RISK PREFERENCES AND PESTICIDE USE BY COTTON FARMERS IN CHINA. JiKun Huang Center for Chinese Agricultural Policy

Transcription

1 RISK PREFERENCES AND PESTICIDE USE BY COTTON FARMERS IN CHINA JiKun Huang Center for Chinese Agricultural Policy Elaine M. Liu * University of Houston September, 2009 Abstract Insect resistant Bt cotton has been lauded for bringing about a reduction in pesticide use. It is a cost saving measure that comes with substantial health benefits. However, several studies have shown that Chinese Bt cotton farmers continue to use excessive amount of pesticide beyond the profit-maximizing optimal level. Using a survey and field experiment, we elicit the risk preferences of 320 Chinese cotton farmers. We find that more risk-averse farmers spray more pesticide. Farmers who are more loss averse use less pesticide. Our findings suggest that when farmers decide on how much pesticide to spray, profit maximization is not their only goal. They do take into account of their own health condition and the possibility of pesticide poisoning. Keywords: Risk Preferences, Prospect Theory, Pesticide Use * corresponding author emliu@uh.edu The authors are indebted to Alan Krueger, Orley Ashenfelter, Anne Case, Angus Deaton, JiKun Huang, Chris Paxson, Carl Pray, Molly Fifer, Analia Scholosser, Tomomi Tanaka, Stephanie Wang and Nate Wilcox for helpful discussions, and seminar and conference participants at Princeton University, Rutgers University, University of Houston, National University of Singapore, National Taiwan University, North America ESA for their suggestions. Special thanks to Raifa Hu, Zijun Wang, Liang Qi, YunWei Cui and other research staff at the CCAP for their help. Financial support from Princeton University Industrial Relations Section is gratefully acknowledged. All errors are our own.

2 1. Introduction Modern agricultural biotechnology has made much progress over the past two decades, enhancing its potential to greatly increase productivity and living standards in developing countries. These advancements include a wide array of genetically modified crops that are insect-resistant, virus-resistant, drought-resistant, and even nutrient-enriched. For example, prior to the invention of Bt cotton, farmers were forced to choose between letting cotton bollworms (the main pest for cotton) erode away their cotton yields or sacrificing their own health by spraying large amounts of pesticide. Bt cotton was devised specifically to counter bollworm infestations and it has been scientifically proven to be effective at pestresistance. Emboldened by this scientific evidence, many policy makers have encouraged the adoption of Bt cotton. However, farmers do not always follow the recommendations given by scientists. One study by Liu (2008) finds that some Chinese cotton farmers waited ten years after the introduction of Bt cotton to switch to this new technology; and in particular, she finds that farmers who are more risk averse switch to Bt cotton later. In addition, several studies find that farmers continued to use excessive amounts of pesticides even after they adopted pest-resistant Bt cotton (Huang et al., 2002b; Pemsl et al., 2005; Wang, 2008; Yang et al., 2005). These findings present a puzzle as to why farmers would deviate from profit maximization, especially considering the fact that spraying pesticide is detrimental to their health. Motivated by that question, this paper investigates the determinants of pesticide use. Several hypotheses seek to explain the overuse of pesticides by Chinese Bt cotton farmers. These include: misinformation from the agricultural extension services, lack of knowledge about the purpose of adopting Bt cotton, substantial uncertainty about the quality and effectiveness of production inputs, or the rise of secondary pests. One important factor that has not come into the discussion so far is the influence of farmers own risk preferences. There is a vibrant amount of literature quantifying the impact of risk preference on agricultural production decisions. Most of these studies rely on the assumption of an objective function and use advanced econometric techniques to impute the coefficient of risk aversion that will fit the model (Saha, 1

3 Shumway and Talpaz, 1994; Chavas and Holt, 1996; Antle, 1980). As suggested by Just and Lybbert (2009), the assumption of utility function and much arbitrary heuristics could cause bias when estimating individual risk aversion. This paper takes a different approach in estimating farmers risk preference. We employ a technique from experimental economics modeled after Tanaka, Camerer and Nguyen (TCN) (2009) to elicit farmers risk preferences. The experiment design elicits risk preference parameters beyond expected utility theory. It incorporates prospect theory components, including loss aversion and nonlinear probability weighting (Kahneman and Tversky, 1979).The major advantage of the TCN design is that it allows us to empirically test whether expected utility theory or prospect theory fits our data better. Along with the field experiments, three hundred-twenty Chinese cotton farmers from 16 villages across 8 counties in 4 provinces were also surveyed. We collect household characteristics, individual characteristics as well as detailed plot information from their harvest and planting in spring and summer of We relate farmers elicited risk preferences to their pesticide use, controlling for their background characteristics. Our main findings are that farmers who are more risk-averse use more pesticides. If the average farmer from this sample becomes risk neutral, he would spray approximately 13% less pesticide (It is equivalent to the effect of 6 more years of education). 1 We also find that farmers who are more lossaverse use less pesticide. It may seem surprising at first glance. However, this is consistent with farmers placing more weight on the importance of their health over the importance of money in the loss domain. We find that more educated farmers seem to better understand the advantages of using Bt cotton since they use less pesticide. For every additional year of education, farmer would reduce pesticide use by 0.56 Kg/Ha (~2%). Attendance of a training session is associated with a dramatic reduction of pesticides used, but over time this effect wanes. Lastly, in terms of the pest-resistant quality of Bt cotton seeds, we find that the pricier seeds are not significantly more pest resistant than the cheaper seeds, and the source of seeds do not make any difference. 1 The average farmer in the sample is risk averse with coefficient of risk aversion equals to

4 This paper proceeds as follows: Section 2 provides background on Bt cotton. Section 3 describes the data set and provides some summary statistics on farmers characteristics. Section 4 describes the game. Section 5 presents our findings on the determinants of pesticide use. Section 6 concludes. 2. Background China has been one of the largest cotton producers in the world. Unlike commercial cotton farmers in the US, Chinese cotton farmers are generally subsistence farmers, who are more risk-averse, less tolerant of pest infestation, and place the highest priority on solving severe pest problems (Bentley and Thiele, 1999; Pray et al., 2002). During the early 1990s, many Chinese cotton farmers experienced failures in controlling bollworm due to continual outbreaks of increasingly pesticide-resistant bollworm infestations. In an attempt to ameliorate the bollworm problem, the provincial government in some parts of China began commercializing Bt cotton seeds in Bt cotton seeds are planted in a similar fashion to traditional cotton seeds, but Bt cotton seeds carry the Bacillus thuringiensis (Bt) toxin that targets cotton bollworm. Using data collected in 2001, Huang et al. (2002a) find that Bt cotton adoption leads to a significant decrease in the use of pesticides. Bt cotton farmers reduce their total pesticide expenditure by 82 percent. In 2006, Chinese scientists tested bollworm pests with Bt cotton and conclude that bollworms found in China s cotton fields have not yet become resistant to Bt cotton (Wu, 2007). While the picture may seem rosy, some problems still exist. First, Chinese cotton farmers are well known for using excessive amounts of highly hazardous pesticides, and this practice has continued even after the adoption of Bt cotton (Huang et al., 2002b; Pemsl et al., 2005; Yang et al., 2005). Huang et al. (2002b) find that Bt cotton farmers applied 11.8 kg per hectare when the optimal pesticide use ranges from kg per hectare. 3 Pemsl et al (2005) find that the optimal input pesticide level was about 5 kg per hectare in 2004, but Chinese farmers applied, on average, 14 kg per hectare. The problem of 2 It was a rolling decision. In some provinces, Bt cotton was approved in kg/ha is based on the estimation with Cobb-Douglas production function. 4.2 kg/ha is based on the estimation with a Weibull damage control function. 3

5 pesticide overuse is further exacerbated by the fact that nearly 40 percent of the pesticides used by Chinese cotton farmers contain active ingredients that are classified as extremely or highly hazardous (Classes 1a or 1b) by the World Health Organization (WHO). 4 There are estimates that some 400 to 500 Chinese cotton farmers die every year from pesticide poisoning (Conko and Parkash, 2004). The discussion above raises the question of why farmers would spray excessive amounts of pesticide if Bt cotton has proven its resistance to the cotton bollworm and if farmers know that spraying pesticide is detrimental to their health? Their uncertainty about the quality of Bt cotton seeds could be a major reason. Existing studies have found that the quality of Bt cotton seeds vary dramatically. Pemsl (2006) collects leaves from cotton farmers in Shandong and found that some of the so-called Bt leaves do not contain the Bt trait. Due to the high demand for Bt cotton seeds, it is not surprising that the unscrupulous are trying to exploit this situation for profit through various nefarious means. Therefore, some lower-quality seeds have permeated the market via different channels. For example, some firms and local research institutions release Bt cotton seeds into the market before obtaining government approval (Yang et al., 2005). Farmers also reproduce the trademarked Bt cotton seeds via on-farm propagation and these self-propagated seeds are of lower quality (Pray et al., 2002). Some seed companies simply repackage their conventional cotton seeds to sell as authentic Bt cotton seeds (LouYang Agricultural News, 2003). To sum up, there were an estimated 140 genetically modified cotton varieties available in 2004, so it could be difficult for farmers to know which seeds are effective Bt cotton seeds a priori (Pray et al., 2006). According to a conversation with a Monsanto representative, seeds produced by Delta & Pineland, Monsanto s joint venture in China, are guaranteed to be 99 percent pure, meaning that 99 out of 100 seeds contain the Bt trait, while seeds from other brands vary greatly in terms of quality (Zhu, 2007). One other reason why farmers could be using excessive amounts of pesticide is proposed by Wang, Just, and Pinstrup-Anderson (2006). They suggest that the population of secondary pests, mainly mirids, has been slowly rising. They examine pesticide use by Chinese cotton farmers using survey data 4 WHO classifies insecticides into 4 classes: Class 1a is extremely hazardous (highest toxicity). Class 1b is highly hazardous. Class 2 is moderately hazardous. Class 3 is slightly hazardous. 4

6 from 2004, and they find that Bt cotton farmers use less pesticide on bollworm, but Bt cotton farmers spray more pesticides in order to target mirids. While it seems reasonable, it is worth noting that their findings primarily come from tabulation of Bt cotton versus non-bt cotton plots without controlling for individual or farm characteristics. 3. Data The Bt cotton survey was designed and collected by the Center for Chinese Agricultural Policy (CCAP), a government-affiliated research agency, in the winter of Four provinces (Shandong, Hebei, Henan and Anhui) with high Bt cotton adoption rates and similar cotton growing seasons (April- October) were selected. CCAP selected 2 counties per province, 2 villages per county, and interviewed twenty households in each village. The CCAP team compensated each participating household 10 Yuan for completing the survey (equivalent of 1/3 of daily wage). We interviewed the head of household or whoever is most responsible for farming. In addition, we obtained detailed information on inputs and outputs used in each cotton plot, perceived pest infestation, incidence of pesticide poisoning, and past training experience. Most farmers own are responsible for multiple plots, as arable land for farming is assigned by the government. Out of the 945 cotton plots belonging to 320 farmers, 930 of them are used to grow Bt cotton, and the remaining 15 are used to grow conventional cotton. 5 In Table 1, the summary statistics at the household level are presented. The average interviewee is about 50 years old and has completed 7 years of education. In China the land is government owned, and it is assigned to every household in the village. The average household in the sample is assigned 0.59 hectares of farmland. In this region, cotton is the major cash crop, which is planted on 0.54 hectares of farmland per household and farmers usually practice rotational cropping with wheat as the primary grain crop (0.33 hectares). Our sample spent most of their time on the farm, and when production on farm stops, they do limited amounts of off-farm work. Ownership of a set of durable goods is used as a proxy 5 In winter 2007, the CCAP collected a subset of cotton seeds from a 2006 sample. When researchers tested the seeds in the lab, they found that farmers often misreported the Bt versus non-bt status of the seeds. Some so-called non-bt seeds in fact contained the Bt gene, and vice versa. 5

7 for wealth in Table 2 presents the summary statistics at the plot level breaking down by Bt versus non Bt cotton. Bt cotton is more expensive than traditional cotton, but farmers who grow Bt cotton spray less pesticide and have higher yield. In this paper, wealth is proxied by the price of durable goods owned per capita. Bt cotton farmers are statistically wealthier than traditional cotton farmers. However, without a baseline survey prior to adoption, we cannot conclude any causality on whether the differential wealth accumulation is related to their decisions to plant Bt cotton. Since not all pesticides are the same, we ideally wish to have a better measure of pesticide use on its level of effectiveness and we need the severity of pest infestations in the region. Then we can use a production function to estimate whether farmers overuse specific pesticides. To do so, we would need information on the active ingredients of pesticides. However, China s pesticide market is extremely fragmented (PWC, 2004), in our survey alone, we find more than 50 brands/formulations. Many farmers purchase the pesticides which are blends of various brands of pesticides. It is also difficult to classify these pesticides by the WHO s measures, since most pesticide bottle labels are missing the active ingredients. Therefore, in our analysis, we will only use the total amount and total cost of pesticides. We also collect information on pesticide poisoning. Farmers were asked if anyone in their household had experienced any health impairments after mixing and spraying pesticides in the past 10 years. Thirty percent (129) of farmers reported that they experienced at least one symptom since 1996, where the most frequently reported symptoms were headaches or dizziness (53 percent), skin irritation (26 percent), restlessness (16 percent) and vomiting (59 percent). We also asked the farmers to report any costs associated with this health impairment. Forty-one farmers reported that they were hospitalized due to pesticide poisoning, and the average medical cost and imputed labor cost was 139 Yuan (approximately five days of wages). 4. Experimental Design A key objective for this paper is to elicit individual risk preference. We conducted a field 6

8 experiment modeled after the one designed by Tanaka, Camerer, and Nguyen (2009). 6 TCN s experiment is similar to Holt and Laury s design economics offering subjects a series of pair-wise lotteries of both risky and safe options, but it elicits three parameters concerning risk preferences risk aversion, loss aversion, and nonlinear probability weighting. Following TCN, we assume a utility function of the following form: ( y) ( p)( ( x) ( y)) U ( x, p; y, q) ( p) ( x) ( q) ( y) x y 0 or x 0 y x y 0 1 x where (x) ( x) 1 for x 0 for x 0 and ( p) exp[ ( ln p) ] σ is the standard measure of risk aversion. A higher sigma indicates a higher degree of risk aversion. λ measures the sensitivity to loss versus gain. ω(p) is the probability weighting function adapting from Prelec (1998). If α< 1, ω(p) has an inverted S-shape. Individuals overweight small probability and underweight large probability. This complicated utility function form gives us flexibility in allowing for the rejection of prospect theory. If α=1 and λ=1, then utility function reform would reduce to the standard expected utility function. Farmers were asked to participate in an experiment after the conclusion of the interview. 7 They were told that they will earn additional real payoff depending on the outcome of the game. The mathematical version of the experiment is presented in Table 3. In reality, we do not use the word probability with our subjects. More specifically, each line in Table 3 is presented to the subject in the following format: 6 The more detailed description of the experiment, analysis and imputation of risk preference parameter can be found in Liu (2008). 7 More details about the game can be found in Liu (2008). 7

9 They understand that there is a bag of 10 numbered balls and depending on whether they have chosen lottery A or lottery B in each line, the numbered ball they draw randomly will determine the payoff. They were presented 35 questions separated in 3 series (see Appendix for answer sheets used in the game). They were asked at which line, from Line 1 to Line 14, they would switch from lottery A to lottery B for each series. We choose Lottery A for Line 1 to. We choose Lottery B for Line to 14. Given that the safe options do not change and the risk option has increasing expected payoff as we move down the line within in Series 1 and Series 2, the more risk tolerant farmer would choose to switch to lottery B later. With the 3 switching points of farmers and the utility function form assumption, we can impute the risk preference parameters. The method of imputation can be found in TCN (2007) and Liu (2008). We reject the null hypothesis that λ = 1 and α =1 at the 1% level. The summary statistics of individual risk preference measures are provided in Table Econometric Framework and Regression Results for Pesticide Use 5.1 Basic Framework To begin, we first replicate Huang et al. s (2002a) results without the risk preference parameters, so we estimate the following equation by ordinary least square (OLS): y ijv o 1 Btijv 2 ( Bt exp) ijv X ijv ' v ijv (2) where i denotes individual, j denotes plot, and v denotes village. y ijv is the amount of pesticides in kg per hectare sprayed for individual i, on plot j, in village v; Bt ijv equals 1 if Bt cotton is planted and zero otherwise. (Btexp) ijv is the number of years that farmer i has planted Bt cotton interacted with Bt cotton term; X ijv is a vector of individual or plot characteristics, such as plot size, age, years of education; μ v is a village fixed effect. The main coefficients of interest are δ 1 and δ 2. δ 1 represents the effectiveness of Bt 8 For more detailed distribution and analysis of the risk preference parameters used in this paper see Liu (2008). 8

10 cotton in reducing pesticide use 10 years after its commercialization. The meaning of δ 2 is more complicated. There are two opposing factors that can affect this coefficient. First, if cotton bollworm builds up a resistance to Bt toxin, more pesticides would need to be used over time, and δ 2 should be positive. In contrast, if farmers become more aware of the benefits of using Bt cotton as they have more experience with planting Bt cotton, δ 2 should be negative. Column 1 of Table 4 shows the results from the estimation of Equation 2. We find that more educated farmers use significantly less pesticides. For every additional year of education, a farmer reduces pesticide use by 0.73 kg/ha, 2.7 percent of total use (of Bt cotton farmers). In other words, farmers who finish elementary school use 16 percent less pesticides compared to farmers with no education. The coefficient on plot size is negative and significant, which could be a sign of economies of scale. In addition, farmers use less pesticide when the price of pesticides is high. The main coefficient of interest (δ 1 ) indicates that the cultivation of Bt cotton reduces pesticide use dramatically. All else equal, the average Bt cotton farmer reduces the use of pesticides by 19.5 kg/ha, on his Bt plot compared with his non-bt plot. The coefficient on the interaction term (δ 2 ) is positive, but it is not statistically significant from zero, which may simply mean that the two factors mentioned above cancel each other out. Overall, we conclude that even a decade after the commercialization of Bt cotton, cotton farmers still use significantly less pesticides on their Bt plots than on their non-bt plots. In all regressions that follow, the standard errors are corrected for heteroskedasticity at the individual level. One very important covariate that is missing from the above estimation is pest severity. Since both pesticide use and the use of Bt cotton seeds are related to pest severity, excluding a measure of pest severity from the model could produce omitted variable bias. Unfortunately, in the survey, we do not have an objective measure of pest severity. However, we asked some questions about the perceived yield loss. The questions were phrased as follows: What do you think your potential yield loss will be if you do not spray any pesticide for controlling bollworm? (0-100 percent) What do you think your potential yield loss will be if you do not spray any 9

11 pesticide for controlling mirids 9? (0-100 percent) We use the answers to these questions as proxies for pest severity, where higher values indicate a more severe pest problem. 10 The regression results, including the pest severity proxies, are reported in Column 2 of Table 4. The positive coefficient on bollworm severity reflects the fact that the higher the perceived yield loss (more severe pest problem), the more pesticides farmers spray. The coefficients on mirids are insignificant and it is probably due to a multicollinearity problem. 11 One other result worth noting is that the magnitude of the coefficient on education is smaller, but it remains positive and significant at the 10 percent level. It implies that education is correlated with farmers perception of pest severity Risk Preference As mentioned before, the above specification, as estimated in Column 2 of Table 4, could be problematic since the pest severity variable is subjective. In particular, since agricultural production is full of uncertainty and risk, it is likely that individual risk preferences play an important role in pesticide use decisions. Typically, in decision making under uncertainty, we would use the neo-classical utility theory, where risk preference is solely characterized by the coefficient of risk aversion. As suggested in the other early agricultural economic research, if farmers follow safety-first principals by setting a target income and minimizing the probability of severe yield loss below that income (Moscardi and de Janvry, 1977; Young, 1979), then it is likely that farmers risk preferences will be best captured by prospect theory instead of neoclassical utility theory. 13 In particular, we have rejected the null hypotheses that λ=1 and α=1 from the experiment result. Therefore, we include the elicited measure of risk preferences in the rest 9 Reported by Wang et al. (2006) as the most serious secondary pest to Chinese Bt cotton farmers. 10 Same methodology is also used by Huang et al. (2002a). 11 Correlation between bollworms and mirids severity is In a separate regression not reported in this paper, when we regress the yield loss on levels of education, controlling for village fixed effects. The coefficient on education is negative and significant at the 1 percent level. Thus, higher levels of education are associated with lower perceptions of yield loss. 13 The applications of these two concepts, safety-first rule and loss aversion, can be found in many behavioral finance studies (Camerer and Kunreuther, 1989; Polkovnichenko, 2005; Campbell and Kräussl, 2007). 10

12 of the estimation. We can rewrite Equation 2 as following: y ijv o 1 Btijv 2 ( Bt exp) ijv 3 i 4 i 5 i X ijv ' v ijv (3) where is the coefficient of risk aversion, λ is a measure of loss aversion, and α is a measure of nonlinear probability weighting. A higher or λ implies greater risk or loss aversion, respectively. α<1 (α>1) implies overweighing (underweighting) of small probability events. The results of estimating Equation 3 are shown in Column 3 of Table 4. The coefficient on of 7.32 indicates that if a farmer is more riskaverse than the average farmer by one standard deviation, he uses 4 kg/ha, or 9 percent, more pesticides than the average farmer. There could be several reasons why we find a positive coefficient on the risk aversion parameter. One is that farmers worry about severe bollworm pest infestations, and this concern is exacerbated by the fact that lower quality seeds are rampant in the seed market. Not knowing whether the seed is effective and not knowing how severe the bollworm problem would be, the more risk averse farmers would likely spray more pesticide. The coefficient on λ of implies that if a farmer is one standard deviation more loss-averse than the average farmer, he uses 1.95 kg/ha, or 7 percent, less pesticides than the average farmer. The negative sign on this coefficient may seem surprising at first glance. However, it is not clear what the sign of this coefficient should be. Let us walk through an exercise to illustrate this point. If we suppose that the farmers reference point is the income they would have earned from planting traditional cotton, then it follows that farmers are particularly sensitive to any loss below that income level. In such a case, it is very unlikely that any Bt cotton farmer (which describes most farmers in our sample) would fall below that target income level, given that the low-quality Bt cotton seeds are no worse than traditional seeds. Therefore, the loss aversion parameter would not dictate their pesticide use. On the other hand, if the loss aversion parameter captures more than loss over incomes; specifically, if farmers are particularly lossaverse with respect to their health, they might spray less pesticide. The coefficient on nonlinear probability weighting parameter is negative but statistically insignificant. It is difficult to predict how the nonlinear probability weighting parameter would impact 11

13 pesticide use without making an assumption on probability distribution of various events. For instance, if severe pest infestation is a low probability event, the farmers who overweight low probabilities should be spraying more than those who underweighting small probability. On the other hand, if the quality of Bt seeds are of main concern and supposing that it is probable the Bt cotton seeds are of low-quality, then those who underweight high probabilities (and who also overweight low probabilities) should spray less than those who overweight high probability (underweighting low probability). Given these two opposite effects, the sign of this coefficient is ambiguous. 5.3 Continued Education Negative and significant coefficients on education in all of the above estimations suggest that less educated farmers did not benefit as much from Bt cotton. In fact, what this education variable captures is the knowledge related to the use of Bt cotton or pest management that farmers possess. While it may be too late to provide formal schooling to adult farmers, one policy intervention to help educate farmers is continuing education, or training sessions. The Chinese government has already provided such services, but the utilization rate of these services is far below 100 percent. In the sample, merely 35 percent of farmers have ever attended a Bt cotton training session. These training sessions are usually provided by agricultural extension services or seed companies in each village. There is at least one farmer in each village who has reported attending a training session, so the low participation rate is not due to the unavailability of training at the village level. Figure 1 shows the distribution of the most recent training session they last attended. Out of all the farmers who have ever attended a Bt cotton session, nearly half of them attended one last year. To investigate the impact of Bt cotton training on pesticide use, we estimate the following regression: y ijv o ' 1 Btijv 2 ( Bt exp) ijv X ijv ' Z ij 3training i 4 ( training * t) i v ijv (4) where all the terms are the same as in Equation 3 except that Z ij is a vector of individual risk preferences, which consist of coefficient of risk aversion ( ), loss aversion (λ), and nonlinear probability weighting (α). 12

14 Training i =1 if farmer i has attended a Bt cotton training in the past and zero otherwise. t equals time elapsed since the last training session, which ranges from zero to 10. δ 3 represents the impact of ever attending a Bt cotton training session. Here, δ 4 represents the difference in pesticide use of each year elapsed since attending a training session. Column 4 presents the results for the estimation of Equation 4. The coefficient on training, δ 3, is negative, which indicates that, all else equal, farmers who have attended a training session use less pesticides. Interestingly, the positive sign on δ 4 indicates that the more time that has elapsed since the training took place, the more pesticides a farmer uses. The average farmer who attended a training session 3 years ago uses 2.2 kg/ha, or 8 percent, less pesticides. Now, the coefficient on education is no longer statistically significant once the training variable is included. In some way, both education and training are proxies for farmers knowledge, and these two variables are likely to be highly correlated. 14 The findings here suggest that training does help farmers reduce their pesticide use. However, slowly over time, farmers seem to forget the knowledge they acquired from the training session. The coefficients on training presented in this section could suffer from upward bias due to an omitted variable. For example, it is possible that the more motivated farmers are more likely to attend training sessions, and they could have more knowledge about Bt cotton even in the absence of training sessions. Unfortunately, we cannot further investigate with the existing data set. However, the findings in this section highlight the importance of knowledge and the channel through which the government can spread this knowledge to farmers. 5.4 Seed Quality As mentioned in Section 2, there is wide variation in the quality of Bt cotton seeds. Figure 2 shows the distribution of the cost of Bt cotton seeds from 2 to 200 Yuan/kg. Figure 3 is a histogram of the 14 In a separate regression and controlling only for village fixed effects, we found that one year of education increases the probability of attending a training session by five percent. 13

15 distribution of 930 Bt cotton plots by the source of the seed, with the average price from each source. 15 The cheapest seeds are those produced from on-farm propagation, or saved seeds. We asked farmers if they know whether the saved seeds are of lower, the same, or better quality than the first generation of Bt cotton seeds. Thirty percent of farmers reported that saved seeds are of the same quality, seven percent reported that they do not know, and 63 percent correctly answered that the saved seeds are of inferior quality. The misperception of the quality of Bt cotton seeds may explain why a full 25 percent of seeds of Bt cotton plots come from on-farm propagation. To investigate whether the source or the price of Bt cotton seeds are determinants of pesticide use, we restricted the sample to only Bt cotton plots. While pesticide use could be an indicator of Bt cotton seed quality, it is not a good indicator for non-bt cotton. The quality of conventional cotton is probably determined by its yield performance, which will not be captured in our estimation. The results from the estimation of Equation 3, where the regressor of interest is the seed price, are presented in Column 1 of Table 5. The coefficient on the price of the seed is not statistically different from zero. The finding here complements Pemsl s (2006) finding, that the more expensive Bt cotton seeds are not significantly more resistant than the cheaper Bt cotton seeds. In Column 2, we include a series of dummy variables for the source of the seeds, where the default source is others. None of the source dummies are statistically different from zero. In an alternative regression specification for which the results are not presented here, we include an interaction term between source indicators and the price of the seeds, and again, none of the coefficients are statistically different from zero. This suggests that the quality of Bt seeds may not be consistent within a type of source Robustness Check So far we have imposed a strong function form on the utility function when we impute the risk preference parameters. As a robustness check, we will relax the utility function form, and simply divide 15 For the saved seeds and the seeds exchanged by neighbors, we asked farmers for the estimate of the market value of seeds. 14

16 farmers into 18 groups depending on their 3 switching points. For instance, group 1 includes the farmers who switch from A to B before line 6 both Series 1 and Series 2 and switch before line 4 in Series 3 (thus group 1 should contain the most risk seeking and least loss averse individuals). Group 18 includes those farmers who switch from A to B after line 11 in both Series 1 and Series 2 and switch after line 4 in Series 3 (thus group 18 should be the most risk averse and loss averse group) (see Figure 4). In a regression, instead of using the 3 risk preference parameters, we include the group dummies. 16 An F-test rejects the null hypotheses that these group dummies jointly equal to zero at the 5% level. Therefore, even without imposing the utility functional form, we know the field experiment can predict pesticide use to some extent. We are inclined to believe that the field experiment design captures individual heterogeneity in risk preferences and the functional form helps to ease the interpretation. 6 Conclusion In this study, we investigate the determinants of pesticide use. Our main findings are that more risk-averse farmers use more pesticide, while more loss-averse farmers use less pesticide. Our finding suggests that when farmers decide on how much pesticide to spray, profit maximizing is not their only goal. They do take into account of the potential risk as well as their own health condition. The findings of this study have important policy implications. They suggest that farmers may not benefit as much from new technology as policy makers and scientists would hope. Simply achieving an adoption rate of 100 percent does not guarantee that farmers know how to fully capitalize on the new technology. In order to ensure that farmers reap all the benefits of modern science, continuing education, such as training sessions provided by the government, is essential. Local governments also need to encourage farmers participation in training sessions. Given that most farmers are risk-averse, offering crop insurance could potentially be desirable. While the government cannot change one s risk preference, it can intervene to mitigate the potential agricultural production risk. In particular, the government can reduce this risk by imposing 16 Regression results are presented in Appendix Table. 15

17 stricter regulations on the seed market to reduce the possibility of low quality seeds or provide some type of crop insurance. Another main finding from this study and the existing study by Liu (2008) reinforce the importance of education. Liu (2008) finds that the more educated farmers are more likely to adopt Bt cotton. In this study, we find that the more educated farmers are also more likely to spray less pesticide and attend training sessions, which is also related to the reduction in pesticide use. While it is possible that there could be unobserved heterogeneity that is correlated with education and their pesticide use and timing of adoption, another interpretation is that education provides an informed basis on which farmers can make better decisions, leading to far reaching benefits. 16

18 Reference "The Characteristics of Bt cotton and the integrated pest management," (March 20, 2003): Louyang Agricultural News. Altieri, M. A. (2004): "Genetic engineering in agriculture : the myths, environmental risks, and alternatives." Food First Books/Institute for Food and Development Policy. Antle, J. M. (1988): "Pesticide Policy, Production Risk and Producer Welfare: An Econometric Approach to Applied Welfare Economics," Johns Hopkins University Press. Barberis, N., M. Huang and T. Santos (2001): "Prospect Theory and Asset Prices," Quarterly Journal of Economics, 116(1), Barwale, R. B., V. R. Gadwal, U. Zehr and B. Zehr (2004): "Prospects for Bt Cotton Technology in India," AgbioForum, 7(1&2), Bennett, R. M., T. J. Buthelezi, Y. Ismael and S. Morse (2003): "Bt cotton, pesticides, labour and health: A case study of smallholder farmers in the Makhathini Flats, Republic of South Africa," Outlook on Agriculture, 32(2), Bentley, J. W. and T. G. (March 1999): "Bibliography: Farmer knowledge and management of crop disease," Agriculture and Human Values, 16, 75-81(7). Binswanger, H. P. (1981): "Attitudes Toward Risk: Theoretical Implications of an Experiment in Rural India," The Economic Journal, 91(364), Binswanger, H. P. (1980): "Attitudes toward Risk: Experimental Measurement in Rural India," American Journal of Agricultural Economics, 62(3), Bond, G. and B. Wonder (1980): "Risk Attitudes Amongst Australian Farmers," Australian Journal of Agricultural Economics, 24, Camerer, C. F. and H. Kunreuther (1989): "Decision Processes for Low Probability Events: Policy Implications," Journal of Policy Analysis and Management, 8(4), Campbell, R. A. and R. Kraussl (2007): "Revisiting the home bias puzzle: Downside equity risk," Journal of International Money and Finance, 26(7), Campbell, R. A. and R. Kräussl (2007): "Revisiting the home bias puzzle: Downside equity risk," Journal of International Money and Finance, 26(7), Chavas, J. and M. T. Holt (1996): "Economic Behavior Under Uncertainty: A Joint Analysis of Risk Preferences and Technology," The review of economics and statistics, 78(2), Conko, G. and C. S. Prakash (December 13, 2004): "Can GM Crops Play a Role in Developing Countries?" Biotechnology and Developing Countries: The potential and the challenge, 2004(2). 17

19 Dillon, J. L. and P. L. Scandizzo (1978): "Risk Attitudes of Subsistence Farmers in Northeast Brazil: A Sampling Approach," American Journal of Agricultural Economics, 60(3), Dohmen, T., A. Falk, D. Huffman, U. Sunde, J. Schupp and G. Wagner (2005): "Individual Risk Attitudes: New Evidence from a Large, Representative, Experimentally-Validated Survey," Institute for the Study of Labor (IZA). F. Camerer, C. and H. Kunreuther (1989): "Decision Processes for Low Probability Events: Policy Implications," Journal of Policy Analysis and Management, 8(4), Fung, P. L. P. (2005): "No Soft Landing," Oxfam Briefing Paper. Huang, J., R. Hu, S. Rozelle, F. Qiao and C. E. Pray (2002a): "Transgenic Varieties and Productivity of Smallholder Cotton Farmers in China," The Australian Journal of Agricultural and Resource Economics, 46(3), Huang, J., S. Rozelle, C. Pray and Q. Wang (2002b): "Plant Biotechnology in China," Science, 295(5555), Just, D. R. and T. Lybbert (2009): "Risk Averters that Love Risk? Marginal Risk Aversion in Comparison to a Reference Gamble," American Journal of Agricultural Economics, 91(3), 612. Kahneman, D. and A. Tversky (1979): "Prospect Theory: An Analysis of Decision under Risk," Econometrica, 47(2), Lipton, M. (2001): "Reviving global poverty reduction: what role for genetically modified plants?," Journal of International Development, 13(7), Liu, E. (2009): "Time to Change What to Sow: Risk Preferences and Technology Adoption Decisions of Cotton Farmers in China," Working Paper,. Loppacher, L. and W. Kerr (2004): "China's Regulation of Biotechnology: Does it Conform to the WTO?," Estey Centre for Law and Economics in International Trade, Mimeo. Morse, S., R. M. Bennett and Y. Ismael (2005): "Genetically modified insect resistance in cotton: some farm level economic impacts in India," Crop Protection, 24(5), Moschini, G. and D. A. Hennessy (2001): "Uncertainty, risk aversion, and risk management for agricultural producers," in Handbook of Agricultural Economics}; Handbook of Agricultural Economics, ed. by B. L. Gardner and G. C. Rausser: Elsevier. Officer, R. R. and A. N. Halter (1968): "Utility Analysis in a Practical Setting," American Journal of Agricultural Economics, 50(2), 257. Pemsl, D. (2006): "Economics of Agricultural Biotechnology in Crop Protection in Developing Countries The Case of Bt-Cotton in Shandong Province, China." Pesticide Policy Project. Pemsl, D. and H. Waibel (2005): "Why do some Bt-cotton farmers in China continue to use high levels of pesticides?," International Journal of Agricultural Sustainability, 3(1),

20 Piao, Y., P. Yang and R. Guo "Status of Bt cotton and its impact on cotton production in China," Proceedings of Regional workshop on Bt cotton study in China,, 2-3. Polkovnichenko, V. (2005): "Household Portfolio Diversification: A Case for Rank-Dependent Preferences," Review of Financial Studies, 18(4), Polkovnichenko, V. (2005): "Household Portfolio Diversification: A Case for Rank-Dependent Preferences," Review of Financial Studies, 18(4), Pray, C. E., J. Huang, R. Hu and S. Rozelle (2002a): "Five years of Bt cotton in China the benefits continue." Plant Journal, 31(4), Pray, C., D. Ma, J. Huang and F. Qiao (2001): "Impact of Bt Cotton in China," World Development, 29(5), Prelec, D. (1998): "The probability weighting function," Econometrica, 497. Qaim, M. and A. De Janvry (2005): "Bt cotton and pesticide use in Argentina: economic and environmental effects," Environment and Development Economics, 10(2), Qaim, M. and D. Zilberman (2003): "Yield Effects of Genetically Modified Crops in Developing Countries," Science, 299(5608), Rozelle, S., J. Huang, H. Lin, R. Hu and C. E. Pray (2006): "Eight Years of Bt Cotton in Farmer Fields in China: Is the Reduction of Insecticide Use Sustainable?," FSI Stanford Publications Working Paper. Ruttan, V. W. (2004): "Controversy about agricultural technology lessons from the green revolution," International Journal of Biotechnology, 6(1), 43. Saha, A., C. R. Shumway and H. Talpaz (1994): "Joint Estimation of Risk Preference Structure and Technology Using Expo-Power Utility," American Journal of Agricultural Economics, 76(2), Shaw, K. L. (1996): "An Empirical Analysis of Risk Aversion and Income Growth," Journal of Labor Economics, 14(4), Tanaka, T., C. F. Camerer and Q. Nguyen (2009): "Poverty, politics, and preferences: Field Experiments and survey data from Vietnam," American Economic Review, Forthcoming. Wang, Z. (2008): "Essays on Bt Cotton,". wang, H., X. Hua and J. Mou (1999): "The Integrated Pest Management of Cotton Bollworm." China Agricultural Press. Wang, S., S. Just and P. Pinstrup-Andersen (2006): "Tarnishing Silver Bullets: Bt Technology Adoption, Bounded Rationality and the Outbreak of Secondary Pest Infestations in China," American Agricultural Economics Association Annual Meeting, Long Beach, CA. July Wu, K. (2007): "Monitoring and management strategy for Helicoverpa armigera resistance to Bt cotton in China," Journal of invertebrate pathology, 95(3),

21 Xia, J., J. Cui, L. Ma, S. Dong and X. Cui (1998): "The role of transgenic Bt cotton in integrated insect pest management," Acta Gossypii(11), Yang, P. and R. Jiang (1995): "Review of Nation-wide IPM strategies for the Control of Cotton Bollworm in 1994 and 1995," in Progress Report of IPM of Cotton Bollworm"Review of Nation-wide IPM strategies for the Control of Cotton Bollworm in 1994 and 1995," in China Agricultural Press. Yang, P., M. Iles, Yan Su and F. Jolliffe (2005): "Farmers knowledge, perceptions and practices in transgenic Bt cotton in small producer systems in Northern China," Crop Protection, 24(3), Young, D. L. (1979): "Risk Preferences of Agricultural Producers: Their Use in Extension and Research," American Journal of Agricultural Economics, 61(5, Proceedings Issue), Young, D. L. (1979): "Risk Preferences of Agricultural Producers: Their Use in Extension and Research," American Journal of Agricultural Economics, 61(5, Proceedings Issue), ZHU, E. (August 13, 2006): "RE: Bt cotton," to the author. 20

22 60 Figure 1 Distribution of the Most Recent Training Session Attended (Year of Attendance) Frequency Year Figiure 2 Cumulative Distribution Function of Seed Price Percent Seed Unit Price (Yuan/Kg) 21

23 Average Price (Yuan) Figure 3 Frequency and Average Price of Bt Cotton by Source Frequency Average Price Frequency 0 Seed Companies Village Office Exchange On Farm with Propagation Neighbors Research Institute Seed Vendors Agricutural Extension Others 0 Source 22

24 Figure 4: Division of Switching Points Note: Group 1 consists of individuals who switch from A to B between Line 1 to 5 in Series 1 and Series 2 and switch from A to B between Line 1 to Line 4 in Series 3. 23

25 Table 1 Summary Characteristics Age (8.89) Education 7.10 (2.96) Female 0.14 (0.35) Size of Household 4.49 (1.45) Time Spent Doing On Farm Work (months) 7.63 (1.76) Time Spent Doing Off-Farm Work (months) 0.13 (0.69) Self-Rated Risk Attitude 2.78 (1=most adventurous, 5= least adventurous) (0.92) σ (Risk Aversion) 0.48 (0.33) λ (Loss Aversion) 3.47 (3.92) α (Probability Weighting) 0.69 (0.23) Total Cotton Sown Area (Ha) 0.54 (0.33) Secondary Cash Crop Sown Area (Ha) 0.12 (0.19) Primary Grain Crop Sown Area (Ha) 0.33 (0.33) Total Crop Sown Area (Ha) 1.07 (0.58) Total Land Owned (Ha) 0.59 (0.29) Cotton Yield (Kg/Ha) 3356 (889.8) Average Year of Bt Cotton Adoption 1998 (1.90) Total Value of DGs Per Capita in 2006 (Yuan) (9.37) Observations 320 Note : Standard deviation are in parentheses. 24

26 Table 2 Summary Characteristics By Seed Type Bt Cotton Non Bt Size of Plot (Ha) (0.15) (0.09) Amount of Pesticide Sprayed (Kg/Ha) 26.37* 37.84* (19.44) (27.34) Total Pesticide Cost (Yuan/Ha) (528.55) (877.49) Cotton Yield (Kg/Ha) 3356* * (889.8) (408.8) Total Cost on Seeds (Yuan/Ha) * * (424.76) (145.54) Wealth a (100 Yuan) 1.03* 0.53* (0.92) (0.58) Total Number of Plots Note : Standard deviation are in parentheses. *statistically difference at 5% level. a. Total Value of Durable Goods owned Per Capita in 2006 (100 Yuan) 25

27 Table 3: Payoff Matrix from the Experiment Series 1 Lottery A Lottery B 1 30% winning 20 Yuan and 70% winning 5 Yuan 10% winning 34 Yuan and 90% winning 2.5 Yuan 2 30% winning 20 Yuan and 70% winning 5 Yuan 10% winning 37.5 Yuan and 90% winning 2.5 Yuan 3 30% winning 20 Yuan and 70% winning 5 Yuan 10% winning 41.5 Yuan and 90% winning 2.5 Yuan 4 30% winning 20 Yuan and 70% winning 5 Yuan 10% winning 46.5 Yuan and 90% winning 2.5 Yuan 5 30% winning 20 Yuan and 70% winning 5 Yuan 10% winning 53 Yuan and 90% winning 2.5 Yuan 6 30% winning 20 Yuan and 70% winning 5 Yuan 10% winning 62.5 Yuan and 90% winning 2.5 Yuan 7 30% winning 20 Yuan and 70% winning 5 Yuan 10% winning 75 Yuan and 90% winning 2.5 Yuan 8 30% winning 20 Yuan and 70% winning 5 Yuan 10% winning 92.5 Yuan and 90% winning 2.5 Yuan 9 30% winning 20 Yuan and 70% winning 5 Yuan 10% winning 110 Yuan and 90% winning 2.5 Yuan 10 30% winning 20 Yuan and 70% winning 5 Yuan 10% winning 150 Yuan and 90% winning 2.5 Yuan 11 30% winning 20 Yuan and 70% winning 5 Yuan 10% winning 200 Yuan and 90% winning 2.5 Yuan 12 30% winning 20 Yuan and 70% winning 5 Yuan 10% winning 300 Yuan and 90% winning 2.5 Yuan 13 30% winning 20 Yuan and 70% winning 5 Yuan 10% winning 500 Yuan and 90% winning 2.5 Yuan 14 30% winning 20 Yuan and 70% winning 5 Yuan 10% winning 850 Yuan and 90% winning 2.5 Yuan Series 2 Lottery A Lottery B 1 90% winning 20 Yuan and 10% winning 15 Yuan 70% winning 27 Yuan and 30% winning 2.5 Yuan 2 90% winning 20 Yuan and 10% winning 15 Yuan 70% winning 28 Yuan and 30% winning 2.5 Yuan 3 90% winning 20 Yuan and 10% winning 15 Yuan 70% winning 29 Yuan and 30% winning 2.5 Yuan 4 90% winning 20 Yuan and 10% winning 15 Yuan 70% winning 30 Yuan and 30% winning 2.5 Yuan 5 90% winning 20 Yuan and 10% winning 15 Yuan 70% winning 31 Yuan and 30% winning 2.5 Yuan 6 90% winning 20 Yuan and 10% winning 15 Yuan 70% winning 32.5 Yuan and 30% winning 2.5 Yuan 7 90% winning 20 Yuan and 10% winning 15 Yuan 70% winning 34 Yuan and 30% winning 2.5 Yuan 8 90% winning 20 Yuan and 10% winning 15 Yuan 70% winning 36 Yuan and 30% winning 2.5 Yuan 9 90% winning 20 Yuan and 10% winning 15 Yuan 70% winning 38.5 Yuan and 30% winning 2.5 Yuan 10 90% winning 20 Yuan and 10% winning 15 Yuan 70% winning 41.5 Yuan and 30% winning 2.5 Yuan 11 90% winning 20 Yuan and 10% winning 15 Yuan 70% winning 45 Yuan and 30% winning 2.5 Yuan 12 90% winning 20 Yuan and 10% winning 15 Yuan 70% winning 50 Yuan and 30% winning 2.5 Yuan 13 90% winning 20 Yuan and 10% winning 15 Yuan 70% winning 55 Yuan and 30% winning 2.5 Yuan 14 90% winning 20 Yuan and 10% winning 15 Yuan 70% winning 65 Yuan and 30% winning 2.5 Yuan Series 3 Lottery A Lottery B 1 50% winning 12.5 Yuan and 50% losing 2 Yuan 50% winning 15 Yuan and 50% losing 10 Yuan 2 50% winning 2 Yuan and 50% losing 2 Yuan 50% winning 15 Yuan and 50% losing 10 Yuan 3 50% winning 0.5 Yuan and 50% losing 2 Yuan 50% winning 15 Yuan and 50% losing 10 Yuan 4 50% winning 0.5 Yuan and 50% losing 2 Yuan 50% winning 15 Yuan and 50% losing 8 Yuan 5 50% winning 0.5 Yuan and 50% losing 4 Yuan 50% winning 15 Yuan and 50% losing 8 Yuan 6 50% winning 0.5 Yuan and 50% losing 4 Yuan 50% winning 15 Yuan and 50% losing 7 Yuan 7 50% winning 0.5 Yuan and 50% losing 4 Yuan 50% winning 15 Yuan and 50% losing 5.5 Yuan 26