split-sample tests of ‘‘no opinion’’ responses in an...

Split-Sample Tests of ‘‘No Opinion’’ Responses in anAttribute-Based Choice Model

Eli P. Fenichel, Frank Lupi, John P. Hoehn, and Michael D. Kaplowitz

ABSTRACT. Researchers using questionnaires toelicit preferences must decide whether to includeresponse options that allow respondents to express‘‘no opinion.’’ Using a split-sample design, weexplore the implications of alternative answerformats including and not including no-opinionresponses in an attribute-based choice experiment.The results indicate that using multiple no-opinionresponses may enable researchers to differentiatebetween respondents who choose no-opinion optionsdue to satisficing and those expressing utilityindifference. Existing literature suggesting no-opin-ion responses may be treated as ‘‘no,’’ but our resultsshow treating no-opinion responses as no yieldssubstantially disparate preference estimates. (JELC25, Q24, Q25, Q51)

I. INTRODUCTION

In surveys eliciting stated preferences,some respondents opt to answer a choicequestion with a response such as ‘‘don’tknow,’’ ‘‘not sure,’’ or ‘‘would not vote.’’These responses are variants of the ‘‘noopinion’’ responses discussed in the broadersurvey research literature (Krosnick 2002).Treatment of no-opinion responses instated preference studies has largely focusedon studies that use the contingent valuationmethod (CVM). The attribute-based meth-od (ABM), also called choice experimentsor stated choice, is a comparatively newtechnique that is related to CVM (Louviere,Hensher, and Swait 2000; Foster andMourato 2003; Holmes and Adamowicz2003). The ABM presents respondents witha set of attributes of a good, where typicallyone attribute is price. The attributes andprices are varied across respondents. Thisdiffers from CVM, where typically only

price is varied across respondents. ABM,therefore, allows the researcher to value theimplicit price for each attribute, much like ahedonic price study (Holmes and Adamo-wicz 2003). Typically, CVM and ABM usediscrete choice responses, and as a resultrandom utility models can be used in theestimation of both methods. Indeed, CVMmay be considered a special case of ABM(Boxall et al. 1996). Although some re-search has addressed treatment of no-opinion responses in CVM studies, thispaper represents a first effort to understandthe significance of no-opinion responses inABM.

In many ABM-based studies, respon-dents are asked to choose between two ormore attribute-price sets. This is similar tothe referendum-style questions commonlyused in CVM, especially in the case whereone attribute-price set is treated as a statusquo. The National Oceanic and Atmo-spheric Administration (NOAA) panel rec-ommended including a ‘‘no vote’’ option forbinary choice CVM studies (Arrow et al.1993). While, this recommendation hasspawned a growing body of research onhow to treat ‘‘would not vote’’ and othertypes of no-opinion responses in the CVMliterature, the issue has received less atten-tion in ABM studies. The increasing use of

Land Economics leco-85-02-09.3d 5/1/09 17:36:13 349

The authors are, respectively, research associate,Quantitative Fisheries Center, Department of Fisheriesand Wildlife; associate professor of Agricultural, Food,and Resource Economics and Fisheries and Wildlife;professor of Agricultural, Food, and Resource Econom-ics; and associate professor of Community, Agriculture,Recreation, and Resource Studies. All are at MichiganState University. The research reported here was sup-ported in part by the Science to Achieve Results (STAR)Program, U.S. Environmental Protection Agency, andthe Quantitative Fisheries Center at Michigan StateUniversity. This is contribution number 2009-01 of theQuantitative Fisheries Center. The authors are solelyresponsible for any errors.

Land Economics N May 2009 N 85 (2): 349–363ISSN 0023-7639; E-ISSN 1543-8325

E 2009 by the Board of Regents of theUniversity of Wisconsin System

ABM studies for policy decisions andvaluation (e.g., Birol, Karousakis, andKoundouri 2006) motivates a more carefulexamination of the use and analysis of no-opinion responses in ABM studies. Thisseems especially relevant since meta-analy-ses have shown that methodology can affectwelfare estimates (Brander, Florax, andVermatt 2006; Johnston et al. 2006). There-fore, researchers’ design and treatment ofno-opinion responses, along with broadermethodological design, must be consideredif ABM welfare results are to be used forbenefit transfer.

There is growing evidence in the CVMbinary choice literature that no-opinionresponses should not be treated as ‘‘for’’votes (Groothuis and Whitehead 2002;Caudill and Groothuis 2005; Carson et al.1998). However, there is not yet agreementas to whether no-opinion responses shouldbe treated conservatively as ‘‘against’’ votes(Carson et al. 1998; Krosnick 2002), orwhether no-opinion responses may repre-sent cognitive difficulties, potentially result-ing from utility indifference, and thereforeshould be treated as a truly unique response(Krosnick et al. 2002; Evans, Flores, andBoyle 2003; Alberini, Boyle, and Walsh2003; Caudill and Groothuis 2005; Champ,Alberini, and Correas 2005). Furthermore,those who support treating no-opinionresponses in CVM as unique responsesseem to largely base their recommendationon improving econometric efficiency, withfew researchers arguing that treating suchresponses as ‘‘against votes’’ (the conserva-tive approach) yields inconsistent estimates.Groothuis and Whitehead (2002) observethat treating no-opinion responses in CVMas unique or ‘‘against’’ votes may depend onwhether the study is attempting to measurewillingness to pay (WTP) or willingness toaccept (WTA).

Arguments for treating no-opinion re-sponses as unique responses in estimationare often based on Wang’s (1997) hypoth-eses on why respondents choose no-opinionresponses. Wang (1997) posits that there arefour general categories of respondents whochoose no-opinion responses: (1) those who

reject the CVM scenario, (2) those whoknow their preference and decline toanswer, (3) those who make an effort andare truly unsure, and (4) those who do notmake an effort and are therefore unsure.

Krosnick et al. (2002) also examine why arespondent may choose a no-opinion re-sponse. Their evidence suggests that no-opinion responses may result from satisfi-cing; that is, the ‘‘work’’ involved withanswering the question is so great that therespondent chooses a no-opinion responsebecause it involves the least work or thelowest risk for him or her.1 Krosnick et al.(2002) also discuss an alternative hypothesisfor a respondent’s no-opinion response:that a respondent’s optimizing processmay result in true indifference. In this case,the respondent is truly unsure when thechoices are close in terms of the associatednet benefits or welfare yields. Accordingly,a respondent may reply with a no-opinionresponse because he or she is indifferent in autility sense. However, it is unlikely thatthere is a clear line between no-opinionresponses resulting from optimizing andfrom satisficing, since a respondent maybegin optimizing but may ‘‘give up’’ beforereaching true indifference.

Investigations by Alberini, Boyle, andWelsh (2003), Caudill and Groothuis(2005), and Evans, Flores, and Boyle(2003) have been aimed at improvingestimation efficiency through sorting no-opinion responses into response categoriesfor estimation purposes. This sorting hasfocused on identifying and making use ofresponses that seem to fall into Wang’s(1997) latter two unsure categories or thatmay be cases of optimizing as per Krosnicket al. (2002). However, there has been littleeffort to sort no-opinion responses that maybe the result of other phenomena, such asno-opinion responses that result fromrespondents being unsure because of utilityindifference or satisficing. Moreover, mostof the work has been focused on ordinal


1 Work requirements may range from reading thesurvey and understanding the question to evaluatingpreferences.

350 Land Economics May 2009

polychotomous-choice and multiboundedquestions, which introduce other types ofdifficulties (Vossler and Poe 2005). Theliterature on ABM does contain a related,but logically distinct, strain of research. Insome ABM studies, respondents are pre-sented with sets of alternatives composed ofvaried attributes and an answer set thatincludes a ‘‘none’’ alternative (Louviere,Hensher, and Swait 2000) or an ‘‘opt out’’alternative (Boxall et al. 1996). In the settingof a product choice, the none option mightbe treated as a ‘‘don’t buy’’ decision (e.g.,Birol, Karousakis, and Koundouri 2006).In a recreational site choice context, thenone option might represent a no-tripdecision, or it might represent a trip to asite not included in the choice set (Banzhaf,Johnson, and Mathews 2001). In othersettings, the none option may be considereda choice to maintain the status quo.Typically, researchers explicitly model thistype of alternative as one of the elements ina multinomial choice model (Birol, Karou-sakis, and Koundouri 2006). In contrast,here we consider a distinct case in the ABM,in which a failure of respondents to choose

an alternative is not a choice for the statusquo. Instead, we examine the setting inwhich respondents’ failure to choose one ofthe ABM alternatives is akin to a no-opinion response.

Some question also remains about thecomparability of ABM studies to CVMstudies (Stevens et al. 2000; Foster andMourato 2003). As a result of their explicitsubstitutes, ABM studies may be cognitive-ly more difficult than CVM studies and mayask respondents to explore their preferencesin more detail (Stevens et al. 2000). Alter-natively, CVM studies tend to be onedimensional (i.e., apart from any scopetests, all else is held constant for therespondent), while ABM studies involvemultidimensional trade-offs that may resultin a larger number of respondents whohonestly don’t know or are closer toindifference relative to CVM. The propor-tion of no-opinion responses may be furtherexacerbated in ABM studies by the com-mon practice of using a main-effects attri-bute design (e.g., Birol, Karousakis, andKoundouri 2006). Such designs can im-prove econometric efficiency but may placerespondents close to utility indifference fora large share of the alternatives in thedesign, thereby increasing cognitive diffi-culty. Figure 1 illustrates the point thatwhen using a main-effects design, fewattribute combinations would arise in thegray area that represents alternatives wherethe levels of each attribute are strictly better(or worse) than those of another designalternative. Choices among alternativeswith strictly better (worse) attributes shouldbe cognitively easy. Instead, many ABMswill involve more difficult comparisons ofalternatives with more subtle attribute leveldifferences (some better, some worse),making indifference more likely.

If the meaning of a no-opinion responseis unclear to the researcher, then increasedeconometric benefits from categorizing orsorting those no-response options will belost. There have been no studies examiningreclassifying no-opinion responses in ABMas ‘‘no’’ responses, which is considered aconservative classification in CVM. That is,


FIGURE 1NONTRIVIAL ATTRIBUTE COMPARISONS LIKE THOSE

COMMON IN MAIN-EFFECTS DESIGNS: OPEN CIRCLE,BASELINE SCENARIO; SOLID CIRCLES, SAMPLE OF THE

POSSIBLE ALTERNATIVES THAT ARE NOT UNIFORMLY

BETTER OR WORSE IN THEIR INDIVIDUAL ATTRIBUTES;GRAY AREAS, ATTRIBUTE COMBINATIONS THAT ARE

BETTER (WORSE) IN ALL ATTRIBUTES

85(2) Fenichel et al.: Split-Sample Tests of ‘‘No Opinion’’ Responses 351

it remains to be tested whether treating no-opinion responses in ABM as ‘‘no’’ re-sponses generates estimates that are consis-tent with similar surveys where a no-opinion option is not offered. Therefore,this paper examines two research questionsconcerning no-opinion responses in ABMstudies. First, does recoding no-opinionresponses as ‘‘against’’ (our ‘‘no’’ responsechoice) provide estimates consistent withthose derived from a survey instrument thatwas the same in every respect expect for anabsence of any no-opinion response cate-gory? Secondly, does offering respondentstwo qualitatively different no-opinion re-sponses allow expressions of welfare indif-ference to be sorted from expressions of noopinion for other reasons? This latter issuemay be generalizable to CVM because itattempts to distinguish Wang’s (1997) thirdtype of response (indifferent or too close tocall) from Krosnick et al.’s (2002) satisficingor other variants of no opinion.

II. SURVEY INFORMATION

A binary choice ABM survey was imple-mented using a web-based method with asplit-sample design. In addition to the usualexperimental design of the attributes, therewere four unique versions of the ABMsurvey that differed in the response options(treatments) respondents received for theirchoice questions:

1. ‘‘Yes’’ and ‘‘no’’ [yes/no treatment]2. ‘‘Yes,’’ ‘‘no,’’ and ‘‘too close to call’’

(TCC) [TCC treatment]3. ‘‘Yes,’’ ‘‘no,’’ and ‘‘not sure’’ (NS) [NS

treatment]4. ‘‘Yes,’’ ‘‘no,’’ TCC, and NS [all-op-

tions treatment]

For each treatment, the last expression, insquare brackets, is what each treatment iscalled in this paper. Collectively, the NS andTCC responses are referred to as no-opinionresponses as a shorthand to refer to respon-dents who did not explicitly choose yes or noin the choice scenario. The TCC response ishypothesized to reflect choice situations forrespondents close to utility indifference.

The web-based ABM survey elicitedpreferences for inland, freshwater wetlandmitigation from adult Michigan residents.Access to potential respondents was pur-chased from Survey Sampling International(SSI), which maintains a web-based panelof individuals who have agreed to completean occasional online questionnaire in returnfor participation in a prize lottery. The SSIpanel is a sample of potential surveyrespondents with known demographiccharacteristics. The response treatmentswere randomly assigned to respondents toprevent systematic differences across treat-ment response groups.2 Similarly, all four ofthe response treatments utilized the sameexperimental design for the survey attri-butes. E-mail invitations to the SSI panel ofpotential respondents resulted in 3,454visits to the welcome page of the web-basedquestionnaire. Of those potential respon-dents who visited the survey’s welcomepage, 58% completed at least one wetlandsmitigation choice question and providedsome demographic information. All togeth-er, there were 9,753 usable choice questionresponses for the reported analyses.

The questionnaire was developed using aseries of focus groups and pretest interviews(described by Kaplowitz, Lupi, and Hoehn2004) and used a policy setting and set ofchoice questions that follow the paperinstrument discussed by Lupi, Kaplowitz,and Hoehn (2002). Each respondent waspresented with the characteristics of acommon wetland that had already beenapproved for drainage (drained wetland)and the characteristics of a wetland beingproposed as compensation (restored wet-land) for the wetland to be drained. Theattributes of the wetlands presented torespondents were wetland type (wooded,marsh, mixed), size (acres), public accessattributes, and habitat attributes (these


2 The percentage of respondents living in a self-described rural region ranged from 28% to 30%. Averageincome ranged between $52,000 and $55,000 annually.Percentage of respondents with a college educationranged from 28% to 32%. Average age ranged from42.2 to 42.9 years, and percentage male ranged from 40%to 47%.


included habitat for songbirds, wadingbirds, amphibians and reptiles, and wildflowers). The respondents were told:

We need your opinion as a member of the citizens’panel on five restoration cases. The cases represent thekind of decisions that are made every day by wetlandrestorers. In each case, the project to drain a commontype of wetland has already been approved. The onlyquestion is whether quality and quantity of a restoredwetland is enough to make up for the loss of thedrained wetland. If the restored wetland is approved,the restoration project goes forward. If the restoredwetland is not approved, the restoration project goesback to the drawing board. A revised project will bereviewed by a different citizens’ panel.

In your opinion, is the restored wetland goodenough to offset the loss of the drained wetland?

The question was framed this way toaddress a tendency, found during surveydevelopment, for some respondents to voteagainst a project without weighing theattributes of the alternatives as a way ofblocking development and wetlands drain-age. Each respondent was asked to makefive wetland comparisons, but each respon-dent had the same response treatment(answer set) for each of his or her choicequestions. Further details of the surveydesign, administration, and general resultswere presented by Hoehn, Lupi, andKaplowitz (2004), and a sample instrumentis available upon request from the authors.

III. RESPONSE FREQUENCY ANALYSIS

The study design included four differentresponse treatments. Response statistics forthe completed choice questions for eachtreatment are presented in Table 1.3 Asexpected, the all-options response treatmentresulted in the highest proportion of no-opinion responses (25%). Chi-square testswere used to compare the proportion of no-opinion responses across the three surveyresponse treatments that offered no-opinionresponses (Table 2).4 Table 2 shows thatthe probability of a no-opinion response issignificantly different when all four re-sponse options are presented to respon-dents, as compared to instances in whichonly one type of no-opinion response isavailable to respondents. This is true at allcommon significance levels. These resultsimply that respondents are more likely tochoose a no-opinion response when boththe TCC and NS options are available. Achi-square test comparing the TCC re-sponse treatment and the NS responsetreatment yielded a low p-value (,0.016).


TABLE 1

RESPONSE DATA

Survey Version/ ResponseTreatment Total Responses

Proportion of

Yes No TCC NS No Opinion

Yes/no 1,586 0.590 0.410 – – –TCC 1,683 0.537 0.272 0.191 – 0.191NS 1,619 0.553 0.288 – 0.159 0.159All options 3,000 0.467 0.287 0.164 0.082 0.246

Note: TCC, too close to call; NS, not sure.

TABLE 2

CHI-SQUARE TESTS COMPARING PROPORTION OF NO-OPINION RESPONSES

NS vs. TCC All Options vs. NS All Options vs. TCC

x2 statistic 5.8360 47.1749 18.3050p-value 0.0157 0.0000 0.0000


3 A total of 4,865 responses were received for the all-options treatment, but 1,865 of these observation wererandomly selected and reserved for later use in assessingthe model predictions.

4 All chi-square tests use the Yates correction forvariables coming from a binomial distribution (Zar 1996).


This result suggests that the TCC and NSresponse options are not viewed as equiv-alent response options by respondents, andindicates that the wording of no-opinionoptions may matter.

In addition to the four response treat-ments, there were two different data codingtreatments, one treating no-opinion re-sponses as no, and the other dropping no-opinion responses from the data set. As aresult, we ended up with seven sets of choicedata with different yes/no ratios (Table 3).The ratio of yes to no responses isimportant because it affects the estimatedparameters, especially the constant term, incommon methods for implementing ran-dom utility models (e.g., logit models).

Carson et al. (1998) used chi-square teststo determine the effect of no-opinionresponses on the proportion of yes to noresponses in a CVM study. We conducted asimilar analysis for the ABM data (Table 4section A). The proportion of yes to noresponses was significantly different, at the95% confidence level, between surveys

treatments that did not allow respondentsto express ‘‘no opinion’’ and those thatoffered either TCC or NS as responseoptions. A chi-square analysis comparingthe ratio of yes to no in the yes/no responsetreatment to the all-options response treat-ment yielded a difference with a p-value of0.07. This difference may not be significantat the traditional 95% confidence level butmay yield different economic results; inother words, the yes’s and no’s from thesetwo groups may produce different estimatesof WTA.

No-opinion responses were then pooledwith no responses, and the yes/no ratio wasretested against the yes/no ratio from theyes/no treatment (Table 4 section B). Allchi-square tests for all of these comparisonsyielded significant differences (p , 0.05).This result implies that pooling no-opinionresponses with no responses, as suggestedby Carson et al. (1998), results in signifi-cantly different yes/no ratios and contrastswith the findings of Carson et al. (1998) forCVM. It remains unclear in the all-options


TABLE 3

RATIO OF YES TO NO RESPONSES BY RESPONSE TREATMENT

Survey Version/ Response Treatment

Ratio of Yes to No

No Opinion, Dropped No Opinion, Pooled with No

Yes/no 1.44 –TCC 1.97 1.16NS 1.92 1.24All options 1.63 0.88


TABLE 4

CHI-SQUARE RESULTS FOR COMPARING YES/NO RATIOS

Comparison Yes/No and TCC Yes/No and NS Yes/No and All Options

A. The ratio of yes to no responses with no-opinion responses dropped compared to yes/no treatment


B. The ratio of yes to no responses with no-opinion responses coded as no compared to yes/no treatment


C. The ratio of yes to no responses compared between no-opinion treatments (no-opinion responses dropped)




response treatment whether both TCC andNS pulled equally from the yes and the noresponses. We also compared yes/no ratiosafter dropping the no-opinion responsesfrom the data set across the three no-opinion survey treatments (Table 4 sectionC).

The ratio of yes to no responses did notchange significantly when TCC or NS wasoffered as the no-opinion response option.The ratio of yes to no responses in the all-options survey treatment was significantlydifferent at the 95% confidence level com-pared to the ratio of yes to no responseswhen only one no-opinion response optionwas presented (TCC and NS survey treat-ments). When more than one no-opinionoption was presented, the ratio of yes to noresponses differed significantly.

These results indicate that survey partic-ipants are influenced by the phrasing,language, or number of no-opinion re-sponse items, supporting the hypothesisthat various no-opinion responses mayrepresent unique types of responses. Fur-ther, these results suggest that no-opinionresponses do not pull evenly from the yesand the no responses and that, unlikeCarson et al.’s (1998) CVM study, theseresponses do not consistently pull from theno responses. It does appear, in thisinstance, that no-opinion responses pullsome from the yes responses and moreheavily from the no responses (see Table 1).Moreover, no-opinion responses seem topull more evenly from the yes and the noresponses when both TCC and NS werepresented as options as opposed to whenonly one type of no-opinion response wasavailable (Table 1). It appears that themarginal impact of adding a second no-opinion response option is to pull moreresponses from the yes than the no respons-es, even though the first no-opinion re-sponse option appears to draw more fromthe no than the yes responses.

Three potential scenarios seem to explainthe apparent divergence in results betweenthis ABM study and previous CVM studies.First, the underlying ABM study focuseson respondents’ WTA compensation (Gro-

othuis and Whitehead 2002) as measuredby in-kind trade-offs. Second, there may besomething unique to the ABM responseformat that is different from CVM studies.Thirdly, it is possible that the additional no-opinion response option causes responses topull more evenly from both the yes and theno pools. The response ratios obtainedimply that TCC and NS responses are goodsubstitutes when only one response optionis available to respondents. When bothTCC and NS are present, the results appearto support the hypothesis that a TCCresponse may involve, perhaps, an attemptby respondents to optimize.

It is possible that one effect of includingno-opinion response options would be tolower item nonresponse rates for the choicequestion. The item nonresponse rates forthe survey treatments were yes/no (0.7%),TCC (1.5%), NS (0.4%), and all options(0.8%). These differences are statisticallysignificant (p , 0.0001). Interestingly, theTCC item nonresponse rate is higher thanthat of the NS treatment. This resultprovides some further evidence that theTCC response option is being used byrespondents differently than the NS re-sponse option. Overall though, none of thetreatments had substantial levels of itemnonresponse.

Response ratios provide some informa-tion about no-opinion responses in ABMstudies. But the ultimate purpose of ABMstudies is to examine welfare estimates andpreference orderings. Therefore, in the nextsection we test whether different questionresponse treatments or data coding treat-ments result in estimates that imply differ-ent compensation measures or differentpreference ordering.

IV. EFFECTS ON WELFARE

The wetlands mitigation survey used inthis study asked respondents to make an in-kind trade-off between acres of drained andrestored wetlands. In essence, respondentswere asked if the restored wetland, withspecified attributes, would compensate forthe loss of an existing wetland, a context



that mirrors the wetland permitting andmitigation process in the United States.Therefore, wetland acres are the unit ofcurrency for this analysis. The drained andrestored wetlands presented to respondentsdiffered in their attributes. The variouswetland quality attributes included in thechoice sets act to shift demand for wetlandacres. Responses were coded into 11 re-sponse variables. These variables includedchange in wetland acreage, dummy vari-ables for capturing changes in wetlands’general vegetative structure, public access,and habitat conditions for amphibians,songbirds, wading birds, and wildflowers(changes could between poor and good, orgood and excellent). Changes in wetlandacres were recorded as the change in thetotal number of acres. Dummy variableswere coded as 1 for a positive change, 0 forno change, and 21 for a negative change.Changes from poor to excellent wereindicated by both the poor to good andgood to excellent dummy variables beingcoded as 1 (other coding followed thispattern). A change from no access to accesswas coded as 1 (21 for the other direction),and changes in wetland type were coded as1 for a change.

In-kind welfare measures were estimatedusing random utility theory (Holmes andAdamowicz 2003). The seven data sets(combinations of each survey and datacoding treatment) were fit using a randomeffects logit model that addressed the paneldata. The random effects logit, estimated bymaximum likelihood, fits parameter valuesso that the probability of response i beingyes (indicated by 1 and meaning that therestored wetland offset the loss of thedrained wetland in comparison i), fromrespondent j is given by

Prob yesi ~ 1ð Þ~L a z bð ACRESrestoredð { ACRESdrainedÞ

z cDXi z vj

�,

where L is the logistic cumulative distribu-tion function; a and b are scalar estimatedparameters; c is an estimated vector ofparameters corresponding to changes in

wetlands attributes,DXi; andvj is a randomeffect that accounts for the lack of indepen-dence among responses provided by thesame respondent.All seven models fit the data, with log-

likelihood ratio tests against a model with asingle choice dummy being significant at allcommon significance levels. Each modelincluded all variables, though not allestimated coefficients were significantlydifferent from zero at the 95% confidencelevel. In all models the parameter associatedwith the change in wetland acres wassignificantly greater than zero at the 95%confidence level.Parameter ratios were used to calculate

in-kind marginal implicit prices in acres ofrestored wetland (Table 5). For the con-stant term, the marginal implicit price is thechange in acres required to maintain thesame level of utility, ceteris paribus. That is,one acre of restored wetland would ade-quately compensate for one acre of drainedwetland if the WTA estimate were zero. Incases where only yes and no options werepresented, a restored wetland could have upto just over eight fewer acres for each acreof the drained wetland, ceteris paribus,before respondents would prefer thedrained wetland (Table 5, first row).This counterintuitive result for the con-

stant term is likely a consequence of thelinear form for the utility function andperhaps a result of the incentive propertiesof the question. The linear form of theutility function is the most common form inthe literature, and hence it is most relevantfor our comparison of response options.However, when one alters the underlyingresponses in a way that increases the shareof yes responses, then all else equal, itlowers the implied compensation. It is wellknown in discrete-choice contingent valua-tion that the linear form does not ensureestimated welfare measures are positive andcan lead to counterintuitive results (Haaband McConnell 2002). Additionally, thequestion format may have encouragedrespondents to accept fewer acres to ensurethat some type of restoration project wasapproved as opposed to not being certain of



getting restoration. That is, it is possiblethat some respondents perceived a no voteas ‘‘dragging out’’ the process and reducingthe chances of any mitigation.

Examination of response patterns ofrespondents revealed that those who chosethe same response throughout the choiceexperiment (serial respondents) were mostlylikely to choose yes, supporting this hy-pothesis.5 The WTA estimates, ceterisparibus, varied greatly across responsetreatments. In treatments with no-opinionresponses, dropping the no-opinion re-sponses from the analysis yielded WTAestimates that were closest to those derivedfrom the yes/no format. The WTA esti-mates showed that substantially less com-pensation was demanded by respondents

when no-opinion responses were droppedas opposed to pooled with no. Recoding theno-opinion responses as no in the all-options response treatment makes the ratioof yes to no less than 1 (Table 1) and causesthe WTA to be positive, that is, one acredrained required more than one acrerestored to offset the loss.6

An important aspect of the ABM is thatit allows the relative importance of theattributes to be ranked. Recoding the datachanged the yes/no ratio and affected theconstants, as expected. However, recodingthe data should not affect the ranking ofattributes if no-opinion responses representsatisficing. To investigate if recoding affect-ed the relative importance of attributes, themarginal implicit prices associated witheach attribute variable (Table 5) wereranked from the largest marginal impact


TABLE 5

MARGINAL IMPLICIT PRICES OF ATTRIBUTES ASSOCIATED WITH THE WILLINGNESS-TO-ACCEPT (WTA) IN-KIND

ACRES COMPENSATION FOR DRAINED WETLANDS

Survey Type

Yes/No

TCC NS All Options

Handling of no opinionPooled

with No DiscardedPooled

with No DiscardedPooled

with No Discarded

WTA, ceteris paribus 28.42 22.54a 211.96 23.55a 211.56 4.81 25.16Change of wetland type 5.86 5.79 7.76 4.80 4.45 2.75 2.66a

Access 26.65 24.89 25.53 27.48 27.03 25.64 27.37Amphibian p R g 27.40 23.25a 24.02 28.78 29.44 27.32 26.74Songbird p R g 210.70 23.36 24.84 210.24 211.02 24.97 26.96Wading bird p R g 27.69 27.68 27.56 29.50 210.20 25.99 26.72Wild-flower p R g 23.79a 23.16a 22.44a 23.66a 23.44a 22.23a 21.55a

Amphibian g R e 25.49 24.39 23.95 26.92 26.79 25.24 25.37Songbird g R e 25.47 23.41 22.19a 21.89a 22.90 25.24 25.30Wading bird g R e 25.49 22.75 23.57 24.53 24.98 23.37 23.50Wild-flower g R e 22.27a 24.12 24.52 24.26 23.81 21.53a 21.59a

Note: Values in this table represent the parameter estimate associated with the listed variable divided by the parameter for changes inarea, p R g 5 poor to good, and g R e 5 good to excellent. TCC, too close to call; NS, not sure.

a Marginal implicit prices that were not significantly different from zero at the 95% confidence level. Confidence intervals weredetermined by bootstrapping the dataset and reestimating the model 10,000 times for each model.

5 Serial respondents answered all five questions usingthe same response: in the yes/no survey responsetreatment 15.5% and 9.5% of respondents seriallyanswered yes and no, respectively; in the TTC surveyresponse treatment 11.9%, 2.7%, and 3.3% of respondentsserially answered yes, no, and TTC, respectively; in theNS survey response treatment 13.1%, 4.7%, and 4.4% ofrespondents serially answered yes, no, and NS, respec-tively; and in the all-options survey response treatment7.8%, 3.7%, 0.5%, and 2.8% of respondents seriallyanswered yes, no, TCC, and NS, respectively (about 4%of respondents evenly distributed across survey treat-ments answered fewer than five questions and were notconsidered to be serial responders).

6 Under maximum likelihood estimation in the logitmodel, the estimated constant parameter ensures that theaverage of the predicted probabilities of yes answersmatches the sample share of yes answers. To the extentthat the no-opinion answers are not being explained bythe other parameters of the models, the under therecoding of responses the newly estimated constantadjusts to match the new sample shares. Thus, recodingof no-opinion responses as no responses has the cleareffect of lowering the constant and hence lowering themarginal implicit prices.


to lowest marginal impact (Table 6).7 Themaximum difference in rank was used todetermine if the different response and datacoding treatments maintained consistentpreference ordering. Low maximum differ-ences indicate that all models order thepreference for a given attribute similarly.Changing wildflower habitat from poor togood had the lowest maximum difference inrank (excluding a change in wetland type,which increased the desired amount ofcompensation), though these marginal im-plicit prices were not significantly differentfrom zero at that 95% confidence level.Improving wading bird habitat from poorto good and increasing access consistentlyranked highly (median rank of 2 and 3, re-spectively) and had a low maximum differ-ence (maximum difference of 3 for both).Other attributes that generally ranked high-ly, however, did so less consistently, largelydue to TCC responses. For example, chang-es in songbird habitat from poor to goodalso had a high median rank, 2, but had amaximum difference in ranks of 5. This

difference is driven by the ranks associatedwith the TCC and all-options responsetreatment when the no-opinion responsesare pooled with no, thereby providingevidence that TCC response may notrepresent satisficing. In contrast, the attri-bute ranks for the NS format are identicalfor all attributes, and very similar to the yes/no format, regardless of whether the NS waspooled with no or dropped. The preferenceordering for the all-options treatment whenthe no-opinion responses are dropped is alsosimilar to the orderings for the yes/notreatment and both NS treatments.

Rank correlations between treatmentsprovide further evidence that TCC and NSresponses are not used interchangeably(Table 7). Preference ordering was general-ly correlated across all treatments except theresponse treatment TCC, with the recodingof TCC responses as no. Moreover, bothNS response treatments when compared tothe yes/no response treatment had correla-tion coefficients of 0.94—the highest corre-lation that yes/no had with any of theresponse treatments. The finding of thesimilarity of results for the NS models andthe yes/no treatments is similar to thefinding of Carson et al. (1998) for CVM.The results also provide some evidence thatNS by itself acts more like a no response,due to satisficing, than an expression ofindifference.


7 It is expected that marginal implicit price for apositive change in wetland quality would result in anegative implicit price, that is, less acres of restoredwetland would be need, if the restored wetland were ofhigher quality. Inspection of Table 5 reveals that all of thequality attributes had the expected signs, though not allwere significant.

TABLE 6

RANKING OF THE MARGINAL IMPLICIT PRICES AND THE MAXIMUM DIFFERENCE IN RANKING ACROSS MODELS;RANKINGS OF 1 HAD THE LARGEST MARGINAL EFFECT AND 10 HAD THE LOWEST MARGINAL EFFECT

Survey type

Yes/No

TCC NS All Options

Handling of noopinion

PooledwithNo Discarded



MedianRank

MaximumDifference

Change of wetlandtype

10 10 10 10 10 10 10 10 0

Access 4 2 2 4 4 3 1 3 3Amphibian p R g 3 7 5 3 3 1 3 3 6Songbird pR g 1 6 3 1 1 6 2 2 5Wading bird p R g 2 1 1 2 2 2 4 2 3Wild-flower p R g 8 8 8 8 8 8 9 8 1Amphibian g R e 6 3 6 5 5 4 5 5 3Songbird g R e 7 5 9 9 9 5 6 7 4Wading bird g R e 5 9 7 6 6 7 7 7 4Wild-flower g R e 9 4 4 7 7 9 8 7 5

Note: TCC, too close to call; NS, not sure; p R g 5 poor to good; g R e 5 good to excellent.


V. OUT-OF-SAMPLE FORECASTS

The evidence presented in the precedingsections indicates that the inclusion andhandling of no-opinion responses is notstraightforward and can substantiallyaffect valuation and preference orderingfrom an ABM. The evidence also suggeststhat inclusion and handling of the TCCresponse option is especially influential.To further investigate possible differencesbetween ‘‘too close to call’’ and ‘‘notsure,’’ we perform out-of-sample forecastsusing the reserved data (see note 2). Therewere 1,865 reserved responses from theall-options survey treatment withheldfrom the estimation data set. We useparameter estimates derived from themodel based on the yes/no responsetreatment to predict the probability of ayes response for the reserved data—datafor which we know the true response. Theforecasted probability of individuali providing a yes response is P̂Pi ~Pi yesjbyn,Xao

i

� �, where b

yn is the estimatedparameter vector from the yes/no modeland Xao

i is i’s vector of change in wetlandattributes. These probabilities were thensorted by known response to the all-response survey to estimate the mean andvariance of predicting a yes response from agiven type of response option.

If the model based on the yes/no treat-ment has the ability to discern yes from novotes, when applied to the reserved data,

then we would expect the mean predictedprobability of a yes given an actual yesresponse to be larger than the meanpredicted probability of a yes given anactual no response, that is,

Pnyes

i ~ 1

P̂Pi yes aoj

nyesw

Pnno

i ~ 1

P̂Pi no aoj

nno,

where nj is the number of actual responsesof type j and j ao is the response type actuallyreport from the all-options survey. Further-more, if respondents chose either TCC orNS as a result of an attempt to optimize butfound the welfare yield to be close to theirlevel of indifference, then we would expectthe mean predicted value associated withTCC and NS responses to lie between themean predicted value associated with yesand no responses.8 This is indeed the case,as shown in Table 8.


TABLE 7

RANK CORRELATION RESULTS BETWEEN TREATMENTS

Yes/No

TCCPooled

with NoTCC

Discarded

NSPooled

with NoNS

Discarded

All Options,No-Opinion

Pooled with No

All Options,No-OpinionDiscarded

Yes/no 1.00 0.39 0.72 0.94 0.94 0.75 0.87TCC pooled with no 0.39 1.00 0.75 0.49 0.49 0.59 0.59TCC discarded 0.72 0.75 1.00 0.85 0.85 0.56 0.75NS pooled with no 0.94 0.49 0.85 1.00 1.00 0.68 0.84NS discarded 0.94 0.49 0.85 1.00 1.00 0.68 0.84No-opinion pooled

with no0.75 0.59 0.56 0.68 0.68 1.00 0.81

No-opinionDiscarded

0.87 0.59 0.75 0.84 0.84 0.81 1.00

Note: A hypothesis test based on a Spearman’s rank correlation suggests that there the correlation coefficient is significantly greaterthan zero at the 95% confidence level if the correlation coefficient .0.65. TCC, too close to call; NS, not sure.

8 In a model that had perfect ability to discriminatethe choices in accord with the theory, we would expect thepredicted yes probabilities for the yes answers (for the noanswers) to be .0.5 (,0.5). For those that selected TCC,we would expect predicted yes probabilities to be aboutequal to 0.5 – the point of indifference implied by randomutility theory. Expectations for the NS option are lessclear. Suppose the NS represent satisficing, then reallyeasy choices would get answered and the easiest to answerare the choices where they are clearly yes or clearly no. Inthis case, we would expect that NS would also tend tohave a predicted value of about 0.5, but that the variancewould be greater than the variance associated with TCC.


A single-factor ANOVA was used to testif there was significant difference amongmeans. The group mean square is 7.94 andthe error mean square is 0.03, yielding an F-statistic of 90.36 with 3 and 1,823 degrees offreedom, which yields a p-value that isessentially zero. This implies that the meanassociated with at least one response type issignificantly different from the mean asso-ciated with at least one other response type.If the model has predictive power, then itshould be expected that at least no and yesresponses were significantly different.

The Tukey test, also known as ‘‘thehonestly significant difference test’’ and‘‘wholly significant difference test,’’ wasused to identify the response options thathad significantly different means in a setof post hoc, pairwise comparisons (Zar1996). Tukey tests allow one to determineif there are pairs of means such that thenull hypothesis of no difference wouldnot be rejected if just those two meanswere tested alone (Table 9). The criticalvalue for the Tukey test with errordegrees of freedom of 1,823, and fourcategories at the 95% confidence level is3.633. All comparisons yielded a Tukey q-statistic greater than the critical valueexcept the NS-TCC comparison (q 52.954). This result supports the hypothe-sized expectation that the predicted meansassociated with yes and no responses areindeed different. These results indicate that

both no-opinion responses are significantlydifferent from both yes and no responses—implying the model has predictive power.This indicates that no-opinion responsesmay reflect that no-opinion respondents arenear utility indifference.

VI. CONCLUSION

To our knowledge, this is the first paperto explore the treatment of no-opinionresponses in an ABM setting that tries todifferentiate between alternative types ofno-opinion responses. Addressing the issueof no-opinion responses has policy implica-tions, especially in the case of benefittransfer. The differences and similaritiesbetween ABM and CVM are well docu-mented (Boxall et al. 1996; Holmes andAdamowicz 2003). Research on how totreat no-opinion responses in the CVMliterature has been advancing since theNOAA commission made its recommenda-tion to include a ‘‘no vote’’ option. Thework presented in this paper providesevidence for ABM that contrasts with theconventional wisdom that no-opinion re-sponses should be treated as no responses inthe CVM literature (Carson et al. 1998).However, comparing the yes/no responsetreatment to the treatments where the onlyno-opinion response was ‘‘not sure’’ led tofindings that are more consistent with thoseof Carson et al. (1998).


TABLE 8

SUMMARY STATISTICS FOR PREDICTED PROBABILITY OF YES GIVEN AN ACTUAL RESPONSE IN THE RESERVED DATA

Prob (Yes)| Yes Prob (Yes)|No Prob (Yes)|TCC Prob (Yes)|NS

Mean 0.669 0.514 0.585 0.602Standard deviation 0.162 0.187 0.165 0.185Total responses 916 490 303 118


TABLE 9

TUKEY TEST USED TO COMPARE THE MEANS FROM OUT-OF-SAMPLE TESTING

Comparison No to Yes No to NS No to TCC Yes to TCC Yes to NS NS to TCC

Difference of means 0.155 0.088 0.070 0.085 0.068 0.017Standard error 0.003 0.005 0.004 0.003 0.004 0.006Tukey q-statistic 48.107 17.872 16.356 24.515 17.948 2.954

Note: The critical value at the 95% confidence level is 3.633. TCC, too close to call; NS, not sure.


There are several hypotheses that mayexplain the results presented here. First, theABM response format may be differentenough so that no-opinion responses rep-resent optimizing and not satisficing. Con-versely, the high dimensionality of ABMstudies and the use of main-effects designsmay be expected to increase proximity toutility indifference and, hence, respondentsuse of no-opinion responses. It is alsopossible that our results have more to dowith the WTA framing of our choicequestion, supporting Groothuis and White-head’s (2002) findings.

The results show that respondents choseno-opinion responses significantly moreoften when the number of no-opinionoptions increased. It does seem likely thatthe inclusion of two no-opinion responsesdrew some respondents that may otherwisehave been leaning in a given direction andpotentially would have answered yes or no.It is also likely that by adding a second no-opinion response option a disproportionatenumber of would-be yes voters switched toone of the no-opinion responses (this maybe true even if a disproportionate number ofwould-be no voters would have chosen no-opinion when only one no-opinion optionwas available).

The effect that TCC had on attributeranks indicates that this option was affect-ing the decision-making process in waysconsistent with utility indifference, and inways that led to pronounced differences inthe preference orderings. Among the codingstrategies examined, dropping no-opinionresponses appears to yield results mostconsistent with surveys that do not offerno-opinion response options. The resultsprovide evidence that when two no-opinionresponse options were offered, respondentsmay have used these options differently,perhaps to express indifference that mayhave resulted from optimizing (‘‘too close tocall’’) as opposed to uncertainty that mayhave resulted from satisficing (‘‘not sure’’).

Our results present a trade-off for re-searchers. Including no-opinion responseoptions means that respondents will selectthem, which reduces the sample size of yes

and no responses. However, if there is a wayto recover information from some no-opinion responses, then adding no-opinionresponse options may be beneficial. Othershave commented on the potential utility ofdeveloping models to allow coding ofnonresponses or no-opinion responses be-tween for and against votes (Kristrom1997). Indeed, Wang’s (1997) ordered prob-it approach treats the no-opinion responsesas ordered responses between yes and no.With our two qualitatively different re-sponse options and the evidence thatrespondents use them differently, there isno natural ordering to the response catego-ries, thereby ruling out the ordered choicemodels. Champ, Alberini, and Correas(2005) use a multinomial logit (unordered)model to analyze the likelihood an individ-ual with a set of demographic characteris-tics chooses each response type. Although,as they note, the model lacks a theoreticalbase in welfare economics and thereforedoes not yield welfare estimates.9 Clearlythere is an empirical need for further modeldevelopment to take advantage of potentialpreference information in some of the no-opinion responses.

The future development and refinementof attribute-based stated preference tech-niques depends on understanding how totreat response options that allow respon-dents to express ‘‘no opinion.’’ ABMincreasingly contributes to our ability tomeasure preferences for goods that havenon-use values or attributes extendingbeyond current conditions. This paper


9 We attempted to use Champ, Alberini, and Cor-reas’s (2005) approach to predict whether NS or TCCresponses should be recoded as yes or no responses.Although the multinomial logit for the unorderedresponse categories had a significant overall model fit,the model failed to correctly predict any of the NS or theTCC responses using the percentage predicted correctlyclassification criteria (i.e., 0% for TCC and 0% for NS).Nevertheless, a few demographic variables were signifi-cant in explaining the response options, providing furtherevidence that TCC and NS are different. Respondentswith a college education were less likely to respond withthe TCC option (p 5 0.007), while urban-dwelling, malecollege graduates (p , 0.01) with higher incomes (p 50.043) were significantly less likely to respond with the NSoption.


provides a first step to understanding howto treat no-opinion responses in the ABMformat, but more work in this area isneeded. Areas of future study includeinvestigating if estimating the probabilityof a ‘‘too close to call’’ response can be usedto estimate indifference and improve theability to predict choices. More case studiesneed to be examined, especially WTP cases.

References

Alberini, A., K. Boyle, and M. Welsh. 2003.‘‘Analysis of Contingent Valuation Data withMultiple Bids and Response Options AllowingRespondent to Express Uncertainty.’’ Journalof Environmental Economics and Management45 (1): 40–62.

Arrow, K., R. Solow, P. R. Portney, E. E. Learner,R. Radner, and H. Schuman. 1993. ‘‘Report ofthe NOAA Panel on Contingent Valuation.’’Federal Register 58 (Jan. 15 [10]): 4601–14.

Banzhaf, M. R., F. R. Johnson, and K. E.Mathews. 2001. ‘‘Opt-Out Alternatives andAnglers’ Stated Preferences.’’ In The ChoiceModelling Approach to Environmental Valua-tion, ed. J. Bennett and R. Blamey. Cheltenham:Edward Elgar Publishing Company.

Birol, E., K. Karousakis, and P. Koundouri. 2006.‘‘Using a choice experiment to account forpreference heterogeneity in wetland attributes:The case of Cheimaditida wetland in Greece.’’Ecological Economics 60 (1): 145–56.

Boxall, P. C., W. L. Adamowicz, J. Swait, M.Williams, and J. Louviere. 1996. ‘‘A Compar-ison of Stated Preference Methods for Environ-mental Valuation.’’ Ecological Economics 18(3): 243–53.

Brander, L. M., R. J. G. M. Florax, and J. E.Vermatt. 2006. ‘‘The Empirics of WetlandValuation: A Comprehensive Summary and aMeta-analysis of the Literature.’’ Environmentaland Resource Economics 33 (2): 223–50.

Carson, R. T., W. M. Hanemann, R. J. Kopp, J. A.Krosnick, R. C. Mitchell, S. Presser, R. A.Ruud, V. K. Smith, M. Conaway, and K.Martin. 1998. ‘‘Referendum Design and Con-tingent Valuation: The NOAA Panel’s No-VoteRecommendation.’’ Review of Economics andStatistics 80 (2): 484–87.

Caudill, S. B., and P. A. Groothuis. 2005.‘‘Modeling Hidden Alternatives in RandomUtility Models: An Application to ‘‘Don’tKnow’’ Responses in Contingent Valuation.’’Land Economics 81 (3): 445–54.

Champ, P. A., A. Alberini, and I. Correas. 2005.‘‘Using Contingent Valuation to Value aNoxious Weed Control Program: The Effectsof Including an Unsure Response Category.’’Ecological Economics 55 (1): 47–60.

Evans, M. F., N. E. Flores, and K. J. Boyle. 2003.‘‘Multiple-bounded Uncertainty Choice Dataas Probabilistic Intentions.’’ Land Economics 79(4): 549–60.

Foster, V., and S. Mourato. 2003. ‘‘ElicitationFormat and Sensitivity to Scope.’’ Environmen-tal and Resource Economics 24 (2): 141–160.

Groothuis, P. A., and J. C. Whitehead. 2002.‘‘Does Don’t Know Mean No? Analysis of‘Don’t Know’ Response in DichotomousChoice Contingent Valuation Questions.’’ Ap-plied Economics 34 (15): 1935–40.

Haab, T. C., and K. E. McConnell. 2002. ValuingEnvironmental and Natural Resources: TheEconometrics of Non-market valuation. North-ampton: Edward Elgar Publishing.

Hoehn, J. P., F. Lupi, and M. D. Kaplowitz.2004. ‘‘Web-Based Methods for ValuingWetland Services.’’ Report R827922. Wash-ington, DC: U.S. Environmental ProtectionAgency.

Holmes, T. P., and W. L. Adamowicz. 2003.‘‘Attribute-Based Methods.’’ In A Primer onNonmarket Valuation, ed. P. A. Champ, K. J.Boyle, and T. C. Brown. Boston: KluwerAcademic Publishers.

Johnston, R. J., M. H. Ranson, E. Y. Besedin, andE. C. Helm. 2006. ‘‘What Determines Willing-ness to Pay Per Fish? A Meta-analysis ofRecreational Fishing Values.’’ Marine ResourceEconomics 21 (1): 1–32.

Kaplowitz, M. D., F. Lupi, and J. P. Hoehn.2004. ‘‘Multiple-Methods for Developing andEvaluating a Stated Choice Survey for Valu-ing Wetland Ecosystems.’’ In QuestionnaireDevelopment, Evaluation, and Testing Methods,ed. A. S. Presser. New York: John Wiley andSons.

Kristrom, B. 1997. ‘‘Spike Models in ContingentValuation.’’ American Journal of AgriculturalEconomics 79 (3): 1013–23.

Krosnick, J. A. 2002. ‘‘The Causes of No-OpinionResponses to Attitude Measures in Surveys:They Are Rarely What They Appear to Be.’’ InSurvey Nonresponse, ed. R. M. Groves, D. A.Dillman, J. L. Eltinge, and R. J. A. Little. NewYork: Wiley.

Krosnick, J. A., A. L. Holbrook, M. K. Berent, R.T. Carson, W. M. Hanemann, R. J. Kopp, R. C.Mitchell, S. Presser, P. A. Ruud, V. K. Smith,W. R. Moody, M. C. Green, and M. Conaway.2002. ‘‘The Impact of ‘No Opinion’ Response



Options on Data Quality, Non-attitude Reduc-tion or an Invitation to Satisfice?’’ PublicOpinion Quarterly 66 (3): 371–403.

Louviere, J. J., D. A. Hensher, and J. D. Swait.2000. Stated Choice Methods: Analysis andApplications. Cambridge, UK: Cambridge Uni-versity Press.

Lupi, F., M. D. Kaplowitz, and J. P. Hoehn. 2002.‘‘The Economic Equivalency of Drained andRestored Wetlands in Michigan.’’ AmericanJournal of Agricultural Economics 84 (5):1355–61.

Stevens, T. H., R. Belkner, D. Dennis, D.Kittredge, and C. Willis. 2000. ‘‘Comparisonof Contingent Valuation and Conjoint Analysis

in Ecosystem Management.’’ Ecological Eco-nomics 32 (1): 63–74.

Vossler, C. A., and G. L. Poe. 2005. ‘‘Analysis ofContingent Valuation Data with Multiple Bidsand Response Options Allowing Respondentsto Express Uncertainty: A Comment.’’ Journalof Environmental Economics and Management49 (1): 197–200.

Wang, H. 1997. ‘‘Treatment of ‘Don’t Know’Responses in Contingent Valuation Surveys: ARandom Valuation Model.’’ Journal of Envi-ronmental Economics and Management 32 (2):219–32.

Zar, J. H. 1996. Biostatistical Analysis, 3rd ed.Upper Saddle River, NJ: Prentice Hall.



split-sample tests of ‘‘no opinion’’ responses in an...

Documents