Amer. J. Agr. Econ. 83(4) (November 2001): 993–1009Copyright 2001 American Agricultural Economics Association

Measuring Producers’ Risk Preferences:A Global Risk-Attitude Construct

Joost M.E. Pennings and Philip Garcia

In applied agricultural economic research various risk-attitude elicitation techniques are used. Here,we investigate whether risk-attitude measures rooted in the expected utility framework are relatedto measures rooted in the multi-item scale framework. Using a second-order factor analyticalmodel, and data obtained from personal computer-guided interviews with 373 farmers, we investi-gate whether the common variance among the (latent) risk-attitude measures can be accounted forby a global risk-attitude construct. We find that the different risk-attitude measures are related, andthat the global risk-attitude construct is significantly related to farmers’ intention to use futures con-tracts. Our research suggests that farmers’ risk attitude is a higher-order characteristic that cannotbe effectively extracted by a single measure.

Key words: expected utility, futures usage, global risk-attitude construct, multi-item scale, second-order factor analysis, validity.

In empirical studies dealing with risk, riskattitudes and producers’ market behaviorvarious risk-attitude measures are used. Twomajor approaches to quantify directly farm-ers’ risk attitudes can be distinguished: mea-sures derived from the expected utility frame-work, and measures derived from responsesto multi-item scales (e.g., Antle; Chavasand Holt; Goodwin and Schroeder; Saha,Shumway, and Talpaz; Smidts).

The purpose of this research is to testwhether risk-attitude measures rooted in theexpected utility framework are related tothe risk-attitude scales rooted in the multi-item scale approach and, if so, whetherthese risk-attitude measures share a common

Joost M.E. Pennings is an associate professor in the Depart-ment of Agricultural and Consumer Economics at the Univer-sity of Illinois at Urbana–Champaign and the AST professorin futures markets in the Department of Marketing and Con-sumer Behavior at the Wageningen University in the Nether-lands. Philip Garcia is a professor in the Department of Agricul-tural and Consumer Economics at the University of Illinois atUrbana–Champaign.

The authors are very grateful for the generous participa-tion of the 373 farmers in our personal computer-assisted inter-views. Financial support provided by the “Algemene StichtingTermijnhandel (AST),” EURONEXT, the Chicago MercantileExchange, the Foundation for Research in Agricultural Deriva-tives, the Niels Stensen Foundation, and the Office for Futuresand Options Research is gratefully acknowledged. The authorsthank J.A. Bijkerk for building a user-friendly interface for thecomputer-assisted personal interviews. The authors express spe-cial thanks to C. Ennew, S.H. Irwin, R.M. Leuthold, M.T.G. Meu-lenberg, A. Smidts, the AJAE editor and two anonymous review-ers who provided helpful comments on the research project andpreliminary versions of the manuscript.

variance that can be attributed to a higher-order factor, a “global risk-attitude construct”(GRAC). Here, we assess the validity andthe relationship among these different risk-attitude measures using recent developmentsin statistics and psychometrics. These pro-cedures, higher-order factor analytical mod-els, allow us to assess whether the differ-ent risk-attitude measures are components ofa GRAC, that is, whether the correlationsamong the different types of risk-attitudemeasures can be accounted for by a GRAC.We hypothesize that the GRAC is relatedto actual farmers’ behavior, a property thatdoes not always holds for single risk-attitudemeasures.

In this article we use four methodsto directly elicit farmers’ risk preferences(Roe).1 Two risk-attitude measures arederived from the expected utility frameworkand two measures are derived from responsesto a multi-item scale. Within the expectedutility framework we use the certainty equiv-alence technique for assessing the utilityfunction. Using both the negative exponen-tial and power functions, we measure the cur-vature of the utility function as a measureof risk attitude (Arrow; Pratt). Several stud-ies have shown that there is a theoretical

1 We focus on directly elicited risk-attitude measures asopposed to risk-attitude measures that are quantified indirectlyfrom observed behavior (Moscardi and de Janvry).

994 November 2001 Amer. J. Agr. Econ.

and empirical difference between the utilityfunction u(x) and the strength of preferencefunction v(x) (e.g., Dyer and Sarin; Ellsberg;Smidts). It is argued that utility function u(x)measures must be adjusted by the strengthof preference v(x) in order to obtain a moreaccurate measure of risk attitude, the intrinsicrisk attitude.We include both the risk attitudeobtained by u(x) and the intrinsic risk atti-tude obtained by relating u(x) to v(x). Withinthe scaling approach we develop two risk-attitude scales based on farmers’ responses tostatements dealing with risk.2

We test the different risk-attitude mea-sures for convergent and nomological valid-ity using statistical tools recently developedin statistics and psychometrics.3 The conver-gent validity of the four measures is testedusing a first-order factor analytical model.This technique allows us to assess the extentto which the different measurements reflectthe same risk-attitude construct (i.e., are posi-tively correlated) (Churchill).4 We then intro-duce a second-order factor analytical modelto investigate whether the risk-attitude mea-sures are components of a GRAC.

We find that the different risk-attitudemeasures used by researchers can beaccounted for by a GRAC. Furthermore, wefind that the GRAC is a better predictor offarmer’s behavior than the single measures.Our research suggests that farmers’ risk atti-tudes are a higher-order characteristic thatcannot be effectively extracted by a singlemeasure.5

Risk-Attitude Measures

Expected Utility Framework

The expected utility model has been usedextensively to investigate behavior under

2 Our purpose is not to determine whether risk-attitude mea-sures rooted in the expected utility model are better than therisk-attitude scales (e.g., Pennings and Smidts). Instead, we inves-tigate whether a combined measure is more useful than its com-ponents in explaining behavior.

3 Convergent validity identifies whether different measure-ments reflect the same construct (i.e., are positively correlated).Nomological validity examines whether measures are related toother constructs in a theoretically meaningful way (Churchill).

4 A construct is a theoretical notion that can be defined con-ceptually but cannot be directly measured or measured withouterror.

5 In spirit, our research follows Robison’s call for furtherresearch on risk-attitude measurement. Robison suggests the useof a function rather than single parameterizations of risk atti-tudes. We use a GRAC that is a function of different risk-attitudemeasurements.

risk. Von Neumann and Morgenstern arethe major contributors to a large body ofwork that provides the justification for theuse of the expected utility model by arational decision maker. The expected utilitymodel views decision making under risk as achoice between alternatives. Decision makersare assumed to have a preference orderingdefined over the probability distributions forwhich a number of axiom’s hold (Fishburn).Risky alternatives can be evaluated underthese assumptions using the expected utilitypreference function, u(x). The curvature ofthe utility function is a measure of risk atti-tude. Pratt and Arrow defined measures thatare independent of a linear transformation ofx. In this article we measure the utility func-tion by means of the certainty equivalencemethod.

Ellsberg introduced the concept of intrin-sic risk attitude to separate marginal valuefrom attitudes toward uncertainty, two fac-tors that are confounded when eliciting riskattitudes in the expected utility framework.Several researchers, among others Dyer andSarin, Krzysztofowicz, Keller, and Smidts,have elaborated on this concept. They con-tend that the assessment of a person’s riskattitude must be done indirectly. That is, itis derived from the assessment of a utilityfunction u(x) and a riskless strength of pref-erence function v(x). They argue that onlyafter removing the nonlinearity in the util-ity function that is the result of the non-linear strength of preference for increasedoutcomes can a person’s preference for riskbe identified. The intrinsic risk-attitude con-cept can be translated into terms of the cer-tainty equivalence technique. The valuationof a lottery is the result of a two-phase pro-cess. First, the outcome in a lottery is trans-formed into subjective values under certaintyby the strength of preference function v(x)(the strength of preference function describesthe intensity of preference a decision makerhas for an outcome). Then, the subjective val-ues are taken as the consequences in the lot-tery which are evaluated under risk, resultingin the utility function u(x). Hence, differencesbetween the utility function and the strengthof preference function are attributed to theinfluence of risk on preferences. Empiricalevidence supports the distinction betweenutility and strength of preference. Smidtsfound strong evidence for the hypothesis thatrisk attitude and strength of preference aretwo distinctive constructs, in a real economic

Pennings and Garcia Measuring Producers’ Risk Preferences 995

setting with a large sample size and a longitu-dinal design. Keller, and Weber and Milliman,also identified empirically a nonlinear rela-tionship between value and utility functions.

Multi-item Scale Approach

In the psychometric approach, risk attitudeis a latent construct (i.e., a not directlyobservable variable) that is measured by aset of observable variables (so-called indi-cators, i.e., questions or items). Guttman,Likert, and Thurstone and Chave proposeddifferent multi-item scaling procedures (see,e.g., Nunnally and Berstein). Because theLikert procedure has performed well withrespect to reliability and validity, it is mostcommonly used. In the agricultural eco-nomics and management literature, severalrisk-attitude measures based on responses toscales have been developed and used suc-cessfully. Goodwin and Schroeder measuredfarmer’s risk attitudes with a binary variablereflecting whether a producer has a prefer-ence for business risk or not. Kunreuther andGinsberg, MacCrimmon and Wehrung, andShapira used several multi-item risk-attitudescales and related them to managers’ atti-tudes and behavior.

Research Method: Measurement ofProducers’ Risk Attitudes

The Risk Context

MacCrimmon, Wehrung and Shapira havedemonstrated that risk attitude is contextspecific. In this article we examine price riskfaced by Dutch hog farmers.6 The Dutch hogindustry, one of the most important industriesin the Dutch agricultural sector, is faced withsubstantial revenue risk. Because of the con-finement production process prevalent in theindustry, the main source of risk is price risk.Consequently, we measure risk in the pricedomain, using hog price fluctuations to gen-erate risk attitudes.

Information on the producers’ risk atti-tudes was obtained by interviews with 373Dutch hog farmers in 1997. A personal inter-view was computerized, taking care to builda user-friendly interface. The software writ-ten for this interview was tested extensively,

6 This implies that farmers might show different risk prefer-ences when asked about the risk regarding chemical use in pro-duction than when asked about price risks faced in selling hogs.

and fifteen test interviews were conducted toensure that the interface was understood bythe farmers and was perceived as “very user-friendly.” The interviewers were thoroughlytrained and were very knowledgeable aboutthe elicitation procedure. Since the elicitationprocedure was fully computerized the inter-viewer was only there to answer questions incase of a misunderstanding. Nine points ofthe utility curve were assessed, correspondingto utilities of 0.125, 0.250, 0.375, 0.500, 0.625,0.750, 0.875 (plus two consistency measure-ments on utilities 0.500 and 0.625). No timeconstraint was imposed on the elicitation pro-cess which lasted about thirty-five minutes.

The interview consisted of several parts.After background questions (e.g., size ofenterprise, previous price-risk managementbehavior), the farmers participated in twoexperiments which measured risk attitude bymeans of the certainty equivalence technique,and strength of preference by means of a rat-ing technique. This was followed by questions(items) to measure farmers’ risk attitude withthe Likert scaling procedure.

The Certainty Equivalence Method

In the certainty equivalence method theresearcher asks the farmer to compare thelottery (xl� p;xh) with a certain outcome,where (xl� p;xh) is the two-outcome lotterythat assigns probability p to outcome xl andprobability 1−p to outcome xh, with xl < xh.The researcher then varies the certain out-come until the respondent reveals indiffer-ence between the certain outcome denotedby CE(p). Substituting in the expected utilitymodel with the von Neumann–Morgensternutility u we obtain u(CE(p)) = pu(xl)+ (1 −p)u(xh) (Keeney and Raiffa).

Several authors have provided conditionsto minimize response biases when elicit-ing a decision maker’s utility function. Inthe agricultural economics literature Bin-swanger, Robison, and Robison et al. offeruseful guidelines. In the management litera-ture Hershey, Kunreuther, and Schoemaker,and Tversky, Sattath, and Slovic provide use-ful guidelines and conditions as well. Robisonet al. argue that biases in the elicitationprocedure may come from different inter-viewers, negative preferences toward gam-bling, absence of realism in the game setting,and compounding of errors in the elicitationprocess. Robison stresses the importance ofconstruction of the choice sets in order to

996 November 2001 Amer. J. Agr. Econ.

obtain the decision maker’s utility function.Hershey, Kunreuther, and Schoemaker, andTversky, Sattath, and Slovic provide similarconditions and guidelines.They argue that theresponse bias is minimized if the followingtwo conditions are met: (1) decision makershave well-articulated preferences and beliefs,and (2) decision makers use a consistent algo-rithm. We designed the elicitation procedurein accordance with these guidelines.

The certainty equivalence technique wasdesigned to resemble the hog producer’sdecision whether or not to fix their pricein advance. We constructed choice sets thatclosely matched the farmer’s daily decision-making process, and hence our choice setreflects the actual choice set facing a farmer(Robison). An important research designissue involves the dimensions of the choiceset. Specifically, what probability and out-come levels should one use in eliciting riskpreferences? The outcome levels were chosenso that all hog price levels that had occurredin the last five years were within the upperand lower outcome levels. Since it has beenargued in the financial literature that com-modity prices follow a random (walk) path,we decided to choose a probability of 0.5,expressing a random walk (prices can go upor down with equal probability).

Farmers were instructed to “read care-fully,” and “to put themselves in the situa-tion of selling their hogs.”7 They were givena choice between three alternatives. Alterna-tive A consisted of a 50/50 lottery (reflectinga spot transaction) where the initial upperand lower bounds were set by the researchers.Alternative B consisted of a fixed price wherethe initial fixed price was randomly gener-ated by the computer within the initial upperand lower bounds. Alternative C consistedof the statement that they were indiffer-ent between alternative A or B. The out-comes in the lotteries were denoted in DutchGuilders per kilogram live weight of hogs.The first lottery presented to the farmers wasa 50/50 lottery with outcomes of 2.34 and 4.29Dutch Guilders as the minimum and maxi-mum price of hogs. The assessment of the cer-tainty equivalent was an iterative process. Ifthe farmer chose alternative A, the computerwould generate a higher fixed price (alterna-tive B) than the previous one, hence making

7 The validity of scenarios and role-playing is well documented(Bem) and is particularly successful when individuals are askedto “play themselves” rather than unfamiliar roles.

alternative B more attractive. If the farmerchose alternative B, the computer would gen-erate a lower fixed price (alternative B) thanthe previous one, hence making alternativeA more attractive. At some point, the farmerwould become indifferent between A or Band would choose alternative C. The nextmeasurement (the next lottery) started afterthe farmer had chosen C.

Our choice set was perceived as realistic bythe farmers, and straightforward.The decisionsituation was very transparent and unambigu-ous. They had clear preferences in choosingbetween options A [receiving a relative highprice or low price (the “lottery”), reflectingsales in the spot market] and B (the fixedprice, reflecting a fixed price contract).

After obtaining the certainty equivalents, autility function is fit to the observations foreach farmer to determine their risk attitude.In the expected utility framework the func-tional form of the utility function u(x) is leftopen. Tsiang refers to Arrow who providesfour conditions for an acceptable utility func-tion: (1) marginal utility of wealth is posi-tive, (2) marginal utility of wealth decreaseswith increasing wealth, (3) marginal absoluterisk aversion is constant or decreasing withincreasing wealth, and (4) marginal propor-tional risk aversion is constant or increasingwith increasing wealth. The negative expo-nential, power, and logarithm functions meetall four conditions. Keeney and Raiffa, andFishburn and Kochenberger have demon-strated that the negative exponential andpower functions perform well relative to thelogarithm function. Moreover, both functionsare easy to work with. After scaling theboundaries of the functions, the estimation ofonly one parameter suffices to characterizea farmer’s risk attitude. We applied both thepower and negative exponential function tospecify the utility function u(x). The negativeexponential function can be expressed as

u(xi) = 1 − e−c(xi−xL)

1 − e−c(xH−xL)(1)

where xL and xH denote the lower and upperbound of the outcome range of the 50/50 lot-tery, respectively, xi stands for the assessedcertainty equivalent, and the parameter cis the risk-attitude coefficient. The negativeexponential function implies a constant abso-lute risk attitude and an increasing propor-tional risk attitude. The power function canbe expressed as

u(xi) = (xi − xL)a

(xH − xL)a

(2)

Pennings and Garcia Measuring Producers’ Risk Preferences 997

where the parameter a measures risk atti-tude. The power function implies a decreasingabsolute risk attitude and a constant propor-tional risk attitude.

Because the certainty equivalents are mea-sured with error and not the utility levels, theinverse functions are estimated. The inverseexponential function can be expressed as

xi = ln(0�5(e−cxl + e−cxh))−c + ei(3)

where xl and xh represent the low and highoutcomes of the 50/50 lottery, respectively.The inverse power function can be expressedas

xi = (xH − xL)

(0�5

((xl − xLxH − xL

)a

(4)

+(xh − xLxH − xL

)a)1/a)+ xL + ei�

The Rating Technique

The strength of preference function v(x) wasassessed by means of a rating technique. Theproducers were asked to value nine price lev-els on a scale from 1 to 10 to reflect theattractiveness of the price to their opera-tions (with 10 being the most attractive). Pro-ducers were also permitted to specify theirvalue in terms of fractions 0.25, 0.50, and0.75. Before the producer started this task,the price range from which the price levelswere drawn and were shown. The price levelswere drawn over the same range as the lotter-ies. This task also was straightforward for ourfarmers because the rating scheme resembledthe grading system used in Dutch schools andwas consistent with how our farmers referto prices. Frequently, they speak of “excel-lent,” “good,” and “bad” prices which is sim-ilar to excellent, good, and bad grades inDutch schools.

The strength of preference function v(x)was also formulated in terms of negativeexponential and power function as reflectedin (1) and (2).8

The Intrinsic Risk Attitude

The assessment of the intrinsic risk atti-tude is indirect, i.e., it is derived fromthe assessed utility function u(x), by means

8 Because the order of prices was random during the ratingprocess, the measurements can be viewed as independent, andestimation of the inverse functions is not necessary.

of the certainty equivalence technique, andthe strength of preference function v(x) bymeans of the rating technique. The intrin-sic risk attitude is determined by relatingthe functions such that u(x) = f(v(x)).The intrinsic risk-attitude measure is definedas −u′′(v(x))/u′(v(x)) (the analogue to thePratt–Arrow coefficient of risk aversion). Itrepresents the remaining nonlinearity in theutility function, after eliminating the nonlin-ear effect related to the strength of prefer-ence function v(x).

The intrinsic risk attitude is estimatedby relating the utility function u(x) to thestrength of preference function v(x) by theexponential and power functions, which areexpressed as

u(x) = 1 − e−cv(x)

1 − e−c and u(x) = v(x)a(5)

respectively.

The Risk-Attitude Scales

We adhered to the iterative procedure rec-ommended by Churchill to obtain reliableand valid scales.9 First, a large pool of ques-tions (i.e., indicators) was generated. Theindicators were based on the literature avail-able, and care was taken to tap the domainof the construct. The indicators were testedfor clarity and appropriateness in person-ally administered pretests. The farmers wereasked to complete a questionnaire and indi-cate any ambiguity or other difficulty theyexperienced in responding to the items, aswell as any suggestions they deemed appro-priate. Based on the feedback received fromthe farmers, some items were eliminated, oth-ers were modified, and additional items weredeveloped. Farmers were asked to indicate ona Likert scale from -4 (“I strongly disagree”)to 4 (“I strongly agree”) the extent to whichthey agreed with the items (statements) dis-played in table 1.

Results of the Risk-AttitudeMeasurements

Expected Utility Framework

In table 2, descriptive statistics of the mea-sured certainty equivalents are shown. The

9 Reliability identifies whether variables are consistent withwhat they are intended to measure. Validity identifies the extentto which a set of measures correctly represents a concept.

998 November 2001 Amer. J. Agr. Econ.

Table 1. Items Representing Farmers’ Risk Attitude

Items

1. When selling my hogs, I prefer financial certainty to financial uncertainty.2. I am willing to take higher financial risks in order to realize higher average returns.3. I like taking financial risks.4. When selling my hogs, I am willing to take higher financial risks in order to realize higher

average returns.5. I like “playing it safe.”6. With respect to the conduct of business, I am risk averse.7. With respect to the conduct of business, I prefer certainty to uncertainty.

order in which the lotteries were presentedto the producer is indicated by the numberin the first column. The second and thirdcolumns show the outcomes used in eachlottery. With the exception of the first lot-tery in which the outcomes are 2.34 and 4.29Dutch Guilders for all producers, the out-comes of the lotteries depend upon the pro-ducers’ answers in previous lotteries. Conse-quently, the expected value of the lottery andrange of the lottery for each level of expectedutility vary among producers.

We had two measurements at u(x) = 0�5and two measurements at u(x) = 0�625 inorder to test the internal consistency of theassessments. If farmers respond in accor-dance with expected utility theory, the samecertainty equivalents should result except forrandom response error. When tested, thedifferences between the assessed certaintyequivalents for the same utility levels werenot significant (p > 0�99; pairwise test) forboth consistency measurements. We also cal-culated the mean absolute deviation betweenthe two consistency measurements (at bothu(x) = 0�5 and u(x) = 0�625) relative to theoutcome ranges XL and XH , which wereonly 0.05 and 0.07 Dutch Guilders/kg, respec-

Table 2. Results of the Assessment of the Certainty Equivalence Technique (Guilders/kg)

Lottery Certainty EquivalentMeasurement Expected Range of E(x)−xi xl xh Utility Mean Median St. Dev. Lottery E(x) CE

1 2.34 4.29 0�5 3�35 3�34 0�44 1�95 3�32 −0�0402 2.34 x1 0�250 3�03 2�91 0�45 1�86 2�85 −0�1783 x1 4.29 0�75 3�72 3�78 0�40 1�70 3�82 0�0984 2.34 x2 0�125 2�82 2�63 0�42 1�82 2�68 −0�1355 x2 x1 0�375 3�21 3�20 0�44 1�86 3�19 −0�0226 x1 x3 0�625 3�57 3�59 0�43 1�73 3�54 −0�0337 x3 4.29 0�875 3�99 4�05 0�29 1�42 4�01 0�0238 x2 x3 0�5 3�44 3�45 0�46 1�90 3�37 −0�0659 x5 x7 0�625 3�65 3�70 0�40 1�84 3�60 −0�054

Note: xl is the relative low price in the lottery, xh is the relative high price in the lottery, E(x) is the expected value of the lottery, and CE is the certaintyequivalent.

tively. These findings support the notion thatfarmers use an EU framework in their deci-sions and are consistent in their choices. Thisfurther substantiates that the design closelyresembles the real business context of thefarmers, thereby minimizing response modeeffects (Payne; Shapira).

Table 2 also identifies the differencebetween the lottery’s expected value E(x)and the certainty equivalent CE. A posi-tive difference indicates risk-averse behavior,while a negative difference points to risk-seeking behavior. Of the farmers, 64.5% showrisk-seeking behavior in lotteries with rela-tively low utility [e.g., u(x) = 0�125, u(x) =0�250 and u(x) = 0�375]. When utility is rel-atively high [u(x) = 0�75 and u(x) = 0�875],39.8% of the farmers show risk-averse behav-ior which is in accordance to Kahneman andTversky’s prospect theory.

In table 3, the descriptive statistics of theaverage parameter estimates are presented[e.g., (1)–(6)].10 Both the negative exponential

10 The parameters in (1)–(6) are estimated with the nonlin-ear least squares routine ZXMIN from the IMSL library thatemploys the Fletcher’s quasi-Newton method.

Pennings and Garcia Measuring Producers’ Risk Preferences 999

Table 3. Average Results of Estimating the Risk Attitude and Strength of Preference forthe Negative Exponential Function and the Power Function (N = 373)

Certainty Equivalence Technique Rating Technique

Exponential Power Exponential Power

Parametera c a c aMean −0�742 −1�211 0�332 0�127Median −0�274 −0�225 0�371 0�202St.dev. 2�545 3�915 0�583 0�412

Fit indicesMean MSE 0�027 0�028 0�011 0�011Median MSE 0�018 0�018 0�007 0�007Mean R2 0�883 0�855 0�903 0�904Median R2 0�916 0�921 0�938 0�941

Classification of respondentson the basis of the t-valueb

Risk averse 34% 31% Decreasing SP 73% 73%Risk neutral 6% 7% Constant SP 5% 5%Risk seeking 60% 62% Increasing SP 22% 22%

Note: SP is the marginal strength of preference.a In the case of the certainty equivalent technique, if c or a > 0 the producer is said to be risk averse and if c or a<0 the producer is said to be risk seeking.In the case of the rating technique, c and a > 0 (< 0) reflect a decreasing (increasing) marginal strength of preference.bA producer is classified risk neutral when the parameter is not significantly different from 0 at the 0.01 level. Note that to perform this test, the residualsmust be normally distributed as well as independently and identically distributed for each individual. Because it is questionable whether the residuals fitthe assumptions, the analysis performed serves illustrative purposes only.

and the power function fit the data equallywell; i.e., the mean sums of squared errors(MSE) are almost equal for both specifi-cations. The similarity of functional fit wasfurther verified by pairwise comparisons ofthe MSEs for the negative exponential andpower functions.

The negative parameter estimates for boththe negative exponential and power func-tion for the farmer’s utility function obtainedin the certainty equivalence technique implythat 60% of the Dutch farmers are (price)risk seeking.11 This finding is in line with var-ious studies in the management literature inwhich a large number of risk-seeking owner–managers have been detected. For exam-ple, Brockhaus; Keller; March and Shapira;MacCrimmon and Wehrung; Miller, Kets deVries, and Toulouse; Shapira; and Smidtsfound risk-seeking behavior. Smidts’ studywas based on Dutch crop farmers using a lon-gitudinal research design and a large sample(N = 253) to study farmers’ decision-makingbehavior in light of their price risk attitude.

11 A reviewer identified that our finding of risk-seeking behav-ior may have been influenced by the fact that decision makersdo not always weight probabilities linearly when evaluating a lot-tery. This can lead to a skewness in the perceived price distribu-tion such that producers who are risk neutral or risk averse mayidentify a certainty equivalent that is greater than the expectedvalue of the lottery.

Several explanations of risk-taking behaviorhave been advanced depending on the spe-cific domain. For example, Jaworski and Kohlifound that responding to market develop-ments entails some degree of risk taking.Han, Kim, and Srivastava found that greatermarket orientation leads to higher degrees ofrisky, innovative behavior.

The positive parameter values obtained forthe strength of preference function obtainedby the rating technique indicate that 73%of the producers are value averse; that is,a farmer values a given increase in DutchGuilders in a relatively low price range morethan the same increase in a relatively highprice range. All ratings of the randomly pre-sented prices were consistent; that is, higherprices were consistently rated higher.

Table 4 shows the estimation results forthe intrinsic risk attitude. The hypothesis ofintrinsic risk attitude, implying that risk atti-tude and strength of preference are differ-ent concepts, is confirmed in our data (thecorrelation between the strength of prefer-ence and risk attitude is 0.09 for the nega-tive exponential function with p = 0�1, and0.07 with p = 0�2 for the power function). Aswas the case for the risk-attitude and strengthof preference functions, both functions fit thedata equally well.

1000 November 2001 Amer. J. Agr. Econ.

Table 4. Average Results of Estimatingthe Intrinsic Risk Attitude for a NegativeExponential Function and a Power Function(N = 373)

Exponential Power

Parameter c aMean −2�084 −1�580Median −1�278 −0�640St.dev. 5�198 4�620

Fit indicesMean MSE 0�013 0�012Median MSE 0�007 0�006Mean R2 0�901 0�903Median R2 0�944 0�945

Classification of respondentson the basis of the t-valueRisk averse 28% 25%Risk neutral 1% 3%Risk seeking 71% 72%

Note: Divide the parameter estimates here by 1.95 (which is the range ofprice levels, i.e., xH − xL) to compare with those in table 3. See table 3for a discussion of the terms and caveats with the analysis.

Scaling Framework

Exploratory factor analysis on the items oftable 1 yielded eigenvalues for the first twofactors of 3.04 and 1.34. The results stronglysupport a two-factor model where the firstfactor explained 39% of the variation in thedata and the second 17%. All the factor load-ings of the items exceeded 0.4 (Bartlett’s testof sphericity = 649, p = 0�00, and KMO mea-sure of sampling adequacy = 0�8).12 The firstfour terms in table 1 make up Scale 1; thelast three terms make up Scale 2. The relia-bility of Scale 1 was 0.75 and that of Scale 2was 0.70, indicating a good reliability for theconstruct measurement.13

Based on these risk-attitude scales, wedivided the sample into risk-averse farmersand risk-seeking farmers. The split was basedon the average sum of the score on the itemsof the two scales. Farmers who had a negativesum score are risk seeking and those who hada positive sum score are risk averse. Farmerswho had a sum score of zero are risk neutral.

12 The null hypothesis in the Bartlett test is that the correla-tion matrix of the items in the scale is an identity. Rejectingthe null hypothesis supports a factor model. The Kaiser–Meyer–Olhin statistic also assesses the appropriateness of factor analy-sis. It ranges from 0 to 1, reaching 1 when each item is perfectlypredicted by the other items.

13 The reliability scale ranges from 0 and 1, with higher valuesindicating greater reliability. With repeated measurement, morereliable scales show greater consistency (Hair et al.).

Table 5. Classification of RespondentsBased on the Sum Scores of the Risk-Attitude Scales

RiskScale Averse Risk Neutral Risk Seeking

Scale 1 62% 6% 32%Scale 2 43% 5% 52%

Note that some of the scale’s items had to berecoded, so that a negative score on the itemsis associated with risk seeking and a positivescore is associated with risk aversion. Table 5presents these results. Again, we find a rela-tively large group of farmers that exhibit risk-seeking behavior, consistent with our findingsof the risk-attitude measures rooted in theexpected utility approach.

When comparing tables 3 and 4 withtable 5, we observe that the classificationbased on Scale 2 is similar to the classi-fication proposed by the certainty equiva-lence method, in that more farmers may beconsidered to exhibit risk-seeking behaviorthan risk-averse behavior. However, for Scale1 more farmers exhibit risk-averse behaviorthan risk-seeking behavior.

Testing for the Global Risk-AttitudeConstruct

To examine whether the four risk-attitudemeasures can be viewed as components of amore comprehensive representation of risk,a GRAC, we use the following procedure.First, we assess the bivariate correlationsbetween the different measures to determinewhether they are related. This provides somemeasure of convergent validity but assumesthat each measurement is equally reliablewhich is unlikely. Second, in the presenceof positive correlations, we estimate a con-firmatory factor analysis model that permitsthe identification of the relationship betweenthe indicators taking measurement error intoaccount (Bollen). Confirmatory factor anal-ysis differs from the standard factor analy-sis, which combines variables based solely oncommon correlation, because the structureof the relationship among the (latent) risk-attitude measures and their particular indi-cators is specified. For example, items 1–4 infigure 1 only are permitted to influence Scale

Pennings and Garcia Measuring Producers’ Risk Preferences 1001

Figure 1. First-order confirmatory factor modelIRA is the instrinsic risk-attitude measure, and RA is the risk-attitude measure.

1. When the (latent) risk-attitude measuresare not orthogonal, it is possible to investi-gate the presence of a common factor acrossthe measures, our global risk-attitude con-struct, through the use of a second-order fac-tor model. The second-order model quantifiesthe presence of a common (latent) factorbased on the correlations across the risk-attitude measures. The second-order model ismore restrictive than standard factor analy-sis models as we impose the structure of howthe indicators can influence the risk-attitudemeasures, and how the risk-attitude mea-sures are combined to form the global risk-

attitude construct. Nevertheless, it is moretheoretically pleasing in that it can providea rationale for correlations. Specifically, ifeach of the four first-order factors (our risk-attitude measures) can be effectively incorpo-rated into the second-order model, the fourfactors can be treated as items representingthe GRAC (Benson and Bandalos). Assess-ment of the effectiveness of the second-ordermodel is based on statistical measures of fit ofboth the first- and second-order factors. Theappendix specifies the first- and second-ordermodels.

1002 November 2001 Amer. J. Agr. Econ.

Results

Table 6 shows the correlation matrix for thedifferent measurement methods. Note that inthe correlation analysis we used the averagesum of the score on the items of Scale 1 andScale 2. Except for the correlation betweenScale 2 and the intrinsic risk attitude, we findthat all measurement methods for risk atti-tude show a significant correlation.

Table 6 establishes convergent validity ofthe four different measurement methods.That is, the measures are reflecting simi-lar dimensions of risk attitude. Although theobserved correlations can give us an indi-cation of convergent validity, this procedureassumes that each indicator has an equalweight in their respective risk-attitude mea-sure which is not likely the case. Therefore,we use a first-order factor model to relatethese four measures.

Figure 1 visualizes the first-order factormodel. The right hand side shows the par-tial regression coefficients between the indi-cators and the risk-attitude measures. The lefthand side shows the correlations between the(latent) risk-attitude measures and their t-values in parentheses. The first-order factormodel has a good fit with a χ2/df of 3.2 (p =0�0), a root mean squared error of approx-imation (RMSEA) of 0.07, a goodness-of-fit index (GFI) of 0.96, and a Tucker LewisIndex (TLI) of 0.92, and it indicates thatthe four risk measures are related, suggestingthe presence of a common variance and thestrong likelihood of an overall (second-orderfactor) GRAC. The RMSEA estimates howwell the fitted model approximates the pop-ulation covariance matrix per degree of free-

Table 6. Correlation Matrix between the Different Risk-Attitude Measurements

Scale 1 Scale 2 CE Technique Intrinsic

Scale 1 1.000p = �

Scale 2 0.3841 1.000p = 0�00 p = �

CE technique 0.1449 0.1087 1.000p = 0�01 p = 0�04 p = �

Intrinsic 0.1502 0.0773 0.7761 1.000p = 0�00 p = 0�14 p = 0�00 p = �

Note: A correlation in bold indicates that the correlation is significant at p < 0�05, where CEtechnique indicates that the risk-attitude measure is obtained using the certainty equivalencetechnique.

dom, the GFI represents the overall degree offit, that is, the squared residuals from predic-tion compared with the actual data, and theTLI is an incremental fit measure that com-bines a measure of parsimony into a compar-ative index between the proposed and nullmodel.14

The hypothesis that the common vari-ance of the four risk-attitude measuresis attributed to a higher-order factor, theGRAC, is tested using a second-order fac-tor analysis. The second-order factor modelis estimated in the maximum likelihood LIS-REL framework developed by Joreskog andSorbom.15 The second-order factor model hada good fit, with a χ2/df of 3.14 (p = 0�00),a RMSEA of 0.07, a GFI of 0.96, and a TLIof 0.92. Focusing on the relationship betweenthe risk-attitude measures and global risk-attitude measure, the second-order confirma-tory factor model is depicted in figure 2.

The right hand side of figure 2 showsthe loadings (partial regression coefficients)of the first-order factors (the risk-attitudemeasures) on the second-order factor (theGRAC). These loadings (which are all signifi-cant) reflect the contributions of the differentrisk-attitude measures to the GRAC. Figure 2reveals the important contribution to GRACthat measures derived from the expected util-

14 For the RMSEA, Browne and Cudeck suggest that a valuebelow 0.08 indicates a close fit. The GFI ranges from 0 (poor fit)to 1.0 (perfect fit). For the TLI Hair et al. recommended a valueof 0.9 or greater.

15 Prior to using LISREL we tested the data for univariate andmultivariate normality. The coefficient of relative multivariatekurtosis was 1.08, indicating that the assumption of multivariatenormality is tenable (Steenkamp and Van Trijp).

Pennings and Garcia Measuring Producers’ Risk Preferences 1003

Figure 2. Second-order confirmatory factor modelIRA is the intrinsic risk-attitude measure, and RA is the risk-attitude measure.

ity framework make. The target coefficientT (first introduced by Marsh and Hocevar),which is the ratio of the chi-square of thefirst-order model to the chi-square of thesecond-order factor model, is 0.99, which isquite large, the upper value of T being 1. Thisindicates that the relations among the first-order factors can be accounted for in termsof the more restrictive second-order factormodel. From the estimation results, it can beconcluded that risk attitude can be concep-tualized as a generalized response of farm-ers in evaluating different situations, which isexplained by the four first-order risk-attitudemeasures.

We hypothesize that the global risk atti-tude is related to farmers’ risk-managementpractices, and that the global risk-attitudeconstruct is better in explaining farmers’behavior than the single risk-attitude mea-sures. In order to test this hypothesis weintroduce a structural equation model inthe next section that includes the globalrisk-attitude construct and farmers’ futuresusage.

The Global Risk-Attitude Construct andRisk-Management Practices

The adoption of the portfolio theoryapproach in the 1960s to decisions in futuresmarkets sets risk at the center of why indi-viduals hedge (e.g., Stein, Johnson). Several(hedging) models in different disciplines(agricultural economics, finance, and eco-nomics) have been used in the past to showthat increasing risk attitude leads to an

increase in futures usage. Further, this liter-ature suggests that an increase in risk aver-sion will lead to an increase of the use ofprice risk-management instruments. Empiri-cal research by Geczy, Minton, and Schrand;Koski and Pontiff; Mian; and Nance, Smith,and Smithson found a significant relation-ship between risk aversion and derivativesusage. Based on this theoretical and empir-ical research we expect to find a positiverelationship between farmers’ intention touse price-risk management instruments andfarmers’ risk aversion.16

We focus on futures contracts as means toreduce the price risk of hogs. In the Nether-lands hog futures contracts are traded atthe Amsterdam Exchange. During the inter-view the farmers were asked to identifytheir attitude toward futures use by respond-ing to three questions (items). These itemsreflect what we call the farmers’ intentionto use futures contracts scale. The first itemmeasured the attitude toward futures (i.e.,whether in general they viewed futures as anattractive marketing strategy), by asking thefarmer to distribute 100 points between usingfutures or not using futures, with more pointsassigned to futures indicating a more posi-tive attitude toward futures. The second item,the intention to use futures, was measured byasking the farmer to identify their probabil-ity of using futures by distributing 100 pointsbetween futures or not using futures. Wechose to measure the attitude and inten-tion toward futures by means of distribut-

16 A negative relationship between risk aversion and hedgingactivity may exist. However, the weight of the evidence is moreconsistent with our discussion.

1004 November 2001 Amer. J. Agr. Econ.

ing 100 points between futures and not usingfutures, in light of the study by Putte van den,Hoogstraten, and Meertens that empiricallydemonstrates that relative measurements ofconstructs such as attitude and intention aresuperior when obtained as direct compar-isons of competing alternatives. Finally, in thethird item, the producers were asked, “sup-pose you have to market your hogs today;would you use futures or not?” The reliabilityof this scale, consisting of these three items,was 0.8, reflecting high reliability and indi-cating that the answers to all the items arehighly related (Hair et al.).

Our measure was added to the model dis-played in figure 2 to relate the GRAC tofarmers’ intentions to use futures contracts.Specifically, a full (latent) model, includingthe indicators for the four different risk-attitude measures, GRAC, and the use offutures, was estimated in a structural equa-tion model framework to identify the effectof risk attitudes on intentions to use futures.See the appendix for the specification of thestructural equation model (also, Pennings andLeuthold).

This model was estimated in the maximumlikelihood LISREL framework (Joreskogand Sorbom). The model had a good fit witha χ2/df of 2.01, (p = 0�0), a RMSEA of 0.06,a GFI of 0.95, and a TLI of 0.93. Further,all the hypothesized relations were supportedby significant t-values [the partial regres-sion coefficients, reflecting the relationshipbetween the (latent) risk-attitude measuresand the GRAC, were similar to the ones asdisplayed in figure 2 and were all significant].The model showed that the farmers’ GRACsignificantly influenced the use futures (β =0�115, t = 2�70, p = 0�004), thereby sup-porting the hypothesis that risk attitude is animportant influence behind farmers’ use offutures, and that risk attitude can be mea-sured by a set of measures rooted in differentdisciplines. The β coefficient indicates that asthe producers become more risk averse thelikelihood that they will increase their use offutures increases.

The model fit further substantiates that theGRAC is able to describe the latent attitudetoward risk and that it affects farmers’ use offutures. The GRAC explains 60% of the vari-ance in the farmers’ intentions to use futuresmeasure. In order to gain more insight intothe contribution of using the GRAC tounderstand risk-management behavior, we

estimated the relationship between the farm-ers’ use of futures measure and each (latent)risk-attitude measure separately. Scale 1 andScale 2 were able to explain 9% and 11%of the variance, respectively. The risk-attitudemeasures rooted in the expected utility modelshowed higher explained variance. The risk-attitude measures obtained by the lotteryexplained 25% of the variance, while theintrinsic risk attitude performed slightly bet-ter with 27%.17 The fact that the measuresrooted in the expected utility framework per-formed best was expected since their contri-bution to GRAC was relatively large. We alsoregressed using least squares the intent touse futures markets on the four independentvariables (risk-attitude measures) and com-pared the adjusted R-squares with the regres-sion where GRAC is the explanatory vari-able. As expected, the GRAC performs (45%compared to 58%) better, further substanti-ating the role of the GRAC as a frameworkto analyze farmers’ behavior.

The analysis demonstrates that the fourrisk-attitude measures reflect dimensions(higher-order factors) of producers’ risk atti-tudes, and that their combined use in aGRAC framework explains producer behav-ior in a manner consistent with theoreti-cal expectations. Results further suggest thatrisk attitude, measured as a second-order fac-tor (i.e., GRAC), can be an important vari-able in explaining farmers’ risk-managementbehavior.

Conclusions

In this article, we have presented an approachto relate the effect of risk attitudes on man-agement decisions. Using (second-order) fac-tor analysis and structural equation model-ing which account for measurement error,we develop a global risk-attitude measurethat is based on risk-attitude measures devel-oped in the expected utility and the multi-item scale frameworks. Based on our GRAC,the hypothesis that farmers’ risk attitudesinfluence their adoption of risk-managementpractices is confirmed. Further, our globalmeasure based on the results from the LIS-REL framework is able to explain 60% ofthe variability in producers’ intentions to usefutures contracts, while the measures con-sidered separately explain between 9% and

17 The details are available on request.

Pennings and Garcia Measuring Producers’ Risk Preferences 1005

27%. It is clear that our research suggeststhat farmers’ risk attitudes are a higher-order characteristic that cannot be effectivelyextracted by a single measure. For appliedeconomists interested in measuring the effectof risk attitudes on producer decisions, thismay mean that it is necessary to considervarious risk-attitude measures, and that theobtained GRAC may be a more accuratemeasure for risk attitude than the single mea-sures. In our view the usefulness of our pro-cedure is most apparent in developing anunderstanding of the factors (e.g., risk atti-tude) that influence decision-maker behav-ior. Inappropriate or incomplete formula-tion of the decision-making process may leadto inaccurate inferences about behavior. Aclearer understanding of the factors affect-ing behavior may lead to an improvement inpredictiveness of behavior and may identifythe areas where theoretical development isneeded.

Several points should be mentioned thataffect the use of our procedure. First, nonlin-earities between the elicitation procedure inthe certainty equivalence technique and thefarmer’s existing portfolio can influence theobtained risk-attitude coefficients (Robisonand Barry). Here, this would imply the pos-sible presence of a nonlinear relationshipbetween hog prices and the farmer’s end-of-period revenue due to insurance or taxes.While a potential problem, in all likelihood,the influence of these factors in our sampleis limited since in the Netherlands no incomeinsurance exists, and for a large portion of theproducers in the sample the tax rate wouldbe relatively flat.

Second, care seems warranted in devel-oping and interpreting the individual riskmeasures, particularly the measures from thecertainty equivalence method. Here, the cer-tainty equivalent technique was conductedwith probability equal to 0.5. Tversky andWakker among others have shown thatresponse bias is minimized when these prob-abilities are used. However, these proba-bilities may not have coincided with theactual probabilities used by the farmers inthe experiments based on their experiences.For example, when prices were near theminimum of the price range, farmers mayhave consciously or subconsciously assigned alower probability than specified in the exper-iment to the low price, leading to skew-ness in their price probability distribution andto more risk-seeking responses. Further, the

individual risk measures may be influencedby factors not explicitly included in the exper-imental design. For example, financial con-straints may have affected the response toa level of price risk. Following Robison andBarry, Shapiro and Brorsen, and Turvey andBaker, farmers under limited financial stress(low debt-to-asset ratios) may be more risktaking in the price domain. Our risk-attitudecoefficients are indeed negatively correlatedwith the debt-to-asset ratio (ρ = −0�15, p =0�045); producers with lower debt-to-assetratios demonstrated a higher willingness toaccept risk.

Third, in response to the issue of biaswhen using the certainty equivalence tech-nique King and Robison introduced the inter-val technique as a measurement procedureto elicit a decision maker’s absolute risk-aversion function. The interval technique isbased on the premise that under certain con-ditions, a choice between two distributionsdefined over a relatively narrow range of out-come levels divides the absolute risk-aversionspace over that range into two regions: oneconsistent with the choice and one inconsis-tent with it. This method can reduce the oper-ational problems of the certainty equivalencetechnique, and in a general context its useshould be considered when the researcheris likely to be confronted with large, uncon-trollable response bias. However, its poten-tial use with our procedure may be some-what limited. The interval technique does notprovide an exact representation of prefer-ences, rather a band of risk attitudes overa range of potential outcomes, which wouldmake comparisons across individuals andacross risk-attitude measures as we do hereexceedingly problematic. Further, its theoret-ical and empirical relationship to the con-cept of intrinsic risk attitude is not readilyapparent. This all suggests that the use ofour procedure requires the development ofan effective research design, including con-sistency checks and validity procedures, toensure the measurement of meaningful risk-attitude measures.

Fourth, we assume that risk attitude is astable construct that does not change overstimuli ranges, in our case price ranges.Kahneman and Tversky, in their prospect the-ory, argue that risk attitude would changeover the price range (individuals in a losssituation would exhibit risk-seeking behav-ior). To gain further insight into farm-ers’ risk-taking behavior we investigated

1006 November 2001 Amer. J. Agr. Econ.

within the expected utility framework foreach lottery the percentage of respondentsthat showed risk-averse, risk-neutral, or risk-seeking behavior. We observed a slight ten-dency of a growing number of risk-aversefarmers as utility increased. Measures in thescaling framework cannot be tested for thistendency as they are measured indepen-dently of a price range. In a similar con-text, due to the nature of the data, weassumed stability of the GRAC over time.Combining the global risk-attitude constructwith work recently conducted (Barry, Robi-son, and Nartea; Lence) on intertemporal riskanalysis may lead to a promising area for fur-ther research.

Finally, a disadvantage of measuring aGRAC in empirical studies, using four risk-attitude measures from two different disci-plines, is the high costs when conducting theresearch with a large sample. The disadvan-tage has to be weighed against the advantageof obtaining more accurate risk-attitude mea-sures. The source of the high costs is the mea-sures rooted in the expected utility frame-work, as they can only be obtained by meansof personal interviews (i.e., experiments). Therisk-attitude measures obtained by the scal-ing framework can be measured by a mailquestionnaire and hence are relatively inex-pensive tools for researchers. Recent devel-opments in statistics and psychometrics innew methods that take measurement errorinto account and that are able to modelhigher-order factor models provide an oppor-tunity and a challenge to identify and developa GRAC that is based solely on the scal-ing framework. However, the likelihood ofsuch a breakthrough may not be large, sinceour study shows that it is precisely the mea-sures derived from the expected utility frame-work that contribute most to the global risk-attitude construct.

[Received January 2000;accepted January 2001.]

References

Antle, J.M. “Econometric Estimation of Produc-ers’ Risk Attitudes.” Amer. J. Agr. Econ.69(August 1987):509–22.

Arrow, K.J. Essays in the Theory of Risk Bearing.Chicago: Markham, 1971.

Bagozzi, R.P., Y. Yi, and D. Nassen. “Represen-tation of Measurement Error in MarketingVariables: Review of Approaches and Exten-sion to Three-Facet Designs.” J. Econometrics89(March/April 1999):393–421.

Barry, P.J., L.J. Robison, and G.V. Nartea, “Chang-ing Time Attitudes in Intertemporal Anal-ysis.” Amer. J. Agr. Econ. 78(November1996):972–81.

Bem, D.J. “Self Perception: An Alternative Inter-pretation of Cognitive Phenomena.”Psych.Rev. 74(May 1967):183–200.

Benson, J., and D.L. Bandalos. “Second-OrderConfirmatory Factor Analysis of the Reac-tions to Tests Scale with Cross-Validation.”Multiv. Beh. Res. 27(3 1992):459–87.

Binswanger, H.P. “Empirical Estimation and Useof Risk Preferences: Discussion.” Amer. J.Agr. Econ. 64(May 1982):391–93.

Bollen, K.A. Structural Equations with Latent Vari-ables. New York: John Wiley & Sons, 1989.

Brockhaus, R.H. “Risk-Taking Propensity ofEntrepreneurs.” Academy Manage. J.23(September 1980):509–20.

Browne, M.W., and R. Cudeck. “Single Sam-ple Cross-Validation Indices for CovarianceStructures.” Multiv. Beh. Res. 24(October1986):445–55.

Chavas, J.P., and M.T. Holt. “Acreage DecisionsUnder Risk: The Case of Corn and Soybeans.”Amer. J. Agr. Econ. 72(August 1990):529–38.

Churchill, G.A. Marketing Research: Methodolog-ical Foundations. Chicago: The Dryden Press,1995.

Dyer, J.S., and R.K. Sarin. “Relative Risk Aver-sion.” Manage. Sci. 28(August 1982):875–86.

Ellsberg, D. “Classic and Current Notions ofMeasurable Utility.” Econ. J. 64(September1954):528–56.

Fishburn, P.C. “Transitive Measurable Utility.” J.Econ. Theory 31(December 1983):293–317.

Fishburn, P.C., and G.A. Kochenberger. “Two-piece von Neumann-Morgenstern UtilityFunctions.” Decisions Sci. 10(October1979):503–18.

Geczy, C., B.A. Minton, and C. Schrand. “WhyFirms Use Currency Derivatives.” J. Finan.52(September 1997):1323–54.

Goodwin, B.K., and T.C. Schroeder. “HumanCapital, Producer Education Programs, andthe Adoption of Forward Pricing Methods.”Amer. J. Agr. Econ. 76(November 1994):936–47.

Guttman, L.A. “Basis for Scaling QuantitativeData.” Amer. Soc. Rev. 9(April 1944):139–50.

Han, J.K., N. Kim, and R.K. Srivastava. “MarketOrientation and Organizational Performance:

Pennings and Garcia Measuring Producers’ Risk Preferences 1007

Is Innovation the Missing Link?” J. Mktg.62(October 1998):30–45.

Hair, J.F., R.E. Anderson, R.L. Tanham, and W.C.Black.Multivariate Data Analysis. EnglewoodCliffs NJ: Prentice-Hall, 1995.

Hershey, J.C., H.C. Kunreuther, and P. Schoe-maker. “Sources of Bias in Assessment Pro-cedures for Utility Functions.” Manage. Sci.28(August 1982):936–54.

Jaworski, B.J., and A.K. Kohli. “Market Orien-tation: Antecedents and Consequences.” J.Mktg. 57(July 1993):53–70.

Johnson, L. “The Theory of Hedging and Specula-tion in Commodity Futures.” Rev. Econ. Stud.27(June 1960):139–51.

Joreskog, K.G., and D. Sorbom. LISREL 8: Struc-tural Equation Modeling with the SIMPLISCommand Language. Chicago: Scientific Soft-ware International, Inc., 1993.

Kahneman, D., and A. Tversky. “Prospect Theory:An Analysis of Decision Under Risk.”Econo-metrica 47(March 1979):263–91.

Keeney, R.L., and H. Raiffa. Decisions With Mul-tiple Objectives: Preferences and Value Trade-offs. New York: John Wiley & Sons, 1976.

Keller, L.R. “An Empirical Investigation ofRelative Risk Aversion.” IEEE Trans.Systems Man Cybernetics 15(July/August1985):475–82.

King, R.P., and L.J. Robison. “An IntervalApproach to Measuring Decision Maker Pref-erences.” Amer. J. Agr. Econ. 63(August1981):510–20.

Koski, J.L., and J. Pontiff. “How Are Deriva-tives Used? Evidence from the MutualFund Industry.” J. Finan. 54(April 1999):791–816.

Krzysztofowicz, R. “Strength of Preference andRisk Attitude in Utility Measurement,” Org.Behavior Human Performance 31(February1983):88–113.

Kunreuther, H.C., and R. Ginsberg.Disaster Insur-ance Protection: Public Policy Lessons. NewYork: John Wiley & Sons, 1978.

Lence, S.H. “Using Consumption and AssetReturn Data to Estimate Farmers’ Time Pref-erences and Risk Attitudes.” Amer. J. Agr.Econ. 82(November 2000):934–47.

Likert, R. “A Technique for the Measurement ofAttitudes.” Archives of Psych. 140(1932).

MacCrimmon, K.R., and D.A. Wehrung. “Charac-teristics of Risk-Taking Executives.” Manage.Sci. 36(April 1990):422–35.

March, J., and Z. Shapira. “Variables Risk Refer-ences and the Focus of Attention.”Psych. Rev.99(January 1992):172–83.

Marsh, H.W., and D. Hocevar. “Application ofConfirmatory Factor Analysis to the Study ofSelf-Concept: First- and Higher-Order FactorModels and Their Invariance across Groups.”Psych. Bull. 97(May 1985):562–82.

Mian, S.L. “Evidence on Corporate Hedging Pol-icy.” J. Finan. Quant. Anal. 31(September1996):419–39.

Miller, D., M.F.R. Kets de Vries, and J. Toulouse.“Top Executive Locus of Control and itsRelationship to Strategy-Making, Structure,and Environment.” Academy Manage. J.25(June 1982):237–53.

Moscardi, E., and A. de Janvry. “Attitudestoward Risk among Peasants: An Econo-metric Approach.” Amer. J. Agr. Econ.59(November 1977):710–16.

Nance, D.R., C.W. Smith, Jr., and C.W. Smithson.“On the Determinants of Corporate Hedg-ing.” J. Finan. 48(March 1993):267–84.

Nunnally, J.C., and I.H. Bernstein. PsychometricTheory, 3rd. ed. New York: McGraw-Hill, 1994.

Payne, J.W. “The Scarecrow’s Search: A CognitivePsychologist’s Perspective On OrganizationalDecision Making.” Organizational DecisionMaking. Z. Shapira, ed., pp. 353–74. New York:Cambridge University Press, 1997.

Pennings, J.M.E., and A. Smidts. “Assessing theConstruct Validity of Risk Attitude.” Manage.Sci. 46(October 2000):1337–48.

Pennings, J.M.E., and R.M. Leuthold. “The Roleof Farmers’ Behavioral Attitudes and Hetero-geneity in Futures Contracts Usage.” Amer. J.Agr. Econ. 82(November 2000):908–19.

Pratt, J.W. “Risk Aversion in the Small and inthe Large.” Econometrica 32 (January/April1964):122–36.

Robison, L.J. “An Appraisal of Expected Util-ity Hypothesis Tests Constructed fromResponses to Hypothetical Questions andExperimental Choices.” Amer. J. Agr. Econ.64(May 1982):366–75.

Robison, L.J., and P.J. Barry. The CompetitiveFirm’s Response to Risk. New York: MacMil-lan, 1987.

Robison, L.J., P.J. Barry, J.B. Kliebenstein, and G.F.Patrick. “Risk Attitudes: Concepts and Mea-surement Approaches.” Risk Management inAgriculture. Peter J. Barry, ed. Ames IA: IowaState University Press, 1984.

Roe, T. “Empirical Estimation and Use of RiskPreference: Discussion.” Amer. J. Agr. Econ.64(May 1982):394–96.

Saha, A., C.R. Shumway, and H. Talpaz. “JointEstimation of Risk-Preference Structure andTechnology Using Expo-Power Utility.”Amer.J. Agr. Econ. 76(May 1994):173–84.

1008 November 2001 Amer. J. Agr. Econ.

Shapira, Z. Organizational Decision Making. NewYork: Cambridge University Press, 1997.

Shapiro, B.I., and B.W. Brorsen. “Factors Affect-ing Farmers’ Hedging Decisions.” N.C.J. Agr.Econ. 10(July 1988):145–53.

Smidts, A. “The Relationship Between Risk Atti-tude and Strength of Preference: A Testof Intrinsic Risk Attitude.” Manage. Sci.43(March 1997):357–70.

Steenkamp, J-B.E.M., and H.C.M. van Trijp. “TheUse of LISREL in Validating MarketingConstructs.” Int. J. Res. Mktg. 8(November1991):283–99.

Stein, J.L. “The Simultaneous Determination ofSpot and Futures Prices.” Amer. Econ. Rev.51(December 1961):1012–25.

Thurstone, L.L., and E. Chove. The Measurementof Attitude. Chicago: University of ChicagoPress, 1929.

Tsiang, S.C. “The Rationale of the Mean-StandardDeviation Analysis, Skewness Preference andDemand for Money.” Amer. Econ. Rev.64(June 1972): 354–71.

Turvey, C.G., and T.G. Baker. “A Farm-LevelFinancial Analysis of Farmers’ Use of Futuresand Options under Alternative Farm Pro-grams.” Amer. J. Agr. Econ. 72(November1990):946–57.

Tversky, A., and P. Wakker. “Risk Attitudes andDecision Weights.” Econometrica 63(Novem-ber 1995):1255–80.

Tversky, A., S. Sattath, and P. Slovic. “ContingentWeighting in Judgement and Choice.” Psych.Rev. 95(July 1988):371–84.

van den Putte, B., J. Hoogstraten, and R. Meertens.“A Comparison of Behavioral AlternativeModels in the Context of the Theory of Rea-soned Action.”British J. Social Psych. 35(June1996):257–66.

von Neumann, J., and O. Morgenstern. Theoryof Games and Economic Behavior, 2nd ed.Princeton: Princeton University Press, 1947.

Weber, E.U., and R.A. Milliman. “PerceivedRisk Attitudes: Relating Risk Perception toRisky Choice.” Manage. Sci. 43(February1997):123–44.

Appendix

The appendix gives the specification of the first-order factor model, the second-order factor model,and the structural equation model.

First-order factor model. Let y be a p ∗1 vectorof observed variables (i.e., items), let η a n∗1 vec-tor of latent (i.e., unobserved) variables or factors

underlying the observed variables, and let ε be ap ∗ 1 vector of error variables. The measurementmodel, representing the relationship between theitems and factors, can be expressed as

y = #yη+ ε(A.1)

where it is assumed that that η′s and ε′s are ran-dom variables with zero means, y is measured indeviations from its means, and the ε′s are uncor-related with the η′s. #y is a p ∗ n matrix of par-tial regression coefficients for the regression of yon η, commonly referred to as factor loadings. Theimplied covariance matrix of y can be expressedas

$ = #y%#′y +&ε(A.2)

where % is the covariance matrix of η and &ε isthe covariance matrix for ε.

For a given specification, and a given identifi-cation of the parameters in #y , % , and &ε, maxi-mum likelihood procedures have been developedfor estimation, and various goodness-of-fit mea-sures are available for evaluating these models(Bagozzi, Yi, and Nassen; Joreskog and Sorbom).

Second-order factor model. In the second-orderfactor model, the q ∗ 1 vector of second-order fac-tors is represented by ξ, the first-order factors bythe m ∗ 1 vector η, and the observed variablesby the p ∗ 1 vector y . The p ∗ m matrix #y con-tains the loadings of the observed variables onthe first-order factors, and * contains the loadingsof the first–order factors on the second-order fac-tors. The covariance matrix of the second-orderfactors is represented by +. The vector of resid-ual variables in the first-order factors is repre-sented by ζ, the unique factors (i.e., error terms)in the observed variables are represented by ε,the variance–covariance matrices of residuals andunique factors are denoted % and &ε, respectively.The relationship between the observed variables interms of the first-order factors can be expressed as

y = #yη+ ε(A.3)

with the first-order factors in terms of the second-order factor as

η = *ξ+ ζ�(A.4)

This second-order factor model in (A.3) and(A.4) hypothesizes a general (second-order) risk-attitude construct, based on the specific risk-attitude measures identified in the confirmatoryanalysis.

The implied variance–covariance matrix of thesecond-order factor model can be expressed as

$ =[#y(*+*

′ +%)+&ε #y*++* ′#′

y +

]�(A.5)

Pennings and Garcia Measuring Producers’ Risk Preferences 1009

Equation (A.5) relates the variances and covari-ances of the observed variables to the parametersof the model. Estimation of the second-order fac-tor model involves finding values for the parame-ter matrices that produce an estimate of $ accord-ing to equation (A.5) that is as close as possible tothe sample matrix S (e.g., the covariance matrix ofthe raw data). The covariance structure in equa-tion (A.5) can be estimated by one of the fullinformation methods: unweighted least squares,generalized least squares, and maximum likelihood(Bollen). The fitting function measures show howclose a given $ is to the sample covariance matrix

S. Because of its attractive statistical properties weuse the maximum likelihood procedure (Bollen).

Structural equation model. The structural equa-tion model resembles (A.3) and (A.4) but adds therelationship between the GRAC and the farmers’intentions to use futures which is reflected as

η = Bη+ *ξ+ ς(A.6)

where B is a (1 ∗ 1) matrix of the coefficient relat-ing the GRAC to the measure of farmers’ inten-tions to use futures.