TO SPRAY OR NOT TO SPRAY: THE ECONOMICS OF WEED AND INSECT MANAGEMENT UNDER EVOLVING ECOLOGICAL CONDITIONS By Braeden Van Deynze 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 2020 ABSTRACT TO SPRAY OR NOT TO SPRAY: THE ECONOMICS OF WEED AND INSECT MANAGEMENT UNDER EVOLVING ECOLOGICAL CONDITIONS By Braeden Van Deynze The protection of crops from insect pests and weeds is fundamentally a problem of ecological management. Modern pesticides used to perform such management are essential to efficient production of corn and soybean, the two most widely gr own crops in the United States. What pesticides are sprayed where, when, and by whom is both shaped by and shapes ecological conditions. This dissertation consists of three essays on how American corn and soybean growers make insect and weed management dec isions, and the impacts of these decisions on the environment. The first essay measures the impact of glyphosate - resistant weeds on farmers' tillage practices using field - level data from across the United States, demonstrating how selection pressure within weed populations can limit the long - term benefits of pesticide technologies. Using a two - stage, panel data econometric model, we estimate that the spread of glyphosate - resistant weeds has led to reduction in the adoption of conservation tillage by soybean growers by as much as 8.5 percentage points. Nationally, we estimate that the reduction in conservation tillage adoption due to glyphosate - resistant weeds has increased soil erosion into water ways by over 65 million metric tons and carbon emissions due t o fuel consumption by 226,000 metric tons. The second essay measures the impact of farmers' pesticide use on butterfly abundance. By examining a full suite of pesticides in a single model , we account for substitution effects between products . We find neoni cotinoids, the most widely used class of insecticides, have a detrimental impact on butterfly populations , both in aggregate and for prominent species such as Monarchs . Overall, our results show that changes in pesticide use between 1998 and 2014 accounted for a 9% decrease in total butterfly abundance. Finally, the third essay examines farmers' decisions to custom hire to spray insecticides rather than performing such field tasks on their own. Using a pilot choice experiment, we demonstrate how the value farmers place on timeliness when custom hiring varies according to farmer characteristics. We find risk - averse farmers are more sensitive to potential delays, while those with more developed social networks are less sensitive . iv ACKNOWLEDGEMENT S Completing this dissertation would not have been possible without the support of too many people to name here. To my advisor Scott Swinton, thank you for being an incredible mentor and friend. Through tough projects and cross - continental meetings, your constant support over the years made this dissertation possible. Your willingness to share your wealth of experience, your keen editorial eye, and at least an hour a week even if it was just to chat, have made me a better researcher and a better person. Yo u have been an incredible role model and friend. Thank you. To my committee members Frank Lupi, David Hennessy, and Trey Malone, thank you for providing advice and critical opportunities to grow as a researcher. To Leslie Reis, thank you for inviting me to work with an incredible dataset and your openness to work across fields with new ideas. To Rick Horan, thank you for being an incredible teacher and a kind friend. To the AFRE staff and administration, particularly Nancy Creed, Ashleigh Booth, and Bob Mye rs, thank you for being the backbone of the department and always supporting grad students as full members of the Michigan State academic community. And to Kellogg Biological Station Long Term Ecological Research Program (NSF DEB 1832042 ), thank you for fu nding social science at the intersection of ecology, including my assistantship. When I drove across the country to begin grad school, Michigan was a state full of strangers. Now I am leaving Michigan as a place full of friends. To my roommate Piotr Szczep anski , thank you for being my daily connection to the outside world. Your friendship and outside perspective always kept me grounded. To Jarrad Farris, Sam Padilla, April Athnos, Asa Wattan, Stephen Morgan, Sophia Tanner, and Mary Doidge, thank you for mak ing coming into v Cook and Olds Halls something to look forward too. And to Sylvia Morse for providing a bed, a warm meal, and a great conversation. To my family, especially my parents Allen and Tracy, thank you for the constant love and support. Most of all , my gratitude goes to Felicia. Even from thousands of miles away, you have been my constant companion through this journey. I love you. vi TABLE OF CONENTS LIST OF TABLES ................................ ................................ ................................ ....................... v i i i LIST OF FIGURES ................................ ................................ ................................ ....................... ix Introduction ................................ ................................ ................................ ................................ ...... 1 REFERENCES ................................ ................................ ................................ ................................ 5 CHAPTER 1. Are Glyphosate - Resistant Weeds a Threat to Conservation Agriculture? Evidence from Tillage Practices in Soybeans ................................ ................................ ................................ .. 8 Abstract ................................ ................................ ................................ ................................ 8 Introduction ................................ ................................ ................................ .......................... 9 Conceptual Model ................................ ................................ ................................ .............. 11 Comparative Statics of Herbicide Resistance ................................ ........................ 13 The Case of Glyphosate and Glyphosate - Resista nt Weeds ................................ ... 15 Empirical Model ................................ ................................ ................................ ................ 16 Data ................................ ................................ ................................ ................................ .... 21 Results and Discussion ................................ ................................ ................................ ...... 28 First - Stage Non - Glyphosate Herbicide Use Model s ................................ .............. 28 No - T ill and C onservation T illage M odels ................................ ............................. 30 Effects of GRWs on Tillage Decisions ................................ ................................ .. 33 Simulation of GRW Effects on Tillage Use ................................ .......................... 36 Environmental Damages Resulting from s to GRWs ................................ ................................ ................................ ................. 38 Conclusion ................................ ................................ ................................ ......................... 41 REFERENCES ................................ ................................ ................................ .............................. 44 CHAPTER 2. Measuring the Effects of Pesticide Technology Change on Midwestern Butterfly Populations ................................ ................................ ................................ ................................ ..... 49 Abstract ................................ ................................ ................................ .............................. 49 Introduction ................................ ................................ ................................ ........................ 50 Conceptual Framework ................................ ................................ ................................ ...... 55 Data ................................ ................................ ................................ ................................ .... 57 Butterfly Abundance Data ................................ ................................ ..................... 57 Pesticide Use Data ................................ ................................ ................................ . 60 Land Cover Data ................................ ................................ ................................ .... 63 Weather Data ................................ ................................ ................................ ......... 64 Data Analysis ................................ ................................ ................................ ..................... 65 Count Models ................................ ................................ ................................ ......... 65 Pesticide Effects by Group ................................ ................................ ..................... 67 Net Pesticide Effects ................................ ................................ .............................. 68 Results ................................ ................................ ................................ ................................ 69 Discussion and Conclus ion ................................ ................................ ................................ 78 vii APPENDICIES ................................ ................................ ................................ .................. 83 APPENDIX A: Monarch Results with Texas Weather Controls ........................... 84 APPENDIX B: Pollard Survey Methods ................................ ............................... 85 APPENDIX C: Supplemental Figures ................................ ................................ ... 86 REFERENCES ................................ ................................ ................................ .............................. 89 CHAPTER 3. The Value of Timeliness: How Soybean Farmers Choose to Custom Hire f or Pest Control ................................ ................................ ................................ ................................ .... 94 Abstract ................................ ................................ ................................ .............................. 94 Introduction ................................ ................................ ................................ ........................ 95 Conceptual Model ................................ ................................ ................................ .............. 98 Base Model ................................ ................................ ................................ ............ 99 Transaction and Timeliness Costs ................................ ................................ ....... 101 Risk Aversion ................................ ................................ ................................ ....... 104 Conjectures and Social Capital ................................ ................................ ............ 105 Choice Experiment and Survey ................................ ................................ ....................... 107 Experimental Design ................................ ................................ ............................ 107 Survey Deployment ................................ ................................ ............................. 111 Additional Survey Data ................................ ................................ ........................ 111 Empirical Analysis ................................ ................................ ................................ ........... 111 M1: Base Model Candidates ................................ ................................ ................ 112 M2: Preference Heterogeneity by Farmer Characteristics ................................ ... 116 Choices by Sprayer Ownership ................................ ................................ ............ 117 Results ................................ ................................ ................................ .............................. 118 Summary Statistics ................................ ................................ ............................... 118 Model Selection and M1 Results ................................ ................................ ......... 12 1 M2 - DU Results with Farmer Characteristics ................................ ....................... 12 3 Choices by Sprayer Ownership ................................ ................................ ............ 124 Willingness - to - Pay for Timely Spraying ................................ ............................. 125 Discussion ................................ ................................ ................................ ........................ 12 7 Conclusion ................................ ................................ ................................ ....................... 132 APPENDICES ................................ ................................ ................................ ................. 135 APPENDIX A: Online Survey Instrument ................................ .......................... 136 APPENDIX B: Survey D eployment and Sample Representativeness ................ 149 REF E RENCES ................................ ................................ ................................ ............................ 152 viii LIST OF TABLES Table 1.1. Descriptions of Variables Included in Empirical Model. ................................ ............. 27 Table 1.2. Results from First - Stage, Non - Glyphosate Herbicide Use Models. ............................. 29 Table 1.3. Results from Second - Stage, Tillage Decision Models. ................................ ................ 31 Table 1.4. GRW Coefficients Under Alternative Specifications. ................................ .................. 35 Table 1.5. Estimated Social and Environmental Damages Resulting from Increased Use of Intensive Tillage in Response to GRWs. ................................ ................................ ....................... 40 Table 2.1. Butterf ly Species - of - Interest and Selected Traits. ................................ ........................ 60 Table 2.2. Butterfly Species - of - Interest Mean Counts per Survey - Hour. ................................ ..... 60 Table 2.3. Poisson Models of Butterfly Abundance. ................................ ................................ ..... 70 Table 2A.1 Poisson Models of Monarch Abundance with Spring Texas Weather Controls. ....... 84 Table 3.1. Levels and Descriptions of Choice Experiment Attributes. ................................ ....... 109 Table 3.2. Summary of Categorical Survey Variables. ................................ ............................... 119 Table 3. 3 . Summary of Numeric Survey Variables. ................................ ................................ .... 120 Table 3. 4 . Response Shares to the Choice Experiment. ................................ .............................. 121 Table 3. 5 . Results of M1 Conditional Logit Models. ................................ ................................ .. 122 Table 3. 6 . Results of M2 - DU, the Preferred Conditional Logit Model with Interactions for Farmer Characteristics. ................................ ................................ ................................ ................ 123 Table 3. 7 . Hypotheses, Relevant Parameters, and Determination of Support. ............................ 128 ix LIST OF FIGURES Figure 1.1. Percentage of Fields in Sample Under No - Till and Conservation Tillage Over Time. ................................ ................................ ................................ ................................ .............. 22 Figure 1.2. Percentage of Fields in Sample Treated with Glyphosate and Non - Glyphosate Herbicides Over Time. ................................ ................................ ................................ ................... 23 Figure 1.3. Price Indices for Glyphosate and Non - Glyphosate Herbicides Over Time. ............... 24 Figure 1.4. Number of Weed Species Resistant to Glyphosate (GRWs) by State. ........................ 26 Figure 1. 5. Predicted Adoption of No - Till and Conservation Tillage by the Number of Glyphosate Resist ant Weeds. ................................ ................................ ................................ ......... 34 Figure 1.6. Increases in Percentage of Soybean Acres Under Conventional Tillage Attributed to GRWs. ................................ ................................ ................................ ...................... 37 Figure 2.1. Counties M onitored in North American Butterfly Monitoring Network. ................... 58 Figure 2.2. Pesticide U se in S ampled C rop - R eporting D istricts. ................................ .................. 62 Figure 2.3. Pesticide Effects by Species and Cropland. ................................ ................................ 73 Figure 2.4. Net Pesticide Effect by Species. ................................ ................................ .................. 76 Figure 2 C .1. Land Cover Patterns by County (Cropland Data Layer) . ................................ ......... 86 Figure 2 C . 2 . Land Cover Patterns by County (NASS Acreage Estimates) . ................................ .. 87 Figure 2 C . 3 . Cropland Variable Geographic Distributio n. ................................ ............................ 88 Figure 3.1. Custom Hiring Trends in Three Soybean - Growing States. ................................ ......... 96 Figure 3.2. Example Choice Scenario. ................................ ................................ ......................... 109 Figure 3.3. Willingness - to - Pay (WTP) for Reductions i n the Probability of Delay. ................... 12 6 Figure 3B.1. Screenshots of Online Survey Instrument. ................................ ............................. 137 1 Introduction Pesticides use is almost ubiquitous in American row crop agriculture. Since the 1980s, over 90% of corn and soybean acres in the United States have been sprayed with one or more pesticides (Fern andez - Cornejo et al., 2014) . Herbicides and insecticides, which provide protection from weed competition and insect pest damage respectively, represent the two most widely applied classes of pesticides in corn and soybeans (Fernandez - Cornejo et al., 2014) . This dissertation consists of three essays on the economics of the use of these pesticides in these two crops, the most widely grown crops in the United States. Farmers spray pesticides because they provide more value over mechanical control, cultural control, and no protection alternatives. Pesticides prevent crop loss. Without control, corn yield would be potentially reduced 50% in the United States and Canada, a production value of over $26 billion annually (Soltani et al., 2016) . Global wheat yields in 2001 - 2003 would been potentially 30% lower without crop protection, 50% lower in corn, and 45% lower in soybeans (Oerke, 2006) . And among pest control options, pesticides protect crops more reliably and at lower financial cost to farmers than other pest control alternatives (Cooper & Dobson, 2007; Swinton & Van Deynze, 2017) . While pesticides provide undeniable benefits to farmers, their use can also impose costs on the environment and to human health. The negative impacts of pesticides were brought to the Silent Spring in 1962, soon after their use became widespread (Carson, 1962) . Certain pesticides have been found to alter the behavior, metabolism, and development of wildlife in detrimental ways, which in turn has led to population declines (Köhler & Triebskorn, 2013) . Pesticides can also be hazardous to 2 human health, with effects ranging from mild and short - term (e.g. headaches, dizziness) and to long - term and debilitating (e.g. asthma, cancer) (Kim et al., 2017) . Economists have long had an interest in pe sticide use as a study system. Economists have contributed to important management tools, including the concept of economic density thresholds for pests after which spraying becomes profitable ( e.g. Auld & Tisdell, 1987; Hueth & Regev, 1974; Marra & Carlson, 1983) . Other avenues of research have examined the roles of uncertainty and farmer attitudes towards risk in driving pesticide decisions (Horowitz & Lichtenberg, 1994; Pannell, 1991) and the impacts of emerging crop protection technologies on the use of alternatives ( e.g. Perry, Ciliberto, et al., 2016; Perry, Moschini, et al., 2016) . This dissertation contains three essays that contrib ute to the literature on the economics of pesticides by examining pest control decisions in corn and soybean fields. Each of the essays relates to either how ecological factors affect or are affected by farmers decisions, and therefore also contribute to t he modeling of agriculture as a managed ecosystem (Swinton et al., 2007) . The results highlighted in these analyses will help policymakers, farmers, and agribusinesses make more well - informed decisions by better projecting the environmental effects of changes in environmental condit ions and technologies. The first essay is an econometric evaluation of the effects of glyphosate - resistant weeds on the adoption of conservation tillage in soybeans. The broad - spectrum herbicide glyphosate, commonly marketed as Roundup, quickly became the most widely applied soybean pesticide following the commercial introduction of varieties with genetically engineered tolerance to the chemical (Fernandez - Cornejo et al., 2014; Perry, Ciliberto, et al., 2016; Swinton & Van Deynze, 2017) . Glyphosate - tolerant soybean see d allowed farmers to more readily adopt conservation tillage practices, as the broad - spectrum weed control value of tillage diminished relative to 3 glyphosate (Perry, Moschini , et al., 2016) . Meanwhile, an overreliance on glyphosate quickly resulted in the evolution of resistance in targeted weed populations (Livingston et al., 2015) . Using panel data from thousands of U.S. soybean growers over 18 years, this essay shows that glyphosate - resistant we eds have diminished the conservation benefits of a glyphosate - based weed control system by reducing the adoption of conservation tillage in soybeans by as much as 8.5 percentage points in some states. The second essay evaluates data from over a decade of b utterfly population and pesticide use tracking surveys to measure the effects of pesticide use decisions on butterfly abundance. Previous studies have found negative impacts on butterfly abundance from specific pesticides such glyphosate (Saunders et al., 2018 ) or neonicotinoid insecticides (Forister et al., 2016; Gilburn et al., 2015) . This study is the first to empirically link butterfly abundance to regional - scale pesticide use measures that account for contemporaneous changes in pesticide technology adoption by farmers. The resulting analysis finds that changes in pesticide use between 1998 and 2014 resulted in a 9% decrease in total butterfly abundance, and a 30%, 46%, and 39% decrease in the abundances of monarchs, silver - spotted skippers, and cabbage whites respectively. The increasing use of neo nicotinoid seed coatings in corn and soybeans drives this result. for pest control. When farmers custom hire for pest control, they expose themselves to increased ri sk of yield loss due to late completion of field operations (referred to as timeliness costs) relative to when they choose to provide pest control on their own. The proposed model is rooted tracting out production activities or completing them on their own is driven by frictions in contracting that can prevent potentially mutually beneficial trades (Coase, 1937; Williamson, 1979) . The model suggests 4 farmer ch aracteristics and resources such as social capital, risk aversion, and equipment capital affect their sensitivity to timeliness losses from custom hiring when considering custom pest control options. The resulting implications are illustrated empirically w ith a pilot choice experiment. Tying these three essays together is a focus on socio - ecological feedbacks in agricultural systems. As farmers adopt new technologies to control weeds and pests, ecosystems respond. Weeds evolve resistance to a popular herbic ide and farmers respond by returning to old technologies. Farmers adopt new insecticidal seed treatments, and butterfly populations fall as new toxins are introduced to their habitats. And the feasibility of custom pest control depends on the potential for timeliness costs, which are derived from the speed at which pest populations can damage a crop. Using economic theory and econometric modelling, this dissertation highlights how socio - ecological linkages shape farmer incentives . T he findings presented the rein can help guide policy attempts to align these incentives with public objectives. 5 REFERENCES 6 REFERENCES Auld, B. A., & Tisdell, C. A. (1987). Economic thresholds and response to uncertainty in wee d control. Agricultural Systems , 25 (3), 219 227. https://doi.org/10.1016/0308 - 521X(87)90021 - 7 Carson, R. (1962). Silent Spring . Houghton Mifflin Harcourt. Coase, R. H. (1937). The Nature of the Firm. Economica , 4 (16), 386 405. https://doi.org/10.1111/j.1468 - 0335.1937.tb00002.x Cooper, J., & Dobson, H. (2007). The benefits of pesticides to mankind and the environment. Crop Protection , 26 (9), 1337 1348. https://doi.org/10.1016/j.cropro.2007.03.022 Fernandez - Cornejo, J., Nehring, R. F., Osteen, C., Wechsler, S., Martin, A., & Vialou, A. (2014). Pesticide Use in U.S. Agriculture: 21 Selected Crops, 1960 - 2008 (No. EIB - 124; Economic Information Bulletin). U.S . Department of Agriculture, Economic Research Service. Forister, M. L., Cousens, B., Harrison, J. G., Anderson, K., Thorne, J. H., Waetjen, D., Nice, C. C., Parsia, M. D., Hladik, M. L., Meese, R., Vliet, H. van, & Shapiro, A. M. (2016). Increasing neonic otinoid use and the declining butterfly fauna of lowland California. Biology Letters . https://doi.org/10.1098/rsbl.2016.0475 Gilburn, A. S., Bunnefeld, N., Wilson, J. M., Botham, M. S., Brereton, T. M., Fox, R., & Goulson, D. (2015). Are neonicotinoid inse cticides driving declines of widespread butterflies? PeerJ , 3 , e1402. https://doi.org/10.7717/peerj.1402 Horowitz, J. K., & Lichtenberg, E. (1994). Risk - Reducing and Risk - Increasing Effects of Pesticides. Journal of Agricultural Economics , 45 (1), 82 89. ht tps://doi.org/10.1111/j.1477 - 9552.1994.tb00379.x Hueth, D., & Regev, U. (1974). Optimal Agricultural Pest Management with Increasing Pest Resistance. American Journal of Agricultural Economics , 56 (3), 543 552. JSTOR. https://doi.org/10.2307/1238606 Kim, K. - H., Kabir, E., & Jahan, S. A. (2017). Exposure to pesticides and the associated human health effects. Science of The Total Environment , 575 , 525 535. https://doi.org/10.1016/j.scitotenv.2016.09.009 Köhler, H. - R., & Triebskorn, R. (2013). Wildlife Ecotoxic ology of Pesticides: Can We Track Effects to the Population Level and Beyond? Science , 341 (6147), 759 765. https://doi.org/10.1126/science.1237591 Livingston, M., Fernandez - Cornejo, J., Unger, J., Osteen, C., Schmimmelpfenig, D., Park, T., & Lambert, D. (2 015). The Economics of Glyphosate Resistance Management in Corn and Soybean Production . ERR - 184 , 52. 7 Marra, M. C., & Carlson, G. A. (1983). An Economic Threshold Model for Weeds in Soybeans (Glycine max). Weed Science , 31 (5), 604 609. JSTOR. Oerke, E. - C. ( 2006). Crop losses to pests. The Journal of Agricultural Science , 144 (1), 31 43. https://doi.org/10.1017/S0021859605005708 Pannell, D. J. (1991). Pests and pesticides, risk and risk aversion. Agricultural Economics , 5 (4), 361 383. https://doi.org/10.1016/0 169 - 5150(91)90028 - J Perry, E. D., Ciliberto, F., Hennessy, D. A., & Moschini, G. (2016). Genetically engineered crops and pesticide use in U.S. maize and soybeans. Science Advances , 2 (8), e1600850. https://doi.org/10.1126/sciadv.1600850 Perry, E. D., Mosch ini, G., & Hennessy, D. A. (2016). Testing for Complementarity: Glyphosate Tolerant Soybeans and Conservation Tillage. American Journal of Agricultural Economics , 98 (3), 765 784. https://doi.org/10.1093/ajae/aaw001 Saunders, S. P., Ries, L., Oberhauser, K. S., Thogmartin, W. E., & Zipkin, E. F. (2018). Local and cross - seasonal associations of climate and land use with abundance of monarch butterflies Danaus plexippus. Ecography , 41 (2), 278 290. https://doi.org/10.1111/ecog.02719 Soltani, N., Dille, J. A., B urke, I. C., Everman, W. J., VanGessel, M. J., Davis, V. M., & Sikkema, P. H. (2016). Potential Corn Yield Losses from Weeds in North America. Weed Technology , 30 (4), 979 984. https://doi.org/10.1614/WT - D - 16 - 00046.1 Swinton, S. M., Lupi, F., Robertson, G. P., & Hamilton, S. K. (2007). Ecosystem services and agriculture: Cultivating agricultural ecosystems for diverse benefits. Ecological Economics , 64 (2), 245 252. https://doi.org/10.1016/j.ecolecon.2007.09.020 Swinton, S. M., & Van Deynze, B. (2017). Hoes t o Herbicides: Economics of Evolving Weed Management in the United States. The European Journal of Development Research , 29 (3), 560 574. https://doi.org/10.1057/s41287 - 017 - 0077 - 4 Williamson, O. E. (1979). Transaction - Cost Economics: The Governance of Contra ctual Relations. The Journal of Law & Economics , 22 (2), 233 261. 8 CHAPTER 1. Are Glyphosate - Resistant Weeds a Threat to Conservation Agriculture? Evidence from Tillage Practices in Soybeans Abstract The use of conservation tillage in American soybean production has become increasingly consumption. This trend has been reinforced by the availability of the general - p urpose herbicide glyphosate and glyphosate - resistant seed genetics since the mid - W eeds have since evolved to resist glyphosate, reducing its effectiveness. In this paper, we provide evidence that the spread of glyphosate - resistant weeds is responsi ble for significant reductions in the use of conservation tillage in soybean production. To capture the effects of glyphosate - resistant weeds on tillage adoption, we estimate a probit model of tillage choice, using a large panel of field - level soybean mana gement decisions from across the United States, spanning 1999 - 2016. We find that while the first two glyphosate - resistant weed species have little effect on tillage practices, by the time that eight glyphosate - resistant weed species are present, conservati on tillage use falls by 5.7 percentage points and no - tillage use falls by 10.0 percentage points. Using a simple benefit transfer model to illustrate how these results can be applied, we conservatively estimate that between 2005 and 2016 , responses to the spread of glyphosate - resistant weeds have caused water quality and climate damages valued at nearly $ 3 90 million. This total is likely to grow as glyphosate - resistance becomes more widespread and farmers continue to turn to tillage for sup plemental weed control. 9 Introduction Since the mid - the conventional production of soybeans and other U.S. field crops. Prior to the first commercial herbicides, farmers typicall y relied on mechanical weed control, characterized by multiple tillage passes to uproot established weeds and disrupt weed seedling emergence. While intensive tillage can provide effective weed control, it comes at a cost to the environment, leading to inc reased soil erosion and energy use, which can impair water quality and increase the carbon footprint of agricultural production (Uri et al., 1999). In this paper, we explore how the declining efficacy of glyphosate, the most widely used herbicide in Americ an soybean production, has led farmers to increase the use of tillage as a weed control tool. When first introduced, herbicides were rapidly adopted by American field crop farmers. Herbicides offered weed control as good or better than tillage at lower co st (Swinton and Van Deynze, 2017). The introduction of soybean varieties genetically engineered to resist glyphosate (and later other herbicides), has further shifted soybean weed control away from tillage (Perry, Moschini, and Hennessy, 2016; Fernandez - Co rnejo et al., 2012). Glyphosate is a broad - spectrum herbicide that could effectively control virtually all weeds when resistant seed varieties were first introduced. Glyphosate - tolerant crops, like Roundup Ready TM soybeans, allow farmers to spray the herbi cide throughout the growing season without damaging their crop. Farmers utilizing these technologies could rely exclusively on glyphosate for weed control, forgoing tillage passes and therefore providing cost savings to farmers and averting environmental d amages. As glyphosate use became more frequent in soybeans and other crops, weeds soon evolved to resist the chemical. In 2000, a population of horseweed growing in a soybean field in Delaware became the first identified case of glyphosate - resistance in we eds (VanGessel, 2001). 10 As of 2017, glyphosate - resistance ha d been identified in 17 weed species in the United States (Heap, 2017). The rise of glyphosate - resistant weeds (GRWs) has led to a growing literature on best practices to delay and manage the onset of herbicide - resistance in weeds ( Beckie, 2006 ; Evans et al., 2015; Bonny, 2016; Beckie and Harker, 2017). The increased use of tillage for weed control is frequently found amongst these recommendations. A smaller literature has focused on how farmers ha ve responded to the onset of GRWs. Livingston et al. (2015) reports the results of cross - sectional surveys of corn and soybean growers in 2010 and 2012, respectively. They find that farmers experiencing problems with GRWs frequently supplemented glyphosate - based weed control with non - glyphosate herbicides, increased their use of glyphosate, and increased the use of tillage. Wechsler et al. (2017), using farm - level cross - sectional data from corn - growing states in 2005 and 2010, find that low numbers of GRWs yields. Perry, Ciliberto, et al. (2016) observe a sharp increase in the use of non - glyphosate herbicides in corn and soybeans from 2007 to 2011 and speculate that this increase is due to GRWs. Most recently, Lambert et al. (2017) find that weed control costs increase by $34 - 55/acre following the emergence of GRWs in upland cotton fields as farmers adopt labor - intensive alternatives to glyphosate. In this paper, we contribute to t he literature on weed management in the face of herbicide resistance by providing the first estimate of the impact of GRWs on the adoption rates of conservation tillage practices in soybeans. We do so first by developing a conceptual model of a cost - minimi zing farmer who chooses among multiple herbicide and tillage options to meet predetermined weed control targets. This model indicates a non - linear response to herbicide - resistance: As more weed species develop herbicide resistance, farmers become increasin gly 11 likely to make major changes to their weed control practices. We test this model empirically with data on the field - level weed control choices of thousands of soybean farmers during 1999 - 2016. Our econometric results indicate that while low numbers of GRWs have little impact on tillage choices, by the time that eight GRWs are present, conservation tillage falls by 5.7 percentage points and no - till adoption falls by 10.0 percentage points. Extrapolating from literature estimates of soil erosion and carbo n emissions from tillage, and their environmental costs, we estimate that the shift towards more intensive tillage practices in response to GRWs has caused water quality and climate damage worth nearly $ 3 90 million. These damages accrued from 2005 - 2016 and have been most acute in the southern states where GRWs are most prevalent. The rest of this paper is structured as follows: We first present a conceptual model of a cost - minimizing farmer who seeks to control several weeds with many herbicide and tillage options. We then present our empirical strategy, followed by a discussion of the data. After presenting of our econometric results, we conduct a benefit - transfer simulation to illustrate potential environmental costs. We close with a discussion of the policy implications of our findings and directions for future research. Conceptual Model - stage cost - minimization problem, assuming a farmer has already determined optimal levels of weed control that are c onsistent with maximization of expected utility (Lichtenberg and Zilberman, 1986). Letting index different weed species, a farmer sets a weed control target for each of their soybean fields, 12 denoted in vector form as . This target re presents the minimum level of control acceptable for each weed in the field. 1 A farmer can achieve these weed control targets through a combination of tillage systems and chemical herbicides. A farmer selects a single tillage system from the choice set , where CT denotes conservation tillage and IT denotes conventional, intensive tillage. A farmer can select any combination of alternative herbicides to supplement weed control provided by his tillage system. Let denote the (non - negat ive) quantity of herbicide , so that a . 2 . . We assume that for all weeds is twice continuously differentiable , that larger quantities of herbicide increase control at a decreasing rate ( and , ), and that intensive tillage provides greater weed control than conservation tillage for any given choice of herbicides ( , ). Notice that when weed has adapted to resist herbicide , then for all quantities of that herbicide. We now turn to the costs of weed control. Denote the per unit costs of herbicide as and the costs of tillage system as . These costs include labor, fuel, and chemical expenses, as well as potential capi tal investments for new tillage equipment if adopting a system for the target. To do so, the farmer first determines the herbicide combination that minimizes tot al weed 1 Farmers and weed control experts typically utilize a maximum acceptable density of weeds in a field measured as individuals per area (e.g. weeds/m 2 efficient (Marra and Carls on, 1983; Swinton and King, 1994). In this model, we instead use a functionally identical concept of minimum acceptable control. 2 Note that farmers can combine different products via tank mixes. We envision as a farmer s herbicide choice set accounting for all feasible tank mixes and other combinations of retail products . 13 control costs for each of the two tillage systems subject to constraints (one for each weed species): ( 1. 1) The optimality conditions for this problem are: ( 1. 2) ( 1. 3) where are Lagrange multipliers for each constraint. Call the solution to the above minimization problem , and call the value function for this solution : ( 1. 4) A farmer then compares the solutions to these fi rst - stage cost - minimization problems for each tillage type and selects the least - cost option: ( 1. 5) - herbicide pairing, . Comparative Statics of Herbicide Resistance Now we use an exercise in comparative statics to consider how a decrease in the effectiveness of a given herbicide against a given target weed , represented by a decrease in , would affect . Let denote the optimal herbicide choices in a scenario with a different, separate kill function denoted where weed has evolved genetic resistance to herbicide . That is, we assume that , ceteris 14 paribus . Under w hat conditions does ? That is, under what conditions does the optimal herbicide regime for a given tillage system differ when one herbicide becomes less effective against a given target weed? If the weed control constraint for weed is bindin g under either kill function (hence ), then , as and therefore, by the continuity and strict monotonicity of , cannot satisfy equation (1 .2 ) if . But if the weed contro l constraint for weed is non - binding in both scenarios (hence in both pre - resistance and post - resistance weed control cost minimization problems), then , as would be multiplied by in equation (1 .2 ) and play no role in the solution. Thus, decreasing herbicide effectiveness from to - Further, this result imp lies that decreasing herbicide effectiveness weakly increases weed control costs for a given tillage choice, and therefore a single weed evolving resistance towards a single herbicide is likely not to influence tillage choices. As more weeds develop resist ance to a herbicide, changes in herbicide use, and hence tillage practices as well, become more likely as farmers seek alternative methods to reach their weed control targets . But because some weeds may and are in fact likely to be over - controlled (i.e. th e weed target constraint is non - binding) the response to herbicide resistance is inherently non - linear. If the herbicide costs associated with conservation tillage outweigh savings in tillage costs, then a farmer will switch to intensive tillage. 15 The Case of Glyphosate and Glyphosate - Resistant Weeds Glyphosate is a broad - spectrum herbicide which, in the absence of genetic resistance, is highly effective at controlling essentially all weeds. The introduction of glyphosate - resistant crop varieties allowed farmers to rely heavily (sometimes exclusively) on this specific herbicide for weed control in soybeans throughout the growing season at a relatively low cost. Glyphosate was rapidly adopted as the use of other herbicides declined (Livingston et al ., 2015). Swinton and Van Deynze (2017) attribute this trend to the cost - dominance of glyphosate - based weed control. When used in conjunction with glyphosate - resistant crops , pre - and post - emergent applications of glyphosate make tillage passes for weed co ntrol redundant, providing no additional weed control but incurring additional fuel, machinery, and labor costs for a farmer. In terms of our conceptual model, this implies that glyphosate has a non - zero marginal weed control effectiveness under conservati on tillage ( ) for all weeds, leading to over control ( i.e. ) for many weeds. When a weed develops resistance to glyphosate, the marginal weed control effectiveness of glyphosate falls . If this weed is not sufficiently controlled by other methods under lower glyphosate resistance (i.e. ), then either the use of another herbicide must increase or the farmer must switch to intensive tillage in order to continue to meet their weed control targets. For a single weed, this can be achieved by adopting a specialized herbicide. However, as more weeds evolve to resist glyphosate, we its advantage as a broad - spectrum weed control method over intensive tillage falls as additional herbicides are necessary to maintain weed control targets. Therefore, we expect compounding pressure to utilize intensive tillage over conservation tillage as glyphosate - resistance becomes more widespread. In other words, as the number of glyphosate - resistant weeds increases, we expect 16 both that intensive tillage becomes more common and that the rate at which it becomes more common in response to the occurrence of glyphosate - resistance to increase. Empirical Model To test the implications of the conceptual analysis presented above, we estimate a dynamic probit model with the tillage decision as the dependent variable. We include farm - level random effects to control for unobserved, time - invariant heterogeneity and a first - stage control function to account for potentially endogenous herbicide use. The unit of analysis, is the field - level ( ) tillage decision on each farm ( ) in a year ( ). With as an indicator for the use of conservation tillage, as the number GRWs, as an indicator for the use of non - glyphosate herbicides, as an indicator for the farms conservation tillage decision in the previous period , as an index for fuel prices , as a vector of farm - level conditioning variables, and as a time - invariant, normally - distributed, farm - level random effect to account for unobserved heterogeneity, the structural function we seek to estimate is the probability that conservation tillage is chosen: ( 1. 6) where is the standard normal cumulative distribution function. In this specification, we account for a non - linear response to additional GRWs suggested by our conceptual model by including the variable in quadratic form. As controls we include variables, , including measures of farm size (for scale economies in use of tillage equipment), soil erodibility (which affects tillage difficulty and soil water retention), and drought incidence (as tillage tends to reduce water retention). We include a time trend to capture the effects of 17 other unobserved time - varying factors that may have contributed to shifts in the use of conservation tillage over time. Before estimating this structural function via maximum likelihood, we must first address two issues: the initial conditions problem induced by including a lagged dependent variable and the potential endogeneity of non - glyphosate herbicide use. Adopting conservation tillage requires significant farmer investment in both learning new skills and acquiring n ew equipment (Krause and Black, 1995 ; Uri, 1999 ). Farmers who have made these investments in previous seasons face lower costs associated with conservation tillage. To account for this effect, fields, , assuming that previously used conservation tillage equipment remains available in following period. However, including the lagged dependent variable in a panel data model forces us to address the initial conditions problem (Arellano and Honore, 2001). Th is problem occurs when the modelled process is not observed from its beginning. Therefore, the initial condition, , is likely correlated with the farm - level random effect, . One approach to address this issue in non - linear models is to explici tly model the distribution of the random effect conditional on the initial condition and the other explanatory variables (Wooldridge, 2005). While this method can take several forms, we follow a specification for the random effect that has been shown to pr oduce unbiased estimates for parameters: ( 1. 7) where is a vector of all initial period explanatory variables and is a vector of explanatory variables averaged across all periods (Rabe - Hesketh and Skrondal, 2013 ). While Wooldridge (2005) suggests including all explanatory variables from all time periods in this auxiliary model, 18 doing so results in a model that is often computationally unwieldly due to the large number of in cidental parameters. Rabe - Hesketh and Skrondal (2013) show that the above constrained model performs similarly to the original Wooldridge solution. In this form, the random effect is constrained to depend on in the same fashion for . But be cause the presence of any non - zero parameters in the tillage model implies that is directly dependent on , we include separately from to account for this potential effect. This expression can be substituted directly into the struc tural equation , Eq uation (1. 6) , and estimation can proceed. The second issue relates to the use of non - glyphosate herbicides, . As herbicide use decisions are made simultaneously with tillage decisions, this variable is potentially endogenous. As our primary goal is to achieve consistent estimation of the parameters on the GRW terms of the tillage model, one could consider omitti ng this variable to avoid the issue of endogeneity entirely. However, the use of non - glyphosate herbicides is almost certainly correlated with GRWs, so its omission would induce omitted variable bias in the parameters of interest. In cases like this one, w here both the dependent variable and potentially endogenous variable are discrete, straight - forward approaches like two - stage least squares are unavailable (Wooldridge, 2015). Alternatives in this setting include bivariate probit models jointly estimated w - - stage model of the potentially endogenous variable are directly included in the structural model (Wooldridge, 2015). The bivariate probit approach is computationally complex e specially when random - effects inconsistently (Wooldridge, 2015). A third option, which we use here, is a control function approach for binary endogenous varia bles in binary dependent variable models known as two - stage residual inclusion 19 (Wooldridge, 2014; Terza et al., 2008). This method offers computational simplicity when compared to jointly - estimated, bivariate techniques. Prior to estimating the tillage mod el, we estimate a first - stage reduced - form model for the distribution of the endogenous variable, calculate generalized residuals of this model, and include these residuals, denoted as in the structural model as an explanatory variable. The idea i s that the residuals serve as a sufficient statistic for the degree of endogeneity in the explanatory variable. The unobserved variables that are the source of the endogeneity, for example unobserved latent weed pressure, are captured in the error term of the first - stage model. By including the residuals of the first - stage model in the second - stage, structural model, we essentially control for endogeneity by including an imperfect but sufficient aggregate measure of the unobserved variables which induce the problem in the first place. The reduced form model we estimate for the first - stage model of non - glyphosate herbicide use is: ( 1. 8) The price variable, , is the difference between the indexed price of glyphosate and an index of non - glyphosate herbicide prices, while is a vector of farm size indicator variables, omitting the soil a nd drought measures included in the tillage model. The farm - level random effect, , is assumed to follow a normal distribution with zero - mean and variance . To account for the joint determination between tillage and herbicide choices, we include la gged tillage choice . This first - stage model is estimated following standard maximum likelihood procedures for probit models with random effects. 20 To ensure identification of the second - stage tillage model , at least one exclusion restriction is required so that the first - stage residuals have independent variation that is not entirely determined by variables already in the model (Wooldridge, 2014). We argue that the indexed price differential between glyphosate and non - glyphosate prices, , satisfies the exclusion restriction. To satisfy the exclusion restriction, must satisfy three conditions: (1) it must not have a direct influence on the dependent variable in the structural model, ; (2) it must be u ncorrelated with omitted explanatory variables in the structural model; and (3) it must be strongly correlated with the potentially endogenous variable, (Terza et al., 2008). We argue these three conditions are satisfied. First, we assume th choices via their effects on the herbicides required for each alternative system, thereby satisfying condition (1). A similar assumption is maintained in Perry, Moschini, and Hennessy (2016), where the premium f or glyphosate - tolerant seed is assumed not to directly affect tillage decisions. The remaining two conditions are addressed in the following sections. With residuals from the first - stage model and the auxiliary model for in hand , the structural functi on we ultimately estimate is: ( 1. 9) This structural function can be estimated using standard maximum likelihood procedures for probit models with random effects. 3 3 We specifically use a Laplace approximation of the likelihood function. Estimation is performed using the R package lme4 (Bates et al., 2015) . 21 Data The core of our data are field - level survey data, representative at the Crop Reporting D istrict level, collected by the market research company Kynetec. These data contain observations on chemical and mechanical weed control practices of 22,151 farmers from 1999 through 2016 in 31 soybean - growing states 4 across the United States with more in tensive sampling in regions where soybeans are more widely grown, for a total of 93,345 field - level supplemented with payment recipient lists from the United States Depa rtment of Agriculture, agricultural publication subscription lists, and the membership lists of state and regional agricultural associations. Survey data were collected via computer assisted telephone interviews. Non - respondents were recontacted a minimum of eight times to reduce non - response error and up to 25 times in areas where response rates were low. Respondents were compensated monetarily upon completion of the interview. All interviews were recorded for verification purposes and data was crosschecke d against established ranges for prices, application rates, and consistency with other reported practices. Many farms provide data for multiple fields per year and responses in multiple years, giving the data an unbalanced panel structure necessary to esti mate the preceding empirical model. Tillage decisions, non - glyphosate herbicide use, herbicide prices, and farm size variables are all sourced from this dataset. The Kynetec survey data include three levels of tillage intensity: conventional, conservation, or no - till. Following Perry, Moschini, and Hennessy (2016), where a shorter 4 The states sampled are: Alabama, Arkansas, Delaware, Florida, Georgia, Illinois, Indiana, Iowa, Kansas, Kentuc ky, Louisiana, Maryland, Michigan, Minnesota, Mississippi, Missouri, Nebraska, New Jersey, New York, North Carolina, North Dakota, Ohio, Oklahoma, Pennsylvania, South Carolina, South Dakota, Tennessee, Texas, Virginia, West Virginia, and Wisconsin. 22 subset of these data are used, we define two binary tillage decision variables: a conservation tillage indicator equal to one when either conservation or no - till is us ed, and no - till indicator equal to one when no - till is used, grouping other conservation tillage practices along with conventional tillage. Because the effect of GRWs on no - till use is of particular interest, we estimate our empirical model twice, once wit h each of our two definitions of tillage practices as the dependent variable. The proportion of fields in the sample classified as no - till and conservation tillage is presented in Figure 1 .1 . The practice data also identify the herbicide products applied o ver each field in each year. We identify the active ingredients in each of these products and define a binary variable equal to one whenever the field is treated with a product containing a non - glyphosate active ingredient. Figure 1.1 . Percentage of Fields in Sample Under No - Till and Conservation Tillage Over Time. Conservation tillage includes no - till fields as well as other forms of reduced tillage. 23 The proportions of fields in the sample treated with glyphosate and non - glyphosate herbicides is presented in Figure 1. 2. Early in the sample period, the use glyphosate became i ncreasingly common, and the use of non - glyphosate products fell rapidly, likely due to the advent of glyphosate - tolerant soybean seed. Starting in 2006, this trend reversed, and non - glyphosate products were used more and more commonly. Glyphosate use reach ed near - saturation in the same year and continued to be used on over 90% of fields through 2016. We use the practice data to compute price indices for both glyphosate and non - glyphosate herbicides. For glyphosate prices, we calculate the mean price paid in dollars per pound each year. Because non - glyphosate herbicides represent a basket of several related products, we Figure 1.2. Percentage of Fie lds in Sample Treated with Glyphosate and Non - Glyphosate Herbicides Over Time. 24 construct a Laspeyr es index of prices and quantities for all non - glyphosate herbicide products used throughout the sample period, with the mean dollar per pound and volume shares from across the full sample used as the base. These indices are scaled so that both equal one in 1999, the first year of our sample. These input price indices enter the empirical model as relative prices and are therefore differenced as . These price indices are presented in Figure 1. 3. 2000 while non - glyphosate prices remained steady, so is negative in all years. During 2007 - 2009, gly phosate prices spiked relative to non - glyphosate prices. Because is Figure 1. 3. Price I ndices for G lyphosate and No n - G lyphosate H erbicides O ver T ime. Both prices normalized to 1 in 1999. 25 driven primarily by patent law and global demand trends, we argue that this variable is uncorrelated with omitted variables in the structural function and therefore satisfi es condition (2) of the exclusion restriction. We address condition (3) in the results section that follows. T he field - level practice dataset describes farm size as one of five categories: less than 100 acres, l00 - 249 acres, 250 - 499 acres, 500 - 999 acres, a nd 1,000 acres or more. These are included as a series of binary variables in the empirical model, with the less than 100 acres category excluded as the baseline. We supplement the field - level practice data with state - level data on the number of reported g lyphosate - resistant weed species at the beginning of the growing season, as reported by the International Survey of Herbicide Resistant Weeds (Heap, 2017). 5 The number of species resistant to glyphosate in each state in our sample in 2004, 2008, 2012, and 2016 is presented in Figure 1. 4. To the best of our knowledge, the ISHRW is the best available measure for this variable, providing consistent reporting on the development of herbicide resistance by mode of action across the full timeframe and the geograp hic region of our panel. As the primary contributors to the ISHRW data are university extension weed scientists, we assume that these counts represent the knowledge available to a typical farmer when making tillage decisions through an extension weed contr ol guide (e.g., Sprague and Burns, 2017). We rely on NASS annual price indices for diesel fuel (National Agricultural Statistics Service, 2018). As conservation tillage typically requires lighter field implements and therefore less fuel, we expect its use to be more frequent when fuel prices are higher (Lal, 2004). 5 Thes e data were provided to us through personal communication with Ian Heap, via email, as a custom report on herbicide - resistance in the United States generated from the ISHRW database. These data are consistently updated and can be viewed publicly on the ISH RW website ( http://www.weedscience.org/ ) . 26 Finally, studies have shown that conservation tillage systems are more likely to be adopted on highly - erodible lands (Uri, 1999; Soule et al., 2000). Past research has also found that the use of conservation tillage (but not no - till) is more likely in years following drought conditions (Ding et Figure 1.4. Number of W eed S pecies R esistant to G lyphosate (GRWs) by State. Prior to 2001, no species had been identified as glyphosate resistant at the start of the growing season. 27 Table 1.1. Descriptions of V ariables I ncluded in E mpirical M odel. Variable Description Geographic Scale Source Tillage Decision, No - Till Binary indicator of use of a no - till system Field GfK Tillage Decision, Cons. Till Binary indicator of use of a conservation tillage system (including no - till) Field GfK Non - Glyphosate Herbicide Use Binary indicator of use of a herbicide other than glyphosate Field GfK GRWs Count of glyphosate resistant weeds at the start of the year State ISHRW Glyphosate Price Average price of glyphosate in dollars per gallon, normalized to 1 in 1999 National GfK Non - Glyphosate Price Laspyres index of non - glyphosate herbicide prices, normalized to 1 in 1999 National GfK Fuel Price Index of diesel fuel prices, normalized to 1 in 1999 National NASS Palmer's Z - Index Index of anomalous moisture conditions, where negative values indicate drier conditions than usual, measured in September of the prior year Climate Division NOAA Soil Erodibility Index Proportion of farmland classified as highly erodible County NRCS Farm Size Acres of farmland operated by farm, categorized into five bins Farm GfK National Resources Inventory has classified as highly - erodible (National Resource - index as a measure of moisture conditions. This value is measured at the climate division level in the September of the prior year , where a more negative Z - index score indicate s drier conditions (National Environmental Satellite, Data, and Information Service, 2018) . In all, we bring together variables from several sources measured at disparate geographic scales. Brief descriptions of each of the variables ultimately included in the empirical model are presented in Table 1 .1 , along with the scale at which they are measured and their original source. 28 Results and Discussion In this section, we present the results of our empirical model. First, we offer a brief discussion of our fir st - stage models of non - glyphosate herbicide use before discussing the coefficients and goodness - of - fit for our second - stage models of tillage adoption. We present two measures of goodness - of - fit: the percentage of observations correctly predicted and pseud o - R 2 measures widely used when generalized linear mixed - effects models are reported (Nakagawa and Schielzeth, 2013). We then turn to the implications of our tillage decision models, examining predicted probabilities of conservation tillage and no - till adop tion at extant GRW species counts. Finally, we use our tillage decision model for conservation tillage adoption to explore a counter - factual scenario in which no weed species adapt to resist glyphosate to get a sense of the degree of environmental damages First - Stage Non - Glyphosate Herbicide Use Models The first - stage model of non - glyphosate herbicide use is estimated twice, once with past no - till use and again with past conservation tillage use as indepe ndent variables for the estimation of control functions for corresponding second - stage models. Results from each are presented in Table 1. 2. In both estimations, coefficients on both GRW terms indicate that glyphosate - resistant weed species are statistical ly significant and similar in scale. The negative coefficient on the linear term and positive coefficient on the quadratic term indicate that although the first GRW species to appear have relatively little impact on the use of non - glyphosate herbicides, th e probability of non - glyphosate herbicide use rises faster as GRW counts reach higher levels. 29 Table 1.2. Results from First - Stage, Non - Glyphosate Herbicide Use Models. Models estimated separately for use with no - till and conservation tillage second - stage models. P - values in bold are less than 0.05. Dep. Var.: Non - Glyphosate Herbicide Use No - Till Model Cons. Till Model Est. p Est. p (Intercept) - 0.782 <.001 - 0.781 <.001 GRWs - 0.043 <.001 - 0.042 <.001 GRWs (squared) 0.026 <.001 0.026 <.001 Glyphosate Price Difference 0.433 <.001 0.434 <.001 Past Tillage Decision 0.033 .017 0.018 .197 Year Trend 0.040 <.001 0.040 <.001 Size (100 - 249 Acres) 0.322 <.001 0.323 <.001 Size (250 - 499 Acres) 0.516 <.001 0.517 <.001 Size (500 - 999 Acres) 0.636 <.001 0.637 <.001 Size (1000 Acres or more) 0.737 <.001 0.738 <.001 Random Effects Farm - level Farm - level Unique Farms 22,151 22,151 Observations 93,345 93,345 Percent Correct (Dep. Var. = 1) 63.1% 62.9% Percent Correct (Dep. Var. = 0) 56.3% 56.4% Percent Correct 59.6% 59.6% Marginal R 2 0.119 0.119 Conditional R 2 0.579 0.579 The coefficient on the price differential between glyphosate and non - glyphosate herbicides is positive and statistically significant for both models. As expected, in years when glyphosate is expensive relative to alternatives, non - glyphosate herbicides are more likely to be used. The statistical significance of this coefficient has been proposed as a test of condition (3) of 30 the exclusion restriction (Wooldridge, 2014). As the coefficient is statistically significant at even very low alpha thresholds, we co nclude that this condition is met and therefore all three conditions for the exclusion restriction are met and the price differential serves as a valid candidate for exclusion in the second - stage models. No - Till and Conservation Tillage Models The results from the second - stage, tillage choice models are presented in Table 1. 3, estimated for both no - till and conservation tillage use as the dependent variable. Both models correctly predict the tillage decision for a field about four - fifths of the time. Further, the models correctly predict tillage decisions at roughly the same rate for fields regardless of the observed outcome. This balance is important for modelling counter - factual scenarios, because if the target, then prediction would be systemically biased Both models explain the majority of the variance in tillage adoption outcomes, as measured by the pseudo - R 2 metrics proposed for generalized linear mixed - effect models by Nakagawa and Schielzeth (2013). Marginal R 2 measures the variance explained by fixed factors alone (i.e. the observed independent variables), while conditional R 2 measures the variance explained by the full model, including random effects . These measures are preferred to - R 2 because (a) they can be interpreted on the same unit - scale as the usual R 2 commonly reported for ordinary least - square models, and (b) they separately identify th e contributions of fixed and random effects. For both models, around two thirds of the total explained variance is accounted for via the observed 31 Table 1.3. Results from Second - Stage, Tillage Decision Models. Models are estimated separately for no - till and conservation tillage. P - values in bold are less than 0.05. Dep. Var.: Tillage Decision No - Till Model Cons. Till Model Est. p Est. p (Intercept) - 1.373 < 0 .001 - 1.000 < 0 .001 GRWs 0.0 09 0 . 488 0.022 0 . 097 GRWs (S quared) - 0.0 10 < 0 .001 - 0.00 7 < 0 .001 Non - Glyphosate Use 0.3 45 0 .00 8 0 .352 0 .00 3 Non - Glyphosate Use (Residuals) - 0.1 37 0 .0 12 - 0. 143 0 .00 4 Fuel Price 0.0 74 < 0 .001 0.0 59 < 0 .001 Past Tillage Decision 0. 637 < 0 .001 0. 770 < 0 .001 Z - Index - 0.001 0 . 826 - 0.00 5 0.079 Soil Erodibility Index 0. 650 < 0 .001 0.4 45 < 0 .001 Year Trend 0.0 26 < 0 .001 0.0 17 <0 .001 Size (100 - 249 Acres) 0.00 9 0 .7 48 0.06 3 0 .0 10 Size (250 - 499 Acres) 0.00 4 0 . 891 0.060 0 .02 3 Size (500 - 999 Acres) - 0.0 27 0 . 370 0.0 56 0 .0 41 Size (1000 Acres or mo re) - 0. 112 < 0 .001 - 0.0 29 0 . 328 Initial Conditions Correction Yes Yes Random Effects Farm - level Farm - level Unique Farms /Observations 22,151 / 93,345 22,151 / 93,345 Percent Correct (Dep. Var. = 1) 72.3% 82.4% Percent Correct (Dep. Var. = 0) 81.2% 73.1% Percent Correct (All Obs.) 77.6% 79.4% Marginal R 2 0.467 0.413 Conditional R 2 0.707 0.625 32 heterogeneity (i.e. the fixed effects) and allowing for a random intercept for each farm to account for unobserved heterogeneity improves model fit substantially. The statistical significance of the residuals from the first - stage, non - glyphosate herbicide use models in both second - stage models allows us to rejec t the null hypothesis that non - glyphosate use is exogenous to tillage decisions (Wooldridge, 2014). The use of non - glyphosate herbicides is positively associated with the use of conservation tillage and no - till practices, as the coefficients on this term a re positive and statistically significant in both models. When farmers move away from intensive conventional tillage practices, they give up a weed control tool and must supplement lost weed control through other means. As glyphosate is used on nearly all fields in our sample regardless of tillage system, this means supplementing with non - glyphosate herbicides. F uel price has a statistically significant coefficient of the expected sign in both models. The positive coefficients on fuel price likely stem from the fact that conservation tillage systems require less fuel than conventional tillage and are therefore more likely to be s elected when fuel is costly (Lal, 2004; Perry, Moschini, and Hennessy, 2016). Previous use of conservation tillage has a statistically significant and positive effect. This conservation tillage today are more likely to use it in the future, perhaps due to increased familiarity with the system (Uri, 1999). T his pattern hold s when no - till is modelled separately from other conservation tillage systems . The remaining coefficient s follow their expected signs. Fields experiencing recent - index values) are more frequently under conservation tillage (though this coefficient is only statistically significant at the 10% level ) , but 33 not no - till. This pattern follows results found in the literature on tillage adoption (Ding et al., 2009). Fields in counties with more highly - erodible land are also more likely to be under conservation tillage systems. The positive time trend likely refle cts the effects of payments through federal conservation programs and state - level extension efforts to promote conservation tillage adoption, as well as increased familiarity with these practices over time. Medium sized farms are slightly more likely to ad opt conservation tillage than the largest (1,000 acres or more) and smallest farms (less than 100 acres), while the largest farms are slightly less likely to adopt no - till. Effects of GRWs on Tillage Decisions The primary focus of this paper is the effec t of glyphosate - tillage practices. In models for both conservation tillage and no - till, the coefficient on the linear term for GRWs is positive but statistically insignificant and the coefficient on the quadratic term is negativ e and statistically significant. This indicates that GRWs have a negative effect on conservation tillage use, and the emergence of additional GRWs has increasing impact. The predicted probabilities of adoption of conservation tillage and no - till for the ob served range of GRW counts, with other variables held at their means, are presented in Figure 1. 5. These curves show the negative and compounding effect of GRWs on the use of conservation tillage, consistent with the expectations of the conceptual model. T hrough the first two glyphosate resistant weed species, the predicted rate of no - till use remains statistically indistinguishable from the rate at zero GRWs (44% adoption). However, by the eighth GRW, the predicted rate of adoption falls by 10.0 percentage points, a 2 2.5 % reduction among no - till users. The impact of GRWs on conservation tillage is similar, though less severe. Through the first two 34 GRWs, conservation tillage is used at rates not statistically different from zero GRWs (66.9% adoption). But by the eighth reported GRW, conservation tillage rates fa ll by 5.7 percentage points, a n 8.6 % reduction among CT users generally. The magnitude of predicted reduction in conservation tillage and no - till use due to eight identified GRWs corresponds with that of the increase in use attributed to the introduction of glyphosate - resistant soybean seed s (Perry, Figure 1.5. Predicted A doption of N o - T ill and C onservation T illage by the N umber of G lyphosate R esistant W ee ds . The shaded region indicates a 95% confidence interval, computed via the delta method. 35 Table 1.4. GRW Coefficients Under Alternative Specifications. P - values are presented in parentheses. P - values in bold are less than 0.05. No - Till Cons. Till Alternative Specification Linear Quadratic Linear Quadratic Machinery Price Included 0.023 (0.141) - 0.010 (<0.001) 0.014 (0.343) - 0.006 ( <0.001 ) Soybean Price Included 0.025 (0.088) - 0.011 (<0.001) 0.013 (0.339) - 0.006 ( <0.001 ) Quadratic Term Omitted - 0.051 ( <0.001 ) - 0.030 ( <0.001 ) Moschini, and Hennessy, 2016). In effect, the advent of GRWs is undoing the stimulus to adopt conservation tillage that was prompted by the introduction of glyphosate - tolerant crop varieties. The negative effect o f GRWs on conservation tillage and no - till adoption is robust to alternative specifications. Table 1. 4 presents estimated coefficients for the linear and quadratic GRW terms for both no - till and conservation tillage models estimated with alternative covari ate structures. In our first alternative specification, we include a NASS machinery price index representing price changes over time for both tillage - related implements and other machinery. Including this covariate from our analysis does not affect the dir ection, significance, or relative magnitude of coefficients on either the linear or quadratic GRW terms. In our next specification, we include soybean prices, measured annually at the state - level in September of the previous year from NASS. Including soybe an prices does not meaningfully change our key result relative to the base model. Finally, excluding the quadratic GRW term results in a negative and statistically significant coefficient on the linear term, corroborating that GRWs have a negative effect o verall on no - till and conservation tillage adoption. To test whether including the quadratic term improves the model fit over the linear terms alone, we conduct a likelihood ratio test of the full model versus a specification where the quadratic terms for GRWs are excluded. For the no - till model the likelihood ratio is 54.886 (p - value for Chi - squared test < 0.001) and for the conservation tillage model the ratio is 23.033 (p - 36 value < 0.001). Both models exhibit significantly better fit when the quadratic ter ms are included, providing further support for the non - linear tillage response to GRWs suggested by the conceptual model. Simulation of GRW Effects on T illage U se To demonstrate the impact that and space, we compute the shares of acres under conservation tillage predicted by the model given realized GRW emergence patterns (denoted in which no w eed species evolve to resist glyphosate, all else equal (denoted for for all observations in a counterfactual dataset, leaving all other variables the same as observed. We - level tillage decisions in the counterfactual scenario, giving us for each field in the sample , the counterfactual predicted probability of conservation tillage use on field , operated by farmer , in year . We then simulate the same predicted probabilities of conservation tillage use under realized GRW emergence patterns (i.e. the original data), denoted for each field as . The shares of soybean acres in each year under conservation tillage in bo th scenarios ( and ) are calculated by summing the predicted probabilities weighted by the number of acres each field represents in the population of soybean acres in a given year, denoted : (1.10) As a display of the spatial variation in the effect of GRWs on tillage decisions over our sample period, the differences between the acre - shares under conservation tillage, , are 37 calculated separately for each state and presented in four maps for 2004, 2008, 2012, and 2016 in Figure 1. 6. On the majority of soybean acres, GRWs have had negligible impact on tillage practices, with increases in intensive tillage adoption of less than 5%. However, t he impact of GRWs on tillage decisions is particularly noticeable where GRWs are most prevalent: southern states such as Mississippi, Missouri, Arkansas, and Tennessee where glyphosate is commonly Fi gure 1.6. Increases in P ercentage of S oybean A cres U nder C onventional T illage A ttributed to GRW s. 38 used as the primary weed control tool on glyphosate - resistant cotton in addition to soybeans and corn. In Mississippi in 2016 for example, conservation tillage would be used on 8.5% more soybean acres ha d GRWs been absent. Environmental Damages Resulting from s to GRWs The use of conservation tillage systems is known to reduce soi l erosion and carbon emissions , two types of agricultural pollution that impair water quality and contribute to global climate change respectively (Uri et al., 1999). A n intuitive follow - up to the pr e ceding analysis of estimate the resulting environmental damages from increased tillage. We develop rough conserva tive estimates of the social costs of increased intensive tillage use on two environmental outcomes, soil erosion and carbon emissions from fuel, by drawing upon values from the literature and applying a simple benefit transfer model to monetize social cos ts (Wilson and Hoehn, 2006). Tillage practices have wide - ranging impacts on the environment (Uri et al., 1999), and a full accounting of these impacts is outside the scope of the present study. However, this exercise suggests that the spread of GRWs is a p roblem not just for farmers, but for society. Our general approach follows the methods presented in Perry, Moschini, and Hennessy (2016). To quantify the soil erosion impact of increased use of conventional tillage, we rely on median erosion rates for soil s under conventional and conservation tillage as reported in a review of 495 studies (Montgomery, 2007). For conventional tillage, the reported median erosion rate is 1.54 mm per acre - year. For conservation tillage, the median erosion rate is 0.08 mm per a cre - year. Assuming a soil density of 1,200kg/m 3 , this implies a 6.8 ton/acre - year reduction in soil 39 erosion in fields under conservation tillage when compared to a conventional tillage baseline (Montgomery, 2007). Conventional tillage leads to increases in carbon emissions over conservation tillage both through increased fuel consumption and by reducing the capacity of the soil to retain carbon. However, given that the potential carbon sequestration ability of soil is highly variable and dependent on the su stained practice of conservation tillage over time, we choose to focus only on carbon emissions from fuel consumption ( Uri et al., 1999) . Lal (2004) synthesizes the literature on fuel consumption required for various tillage operations, reporting the resul ts as mean kilograms CO 2 - equivalent emissions (CE) per hectare. We convert these means to metric tons CE/acre. The resulting mean increase in carbon emissions from fuel consumption when switching from conservation to conventional tillage is 0.0234 metric t ons CE/acre. To monetize the effects of these environmental impacts, we use prices previously used by federal policymakers for benefit - cost analysis. The National Resource Conservation Service estimates the costs of increased soil erosion at $4.93 per ton in water quality damage (National Resource Conservation Service, 2009). For carbon emissions, we rely on the global Social Cost of Carbon (SSC), as reported by the United States Government ( Interagency Working Group on S ocial Cost of Greenhouse Gases, 2016 ). This measure, widely used in policymaking prior to 2017, estimates the social costs of a metric ton of CO 2 released into the atmosphere for each year beginning in 2010. We rely on the reported average SCC estimate at a 3% discount rate, a conservative estimate. As the annual growth in this measure is almost exactly linear, we estimate the SCC for years prior to 2010 by regressing the SCC on a year trend ( R 2 = 0.987 ). These price s are adjusted using the Consumer Price Index to reflect the real value of d amages in each year , and range from $22.73 per ton CO 2 in 2000 to $37.51 in 2016 . 40 Table 1.5. Estimated Social and Environmental Damages Resulting from Increased Use of Intensive Tillage in Response to GRWs. Prior to 2007, GRWs had yet to reach impactful levels in any state. Social Damages Environmental Damages Year Current Value a (USD) Present Value b (USD 2016) Soil Erosion c (Metric Tons) Carbon Emissions d (Metric Tons CE) 2007 2,200,000 2,800,000 450,000 2,000 2008 5,200,000 6,500,000 1,020,000 4,000 2009 13,800,000 16,900,000 2,730,000 9,000 2010 19,200,000 23,000,000 3,730,000 13,000 2011 32,300,000 37,400,000 6,090,000 21,000 2012 41,600,000 46,800,000 7,650,000 26,000 2013 48,200,000 52,600,000 8,730,000 30,000 2014 61,600,000 65,400,000 11,000,000 38,000 2015 63,800,000 65,700,000 11,400,000 39,000 2016 72,100,000 72,100,000 12,770,000 44,000 Total 359,800,000 389,300,000 65,560,000 226,000 a Soil erosion priced at $4.93/ton in 2009 dollars, adjusted to current year prices with CPI (National Resource Conservation Service, 2009); carbon emissions priced following Social Cost of Carbon at 3% discount rate (Interagency Working Group on Social Cost of Greenhouse Gases, 2 016) . b Computed with a 3% annual discount rate . c Assuming a 6.8 ton/acre reduction in soil erosion from conservation tillage use (Montgomery, 2007) . d Accounts only for reduced fuel consumption; assuming a 0.0234 tons/acre reduction in emissions from conservation tillage use (Lal, 2004) . Finally, the conservation tillage acre - share differentials computed in the previous subsection are multiplied by the acres planted to soybean in each year (National Agricultural Statistics Service, 2018), providing an annual estimate of the number of acres that would be under conservation tillage in the absence of GRWs, but are instead under conventional practices. The environmental impact and social value coefficients are applied to these acres, providing an estimate for the value of damages to water quali ty and the climate. Annual social and environmental damages are presented in Table 1. 5. Social damages are presented as lost value in current year price levels and as 2016 present value. In total, we estimate that the net present value of water quality an d climate damage from 3 90 million, 41 accumulated between 2006 and 2016. This social cost has been growing, exceeding $70 million annually in the latest years of our panel. Water qual ity damage will be greatest in regions where GRWs are most prevalent, such as the southern region of the Mississippi Basin, while the climate damage will be realized globally. If weed species continue to evolve to resist glyphosate across the country, and farmers continue increase tillage to achieve similar levels of weed control, we expect the rate at which these damages grow to accelerate. Further, this analysis only considers tillage - related water quality damages and the climate effects of increased fuel consumption so i t is only a partial accounting of the full environmental damages induced by GRWs. For example, increased fuel consumption and soil disturbance under conventional tillage systems may have localized air quality impacts, while herbicide subst itutes for glyphosate may have additional water quality, air quality, and human health impacts. Conclusion Herbicide resistant weeds, GRWs in particular, have become a widespread issue for farmers across the United States. This paper provides new and robu st evidence that farmers respond to the decreasing effectiveness of glyphosate by increasing tillage intensity. We do so by observing the field - level weed control decisions of thousands of soybean farmers across the country during the period that GRWs firs t emerged and subsequently spread. We find evidence - linear pattern. Our empirical model further allows us to estimate the marginal, causal effects of additional GRWs on the use of alternative tillage sys tems. We use these estimates to provide a rough calculation of the scale of social damages that GRWs have caused by increasing tillage in soybean fields. 42 Our approach represents a novel direction in the herbicide resistance literature in two ways. First, w e focus on how farmers have changed their management behavior in response to herbicide resistance, while other economic studies focus on how resistance has affected costs, returns, or yields (Livingston et al., 2015; Wechsler et al., 2017; Lambert et al., 2017). Second, would not be possible without our focus on practices. In doing so, we provide evidence of an evolving technological landscape for farmers, where the efficacy of a ubiquitous weed control tool is waning and additional tools are needed for supplemental control. The environmental damages from these additional tools, partially accounted for in this paper, imply that weed susceptibility to herbicides is a resource that provides value to not only farmers, but the public as well. While this paper focuses on tillage practices, too little is known about how herbicide resistance affects the use of other weed control tools available to farmers. Future research s hould explore which non - glyphosate herbicides farmers are choosing to combat GRWs, which seed traits farmers select, and what those choices imply for environmental quality. Meanwhile, agrochemical companies have responded to GRWs by developing new crop see d genetics resistant to other herbicides (Mortensen et al., 2012; Green, 2014; Bonny, 2016). Farmers remain optimistic that agrochemical companies will develop new solutions that will maintain the simplicity of glyphosate - based weed management (Dentzman an d Jussaume, 2017). However, public weed scientists have questioned whether this path forward is sustainable, as weeds will continue to evolve resistance to more and more biochemical modes of action (Duke, 2011; Mortensen et al., 2012). Davis and Frisvold ( 2017) suggest that the current dominant weed 43 control regime, based on specific herbicides paired with resistant seed, may come to an end within the foreseeable future if action is not taken. Fortunately, numerous solutions have been proposed to all eviate the threat posed by GRWs and weed resistance to other herbicides. Mortensen et al. (2012) call for increased public investment in research and promotion of integrated weed management systems, which rely on a more diverse suite of weed management pra ctices in order to delay the onset of resistance of any specific method. A recent simulation study suggests that this approach can be profit - maximizing for farmers with longer time horizons (Frisvold et al., 2017). Davis and Frisvold (2017) suggest adaptin g current federal subsidies of crop insurance and other conservation programs such as the Environmental Quality Incentive Program to create incentives for the adoption of integrated weed management and other resistance management strategies. Ervin and Fris vold (2016), noting the common pool resource nature of herbicide resistance, envision community - based approaches for encouraging resistance management, modelled after drainage districts and insect eradication programs. Further research into policies to del ay the onset of resistance is needed. Such studies should consider not only the private benefits to farmers from the delayed onset of resistance, but also the public damages to the environment that could result if resistance management is ignored. 44 REFERENCES 45 REFERENCES Panel Data Models: Some Recent Developments Handbook of Econometrics . Vol. 5: 3229 - 3296. Bates , D. , M . Maechler, B . Bolker, and S . Walker (2015). Fitting Linear Mixed - Effects Models Using lme4. Journal of Statistical Software 67(1) : 1 - 48. - Weed Techonlogy 20(3): 793 - 814. Beckie, H.J., and K.N. Harker ( - Resistant Weed Management Pest Management Science 73(6): 1045 - 1052. Bonny, S. ( 2016 ) . Genetically Modified Herbicide - Tolerant Crops, Weeds, and Herbicides: Overview and Impact. Environmental Management 57(1): 31 - 4 8. Pest Management Science 73(11): 2209 - 2220. Management Approac Society & Natural Resources 30(11): 1311 - 1327. Agricultural Production Methods: Does Drought Increase Adoption of Conservation Journal o f Agricultural and Resource Economics 34(3): 395 - 411. Pest Management Science 68(4): 505 - 512. Ervin, D.E., and G.B. Frisvold ( 2016 ) . Are Community - Based Approaches to Manage Herbicide Resistance Wisdom or Folly? Choices 31(4). Evans, J.A., P.J. Tranel, A.G. Hager, B. Schutte, C. Wu, L.A. Chatham, and A.S. Davis (2015). Pest Management Scien ce 72(1): 74 - 80. Fernandez - Cornejo, J., C. Hallahan, R. Nehring, S. Wechsler, and A. Grube (2012). AgBioForum 15(3): 231 - 241. Frisvold, G.B Pest Management Science 73(6): 1110 - 1120. 46 Green, J.M. ( 2014 ) . Current S tate of H erbicides in H erbicide - R e sistant C rops. Pest Management Science 70(9):1351 - 1357. Heap, I . (2017). The International Survey of Herbicide Resistant Weeds . WeedScience. Online: http://www.weedscience.org/ . Accessed Nov. 3, 2017. Interagency Working Group on Social Cost of Greenhouse Gases (2016). Technical Support Document: Technical Update of the Social Cost of Carbon for Regulatory Analysis . United States Government. Online: https://www.epa.gov/sites/production/files/2016 - 12/documents/sc_co2_tsd_august_2016.pdf . Accessed May 9, 2018. Review Food Policy 32(1): 25 - 48. - till Technology in Review of Agricultural Economics 17(3): 299 - 310. Lal, R. (2004). E mission s f rom F a rm O Environment International 30 (7) : 981 - 990. Lambert, D.M., J.A. Larson, R.K. Roberts, B.C. English, X.V. Zhou, L.L. Falconer, R.J. Hogan , J.L. Johnson, and J.M. Reeves ( 2017 ) . : E stimating the C osts of M anaging H erbicid e R esistance as a F irst - O rder Markov P rocess and the C ase of U.S. U pland C otton Pr oducers. Agricultural Economics 48(3): 387 - 396. American Journal of Agricultural Economics 68(2): 261 - 273. Livingston, M., J. Fernandez - Cornejo, J. Unger, C. Osteen, D. Schimmelpfennig, T. Park , an d D. Lambert (2015) . The Economics of Glyphosate Resistance Management in Corn and Soybean Production, ERR - 184. United States Department of Agriculture, Econo mic Research Service. Marra, M . C., and G . A. Carlson (1983). An E conomic T hreshold M odel for W ee ds in S oybeans (Glycine max). Weed Science 31 ( 5 ) : 604 - 609. Proceedings of the National Academy of Sciences 104(33): 13268 - 13272. Mortensen, D.A., J.F. Egan, B.D. Maxwell, M.R. Ryan, BioScience 62(1): 75 - 84. National Agricultural Statistics Service (2018). Quick Stats . United States Department of Agriculture. Online: https://quickstats.nass.usda.gov/ . Accessed Mar. 17, 2018. National Resource Conservation Service (2018). Web Soil Survey . United States Department of Agric ulture. Online: https://websoilsurvey.nrcs.usda.gov/ . Accessed Apr. 23, 2018. 47 National Resource Conservation Service (2009). Interim Final Benefit - Cost Analysis for the Environmental Quality Incentives P rogram (EQIP) . United States Department of Agriculture. Online: https://www.nrcs.usda.gov/Internet/FSE_DOCUMENTS/nrcs143_007977.pdf . Accessed May 9, 2018. National Environ mental Satellite, Data, and Information Service (2018). National Climatic Data Center. National Oceanic and Atmospheric Administration. Online: https://www7.ncdc.noaa.gov/CDO/CDODivisio nalSelect.jsp . Accessed Apr. 30, 2018. Nakagawa , S., and H. Schielzeth ( 2013 2 from Generalized Linear Mixed - Methods in Ecology and Evolution 4(2): 133 - 142. Perry, E.D., F. Ciliberto, D.A. Hen Science Advances 2(8). Tolerant Soybeans and Conservatio American Journal of Agricultural Economics 98(3) : 765 - 784 . Rabe - Hesketh, S. , B iased V S imple S olution to the I nitial C onditions P Economics Letters 120 (2): 346 - 349 . Sprague, C.L. , and E.E. Burns (2017). 2018 Weed Control Guide for Field Crops . Michigan State University Extension Bulletin E0434. Online: https://www.canr.msu. edu/weeds/extension/weed - control - guide/2018WG - table - 12.pdf . Accessed May 9, 2018. Soule, M. J., A. Tegene, and K.D. Wiebe (2000). Land T enure and the A doption of C onservation P ractices. American J ournal of A gricultural Economics 82 (4): 993 - 1005. Swinton, S. M., & King, R. P. (1994). A B ioeconomic M odel for W eed M anagement in C orn and S oybean. Agricultural Systems , 44 (3), 313 - 335. European Journal of Development Research 29(3): 560 - 574. - Stage Residual Inclu sion Estimation: Journal of Health Economics 27(3): 531 - 543. Water, Air, and Soil Pollution 116 (3 - 4): 621 - 63 8. Environmental Geology 38(2): 111 - 125. 48 - Weed Science 49(6): 703 - 705. We Pest Management Science 74(5): 1142 - 1154. vironmental Goods and Services Using Benefit Transfer: The State - of - the - Ecological Economics 60(2): 335 - 342. Nonlinear Panel Data Models with Unobser Journal of Applied Econometrics 20 (1): 39 - 54 . - M aximum L ikelihood E stimation and T esting for N onlinear M odels with E ndogenous E xplanatory V Journal of Econometrics 182 (1): 226 - 234 . Wooldrid ge, J.M. (20 1 Control Functions in Applied Econometrics Journal of Human Resources 50(2): 420 - 445 . 49 CHAPTER 2. Measuring the Effects of Pesticide Technology Change on Midwestern Butterfly Populations Abstract Changing pesticide use is frequently implicated as a driver of declines in butterfly abundance. Existing empirical evidence of linkages between pesticide use and butterfly abunda nce is limited to studies that fail to account for the full suite of pesticides used by farmers, which fail to account for the effects of substitute pesticides. In this paper, we bring together data on the use of the six principal pesticide groups on corn and soybean fields and butterfly abundance data to create a unique county - level panel dataset spanning the 60 counties in the American Midwest over 17 years. We estimate count data models of total butterfly abundance and the abundance of three important sp ecies to measure the effects of each pesticide group. We find that neonicotinoids, a group of systemic insecticides applied to corn and soybean seeds before planting, have a strong negative association with total butterfly abundance and two of our three in dicator species. Further, we find a positive association between the planting Bacillus thuringiensis ( Bt ) traited corn seeds and butterfly abundance, though only in counties with large areas of cropland where interaction between butterflies and affected cr opland is likely. We overall butterfly abundance in the median county in our sample, driven by a shift towards neonicotinoid seed treatments since the mid - 2000s. 50 Introduction Butterfly populations are in decline in the United States and globally. Total butterfly abundance in Ohio declined 33% from 1996 to 2016, a 2% decline per year (Wep prich et al., 2019) . The decline observed in the American Midwest is consistent with a 35% global decline in the abundance of lepidoptera, the taxonomic order including both butterflies and moths, from 1970 to 2010 (Dirzo et al., 2014) . Declines in butterfly populations also coincide with a 45% decline in insect abundance across all taxa during the same period (Dirzo et al., 2014) . Despite ample evidence that butterfly populations are in decline, direct evidence pointing to sp ecific causes remains weak (Belsky & Joshi, 2018; Braak et al., 2018; Fox, 2013) . In this research, we evaluate associations between the decline of butterfly abundance and changes in the levels and types of agricultural pesticides applied and in the American Midwest. Pesticides are agricultural chemicals applied in order to protect crops and include both insecticides which target insect pest s and herbicides which target weeds. American farmers apply hundreds of millions of pounds of pesticides every year to protect the ir crops from pest damage and weed competition . Since the 1980s, pesticides have been sprayed on nearly every field of the most widely grown crops in the Midwest, corn and soybeans, suggesting widespread demand for at least some crop protection (Fernandez - Cornejo et al., 2014) . Historically farmers have always sought to control pest and weed population s, investing heavily in labor - intensive practices to provide even low levels of protection for their crops (Swinton & Van Deynze, 2017) . Because crop protection is both a critical and costly component of production, farmers change their crop production practices as new technologies become available seeking impr oved quality of control and/or lower costs. When synthetic pesticides were 51 first introduced, they were rapidly adopted by farmers as they provided several advantages over status quo crop protection practices (Fernandez - Cornejo, 2013; Swinton & Van Deynze, 2017) . Pesticide use often reduces profit risk by reducing the risk of catastrophic crop loss, making their use attra ctive to risk - averse farmers (Horowitz & Lichtenberg, 1994; Pannell, 1991) . Pesticide - based crop protection systems are also simple r than alternative practices, which often require a combination of several methods (Bastiaans et al., 2008; Castle et al., 2009; Lechenet et al., 2017) . Finally, pesticide - based systems are frequently cheaper and provide better protection than non - chemical alternatives . Just as the first pesticides reduced the risk, cost, and complexity of crop production, new pesticide technologies have improved on older systems. Which pesticides farmers use has changed drastically over time as farmers substitute among different products (Fernandez - Cornejo Swinton & Van Deynze, 2017) . The latest waves of technological change began in the mid - 1990s as genetically engineered seed varieties with herbicide and insect resistant traits were introduced, followed by the introduction of insecticidal seed coatings in the mid - 2000s (Douglas & Tooker, 2015; Perry et al., 2016) . These technologies allowed for novel pest and weed control syste ms that provided protection at historically similar levels at lower costs and reduced complexity. P esticide p roducts associated with these systems were rapidly adopted and the use of other products fell considerably as farmers replaced older technologies (Fernandez - Cornejo et al., 2014; Perry et al., 2016; Perry & Moschini, 2019) . We aim in this paper to distinguish between the effects of different pesticide groups on butterfly abundance while accounting for substitution patterns among pesticide technologies. 52 While pesticides provide clear benefits to farmers, their use is frequently suggested as a driver of declines in butterfly abundance (Agrawal & Inamine, 2018; Belsky & Joshi, 2018; Braak et al., 2018; Thomas, 2016) . In order for pollinator popula tions to be affected by pesticides, they must be exposed to pesticides, either directly or through interaction with environments damaged by pesticides (Sponsler et al., 2019) . Butterflies often rely on habitat in close proxim ity to cropland during different stages of their lifecycles, using vegetation along field edges and in hedgerows for food and shelter (Braak et al., 2018) . Butterflies in such areas can potentially be exposed to pesticides by coming into contact with treated crops or plants contaminated inadvertently through spray drift or translocatio n via water (Braak et al., 2018; Sponsler et al., 2019) . While pesticides can move through the environment, proximity of butterfly populations to treated cropland is likely to impact the effects of the pesticide use. Pesticides represent a diverse class of agrochemicals. There are mult iple mechanisms through which pesticides might affect exposed butterfly populations (Sponsler et al., 2019) . Insecticides can be applied via sprayer or via chemical seed coatings that are incorporated into the crops tissue up on germination. Sprayed insecticides, used specifically for their acute toxicity to insects, pose a self - evident threat when butterflies are directly exposed to spray or residues. Residues from seed - applied neonicotinoid insecticides persist in the soil an d water and can contaminate non - target plants (Douglas et al., 2015; Douglas & Tooker, 2015; Nuyttens et al., 2013) . Herbicides are only applied via sprayer and even though they are not known to cause acute harm to insects, they may have indirect effects on butterfly abundance by reducing suitable habitat and forage (Belsky & Joshi, 2018; Pleasants & Oberhauser, 2013) . Em pirical evidence directly linking spatial patterns of agricultural pesticide use to butterfly abundance is sparse, and past studies have examined the effects of only a single 53 pesticide active ingredient group at a time (Braak et al., 2018) . Saunders et al. ( 2018) find evidence of a negative association between glyphosate use and monarch abundance in areas of Illinois with heavily concentrated agriculture, though this pattern was only present from 1994 to 2003. Forister et al. (2016) find a negative association between neonicotinoid use and butterfly abundance in Northern California since neonicotinoid pesticides were first approved in 1995. Gilburn et al. ( 2015) find a similar negative association between neonicotinoid use and butterfly abundance in the United Kingdom. These paper s each examine the effects of a single pesticide active ingredient group (e.g. glyphosate herbicides, neonicotinoid insecticides) at a time. Therefore they cannot assess the relative impact of different pesticides. Changes in the use of specific pesticides are often associated with changes in the use of substitutes which may have their own negative (or positive) associations with butterfly abundance. As a result, the results of the aforementioned papers are of little use for assessing the net effect of cont emporaneous changes in the use of multiple pesticides. Expected positive effects of reductions in single pesticides may be overstated if the substitute pesticides are more harmful. A related branch of research assesses changes in the relative toxicity of p esticides used by farmers over time by applying active - ingredient specific measures of toxicity to insects and other taxa to pesticide application data, creating a kind of ambient toxicity measure. DiBartolomeis et al. ( 2019) find a 48 - fold increase in the average oral insect toxicity to bees of insecticides used in the United States from 1992 to 2014, driven largely by the large increase in the use of neonicotinoids. Perry and Moschini ( 2019) observe a decrease in ambient insecticide toxicity to bees between 1998 and 2006, attributable to reductions in sprayed insecticide use as farmers 54 adopted Bt seed varieties, though this decrease is offset by a neonicot inoid - fueled return to 1998 levels by 2012. While studies that examine multiple classes of pesticides simultaneously better account for substitution patterns between pesticides, no such studies directly link changes in pesticide use to butterfly abundance. Past studies rely on lab - based acute toxicity measures and cannot account for sub - lethal effects such as reduced fertility, increased predation risk due to behavioral change, or reduced habitat availability (DiBartolomeis et al., 2019) . Further, robust toxicity data is unavailable for butterfly species, so applying res ults from these studies to butterfly populations would require imputation from bee data (Braak et al., 2018) . In this research, we address shortfalls from each of these branches of research by incorporating pesticide use measures for a wide suite of pesticide classes directly into population models of butterfly abundance. By modelling butterfly populations explicitly as a function of pesticide application measures, we capture lethal, sub - lethal, and indirect effects of pesticides on abundance. Our models further allow for the assessment of the relative impact of different clas ses of pesticides on abundance, allowing for comparisons of impact size and direction between substitutes and across pesticide types. Finally, our models can be used to calculate the net impact of observed changes in pesticide use on butterfly abundance ov er time in way that accounts for patterns of substitution between pesticide products. The remainder of this paper is organized as follows. In the next section, we present our conceptual framework, which we use to examine potential linkages between pestici de use and butterfly abundance. We then describe the data used in this research, followed by a description of the statistical methods used to measure pesticide effects on butterfly abundance. Next we present 55 results, before providing a discussion of the fi ndings in context of previous studies of pesticide externalities and policy. Conceptual Framework pesticide application choices and a model of local butterfly population abundance. By examining the intersections of these two models, we identify two key implications for measuring pesticide effects on butterfly abundance which we will use to motivate the remainder of the paper. We model farmers as profit maximizers who choo se the optimal level of control for pest and weed damage , following a two - step procedure where they first determine the profit - maximizing level of damage control and then choose a combination of damage control technologies (i.e. pesticides) to achieve the optimal level of pest control at the least cost (Lichtenberg & Zilberman, 1986). The types of costs farmers consider when choosing among Deynze, 2017), risk (Pannel l, 1991), as well as complexity (Castle, Goodell, & Palumbo, 2009). Different combinations of pesticides have different cost levels for each of these types of costs, so farmers must make trade - offs when choosing between different alternatives. These trade - offs depend on characteristics at the farm (e.g. equipment availability, labor availability, acreage farmed, etc.) and regional (e.g. yield potential, pest and weed pressure, etc.) levels. As new technologies become available to a farmer and the suite of a vailable pesticides expands, farmers will respond and adopt these new technologies if they provide sufficient damage control at lower costs. 56 1997). These social costs i nclude acute and chronic human health effects (Brethour & Weersink, 2001), as well as damage to environmental quality through detrimental effects to non - target plant s , insect s , and wildlife. Not least of these are butterfly species. Butterflies are herbiv ores and rely on vegetation for habitat and forage. The degree to which butterfly habitat and forage range intersects with cropland and cropland - adjacent areas varies between butterfly species. Species characteristics such as migratory behavior and dietary diversity (i.e. specialist or generalist) may influence the likelihood that butterflies of that species interact with cropland. For pesticides to influence butterfly abundance, butterflies must be exposed to pesticides in the environment, either through c ontact during application, consumption of toxic compounds during foraging, or destruction of potential forage or habitat. The likelihood of such exposure depends on the distance between butterfly habitat and the cropland on which farmers apply pesticides. The effect of distance on pesticide effects will depend both on the tendency of specific butterfly species to forage or inhabit in cropland or cropland - adjacent areas and on the specific pesticides use d , which may have different potentials to reach butterf lies and remain in the environment (Sponsler et al., 2019). Our conceptual framework has two key implications for identifying the effects of changes in pesticide use on butterfly abundance. First, farmers substitute between pesticides as new technologies improve on older systems, providing less costly means of achi eving the same goal. As a result, increases in the use of one pesticide are frequently associated with decreases in the use of others that achieve the same goal. Such displacement occurs within pesticide classes (i.e. herbicides and insecticides) rather th an between. Displaced technologies may themselves influence butterfly abundance, so attempting to measure the effects of one pesticide without 57 accounting for changes in the use of others risks biasing estimates due to omitted variables. Further, measures o f the effects of one pesticide without similar measures for the effects of substitutes are of limited policy value, as the damage from one pesticide should be evaluated relative to what will replace it if the full effects of possible regulation are to be a ccounted for (Zilberman & Millock, 1997). Second, the effect of a pesticide on butterfly abundance is a function of the distance between butterfly habitat and where the pesticide is applied. This implies an interaction effect between geographic proximity t o cropland and cropland pesticide use mediating the effects of the pesticide on butterfly abundance. Additionally, the traits of specific butterfly species may affect the strength of this effect, as differences in foraging and migratory behavior may affect how frequently or closely butterflies interact with cropland. Data We bring together data from several sources to construct a unique panel dataset. The unit of observation is a county - year: the base geographic unit in the panel is a county and the base temporal unit is a year. The panel includes observations from 60 counties and 17 years (1998 - 2014). The annual number of monitored counties ranges from 15 to 37 based on data availability. Figure 2.1 shows the counties included in the panel and the number of years they contribute data. Butterfly Abundance Data For butterfly abundance, we use county - year aggregates of monitoring surveys conducted by four volunteer programs associated with the North American Butterfly Monitoring Network. 58 Both the Illinois Butterfly Monitoring Network an d the Ohio Lepidopterists provide data throughout the period of study, while the Iowa Butterfly Survey Network and Michigan Butterfly Network provide data beginning in 2006 and 2011, respectively. At approximately weekly intervals, citizen - scientist volun teers travelled along a fixed path, counting individuals by species Figure 2.1. Counties M onitored in North American Butterfly Monitoring Network. Crop - reporting district (CRD) boundaries indicated in bold. County boundaries indicated by dashed lines. Grey indicates no monitoring. 59 within a 5 - meter buffer of the path ( Pollard & Yates, 1994) . Subsets of these data have been previously analyzed to assess overall butterfly population trends in Ohio (Cayton et al., 2015; Wepprich et al., 2019) and Monarch population trends in Illinois (Saunders et al., 2016, 2018) . A brief summary of Pollard survey methods is provided in Appendix B. Counts by species and in total (across species) are summed across surveys conducted during June through August for each county - year. To better understand how pesticide effects may vary by species, we consider three specific butterfly species: monarchs ( Danaus plexippus ), silver - spotted skippers ( Epargyreus clarus ) , and cabbage white s ( Pieris rapae ). These species are selected for specific examination because of they are consistently observed throughout the study region and timeframe. These three species also exhibit distinct behavioral and life cycle traits that allow for examination of how such traits might influence vulnerability to specific pesticides. Pollinator traits identified as particularly relevant to pesticide exposure include breadth of diet (i.e. specialist vs. generalist) and range (i.e. migratory vs. resident) (Sponsler et al., 2019) . Monarchs are both migratory and host plant specialists. Silver - spotted skippers are residents to the region, host generalists, and are also known to be found feeding in s oybean fields as caterpillars. Cabbage whites are residents, though invasive, host generalists, and the most frequently identified species in the data. Table 2. 1 summarizes behavioral traits for each of these three species . Table 2.2 presents the mean count per hour of sampling during four periods to illustrate changes in populations over time . In accordance with recent studies of subsets of the same survey data, we observe declines in the abundance of each of the three species - of - interest and across all species over the period of our sample (Saunders et al., 2018; Wepprich et al., 2019) . 60 Pesticide Use Data Pesticide application data was collected via paid phone interviews by Kynetec USA, Inc. , a market research company. These data were collected via computer - assisted telephone interviews of soybean and corn growers. Lists of eligible growers were constructed from lists of growers who receive federal payments, membership lists of state and national growers associations, a nd subscription lists to agricultural periodicals. Sampling lists were constructed to ensure representativeness of applications at the level of the crop reporting district (CRD), USDA - designated groupings of counties in each state with similar geography, c limate, and cropping practices. Non - respondents were recontacted at least eight times to reduce non - response error. Respondents were asked to detail their field - level pesticide, tillage, and seed choices during the previous growing season. Respondents were compensated monetarily upon survey completion, and responses were crosschecked against realistic application rates and consistency with other reported practices. Table 2.1. Butterfly Species - of - Interest and Selected Traits. Species Breadth of Diet Mobility Other Notes Monarch Specialist Migratory Silver - Spotted Skipper Generalist Non - migratory Feeds on soybean Cabbage White Generalist Non - migratory Invasive, most common species Table 2.2. Butterfly Species - of - Interest Mean Counts per Survey - Hour. Species 1998 - 2002 2003 - 2007 2008 - 2012 2013 - 2014 Monarch 2.6 3.6 2.2 1.0 Silver - Spotted Skipper 1.3 1.5 1.3 0.6 Cabbage White 5.1 7.1 5.0 4.8 All Species 53.2 49.3 42.7 34.9 61 Pesticide use is measured in area - treatments for each pesticide group. Area - treatments are de fined as the average number of times a pesticide within a group is applied on a field within a defined region in a season (Kniss, 2017) . Area - treatment measures are preferred over volumetric measures because they account for dramatic differences in application rates and associated toxicity between different products (Kniss, 2017; Perry & Moschini, 2019) . More precisely, acre - treatments are calculated for each pesticide group as the sum of soybean and corn acres treated on respondent farms within each CRD divided by the total planted acres of soybean and corn in each CRD for each year. Farmers apply hundreds of distinct pesticide products to soybeans and corn. To simplify our analysis, we identify six group s of pesticides, divided into three classes , herbicides, sprayed insecticides, and systemic insecticides. These groups of pesticides together represent the majority and diversity of pesticide use on these crops. The first class is herbicides, represented b y glyphosate, and non - glyphosate herbicides. The second class is sprayed insecticides, which is comprised of pyrethroids and organophosphates . The final class is systemic insecticides, which includes neonicotinoids and Bt traited seed. Average area - treatme nts of each pesticide on soybean and corn fields for CRDs in our sample are presented in Figure 2.2 . Glyphosate and non - glyphosate herbicides represent farmer use of herbicides. Our glyphosate variable measures all applications of herbicides containing gly phosate as an active ingredient, while non - glyphosate herbicides measures all other herbicide applications. Glyphosate is a broad - spectrum herbicide for which soybean and corn seed with genetically engineered resistance has been available since the late 19 90s. Such seed increased the flexibility of glyphosate, allowing season - wide protection from any weed. Through the early period of our sample, glyphosate use rose while the use of non - glyphosate herbicides fell as farmers adopted 62 glyphosate - r esistant soybean and corn seed. Since 2008, non - glyphosate herbicide use has risen, likely in response to the spread of glyphosate - resistant weeds. Pyrethroids and organophosphates constitute the two sprayed insecticide groups as the two insecticide chemis tries most frequently applied via broadcast spray. These pesticides represent the two primary sprayed insecticide groups used to control insect pests in corn and soybean (Furlan & Kreutzweiser, 2015) . Organophosphate use has declined since 2005 as neonicotinoid seed treatments and Bt tr aited seed has spread, while pyrethroid use has remained steady on average. The final two pesticide groups, neonicotinoids and Bt traited seed, represent the systemic insecticides available to farmers. Systemic insecticides remain present in crop tissue fo r several Figure 2.2. Pesticide Use in Sampled Crop - Reporting Districts. Points represent mean acre - treatments for each pesticide group for crop - reporting districts represented in each year. Vertical lines represent one standard deviation above and below the mean. 63 weeks post application. When pests feed on crops protected with systemic insecticides, they consume compounds toxic to insects. Bt traited seed is genetically modified to produce insecticidal proteins. Such seed has been available only in corn si nce 1996 and is targeted to European corn borer and, since 2003, corn rootworm and earworm. We observe increasing adoption over the period of our study. Neonicotinoids are most frequently applied via seed treatments in the form of an insecticidal dust coat ing corn or soybean seed that is taken up by plant tissue as the crop develops. When neonicotinoids are present in plant tissue, they provide protection from a broad spectrum of insect pests. Their use has grown dramatically since their introduction in 200 4 in response to the appearance of soybean aphid and demand for additional systemic insecticides to supplement Bt traited corn, which target insects were quickly evolving to resist (Dougla s & Tooker, 2015) . Land Cover Data We measure cropland cover as the proportion of land within each county planted to soybeans and corn using the USDA Cropland Data Layer (CDL) ( USDA National Agricultur al Statistics Service Cropland Data Layer , 2019) . The CDL has used satellite images to classify land cover into distinct categories at 30m x 30m resolution consistently across the region of study since 2010 with over 90% accuracy for major crops (Lark et al., 2017) . Because of inconsistent data avail ability from the CDL prior to 2010, we use a time - invariant measure, averaging the proportion of land under soybeans or corn for each county between 2010 and 2014. There is little interannual variability in these measures over this period ( Figure 2C. 1 ), so we use the same value for each county across years, assuming the proportion remains roughly constant through the earliest years of the panel. A similar method is used in 64 Saunders et al. ( 2018) using Illinois land cover data from the Natio nal Land Cover Database over a similar time period. To verify that interannual variability at the county level is not an artifact of the CDL data generating process, we also examine the proportion of total land area planted under soybeans or corn using NAS S acreage estimates for the full sample period (Figure 2C.2), where we find a similar pattern of steady cropland cover was also present from 1998 to 2014 for the sampled counties. The cropland variable ranges from 0.00 0.85 with a mean of 0.44, represent ing a broad spectrum of agricultural intensity ( Figure 2C.3). Weather Data Local weather patterns have been previously shown to affect butterfly distributions, abundances, and the timing of life, though the strength of such associations varies by specie s and land cover (Cayton et al., 2015; Diamond et al., 2014; Saunders et al., 2016, 2018; Zipkin et al., 2012) . To control for potential weather effects on annual butterfly abundance, we generate county - level measures of precipitation and temperature that capture variation between years and within seasons. Dai ly weather data was gathered from NASA Daymet, a 1km x 1km spatial grid of daily weather conditions using data from a network of weather stations ( Thorton et al., 2018) . To aggregate to the county - level, we average daily Daymet data over a 0.2 x 0.2 decimal degree (approximately 10km x 10km) rectangle centered at the centroid of each county. Temperature is measured in growing degree days (GDDs), w hich measures the number of degrees Celsius within a range in which butterflies can develop (11.5°C to 33°C) (Cayton et al., 2015) . Precipitation is measured in millimeters. We partition each season into three intervals: early (March and April), mid (May and June), and late (July and August) season. Daily accumulation of precipitation and 65 GDD are summed over each interval and the resulting variables measure accumulated precipitation and GDDs for each county during each interval for each year. Monarchs annu al migration brings Midwestern populations through Texas each spring, where GDD and precipitation variation has been found to correlate with summer abundance in Illinois (Saunders et al., 2018) . To account for climate factors during Monarchs spring migratio n, we construct annual measures by calculating average accumulated GDD between March 22 and May 2 and precipitation during February, March, and April over Texas. Data Analysis Count Models We develop a Poisson regression model to estimate expected butte rfly counts, for each species - of - interest and in aggregate, in each county - year. For county located in CRD in year , we treat the observed butterfly count ( ) as a Poisson random variable with covariates on the log - link scale. (2.1) For covariates, we include the vector of weather variables ( ), cropland cover ( ), the vector of CRD - level pesticide area - treatments ( ) , and the interactions of pesticide area - treatments and cropland cover ( ). We also include vectors of fixed effects for county ( ) and year ( ) to control for unmeasured temporally invariant factors within each county and spatially invariant factors within each year (Wooldridge, 2010) . Finally, we control for changes in sampling by including the summed duration in minutes of all surveys in each 66 county - year ( ) as an offset. As a result, the exponentiated dependent variable can be interpreted as the rate of butterflies counted per minute. For the total abundance, silver - spotted skipper, and cabbage white models, we include only local weather variables. For the monarch models, we estimate an additional specification of the model including Texas spring weather variables as described in the data section to account We include the interaction between cropland cover and pesticide area - treatments to provide a more proximate measure of pesticide applications th an CRD - level pesticide area - treatments alone. Our measure of cropland cover is measured at the county level for the crops for which we have associated CRD - level pesticide application data (soybean and corn) and distinguishes between counties where soybean and corn dominate the landscape and counties where such land cover is less common. This serves as a proxy for the distance between the habitat and foraging range of the sampled butterfly populations and where pesticides are applied. Including this term ide ntifies how the magnitude, and direction, of the pesticide - abundance relationship varies between counties where butterflies are likely to come into direct contact with fields where pesticides are applied and counties where they are not. Previous studies ha ve made similar use of interaction terms in multiple regression models to identify complex relationships between potential drivers of butterfly abundance, including interactions between temperature and urbanization as well as between glyphosate adoption an d cropland cover (Diamond et al., 2014; Saunders et al., 2018) . By including pesticide area - treatments, cropland proportion, and interaction terms, we measure the impact of agriculture on butterfly abundance in three distinct but related ways. The linear pesticide area - treatment terms capture the effects of pesticide u se overall. The interaction 67 terms capture how the effects of pesticide use change as the likelihood that butterflies come across cropland varies. Finally, the linear cropland proportion term captures how cropland affects butterfly abundance through mechani sms other than pesticide use. 1 We obtain quasi - maximum likelihood coefficient estimates via R 3.5.1 (R Core Team, 2018) . We compute standard errors for coefficients via sandwich estimators that are robust to violations of the typical assumption used in Poisson reg ression that the conditional mean equals the conditional variance 2 (Wooldridge, 2010) - squared an d a likelihood ratio test against the null model with only the fixed effects . We use the robust standard errors to perform z - tests ( ) against the null hypotheses that each coefficient is equal to zero. Pesticide Effects by Group To explore the effect on butterfly abundance as implied by our models. We do so by plotting predicted expected counts per hour over the observed ranges of each pesticide variable. To examine the differen tial impacts of pesticides at different levels of county cropland cover, we plot predictions calculated with the share of county land area planted to corn and soybean set at 0.15 and 0.65, representing the first and third quartiles of the variable, with al l other covariates set at their means (Greene, 1 Note that by including the interaction term, we are including both the effects of pesticide applications per cropped acre at the CRD level ( ) and the effects of pesticide applications per total land acre at the county level ( ) . This is a result of simple unit analysis between the pesticide application measure where units are applications per CRD cropped acre and the cropland measure where units are county cropped acre per county land acre. 2 Fully robust standard errors are obtained from the square - root of the diagonal of the asymptotic variance matrix estimator for , given as , where is the expected value of the Hessian of the log likelihood for observation and is the transposed gradient (i.e. the score). 68 2010) . We include 90% confidence intervals for these predictions, computed using Delta - method standard and the asymptotic variance matrix for the estimated model para meters. Net Pesticide Effects To estimate for the cumulative impact of all changes in pesticides over the course the period of study, we compare predictions in expected values between two pesticide use scenarios for each county - year observation in our pa nel. We compute the difference between predicted population values computed using observed pesticide values ( and predictions computed using pesticide values observed for each county in 1998 ( ), the first year of the period of study. With as the exponentiated right - hand side of Equation (3.1), as a vector of all covariates other than , and as the vector of estimated coefficients, t hese predicted values are calculated as: ; and (3.2) . (3.3) This difference is divided by to compute proportional change: . (3.4) We estimate Delta - method standard errors for both and using the full y robust asymptotic variance matrix. Using these standard errors, we test the hypothesis against the null of no difference using a z - test ( ) to establish when net changes in pesticide use in county between year and 1998 h ave contributed to a statistically significant decline in butterfly abundance . 69 Results Poisson model results are presented in Table 2. 3 . All four models represent a good fit of - squared statistics exceeding 0.80. For all four models, the addition of explicit covariates for pesticide use, cropland cover, and weather provides a significant improvement over equivalent models with only county and year fixed effects, as evidenced by likelihood ratio test s. None of the coefficients on local seasonal precipitation and temperature are statistically significant. This result, suggesting that local weather conditions are not consistently linked to local annual abundance, is consistent with a previous analysis o f a subset of these butterfly abundance data focused on Illinois monarch populations (Saunders et al., 2018) . The cropland coefficient is statistically insignificant in all four models, suggesting that the row crop proportion of county land cover is not con sistently associated with butterfly abundance after accounting for variation in pesticide use and other county fixed effects. For monarch butterflies, spring GDD accumulation and precipitation in Texas is correlated with summer abundance at a statistically significant level, corroborating the findings in Sanders et al. (2018) (Table 2A.1). However, these variables only vary temporally and not over counties in each year. Because the effects of these variables are controlled for equivalently in the base model by including annual fixed effects and their inclusion does not affect the estimates for the pesticide and land cover coefficients, we use the base model to report results for Monarchs. Our Poisson model estimates include statistically significant coeffici ents for at least one pesticide group for all three species - of - interest and total abundance. In the total abundance model, coefficients for all pesticide groups except for glyphosate are statistically significant, 70 Table 2.3. Poisson Models of Butterfly Abundance. Standard errors robust to dispersion assumptions. All models include county and year fixed effects. Species Variable Estimate Std. Error z - score Pr(>|z|) Total Intercept - 1.61 0.71 - 2.25 0.025 Precipitation, early 0.000083 0.00043 0.19 0.847 Precipitation, mid - 0.00054 0.00043 - 1.27 0.204 Precipitation, late 0.000013 0.00030 0.04 0.966 GDD, early - 0.00062 0.00092 - 0.67 0.503 GDD, mid - 0.00007 0.00067 - 0.11 0.914 GDD, late 0.00040 0.00060 0.67 0.503 Cropland 1.60 1.15 1.39 0.163 Glyphosate 0.39 0.27 1.46 0.144 Non - glyphosate - 0.23 0.10 - 2.25 0.024 Pyrethroids 0.90 0.38 2.34 0.019 Organophosphate - 2.24 0.60 - 3.74 < 0.00 1 Bt - 1.47 0.54 - 2.70 0.007 Neonicotinoids - 1.28 0.48 - 2.68 0.007 Cropland X Glyphosate - 0.16 0.86 - 0.18 0.856 Cropland X Non - glyphosate 0.73 0.35 2.09 0.037 Cropland X Pyrethroids - 1.27 1.43 - 0.89 0.374 Cropland X Organophosphate 0.98 2.18 0.45 0.654 Cropland X Bt 3.18 1.55 2.06 0.040 Cropland X Neonicotinoids - 0.31 0.94 - 0.33 0.744 N 401 Pseudo R - squared 0.811 Likelihood ratio (vs. f.e.only) 139 , 793 (Pr(>Chi - squared) < 0.001) Monarch Intercept - 1.71 1.39 - 1.23 0.221 Precipitation, early - 0.000038 0.00088 - 0.04 0.965 Precipitation, mid 0.00083 0.00075 1.10 0.270 Precipitation, late 0.000494 0.00044 1.11 0.266 GDD, early 0.00017 0.00186 0.09 0.928 GDD, mid - 0.00081 0.00137 - 0.59 0.553 GDD, late 0.00074 0.00111 0.67 0.505 Cropland - 3.80 2.10 - 1.81 0.071 Glyphosate 0.85 0.52 1.63 0.102 Non - glyphosate - 0.32 0.20 - 1.64 0.100 Pyrethroids 0.72 1.14 0.63 0.530 Organophosphate - 1.58 1.86 - 0.85 0.397 Bt - 2.09 0.88 - 2.36 0.018 Neonicotinoids - 2.09 0.87 - 2.41 0.016 Cropland X Glyphosate - 1.49 1.37 - 1.09 0.278 Cropland X Non - glyphosate 0.68 0.43 1.59 0.113 Cropland X Pyrethroids - 3.86 2.30 - 1.68 0.093 Cropland X Organophosphate 2.23 3.04 0.73 0.463 Cropland X Bt 7.14 2.10 3.40 0.001 Cropland X Neonicotinoids - 2.83 1.14 - 2.49 0.013 N 396 Pseudo R - squared 0.829 Likelihood ratio (vs. f.e.only) 6,185 (Pr(>Chi - squared) < 0.001) 71 Table 2.3 (cont.). Species Variable Estimate Std. Error z value Pr(>|z|) Silver - Spotted Skipper Intercept - 4.79 1.56 - 3.08 0.002 Precipitation, early 0.000102 0.00118 0.09 0.931 Precipitation, mid - 0.00134 0.00123 - 1.09 0.275 Precipitation, late - 0.001323 0.00069 - 1.93 0.054 GDD, early 0.00230 0.00153 1.50 0.133 GDD, mid - 0.00200 0.00129 - 1.55 0.121 GDD, late 0.00048 0.00105 0.46 0.647 Cropland 0.59 4.23 0.14 0.889 Glyphosate 1.42 0.65 2.19 0.028 Non - glyphosate 0.95 0.28 3.45 0.001 Pyrethroids - 0.18 1.02 - 0.17 0.863 Organophosphate - 5.04 1.20 - 4.19 < 0.001 Bt - 0.98 1.48 - 0.66 0.509 Neonicotinoids - 2.82 1.45 - 1.94 0.053 Cropland X Glyphosate - 3.10 1.56 - 1.98 0.047 Cropland X Non - glyphosate - 1.91 0.82 - 2.34 0.019 Cropland X Pyrethroids - 3.22 3.28 - 0.98 0.325 Cropland X Organophosphate 7.04 3.59 1.96 0.050 Cropland X Bt 6.08 3.56 1.71 0.088 Cropland X Neonicotinoids - 2.71 1.95 - 1.39 0.165 N 388 Pseudo R - squared 0.865 Likelihood ratio (vs. f.e.only) 4,890 (Pr(>Chi - squared) < 0.001) Cabbage White Intercept - 5.73 1.76 - 3.26 0.001 Precipitation, early 0.000550 0.00099 0.55 0.580 Precipitation, mid - 0.00099 0.00079 - 1.25 0.210 Precipitation, late 0.000321 0.00060 0.53 0.593 GDD, early - 0.00070 0.00274 - 0.25 0.799 GDD, mid 0.00220 0.00143 1.54 0.123 GDD, late 0.00198 0.00195 1.02 0.310 Cropland 0.10 2.65 0.04 0.970 Glyphosate - 0.15 0.62 - 0.23 0.815 Non - glyphosate - 1.25 0.36 - 3.48 0.001 Pyrethroids 1.72 1.30 1.32 0.185 Organophosphate - 4.94 2.15 - 2.30 0.021 Bt - 2.11 1.26 - 1.67 0.094 Neonicotinoids - 4.87 1.28 - 3.80 < 0.001 Cropland X Glyphosate 0.41 1.95 0.21 0.835 Cropland X Non - glyphosate 2.03 0.56 3.61 < 0.001 Cropland X Pyrethroids - 6.99 3.67 - 1.91 0.057 Cropland X Organophosphate 8.76 4.26 2.06 0.040 Cropland X Bt 8.30 3.18 2.61 0.009 Cropland X Neonicotinoids 0.21 1.44 0.15 0.882 N 387 Pseudo R - squared 0.904 Likelihood ratio (vs. f.e.only) 21,384 (Pr(>Chi - squared) < 0.001) 72 while the cropland interaction coefficient is significant and positive for non - glyphosate herbicides and Bt traited seed. The positive interaction terms imply that the relationship between these pest control inputs and total butterfly abundance is more pos itive in counties where cropland in more common. For the monarch model, neonicotinoids and Bt traited seed coefficients are statistically significant, both for coefficients on area - treatments and on cropland interaction terms, but neither herbicide group h as a significant effect. By contrast, in the silver - spotted skipper model, both herbicide coefficients are statistically significant for both the area - treatment and cropland interaction terms. Among insecticides, only the organophosphate coefficient is sta tistically significant. Finally, in the cabbage white model, the non - glyphosate herbicides, organophosphate, and neonicotinoid coefficients are statistically significant, as well as the cropland interaction coefficients for non - glyphosate herbicides, orga nophosphate, and Bt seed. The effects of each pesticide group on total abundance and abundance of each species - of - interest are displayed visually in Figure 2.3. Many pesticide effects have notable differences between counties with abundant cropland and one s without (Figure 2.3). Non - glyphosate herbicide use is negatively related to total abundance in areas with low amounts of cropland but positively related in areas with high amounts of cropland. Pyrethroids have a positive association with total abundance at low levels of cropland but display no association at higher levels. Both organophosphates and neonicotinoids have strong negative associations with total abundance at both low and high levels of cropland. For monarchs, neonicotinoid pesticides and Bt tr aited seed, both systemic pesticides, have significant associations with abundance. Neonicotinoids have a strong negative association with monarch populations at both high and low levels of cropland. Bt seed adoption has a weak 73 Figure 2.3. Pesticide Effects by Species and Cropland. Expected counts are predicted using Poisson abundance models with methods described in the text. Color indicates either the primary or cropland interaction coefficient is statistically significant ( ) for the given pesticide for the Poisson abundance models. 74 75 negative association with monarch populations in counties with low amounts of cropland, but a strong positive association in counties with high amounts of cropla nd as a result of the cropland interaction coefficients. Silver - spotted skipper abundance is positively associated with herbicide use, both glyphosate and non - glyphosate products, at low levels of cropland, though slightly, and negatively associated with o rganophosphate use. We do not find these associations at higher levels of cropland, though silver - spotted skippers are rarely observed in high cropland counties. Cabbage white abundance has a strong negative association with neonicotinoid use at both high and low levels of cropland. Cabbage white abundance is also negatively associated with organophosphate, Bt seed, and non - glyphosate use at low levels of cropland. At high levels of cropland, Bt seed use is positively associated with cabbage white abundance , and the negative associations with non - glyphosate and organophosphate use is no longer detected. To examine the net effects of substitution between pesticide technologies over time, we compare the predicted values for every observation in the panel to th e predicted values for the same observations under a counterfactual pattern of pesticide use where the pesticide use variables are held constant at the levels used in that county in 1998 ( Figure 2. 4 ). The net effect of pesticides over time on the abundance of all three species - of - interest, and total abundance, has become increasingly negative over time. Net negative effects are first detected at a statistically significant level in 2003 for silver - spotted skippers, 2004 for monarchs and across all species, and in 2005 for cabbage whites. By 2014, the last year of our panel, over 75% of counties displayed statistically significantly negative net effects from pesticides on silver - spotted skipper, monarch, and total abundance, and 66% of counties display statis tically significant net effects on cabbage white abundance. 76 Figure 2.4. Net Pesticide Effect Since 1998 by Species. Red filled dots indicate a statistically significant difference between predicted values fitted with observed pesticide levels and pestic ide levels from 1998 ( ). 77 78 Changes in pesticide use between 1998 and 2014 account for a 9% decrease in total butterfly abundance in the median county. For monarchs, silver - spotted skippers, and cabbage whites, changes in pesticide use account for median decreases in abundance of 30%, 46%, and 39% respectively. Discussion and Conclusion Our results show that neonicotinoid insecticide use is negatively associated with butterfly abundance both across landscape co nfigurations and across species. Neonicotinoids are negatively associated with total abundance and with the abundance of both monarchs and cabbage whites. Silver - spotted skippers show a similar pattern, but the model driving this result is estimated with s tatistically insignificant parameters (though coefficient on neonicotinoid use is marginally so, with p = 0.053). For all four models, neonicotinoids display the largest magnitude effects at both high and low cropland levels. The finding of a negative association between butterfly abundance and neonicotinoid use is broadly consistent with previous studies. Past studies of regional - level butterfly abundance find a similar negative relationship between abundance and neonicotinoid use in California and the United Kingdom (Foris ter et al., 2016; Gilburn et al., 2015) . These studies examined neonicotinoids but no other pesticide groups; our findings suggest that this result persists even when other pesticide s are accounted for . The increase in neonicotinoid use from 2004 - 2014 c oincides with increasingly negative net pesticide effects and largely drives this result, mirroring results in past studies which rely on bee toxicology data (DiBartolomeis et al., 2019) . The negative effect of neonicotinoids stands in stark contrast to the associations between butterfly abundance and Bt seed use, the o ther systemic pesticide option available to farmers. In 79 all four models, Bt seed use is positively associated with butterfly abundance in those counties with high levels of cropland, though the parameters driving this result for silver - spotted skippers are noisily estimated. Bt adoption is linked to sharp reductions in the use of sprayed insecticides ini, 2019) . It is possible that the positive effect we observe is an indirect effect that results from the adoption of Bt decreasing the frequency of application of sprayed insecticides that may be more harmful to butterflies, though one would expect in cluding sprayed insecticide variables as covariates would control for this effect. Another possibility is that Bt adoption leads to changes in sprayed insecticide use beyond adjusting the frequency of spraying . One possible mechanism include s changes in th e timing of sprayed insecticide application. Changes in the timing of insecticide spraying may lead to applications during periods when butterflies are less vulnerable. Future research could address this hypothesis with data on the precise timing of in - sea son insecticide applications. Our results for sprayed insecticides are less conclusive. Organophosphate use is negatively associated with total butterfly abundance, though unassociated with monarch abundance. Pyrethroid use on cropland is positively asso ciated with total abundance in less agricultural landscapes, but this pattern does not hold in more agricultural counties or for any of the three individual species - of - interest. Our findings indicate that pyrethroids are less harmful than organophosphates in all contexts we examine, and so the decreasing use of organophosphates since 2004 has been a boon to butterfly populations, offsetting some of the negative pressure from increased neonicotinoid use. Our herbicide results suggest that herbicides have li ttle impact on the abundance of any of the species - of - interest and an inconsistent association with total abundance. Most importantly, our results find no impact of herbicide use on monarch abundance. Previous studies have pointed 80 to increased glyphosate u se as a driver of monarch decline. Pleasants and Oberhauser ( 2013) find large declines in both milkweed, the plant species used as a breeding host for monarchs and a weed targeted by corn and soybean farmers, a nd monarchs between 1999 and 2010. Saunders et al. (2018) find a negative relationship between county - level glyphosate purchases and site - level monarch abundance, particularly in areas where agriculture is most intensive , though only prior t o 2005. Our results, which account for changes in glyphosate use as well as contemporaneous changes in the use of other herbicide and pesticides, do not corroborate these past findings and show no link between monarch populations and the use of either glyp hosate or non - glyphosate herbicides. The direction and magnitude of the net pesticide effects we estimate are consistent with observed declines in the abundance of Midwestern butterfly populations reported in the literature. Total butterfly abundance in Oh io declined 33% from 1996 to 2016, and monarch and cabbage white abundance were declining during the same period (Wepprich et al., 2019) . Our findings suggest that changes in pes ticide use patterns, namely the widespread adoption of neonicotinoid insecticides, can account for at least part of these declines. The net pesticide effect on total butterfly abundance between 1998 and 2014 that we estimate accounts for about a third of m agnitude of total decline in abundance over the same period, leaving two - thirds of the decline unexplained. Understanding the full range of externalities associated with the full suite of pesticide technologies available to farmers is critical to understan ding tradeoffs associated with their use and regulation (Zilberman & Millock, 1997) . Our results suggest that the use of pesticide s , notably neonicotinoid s , creates a negative environmental externality by reducing butterfly abundance . By contrast, Bt traited seed adoption creates a positive externality through a positive 81 association with butterfly abundance. These findings improve the understanding of the full social inc entives. To prevent further declines in butterfly abundance, farmers may reduce pesticide use voluntarily. Recent studies in France and the United States suggest that pesticide use could be reduced in row crop production without negatively affecting profit s for the median farm in each region (Lechenet et al., 2017; Mourtzinis et al., 2019) . However, these studies are based on deterministic cost models that do not account for the reductions in risk and complexity that pesticides frequently provide over alternatives. In order to overcome these barriers to reducing pesticide use, farmers may require compensation in the form of a payment - for - environmental - services program. Studies have found that farmers will enter such programs at lower payment rates when farmers believe their participation meaningfully impacts environmental outcomes (Chèze et al., 2020; Ma et al., 2012) . The results of the present study can be used to improve farmer knowledge of the non - target effects of pesticide use. Alternatively, regulatory agencies may consider imposing restrictions on pesticide use. Pesticide externalities, both in terms of butterfly losses and other environmental and health outcomes, vary widely depending on which pesticides are used , where they are used, and ho w they are used, creating challenges for implementing efficient externality taxes (Zilb erman & Millock, 1997) . As a result, pesticide regulation typically takes the form of bans on specific technologies. There is recent precedent for regulating pesticides specifically to protect insect populations. In the European Union, neonicotinoid se ed treatments have been banned since 2013 due to concerns about toxicity to pollinator s (Auteri et al., 2017) . A similar ban in the United 82 States is predicted to lead to pest management substitutions that decrease pesticide toxicity to bees by 19% , albeit wi th offset t ing increases in toxicity to mammals, fish, and birds (Perry & Moschini, 2019) . Our findings here indicate that, like bees, butterflies would benefit from reductions in neonicotinoid use. However, given that farmers would find substitute pesticides, these gains must be balanced against non - butterfly related social costs, including damage to other species and threats to human health, associated with potential substitutes. 83 APPENDICIES 84 APPENDIX A: Monarch Results with Texas Weather Controls Table 2A.1. Poisson Models of Monarch Abundance with Spring Texas Weather Controls. Standard errors robust to dispersion assumptions. All models include county and year fixed effects. Species Variable Estimate Std. Error z - score Pr(>|z|) Monarch Intercept - 6.40 3.14 - 2.04 0.041 Precipitation, early - 0.000038 0.00088 - 0.04 0.965 Precipitation, mid 0.00083 0.00075 1.10 0.270 Precipitation, late 0.000494 0.00044 1.11 0.266 GDD, early 0.00017 0.00186 0.09 0.928 GDD, mid - 0.00081 0.00137 - 0.59 0.553 GDD, late 0.00074 0.00111 0.67 0.505 Precipitation, Texas Spring - 0.0221 0.00540 - 4.08 <0.001 GDD, Texas Spring 0.0249 0.00713 3.50 <0.001 Cropland - 3.80 2.10 - 1.81 0.071 Glyphosate 0.85 0.52 1.63 0.102 Non - glyphosate - 0.32 0.20 - 1.64 0.100 Pyrethroids 0.72 1.14 0.63 0.530 Organophosphate - 1.58 1.86 - 0.85 0.397 Bt - 2.09 0.88 - 2.36 0.018 Neonicotinoids - 2.09 0.87 - 2.41 0.016 Cropland X Glyphosate - 1.49 1.37 - 1.09 0.278 Cropland X Non - glyphosate 0.68 0.43 1.59 0.113 Cropland X Pyrethroids - 3.86 2.30 - 1.68 0.093 Cropland X Organophosphate 2.23 3.04 0.73 0.463 Cropland X Bt 7.14 2.10 3.40 0.001 Cropland X Neonicotinoids - 2.83 1.14 - 2.49 0.013 N 396 Pseudo R - squared 0.829 Likelihood ratio (vs. f.e.only) 6,185 (Pr(>Chi - squared) < 0.001) 85 APPENDIX B: Pollard Survey Methods Each butterfly monitoring network makes use of methods described in Pollard (1977). This method is briefly summarized in this appendix. Routes are designed by volunteers in coordination with network coordinators to (a) transect a variety of habitat types, (b) follow existing pathways so not to disturb habitat, (c) be easily located by other volunteers, and (d) take between 30 minutes to two hours to complete. For each route, a single volunteer walks at a of June, Jul y, and August, with additional runs during these months or others if possible. Runs are conducted between 10:00am and 3:30pm on days with (a) less than 50% cloud cover and (b) light to moderate winds. During the run, the volunteer records all individuals b y species sighted within roughly 20 feet to each side of the route. Volunteers are instructed to only identify species with certainty and not to guess. 86 APPENDIX C: Supplemental Figures Figure 2C.1. Land Cover Patterns by County (Cropland Data Layer). Each line represents the proportion of a county in the sample classified as corn or soybean over the period with consistent Cropland Data Layer availability. 87 Figure 2C.2. Land Cover Patterns by County (NASS Acreage Estimates). Each line represen ts the proportion of a county in the sample planted with corn or soybean over the study period. Sudden drops are the result of missing values for either corn or soybean acreage in a given year. 88 Figure 2C.3. Cropland Variable Geographic Distributio n. Crop Reporting District boundaries (CRD) indicated in bold. 89 REFERENCES 90 REFERENCES Science , 360 (6395), 1294 1296. https://doi.org/10.1126/science.aat5066 Auteri, D., Arena, M., Barmaz, S., Ippolito, A., Linguadoca, A., Molnar, T., Sharp, R., Szentes, C., Vagenende, B., & Verani, A. (2017). Neonicotinoids and bees: The case of the European regul atory risk assessment. Science of The Total Environment , 579 , 966 971. https://doi.org/10.1016/j.scitotenv.2016.10.158 Bastiaans, L., Paolini, R., & Baumann, D. T. (2008). Focus on ecological weed management: What is hindering adoption? Weed Research , 48 (6 ), 481 491. https://doi.org/10.1111/j.1365 - 3180.2008.00662.x Belsky, J., & Joshi, N. K. (2018). Assessing Role of Major Drivers in Recent Decline of Monarch Butterfly Population in North America. Frontiers in Environmental Science , 6 . https://doi.org/10.33 89/fenvs.2018.00086 Braak, N., Neve, R., Jones, A. K., Gibbs, M., & Breuker, C. J. (2018). The effects of insecticides on butterflies A review. Environmental Pollution , 242 , 507 518. https://doi.org/10.1016/j.envpol.2018.06.100 Castle, S. J., Goodell, P. B., & Palumbo, J. C. (2009). Implementing principles of the integrated control concept 50 years later Current challenges in IPM for arthropod pests. Pest Management Science , 65 (12), 1263 1264. https://doi.org/10.1002/ps.1862 Cayton, H. L., Haddad, N. M., Gross, K., Diamond, S. E., & Ries, L. (2015). Do growing degree days predict phenology across butterfly species? Ecology , 96 (6), 1473 1479. https://doi.org/10.1890/15 - 0131.1 reduce pesticide use: A choice experiment. Ecological Economics , 167 , 106349. https://doi.org/10.1016/j.ecolecon.2019.06.004 Diamond, S. E., Cayton, H., Wepprich, T., Jenkins, C. N., Dunn, R. R., Haddad, N. M., & Ries, L. (2014). Unexpected phenological re sponses of butterflies to the interaction of urbanization and geographic temperature. Ecology , 95 (9), 2613 2621. https://doi.org/10.1890/13 - 1848.1 DiBartolomeis, M., Kegley, S., Mineau, P., Radford, R., & Klein, K. (2019). An assessment of acute insecticid e toxicity loading (AITL) of chemical pesticides used on agricultural land in the United States. PLOS ONE , 14 (8), e0220029. https://doi.org/10.1371/journal.pone.0220029 91 Dirzo, R., Young, H. S., Galetti, M., Ceballos, G., Isaac, N. J. B., & Collen, B. (2014 ). Defaunation in the Anthropocene. Science , 345 (6195), 401 406. https://doi.org/10.1126/science.1251817 insecticide travels through a soil food chain, disrupting biologica l control of non - target pests and decreasing soya bean yield. Journal of Applied Ecology , 52 (1), 250 260. https://doi.org/10.1111/1365 - 2664.12372 Douglas, M. R., & Tooker, J. F. (2015). Large - Scale Deployment of Seed Treatments Has Driven Rapid Increase in Use of Neonicotinoid Insecticides and Preemptive Pest Management in U.S. Field Crops. Environmental Science & Technology , 49 (8), 5088 5097. https://doi.org/10.1021/es506141g Fernandez - Cornejo, J., Nehring, R. F., Osteen, C., Wechsler, S., Martin, A., & Vialou, A. (2014). Pesticide Use in U.S. Agriculture: 21 Selected Crops, 1960 - 2008. SSRN Electronic Journal . https://doi.org/10.2139/ssrn.2502986 Forister, M. L., Cousens, B., Harrison, J. G., Anderson, K., Thorne, J. H., Waetjen, D., Nice, C. C., Parsia, M. D., Hladik, M. L., Meese, R., Vliet, H. van, & Shapiro, A. M. (2016). Increasing neonicotinoid use and the declining butterfly fauna of lowland California. Biology Letters . https://doi.org/10.1098/rsbl.2016.0475 Fox, R. (2013). The decline of moths in Great Britain: A review of possible causes. Insect Conservation and Diversity , 6 (1), 5 19. https://doi.org/10. 1111/j.1752 - 4598.2012.00186.x Furlan, L., & Kreutzweiser, D. (2015). Alternatives to neonicotinoid insecticides for pest control: Case studies in agriculture and forestry. Environmental Science and Pollution Research , 22 (1), 135 147. https://doi.org/10.100 7/s11356 - 014 - 3628 - 7 Gilburn, A. S., Bunnefeld, N., Wilson, J. M., Botham, M. S., Brereton, T. M., Fox, R., & Goulson, D. (2015). Are neonicotinoid insecticides driving declines of widespread butterflies? PeerJ , 3 , e1402. https://doi.org/10.7717/peerj.1402 Greene, W. (2010). Testing hypotheses about interaction terms in nonlinear models. Economics Letters , 107 (2), 291 296. https://doi.org/10.1016/j.econlet.2010.02.014 Horowitz, J. K., & Lichtenberg, E. (1994). Risk - Reducing and Risk - Increasing Effects of Pes ticides. Journal of Agricultural Economics , 45 (1), 82 89. https://doi.org/10.1111/j.1477 - 9552.1994.tb00379.x Kniss, A. R. (2017). Long - term trends in the intensity and relative toxicity of herbicide use. Nature Communications , 8 , 14865. https://doi.org/10. 1038/ncomms14865 Lark, T. J., Mueller, R. M., Johnson, D. M., & Gibbs, H. K. (2017). Measuring land - use and land - 92 Cautions and recommendations. International Journal of Applied Ear th Observation and Geoinformation , 62 , 224 235. https://doi.org/10.1016/j.jag.2017.06.007 Lechenet, M., Dessaint, F., Py, G., Makowski, D., & Munier - Jolain, N. (2017). Reducing pesticide use while preserving crop productivity and profitability on arable fa rms. Nature Plants , 3 (3), 1 6. https://doi.org/10.1038/nplants.2017.8 Participate in Payment - for - Environmental - Services Programmes. Journal of Agricultural Economics , 63 (3), 604 626. https://doi.org/10.1111/j.1477 - 9552.2012.00358.x Mourtzinis, S., Krupke, C. H., Esker, P. D., Varenhorst, A., Arneson, N. J., Bradley, C. A., Byrne, A. M., Chilvers, M. I., Giesler, L. J., Herbert, A., Kandel, Y. R., Kazula, M. J., Hunt, C. , Lindsey, L. E., Malone, S., Mueller, D. S., Naeve, S., Nafziger, E., Reisig, D. benefits to US farmers. Scientific Reports , 9 (1), 1 7. https://doi.org/10.1038/s41598 - 019 - 47442 - 8 Nuyttens, D., Devarrewaere, W., Verboven, P., & Foqué, D. (2013). Pesticide - laden dust emission and drift from treated seeds during seed drilling: A review. Pest Management Science , 69 (5), 564 575. https://doi.org/10.1002/ps.3485 Osteen, C. D., pesticide use trends. Pest Management Science , 69 (9), 1001 1025. https://doi.org/10.1002/ps.3529 Pannell, D. (1991). Pests and pesticides, risk and risk aversion. Agricultural Economics , 5 (4), 361 383. https://doi.org/10.1016/0169 - 5150(91)90028 - J Perry, E. D., Ciliberto, F., Hennessy, D. A., & Moschini, G. (2016). Genetically engineered crops and pesticide use in U.S. maize and soybeans. Science Advances , 2 (8), e1600850. https: //doi.org/10.1126/sciadv.1600850 Perry, E. D., & Moschini, G. (2019). Neonicotinoids in U.S. Maize: Insecticide Substitution Effects and Environmental Risk (Working Paper No. 19 - WP 590; CARD Series Papers). Iowa State University, Center for Agirculture and Rural Development. Pleasants, J. M., & Oberhauser, K. S. (2013). Milkweed loss in agricultural fields because of herbicide use: Effect on the monarch butterfly population. Insect Conservation and Diversity , 6 (2), 135 144. https://doi.org/10.1111/j.1752 - 45 98.2012.00196.x Biological Conservation 12, no. 2 (September 1, 1977): 115 34. https://doi.org/10.1016/0006 - 3207(77)90065 - 9 . Pollard, E., & Yates, T. J. (1994). Monitoring Butterflies for Ecology and Conservation: The British Butterfly Monitoring Scheme . Springer Science & Business Media. 93 R Core Team. (2018). R: A language and environment for statistical computing . R Foundation for Statistical Computing. https://www.R - project.org/ Saunders, S. P., Ries, L., Oberhauser, K. S., Thogmartin, W. E., & Zipkin, E. F. (2018). Local and cross - seasonal associations of climate and land use with abundance of monarch butterflies Danaus plexip pus. Ecography , 41 (2), 278 290. https://doi.org/10.1111/ecog.02719 Saunders, S. P., Ries, L., Oberhauser, K. S., & Zipkin, E. F. (2016). Evaluating confidence in climate - based predictions of population change in a migratory species. Global Ecology and Biog eography , 25 (8), 1000 1012. https://doi.org/10.1111/geb.12461 Sponsler, D. B., Grozinger, C. M., Hitaj, C., Rundlöf, M., Botías, C., Code, A., Lonsdorf, E. V., Melathopoulos, A. P., Smith, D. J., Suryanarayanan, S., Thogmartin, W. E., Williams, N. M., Zhan g, M., & Douglas, M. R. (2019). Pesticides and pollinators: A socioecological synthesis. Science of The Total Environment , 662 , 1012 1027. https://doi.org/10.1016/j.scitotenv.2019.01.016 Swinton, S. M., & Van Deynze, B. (2017). Hoes to Herbicides: Economic s of Evolving Weed Management in the United States. The European Journal of Development Research , 29 (3), 560 574. https://doi.org/10.1057/s41287 - 017 - 0077 - 4 Thomas, J. A. (2016). Butterfly communities under threat. Science , 353 (6296), 216 218. https://doi.o rg/10.1126/science.aaf8838 Thorton, P. E., Thorton, M. M., Mayer, B. W., Wei, Y., Devarakonda, R., Vose, R. S., & Cook, R. B. (2018). Daymet: Daily Surface Weather Data on a 1 - km Grid for North America, Version 3 [Map]. https://daac.ornl.gov/cgi - bin/dsviewer.pl?ds_id=1328 USDA National Agricultural Statistics Service Cropland Data Layer . (2019). USDA - NASS. https://nassgeodata.gmu.edu/CropScape/ Wepprich, T., Adrion, J. R., Ries, L., Wiedmann, J., & Haddad, N. M. ( 2019). Butterfly abundance declines over 20 years of systematic monitoring in Ohio, USA. PLOS ONE , 14 (7), e0216270. https://doi.org/10.1371/journal.pone.0216270 Wooldridge, J. M. (2010). Econometric Analysis of Cross Section and Panel Data . MIT Press. http ://ebookcentral.proquest.com/lib/michstate - ebooks/detail.action?docID=3339196 Zilberman, D., & Millock, K. (1997). Pesticide Use and Regulation: Making Economic Sense Out of an Externality and Regulation Nightmare. Journal of Agricultural and Resource Econ omics , 22 (2), 321 332. JSTOR. Zipkin, E. F., Ries, L., Reeves, R., Regetz, J., & Oberhauser, K. S. (2012). Tracking climate impacts on the migratory monarch butterfly. Global Change Biology , 18 (10), 3039 3049. https://doi.org/10.1111/j.1365 - 2486.2012.02751 .x 94 CHAPTER 3. The Value of Timeliness: How Soybean Farmers Choose to Custom Hire for Pest Control Abstract Farmers frequently outsource machinery - intensive field operations to custom operators. In doing so, farmers expose themselves to the risk that fieldwork will not be completed in a timely manner, potentially reducing their yields and revenue. Custom hiring occurs even for activities such as pest control, where losses from late spraying can be particularly large. These potential losses, known as timeliness costs, can be exacerbated when contracting , and therefore can be considered a form of transaction costs . This paper develops a farmer choice model of custom hiring for pest control that is rooted in transaction cost theory. Hypotheses derived from this model are then illustrated through a discrete choice experiment conducted via a web survey of soybean growers in Michigan, Illinois, and Indiana. In this pilot study, farmers respond to a hypothetical pest infestation by choosing between custom operators, spraying on their own, or leaving the field to its fate. Our results imply that, among farmers who choose to spray, the mean willingness - to - pay for marginal increases in timeliness (as defined as the chance of late spraying) ranges from 37 to 52 cents per acre. We also find that farmers who are more averse to risk are more sensitive to custom operator timeliness; farmers with better developed social networks are less sensitive to risk of delay. The results of this pilot study can be used to motiv ate future avenues of research into the drivers of custom hire behavior in pest control and other field operations. 95 Introduction Many fieldwork activities necessary to produce field crops (i.e. corn, soybean, wheat) in the Midwestern United States requir e large, and often expensive, agricultural machinery. While the need for such machinery is ubiquitous, the assignment of property rights over such investments often differs from operation to operation. Some farmers choose to invest in and operate such mach inery themselves, while others choose to hire custom operators, who own and operate their own machinery, to complete specific machinery - intensive activities. Custom hiring is used extensively by farmers across the Midwest, though not universally. Figure 3. 1 presents trends for custom work. Between 2007 and 2012, the number of farms in Illinois, Indiana, and Michigan hiring custom for any fieldwork increased 23%, 37%, and 29% respectively (USDA NASS, 2014) . Expenditures on custom work rose significantly as well during the same period, increasing by 75%, 104%, and 56% in Illinois, Indian a, and Michigan respectively (USDA NASS, 2014) . The aim of this paper is to exam ine why some farmers choose to custom hire while others choose ownership: a question of vertical control (Coase, 1937; Klein, 2005) . Such questions are frequently studied in the context of trans action costs economics (TCE), which is focused on conditions under which vertical control of multiple stages of production is efficient relative to contracting as means of mitigating costs that emerge from conflicting incentives between contracting parties . Whether contracting or vertical control emerges as the efficient institutional arrangement is often a function of both the transaction and the potential participants in question (Williamson, 1979) . While custom hiring i s widespread, the use of custom operators varies widely by production task (e.g. planting, fertilizer application, pesticide application, harvesting). Among 96 Figure 3.1. Custom Hiring Trends in Three Soybean - Growing States. collected in 2017 via a mail survey of 1,478 soybean farmers across Illinois, Indiana, and Michigan. 97 corn and soybean growers in Illinois, Indiana, and Michigan, custom operators are hired to apply distinguished by the degree to which they are vulnerable to unexpected events leading to lapses in work quality or timeliness, which in turn can lead to decreased yields and farm profitability (Allen & Lueck, 2004) . The window for effective completion of many field operations ( referred (Apland, 1993) . For pest control in particular, the window for applying insecticides to is particularly stochastic, as insect pest p opulations, such as soybean aphid ( Aphis glycines ), can arrive unexpectedly and grow exponentially if left untreated . Further, the degree of yield loss if pest control is not completed in a timely manner can be catastrophic, leading to potential soybean yield losses of as much as 50% (Johnson et al., 2009) . From the perspective of the farmer, choosing custom contracting over vertical control adds another layer of uncertainty, as the completion of the task is dependent on the actions of another agent under imperfect observability. Because pest control is especially vulnerable to field day stochasticity and the penalties for lapses in timeliness are so extreme, pest control is an a ttractive task through which to examine the decision to hire custom operators . Previous research has examined drivers of farmer choices to contract at either end of the production cycle. Some studies focus on the choice to access different marketing channe ls and the characteristics of contracts that govern them (Davis & Gillespie, 2007; Dorward, 2001; Franken, Pennings, & Garcia, 2009; Hobbs, 1997; Hudson & Lusk, 2004; Hueth, Ligon, Wolf, & Wu, 1999; Royer, 2011) . Others focus on the control of property rights and contracting characteristics for arable land (Allen & Lueck, 1992, 1992, 1993) . These studies model the choice to contract or among contracts as functions of drivers of transaction costs, or the presence 98 of factors that mitigate them. Such studies typically find support f or the hypothesis that the contracts or organizational structures that minimize transaction costs are ultimately selected, confirming the central hypothesis of transaction cost economics. Both uncertainty and the value of the asset subject to uncertainty a re cited as drivers of transaction costs, creating frictions that prevent efficient contracting (Williamson, 1979) . In this paper, we examine the role of uncertainty in driving transaction costs in pest control, which we expect can explain the relatively low rate of custom hiring (i.e. contracting) for this field operation. In the following section, we build a conceptual model for custom hiring in pest control based around uncertainty and the probability a custom operator provides timely service. Beyond examining the role of uncertainty in increasing transaction costs, we also examine the role of social capital in mitigating such costs by providing information networks and reputational punishment mechanisms to distinguish b etween trustworthy and untrustworthy custom operators (Williamson, 1993; Wilson, 2000) . In the third section, we de scribe a choice experiment conducted with farmers in Michigan, Indiana, and Illinois designed to illustrate the implications of our conceptual model. In the fourth section, we describe our empirical strategy for analyzing the choice experiment data, follow ed by results. We close with a discussion of our findings in the context of our hypotheses and broader transaction costs literature, and implications for the future of custom hiring in pest control and other activities. Conceptual Model In this section, w e first describe a simple choice model in which a farmer faces an acute pest infestation and chooses among three possible responses: (1) spraying with their own 99 equipment, (2) hiring a custom operator to spray for them, or (3) not spraying at all. We then elaborate on the model with transaction costs theory and expected utility theory to motivate models provide insight into the tradeoffs involved in custom hiring fo r pest control and the characteristics of farmers who might be more likely to custom hire. separable from the rest of their production activities. All production costs not related to pest omitted from the analysis. Base Model To begin, assume a profit maximizing farmer with one field growing a generic crop. The field has yield pot ential of ( assumed to be certain for simplicity ) which sells at a price we set to one as the numeraire. the appropriate insecticide on time, the pest infestation does no yield damage. If the insecticide is applied late, the pest inflicts damage eld potential . The farmer chooses a response , where represents the ownership option in which the farmer sprays with their own sprayer, represents the custom hiring option, and represents the option to not spray at all and accept all damages . Each alternative has an associated profit , where is the alternative - specific cost of pest control and is yield realized when option is selected: (3.1) 100 F or now, we assume that choosing either or will lead to on - time treatment with certainty, and so and no damage occurs. When is selected yield damage is realized and no costs associated with s praying are accrued, so . For the remainder of the analysis, we will focus on differences between and , though it should be noted that for values of close to zero and spraying costs sufficiently high, sprayi ng will not occur, consistent with the concept of an economic damage threshold (Rags dale et al., 2018) . It is clear in the above model that the crucial differences between spraying options is the alternative - specific cost of treatment, . These costs are given as: (3.2) . (3.3) In Equation (3.2) and Equation (3.3), insecticides chemical costs, labelled as , are incurred under both alternatives and therefore ignored in the comparative analysis. When is chosen, additional costs incurred include , the wages (or equivalently opportunity cost of time) of the farmer applying the chemicals, and , the cost of owning and operating the sprayer. Equipment costs embody fuel, maintena nce, and depreciation, as well as any costs involved with procuring a sprayer if one is not readily available such as search costs procurement costs can be prohibiti vely high if a farmer has not made prior arrangements via long - term rental or ownership of a sprayer. For , the only additional cost is , the amount paid to the operator in return for services. This simple cost comparison leads to three proposed hypotheses: H1: When on - farm labor is more costly, custom hiring is more likely. 101 H2: When a farmer owns their own sprayer, custom hiring is less likely. H3: When a custom op erator charges a higher fee, they are less likely to be hired . Tran saction and Timeliness Costs between producing its own inputs or procuring them via contracting with another firm. Referred to as the make - or - buy decision, this topic is frequently examined in the transaction costs economics (TCE) literature (Coase, 1937; Klein, Crawford, & Alchian, 1978; Klein, 2005; Shelanski & Klein, 1995; Williamson, 1979) . In the custom pest control context, the input is the application of insecticides on a specific insect - infested field. The farmer chooses betwe en spraying with their own equipment (the vertical integration option) and hiring a custom operator (the contracting option). The key insights of TCE are (a) that transactions require costly governance; (b) that these costs, referred to as transaction cost s, vary among alternative governance structures depending upon the characteristics of the activity or asset exchanged and the identities of the trading partners; and (c) that firms will utilize governance structures that minimize such transaction costs (Shelanski & Klein, 1995) . Transactions vary in many ways, but TCE studies have identified two transaction characteristics as especially important: asset specificity and uncertainty (Klein, 2005) . Asset specificity is typically defined as the degree to which investmen t in assets or actions are specific to the transaction and therefore cannot be recovered should the transaction fall through. Uncertainty in this case relates to the value of said assets or actions and the behavior of trading partners (Klein, 2005) . Asset specificity and uncertainty can create 102 circumstances for trading partners to act opportunistically (Klein, 2005) . Governance structures (e.g. contracts, markets, or vertical integration) can mitigate the incentives to do so, though such structures often create additional administrative costs (e.g. monitoring, enforcement, etc.) (Klein, 2005) . Asset specificity can manifest itself temporally, especially for pesticide application. Once an economically significant pest infestation is recognized, there often exists a critical period during which the infestation can be treated before risking significant yield loss. The value of these losses due to late treatment are referred to as timeliness costs (Allen & Lueck, 2004) . However, the exact dates during the growing season when pests will approach economically damaging levels, or whether a pest infestation will occur at all, is impossible to know a priori (Johnson et al., 2009) . Typically, farmers must choose whether they will spray on their own or hire a custom operator before such uncertainty is resolved. When custom hiri ng, a farmer forfeits control over when and where the sprayer is used , which can increase the likelihood of delays in treatment , amplify ing potential timeliness costs (Allen & Lueck, 2004) . A farmer who owns and operates their own sprayer can more readily apply pesticides precisely when and where they are needed once uncertainty regarding a potential pest infestation is resolved. The custom operator may have other customers with pest infesta tions simultaneously occurring and must choose whose field to treat first. Random occurrences that would lead to delays even if the farmer chose to spray on their own, like unexpected weather, are amplified if they chose to custom hire as they further incr ease the likelihood of overburdening the custom operator. Because of the high degree of uncertainty, limited optimal treatment window, and large potential yield losses surrounding pest control, timeliness costs have the potential to be sizable. 103 We include timeliness costs into the model by introducing uncertainty, from the perspective of the farmer, over whether pest control will be completed within the optimal window. Rather than assuming each is certain, instead now assume that each is a binary random variable with support , the full yield when spraying occurs within the optimal window and the damaged yield when late spraying occurs. Let be the alternative - specific probability that treatment is delayed from the perspective of the f armer, so that . (3.4) Assuming all other variables are known with certainty, then . (3.5) Under assumptions of expected - profit maximization, the farmer chooses the alternative that maximizes Eq uation (3.5). Assume for simplicity that and . For , this assumption is trivial because there is no chance of damage avoidance if no spray ing occurs , so the application is late by definition . For , this assumption is equivalent to assuming that the spray will always occur on time if farmer is doing so themselves. 1 The probability of on - time pest control for custom operators, , is more complex. From the perspective of the farm er, the probability that a custom operator sprays on time is related to the concept of trust. Bhattacharya et al. (1998) prop ose a formal definition of trust an expectancy of positive (or nonnegative) outcomes that one can receive based on the expected action of another party in an interaction characterized by uncertainty In this model of trust, agents h (Bhattacharya et al., 1 Equivalently, the problem be scaled so that is the baseline probability and custom operators are compared to t hat baseline. Then the assumption becomes that all relevant external drivers such as weather influence no more than they do so that all additional uncertainty can be attributed to the custom operator. 104 1998) . In the above model, serve as conjectures in the Bhattacharya et al. sense. Bhattacharya et al. do not hypothesize as to how conjectures are formed. At this stage, we will assu me simply that conjectures exist; factors supporting favorable or unfavorable conjectures will be discussed later in this section. With timeliness costs included in the model as above, it is clear that farmers will be less likely to choose a custom operato r when is closer to one, all else equal. Further, , the damage suffered when spraying is late, now appears in the function, Equation (3.5). Each alternative is associated with a nd when timeliness costs are larger, the alternative becomes less attractive to the farmer. Timeliness costs are composed of the probability of delay, the damage from delay, and the yield potential. We propose the following hypotheses: H4: When the the dam age from delay is higher (i.e. dY is larger ) , custom operators are less likely to be hired. H5: When the probability a custom operator is delayed in spraying is higher (i.e. is closer to one ) , that custom operator is less likely to be hired. Risk Aversion In the development of the preceding model , the farmer is assumed to be risk neutral and would be indifferent between two alternatives: an alternative with a pre - determined profit and one with multiple possible outcomes but the same pr ofit in expectation . In this modification of the model, we allow for a more general case where farmers may be risk averse, preferring non - stochastic options to stochastic alternatives with equivalent expected profit. Rather than assuming that farmers maxim ize expected profit, we instead assume farmers choose the response 105 to potential pest infestation which maximizes their expected utility, , a function of profit conditioned by a risk attitude parameter , where larger values indicate more aversion toward s risk. Each custom operator can be thought of as a lottery, with payoffs and and conjectures about the reliability of the custom operator serving as probabilities of each outcome ive lotteries. Costs serve as the price of each lottery. For each alternative, the expected utility is then represented as: . (3.6) By including risk attitude parameter we allow for a variety of possible behavioral theories, including the curvature of the utility function (Von Neumann & Morgenstern, 1944) , loss aversion and probability weighting schemes (Tversky & Kahneman, 1992) , and models that do not rely on weighted averages of outcomes and allow uncertainty to directly affect utility (Gneezy, List, & Wu, 2006) . If some farmers are more averse to risk than others, then this should be reflected in their custom hiring decisions, and farmers who are more risk averse will be less likely to select risky options. This constitutes our sixth hypothesis: H 6 : More risk averse farmers (i.e. is larger) are more sensitive to potential delays in spraying . Conjectures and Social Capital viewed as a function of not just the relationship between the farmer and the operator themselves, but also other circumstances surrounding the relationship r elevant to the decision in question. In this section, we draw on social capital theory to form an additional hypothesis as to what factors drive custom hiring decisions. Social capital theory proposes that social institutions and relationships between peop le can be viewed as productive assets (Schmid & Robison, 1995) . Personal relationships and social 106 networks can reduce transaction costs by easing the flow of information and establishing informal punishmen t systems for those who violate norms (Schmid & Robison, 1995) . For example, empirical research analyzing over 3,000 cropland rental contracts, both formal and informal, from Nebraska and South Dakota find s evidence supporting reputational enforcement mechanisms in which farmers with more developed social networks are more likely to participate in informal (i.e. unwritten) contracts (Allen & Lueck, 1992b) . In the context of custom pest control, a similar social capital mechanism may exist. Farmers who consistently communicate with many other farmers can both rely on other farmers spraying by a specific operator. Therefore, the probability that a custom operator delays pesticide application, , may depend on t he social network of the farmer who is contracting for the work. To capture this possible effect, we present capital, represented as : . (3.7) When fa rmers have more social capital, they can rely on these networks to punish custom operators who provide late service by damaging their reputation among potential customers. All else equal, such punishment introduces additional costs to the operator for spra ying late. Therefore, from the perspective of the farmer, any given operator is less likely to provide late service, and farmers are likely to weight probabilities of delay downward (i.e. ). This leads to the following propo sed hypothesis: H7: When a farmer has a more developed social network, they are less sensitive to potential delays in spraying by custom operators. 107 Choice Experiment and Survey As an illustration of the implications of the custom hire model presented abov e (i.e. hypotheses H1 - H 7 ) , we deployed a discrete choice experiment. A choice experiment allows the researcher to gather choice data on important decisions even when they occur infrequently and allows the researcher to observe the full choice set because the choice set is designed by the researcher themselves (Hensher, Rose, & Greene, 2015) . We deployed the choice experiment described below as a pilot study to motivate future, in - depth analy sis of the potential roles transaction cost drivers play in the decision to custom hire. While the choice experiment presented in this study does not definitively test hypotheses H1 - H7, it does provide initial data on which of the hypotheses may be most pr omising for further analysis. Experimental Design In this choice experiment, farmers were asked to imagine that their largest soybean field is infested by a generic, unspecified insect pest. The characteristics described for the hypothetical pest were sim ilar to soybean aphid, though the species was not mentioned by name. We utilized soybean as a model crop because soybean farmers are likely to have recent experiences with acute pest infestations during the spread of soybean aphid in the mid - s were presented with the option of hiring one of three custom operators to spray, spraying themselves with their own equipment, or not spraying at all. Farmers were presented with chemical spraying costs (in dollars per acre) and expected soybean price (i n dollars per bushel). These attributes remained fixed for all farmers through all choice scenarios. Farmers were told that they would be responsible for chemical costs in all spraying options (as is typical in custom pesticide application contracts) and i nstructed to assume 108 that all custom options are available even if those specific options are not present in their area. The soybean price and chemical cost values were selected such that spraying dominates not spraying for a profit - maximizing, risk - neutral farmer in the scenario with the lowest yields and the highest damage. The focus of this design, therefore, is on the choice of who sprays, though the option to not spray and allowing damage to occur remains. Respondents each completed eight choice scenar ios. Each scenario included a specific expected pest damage attribute, which referred to the portion of yield lost to insect damage if spraying occurs after a three - day window. This attribute is the variable in the conceptual model. Within each scenario , each custom operator option was presented as one of the following: - farmer. These three classes of custom operators represent the most common providers of cus tom pest control services. All three options were presented in each choice scenario, along with a e field to its fate and damage would be guaranteed. Each of the custom choices had an associated custom fee, presented as a dollar per acre fee paid to the operator, and a percent chance of a three - day delay, representing the probability that the pest dama ge occurs due to late spraying ( ) . Fee levels were based on the range of custom spraying rates reported in extension survey reports from Ohio, Michigan, Illinois, Iowa, and Indiana. Levels for each variable included in the choice experiment are present ed in Table 3.1. An example of a choice scenario as seen within the survey is presented in Figure 3.2. 109 Table 3.1. Levels and Descriptions of Choice Experiment Attributes. Attribute Description Value ( s ) Fixed Chemical costs Costs of insecticides used to spray, measured in dollars per acre $5/ac Soybean price Price of soybean at harvest, measured in dollars per bushel $9/bu Scenario Pest damage Damage the insect pest would induce if spraying is delayed three days, measured in portion of yield potential ( ) 10%, 20%, 30% (3) Choice Custom fee Fee paid to operator for services, measured in dollars per acre ( ) $5/ac, $9/ac, $13/ac (3) Chance of delay Probability spraying occurs three days late and pest damage occurs, measured as a percentage ( ) 20%, 40%, 60% (3) Operator identity Identity of the custom operator Co - op, input dealer, another farmer (3) Figure 3.2. Example Choice Scenario. 110 A 24 - row fractional factorial experimental design was generated using the software package Ngene and split into three blocks of eight scenarios each. After consulting subject - matter experts during the design stage, eight choice scenarios was determined to be the maximum feasible number of scenarios per farmer. Budget considerations pre - empted pilot data collection necessary for the use of priors in the generation of designs targeting efficiency criteria (Hensher et al., 2015) . Because of these constraints, the design was generated by randomly selecting 24 rows from the full factorial design. A common critique of stated preference methods, i ncluding choice experiments, is that they are potentially prone to hypothetical bias. Hypothetical bias occurs when participants respond differently in hypothetical settings than they do when faced with actual decisions (Tonsor & Shupp, 2011) . To mitigate hypothetical bias, respondents were presented with an additional page encouraging farmers to take their time and res pond as if their choices would have script, a method that has been shown to reduce hypothetical bias in a variety of settings (List, 2001; Lusk, 2003; Silva, Nayga, Campbell, & Park, 2011) , including online surveys (Tonsor & Shupp, 2011) . Before deployment, the survey was reviewed by 20 professionals in the agricultural community unassociated with the study, including employees of the Michigan Department of Agriculture and Rural Development, Michigan State University Extension, and the Michigan Soybean Promotion Committee, as well as active farmers. Comments from phone and email interviews with reviewers were incorporated into the survey design to improve clarity and ens ure the choice experiment represented a feasible scenario. 111 Survey Deployment The target population for the survey was farmers with 100 or more acres of soybeans planted in 2017 in Michigan, Indiana, and Illinois , with a focus on Michigan farmers . We employed a web survey design, utilizing both email and postal mail contacts to improve response rates (Dillman et al., 2014) . Sixty - five farmers completed surveys for a total of 519 completed choice scenarios (o ne choice scenario was left incomplete ). The survey deployment system was programmed to randomly assign a block to each respondent in a balanced manner, and so two blocks were completed 22 times while the third was complete d 21 times. Additional Survey Data The experimental data is supplemented with additional data from the survey. Farmers were asked questions about their past spraying and custom hiring activities, their capacity to spray with on - farm equipment, and the characteristics of their farm. Farmers were also asked about their general attitudes towards trust and risk, and the number of other farmers with whom they are comfortable discussing important business matters. The full survey instrument is provided in Appendix A. Empirical Analysis The goal o f this pilot empirical model is to examine the relevance of the proposed hypotheses H1 - 7 that emerge from the conceptual model by translating the conceptual model into an empirically tractable form. To do so, we estimate a series conditional logit models o n the choice experiment data and accompanying survey data on respondent characteristics. First, we estimate a series of candidate models using only variables for the attributes in the choice 112 experiment and select a preferred model based on a number of mode l selection criteria. We then expand on the preferred model by including farmer characteristics, allowing us to explore how preferences for custom hiring vary according to attitudes and resources. M1: Base Model Candidates The conditional logit model is based on the random utility model, in which the farmer is assumed to associate a level of utility with each alternative in a choice set and select the alternative that provides them with the highest utility level (McFadden, 1973) . The level of utility associated with each alternative consists of both a systemic portion for which characteristics (i.e. cost, quality, etc.) are known to the econometrician and a stochastic portion accounting for characteristics unobserved by the econome trician. By estimating conditional logit models using data on observed choices, one can measure farmer preferences over, and willingness - to - pay for, and the characteristics of those alternatives. Random utility models are typically applied in settings wher e the characteristics included in the systemic portion of utility are certain. More recent applications of the random utility model have considered cases where there is uncertainty over whether an alternative possesses one or more characteristics, often by including the probability that an alternative possesses a characteristic as a characteristic itself. Applications include measuring preferences for environmental quality where the outcome of a project is uncertain (Faccioli, Kuhfuss, & Czajkowski, 2019; Glenk & Colombo, 2011, 2013; Lundhede, Jacobsen, Hanley, Strange, & Thorsen, 2015; Makriyannis, Johnston, & Whelchel, 2018; Roberts, Boyer, & Lusk, 2008; Rolfe & Windle, 2015) and preferences for travel time reliability of transportation options (Hensh er, Greene, & Li, 2011; B. Li & Hensher, 2017; H. Li, Tu, & Hensher, 2016; Z. Li, 2018; Z. Li & 113 Hensher, 2013; Z. Li, Hensher, & Rose, 2010) . In these studies, the possible levels of the characteristic and the probability of the level occurring both inf luence how each alternative affects utility. The two are frequently interacted to capture changes in the expected value of the uncertain characteristic. In our setting, the uncertain characteristic is whether the custom applicator arrives on time and prev ents the pest from damaging the crop. This uncertainty is present only for the custom options. When a farmer chooses to spray on their own the probability of delay is assumed to be zero, and when they choose not to spray the probability is assumed to be o ne (i.e. damage is guaranteed), consistent with the framing in the choice experiment and the conceptual model. In our simplest model (M1 - ED, for expected damage), we assuming that farmers hold preferences over outcomes rather than probabilities and that th e utility of outcomes is weighted by the probability they occur (Von Neumann & Morgenstern, 1944) . In our setting, the outcome is the potential pest damage to the farmers crop, . This value is measured in bushels per acre and is computed for each farmer for each choice occasion as the percent of yield loss unique to the choice occasion multiplied by their expected pest - free soybean yield. The probability of this damage occurring is given by for a given alternative. M1 - ED can be characterized by the following (dis)u tility expressions: M1 - E D : (3.8) The alternative - specific constants ( , and , represent coefficients to be estimated. The alternative specific constants capture the average of all unobserved sources of (dis)utility associated with each alternative (Hensher et al., 2015) , including the costs and associated with the option and residual preference for specific custom operator options. 114 Note that not spraying is the base level for the alter native specific constants, the probability of delay ( ) is zero if the farmer chooses to spray on their own, the probability that damage will occur is one if the farmer chooses not to spray at all, and for the non - custom alternatives. Includi ng the interaction between and allows for a measure of the marginal (dis)utility for expected damages, a component of timeliness costs. This model assumes that farmers are risk neutral and that the utility effects of damage and probability of del ay are inseparable. Versions have been used in settings evaluating preferences under outcome uncertainty in environmental and transportation settings (Burghart, Cameron, & Gerdes, 2007; Glenk & Colombo, 2013; Li et al., 2010; Roberts et al., 2008) . An alternative model, M1 - DU (for direct utility), allows the probability of delay to affect utility bot h through the effect on expected damage and a direct effect on utility itself. M1 - DU, known as the direct utility model, is a simple extension of M1 - ED, with the additional term added to the utility expression, where is an indicator variable equal to one for custom hiring alternatives where delay is uncertain (Glenk & Colombo, 2013) . The M1 - DU model is thus characterized as: M1 - DU: (3.9) Including separate from the interaction of with allows for residual preference for timeliness beyond what is measured by the interaction term. If farmers are ri sk neutral and weighting damage exactly according to the probability of its occurrence, then would be estimated at zero. A negative estimate of would suggest that farmers are risk - averse or otherwise have a distaste for uncertainty beyond its effects on increasing expected damage, indicating risk aversion or other behavioral distortions from simple risk - neutral 115 weighting of outcomes. Such direct utility models have been found to outperform expected utility models in various settings, including contracts for irrigation water contingent on uncertain rainfall and emissions reduction programs (Glenk & Colombo, 2011, 2013; Rigby, Alcon, & Burton, 2010) . risky alternatives because of the existence of risk itself, rather than because of the effects of risk on the expected utility of alternative, a result that has been used to expla in apparent violations of the internality axiom of expected utility theory (Gneezy et al., 2006) . M1 - ED and M1 - DU are also considered against a model specification that is linear in the choice attributes. Such a model assumes that all interaction effects are zero and farmers do not condition yield outcomes by the probability of their occurrence (Glenk & Colombo, 2013) . While this is an extreme assumption, the model is retained to ensure as a test that such probability conditioning occurs. This model, M1 - L, is presented below: M1 - L: (3.10) To select a preferred model, we use multiple criteria. Noting that M1 - EU is nested in M1 - DU, we test the linear restriction that in M1 - DU using a likelihood ratio test to establish which of these two models is preferred. To compare M1 - L to the preferred of M1 - DU or M1 - EU, 2 (AIC) and the alternative Bayesian information criteria (BIC), as M1 - L does not nest in either M1 - DU or M1 - EU (Burnham & Anderson, 2002) . 2 Hurvich and Tsai (1989) find that AIC performs poorly in small sample settings. However, the corrected criterion, AICc, introduces a bias correction term that is a function of only the number of parameters of the model and the sample size. Because these values are the same for t he models compared by the AIC criterion, and therefore the rankings of AICc and AIC will be identical, we choose to rely on the simpler AIC criterion. 116 M2: Preference Heterogeneity by Farmer Characteristics To examine how farmer sensitivity to delay varies across farmers and to illustrate hypotheses related to farmer characteristics, we expand on the base model selected from M1 - ED, M1 - DU, and M1 - L by introducing variables that characterize th e farmer. These variables can be interacted with characteristics of the alternatives (i.e. , ) so that the resulting coefficients (i.e. , , , the alternative specific constants) can vary by the characteristics of the farmer responde nt. We build this model from the model that emerges from the model selection process and label the resulting model with interactions M2. We select characteristics relevant to our conceptual model and hypotheses, dividing characteristics into two classes: c haracteristics that affect utility under custom options and characteristics that affect utility when the farmer chooses to spray on their own. social capital to illustrat e H7. This variable measures the number of other farmers, excluding re - rating on a four - point scale have strong predictive validity for risky behavior and responses are simpler to collect in a survey setting than alternative measures of risk attitudes (Beauchamp, Cesarini, & Johannesson, 2017) . To examine how farms of different sizes respond differently to uncertainty, we include the total acres the farmer planted in 2018 of soybeans or any other crop. We interact each 117 selected farmer characteristic with , , and their interaction (for a three - way interaction) to examine the effects of each ch aracteristic on farmer preferences for timeliness in custom hiring. For characteristics relevant when the farmer sprays on their own, we include variables as a p roportion of total income, proxies for the opportunity cost of labor, with a higher proportion of income from off - farm sources assumed to indicate a higher opportunity cost of on - farm labor. This variable allows us to illustrate H3: that higher labor costs lead to more frequent To aid in interpretation in the coefficients for these interactions, we center each non - binary characteristic at zero by subtracting the sample mean. Therefore, the resulting coefficients can be interpreted as piecewise utility changes resulting from a unit change from the mean. Choices by Sprayer Owne rship To illustrate H2, that farmers with their own equipment are more likely to choose to spray between respondents who report owning or leasing a self - propelled sp rayer, those who own a tractor - pulled sprayer, and those who do not own a sprayer at all. Self - propelled sprayers are specialized equipment that can apply chemicals over larger areas more quickly, and at lower equipment and labor costs, than tractor - pulled sprayers. While this method cannot distinguish between equipment and labor cost savings resulting from sprayer ownership, it provides an equipment gradient over which we can examine differences in the likelihood that custom hiring is selected. 118 Results In this section, we first discuss summary statistics for selected farmer characteristics and the raw choice shares from the survey data. We then present results for the M1 models and the results of the model selection process before proceeding to M2, the sel ected model with farmer characteristic interactions. Finally, we present willingness - to - pay measures for reductions in the variable, which we define as willingness to pay for timeliness, and demonstrate the effects of farmer characteristics for these findings. Summary Statistics Most of the responding farmers report possessing both the equipment and certification to perform the tasks on their own. Seventy - five percent of respondents are certified to spray restricted use pesticides and 69 percent own o r lease their own sprayer (Table 3.2). Insecticides are used infrequently among the respondents. The median respondent sprayed insecticides twice in the past ten seasons (mean 3.4) and twice on soybeans specifically (mean 2.7) (Table 3.3). When spraying do es occur, custom spraying is not uncommon; 32 percent of respondents typically hire custom when spraying is needed. Respondents routinely hire custom operators. The median respondent hired a custom operator for any field operation in five of the past ten s easons (mean 5.5), and specifically for spraying pesticides in one of those seasons (mean 2.6) (Table 3.3). Seventy - four percent of respondents reported custom hiring for at least one field operation in at least one of the past ten seasons. 119 Table 3.2 . Summary of Categorical Survey Variables . Variable Description Category N a % of Sample b Who sprays Who sprays insecticides in a typical year Custom applicator 21 32% Employee (family excluded) 2 3% Family member 5 8 % Primary operator (myself) 34 52% Other 3 5 % Certification Whether anyone who works on the farm is certified to spray restricted use insecticides No 16 2 5 % Yes 49 75% Sprayer ownership Whether the farm owns a sprayer No 20 3 1 % Yes 45 69% Farm revenue Gross farm income in 2017 Less than $150,000 8 12% $150,000 - $349,999 25 3 9 % $350,000 - $999,999 16 2 5 % $1,000,000 - $4,999,999 8 12% More than $5,000,000 1 2 % No Response 7 1 1 % Gender Farmer's gender Female 5 8 % Male 56 86% No Response 4 6% Education Farmer's level of education High school graduate 15 23% Some college 12 1 9 % 2 - year degree 14 2 2 % 4 - year degree 18 2 8 % Professional degree 5 8 % No Response 1 2 % Household income Farmer's household income in 2017 Less than $20,000 4 6% $20,000 - $39,999 2 3% $40,000 - $59,999 7 1 1 % $60,000 - $79,999 6 9% $80,000 - $99,999 10 15% More than $100,000 26 40% No Response 10 15% State Farmer's state of residence IL 10 15% IN 22 3 4 % MI 33 5 1 % a Number of responses to each item, accounting for item non - response. Total number of responses is 65. b Percentages may not add to 100% due to rounding. 120 Table 3. 3 . Summary of Numeric Survey Variables. Variable Description Min. 25 th - Perc. Med. 75 th - Perc. Max. Mean Std. Dev. N a Past custom Years in the last ten when custom work was used 0 0.5 5 10 10 5. 5 4.3 63 Past custom, spray Years in the last ten when custom spraying was used 0 0 1 3.5 10 2.6 3. 3 63 Past spray Years in the last ten when any insecticides were used 0 1 2 5 10 3.4 3. 1 63 Past spray, soybeans Years in the last ten when insecti ci des were used on soybeans 0 1 2 3 10 2.7 2.6 59 Age Age in years 27 54 59 66 80 58 Close farmers Not including those who work on your operation, about how many other farmers would you say you feel close enough to discuss important business problems with? 0 2 3 5 15 4.2 3.0 61 Risk score Are you generally a person who is fully prepared to take risks or do you try to avoid taking risks? (Fully prepared to take risks = 1; Unwilling to take risks = 4 ) 1 2 2 3 4 2.3 0. 8 64 Expected yield Expected yield , in bushels per acre, of largest soybean field 35 53 57 65 81 58 9.3 64 Farming income Percent of farmer's household income from agriculture 0 25 52 9 5 100 58 34. 5 62 Planted acres Total planted acres in 2017 150 409 697 1251 3700 997 804 64 a Number of responses to each item, accounting for item non - response. Total number of responses is 65. 121 Most respondents consult with a small circle of other farmers on important issues, though some have larger networks. The median respondent is close enough with three other farmers to discuss important business issues (mean 4.2) and a quarter of respondents are close enough to five or more other farmers to have such disc ussions (Table 3.3). Farmers on average expressed a slight preference for risk - taking behavior, with a mean risk score of 2.3 (Table 3.3). Finally, we report the unconditional rates at which each of the alternatives was selected in the choice experiment (T able 3.4). Respondents chose to spray with their own equipment most frequently (45.5% of choice occasions). Among custom options, input dealers are selected most frequently (24.5%), followed by co - ops (15.2%), and other farmers (11.4%). Not spraying is sel ected rarely (3.5%). While we do not claim that our sample is representative of the population, we present a brief comparison of the demographic statistics of our sample relative to equivalent measures reported by the USDA for the sampled population (Appe ndix B). Model Selection and M1 Results Results for M1 - ED, M1 - DU, and M1 - L are presented in Table 3.5. In the first stage of model selection, we perform a likelihood ratio test of a single linear restriction between M1 - DU Table 3.4. Response Shares to the Choice Experiment. Alternative N % of Responses a Co - op 79 15.2% Input dealer 127 24.5% Farmer 59 11.4% Myself 236 45.5% None 18 3.5% a Percentages may not add to 100% due to rounding. 122 and M1 - ED, . The likelihood ratio test statistic is 19.24, which is larger than the chi - squared critical value at the 5% level with one degree of freedom (3.84). Therefore, we reject the null hypothesis that the restriction holds and proceed with M1 - DU as the preferred mod el over M1 - EU. We then compare M1 - DU to M1 - L by their AIC and BIC, where lower values indicate the preferred model. For both AIC and BIC, M1 - DU is preferred over M1 - L, though by two units or less in both cases, indicating that both models perform about th e same by these metrics. Because M1 - DU both performs marginally better by both model selection criteria and is consistent with the conceptual model, we proceed with M1 - DU as the preferred model. All coefficients for the preferred model are statistically significant at the 5% level or better. The coefficient for , , is negative, suggesting that farmers custom hire less when options are more expensive. The coefficient for expected dam age, , is also negative, indicating that farmers lose utility as the expected timeliness costs associated with the alternative Table 3. 5 . Results of M1 Conditional Logit Models. M1 - EU M1 - DU M1 - L Variable Coef. S.E. Coef. S.E. Coef. S.E. Fee , ($/Acre) - 0 .10*** 0.023 - 0 .11*** 0.023 - 0 .11*** 0.023 Delay P robability , - 2. 6 *** 0.47 - 3.6*** 0.47 Damage , ( bu / a cre) - 0 .03 7 ** 0.017 Expected Damage, - 0 .1 9 *** 0.028 - 0 .08 8 ** 0.017 ASC - Farmer 0 . 0 8 1 ** 0.35 2.3*** 0.37 3.3*** 0.37 ASC - Co - op 1.1*** 0. 33 2. 8 *** 0.38 3.8*** 0.3 8 ASC - Dealer 1.5*** 0.34 3.1*** 0.37 4.1*** 0.3 7 ASC - Self 0.5 2 0.3 7 1.5*** 0.32 2.1*** 0.32 Log Likelihood - 661.5 - 651.6 - 652.5 AIC 1335.1 1317.1 1319.1 BIC 1360.5 1346.8 1348.7 Choice Occasions 511 511 511 Respondents 64 64 64 ***, **, and * indicate statistical s ignificance at the 1%, 5%, and 10% level s . 123 increases. The negative estimate for coefficient indicates that farmers have a distaste for increased probability of delay separate from the effects of on their potential yields. The alternative specific constants are all positive, indicating residual preference for all spraying options over not spraying at all. M2 - DU Results with Farmer Characteristics Results for M2 - DU, the preferred model M1 - DU with farmer characteristics, are presented in Table 3.6. The core results for coefficients estimated without farmer characteristics remain consistent with M1 - DU, including the rankings of the alternative specific constants. Table 3.6. Results of M2 - DU, the Preferred Condition al Logit Model with Interactions for Farmer Characteristics. M2 - DU Variable Coef. S.E. Fee , ($/Acre) - 0.12*** 0.025 Delay Probability, - 3.6*** 0.74 x Close Farmers ( C ount) 0.39*** 0.15 x Risk Score (1 - 4, 4 is risk averse) - 0.016 0.48 x Acres Planted - 0.00021 0.00091 Expected Damage, - 0.15*** 0.053 x Close Farmers - 0.024** 0.011 x Risk Score - 0.073** 0.034 x Acres Planted - 0.00031*** 0.000072 ASC - Farmer 3.9*** 0.69 ASC - Co - op 4.4*** 0.70 ASC - Dealer 4.8*** 0.70 ASC - Self 2.6*** 0.66 x Farm Income Share (proportion) 0.0066* 0.0036 Log Likelihood - 527.9 AIC 1,083.7 BIC 1,141.9 Choice Occasions 471 Respondents 59 ***, **, and * indicate statistical s ignificance at the 1%, 5%, and 10% level s . 124 and statistically significant, indicating that farmers who are close to more farm ers are less sensitive to the probability a custom operator sprays late. On the other hand, the coefficient for broader social networks are more sensitive to increase s in expected damage. The coefficients for the remaining two characteristics interacted with timeliness cost the probability of delay, but negative and stat istically significant when interacted with expected at the 10% l evel, indicating that farmers who receive a larger share of their household income from agriculture are more likely to choose to spray on their own. Choices by Sprayer Ownership Among the 20 respondents who do not own or lease a sprayer, none chose to sp ray on their own in any choice occasion. Among the 24 respondents who own or lease self - propelled sprayers and the 21 who own or lease a tractor - selected in 76 percent and 54 percent of choice occasions. These resu lts suggest that respondents who own more specialized spraying equipment are less likely to custom hire for pest control. Willingness - to - Pay for Timely Spraying We now consider farmers willingness - to - pay (WTP) for reductions in a custom 125 coefficients from M2 - DU, WTP for reductions in the probability of delay at a given damage rat e is calculated as . (3.11) In this expression, is the mean expected soybean yield for respondents (58.1 bushels per acre). Because all characteristics interacted with the probability of delay were centered at zero prior to estimation, and therefore have means of zero, their effects are omitted from this calculation . We divide by 100 to give a measure of WTP measured in dollars per percent change in probability of delay per acre, rather than dollars per probability unit (i.e . on the scale). We find that, on average, farmers are willing to pay $0.37 per acre for a one percent reduction in the probability of delay when the potential damage rate ( ) is 10% of yield, $0.45 per acre when the potential damage rate is 20%, a nd $0.52 per acre when the potential damage is 30%. These WTP for a marginal change in the probability of delay represent 2.7 percent, 3.2 percent, and 3.7 percent of median custom pest control costs per acre as presented in the choice experiment ($9 per a cre in custom fees and $5 per acre in chemical costs). To show how farmer characteristics affect WTP to avoid delay, we illustrate WTP for each respondent in the sample according to farm size (Figure 3.3A), the number of close farmers (Figure 3.3B), and r isk score (Figure 3.3C). Respondent - level WTP is calculated by adding the statistically significant interaction coefficients, multiplied by associated characteristics and potential damage levels, to the numerator of Equation (3.11) (Hensher et al., 2015; Rigby et al., 2010) . The potential damage rate is set at each of the three levels included in the choice experiment to further demonstrate t he impact of increasing damage on WTP to avoid delay. 126 Figure 3.3. Willingness - to - Pay (WTP) for Reductions in the Probability of Delay. WTP is calculated for each respondent at each damage level and presented by (A) acres planted, (B) close farmers, and (C) risk score. Random displacement on the horizontal axis has been added to distinguish between overlapping points. 127 Discussion In this section, we discuss how the results of our pilot study relate to the hypotheses derived from the conceptual model, with particular focus on the roles of timeliness costs, risk s. Table 3.7 reports our assessments of support from the pilot study for each hypothesis. As predicted by H3, higher custom fees reduce custom hire use. However, we find only weak evidence that farmers with higher opportunity costs of labor, as measured by the proportion of household income derived from agriculture, are more likely to custom hire, as predicted by H1. On the other hand, farmers who possess more specialized equipment (i.e. self - propelled sprayers) are less likely to custom hire, a result that provides some support for H2, that sprayer ownership decreases custom use. These results support the implication of the conceptual model that farmers are more likely to custom hire when performing a task on their own is more expensive. Future research sho uld explore whether labor or equipment costs are more important in driving this outcome, which in turn would allow for better predictions of how changes in labor and equipment markets might affect custom hiring demand. The next hypotheses, H4 and H5, rela te to how uncertainty and timeliness costs affect decisions to custom hire. Our results illustrate how farmers are less likely to choose to custom hire when expected damage is higher, supporting H4. Note that both increases in the probability of delay and the increase in the absolute level of damage can drive this effect, so this result also provides weak evidence for H5. By interpreting we can disentangle (dis)taste for uncertainty from their (dis)taste for pest damage and address H5 separately from H4. Changes in can affect utility levels through two mechanisms: a direct effect and an indirect effect through increasing expected dam age. We 128 Table 3.7. Hypotheses, Relevant Parameters, and Determination of Support. Hypothesis Relevant Evidence Support? H1 When on - farm labor is more costly, custom hiring is more likely. M2 - DU: ASC Self x Farm Income Share H2 When a farmer owns their own sprayer, custom hiring is less likely. Results by sprayer ownership H3 When a custom operator charges a higher fee, they are less likely to be hired. M1 - DU/M2 - DU: Fee H4 When the scale of the damage from delay is higher (i.e. is larger), custom operators are less likely to be hired. M1 - DU/M2 - DU: Expected Damage H5 When the probability a custom operator is delayed in spraying is higher (i.e. is closer to one ), that custom operator is less likely to be hired. M1 - DU/M2 - DU: Delay Probability H6 More risk averse farmers (i.e. is larger) are more sensitive to potential delays in spraying. M2 - DU: Delay Probability x Risk Score, Expected Damage x Risk Score H7 When a farmer has a more developed social network, they are less sensitive to potenti al delays in spraying by custom operators. M2 - DU: Delay Probability x Close Farmers, Expected Damage x Close Farmers interpret as the direct utility effect of and as the expected damage effect. E ven for the farmers with the maximum potential damage (81 bushels per acre pest - free yield and 30% pest damage), the direct utility effect from is roughly equal to the utility effect through expected damage from increasing expected damage ( - 3.6 utility units per additional percen t probability of delay for both effects ) . The fact that a direct effect of exists a lso provides potential evidence that farmers do not respond to potential delays in custom hiring in a risk - neutral way under the traditional expected utility model. While the negative estimates for coefficient provide evidence of possible risk aversion among farmers in the context of custom hiring, this finding does not 129 address how individual differences in risk attitudes affect decisions, the question at the core of H6. We use the interaction coefficients from M2 - DU to address this hypothesis. Our findings suggest more risk averse farmers are more sensitive to the probability of delay, but only through its effect on expected damage. If the direct effect of represents risk averse attitudes in the traditional expected utility model sense, we would expect more risk averse farmers to have larger marginal disutility from alone. Our findings do not support this conclusion, so the direct utility effect of can be viewed as potential evidence that alternative models of risk attitudes such as prospect theory (Tversky & Kahneman, 1992) or direct uncertainty effects (Gneezy et al., 2006) may be more relevant in custom hiring scenarios. Future research should directly address alternative models of risk preferences when assessing drivers of custom hiring behavior among growers. that farmers with broader social networks are less sensitive to increases in the probabi lity of dela y in support of our hypothesis . While the coefficient for the interaction with expected damage is negative, the coefficient for the interaction with delay probability is large enough to counteract this negative effect. We int mitigating effect of social capital on the marginal disutility of . At the median potential yield damage (i.e. median pest - free yield expectation of 57 bushels per acr e multiplied by the median pest damage rate of 20%) , the mitigating effect of social capital exceeds the negative impact on the marginal utility from expected damage . This comparison suggest s that social capital ultimately has a mitigating effect on the marginal disutility from the probability of delay for the majority of farmers. In Figure 3.3, panel B, there is a clear downward correlation between 130 respondent WTP for reductions in the probability of delay and the number of farmers the respondent is clos e with, further supporting H7. Therefore, our results on the whole provide some support for H 7. That farmers with more close relationships with other farmers are more likely to custom hire is in line with past findings that U.S. farmers rely on informal social mechanisms to facilitate transactions that may otherwise not occur (Allen & Lueck, 1992b, 2004; Wilson, 2000) . While previous findings regarding the role of social capital in facilitating agricultural contracting focus on land, our results provide some evidence that su ch mechanisms persist and in non - land contexts. Future research should seek to measure the relationships and social networks of both custom operators and the farmers who hire them in order to further explore how social capital might support custom hiring m arkets. sensitivity to timeliness costs, finding that farmers who farm more land are more sensitive to to delay probability separate from its impact on sensitivity to expected damage. There are multiple conflicting theoretical expectations as to the direction of the effect of farm size on sensitivity to uncertainty when custom hiring. Larger farms represent more possible revenue for a custom operator, who are often paid by the acre. If a custom operator sprays late, a farmer can withhold future business and the custom operator would lose out on more revenue if that farm is large. Therefore, we might expect l arger farms to underweight probabilities of delay in a similar way that those with more social capital do, suggesting a positive expectation for this coefficient. On the other hand, a larger farm requires more time to complete field operations. This might lead to increased sensitivity to the probability of delay, because if delays were to occur, then custom operators 131 would struggle more to catch up. Our null finding does not rule out either possibility. If both mechanisms mutually exist, their effects could counteract each other. Future research might examine the connection between farm size, transaction costs, and custom hiring decisions more closely. A common critique of stated preference studies, including choice experiments, is that good in question (Arrow et al., 1993; Lew & Wallmo, 2011) . Our finding that larger farms are more sensitive to greater damage is evidence that farmers are sensitive to scope. The alternative specific constants ( - DU indicate a distinct ranking among alternative providers of custom services (Table 3.5) . The alternative specific constant is largest for input dealers, f ollowed by co - op providers , and then other farmers. All three custom alternatives have larger constants than the constant for the alternative in which farmers spray on their own, indicating that farmers would prefer to have a custom operator spray if there were no custom fees and timely service were guaranteed. Input dealers, and co - ops as well, often offer additional services beyond custom spraying . These includ e agricultural inputs like fertilizers, seed, and pesticides and other custom services such as t he application of fertilizers or harvest . These additional services may provide a possible - ops over similar service from other farmers . Farmer customers of input dealers and co - ops can purchase other goods and services from these custom operators when hiring for custom spraying services (or vice versa), reducing search costs. Another possible explanation for the ordering of preferences between custom service providers is t hat other farmers have their own fields which may require treatment during the same critical window as their customers. Custom operators who 132 manage their own farm s have a clear incentive to prioritize their own fields over their , which may expla in the preference among respondents for non - farmer providers of custom spraying . Conclusion This paper provides a model for how farmers choose between custom hiring and spraying themselves when their fields are threatened by insect pests and examines this model empirically with data from a discrete choice experiment. Although the sample size is not large enough for extrapolation to a farmer population, the results illustrate how timeliness can be an important driver in these decisions, as evid enced by strong marginal disutility from increased probabilities of delay even beyond the effects such increased probability of delay on increased probability of pest damage. - equipment in non - stochastic environments (Ford & Musser, 1994) . However, to our knowledge no previous study has empirically tested transaction cost theories for their ability to explain the common phenomenon of custom hiring in agriculture. This paper begins to fill that gap by proposing a theoretical model of custom pest control hiri ng decisions and providing results from a pilot study to examine the empirical basis for hypotheses regarding the drivers of the custom hiring decision. Using a discrete choice model of custom hiring in a pest control setting, we illustrate how uncertainty over the reliability of custom operators can create timeliness costs, a specific type of transaction cost that is especially relevant in agricultural contexts (Allen & Lueck, 2004) . While previous studies have identifi ed timeliness costs in custom hiring through case study methods 133 (Allen & Lueck, 2004) , we provide indicative evidence that timeliness costs can drive farmers away from custom hiring and towards ownership of equipment. F urther, we illustrate that risk - averse farmers might be more sensitive to these costs. Finally, we illustrate that farmers who are more integrated into the agricultural community (i.e. who have more close friends who farm) are less sensitive to timeliness costs. Understanding which farmers opt to custom hire, and which custom operators they choose when multiple providers are available, can assist in identifying regions where demand for custom services may be high. Pest pressure dynamics, weather patterns, p esticide spraying regulations, and road infrastructure are all regional factors that can affect the ability of firms to provide timely services. For regions threatened by pests that affect many nearby fields concurrently, many farmers are likely to need to apply insecticides at the same time. In such setting, farmers are likely to find custom pest control unattractive, as custom operators will be harder pressed to provide timely services to many farmers concurrently. Areas with highly variable weather are m ore likely to have unexpected delays in spraying for a given field, which exacerbated in states with stricter restrictions on weather conditions suitable for pesticid e applications or regions where road conditions make moving between fields more difficult. Theoretical models and empirical approaches building on this first attempt to characterize the drivers of custom hiring can be applied to scenarios where far mers custom hire for services other than pest control, such as harvest or fertilizer application. These field operations are subject to other forms of ecological uncertainty which may induce timeliness costs in unique ways. Farms across the country make de cisions regarding custom hiring every year for a variety of field operations, providing a rich context to test transaction cost theories and examine what 134 conditions lead to various distributions of property rights between farms and the operators they emplo y. Such studies would provide additional insight into how farmers view timeliness and other transaction costs in the context of different field operations, providing valuable information for custom operators while also building on the broader transaction c osts economics literature. 135 APPENDICIES 136 Appendix A: Online Survey Instrument The following pages contain images of the survey instrument. The survey was multip le versions. Notes have been added to explain where reactive elements exist and to clarify structure. Choice experiment blocks were removed for brevity. Unless otherwise noted, each panel represents a separate page of the survey instrument. 137 Figure 3B.1. Screenshots of Online Survey Instrument. 138 139 140 141 142 [PAGE CONTINUES] 143 Figure 3B.1 [CONTINUED FROM PREVIOUS PAGE] 144 [FOLLOWED BY 8 CHOICE EXPERIMENT QUESTIONS] 145 [OPERATOR CLASS] Page repeats with each class of custom operator for which the 146 147 148 149 Appendix B: Survey Deployment and Sample Representativeness In this appendix, we assess the representativeness of the sample relative to the target population. Emails were purchased from the agricultural marketing data company FarmMarketID for 9,290 Illinois farmers, 4,847 Indiana farmers, and 1,895 Michigan farmers , representing a ll available records with valid emails and 100 or more planted acres of soybeans in 2017 . Farme rs were emailed three times over the course of a week in August and September of 2018. Emails included a link to an online survey hosted by Qualtrics. To further encourage response , a single letter was mailed to all Michigan farmers and 1,234 of the India na farmers directing farmers to visit the online survey . Of the 16,032 email addresses contacted, 388 were immediately returned as undeliverable due to spam blocking software on the receiving end. These returned emails only represent addresses employing sp am blocking software that reports failure of delivery to the sender. We ders where they were unlikely to be read. This hypothesis is supported by considerably larger response rates in Michigan and Illinois where an additional mail contact was employed. Choice experiments targeting farmers are frequently limited by small sample sizes (Chèze, David, & Martinet, 2020) , including the one presented in this paper. The challenge of obtaining large farmer samples for choice experiment surveys is made more difficult by declining trends in farmer respons e rates (Johan sson, Effland, & Coble, 2017) . Further, farmers 150 are often sensitive about discussing pesticide applications due to public concerns over their public environmental and health effects (Chèze et al., 2020) . Direct data on the population of farmer s with 100 or more acres of soybeans planted is not publicly available to the best of our knowledge, so we compare demographic characteristics of our sample to results from the 2017 Census of Agriculture over each state, which includes farms smaller than 1 00 acres and farms that do not grow soybeans (USDA National Agricultural Statistics Service, 2019) . Michigan farmers, and to a lesser extent Indiana farmers, are overrepresented in the sample relative to the population of soybean farmers in the three states (51% of the sample versus 23% of total farms of 100 acres or more for Michigan, 34% of the sample versus 28% of total farms of 100 acres or more for Indiana). This is likely the result of issues with e mail delivery, as Michigan and a subset of Indiana farmers received additional mail invitations to participate in the survey. As a result, the following results should be interpreted as representing mainly the preferences of Michigan and Indiana growers. T acre. This value is considerably higher than the mean yields reported by the Census of Agriculture for Michigan (42 bushels per acre) and Indiana (53 bushels per acre), sug gesting that respondents are either more productive than the population or hold optimistic expectations. Our sampling frame was limited to farms of over 100 planted acres of soybeans while the Census of Agriculture reports yields for all growers of soybean s, including the smallest farms which typically do not operate at a commercial scale and therefore often have lower yields. Such smaller farms make up a large portion of the total population targeted by the Census of Agriculture, which may also explain the discrepancy between the mean yield reported by respondents and the mean yield reported by the Census. 151 Respondents reported planting between 150 and 3,700 total acres across all crops, with a mean of 997 acres planted and median of 697 acres planted. The m edian planted acres for respondents is considerably higher than the median planted acres for all growers with over 100 acres reported by the Census of Agriculture, which lies in the 220 - 259 acres range for all three states. The average age of respondents w as 58, which is slightly older than the average age reported in the Census of Agriculture for farmers in Michigan and Indiana (56.6 and 55.5 respectively), but roughly equal to the average age of farmers in Illinois (58). Overall, our sample represents lar ger and more productive farms than those in the population. 152 REFERENCES 153 REFERENCES Allen, D. W., & Lueck, D. (1992a). Contract Choice in Modern Agriculture: Cash Rent versus Cropshare. The Journal of Law & Economics , 35 (2), 397 426. Reputation, and the S tructure of Farmland Contracts. Journal of Law, Economics, & Organization , 8 (2), 366 376. Retrieved from JSTOR. Allen, D. W., & Lueck, D. (1993). Transaction Costs and the Design of Cropshare Contracts. The RAND Journal of Economics , 24 (1), 78 100. https:/ /doi.org/10.2307/2555954 Allen, D. W., & Lueck, D. (2004). The Nature of the Farm: Contracts, Risk, and Organization in Agriculture . Cambridge: MIT press. Apland, J. (1993). The Use of Field Days in Economic Models of Crop Farms. Journal of Production Agri culture , 6 (3), 437 444. https://doi.org/10.2134/jpa1993.0437 Arrow, K., Solow, R., Portney, P., Leamer, E., Radner, R., & Schuman, H. (1993). Report of the NOAA panel on Contingent Valuation (Vol. 58). Beauchamp, J. P., Cesarini, D., & Johannesson, M. (201 7). The psychometric and empirical properties of measures of risk preferences. Journal of Risk and Uncertainty , 54 (3), 203 237. https://doi.org/10.1007/s11166 - 017 - 9261 - 3 Bhattacharya, R., Devinney, T. M., & Pillutla, M. M. (1998). A Formal Model of Trust B ased on Outcomes. The Academy of Management Review , 23 (3), 459 472. https://doi.org/10.2307/259289 Burghart, D. R., Cameron, T. A., & Gerdes, G. R. (2007). Valuing publicly sponsored research projects: Risks, scenario adjustments, and inattention. Journal of Risk and Uncertainty , 35 (1), 77 105. https://doi.org/10.1007/s11166 - 007 - 9018 - 5 Burnham, K. P., & Anderson, D. R. (2002). Model Selection and Multimodel Inference: A Practical Information - Theoretic Approach (2nd ed.). New York: Springer. Chèze, B., David pesticide use: A choice experiment. Ecological Economics , 167 , 106349. https://doi.org/10.1016/j.ecolecon.2019.06.004 Coase, R. H. (1937). The Nature of the Firm. Economica , 4 (16), 386 405. https://doi.org/10.1111/j.1468 - 0335.1937.tb00002.x Davis, C. G., & Gillespie, J. M. (2007). Factors Affecting the Selection of Business Arrangements by U.S. Hog Farmers. Review of Agri cultural Economics , 29 (2), 331 348. Retrieved from JSTOR. 154 Dillman, D. A., Smyth, J. D., & Christian, L. M. (2014). Internet, Phone, Mail, and Mixed - Mode Surveys: The Tailored Design Method . Retrieved from http://ebookcentral.proquest.com/lib/michstate - eboo ks/detail.action?docID=1762797 Dorward, A. (2001). The Effects of Transaction Costs, Power and Risk on Contractual Arrangements: A Conceptual Framework for Quantitative Analysis. Journal of Agricultural Economics , 52 (2), 59 73. https://doi.org/10.1111/j.14 77 - 9552.2001.tb00925.x Duncan, G. J., & Kalton, G. (1987). Issues of Design and Analysis of Surveys across Time. International Statistical Review / Revue Internationale de Statistique , 55 (1), 97 117. https://doi.org/10.2307/1403273 Faccioli, M., Kuhfuss, L ., & Czajkowski, M. (2019). Stated Preferences for Conservation Environmental Domain. Environmental and Resource Economics , 73 (2), 627 659. https://doi.org/10.1007/s10 640 - 018 - 0276 - 2 Ford, S. A., & Musser, W. N. (1994). The Lease - Purchase Decision for Agricultural Assets. American Journal of Agricultural Economics , 76 (2), 277 285. https://doi.org/10.2307/1243629 Franken, J. R. V., Pennings, J. M. E., & Garcia, P. (2009). Do Transaction Costs and Risk Preferences Influence Marketing Arrangements in the Illinois Hog Industry? Journal of Agricultural and Resource Economics , 34 (2), 297 315. Glenk, K., & Colombo, S. (2011). How Sure Can You Be? A Framework for Considering Delivery Uncertainty in Benefit Assessments Based on Stated Preference Methods. Journal of Agricultural Economics , 62 (1), 25 46. https://doi.org/10.1111/j.1477 - 9552.2010.0027 8.x Glenk, K., & Colombo, S. (2013). Modelling outcome - related risk in choice experiments. Australian Journal of Agricultural and Resource Economics , 57 (4), 559 578. https://doi.org/10.1111/1467 - 8489.12012 Gneezy, U., List, J. A., & Wu, G. (2006). The Unce rtainty Effect: When a Risky Prospect is Valued Less than its Worst Possible Outcome. The Quarterly Journal of Economics , 121 (4), 1283 1309. https://doi.org/10.1093/qje/121.4.1283 Hensher, D. A., Greene, W. H., & Li, Z. (2011). Embedding risk attitude and decision weights in non - linear logit to accommodate time variability in the value of expected travel time savings. Transportation Research Part B: Methodological , 45 (7), 954 972. https://doi.org/10.1016/j.trb.2011.05.023 Hensher, D. A., Rose, J. M., & Gree ne, W. H. (2015). Applied Choice Analysis (2 edition). Cambridge: Cambridge University Press. 155 Hobbs, J. E. (1997). Measuring the Importance of Transaction Costs in Cattle Marketing. American Journal of Agricultural Economics , 79 (4), 1083 1095. https://doi. org/10.2307/1244266 Hudson, D., & Lusk, J. (2004). Risk and Transactions Cost in Contracting: Results from a Choice - Based Experiment. Journal of Agricultural & Food Industrial Organization , 2 (1). https://doi.org/10.2202/1542 - 0485.1046 Hueth, B., Ligon, E., Wolf, S., & Wu, S. (1999). Incentive Instruments in Fruit and Vegetable Contracts: Input Control, Monitoring, Measuring, and Price Risk. Applied Economic Perspectives and Policy , 21 (2), 374 389. https://doi.org/10.2307/1349886 Hurvich, C. M., & Tsai, C. - L. (1989). Regression and time series model selection in small samples. Biometrika , 76 (2), 297 307. https://doi.org/10.1093/biomet/76.2.297 Johansson, R., Effland, A., & Coble, K. (2017, January 19 ). Falling Response Rates to USDA Crop Surveys: Why It Matters. Retrieved December 5, 2019, from Farmdoc daily website: https://farmdocdaily.illinois.edu/2017/01/falling - response - rates - to - usda - crop - surveys.html ., Difonzo, C. D., Swinton, S. M., Dixon, P. M., - Effective Management of Soybean Aphid (Hemiptera: Aphididae) in North America. Journal of Economic Entomology , 102 (6), 2101 2108. https://doi.org/10.1603/029.1 02.0613 Klein, B., Crawford, R. G., & Alchian, A. A. (1978). Vertical Integration, Appropriable Rents, and the Competitive Contracting Process. The Journal of Law & Economics , 21 (2), 297 326. Klein, P. G. (2005). The Make - or - Buy Decision: Lessons from Empi rical Studies. In Handbook of New Institutional Economics (pp. 435 464). Retrieved from https://mospace.umsystem.edu/xmlui/handle/10355/139 Lew, D. K., & Wallmo, K. (2011). External Tests of Scope and Embedding in Stated Preference Choice Experiments: An A pplication to Endangered Species Valuation. Environmental and Resource Economics , 48 (1), 1 23. https://doi.org/10.1007/s10640 - 010 - 9394 - 1 Li, B., & Hensher, D. A. (2017). Risky weighting in discrete choice. Transportation Research Part B: Methodological , 10 2 , 1 21. https://doi.org/10.1016/j.trb.2017.04.014 Li, H., Tu, H., & Hensher, D. A. (2016). Integrating the mean variance and scheduling approaches to allow for schedule delay and trip time variability under uncertainty. Transportation Research Part A: Pol icy and Practice , 89 , 151 163. https://doi.org/10.1016/j.tra.2016.05.014 Li, Z. (2018). Unobserved and observed heterogeneity in risk attitudes: Implications for valuing travel time savings and travel time variability. Transportation Research Part E: Logis tics and Transportation Review , 112 , 12 18. https://doi.org/10.1016/j.tre.2018.02.003 156 Li, Z., & Hensher, D. A. (2013). Behavioural implications of preferences, risk attitudes and beliefs in modelling risky travel choice with travel time variability. Transp ortation , 40 (3), 505 523. https://doi.org/10.1007/s11116 - 012 - 9445 - 2 Li, Z., Hensher, D. A., & Rose, J. M. (2010). Willingness to pay for travel time reliability in passenger transport: A review and some new empirical evidence. Transportation Research Part E: Logistics and Transportation Review , 46 (3), 384 403. https://doi.org/10.1016/j.tre.2009.12.005 List, J. A. (2001). Do Explicit Warnings Eliminate the Hypothetical Bias in Elicitation Procedures? Evidence from Field Auctions for Sportscards. American Eco nomic Review , 91 (5), 1498 1507. https://doi.org/10.1257/aer.91.5.1498 Lundhede, T., Jacobsen, J. B., Hanley, N., Strange, N., & Thorsen, B. J. (2015). Incorporating Outcome Uncertainty and Prior Outcome Beliefs in Stated Preferences. Land Economics , 91 (2), 296 316. https://doi.org/10.3368/le.91.2.296 Lusk, J. L. (2003). Effects of Cheap Talk on Consumer Willingness - to - Pay for Golden Rice. American Journal of Agricultural Economics , 85 (4), 840 856. Retrieved from JSTOR. Makriyannis, C., Johnston, R. J., & Wh elchel, A. W. (2018). Are Choice Experiment Treatments of Outcome Uncertainty Sufficient? An Application to Climate Risk Reductions. Agricultural and Resource Economics Review , 47 (3), 419 451. https://doi.org/10.1017/age.2017.27 McFadden, D. (1973). Condit ional logit analysis of qualitative choice behavior. In P. Zarembka (Ed.), Frontiers in Econometrics (pp. 105 142). New York: Academic Press. Ragsdale, D W, Mccornack, B. P., Venette, R. C., Potter, B. D., Macrae, I. V., Hodgson, E. W., 18). Economic Threshold for Soybean Aphid (Hemiptera: Aphididae). JOURNAL OF ECONOMIC ENTOMOLOGY , 100 (4), 10. America. ANNALS OF THE ENTOMOLOGICAL SOCIETY OF AMERICA , 97 (2), 5. Rigby, D., Alcon, F., & Burton, M. (2010). Supply uncertainty and the economic value of irrigation water. European Review of Agricultural Economics , 37 (1), 97 117. https://doi.org/10.1093/erae/jbq001 Roberts, D. C., Boyer, T. A., & Lusk, J. L. (2008). Preferences for environmental quality under uncertainty. Ecological Economics , 66 (4), 584 593. https://doi.org/10.1016/j.ecolecon.2008.05.010 Rolfe, J., & Windle, J. (2015). Do Respondents Adjust Their Expected Utility in the Presence of an Outcome Certainty Attribute in a Choice Experiment? Environmental and Resource Economics , 60 (1), 125 142. https://doi.org/10.1007/s10640 - 014 - 9760 - 5 157 Royer, A. (2011). Transaction costs in milk marketing: A comparison between Canada and Great Britain. Agricultural Economics , 42 (2), 171 182. https://doi.org/10.1111/j.1574 - 0862.2010.00506.x Schmid, A. A., & Robison, L. J. (1995). Applications of Social Capital The ory. Journal of Agricultural and Applied Economics , 27 (1), 59 66. Shelanski, H. A., & Klein, P. G. (1995). Empirical Research in Transaction Cost Economics: A Review and Assessment. The Journal of Law, Economics, and Organization , 11 (2), 335 361. https://doi.org/10.1093/oxfordjournals.jleo.a036875 Short, C., & Gitu, K. W. (1991). Timeliness Costs for Machinery Selection. Canadian Journal of Agricultural Economics , 39 (3), 457 462. https://doi.org/10.1111/j.1744 - 7976.1991.tb03586.x Silva, A., Nayga, R. M., Campbell, B. L., & Park, J. L. (2011). Revisiting Cheap Talk with New Evidence from a Field Experiment. Journal of Agricultural and Resource Economics , 36 (2), 280 291. Retrieved from JSTOR. Tonsor, G. T., & Shupp, R. S. (2011). Cheap Talk Scripts an d Online Choice Experiments: American Journal of Agricultural Economics , 93 (4), 1015 1031. https://doi.org/10.1093/ajae/aar036 Tversky, A., & Kahneman, D. (1992). Advances in Prospect Theory: Cumulative Representation of Uncertai nty. Journal of Risk and Uncertainty , 5 (4), 297 323. Retrieved from JSTOR. USDA NASS. (2014). 2012 Census of Agriculture, State Data . Retrieved from https://www.nass.usda.gov/Publications/AgCensus/2012/Full_Report/Volume_1,_Chapte r_2_US_State_Level/st99_2_ 003_003.pdf USDA National Agricultural Statistics Service. (2019). NASS - Quick Stats . Retrieved from USDA National Agricultural Statistics Service website: https://data.nal.usda.gov/dataset/nass - quick - stats Von Neumann, J., & Morgenstern, O. (1944). Theor y of Games and Economic Behaviour . Princeton: Princeton University Press. Williamson, O. E. (1979). Transaction - Cost Economics: The Governance of Contractual Relations. The Journal of Law & Economics , 22 (2), 233 261. Williamson, O. E. (1993). Calculativene ss, Trust, and Economic Organization. The Journal of Law & Economics , 36 (1), 453 486. Wilson, P. N. (2000). Social Capital, Trust, and the Agribusiness of Economics. Journal of Agricultural and Resource Economics , 25 (1), 13.