1

A Case Study of Conjoint Analysis: New

Approaches to Product Line Decisions

by

Barry Radler

A thesis submitted in partial fulfillment of requirements for the degree of

Master of Arts

(Psychology)

at

Cleveland State University

1993
 2

INTRODUCTION

Conjoint analysis was first developed in the middle 1960s by mathematical psychologists

Luce and Tukey (1964) and its algorithms solidified by Kruskal and Carmone (Green &

Srinivasan, 1978). Since the early 1970s it has attracted considerable attention in both practical

and academic circles as a technique for decomposing preference structures and predicting

respondent behavior. Conjoint analysis has warranted this fanfare as a flexible method for

portraying consumer decisions realistically as a trade-off among multiattribute products or

services (Hair, Anderson, Tatham, & Black, 1987).

The current study arose from the need of an Eastern agricultural company to redesign one

of its product lines. Conjoint analysis is a particularly applicable tool for such questions in that it

identifies consumer preferences for each component part of the product. Using conjoint analysis,

this study identifies the best product line of calf milk replacers for the client company to offer.

The product line analysis used an integration of conjoint analysis of stated preferences and

directly assessed compositional data, including actual and intended purchases. The conjoint

part-worths and the compositional weights for each product attribute were merged to provide a

stable index of the "real" impact of that attribute.

Specifically, the study addressed the following subjects:

1) Incorporation of a new "pseudo-simulation" of the calf milk replacer market. This pseudo-

simulation, so called because it was not a comprehensive simulation study, specifically addressed

the retention of current customers and, secondarily, the acquisition of new customers. The

pseudo-simulation accounted for low product involvement by integrating conservative weighting

ratios with the conjoint parts-worth scores.

A major problem with conjoint analysis is the susceptibility of the procedure to greatly
 3

overestimate the number of respondents likely to shift to a new product (concept) if it were

actually made available to them. This predicament is most prevalent when disaggregate

(individual) conjoint parameters are interpreted using a "first-choice" or "maximum utility" rule:

that is, it is assumed that respondents will purchase the product with the highest utility score.

This issue is not concerned with the soundness of the statistical algorithm conjoint analysis is

based on (Green & Rao, 1971). Rather, it assumes that systematic biases exist in the consumer's

response data and that these biases are endemic to the procedure as it is usually applied. It was

with these considerations in mind that the pseudo-simulation was constructed.

2) Identification of a "pure" or traditional conjoint output to be juxtaposed with the presumed

more precise pseudo-simulation. The best or most preferred products to offer were derived from

this raw output and what was "best" was operationally indexed by the size of the customer base

(i.e. estimated market share).

3) Documentation of a methodology for accounting for low product involvement and the

accompanying error in a conjoint analysis. The study's methodological design used additional

measures easily incorporated into questionnaires. These approaches are probably widely

applicable to conjoint analyses.

4) A comprehensive--but not exhaustive--review of current literature specific to this type of study

and its problems.

Due to the proprietary nature of this study, pseudonyms were substituted for actual

company and product names.

 4

LITERATURE REVIEW

The goal of consumer research is to predict behavior, and while other methods

(regression, discriminant analysis) attempt to compose a behavioral rule with regard to consumer

action and purchase, conjoint analysis is decompositional in orientation and more closely aligned

with traditional experimentation. Conjoint studies conduct "experiments" with factors identified

as determinant while controlling the levels of these factors (Hair et. al., 1992).

For instance, the additive conjoint model is analogous to the absence of interaction effects

in an N-way ANOVA. ANOVA tests whether or not original cell values are additive

combinations of each row and column. Additive conjoint measurement, essentially a monotone

analogue of main effects ANOVA, attempts to monotonically transform the cell values to achieve

additivity (Green et. al., 1971).

The procedures of conjoint analysis are actually based on models of information

processing and complex decision making (Hair et. al., 1992; Louviere, 1988). The conjoint

approach assumes that consumers are fairly rational about comparing choice alternatives; hence

the technique tends to work best with high-involvement, extended problem-solving situations

rather than low-involvement, impulsive situations (Wyner, 1992).

Complex decision making follows a general pattern beginning with need awareness and

culminating with consequences of behavior (Engel, Blackwell, & Miniard, 1990). This process is

modeled in Figure 1. Conversely, low-involvement situations are typified by limited problem-

solving on the consumer's part. In these situations, it is common to simplify the process and

reduce the number and variety of information sources and alternatives. All of the stages in Figure

1 may still be followed, but with a marked decrease in both extent and rigor. It is important to
 5

note that complex and limited problem solving are extremes on a continuum; a range of possible

behavior exist between the two (Engel et. al., 1990).

Figure 1: Classic complex decision-making pattern.

 6

Consumers' involvement with the decision will also influence the number of criteria used in

alternative evaluation. A lesser number of evaluative criteria are likely to be utilized by the

consumer as involvement decreases (Engel et. al., 1990). Similarly, Wyner (1992) suggests that

conjoint analysis works best in situations where the product attributes are described in easily

understood terms and the number of determinant attributes is small.

The current study used only five variables to define the product and--other than "brand"--

all of the variables were described in discreet terms. Initial information from the client company--

based on past studies of the market--indicated relatively low product involvement. Thus, for the

present study, a congruency existed between conjoint analysis' preferred conditions and the

consumers' assumed behavioral tendencies in low-involvement situations.

Decompositional methods such as conjoint analysis start with measures of preference for

"attribute bundles" and use them to infer the values attached to underlying characteristics. By

contrast, compositional approaches (linear compensatory models) begin with a set of explicit

perceptions or beliefs about characteristics or attributes and use them as the basis for predicting

brand evaluations (Holbrook, 1981). This relationship is demonstrated in Figure 2. Holbrook

(1981) noted that compositional and decompositional methods have developed along largely

separate paths.

Wyner (1992) also has addressed this gap. He advises developing additional scaling

methods that link preference measures to attitudinal measures. The relative parameters generated

by conjoint analysis can then be anchored in some indicator of absolute interest in the product, or

in knowledge or usage of the product. Integration of the two methods, in general, can greatly

enrich conjoint analysis by clarifying the link between objective features and ultimate affect.

With these considerations in mind, the current study incorporated measures of product
 7

usage, knowledge and purchase intent into the questionnaire. The information obtained from

these measures was integrated with obtained conjoint parameters and ultimately used in a pseudo-

simulation.

Looking at Figure 2, it is apparent that consumers form psychophysical, perceptual and

value judgments about attributes and brands (Louviere, 1988). Assumptions must be made

regarding how both compositional and decompositional methods integrate this information before

any link can be made between the two. How these evaluations are combined into an overall

evaluation begs an important question of information integration theory: What decision-making

model is the consumer using?

 8

Figure 2: Combined model of compositional and decompositional techniques.

 9

Assumptions

The underlying assumption of conjoint analysis is that a "composition rule"--a rule used in

combining attributes to produce a judgment of relative value or utility for a product/service--

determines respondents' preferences. It is assumed that any object or concept is evaluated as a

bundle of attributes. These attribute bundles are ultimately judged by combining the separate

amounts of utility provided by each attribute (Hair et. al., 1992). An implicit supposition of this

operation is that such an evaluation occurs within a competitive environment.

Summarily, conjoint analysis refers to any "decompositional method that estimates the

structure of a consumer's preferences given their overall evaluations of a set of alternatives that

are prespecified in terms of different attribute levels" (Green et. al., 1978).

Empirical research has scrutinized whether consumers actually use linear compensatory

decision-making models (e.g. those assumed to be used in conjoint studies) or the evidently

simpler evaluation models such as the lexicographic and conjunctive. This research has found that

simpler rules are usually preferred. However, this apparent problem is subsumed by the fact that

the compensatory model of conjoint analysis can typically approximate the outcomes of other

kinds of decision rules quite closely (Green et. al., 1978).

Further research has indicated three prevailing conditions under which compensatory

decision making models perform (e.g. predict preference functions) well:

1) the preference function is monotone (increasing or decreasing) over increasing

levels of an attribute while holding other attributes constant; 2) there are errors in
the measurement of attribute levels (possibly because of perceptual differences
across consumers); and 3) the attributes tend to be correlated (Green et.
al., 1978).
Algebraically, the basic assumptions of information integration theory can be expressed as:

1) The unknown overall utility a consumer has in their mind regarding the j-th brand is linearly
 10

related to their response on a category rating scale.

Uj = a + bRj + ej, where Uj is the overall value or utility to measure of the j-th brand, Rj is the

observed response on a category-rating scale, and ej is a normally distributed error term with zero

expectation and constant variance, which satisfies assumptions of ANOVA and multiple

regression (Louviere, 1988).

2) A ranking scale used by a consumer under appropriate instruction and task conditions

approximates interval measurement (this will be expounded on later).

3) A consumer's response strategy reveals his or her decision strategy. The response strategy can

be approximated by algebraic conjoint models amenable to experimental investigation and

statistical parameterization (Louviere, 1988).

Indeed, it is the analyst's job to find each attribute's level's part-worth, given some type of

composition rule, that is most consistent with the consumer's responses. The consumer's

evaluation process can be inferred through a conjoint analysis of the way the consumer integrates

attribute evaluations to form overall brand impressions.

Applications

This inference of preference structure explains how important each factor is in overall

preference, and how the differing levels within a factor contribute to overall preference. This

information is used for: 1) definition of the object/concept with the optimum combination of

attribute levels; 2) showing the relative importance of each attribute and level to overall

evaluation; 3) estimating consumer judgments to predict market shares among differing attribute

combinations; 4) definition of potential high and low segments by grouping consumers having

similar preference structures; and 5) exploring the potential for non-existent or hypothetical

attribute combinations (Hair et. al., 1992). Users of conjoint analysis have generally emphasized
 11

predictive validity as of primary importance and have regarded explanation as a desirable, yet

secondary, goal (Green et. al., 1978).

Variable Selection

After the research problem has been stated, a preliminary data collection procedure is

employed identifying those attributes most frequently regarded as relevant. This can be

accomplished by numerous means including customer surveys, focus groups, and consulting

product managers and others knowledgeable about the product/services and its uses.

The task is then to reduce the number of attributes to a reasonable size while still retaining

the strength

to estimate consumer behavior. According to a survey by Cattin and Wittink (1982), in which the

researchers reviewed conjoint procedures frequently used in practical applications, the median

number of attributes used was between 6 and 7. Too many additional attributes complicates the

respondent's job and can introduce unreliable data due to fatigue; too few attributes may not

provide an accurate description of the product.

In the case of continuous attribute spacings, the most frequent practice has been to rely

upon equally spaced attribute levels to represent the appropriate range. Darmon and Rouzies

(1989) have questioned the soundness of this practice and have found that this convention may

not always be appropriate to the study. Specifically, their study suggests that "using smaller

attribute level spacings in the steepest slope range of the utility function will yield more valid

results than using equal or larger spacings." The researchers investigated these effects by varying

level spacings, function range, shape, curvature, and estimation method.

In general, Darmon et. al. (1989) propose that using smaller spacings helps recover the

utility functions' ranges and curvature, and reduces the average error between recovered and true
 12

utilities. They caution, however, that if there is no reason to assume a specific functional form a

priori, then equal attribute level spacings should be used because unequal spacings in the wrong

direction could considerably misrepresent the recovered utilities.

In another paper by Kumar and Gaeth (1991), the authors addressed whether attribute

importance weights changed with the relative position order of the attribute in a conjoint task.

This experiment was specifically meant to investigate the role order effects play in conjoint task

decision-making. Their empirical evidence revealed an absence of order effects for a familiar

product category, but the presence of systematic order effects for an unfamiliar product category.

It was further recommended by Kumar et. al. (1991) that the order of the attributes be

counterbalanced or randomized between subjects to avoid biases. Obviously, this procedure is

appropriate when the researcher is interested in aggregate problems. But attribute order could be

randomized within subjects (Kumar et. al., 1991). While this solution would add unsystematic

variation and inflate the conjoint model's error term, it also would guarantee coefficients not

biased. This procedure tends to improve with a disaggregate interpretation of the utilities.

Another solution suggested by Kumar et. al. (1991), maintained by Page and Rosenbaum

(1989) and useful to remember when constructing the conjoint task and stimuli, is to simply

present the attributes in the natural order they occur when consumers encounter the products.

Again, the correct solution depends upon the nature of study and the proposed research

questions. Regardless, the attributes and their levels must be realistic, distinct and represent a

single concept while at the same time accounting sufficiently well for consumer preferences and

avoiding multicollinearity (Green et. al., 1978, Hair et. al. 1992).

Preference Model

A preference or composition model is then specified and an assumption made about

 13

customer information processing. The issue is whether the predictive validity of the model with

interactions would be better because of increased realism or worse because of the estimation of

several additional parameters, a common trade-off problem in social science. There is some

evidence that the model with interaction items often leads to lower predictive validity and that this

is due to the inclusion of additional parameters in the model (Green and Srinivasan, 1990).

The additive model already discussed assumes that consumers simply add up the

part-worths, or utility associated with each level of each attribute, to get the total worth of the

product/service. An interactive model assumes the total worth is more, or less, than the sum of its

part-worths. The interactive model may be a more accurate representation of the customer

decision-making process, but the additive model allows better estimates of part-worths.

Once the general model is chosen, the part-worth relationship must be specified. While

the composition model decides how the attributes are related, defining the part-worth relationship

indicates how the levels of a attribute are related. The part-worth relationship can be estimated

three ways; vector (linear), quadratic (ideal point) or separate part-worth.

The separate part-worth model provides the greatest flexibility in allowing different shapes

for the preference function along each of the attributes in that both the vector and ideal-point

models can be derived from it. However, this flexibility is delivered at the cost of estimating

additional parameters and approximating intermediate levels by linear interpolation (Green et. al.,

1978). To determine a part-worth value outside the range of estimation, extrapolation of the

linear function would be needed. The validity of this procedure is disputable, hence it is important

to initially incorporate an inclusive range of the attributes when possible.

 14

In choosing the appropriate model, the flexibility of the shape of the preference model

becomes greater as we go from the vector to the ideal point to the part-worth function models.

Derivation of degrees of freedom, in which part-worth models have the fewest degrees of

freedom, also follows this pattern. In fact, the typical conjoint study using a part-worth model

often has no degrees of freedom (Green et. al., 1990). The reliability of the estimated parameters,

however, is likely to improve in the reverse order.

From the standpoint of predictive validity, the relative effectiveness of each model is

generally unclear and most often depends upon a priori conclusions about the variables. It is

possible to incorporate a mixed model where some attributes are best represented using a vector

model while other attributes--categorical variables for instance--may require a part-worth model.

Lastly, in the Cattin et. al. (1982) survey, the part-worth was the most common model used, an

indication of its flexibility and robustness.

Stimulus Set Construction

Once the attributes and corresponding levels have been selected, the stimulus set must be

created. The stimuli can be combined in a factorial design where all possible combinations are

included. However, this design becomes impractical as the number of attributes and levels

increases. For this reason, it is common for only a subset, called a fractional factorial design, of

all possible combinations to be used in the experiment.

A special case of the fractional factorial design is the orthogonal array. An orthogonal

array is a highly fractionated factorial design of the attribute levels and assumes away all higher-

order interactions. In an orthogonal array, each level of an attribute occurs with each level of

another attribute with equal or proportional frequencies, which is a sufficient condition for the

main effects of any two factors to be estimated on an uncorrelated basis (Green, 1974). Such a
 15

design assures the independence of the main effects and represents the most parsimonious way to

estimate all main effects (Green, Carroll & Carmone, 1978). Steckel, DeSarbo and Mahajan

(1991) state "orthogonality guarantees that the resulting parameter estimates obtained from the

analysis would have maximum 'efficiency' since the attributes would be devoid of statistical

correlation."

To capitalize on the fractional design's efficiency, then, factor independence should be

sought when possible. However, the presence of interattribute correlations per se does not

violate any assumptions of conjoint analysis. In this manner conjoint analysis is analogous to

multiple regression with a strictly additive model, where it is implicitly assumed the predictor

variables are not perfectly orthogonal.

However, Green et. al. (1978) believe that if substantial environmental correlations do

exist, then orthogonalizing these attributes is likely to result in unbelievable or unrepresentative

objects. In fact, it is possible that if there is an interaction between two variables and only one is

presented, the respondent reacts towards the presented variable as they would in its normal,

dichotomous context (Hair et. al., 1992). Obviously, these stimuli may have an adverse effect on

the development of the conjoint utility function(s) and any corresponding predictions.

Defining or identifying certain stimuli as unrepresentative or unbelievable appears to be an

area for some research (Stekel et. al. 1991). If high inter-attribute correlations do exist though,

resulting in what the researcher assumes to be unbelievable stimulus profiles, the stimulus displays

can be changed by permuting the set of attribute levels, or by modifying or deleting the unrealistic

profile(s)--a far more common practice (Green et. al., 1990; Green et. al., 1978). Where an

interaction of categorical variables is assumed a priori, one possible solution is to combine the two

variables into one. Whenever these changes are performed, care should be taken that the data
 16

remain reasonably orthogonal and analyzable. This is usually indicated by a condition number (a

measure of factor independence found with many software packages (Steckel et. al., 1991)).

Finally, according to Louviere (1988), it is useful to know that the above considerations

rest upon the following generalizations about significant effects:

a) main effects explain the largest amount of variance in response data, often 80%
or more; b) two-way interactions account for the next largest proportion of
variance, although this rarely exceeds 3%-6%; c) three-way interactions account
for even smaller proportions of variance, rarely more than 2%-3% (usually 0.5%-
1%); and d) higher-order terms account for minuscule proportions of variance.

Data Collection

Conjoint data are usually gathered by two basic methods, two factor evaluation (TFE), or

trade off method, and multiple factor evaluation (MFE), or full profile method (Hair et. al., 1992).

TFE compares attributes two at a time by ranking all combinations of the levels of those two

attributes. MFE describes each bundle of attributes separately and asks the respondent to rank

order or rate the profiles presented to them.

TFE is beneficial when there are few variables, so that a true factorial design may be used.

However, TFE is less realistic than MFE in that rarely do consumers evaluate a product two

attributes at a time. In fact, there may be naturally correlated variables that are not presented

together.

Among the advantages of MFE are a more realistic description of the product/service by

defining level of each attribute, an explicit portrayal of the trade-offs among all factors, and the

existing environmental correlations among attributes. The disadvantage is primarily the number of

stimuli involved. MFE requires some type of fractional factorial design be utilized or,

consequently, respondents may become fatigued and/or resort to simplifying the conjoint task
 17

(Hair et. al., 1992; Green et. al., 1978).

Research indicates that the stimulus presentation method should be as congruent as

possible with the respondent's actual experience with the object. That is, placing the conjoint

study design within the experience of most respondents serves to enhance its realism (Page et. al.,

1989). The increased realism of the full profile method, as well as the advent of computer

programs which can easily analyze the respondent's more complex evaluations of MFE profiles,

has contributed in recent years to a decrease in use of TFE (Wittink et. al., 1989).

A relatively new method for conjoint data collection is the telephone-mail-telephone

(TMT) procedure. Respondents are recruited by telephone screening. The main interview

materials, including questionnaires, stimulus cards, incentive gifts, and other items, are then sent

by mail or by an express service. An approximate time is set for collecting all data by telephone.

The conjoint exercise usually is reserved for the telephone interview. The easier questions can be

self-administered by the respondent; the answers are simply recorded by the interviewer during

the main interview (Green et. al., 1990).

The advantages of the TMT interview are: 1) selection bias is reduced because sampled

populations can be defined and probability sampling methods employed; 2) any difficulties

encountered can be eased by the presence of visual materials and the interviewer; 3) once

respondents are recruited, completion rate is quite high; and 4) all questionnaires will contain

complete data (Green et. al. 1990).

Stimulus Presentation

There are three main methods for stimulus presentation: verbal description (stimulus

cards), paragraph description, and pictorial representation. According to Cattin et. al. (1982),

verbal and paragraph descriptions of hypothetical objects are the most commonly used methods of
 18

presenting the stimuli. More recently, profile cards have become by far the most popular stimulus

presentation method (Green et. al, 1990). Again, the presentation method should closely mimic

the respondent's experience in the real world.

The cards are ranked from most to least preferred, or most to least likely to purchase, etc,

depending on the purpose of the study. In addition, if cards are used, they should be shuffled

before being presented to the respondent. Ranking (nonmetric) is likely to be more reliable

because it is easier for the respondent to say they prefer X more than Y, as opposed to expressing

how much more preferable X is than Y. It also provides more flexibility in estimating different

types of composition rules i.e. a multiplicative model can be derived from an additive one via a

logarithmic, monotone transformation. Rating allows metric measurements, which are easily

analyzed and administered, and it allows conjoint estimation to be performed by multi-attribute

regression.

When comparing ordinal and interval scales, it should be noted that the estimated

parameters derived from nonmetric dependent variables tend to closely satisfy interval-scaled

properties. The main advantage of metric methods is the increased information content

potentially available in these scales.

Estimation Methods

Selecting an estimation technique is contingent upon the type of data collected. Rank

order evaluations require a monotonic analysis of variance (MONANOVA) which is a modified

form of the ANOVA specifically designed for ordinal data. Such analyses estimate attribute part--

worths such that the rank order of their total worth for each stimulus is correlated as closely as

possible with the observed rank order. Among algorithms designed for ordinal data,

MONANOVA is restricted to the part-worths function model.

 19

For data that are assumed to be at least intervally-scaled, many methods, including

primarily ANOVA and ordinary least squares (OLS) regression, can estimate the part-worths of

each level. The important advantage of OLS procedures is that they provide standard errors for

the estimated parameters. As a regression procedure, however, OLS is subject to problems such

as retaining enough degrees of freedom in estimation. The situation is exacerbated in conjoint

studies because the ratio of observation to variables tends to be small. There is evidence that

when the ratio of observations to variables falls too low and the residual degrees of freedom are

too low, the coefficients derived from OLS are unstable (Hinta, 1990; Mollet, 1989; Tabachinick

& Findell, 1986; cited in Chrzan, 1990).

In addition, there are estimation methods for paired-comparison data. These are primarily

choice probability models which include LOGIT and PROBIT. Among their limitations, they

assume that the paired comparisons are "probabilistically independent." The appropriateness of

this model to estimating data gathered in the paired-comparison or TFE method is offset by the

inefficiency of this data collection and stimulus presentation method. Also, if the data are

gathered as ranked, parameter estimation is probably woefully unrealistic. In their behalf, the

choice probability models appear to have very good predictive validity when used under

appropriate conditions (Green et. al., 1978).

According to Cattin et. al. (1982), regression analysis became the most common method

during the early 1980s. The reason for this development is that simulation research studies

(Cattin & Wittink, 1982; Carmone, Green & Jain, 1978) have found that OLS applied to integer

ranks (where the rank ordered dependent variable is redefined as a "pseudo-interval-scaled

variable," e.g. 1, 2 etc. depending on the stimuli's rank) results in parameter estimates that are

very close, in predictive validity, to those obtained by nonmetric algorithms such as

 20

MONANOVA. In fact Cattin and Wittink (1982) report that the results from OLS and

MONANOVA were virtually indistinguishable. This evidence and that presented by information

integration theory suggest that it is safe to assume ranking data to be intervally scaled, though the

standard errors and statistical tests derived from an OLS analysis with ranked data are not strictly

valid.

Specifically, OLS appears to be the better approach with a compensatory decision making

model while the other estimation procedures are preferred when a lexicographic structure is

assumed. Even these methods differ by only very small amounts (Green et. al., 1978).

Summarily, the estimation methods do not seem to differ much in their predictive validities other

than under the aforementioned conditions. The best (albeit often impractical) way to determine

which method is most robust is to use both a metric and nonmetric method in estimation.

Estimated Parameters

Conjoint analysis can not only assess each attribute level's part-worth value, but can also

assess the importance of each attribute relative to the other attributes. Since part-worth estimates

are on a common scale, the attribute with the greatest contribution to overall utility or the highest

range of part-worths will be the most important attribute. This is accomplished by dividing each

attribute's range value by the sum of all range values. This results in a relative importance value

for each attribute.

Within each attribute, conjoint analysis derives relative importance scores for each

attribute level from the ranking or rating data. These utility scores are analogous to regression

coefficients and their range is used to find the relative importance of each factor. This information

is useful when deciding which combination of attribute levels is best for a product/service or

predicting sales given specific combinations of attribute levels (Hair et. al., 1992).
 21

The conjoint function can then be applied at the aggregate (group) or disaggregate

(individual) level. In the disaggregate approach, each respondent is modeled separately and the

researcher appraises the behavior of each respondent relative to the model's assumptions. This

approach also allows for the exclusion of respondents who demonstrate such poor preference

structure that it is assumed they did not perform the preference task correctly (Hair et. al., 1992).

When using the aggregate approach, the analysis fits one model to the entire set of

respondents. This approach is not useful for predicting individual behavior or interpreting

attribute values for any single person. Unless the researcher is definitely dealing with a population

relatively homogeneous in behavior with regard to the attributes, is interested in aggregate

behavior (i.e. market share), or is constrained to use an aggregate approach for non-statistical

considerations, aggregate analysis is not an appropriate line of action. Thus application to the

individual or group level depends on the primary purpose of the study.

Simulation

At this point the researcher has an understanding of the relative importance of each

attribute and the impact of differing levels of that attribute at either the group or individual level.

It is common for many commercial conjoint studies to have as their ultimate objective the

prediction of respondent's behavior contingent upon various changes in the levels of determinant

attributes. This is accomplished through the use of market simulators.

The word "simulator" is used here to mean prediction of individual behavior under

hypothetical, constructed conditions. Typically, choice simulators are computer programs written

for each competitive scenario which attempt to accomplish three things: 1) predict the expected

overall utility of each individual to each treatment combination in each competitive scenario; 2)

identify the best treatment for each individual in each scenario (best is usually defined as the
 22

treatment receiving the highest utility); and 3) simulate the choices of all respondents for specific

stimuli sets and predict market share for each stimulus by aggregating the data (Hair et. al., 1992;

Louviere, 1988).

In most of its applications, each individual's part-worth utilities and background

description (current purchase behavior, socioeconomic, demographic, and psychographic

characteristics) are entered into the simulator. Then, a competitive scenario is constructed by

creating products consisting of various combinations of the attribute levels which were varied in

the conjoint task. An important assumption at this juncture is that attribute information about the

brand(s) has been adequately communicated to the market and that the market accepts or believes

this information (Louviere, 1988). Also, other marketing mix factors such as advertising,

distribution, and promotion may affect actual market shares but are not manifested in the product

attribute evaluations (Wyner, 1992).

It is in this context of competing products/services that a utility for each of the competing

items is computed for each individual (again, using a first-choice rule, the individual is assumed to

choose that item displaying the highest utility to them). The frequency of "first-choices" for each

individual are then summed and expressed in an aggregate context (Green et. al., 1978). The

expected market share of a treatment combination is estimated to be a ratio of the number of

individuals for whom a particular treatment was best by the total number of individuals (Louviere,

1988).

Many variations on this relatively simple procedure have been used in proprietary studies.

If current market information on brands is available, it is possible to examine switching behavior

and cannibalization as new products are entered into the simulation environment. Also, if

background data such as purchase behavior is available, it is possible to obtain consumer

 23

segments; hypothetical market shares or shares of choice and then be cross-tabulated by these

segments (Green et. al., 1978). Among the segments that can be analyzed are current brand users

and competitive brand users. This allows separate strategies to be created for retaining current

customers and attracting new ones (Wyner, 1992).

Validity

The internal validity of conjoint analysis is usually reported in terms of a correlation

coefficient measuring the relationship between the original versus the estimated values of the

dependent variable. Assessing predictive validity involves predicting the respondent's rank order

of the choice set using the estimated preference function. By repeating the predictions for each

respondent in the sample, a frequency distribution of the number of respondents choosing the

first, second, etc. most preferred stimulus is obtained.

According to Scott and Wright (1976), other consistency checks can test the validity of

the parameter estimates. First, the signs of the parameters should agree with a priori

expectations. Second, the parameter estimates for different sub-populations should be in the

hypothesized direction. These and other measures of face validity are acceptable and useful

(Green et. al., 1978).

The primary goal of the study was the incorporation of a new pseudo-simulation into a

conjoint analysis on the grounds that the procedure was likely to overestimate preference

parameters within the context of relatively low involvement. Since the technique was based on

models of information processing, this inclusion was plainly missing in other studies. The

literature supported the contentions that 1) low involvement information processing differs from

complex decision making in both rigor and extent (Wyner, 1992; Engel et. al., 1990); and that 2)

using both compositional and decompositional techniques, it was possible to account for the
 24

assumed systematic bias of customers.

A review of the current literature also found that although the technique is robust in its

statistical assumptions, it is uniquely dependent upon a carefully designed conjoint task which

allows the easy communication (on the respondent's part) and accurate estimation (on the

experimenter's part) of preference structure.

Specific to a conjoint analysis of this product line (calf milk replacers), the literature

supported the following: 1) a reasonable (4-8) number of attribute levels which were theoretically

and practically appropriate (Cattin et. al., 1982; Page et. al., 1989). This in turn allowed a

smaller stimulus set to be constructed and made the respondent's job much easier; 2) an additive

part-worth model allowing for estimation of discrete variable parameters, the most robust for

ordinal data and the model best representing respondent preference structure in these contexts

(Cattin et. al., 1982); 3) a rank order conjoint task with product attributes presented in a realistic

medium, again for accurate and easy responses (Kumar et. al., 1991; Hair et. al., 1992; Green et.

al., 1990); 4) a robust OLS algorithm for conjoint parameter estimation (Chrzan, 1990; Green

et. al., 1978); and 5) a TMT procedure which would allow accurate data collection (Green et. al.,

1990). A detailed account of the application of these suggestions follows.

 25

METHOD

Subjects

The sample consisted of dairy farmers in New three eastern states. A list of farmers was

obtained from a national agricultural data bank, and it was from this list that the sample was

drawn. To ensure geographical representativeness of the market, a proportional quota sampling

scheme was employed.

In order to qualify for participation, the farmers were required to have at least 50 cows

and to have purchased calf milk replacer in the last 6 months. Complete data was obtained from

358 farmers with 195 in one state, 127 in the second state and 36 in the third.

Data Collection

A field research service was contracted to handle all data collection responsibilities and the

aforementioned TMT procedure was used. Dairy farmers were contacted over the telephone and

administered a screener, found in Appendix A. If they qualified (50 cows and calf milk replacer

purchased in the last six months) and consented to participate in the study, they were asked for

additional information about their calf milk replacer use, including primary supplier, fat content,

protein content, protein source and price--the five factors to be included in the conjoint design.

The farmers were asked to respond "off the top of your head" without any outside references.

This was later used as an indication of product involvement and in the creation of a knowledge

discount variable for weighting self-explicated values.

Participants were then sent the materials needed for the telephone interview and were

contacted within two weeks of the receipt of the materials. Included in the mailing were a copy of

the questionnaire, 16 index cards which constituted the conjoint task, and five dollars. The

respondents completed the questionnaire and performed the card sort at their leisure and either
 26

recorded the ranked stimulus cards on the questionnaire or simply relayed them to the interviewer.

During the interview any questions or problems on the farmers' part were answered.

Survey Instrument

The survey instrument consisted of a five-page questionnaire (found in Appendix B). The

first section of the questionnaire asked respondents about their calf herd, milk production, and calf

milk replacer consumption. Participants were also asked to refer to a purchase tag from the bag

of the calf milk replacer they most recently purchased. These tags contained "referenced"

information relevant to the formation of a knowledge discount to be explained later. Respondents

were then again asked about the five primary characteristics of calf milk replacer: primary

supplier, fat content, protein content, protein source and price.

The second section dealt with a new product concept. Participants read a description of

the possible new product and then indicated, on a five-point Likert scale, how willing they would

be to buy the product under various circumstances. Also, in open-ended questions, they indicated

what they liked and disliked about the product. The nature of the product description is

proprietary to the study sponsor and cannot be released in this report. Let it suffice to note that it

defined the characteristics and benefits of a new protein source labeled "Soygrow."

The last section contained a proprietary questionnaire designed to elicit psychographic and

lifestyle information from the dairy farmers. This section was not designed to be used with the

conjoint analysis and was excluded from further analysis.

The Conjoint Task

The calf milk replacers were described in terms of five factors. Factors were chosen on

the basis of past milk replacer studies and a priori client specifications. Attributes used were

company name, protein source, protein level, fat level, and price (per 50 lb. bag). There were
 27

four attribute levels within each factor.

For protein level, fat level, and price, the attributes chosen were determined to be realistic

levels which one could expect to find when purchasing calf milk replacer. The same rationale was

used for protein source, but with the addition of "SoyGrow." Before ranking the cards,

participants were told that some of the cards mentioned "SoyGrow" as a protein source and that

this was the same ingredient described in the new product concept. For company names,

"Farmco," "Plowshare," and "Harvest" were chosen because of their relatively strong, and rival,

market shares. A fourth level was "Other" which was used as a variable measure of company.

With these factors and respective levels, an orthogonal array was produced using the

Bretton-Clark Conjoint Designer program. A part-worth model was specified for all factors. The

program generated 16 treatment combinations and printed them onto individual cards. After

initial cards were produced, some of the combinations of levels were changed to produce more

realistic combinations. Generation of this new orthogonal design produced some weak,

systematic correlations among the factors. The program indicated--through the condition

number--that these correlations were not substantial and that the resultant data would be

analyzable.

For the conjoint task, a full profile method was used. As stated, participants were mailed

16 index cards on which were profiled 16 different possible product combinations. These product

or treatment combinations were composed of the five factors to be tested by the conjoint analysis.

Respondents were instructed to arrange the 16 cards in rank order according to their

willingness to purchase each product if it were actually available locally. The top card (rank 1)

should be the product they would be most willing to buy and the last card (rank 16) should be the

one least likely to be purchased under such conditions. The cards were labeled with randomly
 28

picked numbers. Appendix C lists the 16 products which were printed on the randomly labeled

cards.

Method of Analysis

Upon collecting the questionnaire data, the contracted field service produced a cleaned

data disk for analysis. Initial frequencies were run on the data for verification purposes. All

erroneous data--impossible product features, redundant rankings, data entry mistakes--were

deleted from the database, leaving a total sample of 307 respondents.

With a clean data set, an ordinary least squares regression analysis was performed on the

rankings using SPSS Categories software. This produced raw, aggregate utilities for each

attribute and relative importance values for each factor. The same model was applied to each

individual to obtain disaggregate utilities and values to be used later. The utilities were then

standardized by multiplying them by 10 and adding 100. This was done to eliminate negative

signs, excessive decimal places, and means of zero (which all factors had). With the completion

of this step, one has in their hands raw conjoint output on which a variety of traditional operations

can be performed.

Interpolation

Next, the intermediate levels between the continuous variables of protein level, fat level,

and price were constructed by interpolation. Specifically, this was accomplished by taking the

absolute utility difference between two consecutive attribute levels. This difference was divided

by the number of percentage (or dollar) increments separating the two. This number was then

added to the lower of the two levels (it could also be subtracted from the higher). Linear

interpolations were produced in this manner for these variables.

The resultant data set was divided into Farmco customers and non-customers. Customer
 29

status was defined as farmers who stated that Farmco was their primary supplier of calf milk

replacer, and so used one of the four Farmco replacers. Customers were further categorized into

either "Creamy" "Creamier," or "Creamiest" groups based on which Farmco replacer they used.

A fourth Farmco calf milk replacer was excluded from any further consideration based on its

insignificant market share.

This information was primarily for use in the pseudo simulation and was constructed using

the referenced tag information. The customers described their calf milk replacer's protein source,

protein and fat level. It was inferred from these reports which of the four Farmco replacers they

used.

After this characterization, two data sets now existed, one at the aggregate level and one

at the disaggregate level. Each consisted of interpolated utilities for both customers and non-

customers.

Weighting

Two weights were now applied to each individual's utility scores. The first weight was

applied to better represent the actual market and to emphasize heavy users' preferences.

Specifically, it was a ratio constructed by dividing the number of calves each farmer owned by the

total number of calves within the sample. This ratio proved too stringent in that the largest

weights (derived from those farmers with the most calves) amounted to no more than 3%. The

maximum number of calves in the sample--200--was substituted as the denominator of the ratio,

correcting for this depreciation.

The second weight was applied separately but not differently for customers and non-

customers (it was hypothesized that customers and non-customers would differ considerably as

market segments) and was labeled a "knowledge discount." It was an accuracy adjustment
 30

calculated by using the "top of mind" information recorded during the initial screening phase and

the referenced tag information obtained during the interview. This knowledge discount was for

the purpose of taking into account and correcting for the low involvement hypothesized to be a

characteristic of this product class.

The discrepancy between "top of mind" recall and tag information was calculated on all

five factor for every individual. People were scored "1" if they were accurate, and "0" otherwise.

If respondents did not remember off the top of their head who manufactured their calf milk

replacer, then the utility scores for company name should be discounted to that degree, since for

those respondents the name of the manufacturer would not likely be an important determinant of

product purchase. This produced a frequency distribution which indicated how many farmers

were accurate or knowledgeable in their use of calf milk replacer.

A three-percentage-point margin was allowed for fat and protein level and $2.50 was

allowed for price. That is, a respondent was considered accurate if his assessment of the

product's protein or fat were within 3% of the true (tag) level of those ingredients. Similarly, if

their estimates of price were within $2.50 of the true price, they were judged accurate. For the

discrete variables of protein source and company, no error was allowed.

The results of this process allowed a farmer to be inaccurate about his calf milk replacer's

protein level, for instance, by three percentage points and still be classified as accurate enough for

our purposes. The final weight was the percent of respondents who correctly identified their calf

milk replacer's company, protein levels, etc. The knowledge discount procedure is summarized:

KD = 1 - (Dk + Wrong) / N

where Dk = number of respondents who indicated they did not know the information off the top

of their head; Wrong = number of respondents who inaccurately reported information when
 31

comparing "top of mind" and tag information; and N = number of customers/noncustomers.

These weights were then applied by simply multiplying them as ratios to the standardized,

individual utility scores. Applying the calf weight and knowledge discount to the matrix soon

produced numbers which appeared uninterpretable due to the large ratios employed.

It was decided to standardize the procedure once again to eliminate negative signs and

excessive decimals. To do this, the standardized utility scores were multiplied by the calf weight.

This product was standardized by multiplying it by 10 and then adding 100. The knowledge

discount was multiplied by this product and the resultant scores appeared intuitively and logically

valid. The algorithm for this procedure is summarized:

Uc = [{uc * 10} + 100 * {Cw}] * 10 + 100 * {Kd}

where Uc = the weighted utility for an attribute level; uc = the original raw utility for an attribute

level; Cw = calf weight; and Kd = knowledge discount.

The first subject's raw utilities for the company factor are shown below. Using these

scores as an example, the algorithm is applied:

ID = 0001 Company1 Company2 Company3 Company4

Raw -0.71 -0.21 0.78 0.15

Standardized 92.81 97.81 107.81 101.56

Calf Weight (ID0001 = .075) 6.96 7.33 8.08 7.61

Standardized 176.17 169.6 173.35 180.85

Knowledge Discount (ID0001 = 73.27 74.89 78.13 76.1

.4319 for Company)

The calf weight for this individual was .075 because he/she had 15 calves (15/200 = .075).

The customer's knowledge discount for the company factor was .4319. Notice that the final
 32

utility scores retain a relative magnitude and the same rank order.

Maximum utilities were identified for each individual for each factor used to identify

current products. Frequencies of the maximum utility along with its corresponding level were

obtained for customers and non-customers; customers were further delineated by Creamy (n=12)

Creamier (n=27) and Creamiest (n=15) status.

This information had two purposes. The first was to examine which level of each attribute

the sample preferred most, indicated by the relative market share of the sample. The second use

for this information concerned the pseudo-simulation. Customer preferences for each of the five

factors were profiled based on their maximum utility. Within these customer groups, relative

market shares were obtained from the client: Creamy (22%), Creamier (55%), and Creamiest

(20%), and this information was used to weight customers for the subsequent overstatement

discount (see Table I).

Pseudo-Simulation

At this point in the analysis, the client submitted a battery of hypothetical products (a full

profile of brand, fat, protein, protein source and price) for analysis in various simulated contexts.

These concepts were derived after examination of the final conjoint utilities. The simulation

environment consisted of primarily Farmco calf milk replacers; no information was obtained

about competitors' products and retaliatory actions (responses of competitors to Farmco

reactions) were ignored. This portion of the analysis used the compositional information obtained

in the questionnaire. By obtaining this information from both customers and non-customers, it

was possible to predict the general reaction of the market under prescribed conditions.

The primary purpose of this pseudo-simulation was to examine the utility of each product

(hypothetical or real) within different contexts. Using the market share and other information
 33

provided by the client, it was also possible to examine switching behavior and cannibalization

among customers for each competitive scenario of interest. Additional weights needed to be

calculated to accurately simulate respondent behavior in the "real world." Using non-conjoint

information obtained through the survey, these considerations were handled differently for

customers and non-customers.

Customers

An "overstatement discount" was created to conservatively weight the halo effect of

Farmco SoyGrow products. The rationale for the attention to possible overstatement was that

respondents would overstate preferences for issues/items that were "obviously" considered a good

idea by the research sponsor. Simply, these were instances of demand characteristics.

Because "SoyGrow" was introduced in the product concept prior to the conjoint task, and

since "SoyGrow" was presented in a relatively positive light by the research sponsor, it was

reasonable to assume that reactions to "SoyGrow" during the conjoint task included some degree

of overstatement.

Conjoint tasks inherently include some amount of noise due to lack of information of

available products, lack of attention to characteristics of current products, to the availability of

other products that might work and low product involvement. The overstatement discount was

used to account for some of this accumulated noise.

The correction for customers was calculated by figuring the percentage of Creamy,

Creamier and Creamiest customers, defined by their current product, who had as their maximum

utility some other actual Farmco product in the pseudo-simulation environment; that is, they

preferred another Farmco product over their current Farmco product. This percentage was

averaged across the current product line for each of the three customer groups (the fourth Farmco
 34

product was reentered into the pseudo-simulation--not the customer group, but the product).

Each of these three figures were then multiplied by their respective market shares and once again

averaged. The resultant figure was subtracted from the percent of customers projected to shift to

another (hypothetical) product offered by the client. That is, this figure was subtracted from the

percentage of Creamy, Creamier, and Creamiest customers who indicated through their utility

scores that they preferred another product (concept).

This overstatement correction was generated twice, the first assuming that Creamiest and

Creamier actually contained "milk and soy." The second assumed that Creamier and Creamiest

actually contained Soygrow. The rationale for this operation was information obtained from the

client company which suggested Creamier and Creamiest customers were not aware that

SoyGrow was already an ingredient in their calf milk replacers. The process by which these two

likelihood discounts were created is summarized in Table I. Non-customers

Considering the popular assumption of conjoint models that, if a buyer's utility score for a

new hypothetical product concept is higher (more attractive) than is the utility score for his/her

presently purchased product, this buyer will purchase the new product rather than the current one.

It is often reported that when the percentage of buyers likely to shift is calculated from this

assumption, the percentage shifting is far too high. That is, the calculated percentage is far higher

than the actual percentage that does shift when that new product is actually introduced in the

market.

Why does this occur? One explanation is that the assumption does not take into account

such things as inertia (laziness, apathy) and low involvement among buyers. Another explanation

might be that a substantial increment must be present before the new product is perceived to be

truly superior to the current (analogous to the concept of "just noticeable difference" in
 35

psychophysics).

An indication of the magnitude, if any, of such a bias was the readiness of customers to

buy a given product with a lower utility score even though another product with a higher utility

score is readily available to them. One of the major problems in a conjoint analysis is the

susceptibility of the procedure to greatly overestimate the number of buyers likely to shift to a

new product (concept) if it were to be made available. For the pseudo-simulation to generate

useful results, it was necessary to reduce the impact of this bias.

A "likelihood discount" was calculated to get some indication of how many non-customers

would be willing to shift to customer status. This weight, too, was created using respondent's

direct assessments of product concepts. Recall that respondents were asked to react to a new

product concept which may or may not be available, and which may or may not be manufactured

by Dairyear, a well-known agriculture company.

Respondents were asked their willingness to purchase the new product if it was manufactured by

their primary supplier and if it was manufactured by a supplier other than their own.

The discount itself was calculated by taking the ratio of the proportion of non-customers

who indicated that they 'definitely' or 'probably' would buy the new product if manufactured by a

supplier other than their own by the proportion of non-customers who indicated they 'definitely' or

'probably' would purchase the new product if manufactured by their primary supplier.

Sixty-three percent (63%) of non-customers indicated that they 'definitely' or 'probably'

would purchase the new product if manufactured by their primary supplier. Thirty percent (30%)

of non-customers indicated they 'definitely' or 'probably' would purchase the new product if

manufactured by a supplier other than their own.

 36

Table I: Summary of the derivation of the likelihood discount for customers using protein sources

B and C.

Customers
Creamy Creamier Creamiest
Market share 22% 55% 2%

Protein source level B

Creamy NA 44% 53%
Creamier 8% NA 20%
Creamiest 8% 70% NA
Fourth 0% 33% 20%

Mean: 5% 49% 31%

Weighted by market share: 1.0% + 27.0% + 7.0% = 35%

Protein source level C

Creamy NA 33% 27%
Creamier 0% NA 20%
Creamiest 50% 70% NA
Fourth 0% 19% 0%

Mean: 17% 41% 16%

Weighted by market share: 4.0% + 23% + 3.0% = 30%
 37

The ratio, then, would be: 30% / 63% = .4761

The rationale for this procedure was to get a rough estimate of the number of non-

customers who might be willing to switch to the client's product line of calf milk replacers.

Respondents' weighted current utilities were used to calculate a mean utility for each respondents'

current product (as indicated by tag information).

The likelihood discount was applied to only those non-customers who had a current

product mean utility which was lower than the mean utilities of any Farmco products. That is, the

likelihood discount was applied only to respondents who, according to their utilities, might

possibly switch to a Farmco calf milk replacer. A percent of non-customers preferring, for

example, "Creamy" calf milk replacer over their own would be multiplied by .47. This was done

regarding "Creamier" and "Creamiest" products also, which were calculated as having both milk

and soy, and, milk and Soygrow.

For instance, if 32%, 20% and 15% of non-customers indicated they would shift to

Creamy, Creamier and Creamiest calf milk replacers, respectively, these percentages would be

multiplied by .47. This would produce percentage amounts of 15, 9 and 7 which would be

averaged: 10. This average percentage, 10%, would then be treated as the overstatement

discount and would be subtracted from the percent non-customers shifting. The actual

percentages were 19% for milk and Soygrow tables and 18% for milk and soy tables.

Environment

Hypothetical products were entered into the pseudo-simulation environment three at a

time. To this were added the three current Farmco calf milk replacers. Then, customers' and

non-customers' weighted overall utilities were composed for each product in the environment.

For instance, if a hypothetical calf milk replacer was entered with protein level B, fat level B,
 38

protein source D and priced at $32, each respondent's respective, weighted utilities for those

factors were summed to obtain overall preference.

Finally, the amount of respondents who had greater overall preference for at least one of

the hypothetical products than for their own current product was expressed as a percentage of the

entire sample. This was an indication of the market share that could be captured by that particular

product.
 39

RESULTS

Traditional

Table II shows the original output from the OLS procedure. As can be seen from the

table under the first column, protein source was the most important attribute relative to the

others. Within the protein source attribute, level A has the lowest utility to the sample, while

protein source level D obtained the highest utility. By comparison, the company attribute had the

lowest relative importance, with the variable company name "other" having the highest utility.

From the information in Table II, it could be inferred that the product with the maximum

amount of utility for the entire sample would be comprised of protein source D, protein level B,

fat level C, would be priced at $26 and would be the farmers' own brands.

A utility function was also obtained at the disaggregate level and Table III is an example

of one individual's summary table. This individual also attaches the most importance to protein

source as an attribute. They also have the same utility profile as the aggregate output, with

protein source A as the lowest and protein source D as the highest. However, this individual

values company name more than the sample as a whole, and within this attribute they have a high

utility for Harvest.

Table IV is the result of separating the sample into customer and non-customer segments

and averaging the disaggregate results. It also shows the results of linear interpolation between

attribute levels, standardization and application of the calf weight and knowledge discounts.

The first page of the table shows the results for protein source and company. Notice that

for both customers and non--customers protein source D has the highest utility followed by

protein source C. Regarding the company attribute, for Farmco customers, "Farmco" has the
 40

highest utility and "other" has the lowest. For non-customers, "other" has the highest followed by

Harvest, Farmco and Blue Seal, respectively. These findings are as expected.

The second page of the table shows the interpolated attribute levels for the continuous

variables. The original attribute levels and their respective utilities are in bold. For protein level,

both groups show a corresponding order of preference, with level B having the highest utility

followed by level D, level C, and level A.

For fat level, there is a difference between the groups, with customers having a higher

utility for fat level D over fat level C. For non-customers this order is reversed. However both

groups have fat level A as their lowest utility with level B having a bit more utility.

Perhaps the most interesting attribute is price. Here the lowest price ($26) have the

highest utility for both groups. The next highest price ($30) reduces the utility a bit. But when

the price is raised even more to $34, the utility for both groups increases. For the highest price

($42) the utility for both groups plummets. The shape of this function seems to suggest that it

would be better to raise the price of a calf milk replacer above $30 than keeping it below $30.

This could keep the utility reasonably high without substantially losing customers.

Table IV also shows that for customers, Farmco has the highest utility. This appears

intuitively correct. Farmco customers also have generally higher utility scores for company name

than do non-customers. As for protein source, both customers and non-customers have the same

rank order of preference, preferring protein source D most and source A the least. Notice again

that customers attach more weight to protein source than non-customers.

For protein and fat levels, though, the magnitude of preference is reversed, with non-

customers attaching more weight to both attributes. For protein level, the order of

preference is level B, level D, level C and level A for both groups. For fat level, customers prefer
 41

Table II: Aggregate results of conjoint analysis showing relative importance of attributes and

overall utilities for attribute levels.

Factor Relative Attribute Utility

importance levels scores

Company 3.41%
Plowshare -0.3224
Farmco 0.1003
Harvest 0.0579
Other 0.1642

Protein source 37.49%

Level A -3.2751
Level B -0.4311
Level C 1.6239
Level D 2.0823

% Protein 11.02%
Level A -1.076
Level B 0.4985
Level C 0.1247
Level D 0.4529

% Fat 25.91%
Level A -2.2755
Level B -0.3908
Level C 1.4268
Level D 1.2395

Price 22.17%
$26 1.0823
$30 0.4382
$34 0.5652
$42 -2.0858

Sums to 100% Constant =

8.1848
 42

Table III: Summary table of conjoint analysis showing relative importance of attributes and

overall utilities for attribute levels for an individual respondent.

Factor Relative Attribute Utility

importance levels scores

Company 14.86%
Plowshare -1.4063
Farmco -0.4062
Harvest 1.3438
Other 0.4687

Protein source 60.81%

Level A -5.9063
Level B -2.2813
Level C 2.8438
Level D 5.3438

% Protein 7.43%
Level A 0.7187
Level B 0.0937
Level C -0.6562
Level D -0.1562

% Fat 15.54%
Level A -1.7812
Level B 1.0938
Level C 0.0938
Level D 0.5937

Price 1.35%
$26 0.0938
$30 0.0937
$34 -0.1563
$42 -0.0312

Sums to 100% Constant =

8.4063
 43

Table IV: Mean utility scores for customers and non-customers by attribute level.

Mean utilities Mean utilities

Customers Non-customers

Company
Plowshare 168.96 91.67
Farmco 175 94.21
Harvest 170.51 94.8
Other 166.17 95.55

Protein source
Level A 72.59 48.7
Level B 84.88 56.77
Level C 92 64.32
Level D 95.17 65.34

% Protein
Level A 39.14 88.93
interpolation 41.22 92.94
Level B 43.31 97.05
interpolation 42.92 95.69
Level C 42.54 94.33
interpolation 42.56 94.53
interpolation 42.57 94.73
interpolation 42.59 94.93
Level D 42.6 95.13

% Fat
Level A 39.21 72.75
interpolation 39.86 74.54
interpolation 40.51 76.33
interpolation 41.16 78.13
interpolation 41.18 79.92
Level B 42.47 81.71
interpolation 43.4 83.64
interpolation 44.33 85.58
interpolation 45.26 97.51
interpolation 46.18 89.44
Level C 47.11 91.37
interpolation 47.28 90.41
Level D 47.46 89.46
 44

level D the most with a corresponding drop in preference for each lower level. Non-customers

prefer fat level C the most, followed by level D and B and A.

Regarding price, Figure 3 shows the average utility of non-customers and each customer

group. The dashed vertical lines labeled "Creamy," "Creamier," and "Creamiest" are the

respective current prices for each product. As can be seen, there is a noticeable ogive shape to

the price function for each group. In fact, "Creamy" customers have $34 as their most preferred

attribute level. This graph also displays differences in magnitude of preference, with "Creamiest"

customers attaching the least weight to price and non-customers and "Creamy" customers

attaching the most weight.

Figure 4 is a bar graph showing the distribution of maximum utilities within each factor,

except for price, for the entire sample. This distribution is shown as a percentage of the sample

after all respondents with ties or draws among their utility values were excluded. This left 191

respondents. This exclusion was deemed necessary because including their deadlocked scores

would have been difficult to interpret and would have resulted in the generation of an inordinate

number of graphs. Judging from this truncated sample then, the best product to offer based on a

first-choice rule would have fat level C, protein level D, protein source D (possibly protein source

C, but the utility scores these percentages are based on have not been discounted), and wold be a

Farmco brand or the farmers' own brand. These results are very complementary with those

obtained from the aggregate output summarized in Table II.

 45

Figure 3: Mean utility by customer status and Farmco brand used.

 46

Figure 4: Distribution of maximum utilities for each factor level except price.
 47

Pseudo-Simulation Output

Tables V through XII show the results of the pseudo-simulation. Specifically, Tables V

through VIII show the simulation calculated assuming Creamier and Creamiest customers are

aware of protein level C in their current calf milk replacers. Recall that there was doubt that the

farmers knew this product attribute was currently in their calf milk replacers. Tables IX through

XII were created assuming the opposite, that the Creamier and Creamiest customers were aware

of only protein level B in their products.

Each tables exhibits across the top five rows the products entered into the pseudo-

environment. The first three products represent the current product line and were entered to

examine switching behavior among the current products as they stand. The rest are the product

concepts offered by the client company. These are categorized into three rough groups. Products

1A through 1C are derivation of the Creamy calf milk replacer. Products 2A through 2E are

derivations modifying price, protein and fat levels for a protein source C product. Products 3A

through 3C are derivations using protein source B and a lower fat level.

The left column of the tables show segments of interest to the client company, including

the three customers groups, all non-customers, and non-customers who indicated in the

questionnaire that their current product was protein level B, fat level C and protein source D.

This last segment was included because the client company wished to know the probable buy-in of

this sizable group of farmers.

 48

Table V: Simulation output

COMPANY: FARMCO (CREAMY) FARMCO (CREAMIER) FARMCO (CREAMIEST)

PROTEIN SOURCE: C C C
% PROTEIN: C B C
% FAT: A C C
PRICE: $38.00 31.00 $34.00
CREAMY 104 100.1 102.2
0.00% 0.00% 28.00%
CREAMIER 86.3 87.6 88.5
11.00% 0.00% 48.00%
CREAMIEST 78.1 76.7 78.2
5.00% 0.00% 0.00%
NON-CUSTOMERS 89.5 88.4 90.7
0.00% 0.00% 5.00%
Protein source D, 88.2 87 90
% Protein B, % Fat C 0.00% 0.00% 10.00%

Table VI: Simulation output

COMPANY: FARMCO (1A) FARMCO (1B) FARMCO (1C)

PROTEIN SOURCE: C C B
% PROTEIN: D D C
% FAT: A A A
PRICE: $38.00 $38.00 $38.00
CREAMY 104 104.3 104.4
0.00% 0.00% 23.00%
CREAMIER 86.3 86.4 86.5
43.00% 35.00% 0.00%
CREAMIEST 78.1 78.2 78.1
17.00% 0.00% 0.00%
NON-CUSTOMERS 89.1 89.7 90.1
5.00% 0.00% 1.00%
Protein source D, 88.2 88.9 88.9
% Protein B, % Fat C 0.00% 0.00% 3.00%
 49

Table VII: Simulation output

COMPANY: FARMCO (2A) FARMCO (2B) FARMCO (2C) FARMCO (2D) FARMCO (2E)
PROTEIN SOURCE: B B B B B
% PROTEIN: C C C C C
% FAT: C C C B C
PRICE: $31.00 $34.00 $33.00 $33.00 $32.00
CREAMY 101.5 102.6 102.2 100.8 101.8
3.00% 28.00% 0.00% 0.00% 0.00%
CREAMIER 88.5 88.7 88.6 87.8 88.5
45.00% 52.00% 9.00% 39.00% 39.00%
CREAMIEST 77.9 78.3 78.1 77.3 78
0.00% 45.00% 18.00% 25.00% 25.00%
NON-CUSTOMERS 91 91.3 91.2 89.6 91.1
6.00% 10.00% 0.00% 7.00% 7.00%
Protein source D, 89.4 89.3 90 87.7 89.9
% Protein B, % Fat C 2.00% 3.00% 9.00% 0.00% 10.00%

Table VIII: Simulation output

COMPANY: FARMCO (3A) FARMCO (3B) FARMCO (3C)

PROTEIN SOURCE: B B B
% PROTEIN: C-2% C-2% C-3%
% FAT: B B B
PRICE: $27.00 $28.00 $28.00
CREAMY 101.6 101.3 99.7
0.00% 0.00% 15.00%
CREAMIER 89 88.7 87.8
48.00% 43.00% 35.00%
CREAMIEST 77.9 77.7 76.6
18.00% 18.00% 32.00%
NON-CUSTOMERS 90.6 90.5 88.5
4.00% 5.00% 11.00%
Protein source D, 89.4 89.2 86.6
% Protein B, % Fat C 8.00% 3.00% 0.00%
 50

Under each product are the group of interests' average utility for that product and, below

that, the percent likely to buy that product. (These percents are the result of applying the

overstatement and likelihood discounts.) Looking at the protein level C tables, Creamy customers

appear to be the least likely of the customers to shift to another product. The protein level B, fat

level C, protein source D group also appears to be quite content with their product. However,

the Creamier customers appear to be very likely to switch products, regardless of what product is

offered.

Other findings of interest include the fact that product 2B would generate much switching

behavior among customers and a sizable shift from non-customers. Product 3C would appear to

capture the greatest amount of non-customers as well as a sizable amount of customers. It must

be remembered that these scenarios are premised on the assumption that the Creamier and

Creamiest customers--as well as all the other groups--are aware of the product characteristics in

each calf milk replacer, especially protein source C.

Looking now at the protein source B tables, there appears to be less switching in general

among customers, but more among the two non-customer groups. This is probably because the

customer groups were weighted down by the less preferred protein source B which Creamier and

Creamiest customers were assumed to be aware of. Conversely, the non-customers' discount was

slightly lower when calculated using protein source B as Creamier and Creamiest attributes.

According to these tables, Creamier customers would still be likely to engage in much switching

behavior and Creamy customers still show much stability. Once again, 2B would appear to be

responsible for much switching behavior, capturing 11% of the non-customers and 11% of the

protein level B/fat level C/protein source D customers.

 51

Table IX: Simulation output

COMPANY: FARMCO (CREAMY) FARMCO (CREAMIER) FARMCO (CREAMIEST)

PROTEIN SOURCE: A B B
% PROTEIN: C C C
% FAT: C B C
PRICE: $38.00 31.00 $34.00
CREAMY 104 98.7 100.9
0.00% 0.00% 0.00%
CREAMIER 86.3 86.8 87.7
9.00% 0.00% 35.00%
CREAMIEST 78.1 75.8 77.4
18.00% 0.00% 0.00%
NON-CUSTOMERS 89.5 86.8 89.1
0.00% 0.00% 0.00%
Protein source D, 88.2 84.7 87
% Protein B, % Fat C 0.00% 0.00% 0.00%

Table X: Simulation output

COMPANY: FARMCO (1A) FARMCO (1B) FARMCO (1C)

PROTEIN SOURCE: A A A
% PROTEIN: C C B
% FAT: D D C
PRICE: $38.00 $38.00 $38.00
CREAMY 104 104.3 104.4
7.00% 15.00% 23.00%
CREAMIER 86.3 86.4 86.5
0.00% 9.00% 9.00%
CREAMIEST 78.1 78.2 78.1
8.00% 8.00% 9.00%
NON-CUSTOMERS 89.1 89.7 90.1
0.00% 0.00% 1.00%
Protein source D, 88.2 88.9 88.9
% Protein B, % Fat C 0.00% 1.00% 4.00%
 52

Table XI: Simulation output

COMPANY: FARMCO (2A) FARMCO (2B) FARMCO (2C) FARMCO (2D) FARMCO (2E)
PROTEIN SOURCE: C C C B C
% PROTEIN: B C B B B
% FAT: C C C C C
PRICE: $31.00 $34.00 $33.00 $33.00 $32.00
CREAMY 101.5 102.6 102.2 100.8 101.8
0.00% 15.00% 0.00% 0.00% 0.00%
CREAMIER 88.5 88.7 88.6 87.8 88.5
39.00% 35.00% 35.00% 35.00% 35.00%
CREAMIEST 77.9 78.3 78.1 77.3 78
25.00% 32.00% 32.00% 8.00% 25.00%
NON-CUSTOMERS 91 91.3 91.2 89.6 91.1
7.00% 11.00% 9.00% 1.00% 9.00%
Protein source D, 89.9 90 90 87.7 89.9
% Protein B, % Fat C 9.00% 11.00% 10.00% 0.00% 11.00%

Table XII: Simulation output

COMPANY: FARMCO (3A) FARMCO (3B) FARMCO (3C)

PROTEIN SOURCE: B B B
% PROTEIN: B B B
% FAT: C-2% C-2% C-3%
PRICE: $27.00 $28.00 $28.00
CREAMY 101.6 101.3 99.7
0.00% 0.00% 0.00%
CREAMIER 89 88.7 87.8
43.00% 43.00% 35.00%
CREAMIEST 77.9 77.7 76.6
18.00% 18.00% 0.00%
NON-CUSTOMERS 90.6 90.5 88.5
5.00% 5.00% 0.00%
Protein source D, 89.4 89.2 86.6
% Protein B, % Fat C 3.00% 3.00% 0.00%
 53

DISCUSSION

From the pseudo-simulation information, the client company could choose the product line

that would maximize market share while still retaining customers. One possible method for doing

so might be to simply add up the average utilities for the desired product line, choosing the line

with maximum utility. One liability with using this method in the pseudo-simulation is the absence

of measured interaction effects in the conjoint parameters. Another problem may be that the

mean utility values for each customers group are skewed by extreme preference score values.

Also, the means may be skewed to outliers (those with especially high preference scores).

Another, and far more important issue for the client company, is cannibalization within the

current product line. It is difficult to make a prediction of switching behavior relative to every

possible product line. One way of circumventing this situation is, again, to choose the four-

product-line which maximizes utility or profit. Judging from the pseudo-simulation, keeping a

derivation of the Creamy product would behoove the company; Creamy has the most stable

customer base and is the most expensive product in the current line.

Addressing prediction models, Tables V through XII were computed using two of three

general choice rules: 1) first choice; and 2) share of preference. The third possible choice rule,

likelihood of purchase, was not used though nothing appears to be impeding its possible future

implementation by the client company.

At the same time, a general measure of likelihood of purchase was taken in the
 54

questionnaire and this was used in the development of the likelihood discount used for customers.

It must be stressed that this was a general measure removed from the conjoint task; one

indigenous obstacle in conjoint studies has been the issue of non-choice. What provision is made

for the consumer who likes none of what is offered or who likes and would purchase more than

one of what is offered? Persons were screened for calf milk replacer use, so all buy one product.

Also, extensive proprietary research on the client's part revealed that farmers typically buy one 50

pound bag at a time and do not buy in bulk. Overall, in the realm of predicting respondent

behavior, the study has appeared to address most prevailing considerations reasonably well.

One other imperfection of the study was the lack of statistics or any feedback on the TMT

data collection. Using a quota sampling technique, it is very important to obtain information on

non-respondents, the lack of which could seriously impede the generalizability of the study.

Besides the issue of representativeness, the stated intent of the study was to improve upon

conjoint's use with a low involvement product. Surely, the presence of an interviewer at the time

of the conjoint task would increase respondent accuracy if not involvement itself. While these

issues were not addressed here, future studies would benefit by substantiating the procedure's

claims. While the substantive implications could be of considerable interest to the client company,

the main thrust of the study was methodological. Regrading the pseudo-simulation, it appears

that the model did indeed take into account the systematic error so inherent in conjoint analysis

The likelihood and overstatement discounts differentially compensated for this error and

the results are intuitively appealing. The client company also found the pseudo-simulation results

attractive, and used them in developing a new product line.

This calf milk replacer example can also serve to somewhat account for the mediating role

of affect or attitude on perceptions. The documented methodology, while lengthy, provides a

 55

wealth of crucial information well worth the extra labor. While the conventional or traditional

output of this conjoint study estimated the utility of product features, the integrative method

extended these conventional applications by providing information which helped elaborate upon

the conjoint's feature effects vis-a-vis a low involvement product category.

These simple measures of affect, product involvement and usage should accompany any

pragmatic conjoint research which has as its goal the accurate prediction of consumer behavior.

An accurate and thorough review of conjoint analysis literature, its history and assumptions would

preclude the absence of these fundamental measures.

The study appears to have covered the intended main issues. Conjoint analysis is typically

used to isolate attributes which determine preference. Such analysis applied to low involvement

products can be greatly enriched if this decompositional perspective is broadened to include

measures beyond simple product preference measures. This approach is likely to apply in other

sets of consumption situations where involvement is high, probably more aptly so than in low

involvement situations. On the negative side, the present study's simulation did not take into

account competitor market shares or possible retaliatory efforts. The possibility exists for

incorporating such scenarios into the present study.

 56

REFERENCES

Carmone, F. J., Green, P. F., & Aryan, K. J. (1978). Robustness of conjoint analysis: Some

Monte' Carlo results. Journal of Marketing Research, 15, 300-303.

Cattin, P., & Wittink, D. R. (1982). Commercial use of conjoint analysis: A survey. Journal of

Marketing Research, 46, 44-53.

Chrzan, K. (1990). A survey version of full profile conjoint analysis. Proceedings of the Society

for Consumer Psychology, 125-129.

Curry, J. (1990). Understanding conjoint analysis: Predicting choice. Quirk's Marketing

Research Review, May, 10-13.

Darmon, R. Y., & Rouzies, D. (1989). Assessing conjoint analysis internal validity: The effect

of various continuous attribute level spacings. International Journal of Research in

Marketing, 6, 35-44.

Engel, J. F., Blackwell, R. D., & Miniard, P. W. (1990). Consumer Behavior. Chicago: The

Dryden Press.

Green, P. E. (1974). On the design of choice experiment involving mulitfactor alternatives.

Journal of Consumer Research, 1, 61-68.

Green, P. E., & Krieger, A. M. (1991). Segmenting markets with conjoint analysis. Journal of

Marketing, 55, 20-31.

Green, P. E., & Rao, V. R. (1971). Conjoint Measurement for quantifying judgmental data.
 57

Journal of Marketing Research, 8, 355-363.

Green, P. E., & Srinivasan. (1978). Conjoint analysis in consumer research: Issues and outlook.

Journal of Consumer Research, 5, 103-123.

Green, P. E., & Srinivasan. (1990). Conjoint analysis in marketing: New developments with

implication for research and practice. Journal of Marketing, 54, 3-19.

Haggerty, M. R. (1985). Improving the predictive power of conjoint analysis: The use of factor

analysis and cluster analysis. Journal of Marketing Research, 22, 168-184.

Hair, J. F., Anderson, R. E., Tatham, R. L., & Black, W. C. (1992). Mulitvariate Data Analysis.

New York: Macmillan Publishing Company.

Holbrook, M. B. (1981). Integrating compositional and decompositional analyses to represent

the intervening role of perceptions in evaluative judgements. Journal of Marketing

Research, 18, 13-28.

Kumar, V., & Gaeth, G. J. (1991). Attribute order and product familiarity effects in decision

tasks using conjoint analysis. International Journal of Research in Marketing, 8, 113-124.

Louviere, J. J. (1988). Analyzing decision making: Metric conjoint analysis. Newbury Park,

California: Sage.

Page, A. L. & Rosenbaum, H. F. (1989). Redesigning product lines with conjoint analysis: A

Reply to Wittink. Journal of Product Innovation Management, 6(4), 293-296.

Steckel, J., DeSarbo, W. S., & Mahajan, V. (1991). On the creation of acceptable conjoint

analysis experimental design. Decision Sciences, 22, 435-432.

SPSS. (1993). SPSS Statistical Algorithms, Second Edition, 39. Chicago: SPSS, Inc.

Wittink, D. R., & Cattin, P. (1989). Commercial use of conjoint analysis: An update. Journal of

Consumer Research, 53, 91-96.

 58

Wyner, G. A. (1992). Uses and limitations of conjoint analysis--Part I. Marketing Research: A

Magazine of Management and Applications, 4(2), 42-44.

Wyner, G. A. (1992). Uses and limitations of conjoint analysis--Part II. Marketing Research: A

Magazine of Management and Applications, 4(3), 46-47.

 59

APPENDICES

APPENDIX A
60
61
 62

APPENDIX B
63
64
65
66
 67

APPENDIX C

CONJOINT TASK

CARDS TO BE SORTED

Company .....................Plowshare
Protein source ..............B
Protein level ...............B
Fat level ...................B
Price .......................B

Company .....................Farmco
Protein source ..............B
Protein level ...............D
Fat level ...................B
Price .......................D

Company .....................Plowshare
Protein source ..............C
Protein level ...............A
Fat level ...................D
Price .......................C

Company .....................Harvest
Protein source ..............B
Protein level ...............D
Fat level ...................D
Price .......................A

Company .....................Plowshare
Protein source ..............D
Protein level ...............C
Fat level ...................A
Price .......................A

Company .....................Farmco
Protein source ..............A
Protein level ...............A
Fat level ...................B
Price .......................A
 68

Company .....................Your supplier

Protein source ..............C
Protein level ...............B
Fat level ...................C
Price .......................A

Company .....................Your supplier

Protein source ..............A
Protein level ...............C
Fat level ...................D
Price .......................B

Company .....................Harvest
Protein source ..............D
Protein level ...............A
Fat level ...................C
Price .......................B

Company .....................Farmco
Protein source ..............B
Protein level ...............C
Fat level ...................C
Price .......................C

Company .....................Harvest
Protein source ..............A
Protein level ...............B
Fat level ...................A
Price .......................C

Company .....................Plowshare
Protein source ..............A
Protein level ...............D
Fat level ...................C
Price .......................D

Company .....................Harvest
Protein source ..............C
Protein level ...............C
Fat level ...................B
Price .......................D
 69

Company .....................Your supplier

Protein source ..............D
Protein level ...............D
Fat level ...................B
Price .......................C

Company .....................Farmco
Protein source ..............C
Protein level ...............D
Fat level ...................A
Price .......................B

Company .....................Your supplier

Protein source ..............B
Protein level ...............D
Fat level ...................D
Price .......................D