A MARKET-ORIENTED APPROACH TO MAXIMIZING PRODUCT BENEFITS … · 2013-12-08 · A MARKET-ORIENTED...

A MARKET-ORIENTED APPROACH TO

MAXIMIZING PRODUCT BENEFITS:CASES IN U.S. FOREST PRODUCTS

INDUSTRIES

VIJAY S. REDDY, ROBERT J. BUSH & RONEN ROUDIK

Conjoint analysis, a decompositional customer preference modellingtechnique, has seen little application to forest products. However, thetechnique provides useful information for marketing decisions byquantifying consumer preference functions for multiattribute productalternatives. The results of a conjoint analysis include thecontribution of each attribute and attribute level to overall productutility, the relative importance of product attributes, and marketshare estimates for hypothetical product attribute combinations. Inthis paper, the technique is first examined by discussing the series ofsteps involved in a conjoint study. Following this, the application ofconjoint analysis in two studies of forest products is demonstrated.

INTRODUCTION

Consider the following situation:You are a producer of softwood lumberand a competing company has made adecision to provide wane free lumberand to increase the grading accuracy ofits lumber. How should you respond?

Perhaps you are a supplier ofrepaired pallets and your customers areasking for a better quality product.However, you are not certain how toobtain the greatest increase in quality forthe cost. You might eliminate stringerrepairs or tighten your restrictions ondeckboard splits. Will this increase

Authors are, respectively, Post-DoctoralResearch Associate, Associate Professor andDirector, and a former graduate student, Center forForest Products Marketing, Department of WoodScience and Forest Products, Virginia PolytechnicInstitute and State University, Blacksburg, VA24061-0503, Tel. +1-703-231-5876, Fax. +1-703-231-8868, E-mail: [email protected]

quality as perceived by your customers?When faced with such problems,

market-oriented companies will strive tounderstand the impact that productchanges have on customer perceivedquality. In other words, they will seek tounderstand customer preference struc-ture and provide products which matchthese preferences. Much of the informa-tion available concerning customer pref-erences will be qualitative. However,companies may wish to supplement suchinformation with the results of quantita-tive studies.

This paper describes one approach toobtaining and analyzing quantitativedata concerning customer preferencestructures: conjoint analysis. We beginwith a brief overview of preferencemodels and a description of howconjoint analysis differs from alternativetechniques. Conjoint analysis involvesspecific steps, each involving methodo-

VIJAY S. REDDY, ROBERT J. BUSH & RONEN ROUDIK 19

logical choices. These steps and choicesare reviewed in the next segment of thepaper. Finally, redescribe the applica-tion of conjoint techniques in twostudies of forest products: softwoodlumber used by manufacturers ofpreservatively treated products andrepaired wood pallets used by the U.S.grocery distribution industry.

The primary purpose of this paper isto provide the marketing researcher inthe area of forest products with a non-technical discussion of evolution ofconjoint analysis and the various pathsthat marketing researchers have taken tomultiattribute utility measurement. Thepaper is not meant to provide a compre-hensive discussion or “how-to” ofconjoint analysis. Rather it is meant tostimulate interest in this powerful tool.

BACKGROUND

PREFERENCE MODELS

There are two basic approaches tomodeling buyer preferences: composi-tional and decompositional models. Incompositional models, the total utility orpreference for a product is estimatedfrom evaluative data on individualattributes. In other words, this modelbegins with a set of explicit perceptionsor beliefs about product attributes anduses them as a basis for predictingoverall product evaluations (Holbrook198 1). The value expectancy model(Fishbein 1967) and linear compensa-tory attitude model (Wilkie & Pessemier1973) fall within this category.

When using decompositional mod-els, utilities for individual attributelevels are estimated from data whichdescribes overall customer evaluationsof products. These models start withpreference judgments for attribute bun-dles (i.e., products consisting of specific

combinations of attributes and attributelevels) and use these judgments to inferthe values associated with underlyingattributes (Holbrook 1981). Conjointanalysis (Green & Rao 1971) is a de-compositional model.

Although buyer preference modelsdeveloped using compositional tech-niques provide an understanding ofcustomer attitudes toward particularproduct attributes, they usually do notexplain how buyers make trade-offsbetween the features (attributes) andprice when bundled together withinproducts. Interdependencies amongattributes are not entirely explained byresearch techniques which examine eachattribute independently. In reality, buy-ers do not purchase individual productattributes. Instead, buyers most oftenchoose one product from several alter-native products (bundles of attributes)based on the entire product and theability of that product to satisfy theirneeds better than the available alterna-tive products. Conjoint analysis (adecompositional model) can be used tounderstand the trade-off approach toproduct choice.

CONJOINT ANALYSIS

Conjoint analysis1 is a measurementtechnique developed in the field ofmathematical psychology by Lute andTukey (1964) and introduced into mar-keting research by Green and Rao(1971). Since its introduction in theearly 1970s, conjoint analysis hasemerged as the most widely appliedmarketing tool for measuring buyerpreferences among multiattribute alter-natives and determining the trade-offsconsumers make among multiattribute

1 Conjoint analysis is sometimes called trade-offanalysis since this procedure is concerned withtrade-off judgments.

20 MAXIMIZING PRODUCT BENEFITS

products. According to Wittink andCattin (1989), commercial applicationsof conjoint analysis from 1971 to 1981and from 1981 to 1985 were 698 and1062, respectively. Most commercialapplications of conjoint analysis were innew product concept identification,competitive analysis, pricing, segmenta-tion, and product repositioning (Kohli &Mahajan 1991).

In conjoint analysis, overall judg-ments of a set of alternatives are decom-posed into separate and compatibleutility scores from which the originaljudgments can be reconstituted. Theresulting information describes the rela-tive importance of the product attributesand the value (utility) of the levels ofeach attribute. In other words, conjointanalysis provides a quantitative estimateof how each attribute level impacts abuyer’s judgment (i.e., purchase deci-sion).

The word “conjoint” is used becausethe relative values of attributes consid-ered jointly can be determined eventhough they might not be measurable iftaken one at a time (Churchill 1991). Inconjoint analysis, the characteristics(key dimensions) of a product aredescribed in terms of attributes. Vari-ations within an attribute are describedas levels. For example, exterior color ofan automobile is an attribute, and thealternative colors (e.g., red, white,green) are levels of the color attribute. Inother types of analyses, attributes arecommonly known as independent vari-ables and attribute levels are the valuesof the independent variables.

A researcher must follow certainsteps in conducting a conjoint study. Thefollowing sections focus on the stepsinvolved in conjoint analysis and thealternative available at each step.

Selection of Attributes

The first step in a conjoint study isthe selection of product attributes andattribute levels. This selection should bedone carefully since failure to includethe appropriate attributes will result in aspurious model. Obviously, all attributesthat are relevant to the consumer whenselecting a product should be included(within the limitations described in thefollowing sections), and all attributesthat are not relevant to the purchasedecision should be excluded. In practice,identifying the relevant attributes can bea significant task. Possible approachesinclude: carefully reviewing the previ-ous studies that have investigated thesame or similar products, interviewingpeople familiar with the product, usingfocus-groups of customers, and/or con-ducting a pilot study. A difficult andoften subjective task is the reduction ofthe number of attributes to a level whichis manageable yet sufficiently accountsfor buyer preferences (Green & Srini-vasan 1978).

Preference Models

The next step in conjoint analysisinvolves choosing a suitable preferencemodel for the study. There are threepreference models in common use, eachof which differs in terms of the shape ofthe function relating attribute values topreference structure. The vector, ideal-point, and part-worth models treatvalues of each attribute as linear, linearplus quadratic (curvilinear), and piece-wise linear, respectively. It is possiblefor a researcher to specify a mixedmodel in which some attributes followthe part-worth function model whileother attributes follow vector and/orideal point models (Green & Srinivasan1990).


In general, the vector or ideal-pointmodels should be specified for attributeswhose values are metric (i.e., continu-ous). For example, the vector modelwould be appropriate for price (wherethe utility of a product could be theo-rized to decrease in proportion to a priceincrease). The ideal-point model isappropriate when a customer’s prefer-ence for a product diminishes as theattribute levels reach extreme values. Inthis situation, customer preference for aproduct will be high when the attributelevel value is closer to the ideal point.For example, some people prefer mod-erate levels of sweetness in coffee, andpreference will decrease when sweetnessdecreases or increases from themoderate level.

The part-worth model is specifiedwhen attribute measurement is non-metric. This model allows differentpreference function shapes along each ofthe attributes (Green & Srinivasan1978). This model also allows the esti-mation of a part-worth or utility valuefor each level of each attribute. Thegeneral forms of the three models aredescribed below:

First, let

U jn =

i =

V i =X jn =

P i =

f i =

t =

the total utility (preference) of the jth

stimulus for the nth respondent1, 2, . . . , t denote the set of tattributes or factors that have beenchosenvector coefficient for the attribute ithe level of the ith attribute for the jth

stimulus facing nth respondentideal-point coefficient for theattribute ithe function denoting the part-worthof different levels of X jni for the ith

attributetotal number of attributes

Vector model

(1)

Ideal-point model:

(2)

Part-worth model:

(3)

Conjoint Designs

The next step in conjoint analysis isthe selection of a design for the collec-tion of data. There are four main con-joint designs through which data can begathered: (1) the pairwise trade-offmethod; (2) the full profile method; (3)the hybrid method; and (4) adaptiveconjoint analysis.

Pairwise trade-off designs

The pairwise trade-off method gath-ers data on a two attributes-at-a-timebasis. The respondent is asked to rankeach pair of attributes at different levelsfrom the most preferred to the leastpreferred (Green & Srinivasan 1978).This conjoint design method is simpleand easy to apply while reducing poten-tial information overload on the part ofthe respondent. However, by decompos-ing the product attributes into two-at-a-time combinations, this model sacrificessome realism. In addition, respondentsmay be unclear as to what should beassumed about the other attributes thatare not shown in the two-attribute trade-off matrix. When the attributes of aproduct correlate, it is unclear what theobtained data really mean. In particular,a problem arises when one of the attrib-utes presented in the stimulus is corre-


lated with other relevant attributes notincluded in the trade-off matrix. Perhapsbecause of these limitations, the use ofthe trade-off conjoint design method hasdecreased substantially (Wittink & Cat-tin 1989). According to Johnson (1987,p. 257), the tradeoff method ". . . hasbecome nearly obsolete."

Full-profile designs

In contrast to the pairwise trade-offdesign, the full-profile design utilizesseveral attributes of a product at a time.In this method, a set of full-profiles areconstructed where each profile consistsof a certain combination of attributelevels (one per attribute). The respon-dent evaluates the set of full-profiles(also called stimuli) and ranks or rateseach profile in terms of preference. Thefull-profile method is an improvementover the trade-off method because itprovides more realistic descriptions ofstimuli. The full-profile conjoint designwas used in almost two thirds of thecommercial studies conducted between1981 and 1985 (Wittink & Cattin 1989).

The use of a full-profile designbecomes problematic when the numberof attributes and/or the number of levelsof each attribute becomes large. Forexample, if there are six attributes withthree levels each, the number of possiblecombinations a respondent might berequired to evaluate will be 729 (36).This would undoubtedly cause informa-tion overload and may tempt respon-dents to simplify the experimental taskby ignoring variations in the lessimportant attributes or by simplifyingthe attribute levels themselves (Green &Srinivasan 1978). As a result, mostconjoint studies rely on some type offractional factorial design. The primarypurpose of this design is to reduce thenumber of evaluations required of

respondents. The fractional factorial de-sign selects a subset of all possiblecombinations which still allows estima-tion of model parameters (part-worths orutilities) for all main effects withoutsacrificing the predictive power con-tained in the original design.

The full-profile method is generallylimited to small number of attributes(usually no more than seven) in anyspecific experiment. Researchers shouldincorporate "bridging" factors in studiesthat require a larger number of attributes(Bretton-Clark 1992, Green & Srini-vasan 1978). In bridging, the full set ofattributes are first split into subsets offive or six attributes each. Each stimulus(card) is then composed of attributecombinations that involve those five orsix attributes. In each subset one or twoattributes are common to provide a basisfor linking part-worth functions acrossthe various subsets of attributes (Green& Srinivasan 1978).

Hybrid conjoint designs

Hybrid conjoint designs were devel-oped to reduce the complexity of thedata collection task in conjoint analysiswhile retaining the individual differ-ences in preference (utility) estimation.There are two steps in hybrid conjointanalysis: (1) self-explicated data; and (2)traditional fill-profile stimuli rating.

The self-explication approach is acompositional technique similar to ex-pectancy-value models of attitude theory(Wilkie & Pessemier 1973). While usingthis approach, the respondent firstevaluates the levels of each attribute ona numeric scale (e.g., 0 to 10 desirabilityscale with the value of 10 for the mostdesirable level and 0 for the leastdesirable level). In addition, the respon-dent allocates 100 points among allattributes so as to reflect his/her


desirability ratings. The part-worth ofeach attribute level is calculated bymultiplying the importance weights withattribute-level desirability, and the re-spondents overall preference for a givenalternative is then calculated using theadditive model (Green 1984).

(4)

Where

U j =

W i =

U ki =

In

the total utility (preference) ofalternative (stimulus) jthe self-explicated importanceweight for attribute ithe desirability score for level k ofattribute i

the second step of hybrid conjointanalysis (i.e., traditional full-profilestimuli rating), each respondent providesevaluations of a limited number of pro-files. A small number of profiles aredrawn from a master full-profile designin a manner that permits orthogonalestimation of all main effects andselected t we-way interactions. The pro-files are selected in such a way that eachstimulus in the larger set is evaluated bya subset of the respondents. The hybridpart-worths are estimated through mul-tiple regression by relating the overallpreference judgments for the fullprofiles to the self-explicated utilities(Green 1984).

Adaptive conjoint analysis designs

Adaptive conjoint analysis (ACA)was developed by Richard Johnson atSawtooth Software, Inc., Evanston, IL.The technique involves collecting pref-erence data using microcomputers whilecustomizing the stimuli presentationsaccording to a respondent’s priorevaluation of attributes and levels ofattributes. It incorporates some of the

strengths of both trade-off and full-pro-file approaches (Johnson 1987).

Green and Srinivasan (1990) suggestthat the full-profile method should bepreferred as a data collection tool whenthe number of attributes included in thestudy is six or less. They recommend theuse of the trade-off method or bridgingtechnique (if respondents are willing todo multiple card sorts) if the number ofattributes is larger. They suggest the useof the self-explication approach or thecombination methods such as Hybrid orACA when the number of attributes is10 or more. Indeed, Wittink, Vriens, andBurhenne (1994) report an increase inthe application frequency of ACA as thenumber of attributes increases.

Stimulus Presentation

The next step in conjoint analysis isto select the method of presentingstimuli to respondents. The followingare the primary presentation methods:(1) verbal description or profile cards;(2) paragraph description; (3) pictorialrepresentation; (4) actual product orprototype product. In the verbal descrip-tion (profile cards) approach, the re-spondent is presented with a number ofstimuli with concise attribute leveldescriptions. This is, by far, the mostwidely used approach to presentingstimuli (Green & Srinivasan 1990, Wit-tink et al. 1994). In the paragraphdescription approach a more completedescription of the stimulus is provided.The disadvantage of this method is thatit limits the number of descriptions thatcan be presented, possibly making theestimated parameters less accurate(Green & Srinivasan 1978).

Pictorial representation using vari-ous kinds of visual props or threedimensional models is increasing inpopularity (Green & Srinivasan 1990,


Wittink et al. 1994). This method con-veys information in less ambiguousways, reduces information overload byrelieving respondents of the reading taskand the need to visualize a large amountof information, makes stimuli morerealistic, and makes the task interestingfor respondents (Green & Srinivasan1978). The major disadvantage of thismethod is cost and the possibility of apicture displaying the information in away that differs from the researcher’sintention. In a limited number of studiesactual products or product prototypeswere used as a stimuli (Green & Srini-vasan 1990, Wittink et al. 1994). Thismethod has all the advantages and dis-advantages of pictorial representations.

Measurement Scale for the DependentVariable

Once the stimulus presentationmethod has been selected, the researchermust choose the measurement scale forthe dependent variable. The two types ofscales commonly used are ranking andrating. The rating scale is the mostpopular measurement scale used inconjoint studies (Wittink & Cattin 1989,Wittink et al. 1994) and has the advan-tage of being easily implemented inmailed questionnaires. Depending on thepurpose of the study, both types ofscales can be used to measure eitherpreference for a product or intention topurchase. A third scale type that isgaining popularity is graded pairedcomparisons in ACA (Green & Srini-vasan 1990). In this method, respon-dents rate preference for profiles showntwo at a time.

Estimation Procedures

Depending on the type of datacollected, one of the following three

procedures can be used to estimate thepreference values (utilities or part-worths) of attribute levels: MONANO-VA (Kruskal 1965), LINMAP (Shocker& Srinivasan 1977), and OLS regres-sion. Initially, researchers favored non-metric estimation procedures (MON-ANOVA and LINMAP) for rank orderdata. However, Green and Srinivasan(1978) have shown that OLS applied torank order data provides resultscomparable to non-metric procedures.Because of i t s s impl ic i ty andavailability, OLS is widely used toderive utilities (Wittink et al. 1994;Wittink & Cattin 1989). Darmon andRouzies (1991) report that OLS yieldsthe least distorted weights whencompared to other methods (i.e., LIN-MAP, MONANOVA) irrespective ofthe type of stimulus presentation usedwhile collecting data. In addition, underfractional factorial design, OLS is by farthe most preferable procedure as far asattribute importance weight estimatesare concerned (Cattin & Bliemel 1978;Darmon & Rouzies 1991; Green &Srinivasan 1978).

Simulations and Sensitivity Analysis

Since conjoint analysis uses modelsdeveloped at the individual respondentlevel, simulations and sensitivity analy-ses can be conducted. Simulation algo-rithms use a matrix of individual levelutilities and a set of user-specifiedproduct profiles as inputs and providethe proportion of choices received byeach product (i.e., its market share) asoutput. Using these algorithms, one canperform sensitivity analysis by examin-ing systematically the interplay of utili-ties and product profiles. Most choicesimulators currently available utilize amatrix of respondent background char-acteristics, such as the product currently


used and demographics, in addition to amatrix of respondent utility scores and aset of product descriptions (Green &Krieger 1988).

Simulation algorithms use one ofthree rules, max-utiliy, share of utility(BTL), and Logit. 2 However, the max-utility rule (which assumes that buyerschoose the product with the highestutility) has been widely used in marketresearch. Basic sensitivity analysesalgorithms typically assume a max-utility choice rule (Green & Kreiger1993).

CONJOINT STUDIES IN U.S. FOREST

PRODUCTS INDUSTRIES

SITUATIONAL CONTEXT OF THE

STUDIES

Study 1: Softwood Lumber

Softwood lumber producers main-tain lumber quality by adhering to vol-untary lumber grading rules. Thesegrading rules provide a common methodfor maintaining certain quality charac-teristics and for conducting marketingactivities. However, they provide little(if any) incentive for lumber suppliers toenhance the minimum required qualitylevel. Since lumber grading rules allowvariation within each grading category,suppliers may maximize these variationsto increase lumber recovery. As a resultof this concentration on lumber recov-ery, mills may produce lumber thatbarely passes the minimum standards setby the grading rules. A lumber recoveryorientation is further encouraged by rawmaterial costs which can account for 60to 80% of the total cost of lumber pro-duction.

2 See Green and Krieger (1988) for further informa-tion about these choice-rules.

Most softwood lumber grading rulesare based on structural properties. How-ever, lumber markets are expanding andsome market segments would like topurchase lumber based on appearance inaddition to structural properties. Onesuch market segment is the woodpreservation industry. Consequently, astudy was commissioned to investigatethe value of various attributes to buyersof softwood lumber for preservativetreatment, how buyers trade-off variousattributes to realize the best value, andthe willingness to pay a higher price ifthe quality of lumber is increasedbeyond the minimum levels specified inthe grading rules. Conjoint analysis is anappropriate tool for investigating thesequestions and was used in this study.

Study 2: Repaired Pallets

The pallet and container industry inthe U.S. uses significant amounts ofwood materials, particularly hardwoodand softwood lumber. This industry isthe single largest market for U.S. hard-woods. In 1993, approximately 4.82billion board feet of solid hardwood(lumber, cants, parts, and shook) wereconsumed by the pallet and containerindustry, accounting for approximately44 percent of U.S. hardwood lumberproduction (Bush et al. 1994). Thisindustry also utilized 2.12 billion boardfeet of solid softwoods in 1993 (Bush etal. 1994).

However, the desire to maintain orreduce costs has resulted in increasedutilization of repaired pallets. To marketrepaired pallets properly, it is necessaryto understand user preferences. Infor-mation concerning how buyers maketrade-offs among repair techniques,pallet condition, and price is particularlyuseful. Since the grocery industry is oneof the major users of pallets in the U.S.,


buyers of recycled and repaired palletsin this market segment were contacted toinvestigate their preferences. Again,conjoint techniques were used to gatherand analyze the data.

METHODS

The following section detail thesteps taken to conduct the two studiesoutlined above.

Selection of Independent Variables

The preliminary variables of interestin the softwood lumber value study weredrawn from Hansen’s (1994) study. InHansen’s study buyers of softwoodlumber for preservative treatment wereasked to rate attributes based on theirimportance to the quality of lumber.Using these data as a starting point, thefinal lumber attributes and levels weredeveloped after reviewing the literatureconcerning grading rules and conductingin-depth field interviews with peopleresponsible for buying softwood lumberfor preservative treatment. These attrib-utes and their levels are shown in Table1. Since softwood lumber prices can bequite volatile, price levels weredescribed relative to the prevailingmarket price.

The attributes of repaired pallets andlevels for these attributes were finalizedafter consulting existing literature,people in the industry, trade associationpersonnel, and academic professionals.Table 2 provides a list of attributes andlevels used in this study.

Pallet attributes were chosen torepresent three general areas with whichbuyers are most concerned: repairtechniques, condition of the pallet, andcost of the pallet. The repair techniquecategory focuses on the methods palletrecyclers use to repair broken stringers.

Recyclers repair a broken stringer byattaching a full or half stringers or byusing metal plating. The condition of thepallet is related to the state of t o pdeckboards and the presence ofprotruding fasteners. The primary de-terminant of the cost of a pallet is theinitial purchase price.

Preference Models

A mixed model was specified forboth studies. Spesifically, a part-worthmodel was specified for all non-metricvariables (attributes), and a vector modelwas specified for the price attribute. Themodel specification for each attributewas verified after data collection and themixed model was found to generate thebest fit with the data. The general formof the two models is:

(5)

Where

P jn =

t =

X jn =

f i =

V p =

X jnp =

perceived value or likelihood of pur-chase of the jth (lumber pack or re-paired pallet product) stimulus orprofile for the nth respondenttotal number of categorical (non-metric) attributes (five in bothstudies)the level of the ith attribute for the jth

stimulus facing nth respondentthe function denoting the part-worthof different levels of X jni for the ith

attributevector coefficient for the attributeprice (p)the level of price (p) attribute for thejth stimulus facing n th respondent

Conjoint Designs

In both studies, the full-profilemethod was used to collect data. Anorthogonal fractional factorial design(Green 1974) was utilized in bothstudies to reduce the judgement burden


imposed on respondents. This type ofdesign is useful for investigating themain effects of attributes on respondentjudgments. Sixteen and twenty fivestimuli were generated for study 1(lumber value) and study 2 (repairedpallets), respectively, using the"Conjoint Designer" software developedby Bretton-Clark (1990). These numbersrepresented the minimum possiblenumber of stimuli, given the number ofattributes and levels, and correspond tothe typical practice of attempting tominimize respondent fatigue (Green1974, Greenberg 1986).

Four additional profiles were in-cluded in each investigation to serve asholdout cards. The purpose of holdoutcards is to assess the reliability of theestimated parameters. Respondents ratethese holdout cards but their ratings arenot used to estimate the parameters. ofthe models. Holdout cards facilitatetesting of the stability of the modelbeyond the data with which the parame-ters were estimated and are the mostcommonly used method of assessing thecross-validity or reliability of estimatedparameters (Bateson et al. 1987).

Stimulus Presentation

In study 1, a verbal description(profile card) approach was adapted topresent each profile to respondents(buyers of softwood lumber). All twentystimuli were presented to lumber buyersin a questionnaire. This method wasutilized as some of the attributes wereabstract and, therefore, could not be pre-sented in pictorial form. Figure 1provides an example of a profile cardused in study 1.

In study 2, pictorial props were usedto collect preference data from the usersof repaired pallets. Pictures make therespondent’s task easier and convey

information in an unambiguous manner.Figure 2 depicts one of the stimuluscards used in the study. One masterpallet was altered in accordance with thespecification of each pallet profile. Ineach prop, up to three pictures wereprovided. The first picture provided anoverall view of the pallet, and theremaining two pictures (where neededfor clarity) were taken at close range tohighlight those attributes that werealtered in each profile.

Dependent Variable andMeasurement Scale

In study 1, the dependent variable(perceived value) was measured using aseven-point rating scale, where 1indicated poor value and 7 indicatedexcellent value. A general definition ofvalue (quality for the price) wasprovided in the survey instrument.

In study 2, the dependent variablewas likelihood of purchasing therepaired pallet. Each respondent waspresented one card at a time and askedto indicate their likelihood of purchaseon a nine point rating scale, where 1indicated very unlikely to purchase and9 indicated very likely to purchase.

Data Collection

In study 1, data were collectedthrough a mail questionnaire to buyersof softwood lumber for preservativetreatment in the U.S. A pilot study wasconducted by mailing the questionnaireto fifty subjects selected randomly fromthe sample frame. This was done toassess the response rate and improve thequality of questions. Based on the resultsof the pilot study, minor modificationswere made to the survey instrument.Dillman's (1978) suggestions were fol-lowed while developing the survey


instrument, postcards, cover letters, andfor mailing postcards and question-naires. The modified questionnaire wassent to all remaining buyers in thesample frame. This resulted in 151responses, which is art adequate numberwhen the sample is relatively homoge-n e o u s3. After adjusting for badaddresses and buyers who did not treatsoftwood lumber, the response rate was43.6 percent (151 out of 346).

In study 2, sixty buyers of recycledand repair pallets in the U.S. groceryindustry were personally interviewed.One of the authors traveled to theparticipants’ facilities to conduct theinterview and collect rating data usingthe profile cards. This allowed us toprobe the respondent to gain insightsand information not available from amail questionnaire.

Estimation Procedures

Both studies used the ConjointAnalyzer software distributed by Bret-ton-Clark (1992) for data analysis. Thissoftware uses ordinary least square(OLS) regression to estimate the modelparameters. Respondent’s rating scoreswere treated as dependent variables andattribute levels were treated as values ofindependent variables in the regression.The levels of the non-metric independ-ent variables were dummy coded priorto the regression. The estimated parame-ters are referred to as part-worths orutility scores. They indicate the degreeof importance of each attribute level tothe buyer in the formation of prefer-ences. As mentioned previously, themodel parameters were estimated on an

3 The sample was assumed to be homogeneous sincerespondents are one softwood lumber customergroup. Cluster analysis of the data confirmed thatrespondent preferences are relatively similar.

individual respondent basis and aggre-gated to estimate the overall model.

Simulation and Sensitivity Analysis

A first-choice (max-utility) rulesimulator was used to perform simula-tions and sensitivity analyses in bothstudies and the sensitivity of respondentsto changes in attribute levels and pricewas determined. While manipulatingattribute levels, one attribute at a time,and price, the proportion of respondentspreferring each product profile in thescenario was noted. This type of analysishelped to determine how respondentstrade-off quality for price or vice versa.

INTERPRETATION

Utility or part-worth scores for eachattribute are the primary results aresearcher obtains in a conjoint study.How can one interpret these results?What else can a researcher do with theseresults? The following sections answersthese questions by interpreting theresults obtained in studies 1 and 2.

Utility Scores

Utility scores indicate the influenceof each attribute level in the formationof respondent preferences (or anydependent variable) for the overallproduct. In other words, utility scoresrepresent a respondent’s degree of pref-erence for each level of each attribute.Utility is measured in common interval-scaled units. Consequently, preferencesfor different attributes can be comparedand added together to determine theoverall preference for a product.” Con-joint analysis uncovers a separate utilityfunction for each attribute and for eachrespondent.


In both studies, the individualrespondent’s utility scores were aggre-gated and averaged to obtain an overallpreference structure (utility function) foreach attribute. All respondents’ utilityscores were averaged as the results ofcluster analyses of utility scores indi-cated that respondent preference struc-tures for each attribute were similar.Figures 3 and 4 illustrate the preferencestructure for each attribute in studies 1and 2, respectively. The utility scores inthese figures were adjusted so that theleast preferred level of an attribute willalways have a zero utility score.

Study 1

The utility scores shown in Figure 3reveal that softwood lumber buyersprefer the following level of each of thefive attributes: no wane, $5/mbf belowthe current market price, a lumber packwith 99% of its pieces on-grade orbetter, straight lumber, and a lumberpack with no forklift damage. Althoughthese utility functions confirmed theobvious (i.e., buyers prefer higherquality over lower quality and lowerprice over higher prices), they alsoexpress buyer preferences in quantifi-able terms and show the strength of thepreference. For example, the utilityfunction for wane shows that buyers’preference for lumber increasesapproximately in proportion to thedecrease in the amount of wane inlumber from maximum allowable waneto pencil wane and pencil wane to nowane.

Let us consider the utility functionsfor wane and price. First, the conjointanalysis shows that the average respon-dent’s preference for lumber is highest ifthe lumber pack has no wane a n ddeclines with increases in the amount ofwane. The analysis also shows that the

respondent’s preference for lumber packis strongest at $5/mbf below the currentmarket price and decreases as the priceincreases. Second, it shows that wanehas a stronger influence (range of utilityscores) on the respondent’s overallpreference for lumber than does price.

Third, the utility functions allow usto compute the degree of preference thebuyer has for any of the attributecombinations. For example, the no waneand $5/mbf above the market pricecombination of profile has a combinedutility score of 1.97 (1.49 + 0.48),whereas the pencil wane and same asthe market price/mbf has a combinedutility score of 1.63 (0.67+ 0.96).

Clearly, the respondent exhibits astronger preference for the firstwane/price combination. The overallutility of these two combinations alsoillustrate the kinds of trade-offs respon-dents make in deciding among multiat-tribute products. Here, even though thesecond profile has lower price, theaverage respondent is willing to trade-off some of the utility of low price forthe added utility of no wane. This isevident since the combined utility of thefirst combination is higher than that ofsecond combination. Respondent'strade-offs for the remaining attributecombinations can be observed using theutility scores.

Study 2

The utility scores in this study indi-cate the impact of each level on thelikelihood of respondents purchasing therepaired pallet. The preference structureof an average repaired pallet user foreach attribute is shown in Figure 4. Theresults indicate that pallet end-usersprefer a repaired pallet that has no halfor full stringer attached to its originalstringer, has no metal plate, has no


protruding fasteners, has all top deck-boards intact, and costs $3.00. Againthis combination represents the predict-able result of preferring high quality atlow price.

Preference (utility) is greatest for apallet with no half attached stringer.The preference for a repaired palletdecreases as the number of half attachedstringers increases from none to two.The steep slope of the utility functionfor half attached stringer shows that anytime a half stringer is attached to apallet, buyers' preference toward thatpallet diminishes substantially. Thisutility function also shows that buyersare more sensitive to the introduction ofthe first half attached stringer than to theintroduction of any additional halfattached stringers. The results shown inFigure 4 can be interpreted in the similarfashion for the remaining attributes.

The utility functions for variouspallet attributes indicate how usersperceive trade-offs between attributes.For instance, the average user ofrepaired pallets will trade-off protrudingfasteners to obtain a pallet with no halfattached stringer. This is because theadditional utility to the pallet user that isderived by moving from one halfattached stringer to no half attachedstringer is greater than what is lost fromtrading down to a pallet with protrudingfasteners from a pallet with no protrud-ing fasteners. Figure 4 also indicatesthat the pallet user will trade-off lowerprices in favor of obtaining the mostpreferred levels of all other attributesexcept metal plating (plated stringers).

Relative Importance

The estimated utility scores also helpdetermine the relative importance ofeach attribute in the preference structure.Relative importance indicates the posi-

tion of each attribute, in relation to allother attributes, in influencing a respon-dent’s decision. If the relative impor-tance of one attribute is found to betwice that of the second, it can beinferred that the first attribute has twicethe influence of the second. The relativeimportance score of an attribute can becalculated by dividing the utility scorerange4 of that attribute by the sum ofutility score ranges of all attributes. Thiscan be indicated in percent by multiply-ing the resulting value by 100. Thus, thesum of relative importance of allattributes equals to 100.

Study 1

The relative importance of variousattributes in the determination of lumbervalue are shown in Figure 5.

FIGURE 5. RELATIVE IMPORTANCE OF

ATTRIBUTES IN THE DETERMINATION OF

LUMBER VALUE.

Wane

Price

Accuracy ofgrading

Damage tolumber pack

Lumberstraightness

Moisturecontent

Relative Importance

The largest bars for wane and priceindicate that these attributes are the mostimportant determinants of perceivedvalue to buyers of softwood lumber for

4 The difference between the utility score of themost preferred level and the utility score of theleast preferred level of the attribute.


preservative treatment. Buyers trade-offbetter levels of attributes that havesmaller bars for the better levels ofattributes that have larger bars in thegraph. The data in Figure 3 indicate thatlumber buyers prefer lumber with 19%moisture content over lumber with 25%moisture content. 5 However, Figure 5indicates that this attribute is of verylittle importance to softwood lumberbuyers when compared to other attrib-utes. In fact, the utility functions indi-cate that the average buyer would trade-off 19% moisture content for betterquality levels of more preferred attrib-utes, such as wane, price, lumberstraightness.

Study 2

The relative importance shown inFigure 6 indicates how much eachattribute contributes, in relation to otherattributes, to the formation of buyerpreferences (likelihood of purchase) fora repaired pallet.

Of all attributes, half attachedstringers have the greatest impact onbuyers’ preference for a repaired pallet.The relative importance of this attributewas 28.4 percent. The top deckboardwith a relative importance of 20.3percent represented the second mostsignificant factor in purchase decisions.Collectively, these two attributesaccounted for almost one-half of therespondents’ repaired pallet preference.Figure 6 illustrates the relative impor-tance of all of the attributes included inthe study.

5 Figure 3 also shows that the average respondentpreferred 25% MC over 22% MC. However, thereis no statistically significant difference between theutility scores of these two levels.

FIGURE 6. RELATIVE IMPORTANCE OF

VARIOUS ATTRIBUTES IN THE DETER-

MINATION OF PREFERENCE FOR REPAIRED

PALLETS.

Half attachedstringer

Top deckboard

Protrudingfasteners

Full attachedstringer

Price

Plated stringer

Relative Importance

Trade-offs Between Quality and Price

Simulation analyses were used tounderstand buyer trade-offs betweenquality and price and to learn the differ-ential price point at which buyers prefertwo products equally, even though thequality of one product is superior to thequality of the other product.

Study 1

Figure 7 illustrates the additionalamount softwood buyers will pay permbf for lumber with the most preferredlevel as compared to the least preferredlevel of each attribute. For example,buyers will pay up to $15.00/mbf morefor wane free lumber than for lumberwith maximum allowable wane. At theprice difference of $15.00/mbf, buyerswill be indifferent between the twoofferings (no wane lumber pack andmaximum allowable wane lumber pack),assuming similar knowledge and marketaccess (awareness and distribution, forexample). At this price difference, thepercentage of respondents preferring no


wane lumber and maximum wanelumber becomes roughly equal.

Softwood lumber buyer trade-offsbetween other quality attributes andprice can be interpreted in the similarfashion. For instance, suppliers whoprovide lumber with no forklift damagecan command a price premium of $7.00over competitors who provide lumberwith minor forklift damage.

FIGURE 7. THE ADDITIONAL AMOUNT

TREATERS ARE WILLING TO PAY PER MBF

FOR THE MOST PREFERRED LEVEL AS

COMPARED TO THE LEAST PREFERRED

LEVEL OF EACH LUMBER ATTRIBUTE.

Wane

Damage tolumber pack

Accuracyof grading

Lumberstraightness

Moisturecontent

Study 2

The sensitivity of repaired palletusers to various attributes and price wasdetermined in a manner parallel to thatused for the softwood lumber study.Pallet users were most sensitive to halfattached stringers and traded low priceto obtain this attribute. Even at the pricedifference of $3.90 between the palletwith no half attached stringers and thepallet with two half attached stringers,given the quality of all other attributes isheld constant, a significantly higherpercentage of pallet users (74% vs. 26%)preferred the former pallet (no halfstringers). Also at a price difference of$3.90, the percentage of respondents

preferring no half attached stringerpallet was significantly higher (68% VS.32%) than those preferring one halfattached stringer. However for theattributes protruding fasteners, metalplating, and top deckboards, t h edifference between the percentage ofrespondents preferring a pallet with theleast preferred level and the percentageof respondents preferring a pallet themost preferred level becomes insignifi-cant at a price difference between $1.20to $2.70.

SUMMARY AND CONCLUSIONS

Conjoint analysis can be used tounderstand the preference structure usedby customers when selecting products.This paper provided a non-technicaldescription of various steps involved ina conjoint study while briefly explainingthe different options available in eachstep. The results of two studies wereused to discuss the various steps andresults of conjoint analysis.

A researcher preparing to useconjoint analysis in a study should plancarefully to obtain reliable data.Attribute selection is the most crucialstep since researchers must makesubjective judgments and the remainderof the analysis is predicated on thesignificance of the included attributes.The availability of several specializedconjoint software packages makes thetask of undertaking the remaining stepsin a conjoint study easier. These stepsinclude selecting preference models foreach attribute, determining the stimulusform to use for data collection, selectingan approach to present stimuli tosubjects, deciding on a measurementscale for dependent variables, selecting aprocedure to estimate utility scores, anddetermining the type of choice simula-


Green, Paul E. 1984. Hybrid models fortors to be used to investigate sensitivityto various quality and price combina-tions.

We hope that this paper will sparkthe interest in this useful techniqueamong forest products marketers.Further, we hope that this paper willserve as a resource for people beginningto investigate conjoint analysis.

LITERATURE CITED

Bateson, John E., Reibstein, David & Boulding,William. 1987. Conjoint analysisreliability and validity: A framework forfuture research. In Review of Marketing1987, Michael J. Houston ed. AmericanMarketing Association, Chicago, IL.451-81.

Bretton-Clark. 1990. Conjoint Designer(Version 3). Morristown, NJ.1992. Conjoint Analyzer (Version 3).Morristown, NJ.

Bush, Robert, Hansen, Eric & Punches, John.1994. Wood Use in the U.S. Pallet andContainer Industry: 1993. A reportsubmitted to the members of the Centerfor Forest Products Marketing, VirginiaTech, Blacksburg, VA.

Cattin, Philippe& Bliemel, Friedhelm. 1978.Metric vs. nonmetric procedures formultiattribute modeling: Somesimulation results. Decision Sciences9(3):472-80.

Churchill, Gilbert A., Jr. 1991. MarketingResearch: Methodological Foundations,Fifth edition. The Dryden Press, NewYork.

Darmon, Rene Y. & Rouzies, Dominique. 1991.Internal validity assessment of conjointestimated attribute importance weights.Journal of Academy of MarketingScience 19(4):315-22.

Dillman, Don A. 1978. Mail and TelephoneSurveys: The Total Design Method.John Wiley & Sons, New York, NY.

Fishbein, Martin. 1967. A behavior theoryapproach to the relations between beliefsabout an object and the attitude towardthe object. In Readings in AttitudeTheory and Measurement, M. Fishbein,ed. John Wiley & Sons, New York, 389-99.

conjoint analysis: An expository review.Journal of Marketing Research21(May):155-9.& Krieger, A.M. 1988. Choice rules andsensitivity analysis in conjointsimulators. Journal of AcademicMarketing Science 16(Spring):114-127.

. 1993. Conjoint analysis withproduct-positioning applications. InHandbooks in Operations Research andManagement Science, G.L. Nemhauserand A.H.G Rinnooy Kan, Eds. Volume 5,Elsevier Science Publishers B.V.,Amsterdam, The Netherlands, 467-515.& Rae, V.R. 1971. Conjointmeasurement for quantifying judgmentaldata. Journal of Marketing Research8(3):355-63.& Srinivasan, V. 1978. Conjoint analysisin consumer research: Issues andoutlook. Journal of Consumer Research5(2):103-123.& Srinivasan, V. 1990. Conjoint Analysisin marketing: New developments withimplications for research and practice.Journal of Marketing 54(4):3-19.

Greenberg, Marshall G. 1986. Commentary on'cost of simplifying preference models'.Marketing Science 5(3):320-1

Hansen, Eric. 1994. Empirically deriveddimensions of total product quality forsoftwood lumber. Unpublished doctoraldissertation, Virginia PolytechnicInstitute and State University,Blacksburg, VA.

Holbrook, Morris B. 1981. Integratingcompositional and decompositionalanalyses to represent the intervening roleof perceptions in evaluative judgments.Journal of Marketing Research 18(1):13-28.

Johnson, Richard M. 1987. Adaptive conjointanalysis. Proceedings of the SawtoothSoftware Conference on PerceptualMapping, Conjoint Analysis, andComputer Interviewing, Sun Valley,Idaho, March. 253-65.

Kohli, Rajeev & Mahajan, Vijay. 1991. Areservation-price model for optimalpricing of multiattribute products inconjoint analysis. Journal of MarketingResearch 28(3):347-54.

Kruskal, J. B. 1965. Analysis of factorialexperiments by estimating monotone


transformations of the data. Journal ofRoyal Statistics Social Services, Series B27(2):251-263.

Lute, R. Duncan& Tukey, John W. 1964.Simultaneous conjoint measurement Anew type of fundamental measurement.Journal of Mathematical Psychology1(2): 1-27.

Roseberg, Milton J. 1956. Cognitive structureand attitudinal affect. Journal ofAbnormal and Social Psychology53(4):367-372.

Shocker, A.D. & Srinivasan, V. 1977. Aconsumer-based methodology for theidentification of new product ideas.

Management Science 20(February):921-937.

Wilkie, William L. & Pessemier, Edgar A.1973. Issues in marketing’s use ofmuliattribute attitude models. Journal ofMarketing Research 10(4):428-41.

Wittink, Dick R. & Cattin, Philippe. 1989.Commercial use of conjoint analysis: Anupdate. Journal of Marketing 53(3):91-6., Vriens, Marco & Burhenne, Wire. 1994.Commercial use of conjoint analysis inEurope: Results and critical reflections.International Journal of Research inMarketing 11(11):41-52.

APPENDIX

TABLE 1. ATTRIBUTES AND LEVELS USED TO ESTIMATE SOFTWOOD LUMBER VALUE.


TABLE 2. ATTRIBUTES AND LEVELS USED TO QUANTIFY THE LIKELIHOOD OF PURCHASING

REPAIRED PALLETS.

FIGURE 1. AN EXAMPLE PROFILE CARD USED TO COLLECT DATA CONCERNING THE

PERCEIVED VALUE OF SOFTWOOD LUMBER.

95% of the pieces in the lumber pack are on-grade or better.All the pieces in the lumber pack have no wane.All the pieces in the lumber pack have 25% or less moisture content.All the pieces in the lumber pack are straight.The lumber pack has no forklift damage.Price: same as the current market price/cbm.

Please circle the number that best represents the value (i.e., quality for the price)of the lumber pack described. Please assume the lumber package differs only onthe listed characteristics, and all the other characteristics remain the same.

Poor Excellentvalue value1 2 3 4 5 6 7


FIGURE 2. EXAMPLE PROFILE CARD USED TO COLLECT LIKELIHOOD OF PURCHASING

DATA FOR REPAIRED PALLETS.

FIGURE 3. AVERAGE SOFTWOOD LUMBERR BUYER'S PREFERENCE STRUCTURE FOR EACH

LUMBER ATTRIBUTE AND PRICE.

Attribute Levels


FIGURE 4. AVERAGE RESPONDENT’S PREFERENCE STRUCTURE FOR VARIOUS ATTRIBUTES

OF REPAIRED PALLETS.


UNIVERSITY OF HELSINKI

DEPARTMENT OF FOREST ECONOMICS

PUBLICATIONS NO. 4

ENVIRONMENTAL ISSUES AND MARKET ORIENTATION

CURRENT TOPICS IN FOREST PRODUCTS MARKETING

PROCEEDINGS OF

FOREST PRODUCTS MARKETING GROUP (P5.06-00)IUFRO XX WORLD CONGRESS , 6-12 AUGUST 1995

EDITED BY

HEIKKI JUSLIN AND MIIKKA PESONEN

A MARKET-ORIENTED APPROACH TO MAXIMIZING PRODUCT BENEFITS … · 2013-12-08 · A MARKET-ORIENTED...

Documents

Transcript of A MARKET-ORIENTED APPROACH TO MAXIMIZING PRODUCT BENEFITS … · 2013-12-08 · A MARKET-ORIENTED...