A.C., Sherman, S.J., 2002. Reactance, compliance and anticipated regret. Journal of Experimental Social Psychology 38 (1), 56 63. Dekker , T. and Chorus, C.G., 2018. Consumer surplus for ran dom regret minimisation models. Journal of Environmental Economics and Policy , pp.1 - 18. Dekker, T. and Chorus, C.G., 2018. Consumer surplus for random regret minimisation models. Journal of Environmental Economics and Policy , pp.1 - 18. Fishburn, P.C., 1982. Non - transitive measurable utility. Journal of Mathematical Psychology 26 (1), 31 67. Glenk, K., 2011. Using local knowledge to model asymmetric preference formation in willingness to pay for environment al services. Journal of environmental management, 92( 3), pp.531 - 541. Hensher, D., Greene, W., Chorus, C., 2013. Random regret minimisation or random utility maximisation: an exploratory analysis in the context of automobile choice. J. Adv. Transp. 99 Hess, S. , Beck, M.J., Chorus, C.G., 2014. Contrasts between utility maximisation and regret minimisation in the presence of opt out alternatives. Transport. Res. Part A: Policy Pract. 66 (0), 1 12. Hess, S., Stathopoulos, A. and Daly, A., 2012. Allowing fo r hetero geneous decision rules in discrete choice models: an approach and four case studies. Transportation , 39 (3), pp.565 - 591. Kahneman, D., & Tversky, A. 1979. Prospect theory: an analysis of decisions under risk. Econometrica, 47, 263 - 291. Kahneman, D. , Knetsc h, J.L. and Thaler, R.H., 1991. Anomalies: The endowment effect, loss aversion, and status quo bias. Journal of Economic perspectives , 5 (1), pp.193 - 206. Kataria, M., Bateman, I., Christensen, T., Dubgaard, A., Hasler, B., Hime, S., Ladenburg, J., L evin, G. , Martinsen, L. and Nissen, C., 2012. Scenario realism and welfare estimates in choice experiments A non - market valuation study on the European water framework directive. Journal of environmental management, 94(1), pp.25 - 33. Kivetz, R., Netzer, O. and Srin ivasan, V., 2004. Alternative models for capturing the compromise effect. Journal of marketing research , 41 (3), pp.237 - 257. , B., Rabin , M., 2006. A Model of Reference - Dependent Preferences. The Quarterly Journal of Economics, Volume 121, Issue 4, 1 November 2006, Pages 1133 1165. Loomes, G., Sugden, R., 1982. Regret theory: an alternative theory of rational choice under uncertainty. The Economic Journal 92, 805 824. Louviere, J.J. and Henshe r, D.A., 1982. On the design and analysis of simulated choice or allocation experiments in travel choice modelling. Transportation research record , 890 (1982), pp.11 - 17. Louviere, J.J. and Woodworth, G ., 1983. Design and analysis of simulated consumer choic e or allocation experiments: an approach based on aggregate data. Journal of marketing research , pp.350 - 367. Lusk, J.L. and Schroeder, T.C., 2004. Are choice experiments incentive compatible? A test w ith quality differentiated beef steaks. American Journal of Agricultural Economics , 86 (2), pp.467 - 482. Machina, M.J., 1987. Choice under uncertainty: problems solved and unsolved. Journal of Economic Perspectives 1, 121 154. McFadden, D., 1974. Conditional logit analysis of qualitative choice behaviour. In: Zar embka, P. (Ed.), Frontiers in Econometrics. Academic Press, pp. 105 142. 100 Metrics, C., 2012. Ngene 1.1. 1 user manual & reference guide. Sydney, Australia: ChoiceMetrics, 19, p.20. Metrics, C., 2012. Ngene 1.1. 1 user manual & reference guide. Sydney, Austr alia: Choice Metrics . Myerson, R.B., 1979. Incentive compatibility and the bargaining problem. Econometrica: journal of the Econometric Society , pp.61 - 73. Quiggin, J., 1994. Regret theory with general choice sets. Journal of Risk and Uncertainty 8 (2), 153 165. Rakotonarivo, O.S., Schaafsma, M. and Hockley, N., 2016. A systematic review of the reliability and validity of discrete choice experiments in valui ng non - market environmental goods. Journal of environmental management , 183 , pp.98 - 109. Samuelson, Wi Risk and Uncertainty, 1988, 1, 7 59. Swait, J., 2001. A non - compensatory choice model incorporating attribute cutoffs. Transportation Research Part B: Methodological , 35 (10), p p.903 - 928. Taylor, K.A., 1997. A regret theory approach to assessing consumer satisfaction. Marketing Letters 8 (2), 229 238. Thaler, R. , 1980. Toward a positive theory of consumer choice. Journal of Economic Behavior & Organization , 1 (1), pp.39 - 60. Thiene , M., Boeri, M. and Chorus, C.G., 2012. Random regret minimization: exploration of a new choice model for environmental and resource eco nomics. Environmental and resource economics , 51 (3), pp.413 - 429. Tian, Q., Zhao, J., Reimer, A., and Lupi, F., 2018. Far mers Decision on Nitrogen Application: Testing Information Treatment Effect and Commitment Cost Theory. Working paper. Tonsor, G.T., 20 18. Producer Decision Making under Uncertainty: Role of Past Experiences and Question Framing. American Journal of Agric ultural Economics . Tversky, A., & Kahneman, D. 1981. The framing of decisions and the psychology of choice. Science, 211, 453 - 458. Tvers ky, A., & Kahneman, D. 1986. Rational choice and the framing of decisions. The journal of Business, 59, S251 - S278. Tvers ky, A., & Kahneman, D. 1991. Loss aversion and riskless choice: a reference dependent model. Quarterly Journal of Economics, 106, 1039 - 1 061. 101 Van Cranenburgh, S., Guevara, C.A. and Chorus, C.G., 2015. New insights on random regret minimization models. Tran sportation Research Part A: Policy and Practice , 74 , pp.91 - 109. Van Cranenburgh, S., Rose, J.M., Chorus, C.G., 2018. On the robustness o f efficient experimental designs towards the underlying decision rule. Transport. Res. Pol. Prac 109, 50 64. Zhang, J., Timmermans, H., Borgers, A. and Wang, D., 2004. Modeling traveler choice behavior using the concepts of relative utility and relative i nterest. Transportation Research Part B: Methodological, 38(3), pp.215 - 234. 102 CHAPTER 4 . REGRET MINIMIZATION, PATH DEP ENDENCE, AND ATTRIBUTE NON - ATTENDANCE IN DISCRETE CHOICE EXPERIMENTS 4 .1. Introduction Discrete Choice Experiment (DCE), developed by Louviere and Hensher (1982) and Louviere and Woodworth (1983), is a popular stated preference approach to elicit environm ental preferences (Louiviere et al., 2000; Kanninen, 2007). A hypothetical decision - maki ng scenario is introduced through a survey experiment. The survey consists of repeated choice scenarios. Each choice scenario consists of multiple alternatives, with ea ch described by a combination of e DCE, estimation models such as preferences over the individual attributes. The workhorse estimation model of DCE, i.e., RUM, is based on the notion of rational a gents making choices to maximize expected utilities. However, recent developments in behavioral economics have identified a range of alternative decision strategies tha t challenge the rational decision assumption that endorses the RUM framework. These deci sion strategies include attribute cancellation/exclusion when attributes are in common levels (e.g., Layton and Hensher, 2010), reduced attention to a subset of attribu tes (e.g., Houston and Sherman, 1995), imposing thresholds of acceptable levels on attri butes (e.g., Swait, 2001; Hensher and Rose, 2012), and reference dependence around a recent or past experience (e.g., Chorus, 2008, 2010; Caputo, Lusk, and Nayga, 2018, 2020). In sum, there are two major categories of behavioral decision strategies reduced attention and reference dependence. Previous research shows these strategies are context specific (Gilovich et al., 2002), whereas there is a lack of research directly investigating and comparing the performances of these alternative strategies within a s ingle study. 103 This paper examines the interlinkage and implications of these two major behavioral strategies, namely reduced attention and reference dependence, in the hope of making up the absence of DCE literature that investigates the relationship betwe en the two approaches. To account for these two strategies within discrete choice modeling, I followed the frameworks of Attribute non - Attendance (ANA) (see, Hensher et al., 2005 for an initial introduction) and Random Regret Minimization (RRM) (see, Choru s et al., 2008 for an initial introduction), and developed a model to examine these two strategies within a single framework. Specifically, ANA literature describes an information processing strategy wherein respondents ignore specific attributes in compar ing alternatives in a DCE setting. The idea of ANA can be traced back to the lexicographic heuristic strategy (Tversky, 1969), which assumes that choices are made based on the essential attributes while ignoring all other information. The drivers for ANA b ehavior can be either subjective ignorance or unconscious ignorance decision - makers intentionally eliminate the irrelevant/unimportant attributes to simplify choice tas ks, or they unconsciously ignore the attributes due to inattention. RRM (Chorus, 2008, 2010) argues that decision making is based on pairwise comparison of the chosen and foregone alternatives (Bell, 1982; Fishburn, 1982; Loomes and Sugden, 1982). Rooted in Regret Theory (RT), it posits that the foregone alternative performing better or wors e than the chosen alternative gives people a feeling of loss or gain from that decision. The regret aversion emotion leaves individuals more sensitive to loss than to g ain generated from the bilateral comparison. Its consistency with real behavior has made RT a popular alternative of utility maximization for choice analysis (e.g, Loomes and Sugden, 1983, 1987; Machina, 1987; Quiggin, 1994; Hey and Orme, 1994; Starmer, 20 00; Hart, 2005). 104 This paper investigates the relationship between ANA and RRM in a singl e framework and how the relationship evolves across the repeated choice scenarios in the DCE setting. The empirical analysis uses data from a Willingness - To - Accept (WTA ) DCE survey on incentives to adopt Conservation Agriculture (CA) 12 practices among corn growers in the Midwest U.S. I showed that the reference - better than the utility maximization framework in terms of statistical criteria. I also found a decision - us choice scenario of the survey, if the choice set is not the first one, is an important Reference Point (RP) of decision making. Specifically, decision - makers shift th s they collect information through making repeated choices. Lastly, I found evidence of ANA behavior in DCE under both RUM and RRM frameworks, but this behavior vanishes after this path dependence behavior is accounted for. Before going into the details o f model development, which will be presented in the next section, I will first take the space to introduce the frameworks of ANA and RRM as well as to discuss the ration ales of developing a new model based on these two frameworks. To begin with, RRM (Chor us, 2008, 2010) describes a simplifying decision strategy of reference dependence. It assumes a decision - maker chooses the alternative with the lowest regret from the gi ven choices and follows a similar framework of RUM for econometric modeling. RRM distin guishes from RUM in that RRM adopts a non - linear function form for the observed part of utility (or regret), as opposed to the linear combination function form adopted b y RUM. It, therefore, captures the asymmetrical weighting in loss and gain. Due to the advantages in 12 Conservation Agriculture (CA) is a term defined by the Food and Agricultural Organization of the United Nations - saving agricultural crop production that strives to achieve acceptable profits together with high and sustained production levels while c oncurrently conserving the environ 105 capturing behavior features with no additional requirement of estimation, RRM has gained wide attention from literature in fields such as transportation, m arketing, and environmental economics (e.g., Hensher et al., 2013; Thiene et al., 2012; Chorus and Bierlaire, 2013; Boeri et al., 2014; Adamowicz et al., 2014). Notwithstanding the contribution of the RRM framework in incorporating reference - dependent beh avior within discrete choice modeling, the RRM literature does not explicitly discuss t he roles of different information plays as reference points. Chapter 3 of this dissertation addresses this issue explicitly. There is a multitude of information delivere d in the survey with the potential of playing as an RP. The information includes the hy pothetical alternatives described in the choice set, the Status Quo (SQ) of the decision - maker, as well as the information acquired from the previous choice set(s). Diff erent RPs may contribute to regret generation with different weights. Besides, if infor mation learned from the previous choice set(s) serves as a significant RP, the reference strategy will be dynamic because details of the previous choices are accumulated across the repeated choice sets. This paper will explicitly discuss the roles of diffe rent RPs and explore the dynamic pattern of reference dependence behavior. To continue, ANA is another simplifying strategy that captures the reduced attention behavior . Decision making is based on a lexicographic heuristic, and only important attributes are given attention. One type of ANA behavior is that an attribute is always ignored regardless of its value. Another type of ANA behavior is based on the Elimination by Aspects (EBA) rule (Tversky, 1972). This rule starts by setting a cutoff value for eac h attribute, and then all attributes below the cutoffs are eliminated and thus ignored. The later type coincides with the reference dependence strategy. Under the RRM fr amework, the attribute(s) of the considered alternative performing better than that of the forgone alternative generate(s) zero regret and thus 106 will be statistically eliminated. Under such circumstances, an attribute contributing no regret will statistical ly be identified as ignored if it is modeled with the ANA framework. This phenomenon br ings up the necessity of accounting for reference dependence behavior in the ANA framework to avoid accidentally identifying reference dependence behavior as ANA behavio r. There are two primary methods to identify the ANA information processing strategy s tated ANA, which asks respondents which attributes have been ignored, and inferred ANA, which identifies ANA behavior through statistical inference without directly aski ng (Caputo et al., 2018). Whilst the stated approach is straightforward, it suffers the problem of inconsistency between stated and actual behaviors. On the one hand, the stated ignored attribute is not necessarily totally ignored. On the other hand, if th e question of ignorance is asked at the end of the serial choice sets, there is no guar antee that ignorance behavior has not changed across repeated choices. If the question of ignorance is inserted right after each choice set, the question itself is likel intere st in identifying the role of ANA through model inference (e.g., Caputo et al., 2013; Scarpa et al., 2010; Hess and Hensher, 2010; Hensher and Greene, 2010; Hole, 2011). This approach assumes multiple latent classes, which represent a different combination of attended attributes. The most popular latent class model is the Equality - Constrained Latent Class model (ECLC). in a latent class are restricted to zeros and either the same or different across classes (Scarpa et al., 2009; Caputo et al., 2013). This paper follows the inferred ANA framework to identify ANA behavior. Despite the popularity of examining ANA behavior using stated or inferred approaches in empiri cal DCE literature, there are several problems unsolved within the ANA framework. One 107 problem, as pointed by Hess et al. (2013), is if there exists substantial preferenc e heterogeneity that is not related to ANA, then setting a restriction to zero for para meters of the latent class model will produce results that confound non - attendance and taste heterogeneity. Besides, there lacks of discussion about the changing pattern s of ANA behavior over repeated choices in both stated and inferred ANA literature. Tha t is, ANA behavior is assumed to be respondents heterogeneous, but for each individual decision - maker, ANA behavior is assumed to be unchanged across choice sets. Moreov er, an inferred ANA behavior could be due to the failure of modeling alternative strate gies rather than due to true ANA behavior. To the best of my knowledge, all existing ANA literature builds on the utility maximization framework rather than the alternat ive. This paper solves the above problems by allowing for reference dependence behavior within the inferred ANA framework and discusses the changing pattern of ANA behavior, if there are any. The remainder of the paper is organized as follows. Section 2 i ntroduces models to account for different decision rules and discusses how to test thes e decision rules with these models. The survey and data used for empirically examining the models are introduced in section 3. Empirical estimation results are reported in section 4. Section 5 discusses the implications of incorporating behavioral factors for DCE. 4 .2. Econometric Frameworks and Hypothesis Testing There are three components of a DCE: a choice scenario composed of several alternatives, a decision rule de fined by a function that describes the observed utility, and an error term that describ es the unobserved utility and the associated distribution. This study focuses on identifying 108 the underlying decision rule through model development and hypothesis testin g. In this section, I discussed the models I developed to explore the decision rule. I began by introducing the traditional discrete choice modeling approach, i.e., RUM. I then discussed the RRM framework and the associated innovations as well as the ANA framework. I concluded by introducing an integrated framework to test the decision stra tegy. 4 .2.1. Random Utility Maximization (RUM) Assume the following choice scenario: a decision - maker, i , faces a choice scenario, s , composed of J alternatives, with each alternative, j , being described in terms of the attribute, m , i.e., . Utility Maximization Theory (McFadden, 1974) postulates that utility from alternative j is independent of other alternatives within the choice set, and the alternative with the highest utility will be chosen. The utility of each alternative i s described by a linear combination of observable attributes. A random error term is added to the utility to represent the inability to capture all factors that determine a decision - i alternat ive j in cho ice scenario s with taste parameters can be described as follows: Or Note that the RUM model restricts and by assuming that preferences are consistent over repeated choice scenarios. An SQ constant term is added to capture the status quo effect with when is the SQ alternative and otherwise. Here I 109 are different, and therefore I separately specify the corresponding preference parameters as an d = will reduced to conventional RUM . Under the assumption that the error term follows an independent and identically distributed, i. e. i.i.d., Extreme Value Type I with variance equaling , a multinomial logit (MNL) model can be used for model estimation (McFadden, 1974). The choice probability for alternative j is: . 4 .2.2. Random Regret Minimization (RRM) RRM framework, based on Regret Theory (RT) (Bell, 1982; Fishburn, 1982; Loomes and Sugden, 1982) , assumes that decision making depends not only on the performance of the consi dered alternative but also on that o f the foregone alternatives. The regret aversion emotion leaves individuals focusing on the loss rather than the gain generated from the bilateral comparison. As the counterpart of the RUM framework, RRM postulates that a decision - maker will choose the alt ernative with the lowest regret from the given choice set, and the regret is composed of a systematic regret described by observed attributes and an i.i.d random error . Regardless of the RUM or RRM framework, there exists a decision strategy for decision - underlying decision rule. 4 .2.2.1. Random Regret Minimization (RRM) A utility function is needed to translate the observable attributes into comparable lev els of regret as well as to account for the regret minimization assumption. Chorus (2010) defines attribute - 110 level regret as 13 where . That is, for an individual i in c hoice set s , the regret of alternative j for attribute m is determined by bilaterally comparing alternative j with every other referred alternative k . The regret is a function of loss based on th e comparison, i.e., X = . A non - linear log function form is used to represent that individuals respond more to loss (X>0) than to gain (X<0). This paper will focus on the case when , because when , RRM reduces to linear RU M (see Chorus, 2010, for a formal proof). Figure 3 . 1 plots regret over the loss, i.e., X = and explicitly illustrates how the non - linear function form of RRM captures the asymmetrical weights in loss and gain as opposed to the linear function form of RUM. Note that th is figure can also plot regret over the loss for the RUM framework because the regret function reduces to produces equal estimates as a RUM framework does (see, Chorus, 2014, for a formal proof). With attribute - level regret defined, an individual i j from choice set s can be defined as follows: Minimizing the regret is mathematically equivalent to maximizing the negative of the regret defined in equation (2). As such, the SQ constant term parameter, , has the opposite sign of the conventional RUM model as defined in equation (1). An MNL regression can be used to estimate the parameters of RRM in equation (2). 13 An alternative regret minimization model is defined in Chorus (2008) as follows: , where : is the preference parameter of attribute m. This formulation implies that rejoice gains zero weight in formu lating regret: when a considered alternative outperforms its competing alternative, i.e., , the regret is zero. 111 4 - RRM) An essential set forced choice and thus guarantees proper welfare measures (Hanley et al., 2001). This setting calls the issue s of the endowment effect the fact that people demand more to give up an object than they would be willing to acquire it (Thaler, 1980) and status quo bias the fact that a preference for the current state biases the decision - makers against foregoing the cu rrent status (Samuelson and Zeckhauser, 1988). Hence, the inclusion of an SQ specific constant term as I defined in the model specification is imp ortant. Besides the endowment effect and status quo effect, the inclusion of an SQ alternative also incurs th e behavior of reference dependence since SQ can also play as an RP. As discussed in the RRM section, giving up a hypothetical alternative that per forms better than the considered one can generate emotions of regret. Similarly, leaving the SQ that performs better than the considered alternative can cause regret as well. The only difference is that SQ serves as an internal reference point since it has been endowed with the decision - makers; in contrast, the hypothetical alternatives serve as an external refere nce point since they are introduced in the choice set design. This difference makes it reasonable to differentiate the impacts of SQ and hypotheti cal alternatives in serving as RPs. Chapter 3 explicitly investigates the contributions of SQ as RP and discus ses the importance of differentiating the contribution of SQ alternatives - RRM model that describes the r egret function as follows: 112 This mode l differentiates from the RRM model in th at it allows for different preference estimates for the comparisons with hypothetical and SQ alternatives, i.e., and . Additionally, it introduces a new parameter , where , such that it al lows the model to be reduced to a traditional uti lity maximization framework if the underlying behavior is utility maximization when 14 . 4 .2.2.3. Random Regret Minimization in the Presence of Path Dependence (P - RRM) If decision making is reference - dep endent, it is necessary to explore all possible RPs comprehensively. As I discussed above, both hypothetical and SQ can play as RP. The next question is whether the previous choice set(s) can also contribute as an RP if t he decision - makers are making repea ted choices in the survey. Specifically, when a decision - maker faces a choice scenario s, where , it is possible that the path, i.e., previous choice scenarios - 1 , will also play as an RP. To account for this path dependence behavior, Chapter 3 develops a Path Dependent Random Regret Minimization (P - RRM) model where the chosen alternativ e from the previous choice scenario s - 1 plays as the RP from the path. I, therefore, followed this work and defined the P - RRM as follows: 14 See Appendix B for a formal proof. 113 The innovation of this P - RRM model from th e previous RRM models is that it sets an additional part of regret generated from comparison with the last round chosen alternative, which is described by for each attribute m. 4 .2.3. Attributes Non - Attendance (ANA) Besides regret mi nimizati on behavior, ANA is another simplifying strategy for decision making. It assumes that decision - makers strategically ignore some of the attributes when evaluating the alternative provided in the choice tasks. The unattended attributes will be given zero/red uced weights in assessing the alternatives. ANA strategy is not necessarily exclusive to but can coincide with the reference dependence strategy. Under such a circumstance, decision - makers set the forgone alternative as the RP and eliminate the att ribute(s ) which perform(s) better, i.e., generate(s) no regret, for the considered alternative rather than for the forgone alternative. To explore the incidence of ANA behavior, I followed an Equality Constrained Latent Class (ECLC) framework (Scarpa et al ., 2009) to account for heterogeneous attention behaviors. The ECLC model assumes that the population of the respondents can be divided into a set number (Q) of classes with heterogeneous preferences. An individual belongs to each class with a certain prob ability where the probability belonging to each class sums up to one. Following the RUM framework for discrete choice modeling, the probability of individual i choosing alternative j when i belongs to class q can be described as follows: The probability that an individual belongs to a certain class q is given as: 114 where is the class paramet er to be estimated and is normalized to zero to secure identification of the model (Greene and Hensher, 2003). The probability that alternative j is chosen from J alternatives is a weighted average over the Q classes with weight Different from the standard latent class model, which is intended to explore preferen ce heterogeneity, the ECLC mode l is based on classes embedding different forms of attendance to attributes. Hence, the preference coefficients of the unattended attribute(s) belonging to a particular class are(is) restricted to zero(s). A stepwise approach (Lagarde, 2013) is used in thi s paper to avoid too many classes being generated. That is, I started with one single class with all attributes being attended. I then added additional classes with one attribute of each class not being attended. Furthermore, I kept the classes with non - ze ro probability from the previous step and added additional classes with one more attribute not being attended. I continued the process to exhaust the combination of all ANA classes. The existing ECLC framework of identifying heterogeneous attention classe s is built on the RUM framework. Hence, the utility function in equation (5) is defined as equation (1). But this does not restrict the ECLC framework from being extended to the RRM framework. That being said, I can ta ke the regret function defined in equations (2) (4) into equation (5) to conduct ECLC es timation. 4 .2.4. Decision Rule Testing To investigate the decision rule of DCE, I started with separately running each model, i.e., RUM, - RRM, and P - RRM, and I then allowed ANA behavior in each of these models. The 115 null hypothesis of running each m odel is that this model has captured the underlying decision rule. Furthermore, to explore the changing pattern of the decision rule over repeated choice sets of DCE, I separately ran each model over every single choice set and compared the corresponding performance over repeated choices. The n ull hypothesis is that the decision strategy does not change over repeated choices. The alternative hypothesis is that the decision strategy changes over repeated choices. The rationale of changing strategy is that s urvey respondents collect information fr om the repeated choices, and they may unconsciously or strategically use this information to make better decisions as well as reduce decision making cognitive burden. If the decision strategy is dynamic, failing to c apture such behavior will produce incons istent estimations over different choice sets. If there exists a dynamic decision - making strategy, one possible change is the general decision rule, and the other possible change is the information processed under ea ch decision rule. For instance, decision - makers might switch from utility maximization strategy to simplifying strategies such as regret minimization and ANA because the information acquired from the previous choices could help with simplifying the decisio n. Besides the decision rule, decision - m akers might also change the information processed, such as the RPs or the attended attributes. In the section on empirical analysis, I will explicitly investigate these decision patterns. 4 .3. Survey and Data As an empirical illustration of the approach, I used data from a choice experiment that elicits the farmers to adopt CA practices to reduce nitrogen fertilizer leakage into t he environment. To 116 illingness to adopt these CA practices, a mail survey with $2 cash incentives was conducted amongst corn growers in the Midwestern U.S., specifically in Michigan, Iowa, and Indiana in 2016. Mailing addresses for the survey were randomly drawn from the Farm Service Agency (FSA) 15 . With a response rate of 27%, there are 1,294 completed surveys. The survey contains four repeated choice experiment tasks, with each described by two hypothetical alternatives and one SQ alter native. Each alternative is described b y a payment vehicle and three CA practices plus an expected nitrogen saving. The payment level is suggested by a focus group study among farmers and adjusted after a pilot study of this survey, which was conducted in 2015. Attributes and attribute levels a re defined in Table 2. 2. The first three imposed. Expected nitrogen saving is decided by the combination of the three CA pr actices calculated by agronomy and environmental experts. Beside s the choice tasks, the survey contains - quo so that the SQ alternative can be linked with individual stated status quo values. To be compatible with a r eal decision scenario, the survey A practices after the choice tasks to avoid the questions affecting decision making. The individual status quo CA adoption levels as well as associated expected Nitrogen saving levels will be incorporated into the dataset for the later empirical estimation . A Bayesian efficiency design that minimizes D - error based on priors from the pilot survey and contains 24 choice sets is generated using Ngene software (Choice Metrics, 2012). I used a bloc k design with six blocks containing four choice sets for each to avoid fatigue effects. A respondent was randomly assigned to one of the blocks. The order of presentation and allocation to 15 FSA is the payment services agency within USDA. FSA has records for ever y farmer who receives any form of payment (direct payments, crop insurance subsidies, disaster payments, conservation payments, etc) through USDA . 117 respondents of the various choice sets is randomized. A sample of t he survey is attached in Figure 2. 5. Beyond discrete choice ques social demographic status, attitude toward the environment policy, different resources for information, as well as far ristics. Further details of this survey can be found in Chapter 2 . Sample characteristics by the state are summarized in Table 3. 2. After excluding the incomplete responses, the response rate is highest in Michigan, i.e., 27%, and lowest in Indiana, i.e., 21%. Opting out rate, i.e., the percentage of farmers choosing SQ alternative among the three alternatives, reaches its highest level in Iowa, i.e., 43%, and lowest in Indiana, i.e., 37%. The conservation tillage rate is significantly lower in Michigan th an that of the other two states, while the reduced tillage rate is highest in Michigan. Conservation program enrollment is relatively higher in Iowa. The distributions of age, gender, farming experience, and days off - farms are generally consistent across t he three states. Iowa has a higher percentage of farmers who com pleted Associate or higher - level degrees. Both farm product values and household incomes are higher in Iowa. Iowa and Indiana have more large farm owners. The status quo CA adoptions are diver ging across states: Iowa has the lowest rates of covering crops in the winter and side - dressing fertilizer but has the highest rate of avoiding fertilizer application in the fall. 4 .4. Empirical Estimation Results To test the decision rules, I ran MNL mo - RRM, and P - RRM specifications using Python Bi ogeme. I also examined these models separately on every single choice set to identify the decision strategy changing pattern over repeated choice scenarios. 118 4 - RRM and P - RRM Estim ations The estimation results of RUM - RRM, and P - RRM are rep orted in Table 4 .1 - 4 .3. For the 16 , the signs and the scales of the preference parameters are generally as expected. The three CA practices, i.e., covering crops in the winter ( Wint er ), avoid applying fertilizer in the fall ( Fall ), and side - dressing fertilizer ( Side ), all reduce utility, nitrogen - saving ( Nitrogen ) and payment ( Pay ) both increase utility, and leaving status quo (SQ) decreases utility. Note that is co nstant as 0 because the SQ alternative has no variation wit h $0 payment. The sensitivity with respect to attributes in SQ is stronger and more significantly than that in hypothetical alternatives. In addition to the sensitivity, SQ alternative also adds a positive Checki ng the scales of the parameters, I found that Winter is the least favored, followed by Side and Fall . This is consistent with the difficulty of adopting each practice, as suggested by the agronomy experts as well as the actual adoption rate claimed in the survey. 22 %, 85%, and 44 % of the survey respondents have met the requirements of Winter , Fall , and Side , respectively. This finding persists after running the RUM model on every single choice set separately. - RRM model, which allows for within choice set reference dependence behavior. Here instead of estimating in Equ (3), I set = 1 following the RRM framework because regret minimization behavior has been verified in Chapter 3 and measuring regret extent, i.e. , is out of the scope of this chapter. I did the same for P - RRM modeling defined in Equ (4). Not - RRM, I a llowed hypothetical alternatives and SQ alternatives to make different contributions to regret generation. Compa red with the RUM model, the - estimations of preference parameters are consistent except for 16 is rejected in chapter 3 . 119 attribute Fall , but the estimations of are only consistent for Nitrogen and Pay . Specifically, when referring to an SQ alternative, Winter and Side significantly contribute to increasing regret level, and Nitrogen and Pay significantly contribute to decre asing regret level. However, th is is not true for Winter , Fall, and Side when referring to the hypothetical alternatives. Further checking the performances of this same model on each single choice set, I found similar results. One reason for the unexpected signs of and could be that the model confou nds the estimations of with that of due to the restriction of the sample size. As SQ is a more critical RP, is more likely to be significantly identified whe n the sample size is small. Lastly, comparing the statistical criteria, I found t - RRM slightly outperforms RUM in terms of the likelihood value , the Akaike information criterion (AIC), and the Bayesian information criterion (BIC). Furthermore, I allowed decision making to be path - dependent by incorporating the last information for the first choice set, I tested this model based on choice sets 2, 3, and 4. I found that the last round chosen alternative is the most important RP. It produces expected signs of estimates, and the estimations are all significant at the 1% level . SQ alternative produces consistent estimations for all attributes except for Fall . Hypothetical alternative produces consistent estimations for all attributes e xcept for Winter and Side . Besides, the scales of estimations for the same attribute are largest for the last round chosen alternative and smallest for the hypothetical alte rnative. The expected signs, scale of estimations, significance of estimations for the last round chosen alternative persist in each single choice set. Besides that, this P - RRM model significantly outperforms RUM and SQ - RRM in terms of the likelihood value , AIC, 120 and BIC. These findings raise the necessity of accounting for path dependence decision behavior in a DCE. In summary, after comparing the estimations of models RUM - RRM, and P - RRM, I easonable RP if the current choice set is not the first one. Besides regret generation in terms of the attributes Winter , Side, Nitrogen, and Pay . The hypothetical alternative contribut es to regret generation in terms of the attribute s Fall, Nitrogen, an d Pay . There are a few confounding estimates regarding the signs and significance levels considering the SQ and hypothetical alternatives as RPs. A possible explanation is the increase in model complexity significantly increases the minimal sample size req uired for statistical identification. Restricted by the sample size, only the most important parameters are significantly identified. 121 Table 4. 1. RUM' Estimations Sample Set = 1, 2, 3 , 4 Set = 1 Set = 2 Set = 3 Set = 4 - 0.718*** a - 0.969*** - 0.842*** - 0.437 - 0.609** (0.145) b (0.312) (0.283) (0.296) (0.302) 0.00343 - 0.222 - 0.0309 0.175 0.117 (0.0821) (0.179) (0.166) (0.157) (0.172) - 0.0133 - 0.222 - 0.154 0.344 - 0.0522 * (0.196) (0.429) (0.376) (0.402) (0.403) 1.06 2.75 1.71 - 0.892 0.615 (0.813) (1.8) (1.55) (1.64) (1.7) 0.00827*** 0.00648*** 0.00773*** 0.00828*** 0.00941*** (0.000472) (0.00108) (0.000961) (0.00097) (0.000909 ) - 0.752*** - 0.722 - 0.853* - 0.475 - 0.915* (0.249) (0.518) (0.476) (0.506) (0.51) - 0.231 0.00952 - 0.25 - 0.298 - 0.275 (0.173) (0.357) (0.326) (0.357) (0.35) - 1.26*** - 0.904 - 1.51* - 1.01 - 1.4 1 (0.434) (0.899) (0.825) (0.89) (0.883) 2.44 0.242 3.8 1.35 3.48 (1.72) (3.6) (3.26) (3.5) (3.49) - - - - - 1.2*** 1.02*** 1.15*** 1.15*** 1.3*** (0.0759) (0.161) (0.16) (0.148) (0.153) Model Statistics - 4434 - 1109 - 1108 - 1112 - 1085 AIC/N 2.07 2.08 2.08 2.18 2.04 BIN/N 2.08 2.13 2.13 2.14 2.09 N 4300 1075 1075 1075 1075 a. *** , ** , * . b. Robust standard error is reported in the bracket. 122 Table 4. 2. G' - RRM Estimatio ns Sample Set = 1, 2, 3, 4 Set = 1 Set = 2 Set = 3 Set = 4 0.435*** a 0.623*** 0.283 0.787*** 0.384 (0.121) b (0.218) (0.225) (0.312) (0.213) - 0.134** - 0.237 - 0.093 - 0.0517 - 0.0316 (0.0589) (0.273) (0.107) (0.0929) (0.106) 0.58*** - 0.777*** 0.496** 0.81*** 0.497** (0.0866) (0.171) (0.205) (0.195) (0.157) 0.738*** 0.0875 0.821 0.284 0.762 (0.22) (0.331) (0.593) (0.463) (0.417) 0.00498*** 0.00416*** 0.00395* ** 0.00531*** 0.00597*** (0.000374) (0.00077) (0.000789) (0.000767) (0.000687) - 1.95*** - 2.07*** - 1.74*** - 2.3*** - 1.85*** (0.154) (0.287) (0.294) (0.354) (0.298) 0.033 0.263 - 0.0257 - 0.00537 0.00402 (0.042 ) (0.453) (0.11) (0.0176) (0.0202) - 1.79*** - 1.77*** - 1.59*** - 2.08*** - 1.81*** (0.117) (0.229) (0.253) (0.238) (0.236) 1.1*** 0.144 1.14 0.453 1.16 (0.383) (0.531) (1) (0.765) (0.782) 0 .037*** 0.00989*** 0.0327** 0.0402*** 0.0384** (0.00638) (0.0038) (0.0152) (0.00813) (0.00839) - 2.13*** - 1.63*** - 2.13*** - 2.08*** - 2.33*** (0.0656) (0.131) (0.146) (0.112) (0.115) Model Statistics - 4418 - 1122 - 1110 - 1097 - 1073 AIC/N 2 .06 2.11 2.09 2.06 2.02 BIN/N 2.05 2.16 2.14 2.11 2.07 N 4300 1075 1075 1075 1075 a. *** , ** , * . b. Robust standard error is reported in the bracket. 123 Table 4. 3. P - RRM Estimations Sample Set=2, 3, 4 Set=2 Set=3 Set=4 1.12*** a - 0.257 1.72*** 0.201 (0.114) b (0.213) (0.247) (0.242) - 0.838*** - 0.871*** - 1.01*** - 0.906*** (0.105) (0.172) (0.162) (0.177) 0.94*** 0.107 - 0.0363 0.185 (0.103) (0.293) (0.27) (0.192) 0.825*** - 0.302 0.254 0.255 (0.227) (0.648) (0.569) (0.393) 0.00637*** 0.00543*** 0.00708*** 0.00717*** (0.000416) (0.000856) (0.00088) (0.000802) - 1.7*** 1.63*** - 1.76*** 1.31*** (0.173) (0.32) (0.409) (0.39 7) - 0.192 0.0186 0.00922 0.0658 (0.119) (0.213) (0.14) (0.184) - 1.46*** - 1.06*** - 1.8*** - 1.3*** (0.158) (0.36) (0.281) (0.296) 1.86*** - 0.24 0.297 0.368 (0.59) (0.56) (0.67) (0.581) 0.0285*** 0.0345* 0.0275*** 0.0403*** (0.00799) (0.0192) (0.00782) (0.0162) - 2.07*** - 2.31*** - 2.72*** - 2.6*** (0.111) (0.21) (0.22) (0.202) - 1.79*** - 2.01*** - 2.19*** - 1.7*** (0.111) (0.211) (0.202) (0.243) - 1.86*** - 2.32*** - 2.42*** - 2.45*** (0.116) (0.249) (0.213) (0.194) 2.63*** 4.96*** 1.65 1.21 (0.654) (1.22) (1.73) (1.52) 0.0025*** 0.00277*** 0.00316*** 0.00227*** (0.0004 37) (0.000758) (0.000835) (0.000906) - 2.3*** - 2.36*** - 2.24*** - 2.42*** (0.0863) (0.168) (0.136) (0.151) Model Statistics - 4198 - 997 - 971 - 990 AIC/N 1.96 1.88 1.84 1.87 BIC/N 1.98 1.95 1.91 1.95 N 4300 1075 1075 1075 a. *** , ** , * . b. Robust standard error is reported in the bracket. 124 4 .4.2. RUM - RRM, and P - RRM Estimations Accounting for the ANA Behavior changing pattern over repeated choices. I allowed for ANA behavior with RUM - RRM, and P - restricted to zero. To figure out a proper number of classes for the ANA model, I followed a stepwise a pproach (Lagarde, 2013) by starting with one class that all attributes are attended, and then added additional classes that only one attribute of each class is not attended. Furthermore, I kept the classes with non - zero probability from the pre vious step a nd continued to add additional classes with more attributes not being attended until I exhausted the combination of all ANA classes. After this stepwise approach, I found that only five classes are common scenarios. That is, I have the first cl ass with all of the five attributes attended and the other four classes with one of Winter , Fall , Side , and Nitrogen non - attended. Further checking the probability of each class, I found that Fall non - attended is the only class with significant non - zero pr obability. T herefore, I proceeded with the model examination with only two classes, i.e., all attended and only Fall non - attended. With the number of classes defined, I ran RUM - RRM, and P - RRM defined in Equ (1) (3) with and two latent classes to examine t coefficients are reported in Table 4 . 4 - 4 . 6 . Note that the class probability parameter for all attributes attended (AA) class is fixed at zero for identification. I found that the model that accounts for the ANA behaviors are generally consistent with the model that does not account for ANA behaviors with respect to parameter estimation. Checking the class probability for the RUM - ANA model, I found that the class that Fall is unattended is significant wi th high probability starting from choice set 2 . It is worth to mention that after the ANA behavior is 125 accounted for, the estimation of Fall expected. One exp lanation is that Fall is not important fo explanation is that the sample size is too small to identify Fall Fall non - attended observations are excluded; a third explanation is that Fall is assessed in a way not correctly specified by this - RRM model in Table 4. 4, I found that the class that Fall is non - attended is still close to 100%, and the associated estimation is significant on any single choice set . Finally, I explored the ANA behaviors w ith the P - RRM - ANA model. I no longer observed any significant or greater than zero probability of ANA class after controlling for the path dependence behavior. The last round chosen alternative pl ays a significant r ole in regret generation with correct sig ns . Note that ANA models that do not account for path dependence RUM - - RRM - ANA indicate that ANA behavior emerges starting from the second or the third choice set. However, after controlling for path depend ence behavior, ANA behavior no longer per sists. for all the five attributes are significantly identified with correct signs. Meanwhile, Fall is correctly and significantly identified for all three positions, i.e., The findin gs above imply that it is not that respondents gradually lost attention or some attributes are not important , but that respondents gradually shifted their RPs to process each attribute and alternative . This changing pattern of ANA behavior across choice se ts manifests the importance of accounting for the path dependence behavior in discrete choice modeling. Attributes can be mistakenly interpreted as insignificant when they are actually important if the actual underlying decision rule is not accounted for. Finally, I found that the statistical performances of the three models - RRM, and P - RRM do not change after acco unting for ANA behavior. The likelihood value, AIC, and 126 BIC of ANA models all stay basically the same as the models without controlling f or ANA behavior . This again justifies the unnecessity of controlling for ANA behavior. To conclude, the findings above show that the inferred ANA approach is sensitive to model specification. Imposing model structures that best describe the underlying dec ision strategies is the prerequisite of correctly identifying ANA behaviors. After path dependence behavior is accounte d for, I no longer observed evidence of ANA behavior. Besides, reference dependence behavior is consistent over repeated choices. Decisio n - makers gradually shift their RP from the current choice set to the previously chosen alternative as they gradually co llect information from the path. 127 Table 4. 4. RUM' Estimations with Inferred ANA Sample Set = 1, 2, 3, 4 Set = 1 Set = 2 Set = 3 Set = 4 Class AA ANA - Fall AA ANA - Fall AA ANA - Fall AA ANA - Fall AA ANA - Fall - 0.453*** - 0.453*** - 0.665*** - 0.665*** - 0.497*** - 0.497*** - 0.484*** - 0.484*** - 0.772*** - 0.772*** (0.0521) (0.0521) (0.0766) (0.0766) (0.114) (0.114) (0.0997) (0 .0997) (0.136) (0.136) 0.141*** 0 - 0.0254 0 0.258*** 0 0.152*** 0 0.0844 0 (0.0336) - (0.0937) - (0.101) - (0.0953) - (0.0805) - - 0.361*** - 0.361*** - 0.206 - 0.206 - 0.36* - 0.36* - 0.275** - 0.275** - 0.283 - 0.283 (0.102) ( 0.102) (0.165) (0.165) (0.192) (0.192) (0.185) (0.185) (0.111) (0.111) 1.55*** 1.55*** 1.42** 1.42** 0.693 0.693 1.39*** 1.39*** 1.87*** 1.87*** (0.298) (0.298) (0.618) (0.618) (0.876) (0.876) (0.664) (0.664) (0.448) (0.448) 0.00826** * 0.00826*** 0.00652** * 0.00652** * 0.00764** * 0.00764** * 0.00818** * 0.00818*** 0.00931** * 0.00931*** (0.000392) (0.000392) (0.001) (0.001) (0.000615) (0.000615) (0.000796) (0.000796) (0.000864) (0.000864) - 0.36 6** - 0.366** - 0.793*** - 0.793*** - 0.459*** - 0.459*** - 0.266 - 0.266 - 0.478** - 0.478** (0.188) (0.188) (0.235) (0.235) (0.178) (0.178) (0.188) (0.188) (0.196) (0.196) - 0.0537 0 - 0.0228 0 0.333 0 - 0.219 0 - 0.0912 0 (0.108) - (0.187) - (0.6) - (0.196) - (0.127) - - 0.569*** - 0.569*** - 0.997*** - 0.997*** - 0.787** - 0.787** - 0.63*** - 0.63*** - 0.642** - 0.642** (0.0757) (0.0757) (0.24) (0.24) (0.213) (0.213) (0.16) (0.16) (0.179) (0.179) 0.683 0.683 1.07 1.07 1.11 1.11 0.267 0.267 0.53 0.53 (0.612) (0.612) (0.936) (0.936) (0.965) (0.965) (0.699) (0.699) (0.669) (0.669) 0 0 0 0 0 0 0 0 0 0 - - - - - - - - - - 1.2*** 1.2*** 1.06*** 1.06*** 1.07*** 1.07*** 1.12*** 1. 12*** 1.26*** 1.26*** (0.114) (0.114) (0.146) (0.146) (0.173) (0.173) (0.189) (0.189) (0.112) (0.112) ANA Class Estimations Class Par a 0 b 38.7*** 0 0.136 0 1.59** 0 33.5*** 0 1.3*** 128 - (0.00000319) - (2.12) - (0.153) - (0.00000116 ) - (0.0000272) Class Prob 0% 100% 47% 53% 17% 83% 0% 100% 21% 79% Model Statistics l - 4422 - 1108 - 1109 - 1111 - 1084 AIC/N 2.06 2.08 2.08 2.09 2.04 BIC/N 2.08 2.14 2.14 2.15 2.09 N 4300 1075 1075 1075 1075 a. Latent class parameter s in Equ (6). b. Latent class parameter, i.e., s, for the all attributes attended (AA) class is fixed at zero. 129 Table 4. - RRM Estimations with Inferred ANA Sample Set = 1, 2, 3, 4 Set = 1 Set = 2 Set = 3 Set = 4 Class AA ANA - Fall AA ANA - Fall AA ANA - Fall AA ANA - Fall AA AN A - Fall 0.86*** 0.86*** 0.704*** 0.704*** 0.229* 0.229* 1.01*** 1.01*** 0.611*** 0.611*** (0.158) (0.158) (0.127) (0.127) (0.128) (0.128) (0.271) (0.271) (0.293) (0.293) - 0.38*** 0 - 0.119 0 1.12 0 - 2.04 0 8.66 0 (0.119) - (0 .288) - (1.8) - (2.45) - (24.26) - - 0.545*** - 0.545*** - 0.618*** - 0.618*** - 0.192 - 0.192 - 1.07*** - 1.07*** - 0.723*** - 0.723*** (0.116) (0.116) (0.11) (0.11) (0.13) (0.13) (0.11) (0.11) (0.0874) (0.0874) 1.92*** 1 .92*** - 0.811 - 0.811 1.96*** 1.96*** 3.29*** 3.29*** 3.21*** 3.21*** (0.188) (0.188) (0.561) (0.561) (0.453) (0.453) (0.484) (0.484) (0.363) (0.363) 0.0075*** 0.0075*** 0.00616*** 0.00616*** 0.00614*** 0.00614*** 0.00815*** 0.00815*** 0.00 846*** 0.00846*** (0.00037) (0.00037) (0.000732) (0.000732) (0.000586) (0.000586) (0.000577) (0.000577) (0.000679) (0.000679) - 1.98*** - 1.98*** - 2.03*** - 2.03*** - 1.84*** - 1.84*** - 2.24*** - 2.24*** - 1.88*** - 1.88*** (0.199) (0.1 99) (0.209) (0.209) (0.237) (0.237) (0.323) (0.323) (0.373) (0.373) - 0.132 0 0.15 0 6.88 0 10 0 - 10 0 (0.173) - (0.512) - (10.3) - (1.80e+308) - (1.80e+308) - - 1.8*** - 1.8*** - 1.77*** - 1.77*** - 1.49*** - 1.49*** - 2. 01*** - 2.01*** - 1.72*** - 1.72*** (0.0916) (0.0916) (0.157) (0.157) (0.226) (0.226) (0.186) (0.186) (0.138) (0.138) 5.52*** 5.52*** 0.734 0.734 9.51*** 9.51*** 7.52*** 7.52*** 5.96*** 5.96*** (0.234) (0.234) (0.731) (0.731) ( 0.808) (0.808) (0.563) (0.563) (0.659) (0.659) 7.03*** 7.03 6.39** 6.39** 5.47*** 5.47*** 3.05*** 3.05*** 2.94*** 2.94*** (2.97) (2.97) (3.40) (3.40) (0.0786) (0.0786) (0.0741) (0.0741) (0.0713) (0.0713) - 1.26*** - 1.26*** - 0.8 62*** - 0.862*** - 1.28*** - 1.28*** - 1.27*** - 1.27*** - 1.45*** - 1.45*** (0.0539) (0.0539) (0.0952) (0.0952) (0.101) (0.101) (0.0935) (0.0935) (0.0953) (0.0953) ANA Class Estimations Class Par a 0 b 21.7*** 0 32.2*** 0 48.5*** 0 2.45** 0 33.7*** - (1.12) - (0.0153) - (0.0183) - (1.27) - (0.158) 130 Class Prob 0% 100% 0% 100% 0% 100% 8% 92% 0% 100% Model Statistics - 4410 - 1123 - 1066 - 1083 - 1072 AIC/N 2.06 2.11 2.01 2.04 2.02 BIC/N 2.07 2.17 2.06 2.09 2.07 N 4300 1075 1075 1075 1075 a. Latent class parameter s in Equ (6). b. Latent class parameter, i.e., s, for the all attributes attended (AA) class is fixed at zero . 131 Table 4.6 . P - RRM Estimations with Inferred ANA Sample Set = 2, 3, 4 Set = 2 Set = 3 Set = 4 Class AA ANA - Fall AA ANA - Fall AA ANA - Fall AA ANA - Fall 1.33*** 1.33*** 1.15*** 1.15*** 1.8*** 1.8*** 1.27*** 1.27*** (0.169) (0.169) ( 0.143) (0.143) (0.217) (0.217) (0.448) (0.448) - 1*** 0 - 1.05*** 0 - 1.05*** 0 - 0.997*** 0 (0.0749) - (0.14) - (0.138) - (0.148) - 0.167 0.167 0.251 0.251 0.11 0.11 0.283* 0.283* (0.111) (0.111) (0.183) (0.183) (0.249) ( 0.249) (0.154) (0.154) 0.243 0.243 - 0.199 - 0.199 - 0.0288 - 0.0288 0.115 0.115 (0.259) (0.259) (0.592) (0.592) (0.516) (0.516) (0.104) (0.104) 0.00336*** 0.00336*** 0.00137 0.00137 0.00354*** 0.00354*** 0.00478*** 0. 00478*** (0.000693) (0.000693) (0.00129) (0.00129) (0.000983) (0.000983) (0.000898) (0.000898) - 1.5*** - 1.5*** - 1.56*** - 1.56*** - 1.88*** - 1.88*** - 1.14 - 1.14 (0.301) (0.301) (0.291) (0.291) (0.383) (0.383) (0.978) (0.978) - 0.0611 0 - 0.0297 0 - 0.0265 0 - 0.0246 0 (0.106) - (0.208) - (0.158) - (0.128) - - 1.4*** - 1.4*** - 1.04*** - 1.04*** - 1.85*** - 1.85*** - 1.39*** - 1.39*** (0.119) (0.119) (0.29) (0.29) (0.21) (0.21) (0.167) (0.167) 2.17* 2.17* 6.14*** 6.14*** 1.55 1.55 0.27 0.27 (1.27) (1.27) (1.19) (1.19) (2.66) (2.66) (0.251) (0.251) 5.32*** 5.32*** 7.53 7.53 3.66*** 3.66*** 2.63*** 2.63*** (0.0959) (0.0959) (1.80e+308) (1.80e+308) (0.193) (0.193) (0.107) (0.107) - 2.44*** - 2.44*** - 2.14*** - 2.14*** - 2.68*** - 2.68*** - 2.44*** - 2.44*** (0.111) (0.111) (0.143) (0.143) (0.207) (0.207) (0.245) (0.245) - 1.96*** 0 - 2.38*** 0 - 2.18*** 0 - 1.67*** 0 (0.177) - (0.186) - (0.217) - (0.269) - - 2.4*** - 2.4*** - 2.27*** - 2.27*** - 2.4*** - 2.4*** - 2.35*** - 2.35*** 132 (0.114) (0.114) (0.232) (0.232) (0.22) (0.22) (0.225) (0.225) 0.506 0.506 1.29 1.29 1 .052 5 1 .0525 0.264 0.264 (0.838) (0.838) (1.03) (1.03) (0.915) (0.915) (0.424) (0.424) 0.0149*** 0.0149*** 0.0164*** 0.0164*** 0.0155*** 0.0155*** 0.0162*** 0.0162*** (0.00129) (0.00129) (0.00185) (0.00185) (0.00211) (0.00211) (0.00156) (0. 00156) - 1.35*** - 1.35*** - 1.37*** - 1.37*** - 1.22*** - 1.22*** - 1.55*** - 1.55*** (0.0688) (0.0688) (0.116) (0.116) (0.0948) (0.0948) (0.0948) (0.0948) ANA Class Estimations Class Par a 0 b - 25.3*** 0 - 0.53 0 - 59.3*** 0 - 38.8*** - (0.00000126) - (0.83) - (0.00000588) - (0.00000257) Class Prob 100% 0% 63% 37% 100% 0% 100% 0% Model Statistics - 2959 - 985 - 965 - 987 AIC/N 1.85 1.86 1.83 1.88 BIC/N 1.88 1.94 1.91 1.95 N 3225 1075 1075 1075 a. Latent class parameter s in Equ (6). b. Latent class p arameter, i.e., s, for the all attributes attended (AA) class is fixed at zero. 133 4 .5. Conclusions There is increasing literature incorporating behavioral strategies into discrete choice modeling. This study discusses how the respondents in a DCE use these behavioral strategies to make decisions, how such strategies evolve over repeated choice tasks, an d how failing to identify these strategies leads to confounding conclusions. One behavioral strategy developed in the DCE literature is ANA. Respondents str ategically or unconsciously ignore some attributes in a choice task to reduce cognitive burdens of decision making. Another behavioral strategy is RM. Respondents use RP to cancel out shared attributes and make choices depending on the net gain or loss fro m the bilateral comparisons between alternatives. The asymmetric weights on gain and loss leave som e attributes of certain alternatives appearing to be non - attended. This paper, for the first time, discusses the relation of RM with ANA. The empirical anal ysis is based on a DCE survey conducted in the Midwest states of Michigan, Iowa, and Indiana on far - paid fertilizer management program. This is among the first few literatures investigating RM or ANA behaviors in a WTA choice scena rio as well as in environmental economics. However, the findings have general implications in field s outside the environmental economics literature. In addition, even though the proposed government program in the survey provides public benefits, the adopti on of this program from the perspective of survey respondents, i.e., corn growers, is still an indi vidual decision. So, the behavior identified in this paper will also apply to other private good decision - making scenarios. The first finding is that decisi on making is choice set composition as well as path dependent. The hypothetical alternatives, the i nformation carried in the SQ alternative, and the 134 the last rou - makers are first exposed to the ch oice set where there is no last round of information delivered, the SQ is the most important RP. As survey respondents gradually collect information over rep eated choices, the decision - making shifts to path dependence. That is, decision making evolves from the current choice set dependent to across choice sets dependent. Moreover, the RPs are attributes specific. For instance, Winter and Fall contribute to reg ret generation in all referred scenarios, but Side only contributes to regret in the SQ and the las t round chosen referred scenarios. To continue, although I found evidence of ANA behavior based on RUM - RRM specifications, I no longer observed this behavior once I allowed respondents to condition their current decision on their choices in the pr evious choice scenario(s) in the extended RRM model , i.e, P - RRM . For instance, the RUM model shows significant ANA behavior starting from the choice set 2. This inferred ANA behavior can be interpreted as reduced attention over repeated choices due to the fatigue effect, as discussed in previous ANA literature. However, after I longer significant. This implies that the attributes that are otherwise interpreted to be non - atte nded might be , in fact, attended in a path - dependent regret minimization manner. Thus, ANA behavior can be the result of model misspecification rather than a true decision strategy. Finally, this study suggests that P - RRM can be a guiding choice in DCE mo deling. I showed that the reference - dependent regret minimization model matches respondent behavior better than RUM . Meanwhile, P - RRM captures the behavior strategies that are otherwise identified as ANA in a more informative manner. More importantly , thi s reference - dependent behavior brings a new perspective to evaluate the incentive compatibility property of the DCE 135 method . Discrete choice researchers need to account for the across choice sets dependent decision behaviors in the DCE survey design. Otherw ise, the preference estimation from the DCE would be survey design dependent and cannot be adapted to general decision - making predictions. There are several issues left with future discussions. First, we need to be cautious that these conclusions are base d on an experimental study. Whether the decision - makers adopt similar decision strategies in a real choice scenario remains further examination. Second, the rejection of the ANA strategy in favor of the path - dependent RM strategy is built on a four - choice sets survey with five attributes. Whether there exists an ANA behavior in a survey that includes more choice tasks and/or more attributes remains further inv estigation. It is also worth mentioning that the conclusions are based on an ECLC model, which rest ricts the coefficients of non - attended attributes to be zeros. It is possible that the non - discounted but non - zero weights. Whether survey respondents use ANA strategies and under what circumstance they use these strategies are left for further research. Last, among the the RP. Fu ture work is needed to examine the role of other previously delivered information. 136 REFERENCES 137 REFERENCES Adamowicz, W.L., Glenk, K. and Meyerhoff, J., 2014. Choice modelling research in environmental and resource economics. Chapters, pp.66 1 - 674. Bell, D.E., 1982. Regret in decision making under uncertainty. Operations Research 30 (5), 961 981. Boeri, M., Scarpa, R., & Chorus, C. G., 2014. Stated choices and benefit estimates in the context of traffic calming schemes: Utility maximization, r egret minimization, or both? Transportation Research A, 61, 121 135. 2014.01.003. Caputo, V., Lusk, J.L. and Nayga Jr, R.M., 2018. Choice experiments are not conducted in a vacuum: The effects of external price information on choice behavior. Journal of Ec onomic Behavior & Organization, 145, pp.335 - 351. Caputo, V., Lusk, J.L. and Nayga, R.M., 2020 Decision Making When the Reference Price Is Uncertain. American Journal of Agricultural Economics, 102(1), pp.132 - 1 53. Caputo, V., Nayga Jr, R.M. and Scarpa, R., 2013. Food miles or carbon emissions? Explorin g labelling preference for food transport footprint with a stated choice study. Australian Journal of Agricultural and Resource Economics, 57(4), pp.465 - 482. Caput o, V., Van Loo, E.J., Scarpa, R., Nayga Jr, R.M. and Verbeke, W., 2018. Comparing serial, and experiments. Journal of Agricultural Economics, 69(1), pp.35 - 57. Chorus, C., 2014. A Generalized Random Regret Minimization model. Trans. Res. Part B, Pages 224 - 238. Chorus, C., Arentze, T., and Timmermans, H., 2008. A random regret minimization model of travel choice. Trans. Res. Part B, 42(1), 1 18. Chorus, C.G. and Bierlaire, M., 20 13. An empirical comparison of travel choice models that capture preferences for compromise a lternatives. Transportation, 40(3), pp.549 - 562. Chorus, C.G., 2010. A new model of random regret minimization. EJTIR, 10 (2), 2010. Fishburn, P.C., 1982. Non - trans itive measurable utility. Journal of Mathematical Psychology 26 (1), 31 67. Gilovich, T., Gri ffin, D. and Kahneman, D. eds., 2002. Heuristics and biases: The psychology of intuitive judgment. Cambridge university press. Greene, W.H. and Hensher, D.A., 2003 . A latent class model for discrete choice analysis: contrasts with mixed logit. Transportati on Research Part B: Methodological, 37(8), pp.681 - 698. 138 Hanley, N., Mourato, S. and Wright, R.E., 2001. Choice modelling approaches: a superior alternative for envi ronmental valuatioin?. Journal of economic surveys, 15(3), pp.435 - 462. Hart, S., 2005. Adaptive heuristics. Econometrica 73 (5), 1401 1430. Hensher, D., Greene, W., Chorus, C., 2013. Random regret minimisation or random utility maximisation: an exploratory analysis in the context of automobile choice. J. Adv. Transp. Hensher, D.A. and Greene, W.H., 2010. Non - attendance and dual processing of common - metric attributes in choice analysis: a latent class specification. Empirical economics, 39(2), pp.413 - 426. He nsher, D.A. and Rose, J.M., 2012. The influence of alternative acceptab ility, attribute thresholds and choice response certainty on automobile purchase preferences. Journal of Transport Economics and Policy (JTEP), 46(3), pp.451 - 468. Hensher, D.A., Rose, J . and Greene, W.H., 2005. The implications on willingness to pay of res pondents ignoring specific attributes. Transportation, 32(3), pp.203 - 222. Hess, S. and Hensher, D.A., 2010. Using conditioning on observed choices to retrieve individual - specific attrib ute processing strategies. Transportation Research Part B: Methodologic al, 44(6), pp.781 - 790. te non - Attendance and taste heterogeneity. Transportation, 40(3), pp.58 3 - 607. Hey, J.D., Orme, C., 1994. Investigating generalizations of expected utility theory using experimental data. Econometrica 62 (6), 1291 1326. Hole, A.R., 2011. A discrete choice m odel with endogenous attribute attendance. Economics Letters, 110(3), p p.203 - 205. Houston, D.A. and Sherman, S.J., 1995. Cancellation and focus: The role of shared and unique features in the choice process. Journal of Experimental Social Psychology, 31(4), pp.357 - 378. Kanninen, B.J. ed., 2007. Valuing environmental amenities using stated choice studies: a common sense approach to theory and practice (Vol. 8). Springer Science & Business Media. s consequences in choice experiments with latent class models. Health e conomics, 22(5), pp.554 - 567. Layton, D. and Hensher, D.A., 2010. Aggregation of common - metric attributes in preference re velation and implications for willingness to pay. Transportation Research Part D: Transport and Environment, 15(7), pp.394 - 404. Loomes, G., Sugden, R., 1982. Regret theory: an alternative theory of rational choice under uncertainty. The Economic Journal 92 , 805 824. 139 Loomes, G., Sugden, R., 1983. A rationale for preference reversal. The American Economic Review 73 (3), 428 432. Loomes , G., Sugden, R., 1987. Some implications of a more general form of regret theory. Journal of Economic Theory 41 (2), 270 287. Louviere, J.J. and Hensher, D.A., 1982. On the design and analysis of simulated choice or allocation experiments in travel choice modelling. Transportation research record, 890(1982), pp.11 - 17. Louviere, J.J. and Woodworth, G., 1983. Design and analysis o f simulated consumer choice or allocation experiments: an approach based on aggregate data. Journal of marketing research, pp.350 - 367. Louviere, J.J., Hensher, D.A. and Swait, J.D., 2000. Stated choice methods: analysis and applications. Cambridge universi ty press. Machina, M.J., 1987. Choice under uncertainty: problems solved and unsolved. Journal of Economic Perspectives 1, 121 154 . McFadden, D., 1974. Conditional logit analysis of qualitative choice behaviour. In: Zarembka, P. (Ed.), Frontiers in Econome trics. Academic Press, pp. 105 142. Metrics, C., 2012. Ngene 1.1. 1 user manual & reference guide. Sydney, Australia: ChoiceMetric s, 19, p.20. Quiggin, J., 1994. Regret theory with general choice sets. Journal of Risk and Uncertainty 8 (2), 153 165. Samue and Uncertainty, 1988, 1, 7 59. Scarp a, R., Gilbride, T.J., Campbell, D. and Hensher, D.A., 2009. Modelling Attribute non - Attendance in choice experiments for rura l landscape valuation. European review of agricultural economics, 36(2), pp.151 - 174. Scarpa, R., Thiene, M. and Hensher, D.A., 201 0. Monitoring choice task attribute attendance in nonmarket valuation of multiple park management services: does it matter?. L and economics, 86(4), pp.817 - 839. Starmer, C., 2000. Developments in non - expected utility theory: The hunt for a descriptive theor y of choice under risk. Journal of economic literature, 38(2), pp.332 - 382. Swait, J., 2001. A non - compensatory choice model in corporating attribute cutoffs. Transportation Research Part B: Methodological, 35(10), pp.903 - 928. Thaler, R., 1980. Toward a positive theory of consumer choice. Journal of Economic Behavior & Organization, 1(1), pp.39 - 60. Thiene, M., Boeri, M. and Chorus, C.G., 2012. Random regret minimization: exploration of a new choice model for environmental and resource economics. Environme ntal and resource economics, 51(3), pp.413 - 429. 140 Tian, Q., and Zhao., J. 2019. Regret Minimization in Decision making: Implications for Choice Modeling and Policy Design. Working paper. Tian, Q., Zhao, J., Reimer, A., and Lupi, F., 2019. Farmers Decision o n Nitrogen Application: Testing Information Treatment Effect and Commitment Cost Theory. Working paper. Tversky, A., 1969. Intran sitivity of preferences. Psychological review, 76(1), p.31. Tversky, A., 1972. Elimination by aspects: A theory of choice. Psy chological review, 79(4), p.281. 141 CHAPTER 5. C ONCLUSIONS As global nitrogen fertilizer use has increased dramatically in the past decades, and most crops only take a small proportion of the nitrogen applied, nitrogen leakage from the farming system has brought significant ecological consequences. CA practices an d tools have been developed to reduce nitrogen leakage, but the ado ption rates are strikingly low. For instance, agronomists have developed three corn growing CA practices, i.e., coving crops in the winter, forbidding applying fertilizer in the fall, and s ide - dressing fertilizer, to effectively reduce nitrogen usage by up to 50% corn growers, only 11%, 85%, and 32% of corn growers, respectively, have applied each of the three practices in at least one of the past three years. A pol icy to incentivize CA adoption is doption decisions in a paid - to - participate program through conducti ng a DCE. This thesis finds that payment incentives are critical and effective to encourage CA adoption, and the necessary amount of payment is associated with the difficulty level of ado pting each practice. In addition to the payment directly associated with compensating the adoption of change behaviors from their status quo. Removing the concer ns of committing to a CA program, if the concerns are clearly studi ed, can potentially increase the adoption rate without increasing the policy costs. In addition, payment has its sweet point as well as limitation in incentivizing adoption. As shown in Cha pter 3 ad opters who have experience with the CA practices or are more sensitive to payment; however, 142 s, who do not want to commit to a program regardless of how high th e payment is. Future researchers can work on solving an optimal policy adoption target to balance the adoption rate, policy cost, and social welfare to improve policy efficiency. A couple of factors associated with the policy design as well as farmers ar e found to affect the WTA. For instance, I found that emphasizing the environmental consequence s of not taking CA practices works better than emphasizing the benefits and contributions of e nrolling in a CA program . Giving farmers the opportunity to delay t he decision - making and collect more information can nudge CA adoption effectively compared to forcing farmers to make decisions immediately. A reference number of expected nitrogen savings provided in the policy can also effectively reduce the WTA. Besides , as people use behavioral strategies, i.e., reference dependence, to reduce the cognitive burden of making decisions, when it comes to proposing a policy, a nudge program can be carefully designed and combined with the target program to increase the targe - to - effic iency. Going beyond empirical analysis, this thesis investigates the fundamental assumptions of DCE modeling. The DCE modeling traditionally relies on the RUM framework, which assumes rational decision - making. This thesis investigates alternative decision strateg ies through developing a behavioral decision framework that nests rational decision - making as a special case. An adaptive decision pattern that gradually shifts the reference dependence points is justified in the DCE composed of repeated choice scenarios. This finding sets caveats for constructing DCE 143 modeling without controlling behavioral strategies and inspires future researchers to explore better DCE design and modeling strategies. - making analysis is built on an experimental a pproach, whether the decision - makers adopt similar decision strategies in the real - world choice scenario remains to be examined further. Meanwhile, even though the behavioral DCE modeling specification makes DCE possible to reveal the preference better, th e complexity of modeling introduces parameter identification issues given the limited sample size. This may restrict the application of the behavioral DCE framework concerning the sample size. Lastly, as nudging can effectively increase policy adoption due to the behavioral decision pattern, future research can be carried out on quantifying the impacts of nudging to generate more specific policy suggestions. 144 APPENDICES 145 APPENDIX A . G - RRM REDUCES TO RUM WHEN 17 Given a choice set of J alternatives, the probability of choosing an alternative generated by G - RRM ( with preference parameters and SQ specific constant term parameter will be equal to that generated by RUM with preference par ameters and SQ specific constant term parameter . Proof: first when , = = . So, the regret of alternative j can be written as: + So, the probability of alternative j is chosen will be: This probability is equal to that generated by RUM with preference parameters of . 17 Chorus, 2014 gives a formal proof of how G - RRM reduces to RUM when =0. 146 APPENDIX B. - RRM REDUCES TO RUM WHEN Given a choice set of J alternatives, the probabilit - RRM ( with preference parameters and SQ specific constant term parameter will be equal to that generated by RUM with preference parameters and SQ specifi c constant term parameter . Proof: first, when , = . So, the regret of alternative j can be written as: So, the probability of alternative j is chosen will be: This probability is equal to that generated by RUM with prefere nce parameters and SQ specific constant term parameter . 147 APPENDIX C. P - RRM REDUCES TO RUM WHEN Given a choice set of J alternatives, the probability of choosing an alternative generated by P - RRM ( with prefer ence parameters and SQ specific constant term parameter will be equal to that generated by RUM with preference parameters and SQ specific constant term parameter . Pr oof: first, when , = , and . So, the regret of alternative j can be written as: So, the probability of alternative j is chosen will be: