Citation
Attentional Contrast During Sequential Judgments: An Explanation for the Number-of-Levels Effect

Material Information

Title:
Attentional Contrast During Sequential Judgments: An Explanation for the Number-of-Levels Effect
Creator:
DE WILDE, ELS ( Author, Primary )
Copyright Date:
2008

Subjects

Subjects / Keywords:
Attribution theory ( jstor )
Capacity costs ( jstor )
Digital cameras ( jstor )
Experiment design ( jstor )
Imaging ( jstor )
Marketing ( jstor )
Refrigerators ( jstor )
Sensors ( jstor )
Term weighting ( jstor )
Verbs ( jstor )

Record Information

Source Institution:
University of Florida
Holding Location:
University of Florida
Rights Management:
Copyright Els De Wilde. Permission granted to University of Florida to digitize and display this item for non-profit research and educational purposes. Any reuse of this item in excess of fair use or other copyright exemptions requires permission of the copyright holder.
Embargo Date:
7/30/2007
Resource Identifier:
71376114 ( OCLC )

Downloads

This item has the following downloads:


Full Text












ATTENTIONAL CONTRAST DURING SEQUENTIAL JUDGMENTS:
AN EXPLANATION FOR THE NUMBER-OF-LEVELS EFFECT

















By

ELS DE WILDE


A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL
OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT
OF THE REQUIREMENTS FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY

UNIVERSITY OF FLORIDA


2005

































Copyright 2005

by

Els De Wilde
































This dissertation is dedicated to my parents.















ACKNOWLEDGMENTS

This dissertation leaves me indebted to many people. First, I would like to thank

my advisor, Chris Janiszewski, for taking up the mentorship role and guiding me through

the dissertation process. Interactions with other members of my committee were also

invaluable in shaping this dissertation. In particular, I would like to thank Alan Cooke for

his conceptual guidance and Richard Lutz and Asoo Vakharia for their feedback. My

thanks also go to my friends in Belgium and in the doctoral program and my colleagues

at HEC Montreal for their support. I would like to thank JoAndrea Hoegg and Andrew

Kuo for their assistance with the data collection while I was at long distance. I am

extremely thankful to my family members for their encouragement throughout the entire

duration of my doctoral studies. Finally, I would like to thank Roberto for giving me the

best reason to finish this dissertation.
















TABLE OF CONTENTS

page

A C K N O W L E D G M E N T S ......... ................................................................................... iv

LIST OF TABLES .............. ............. .......... ............ vii

LIST OF FIGURES ........................ .. .................... viii

ABSTRACT ........................ ............... ................... .........ix

CHAPTER

1 IN TR O D U C TIO N ......................................................................... .... .. ........

2 THEORETICAL BACKGROUND....................................................................4

The Number-Of-Levels Effect.................................. .............................. 4
Existing Accounts of the Number-Of-Levels Effect ................................................4
A ttentional C ontrast........... ................................................................ .......... .. ..

3 EXPERIMENT 1: DEMONSTRATION OF THE NUMBER-OF-LEVELS
EFFECT WITH EXPANDED DESIGN .............................................................16

M eth o d ............................................................................................. 16
R e su lts ...................................... .................................................... 1 7
D isc u ssio n .............................................................................................................. 2 0
E x p e rim e n t 1A ............................................................................................................ 2 1

4 EXPERIMENT 2: TEST OF ATTENTIONAL CONTRAST ACCOUNT WITH
NO DIFFERENCE IN NUMBER OF LEVELS .............. ................................ 28

M eth o d ..............................................................................................2 8
R e su lts ...................................... .................................................... 3 0
D iscu ssio n ...................................... ................................................. 3 1

5 EXPERIMENT 3: TEST OF ATTENTIONAL CONTRAST ACCOUNT WITH
REVERSED NUMBER OF LEVELS.................................................................35

M eth o d ............................................................................................. 3 5
R e su lts ...................................... .................................................... 3 6
D discussion .................................... ................................... .......... 38









6 G E N E R A L D ISC U SSIO N ................................................................ ....................4 1

7 LIMITATIONS AND FUTURE RESEARCH ................................. ...............45

L IST O F R E F E R E N C E S .......................................................................... ....................47

B IO G R A PH IC A L SK E TCH ...................................................................... ..................49
















LIST OF TABLES

Table p

1. Number-of-levels manipulations used in published studies .......................................15

2. A tribute levels from experim ent 1 ........................................... ......................... 26

3. Mean difference scores and simple effects for experiment 1 .............. ...................26

4. Mean difference scores and simple effects for experiment 1A...................................27

5. Mean difference scores and simple effects for experiment 2 .....................................34

6. A tribute levels from experim ent 3 ........................................... ......................... 40

7. Mean difference scores and simple effects for experiment 3 .....................................40
















LIST OF FIGURES

Figure page

1. D esign for experim ents 1 and 1A ....................................................... .....................25

2 D design for experim ent 2 ......................................................................... .................. 33

3. Design for experiment 3............. ................................................................39















Abstract of Dissertation Presented to the Graduate School
of the University of Florida in Partial Fulfillment of the
Requirements for the Degree of Doctor of Philosophy


ATTENTIONAL CONTRAST DURING SEQUENTIAL JUDGMENTS:
AN EXPLANATION FOR THE NUMBER-OF-LEVELS EFFECT

By

Els De Wilde

August 2005

Chair: Chris Janiszewski
Cochair: Alan D. J. Cooke
Major Department: Marketing

Normatively, the importance of an attribute does not depend on the number of

levels on which it is defined, exterior levels being fixed. This dissertation examines a bias

identified in the conjoint literature known as the number-of-levels effect, which occurs

when consumers' derived importance weights increase with the number of attribute

levels. Borrowing from research on attention during category learning, this dissertation

examines psychological mechanisms that underlie the number-of-levels effect.

In contrast to earlier attentional accounts that have been offered for the bias, I

propose that the weighting effect found in the number-of-levels studies is due to the

attentional processes that are inherent to the sequential evaluation procedure.

Analogously to attentional processes during category learning, consumers divide their

attention between attribute dimensions and attribute levels when making sequential

evaluations. Within a given trial, a mechanism of attentional contrast directs attention









away from attribute levels that have been associated with an outcome. I suggest that the

number-of-levels effect is caused by this attentional contrast strategy. When attributes

differ on the number of levels on which they are defined, they also differ on the number

of times a given attribute level has been associated with a judgment (i.e., attribute level

novelty). A process of attentional contrast away from the more familiar attribute level

within each offer results in differential importance weights for the attributes.

In a series of experiments, I examine this attentional contrast account for the

"number-of-levels" effect. Consistent with this account, the weighting effect is shown to

appear when attentional contrast is created in a situation in which there is no difference in

number of levels. Moreover, the number-of-levels effect reverses when attentional

contrast is created on the attribute with the lower number of levels.














CHAPTER 1
INTRODUCTION

Normatively, the importance of an attribute should not depend on the number of

levels on which it is defined, exterior levels being fixed. Yet, several studies from the

conjoint literature show that derived importance weights increase with the number of

attribute levels, a bias which has generally been labelled the number-of-levels effect (e.g.,

Currim, Weinberg and Wittink 1981; Verlegh, Schifferstein and Wittink 2002; Wittink et

al. 1992b). In searching for a method to eliminate this bias, several accounts have been

offered for it. Methodologically, there is evidence that the number-of-levels effect is

related to the data collection method (e.g., Wittink et al. 1992b), the measurement scale

for the dependent variable (e.g., Steenkamp and Wittink 1994; Wittink et al. 1992b), and

the parameter estimation procedure (Wittink, Krishnamurthi and Reibstein 1989).

Although tests of these methodological accounts reduced the number-of-levels effect,

they have never completely eliminated the bias.

The possibility that psychological processes are at least partly responsible for the

number-of-levels effect has been acknowledged since the effect was first documented.

Currim et al. (1981) state that "the respondent may react to the number of levels used for

a given attribute by consciously or unconsciously weighting an attribute more heavily as

the number of levels increases" (p. 72). Although a test of this account has been included

in several studies, it has not received much support during the past two decades. An

assumption underlying tests of this attentional account is that there should be a number-

of-levels effect not only in sequential judgments or choices (e.g., full-profile rank orders,









full-profile paired comparisons, full-profile ratings), but also in self-explicated priors

(Wittink et al. 1992a) and stated importance measures (Steenkamp and Wittink 1994;

Verlegh et al. 2002).

However, I propose that the weighting effect found in the number-of-levels studies

may be due to the attentional processes that are inherent to the sequential evaluation

procedure. Studies in which people learn new categories by making sequential

classifications distinguish between attention allocated to attributes and to attribute levels

(Kersten, Goldstone and Schaffert 1998). They show that within a given trial, a

mechanism of attentional contrast directs attention away from attribute levels that have

been associated with an outcome. I propose that this attentional contrast strategy can at

least partly account for the number-of-levels effect. When attributes differ on the number

of levels on which they are defined, they also differ on the number of times a given

attribute level appears across offers during a sequential judgment task (henceforth

referred to as relative attribute level novelty). More specifically, the levels of attributes

that are defined on more levels are overall relatively more novel than the levels of

attributes that are defined on fewer levels. A process of attentional contrast away from

the less novel attribute level within each offer may result in differential importance

weights for the attributes.

This dissertation proceeds as follows: Chapter 2 discusses the number-of-levels

effect and the methodological and psychological accounts that have been offered for it in

the literature. It also introduces the attentional contrast account from the category

learning literature. In chapters 3 to 5, I examine if attentional contrast can account for the

number-of-levels effect. In chapter 3, I extend the designs that have typically been used






3


in studies that examine the number-of-levels effect and find evidence for an account in

terms of attentional contrast. In chapter 4, I manipulate attentional contrast directly to

examine its possible impact when there is no difference in number of levels. In chapter 5,

I create attentional contrast on the attribute with the lower number of levels. Finally,

chapter 6 discusses the findings and chapter 7 examines the limitations and avenues for

future research.














CHAPTER 2
THEORETICAL BACKGROUND

The Number-Of-Levels Effect

When an attribute has a higher number of levels (exterior values being fixed),

derived importance weights for that attribute increase. For example, Currim et al. (1981)

examined subscription series to performing arts events, using three attributes that were

defined on two levels and three attributes that were defined on three levels. They asked

participants to make trade-offs between pairs that were defined on two attributes: a two-

level attribute and a three-level attribute. The two-level attributes had derived importance

weights between 0.36 and 0.45, whereas the three-level attributes had derived importance

weights between 0.55 and 0.66.

Originally identified between attributes using part-worth estimation within conjoint

analysis, this number-of-levels effect has been replicated using a variety of data

collection methods, measurement scales for the dependent variable, and estimation

methods. Accounts for the number-of-levels effect include methodological and

psychological factors. I will first review the existing accounts for the number-of-levels

effect and then propose a new account based on attentional contrast from the category

learning literature.

Existing Accounts of the Number-Of-Levels Effect

The greater share of research on the number-of-levels effect has dealt with

methodological accounts. This research attributes the number-of-levels effect to the









properties of the measurement scale, the usage of the measurement scale, or the method

of parameter estimation.

Properties of the Measurement Scale

Several accounts of the number-of-levels effect relate to the properties of the

measurement scale. Wittink, Krishnamurthi, and Reibstein (1989) hypothesized that the

number-of-levels effect was partly due to the use of a rank-order measurement procedure.

They showed analytically why the effect occurs for rank-order preferences using least-

squares analysis on a multitude of full factorial and fractional factorial designs. Since it is

impossible to analytically show the same effect exists for ratings, Wittink et al. (1989)

examined whether the number-of-levels effect exists for rating scales empirically,

comparing the results from full-profile ratings with the results from rank-order

preferences obtained by paired comparisons and trade-off matrix rank orders. The full-

profile preferences were obtained using a ten-point scale. Wittink et al. (1989) compared

derived importance of a three-level and five-level price attribute in five categories and

found that the number-of-levels effect was equally strong using rank-order or rating scale

measurement procedures, the implication being that rank-order measurement was not the

sole source of the number-of-levels effect.

Steenkamp and Wittink (1994) proposed that the number-of-levels effect occurs

with a rating scale measurement procedure because the preference ratings lack metricity.

They hypothesized that rating scales' properties are closer to ordinal than to interval

measurement. Since the method of magnitude estimation has been claimed to result in

interval-scale data, Steenkamp and Wittink (1994) used this method to test their

hypothesis. The method of magnitude estimation consists of asking participants to match

their evaluation of a profile to the perceived magnitude expressed on a physical medium.









Steenkamp and Wittink (1994) used numerical estimation (i.e., participants assign

numbers to profiles) and line production (i.e., participants draw lines) as response

modalities. They found that a nine-point rating scale and a magnitude estimation

procedure provide number-of-levels effects of comparable size. However, the magnitude

estimation procedure allows for a test of the metricity of the results of each respondent

separately. Categorizing the respondents as metric versus non-metric, Steenkamp and

Wittink (1994) found a reduced number-of-levels effect in participants whose responses

satisfied the criteria for metricity. In other words, the lack of metricity of the rating

method per se had no impact on the number-of-levels effect, but individual differences in

metricity were identified as a source of the effect.

Usage of the Measurement Scale

Other measurement procedure explanations relate to respondents' use of the

measurement scale. Wittink et al. (1992b) included a balancing manipulation in their

study (i.e., they constructed a condition in which dominated pairs were eliminated from

the procedure), provided respondents made pair-wise comparisons among alternatives.

Using adaptive conjoint analysis, Wittink et al. (1992b) gathered full factorial data on one

product category (notebook computers) defined on six attributes, of which four attributes

alternated between two and four levels, depending on the condition. Wittink et al.

(1992b) found that the number-of-levels effect is considerably reduced when dominated

pairs are eliminated from the procedure. Wittink et al. (1992b) explain this result in terms

of participants' tendencies to give preference ratings toward the midpoint of the scale.

Since preference ratings for dominated pairs are expected to be more extreme than those

for non-dominated pairs, the exclusion of dominated pairs provides less opportunity for

distortion.









Verlegh et al. (2002) propose the number-of-levels effect may be a consequence of

a person's tendency to perceive a constant, absolute distance between levels of an

attribute (i.e., exhibit a uniform distribution tendency). They tested this hypothesis by

comparing results on five different measures of attribute importance, using one product

category (color TVs). The product category was defined on six attributes, of which two

alternated between two and four levels, depending on the condition. Verlegh et al. (2002)

found significant number-of-level effects on two measures of attribute importance:

difference in part worths in a full-profile conjoint and attractiveness ratings for individual

attribute levels. On the remaining three measures (stated importance of a set of attributes,

linear regression coefficients, and paired comparisons varying one attribute's exterior

levels at a time), no significant number-of-levels effect was observed. Verlegh et al.

(2002) interpreted their results as evidence for respondents' tendency to distribute

stimulus representations uniformly over a restricted internal continuum (i.e., the distance

between each level of an attribute was constant) in full-profile and attribute level ratings.

Ceteris paribus, the addition of intermediate levels will lead respondents to distribute the

exterior attribute levels more toward the extremes of the internal continuum.

Method of Parameter Estimation

A second potential source of the number-of-levels effect is the method of parameter

estimation. In their seminal paper, Currim, Weinberg, and Wittink (1981) mentioned the

method of parameter estimation as a possible source of the number-of-levels effect. For

the 2x3 trade-off matrix they used to obtain their empirical results, they estimated the

parameters based on permissible configurations of preferences using ordinary least

squares. They derived the importance weights for two extreme cases: a dominant two-

level attribute and a dominant three-level attribute. They found that the two-level









attribute's importance weights varied between 0.20 (i.e., when the three-attribute level

attribute was dominant) and 0.60 (i.e., when the two-level attribute was dominant).

Analogously, the three-level attribute's importance weights ranged from 0.40 (i.e., when

the two-level attribute was dominant) to 0.80 (i.e., when the three-level attribute was

dominant). In other words, while the range of possible derived importance weights was

not affected by the number of levels, its size was affected by it. Currim et al. (1981) used

this analytical evidence as support for their hypothesis that the method of parameter

estimation is a source of the number-of-levels effect.

Subsequent studies have provided additional support for the method of parameter

estimation account of the number-of-levels effect. Wittink et al. (1989) showed

analytically why the number-of-levels effect occurs on rank-order preferences on a

multitude of full factorial and fractional factorial designs. Assuming an additive, main-

effect model, they calculated raw attribute importance weights for any number of

attributes and any attribute levels using least-squares analysis. Based on the raw attribute

importance weights, they estimated maximum relative weights for each attribute, and the

differences and ratios between the attributes within a design. They found that maximum

relative weights increase with the number of levels and concluded that the method of

parameter estimation is a source of the number-of-levels effect. Wittink, Krishnamurthi,

and Nutter (1982) gathered participants' preference data for hypothetical summer jobs,

defined on four attributes, of which two attributes alternated between two and four

attribute levels, depending on the condition. Two data collection methods were used: a

fractional factorial profile evaluation and a trade-off matrix. Wittink et al. (1982)

estimated the parameters using a non-metric method (LINMAP, Srinivasan and Shocker









1973) and found that the derived importance weights vary with the number of levels on

which an attribute is defined. They concluded that the number-of-levels effect is at least

partially accounted for by the method of parameter estimation.

Perceptual Processes

Since the initial identification of the number-of-levels effect, researchers have

acknowledged that the effect could be a consequence of perceptual processes (Currim et

al. 1981). More specifically, it has been proposed that respondents may consciously or

unconsciously assign more weight to attributes with more levels. Several studies have

included one or more tests for this weighting explanation, but have failed to find support

for the hypothesis (Steenkamp and Wittink 1994; Verlegh et al. 2002; Wittink et al.

1992b). For example, Verlegh et al. (2002) hypothesized that increased attention should

affect the stated importance of a set of attributes and paired comparisons varying one

attribute's exterior levels at a time. However, no number-of-levels effect was observed on

these two measures.

Attentional Contrast

Several of the proposed accounts of the number-of-levels effect can partially

explain the effect. However, no single explanation can wholely account for the number-

of-levels effect. I propose that attentional contrast may be an additional source of the

number-of-levels effect.

The proposed attentional contrast explanation is based on two observations. First,

the occurrence of the number-of-levels effect depends on the preference elicitation

method. Decompositional methods ask respondents to provide multiattribute judgments

for the full-profile descriptions of alternatives and then derive individual attribute level

values from the judgments. The number-of-levels effect is typically found using









decompositional methods. Compositional methods ask respondents to assign values to

each level of an attribute. The values are combined to construct an overall evaluation of

an alternative. The number-of-levels effect is typically not found using compositional

methods.

There are three studies that have used a combination of decompositional and

compositional methods. Wittink et al. (1992a) showed that the magnitude of the number-

of-levels effect was greater for a full-profile procedure than for an adaptive conjoint

analysis. Whereas full-profile procedure is an entirely decompositional method, adaptive

conjoint analysis is a combination of a decompositional and compositional procedure. No

number-of-levels effect was observed in the compositional data. Verlegh et al. (2002)

used a decompositional and compositional procedure and found evidence for a number-

of-levels effect in the compositional data. Yet, the compositional data were collected

after the decompositional data, suggesting the decompositional procedure may have

biased the values provided in the compositional procedure. Steenkamp and Wittink

(1994) also found minimal evidence for a number-of-levels effect on stated attribute

importance that were gathered after the decompositional data.

Johnson (1992) found evidence for a number-of-levels effect using only a

compositional method. He asked participants to assess attribute values directly. However,

the values of two attributes were assessed at the same time. More specifically,

participants were told that a base TV with 17 inch screen and monophonic sound was

worth $200. Then, they were asked to evaluate what other options were worth: 19 inch

screen, 21 inch screen and excellent stereo sound in the 3+2 condition and 21 inch screen,

good stereo sound and excellent stereo sound in the 2+3 condition. Hence, participants









could interpret this task as full-profile evaluation. For example, when evaluating the 19

inch screen, it is likely that they imagined a monophonic sound TV.

Second, the designs used in studies on the number-of-levels effect invariably

compare conditions in which the number of levels of more than one attribute is varied at

once (i.e., varying the number of levels of attribute 1 is confounded with varying the

number of attribute levels of attribute 2). For example, consider an experiment in which

participants respond to a 2x4 design (i.e., attribute 1 has two levels and attribute 2 has

four levels) or a 4x2 design. If one finds that the derived importance weight on attribute 1

is larger in the 4x2 condition than the 2x4 condition, then the weighting difference may

be due to the increased number of levels in attribute 1 or to the accompanying decrease in

the number of levels in attribute 2. In other words, the weighting effect that has been

interpreted as a number-of-levels effect may not (solely) be caused by the increased

number of levels in that attribute. It may also be caused by the reduced number of levels

of other attributes. It is important to note that all published demonstrations of the number-

of-levels effect have this characteristic built into their manipulations. An overview of the

designs can be found in Table 1.

Based on these two observations, I will now propose a new account for the number-

of-levels effect: the attentional contrast account. The attentional contrast account depends

on the idea that people direct attention away from attribute levels they have seen more

often and towards attribute levels they have seen less often. For example, in a 4x2 design,

the first attribute has four levels and each of these levels will be experienced twice in a

full-factorial design. The second attribute has two levels and each of these levels will be

experienced four times in a full-factorial design. Thus, the levels of the first attribute will









be relatively more novel than the levels of the second attribute. The attentional contrast

account proposed that attention will contrast away from the more common levels (i.e.,

attribute 2) and toward the less common levels (i.e., attribute 1). Thus, the attribute with

more levels will be perceived as more important.

The attentional contrast hypothesis has its foundation in the categorical learning

literature. The most successful theories in categorization and learning incorporate some

notion of selective attention (Nosofsky 1986, 1991; Sutherland and Mackintosh 1971).

More specifically, consumers increasingly pay attention to attributes and attribute levels

that are important, and decreasingly pay attention to irrelevant attributes, irrelevant

attribute levels, or both (Goldstone and Steyvers 2001). For example, Goldstone and

Steyvers (2001) showed that selective attention can be directed toward arbitrary

dimensions created by randomly pairing faces. They selected four photographs of bald

heads and generated two dimensions by morphing between the photographs. Each of the

sixteen created faces varied in the proportion of each of the four faces. Based on category

relevance, both sensitization of relevant dimensions and desensitization of irrelevant

dimensions were observed. Moreover, Goldstone and Steyvers (2001) showed that once a

participant had learned to direct attention toward or away from a dimension, this learning

extended to different stimuli.

Kersten, Goldstone, and Schaffert (1998) provide additional evidence that

differential selective attention mechanisms operate on attributes and attribute levels. They

show that within a given trial, a mechanism of attentional contrast directs attention away

from attribute levels that have been associated with an outcome. In a series of

experiments, participants learned verbs that were defined on two attributes: path and









manner. Each verb was always accompanied by a particular value of one or both of these

attributes. One group of participants learned conjunctive verbs (i.e., verbs that could be

distinguished on path and manner). Another group first learned path verbs (i.e., verbs that

could only be distinguished on the basis of path). When this second group learned new

conjunctive verbs (i.e., verbs that could be distinguished based on manner as well as the

same values of path), they attended more to level of manner than the participants who

first learned conjunctive verbs. In other words, since the path values had been associated

with an outcome (i.e., a verb), attentional contrast was directed toward the new manner

values within subsequent trials.

The attentional contrast account is consistent with the observation that the

occurrence of the number-of-levels effect depends on whether the method is

compositional or decompositional. In a compositional method, each level of an attribute

is repeated only once and it is repeated out of context, so no attentional contrast can

operate. In a decompositional method, attribute levels are repeated and this repetition

occurs in the context of other attribute levels that have been more or less repeated, so

attentional contrast can operate.

The attentional contrast account also makes specific predictions in the event the

levels of attributes are unconfounded in an experimental design. For example, imagine

three designs are investigated: a 4x2, a 2x4, and a 4x4. The derived importance weight of

attribute one (two) should be greater (smaller) in the 4x2 than the 2x4 studies, as has been

observed in prior studies. However, the attentional contrast account predicts that the

derived importance weight of attribute one should be greater in the 4x2 design than the

4x4 design because the number of levels of attribute two are more frequent in the 4x2









condition. The attentional contrast account also predicts that the derived importance

weight of attribute two should be equal in the 4x2 and the 4x4 conditions because the

number of levels of attribute one are equivalent. Similar predictions can be made when

comparing the 2x4 and the 4x4 conditions.

The predictions of the attentional contrast hypothesis are not consistent with the

predictions that might be expected if one were simply to count the levels of an attribute.

For example, if the predictions were based solely on the number of levels of an attribute,

the derived importance weight of attribute one should be the same in the 4x2 condition

and the 4x4 condition because the number of levels of attribute one are equivalent. The

derived importance weight of attribute two should be less in the 4x2 condition than the

4x4 condition because the number of levels of attribute two is less in the 4x2 condition.

Thus, an explanation based solely on attribute levels makes a set of qualitatively different

predictions than the attentional contrast hypothesis.

In a series of four experiments, I will test this attentional contrast account for the

number-of-levels effect. In experiments 1 and 1A, I will extend the designs that have

typically been used to examine the number-of-levels effect to test an attentional contrast

account. In experiment 2, I will manipulate attentional contrast directly to examine its

potential impact when there are no differences in number of levels. In experiment 3, I

create attentional contrast on an attribute with a lower number of levels.











Table 1. Number-of-levels manipulations used in published studies
Authors Number-Of-Levels Manipulation
number-of-levels manipulation
Currim, Weinberg and Wittink (1981) between attributes (2 or 3 levels)
Johnson (1992) 3+2 vs. 2+3 (used lists)
number-of-levels manipulation
Orme (1998) between attributes (2 to 5 levels)
Steenkamp and Wittink (1994) 2x4 vs. 4x2
Verlegh, Schifferstein and Wittink (2002) 4x2x2x2x3x2 vs. 2x4x2x2x3x2
3x4x2x2x4x2x2x2x2 vs.
Wittink, Huber, Fiedler and Miller (1992a) 3x2x4x4x2x2x2x2x2
Wittink, Huber, Zandan and Johnson (1992b) 2x2x4x4 vs. 4x4x2x2
Wittink, Krishnamurthi and Nutter (1982) 4x4x2x2 vs. 4x2x4x2
Wittink, Krishnamurthi and Reibstein (1989) 3 vs. 5 levels














CHAPTER 3
EXPERIMENT 1: DEMONSTRATION OF THE NUMBER-OF-LEVELS EFFECT
WITH EXPANDED DESIGN

The objective of the first experiment was to test the attentional contrast account of

the number-of-levels effect by expanding the experimental design that has typically been

used to examine the bias. More specifically, the number of levels of only one attribute

was varied across conditions. The attentional contrast hypothesis (Kersten et al. 1998),

predicted that the number-of-levels effect would be sensitive to the relative novelty of

each attribute level.

Method

Design

The design was a 3 cell between-subjects design (4x2, 2x4, and 4x4 attribute

levels) with three between-subject product replicates (refrigerators, LCD televisions, and

digital cameras). Each offer was defined on two attributes. Participants were asked to

provide their reservation price for each product offer. This dependent variable should not

be sensitive to the uniform distribution problem associated with a fixed scale (Verlegh et

al. 2002) or suffer from a lack of metricity (Steenkamp and Wittink 1994). The design of

experiment 1 is presented in Figure 1.

Stimuli

The product categories were refrigerators, LCD televisions, and digital cameras.

Each product offer was defined on two attributes for which participants were known to

have monotonic preferences. The attributes were the following: annual energy cost and









capacity (refrigerators), screen size and horizontal viewing angle (LCD televisions),

imaging sensor resolution and image storage capacity (digital cameras). Each attribute

had two or four possible attribute levels (see Table 2 for an overview of the attribute

levels). These attribute levels were either exterior (i.e., the two values in a 2-level

attribute or the two extreme values in a 4-level attribute) or interior (the two inserted

levels in a 4-level attribute). Exterior offers consisted of the exterior levels of the two

attributes (e.g., [1,1], [1,4], [2,1], [2,4] in a 2x4 design). Intermediate offers consisted of

the interior levels of the two attributes (e.g., [2,2], [2,3], [3,2], [3,3] in a 4x4 design), or

of an exterior level of one attribute and an interior level of a second attribute (e.g., [1,2],

[1,3], [2,2], [2,3] in a 2x4 design).

Procedure

Participants were seated at personal computers. Participants read instructions

stating that they would be stating reservation prices for products described on two

dimensions. Participants were asked to assume that the products were identical on all

non-defined attributes and were informed of the average price of the product. Then, the

product offers were presented. First, the participants stated their reservation price for the

intermediate offers. These four stimuli were presented in random order. Next, participants

stated their reservation price for the exterior offers. Again, the stimuli were presented in

random order. Thus, each participant provided reservation price judgments for eight

stimuli with the last four being the identical exterior stimuli.

Results

The analysis was conducted using the mean difference scores of the exterior offers.

For example, attribute one's differences score in a 2x4 design is the difference between

the average of [1,1] and [1,4] and the average of [2,1] and [2,4]. Similarly, attribute two's









differences score in a 2x4 design is the difference between the average of [1,1] and [2,1]

and the average of [1,4] and [2,4]. The four exterior offers were identical across

conditions, so these difference scores could be compared. The results of participants

whose ratings on the exterior offers were non-monotonic were removed from the dataset.

Table 3 presents the mean difference scores per attribute for each condition.

Refrigerators

Data of 91 participants were retained for analysis. The analysis for each attribute

was performed separately. The first analysis was on attribute one (annual energy cost).

The mean difference scores of the 2x4 condition (X = $78.97) and the 4x2 condition (X =

144.24) were significantly different (F(1, 88) = 10.38, p < .01). This replicates the

number-of-levels effect using a design that is common in the academic literature. The

mean difference scores of the 2x4 condition (X = $78.97) and the 4x4 condition (X =

$63.84) were not significantly different (F(1,88) = 0.63, p > .05). This result is

inconsistent with the hypothesis that the number-of-levels effect is sensitive only to the

number of levels of an attribute. The mean difference scores of the 4x2 condition (X =

$144.22) and the 4x4 condition (X = $63.84) were significantly different (F(1, 88) =

17.16, p < .01). This result is consistent with the attentional contrast hypothesis.

The second analysis was on attribute two (capacity). The mean difference scores of

the 2x4 condition (X = $122.55) and the 4x2 condition (X = $120.17) were not

significantly different (F(1, 88) = 0.01, p > .05). The mean difference scores of the 2x4

condition (X = $122.55) and the 4x4 condition (X = $100.93) were not significantly

different (F(1, 88) = 0.96, p > .05). The mean difference scores of the 4x2 condition (X =

$120.17) and the 4x4 condition (X = $100.93) were not significantly different (F(1, 88) =

0.73, p> .05)









LCD Televisions

Data of 90 participants were retained for analysis. The first analysis was on

attribute one (screen size). The mean difference scores of the 2x4 condition (X =

$176.20) and the 4x2 condition (X = $257.37) were significantly different (F(1, 87) =

3.98, p < .05). The mean difference scores of the 2x4 condition (X = $176.20) and the

4x4 condition (X = $196.82) were not significantly different (F(1, 87) = 0.26, p > .05).

The mean difference scores of the 4x2 condition (X = $257.37) and the 4x4 condition (X

= $196.82) were not significantly different (F(1, 87) = 2.21, p > .05). These results are

not supportive of the number of levels hypothesis or the attentional contrast hypothesis.

The second analysis was on attribute two (horizontal viewing angle). The mean

difference scores of the 2x4 condition (X = $140.64) and the 4x2 condition (X = $130.37)

were not significantly different (F(1, 87) = 0.26, p > .05). The mean difference scores of

the 2x4 condition (X = $140.64) and the 4x4 condition (X = 122.82) were not

significantly different (F(1, 87) = 0.78, p > .05). The mean difference scores of the 4x2

condition (X = $130.37) and the 4x4 condition (X = $122.82) were not significantly

different (F(1, 87) = 0.14, p > .05).

Digital Cameras

Data of 117 participants were retained for analysis. The first analysis was on

attribute one (imaging sensor resolution). The mean difference scores of the 2x4

condition (X = $134.90) and the 4x2 condition (X = $169.31) were significantly different

(F(1, 114) = 3.18, p = .08). The mean difference scores of the 2x4 condition (X =

$134.90) and the 4x4 condition (X = $131.49) were not significantly different (F(1, 114)

= 0.03, p > .05). The mean difference scores of the 4x2 condition (X = $169.31) and the









4x4 condition (X = $131.49) were significantly different (F(1, 114) = 4.07, p < .05).

These results are supportive of the attentional contrast hypothesis.

The second analysis was on attribute two (image storage capacity). The mean

difference scores of the 2x4 condition (X = $110.42) and the 4x2 condition (X = $106.81)

were not significantly different (F(1, 114)= 0.07, p > .05). The mean difference scores of

the 2x4 condition (X = $110.42) and the 4x4 condition (X = $95.86) were not

significantly different (F(1, 114) = 1.15, p > .05). The mean difference scores of the 4x2

condition (X = $106.81) and the 4x4 condition (X = $95.86) were not significantly

different (F(1, 114) = 0.63, p > .05).

Discussion

The data from experiment 1 replicate the number-of-levels effect that has been

reported in the literature and provide some new insights as to the potential source of this

effect. First, all three replicates showed a number-of-levels effect on attribute one when

the analysis was restricted to the 2x4 and 4x2 conditions. This type of design is common

in the number-of-levels literature. Second, when the 4x4 condition was also included in

the analysis, the mean difference score for attribute one was similar to the difference

score in the 2x4 condition, but less than the difference score in the 4x2 conditions.

Although this pattern of results held for only the refrigerator and digital camera

replicates, and only for attribute one, the result is consistent with the attentional contrast

hypothesis.

The most perplexing result in experiment 1 is the apparent attentional contrast

effect on attribute one and the lack of the attentional contrast effect on attribute 2. One

possible reason for this finding is that attentional contrast only operates on attributes that

are perceived to be important (i.e., worth attending to). In other words, an attribute must









be perceived to be important if its attribute levels are to exhibit an attentional contrast

based number-of-levels effect. The implication is that attribute two could also exhibit a

number-of-levels effect if it were to be perceived as more important. Experiment 1A will

examine this hypothesis.

Experiment 1A

The goal of experiment 1A was to assess whether the differential effects observed

on attributes one and two in experiment 1 could be a consequence of a (perceived)

difference in importance between the two attributes. In Experiment 1A, the (relative)

importance of the second attribute was increased in two ways. First, the range of the

attribute two's levels was increased and the range of attribute one's levels was decreased.

Second, participants were asked to elaborate on product usage. This second strategy was

adopted from Huffman (1997), who showed that elaboration on product usage increases

the weight of usage-related attributes whereas elaboration on product purchase increases

the weight of purchase-related attributes. Of the three replicates, only the refrigerator

category had a purchase-related attribute (annual energy cost) and a usage-related

attribute (capacity). Given the failure to find a number-of-levels effect on the usage-

related attribute in this category, participants were asked to elaborate on usage.

Method

The design was identical to experiment 1 with the limitation that only one of the

replicates (refrigerators) was investigated. The stimulus were adjusted so that the range of

attribute two (capacity) was doubled and the range of attribute one (annual energy cost)

was reduced. The attribute levels for capacity were 13, 17, 21, and 25 cubic feet. The

attribute levels for annual energy cost were $90, $85, $80, and $75.









The procedure was identical to experiment 1 except for one change. Prior to being

exposed to the eight product offers, participants were asked to elaborate on product

usage. Participants were told to "Think back to the times you have used a refrigerator.

Picture, in your mind's eye, the situations in which you have used it. Below, please

describe your thoughts and feelings about using the refrigerator. What was important to

you when you were using the refrigerator?"

Results

Table 4 presents the mean difference scores per attribute for each condition.

Data of 120 participants were retained for analysis. The first analysis was on

attribute one (annual energy cost). T the mean difference scores of the 2x4 condition (X

= $74.25) and the 4x2 condition (X = $98.37) were significantly different (F(1, 117) =

4.09, p < .05). The mean difference scores of the 2x4 condition (X = $74.25) and the 4x4

condition (X = $71.30) were not significantly different (F(1, 117) = 0.05, p > .05). The

mean difference scores of the 4x2 condition (X = $98.37) and the 4x4 condition (X =

$71.30) were significantly different (F(1, 117) = 5.70, p < .05). These results are

supportive of the attentional contrast hypothesis.

The second analysis was on attribute two (capacity). The mean difference scores of

the 2x4 condition (X = $265.13) and the 4x2 condition (X = $180.93) were significantly

different (F(1, 117) = 7.10, p < .05). The mean difference scores of the 4x2 condition (X

= $180.93) and the 4x4 condition (X = $220.72) were not significantly different (F(1,

117) = 1.75, p > .05). The mean difference scores of the 2x4 condition (X = $265.13) and

the 4x4 condition (X = $220.72) were not significantly different (F(1, 117) = 1.76, p >

.05). These results are supportive of the attentional contrast hypothesis.









Discussion

Experiment 1A shows that a number-of-levels effect can be produced on both

attributes in a two-factor conjoint design, provided the participant perceives each attribute

as important. When the importance of a second attribute was increased by increasing its

range and by getting consumers to elaborate on situations in which the attribute might be

important, a number-of-levels effect was obtained. In addition, the mean difference score

for attribute one in the 4x4 condition was similar to the difference score in the 2x4

condition, but less than the difference score in the 4x2 conditions. Similarly, the mean

difference score for attribute two in the 4x4 condition was similar to the difference score

in the 2x4 condition, but less than the difference score in the 2x4 conditions. These

results are consistent with the attentional contrast hypothesis.

Experiments 1 and 1A show that the "number-of-levels" effect does not occur

because of increasing levels on the focal attribute, but because of the decreasing levels on

the accompanying attribute. Moreover, the effect depends on the (perceived) importance

of an attribute. These results are consistent with an attentional contrast account, which

predicts that within subsequent offers, attentional contrast is produced on the attribute

that has more levels, as its levels are relatively more novel than the attribute levels of the

attribute with the lower number of levels.

However, experiments 1 and 1A were designed to manipulate the number of levels

and are, therefore, not the most direct test of the attentional contrast account. While the

conditions that were compared differed on attentional contrast on the focal attributes,

they also differed on the number of levels of the accompanying attribute. Experiment 2

provides a more direct test of the attentional contrast account by showing that it can






24


create a weighting difference while the number of levels on both focal and accompanying

attributes is kept constant.












Attribute 2
2 3


Attribute 1 2
3
4


Attribute 2


1
2
3
4


Attribute 1


Attribute 2


Attribute 1 [ 43



Note.-Exterior offers are colored in black, intermediate offers are shaded in grey, and
interior offers are hatched.

Figure 1. Design for experiments 1 and 1A. A) 4x2 Condition. B) 4x2 Condition. C) 4x4
Condition












Table 2. Attribute levels from experiment 1


Product category
Refrigerators


LCD televisions


Digital cameras


Attribute
annual energy cost ($n)
capacity (n cubic feet)


screen size (n inches)


Attribute level
1 2 3 4
100 90 80 70
16 18 20 22


20 24 28 32


horizontal viewing angle (n off-center) 45


imaging sensor resolution (n MegaPixels) 3.2


55 65 75


4.4 5.6 6.8


image storage capacity (up to n MB) 128 256 384 512



Table 3. Mean difference scores and simple effects for experiment 1


Attribute 1
Annual energy cost
(Refrigerators)
Screen size
(LCD televisions)
Imaging sensor resolution
(Digital cameras)


Attribute 2
Capacity
(Refrigerators)
Horizontal viewing angle
(LCD televisions)
Image storage capacity
(Digital cameras)


Mean difference score
4x2 2x4

144.24 78.97


F values
4x2 vs 2x44x2 vs 4x42x4 vs 4x4


63.84 10.38a 17.16a 0.63


257.37 176.20 196.82 3.98b


169.31


2.21 0.26


134.90 131.49 3.18c 4.07b 0.03


120.17 122.55 100.93


130.37

106.81


a implies significance at 99% level of confidence
b implies significance at 95% level of confidence
c implies significance at 90% level of confidence


4x2 vs 2x44x2 vs 4x42x4 vs 4x4

0.01 0.73 0.96


140.64 122.82 0.26 0.14 0.78


110.42 95.86 0.07 0.63







27




Table 4. Mean difference scores and simple effects for experiment 1A


Mean difference score


F values


Attribute 1
Annual energy cost
(Refrigerators)


4x2

98.37


2x4

74.25


4x4 4x2 vs 2x4 4x2 vs 4x4 2x4 vs 4x4


71.30


Attribute 2 4x2 2x4 4x4
Capacity
(Refrigerators) 180.93 265.13 220.72
a implies significance at 99% level of confidence
b implies significance at 95% level of confidence


0.05


4x2 vs 2x4 4x2 vs 4x4 2x4 vs 4x4


7.100














CHAPTER 4
EXPERIMENT 2: TEST OF ATTENTIONAL CONTRAST ACCOUNT WITH NO
DIFFERENCE IN NUMBER OF LEVELS

The goal of experiment 2 was to examine if attentional contrast can create a

weighting effect when there is no difference in number of levels. In experiments 1 and

1A, we found evidence for an attentional contrast account by varying the number of

levels of only one attribute at a time. However, attentional contrast was confounded with

the number of levels manipulation. By manipulating attentional contrast in a context in

which there is no difference in number of levels, we want to obtain more direct evidence

that the attentional contrast by itself can account for the weighting difference. In addition,

the three replicates and procedure from experiment 1 were used in the experiment. It was

felt that demonstrating a robust attentional contrast effect on one attribute was superior to

demonstrating an attentional contrast effect on two attributes of a single replicate.

Method

Design and Stimuli

The design was a 3 cell between-subjects design (contrast on attribute 1, contrast

on attribute 2, balanced contrast) with three between-subject replicates (refrigerator, LCD

televisions, digital cameras). In all three experimental conditions, each attribute was

defined on 4 levels (see Figure 2). The cells differed on the attribute on which attentional

contrast was created. In the contrast on attribute 1 condition, attentional contrast was

created on attribute 1. This contrast was created via the selection of offers that were

initially rated by the participant. The intermediate offers were [2,1], [2,2], [2,3], [2,4],









[3,1], [3,2], [3,3], and [3,4]. Thus, when the participant saw the exterior offers of [1,1],

[1,4], [4,1], and [4,4], the values of the first attribute were novel and the values of the

second attribute were familiar. In the contrast on attribute 2 condition, the intermediate

offers were [1,2], [1,3], [2,2], [2,3], [3,2], [3,3], [4,2],and[4,3]. Thus, when the participant

saw the exterior offers of [1,1], [1,4], [4,1], and [4,4], the values of the second attribute

were novel and the values of the first attribute were familiar. In the balanced contrast

condition, the intermediate offers were [2,2], [2,3], [3,2], and [3,3]. Thus, when the

participant saw the exterior offers of [1,1], [1,4], [4,1], and [4,4], the values of both

attributes were on average equally novel, hence there was no contrast. Note that this last

condition is particularly important because it will show that the number-of-levels effect is

not an issue of novelty (i.e., seeing levels that have not been seen before), but an issue of

attentional contrast (i.e., seeing levels that are novel relative to the levels seen on the

other attribute). In all other respects, the procedure was identical to the one used in

experiment 1.

Predictions

The attentional contrast hypothesis makes predictions for attributes one and two.

With respect to attribute one, the attentional contrast hypothesis predicts that the mean

difference score should be greater in the contrast on attribute 1 condition than in the

contrast on attribute 2 condition or the balanced contrast condition. With respect to

attribute two, the attentional contrast hypothesis also predicts that the mean difference

score should be greater in the contrast on attribute 2 condition than in the contrast on

attribute 1 condition or the balanced contrast condition. Given the results of experiments

1 and 1A, it was anticipated that the attentional contrast effect would be obtained on

attribute 1, but not attribute 2.











Results

Table 5 presents the mean difference scores per attribute in each condition.

Refrigerators

Data of 104 participants were retained for analysis. The analysis for each attribute

was performed separately. The first analysis was on attribute one (annual energy cost).

The mean difference scores of the contrast on attribute 1 condition (X = $99.37) and the

contrast on attribute 2 condition (X = $68.60) were significantly different (F(1, 101) =

4.96, p < .05). The mean difference scores of the contrast on attribute 1 condition (X =

$99.37) and the balanced contrast condition (X = $70.04) were significantly different

(F(1, 101) = 4.87, p < .05). The second analysis was on attribute two (capacity). The

mean difference scores of the contrast on attribute 2 condition (X = $94.02) and the

contrast on attribute 1 condition (X = $126.58) were not significantly different (F(1, 101)

= 2.65, p > .05). The mean difference scores of the contrast on attribute 2 condition (X =

$94.02) and the balanced contrast condition (X = $99.79) were not significantly different

(F(1, 101) = 0.08, p > .05).

LCD televisions

Data of 111 participants were retained for analysis. The first analysis was on

attribute one (screen size). The mean difference scores of the contrast on attribute 1

condition (X = $286.04) and the contrast on attribute 2 condition (X = $228.72) were not

significantly different (F(1, 108) = 2.13, p > .05). The mean difference scores of the

contrast on attribute 1 condition (X = $286.04) and the balanced contrast condition (X =

$192.99) were significantly different (F(1, 108) = 5.27, p < .05). The second analysis was

on attribute two (horizontal viewing angle). The mean difference scores of the contrast on









attribute 2 condition (X = $133.74) and the contrast on attribute 1 condition (X =

$146.52) were not significantly different (F(1, 108) = 0.26, p > .05). The mean difference

scores of the contrast on attribute 2 condition (X = $133.74) and the balanced contrast

condition (X = $101.04) were not significantly different (F(1, 108) = 1.75, p > .05).

Digital cameras

Data of 133 participants were retained for analysis. The first analysis was on

attribute one (imaging sensor resolution). The mean difference scores of the contrast on

attribute 1 condition (X = $190.12) and the contrast on attribute 2 condition (X =

$143.15) were significantly different (F(1, 130) = 5.78, p < .05). The mean difference

scores of the contrast on attribute 1 condition (X = $190.12) and the balanced contrast

condition (X = $144.37) were significantly different (F(1, 130) = 6.14, p < .05). The

second analysis was on attribute two (image storage capacity). The mean difference

scores of the contrast on attribute 2 condition (X = $128.43) and the contrast on attribute

1 condition (X =$114.12) were not significantly different (F(1, 130) = 0.89, p > .05). The

mean difference scores of the contrast on attribute 2 condition (X = $128.43) and the

balanced contrast condition (X = $119.46) were not significantly different (F(1, 130) =

0.34, p> .05).

Discussion

Experiment 2 manipulated attentional contrast directly by varying the relative

novelty of the attribute levels while keeping the number of levels of the attributes

constant. The results showed that attentional contrast can create a weighting difference

when there is no difference in number of levels. While this effect was consistently

produced on attribute 1, the results on the second attribute were non-significant. Since the

stimuli used in this experiment were identical to the ones used in experiment 1, we






32


attribute this lack of significance to the fact that attribute 2 was perceived to be less

important (see experiment 1A).










Attribute 2



Attribute 1

j4


Attribute 2
1 2 3 4


Attribute 1


Attribute 2


Attribute 1__I431


Note.-Exterior offers are colored in black, intermediate offers are shaded in grey. The
hatched offers are presented twice.

Figure 2. Design for experiment 2. A) Contrast on Attribute 1 Condition. B) Contrast on
Attribute 2 Condition. C) Balanced Contrast Condition.













Table 5. Mean difference scores and simple effects for experiment 2


Mean difference score


Attribute 1
Annual energy cost
(Refrigerators)
Screen size
(LCD televisions)
Imaging sensor resolution
(Digital cameras)


contrast
on
attribute 1


contrast
on balanced
attribute 2 contrast


F values
contrast on
attribute 1 vs
contrast on
attribute 2


99.37 68.60 70.04 4.96b

286.04 228.72 192.99 2.13

190.12 143.15 144.37 5.78b


contrast
on


contrast
on balanced


Attribute 2 attribute 1 attribute 2 contrast
Capacity
(Refrigerators) 126.58 94.02 99.79
Horizontal viewing angle
(LCD televisions) 146.52 133.74 101.04
Image storage capacity
(Digital cameras) 114.12 128.43 119.46
a implies significance at 99% level of confidence
b implies significance at 95% level of confidence


contrast on
attribute 1 vs
contrast on
attribute 2


0.26


contrast on
attribute 1 vs
balanced
contrast


6.14a


contrast on
attribute 2 vs
balanced
contrast


4.87b














CHAPTER 5
EXPERIMENT 3: TEST OF ATTENTIONAL CONTRAST ACCOUNT WITH
REVERSED NUMBER OF LEVELS

The goal of experiment 3 was to show that the "number-of-levels" effect can be

reversed when attentional contrast is created on the attribute with the lower number of

levels. In experiments 1 and 1A, attentional contrast was shown to create weighting

differences but the number of levels of the accompanying attribute was varied, but not the

number of levels of the focal attribute. In experiment 2, attentional contrast created

weighting differences when there was no difference in number of attribute levels. In

experiment 3, the test of the attentional contrast account is made more rigorous. More

explicitly, I test whether the number-of-levels effect can be obtained on the attribute with

the lower number of levels.

Method

Design and Stimuli

The design was a 3 cell between-subjects design (4x6, 6x4, and 6x6 attribute

levels) with three between-subject product replicates (refrigerators, LCD televisions,

digital cameras). Each attribute had four or six possible attribute levels. For an overview

of the attribute levels, see Table 6. The intermediate offers were [3,1], [3,2], [3,3], [3,4],

[3,5], [3,6], [4,1], [4,2], [4,3], [4,4], [4,5], and [4,6] in the 4x6 condition, [3,1], [4,1],

[3,2], [4,2], [3,3], [4,3], [3,4], [4,4], [3,5], [4,5], [3,6], and [4,6] in the 6x4 condition, and

[1,3], [1,4], [2,2], [2,5], [3,1], [3,6], [4,1], [4,6], [5,2], [5,5], [6,3], and [6,4] in the 6x6

condition. The four exterior offers were [1,1], [1,6], [6,1], and [6,6]. The design of









experiment 3 is presented in Figure 3. In the 4x6 condition, the levels of attribute 1 were

more novel than the attribute levels of attribute 2 at the time the participants evaluated the

exterior offers. In other words, attentional contrast was created on attribute 1. In the 6x4

condition, the levels of attribute 2 were more novel than the attribute levels of attribute 1

at the time the participants evaluated the exterior offers. In other words, attentional

contrast was created on attribute 2. In the 6x6 condition, the attribute levels of attribute 1

and attribute 2 were on average equally novel when participants evaluated the exterior

offers. In sum, attentional contrast was created on the four-level attributes. If the number-

of-levels effect depends only on the number of levels, the four-level attributes will have

lower derived importance weights than the six-level attributes. If the number-of-levels

effect depends on attentional contrast, then the four-level attributes will have higher

derived importance weights.

Predictions

The attentional contrast hypothesis makes predictions for attributes one and two.

With respect to attribute one, the attentional contrast hypothesis predicts that the mean

difference score should be greater in the 4x6 condition than in the 6x4 condition or the

6x6 condition. With respect to attribute 2, the attentional contrast hypothesis predicts that

the mean difference score should be greater in the 6x4 condition than in the 4x6 condition

or the 6x6 condition. Given the results of experiments 1, 1A, and 2, it was anticipated that

the attentional contrast effect would be obtained on attribute 1, but not on attribute 2.

Results

Table 7 presents the mean difference scores per attribute for each condition.









Refrigerators

Data of 87 participants were retained for analysis. The analysis for each attribute

was performed separately. The first analysis was on attribute one (annual energy cost).

The mean difference scores of the 4x6 condition (X = $119.89) and the 6x4 condition (X

= $77.59) were significantly different (F(1, 84) = 7.52, p < .05). The mean difference

scores of the 4x6 condition (X = $119.89) and the 6x6 condition (X = $76.75) were

significantly different (F(1, 88) = 6.72, p < .05). The second analysis was on attribute two

(capacity). The mean difference scores of the 6x4 condition (X = $110.09) and the 4x6

condition (X = $132.15) were not significantly different (F(1, 84) = 1.34, p > .05). The

mean difference scores of the 6x4 condition (X = $110.09) and the 6x6 condition (X =

$94.92) were not significantly different (F(1, 84) = 0.55, p > .05).

LCD televisions

Data of 96 participants were retained for analysis. The analysis for each attribute

was performed separately. The first analysis was on attribute one (screen size). The mean

difference scores of the 4x6 condition (X = $246.95) and the 6x4 condition (X = $201.55)

were not significantly different (F(1, 93) = 1.07, p > .05). The mean difference scores of

the 4x6 condition (X = $246.95) and the 6x6 condition (X = $228.70) were not

significantly different (F(1, 93) = 0.15, p > .05). The second analysis was on attribute two

(horizontal viewing angle). The mean difference scores of the 6x4 condition (X =

$138.15) and the 4x6 condition (X = $129.30) were not significantly different (F(1, 93) =

0.16, p > .05). The mean difference scores of the 6x4 condition (X = $138.15) and the

6x6 condition (X = $133.70) were not significantly different (F(1, 93) = 0.05, p > .05).









Digital cameras

Data of 123 participants were retained for analysis. The analysis for each attribute

was performed separately. The first analysis was on attribute one (imaging sensor

resolution). The mean difference scores of the 4x6 condition (X = $175.06) and the 6x4

condition (X = $169.20) were not significantly different (F(1, 120) = 0.09, p > .05). The

mean difference scores of the 4x6 condition (X = $175.06) and the 6x6 condition (X =

$153.29) were significantly different (F(1, 120)= 1.17, p > .05). The second analysis was

on attribute two (image storage capacity). The mean difference scores of the 6x4

condition (X = $157.11) and the 4x6 condition (X = $126.92) were not significantly

different (F(1, 120) = 2.50, p > .05). The mean difference scores of the 6x4 condition (X

= $157.11) and the 6x6 condition (X = $111.00) were significantly different (F(1, 120) =

5.54, p <.05).

Discussion

The results of experiment 3 point in the direction of an attentional contrast

explanation. On the first attribute, the means are in the direction predicted by the

attentional contrast account, though the results are significant in only one of three product

categories (refrigerators). On attribute 2, the results of the refrigerators category differs

from the predictions of both number-of-levels effect and attentional contrast. The means

of the LCD televisions category are in the direction predicted by attentional contrast,

though non-significant. The results for the digital cameras category are in the direction

predicted by attentional contrast. Since this experiment used the same exterior levels as

experiments 1 and 2, the partial lack of significance on attribute 2 may be attributed to

the perceived lack of importance of attribute 2 (see experiment 1A).





















Attribute 1


A

Attribute 2


1

2
3

4

5


B
Attribute 2

22 3 4 5


2

Attribute 1
4
5
6


C
Attribute 2




2

Attribute 1



6

Note.-Exterior offers are colored in black, intermediate offers are shaded in grey.

Figure 3. Design for experiment 3. A) 4x6 Condition. B) 6x4 Condition. C) 6x6
Condition.












Table 6. Attribute levels from experiment 3
Attribute level
Product
category Attribute 1 2 3 4 5 6
Refrigerators annual energy cost ($n) 100 95 90 80 75 70
capacity (n cubic feet) 16 17 18 20 21 22

LCD
televisions screen size (n inches) 20 22 24 28 30 32
horizontal viewing angle (n off-center) 45 50 55 65 70 75


Digital imaging sensor resolution (n
cameras MegaPixels) 3.2 3.8 4.4 5.6 6.2 6.8
image storage capacity (up to n MB) 128 192 256 384 448 512
Note.-The underlined attribute levels are the levels that are added when an attribute has
6 levels.


Table 7. Mean difference scores and simple effects for experiment 3


Mean difference score


F values


Attribute 1
Annual energy cost
(Refrigerators)
Screen size
(LCD televisions)
Imaging sensor resolution
(Digital cameras)


4x6 6x4 6x6 4x6 vs 6x4 4x6 vs 6x6

119.89 77.59 76.75 7.52a 6.72a

246.95 201.55 228.70 1.07 0.15

175.06 169.20 153.29 0.09 1.17


Attribute 2 4x6 6x4
Capacity 132.15 110.09
(Refrigerators)
Horizontal viewing angle 129.30 138.15
(LCD televisions)
Image storage capacity 126.92 157.11
(Digital cameras)
a implies significance at 99% level of confidence
b implies significance at 95% level of confidence


6x6

94.92


4x6 vs 6x4 6x4 vs 6x6

1.34 0.55


133.70 0.16

111.00 2.50


0.05

5.54b














CHAPTER 6
GENERAL DISCUSSION

Experiment 1 provides evidence that the weighting effect as observed in published

studies may not be a "number-of-levels" effect, but that it may be due to attentional

contrast away from the less novel attribute level within each trial. Experiment 2 shows

that attentional contrast can create a weighting effect when there is no difference in the

number of levels. Experiment 3 shows that attentional contrast can overrule the number-

of-levels effect, i.e., that it can result in greater derived importance weights on the

attribute with the lower number of levels. In the three experiments, results were

consistent on attribute 1, but did not reach significance on attribute 2. In experiment 1A, I

showed that this lack of significance was due to the perceived relative unimportance of

the second attribute.

In experiment 1A, the distinction between attribute level importance and attribute

importance is highlighted. Simonson (1993) shows that consumers often do not possess

well-formed beliefs about attribute importance. Instead, they construct their preferences

during judgment and choice (Payne, Bettman, and Johnson 1992). First, the experiments

in this dissertation show that attribute importance is influenced by the attribute levels that

are presented. Second, experiment 1A shows that this influence only occurs when the

attribute is perceived to be important. While in experiments 1, 2 and 3 attentional contrast

only created a weighting effect on attribute 1, increasing the accessibility of consumers'

product experiences in experiment 1A resulted in a weighting effect on attribute 2. In the

number-of-levels effect literature, price was often used as one of the attributes. Since this









dissertation uses price as a dependent variable, one of the product attributes that are

usually perceived as being inherently important was not eligible, which may have

reduced the overall perceived importance of the attributes used in this dissertation

compared to attributes used in earlier studies. This effect may be enhanced by the fact

that the choice of attributes was restricted to attributes that have numerical values and for

which evaluations are known to be monotonically distributed.

The attentional contrast explanation is consistent with all the results that have been

published in the number-of-levels effect literature. First, all empirical studies (except

Johnson 1992) used a sequential evaluation procedure. Attentional contrast can only

operate when attribute levels are repeated and when this repetition occurs in the context

of other attribute levels that have been more or less repeated. Second, attentional contrast

is consistent with the observation that the number-of-levels effect was not produced in

the compositional measures that were added before the sequential evaluation procedure

(Wittink et al. 1992a) and that only minimal effects were produced when the

compositional measures were added after the sequential evaluation procedure

(Steenkamp and Wittink 1994; Verlegh et al. 2002). Finally, the reason why attentional

contrast was not detected in earlier studies is that they invariably compared conditions in

which the number of levels of more than one attribute was compared at once (see Table

1).

While the main contribution of this dissertation is the proposition of a new account

for a repeatedly published effect, two additional contributions can be mentioned. First,

until now, the process of attentional contrast had only been documented in a category

learning task. Based on the shared characteristic of sequential trials between the









classification learning and conjoint or multiattribute tasks, I showed that attentional

contrast also occurs in an evaluation (versus categorization) task. Second, in previous

studies (Kersten et al. 1998), attentional contrast was documented using pictorial values. I

generalized its occurrence to the use of metric values.

While I conclude that attentional contrast is at least partly responsible for the

weighting effect that was observed in studies examining the number-of-levels effect, the

data do not provide information as to how attentional contrast creates a weighting

difference. One possibility is that participants attend longer to the relatively novel

attribute levels, and that attribute levels that are attended to for longer durations receive

more weight. However, it is possible that it is not mere attention which results in the

weighting difference. It is likely that participants elaborate more on the relatively novel

attribute levels. For example, when asked to evaluate an LCD television with a horizontal

viewing angle which is 60 degrees off-center, participants may imagine themselves in

their apartments and try to assess mentally from which area they would have a decent

view of the screen.

Aside from the question how attentional contrast works, one may ask when it

works. First, the impact of the attentional contrast effect is likely to be dependent on the

assumptions with regards to the total pool of attention. In the experiments presented in

this dissertation, students evaluated the trials at their own pace. In other words, the pool

of attention was variable. This means that increased attention to one attribute is not

necessarily accompanied by decreases in attention to a second attribute. In contrast, in a

model in which the total pool of attention is fixed, increased attention to a first attribute

results in decreases in attention to the second attribute. A model with a fixed pool of









attention is more closely operationalized by a computer-paced experiment. Using a

computer-paced experiment may enhance the chances that attentional contrast creates a

weighting effect on both attributes. Second, as shown in experiment 1A, the occurrence

of the weighting effect depends on the importance of the attribute. The prior level of

attention to an attribute may influence the impact of changes in attention towards the

attribute levels. Attributes that normally receive a low level of attention may show small

reductions in attention as a result of attentional contrast away from the attribute levels.

Analogously, attributes that tend to receive high levels of attention may show small

increases in attention as a result of attentional contrast towards the attribute levels.

Therefore, attentional contrast is likely to have the greatest impact when attributes are

moderately important.

Finally, a main focus of many number-of-level effect studies was to find a way to

eliminate the bias. Based on the account offered in this dissertation, one should try to

keep the relative novelty of attribute levels in subsequent trials constant. Since this is

impossible when attributes are defined on unequal numbers of levels, one may prefer to

use a data collection method that does not involve subsequent trials. For example, the use

of a trade-off matrix may be preferred in case of an unequal number of levels.














CHAPTER 7
LIMITATIONS AND FUTURE RESEARCH

While this dissertation takes an important first step towards proposing a new

account for the number-of-levels effect, there are several issues that remain unexamined

and that could be extended to potentially useful future research.

There are several characteristics of the experiments that limit the generalizability of

the findings. First, only one dependent measure was used in all studies. Participants stated

their reservation price for product offers. While the results of these studies supported an

attentional contrast explanation, the number-of-levels effect has been identified using a

variety of dependent measures. To generalize the results of this dissertation, it would be

useful to examine whether attentional contrast can also explain the results of studies that

use other dependent measures such as paired comparisons and rating scales. Second, no

compositional measures were collected in this dissertation. While there is considerable

evidence from the literature that the number-of-levels effect is not produced on

compositional measures, it could be useful to gather additional evidence that shows that

the number-of-levels effect is inherent to the sequential evaluation procedure. Moreover,

it could be useful to examine if the weighting effect caused by attentional contrast carries

over to a posteriori collected compositional measures.

A final extension of this dissertation has the potential of being useful. This research

shows that attentional contrast does occur not only during category learning tasks, but

also during multiattribute judgment tasks. The role of attentional contrast during other






46


phenomena that have been identified in the multiattribute judgment literature (e.g.,

frequency effect) would be worth examining.















LIST OF REFERENCES


Currim, Imran S., Charles B. Weinberg, and Dick R. Wittink (1981), "The Design of
Subscription Programs for a Performing Arts Series," Journal of Consumer
Research, 8 (June), 67-75.

Goldstone, Robert L. and Mark Steyvers (2001), "The Sensitization and Differentiation
of Dimensions During Category Learning," Journal ofExperimental Psychology:
General, 130 (March), 116-139.

Green, Paul E. and V. Srinivasan (1978), "Conjoint Analysis in Consumer Behavior:
Issues and Outlook," Journal of Consumer Research, 5 (September), 103-123.

Green, Paul E. and V. Srinivasan (1990), "Conjoint Analysis in Marketing Research:
New Developments and Directions," Journal ofMarketing, 54 (October), 3-19.

Huffman, Cynthia (1997), "Elaboration on Experience: Effects on Attribute Importance,"
Psychology & Marketing, 14 (August), 451-474.

Johnson, Richard M. (1992), "Comment on Attribute Level Effects," SecondAnnual
Advanced Research Techniques Forum Proceedings, Rene Mora (ed.), Chicago:
American Marketing Association, 62-64.

Kersten, Alan W., Robert L. Goldstone, and Alexandra Schaffert (1998), "Two
Competing Attentional Mechanisms in Category Learning," Journal of
Experimental Psychology: Leiaruiu Memory, and Cognition, 24 (6), 1437-1458.

Nosofsky, Robert M. (1986), "Attention, Similarity, and the Identification-Categorization
Relationship," Journal ofExperimental Psychology: General, 115 (March), 39-57.

Nosofsky, Robert M. (1991), "Tests of an Exemplar Model for Relating Perceptual
Classification and Recognition Memory," Journal ofExperimental Psychology:
Human Perception and Performance, 17 (February), 3-27.

Orme, Bryan K. (1998), "Reducing the Number-of-Attribute-Levels Effect in ACA with
Optimal Weighting," 1998 .Sen tiI ,th Software Conference Proceedings, Ketchum,
ID: Sawtooth Software.

Payne, John W., James R. Bettman, and Eric J. Johnson (1992), "Behavioral Decision
Research: A Constructive Processing Perspective," Annual Review ofPsychology,
43, 87-131.









Simonson, Itamar (1993), "Get Closer to Your Customers by Understanding How They
Make Choices," California Management Review, 35 (4), 68-84.

Srinivasan, V. and Allan D. Shocker (1973), "Estimating the Weights for Multiple
Attributes in a Composite Criterion Using Pairwise Judgments," Psychometrika, 38
(December), 473-493.

Steenkamp, Jan-Benedict E.M. and Dick R. Wittink (1994), "The Metric Quality of Full-
Profile Judgments and the Number-of-Attribute-Levels Effect in Conjoint
Analysis," International Journal ofResearch in Marketing, 11 (June), 275-286.

Sutherland, N.S. and N.J. Mackintosh (1971). Mechanisms of Animal Discrimination
Learning. New York: Academic Press.

Verlegh, Peeter W.J., Hendrik N.J. Schifferstein, and Dick R. Wittink (2002), "Range
and Number-of-Levels Effects in Derived and Stated Measures of Attribute
Importance," Marketing Letters, 13 (February), 41-52.

Wittink, Dick R., Joel Huber, John A. Fiedler, and Richard L. Miller (1992a), "Attribute
Level Effects in Conjoint Revisited: ACA versus Full Profile," Second Annual
Advanced Research Techniques Forum Proceedings, Rene Mora (ed.), Chicago:
American Marketing Association, 51-61.

Wittink, Dick, Joel Huber, Peter Zandan, and Richard M. Johnson (1992b), "The Number
of Levels Effect in Conjoint: Where Does It Come From and Can It Be
Eliminated?" 1992 .i eni ,,th Software Conference Proceedings, Ketchum, ID:
Sawtooth Software.

Wittink, Dick R., Lakshman Krishnamurthi, and Julia B. Nutter (1982), "Comparing
Derived Importance Weights Across Attributes," Journal of Consumer Research, 8
(March), 471-474.

Wittink, Dick R., Lakshman Krishnamurthi, and David J. Reibstein (1989), "The Effects
of Differences in the Number of Attribute Levels on Conjoint Results," Marketing
Letters, 1, 113-123.















BIOGRAPHICAL SKETCH

Els De Wilde was born in Gent, Belgium, on July 18, 1975. After attending high

school at the Onze-Lieve-Vrouw-Presentatie in Sint-Niklaas, Belgium, she attended the

Katholieke Universiteit Leuven in Louvain, Belgium. At this university, she obtained the

degrees of Kandidaat in de Psychologie (1996), graduating "with distinction," and

Licentiaat in de Psychologie (1999), for which she graduated "with great distinction." As

part of her studies for the Licentiaat degree, she spent five months at the Universidad

Complutense de Madrid, Spain. She wrote a thesis on categorization under the guidance

of Professor Gert Storms. She spent seven months on a research internship, split between

the Laboratory of Experimental Psychology at the Catholic University of Louvain, where

she worked with Karl Verfaillie, and the marketing research company TNS Dimarso in

Brussels, Belgium, where she was a trainee at the Fast Moving Consumer Goods unit.

These research experiences were instrumental in shaping her interest in pursuing doctoral

studies in the area of marketing, which she started at the University of Florida in August

1999. During her four years in the Ph.D. program at the Department of Marketing, she

collaborated with Professors Chris Janiszewski and Alan Cooke. In 2001, Els taught an

undergraduate course in consumer behavior at the University of Florida. Since Summer

2003, Els has been working as an Assistant Professor of Marketing at HEC Montreal,

Canada.




Full Text

PAGE 1

ATTENTIONAL CONT RAST DURING SEQUENTIAL JUDGMENTS: AN EXPLANATION FOR THE NU MBER-OF-LEVELS EFFECT By ELS DE WILDE A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLOR IDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2005

PAGE 2

Copyright 2005 by Els De Wilde

PAGE 3

This dissertation is dedicated to my parents.

PAGE 4

ACKNOWLEDGMENTS This dissertation leaves me indebted to many people. First, I would like to thank my advisor, Chris Janiszewski, for taking up the mentorship role and guiding me through the dissertation process. Interactions with other members of my committee were also invaluable in shaping this dissertation. In particular, I would like to thank Alan Cooke for his conceptual guidance and Richard Lutz and Asoo Vakharia for their feedback. My thanks also go to my friends in Belgium and in the doctoral program and my colleagues at HEC Montral for their support. I would like to thank JoAndrea Hoegg and Andrew Kuo for their assistance with the data collection while I was at long distance. I am extremely thankful to my family members for their encouragement throughout the entire duration of my doctoral studies. Finally, I would like to thank Roberto for giving me the best reason to finish this dissertation. iv

PAGE 5

TABLE OF CONTENTS page ACKNOWLEDGMENTS .................................................................................................iv LIST OF TABLES ............................................................................................................vii LIST OF FIGURES .........................................................................................................viii ABSTRACT ......ix CHAPTER 1 INTRODUCTION........................................................................................................1 2 THEORETICAL BACKGROUND..............................................................................4 The Number-Of-Levels Effect......................................................................................4 Existing Accounts of the Number-Of-Levels Effect....................................................4 Attentional Contrast......................................................................................................9 3 EXPERIMENT 1: DEMONSTRATION OF THE NUMBER-OF-LEVELS EFFECT WITH EXPANDED DESIGN....................................................................16 Method........................................................................................................................16 Results.........................................................................................................................17 Discussion...................................................................................................................20 Experiment 1A............................................................................................................21 4 EXPERIMENT 2: TEST OF ATTENTIONAL CONTRAST ACCOUNT WITH NO DIFFERENCE IN NUMBER OF LEVELS........................................................28 Method........................................................................................................................28 Results.........................................................................................................................30 Discussion...................................................................................................................31 5 EXPERIMENT 3: TEST OF ATTENTIONAL CONTRAST ACCOUNT WITH REVERSED NUMBER OF LEVELS........................................................................35 Method........................................................................................................................35 Results.........................................................................................................................36 Discussion...................................................................................................................38 v

PAGE 6

6 GENERAL DISCUSSION.........................................................................................41 7 LIMITATIONS AND FUTURE RESEARCH..........................................................45 LIST OF REFERENCES...................................................................................................47 BIOGRAPHICAL SKETCH.............................................................................................49 vi

PAGE 7

LIST OF TABLES Table page 1. Number-of-levels manipulations used in published studies..........................................15 2. Attribute levels from experiment 1................................................................................26 3. Mean difference scores and simple effects for experiment 1........................................26 4. Mean difference scores and simple effects for experiment 1A......................................27 5. Mean difference scores and simple effects for experiment 2........................................34 6. Attribute levels from experiment 3................................................................................40 7. Mean difference scores and simple effects for experiment 3........................................40 vii

PAGE 8

LIST OF FIGURES Figure page 1. Design for experiments 1 and 1A..................................................................................25 2. Design for experiment 2.................................................................................................33 3. Design for experiment 3.................................................................................................39 viii

PAGE 9

Abstract of Dissertation Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy ATTENTIONAL CONTRAST DURING SEQUENTIAL JUDGMENTS: AN EXPLANATION FOR THE NUMBER-OF-LEVELS EFFECT By Els De Wilde August 2005 Chair: Chris Janiszewski Cochair: Alan D. J. Cooke Major Department: Marketing Normatively, the importance of an attribute does not depend on the number of levels on which it is defined, exterior levels being fixed. This dissertation examines a bias identified in the conjoint literature known as the number-of-levels effect, which occurs when consumers derived importance weights increase with the number of attribute levels. Borrowing from research on attention during category learning, this dissertation examines psychological mechanisms that underlie the number-of-levels effect. In contrast to earlier attentional accounts that have been offered for the bias, I propose that the weighting effect found in the number-of-levels studies is due to the attentional processes that are inherent to the sequential evaluation procedure. Analogously to attentional processes during category learning, consumers divide their attention between attribute dimensions and attribute levels when making sequential evaluations. Within a given trial, a mechanism of attentional contrast directs attention ix

PAGE 10

away from attribute levels that have been associated with an outcome. I suggest that the number-of-levels effect is caused by this attentional contrast strategy. When attributes differ on the number of levels on which they are defined, they also differ on the number of times a given attribute level has been associated with a judgment (i.e., attribute level novelty). A process of attentional contrast away from the more familiar attribute level within each offer results in differential importance weights for the attributes. In a series of experiments, I examine this attentional contrast account for the number-of-levels effect. Consistent with this account, the weighting effect is shown to appear when attentional contrast is created in a situation in which there is no difference in number of levels. Moreover, the number-of-levels effect reverses when attentional contrast is created on the attribute with the lower number of levels. x

PAGE 11

CHAPTER 1 INTRODUCTION Normatively, the importance of an attribute should not depend on the number of levels on which it is defined, exterior levels being fixed. Yet, several studies from the conjoint literature show that derived importance weights increase with the number of attribute levels, a bias which has generally been labelled the number-of-levels effect (e.g., Currim, Weinberg and Wittink 1981; Verlegh, Schifferstein and Wittink 2002; Wittink et al. 1992b). In searching for a method to eliminate this bias, several accounts have been offered for it. Methodologically, there is evidence that the number-of-levels effect is related to the data collection method (e.g., Wittink et al. 1992b), the measurement scale for the dependent variable (e.g., Steenkamp and Wittink 1994; Wittink et al. 1992b), and the parameter estimation procedure (Wittink, Krishnamurthi and Reibstein 1989). Although tests of these methodological accounts reduced the number-of-levels effect, they have never completely eliminated the bias. The possibility that psychological processes are at least partly responsible for the number-of-levels effect has been acknowledged since the effect was first documented. Currim et al. (1981) state that the respondent may react to the number of levels used for a given attribute by consciously or unconsciously weighting an attribute more heavily as the number of levels increases (p. 72). Although a test of this account has been included in several studies, it has not received much support during the past two decades. An assumption underlying tests of this attentional account is that there should be a number-of-levels effect not only in sequential judgments or choices (e.g., full-profile rank orders, 1

PAGE 12

2 full-profile paired comparisons, full-profile ratings), but also in self-explicated priors (Wittink et al. 1992a) and stated importance measures (Steenkamp and Wittink 1994; Verlegh et al. 2002). However, I propose that the weighting effect found in the number-of-levels studies may be due to the attentional processes that are inherent to the sequential evaluation procedure. Studies in which people learn new categories by making sequential classifications distinguish between attention allocated to attributes and to attribute levels (Kersten, Goldstone and Schaffert 1998). They show that within a given trial, a mechanism of attentional contrast directs attention away from attribute levels that have been associated with an outcome. I propose that this attentional contrast strategy can at least partly account for the number-of-levels effect. When attributes differ on the number of levels on which they are defined, they also differ on the number of times a given attribute level appears across offers during a sequential judgment task (henceforth referred to as relative attribute level novelty). More specifically, the levels of attributes that are defined on more levels are overall relatively more novel than the levels of attributes that are defined on fewer levels. A process of attentional contrast away from the less novel attribute level within each offer may result in differential importance weights for the attributes. This dissertation proceeds as follows: Chapter 2 discusses the number-of-levels effect and the methodological and psychological accounts that have been offered for it in the literature. It also introduces the attentional contrast account from the category learning literature. In chapters 3 to 5, I examine if attentional contrast can account for the number-of-levels effect. In chapter 3, I extend the designs that have typically been used

PAGE 13

3 in studies that examine the number-of-levels effect and find evidence for an account in terms of attentional contrast. In chapter 4, I manipulate attentional contrast directly to examine its possible impact when there is no difference in number of levels. In chapter 5, I create attentional contrast on the attribute with the lower number of levels. Finally, chapter 6 discusses the findings and chapter 7 examines the limitations and avenues for future research.

PAGE 14

CHAPTER 2 THEORETICAL BACKGROUND The Number-Of-Levels Effect When an attribute has a higher number of levels (exterior values being fixed), derived importance weights for that attribute increase. For example, Currim et al. (1981) examined subscription series to performing arts events, using three attributes that were defined on two levels and three attributes that were defined on three levels. They asked participants to make trade-offs between pairs that were defined on two attributes: a two-level attribute and a three-level attribute. The two-level attributes had derived importance weights between 0.36 and 0.45, whereas the three-level attributes had derived importance weights between 0.55 and 0.66. Originally identified between attributes using part-worth estimation within conjoint analysis, this number-of-levels effect has been replicated using a variety of data collection methods, measurement scales for the dependent variable, and estimation methods. Accounts for the number-of-levels effect include methodological and psychological factors. I will first review the existing accounts for the number-of-levels effect and then propose a new account based on attentional contrast from the category learning literature. Existing Accounts of the Number-Of-Levels Effect The greater share of research on the number-of-levels effect has dealt with methodological accounts. This research attributes the number-of-levels effect to the 4

PAGE 15

5 properties of the measurement scale, the usage of the measurement scale, or the method of parameter estimation. Properties of the Measurement Scale Several accounts of the number-of-levels effect relate to the properties of the measurement scale. Wittink, Krishnamurthi, and Reibstein (1989) hypothesized that the number-of-levels effect was partly due to the use of a rank-order measurement procedure. They showed analytically why the effect occurs for rank-order preferences using least-squares analysis on a multitude of full factorial and fractional factorial designs. Since it is impossible to analytically show the same effect exists for ratings, Wittink et al. (1989) examined whether the number-of-levels effect exists for rating scales empirically, comparing the results from full-profile ratings with the results from rank-order preferences obtained by paired comparisons and trade-off matrix rank orders. The full-profile preferences were obtained using a ten-point scale. Wittink et al. (1989) compared derived importances of a three-level and five-level price attribute in five categories and found that the number-of-levels effect was equally strong using rank-order or rating scale measurement procedures, the implication being that rank-order measurement was not the sole source of the number-of-levels effect. Steenkamp and Wittink (1994) proposed that the number-of-levels effect occurs with a rating scale measurement procedure because the preference ratings lack metricity. They hypothesized that rating scales properties are closer to ordinal than to interval measurement. Since the method of magnitude estimation has been claimed to result in interval-scale data, Steenkamp and Wittink (1994) used this method to test their hypothesis. The method of magnitude estimation consists of asking participants to match their evaluation of a profile to the perceived magnitude expressed on a physical medium.

PAGE 16

6 Steenkamp and Wittink (1994) used numerical estimation (i.e., participants assign numbers to profiles) and line production (i.e., participants draw lines) as response modalities. They found that a nine-point rating scale and a magnitude estimation procedure provide number-of-levels effects of comparable size. However, the magnitude estimation procedure allows for a test of the metricity of the results of each respondent separately. Categorizing the respondents as metric versus non-metric, Steenkamp and Wittink (1994) found a reduced number-of-levels effect in participants whose responses satisfied the criteria for metricity. In other words, the lack of metricity of the rating method per se had no impact on the number-of-levels effect, but individual differences in metricity were identified as a source of the effect. Usage of the Measurement Scale Other measurement procedure explanations relate to respondents use of the measurement scale. Wittink et al. (1992b) included a balancing manipulation in their study (i.e., they constructed a condition in which dominated pairs were eliminated from the procedure), provided respondents made pair-wise comparisons among alternatives. Using adaptive conjoint analysis, Wittink et al. (1992b) gathered full factorial data on one product category (notebook computers) defined on six attributes, of which four attributes alternated between two and four levels, depending on the condition. Wittink et al. (1992b) found that the number-of-levels effect is considerably reduced when dominated pairs are eliminated from the procedure. Wittink et al. (1992b) explain this result in terms of participants tendencies to give preference ratings toward the midpoint of the scale. Since preference ratings for dominated pairs are expected to be more extreme than those for non-dominated pairs, the exclusion of dominated pairs provides less opportunity for distortion.

PAGE 17

7 Verlegh et al. (2002) propose the number-of-levels effect may be a consequence of a persons tendency to perceive a constant, absolute distance between levels of an attribute (i.e., exhibit a uniform distribution tendency). They tested this hypothesis by comparing results on five different measures of attribute importance, using one product category (color TVs). The product category was defined on six attributes, of which two alternated between two and four levels, depending on the condition. Verlegh et al. (2002) found significant number-of-level effects on two measures of attribute importance: difference in part worths in a full-profile conjoint and attractiveness ratings for individual attribute levels. On the remaining three measures (stated importance of a set of attributes, linear regression coefficients, and paired comparisons varying one attributes exterior levels at a time), no significant number-of-levels effect was observed. Verlegh et al. (2002) interpreted their results as evidence for respondents tendency to distribute stimulus representations uniformly over a restricted internal continuum (i.e., the distance between each level of an attribute was constant) in full-profile and attribute level ratings. Ceteris paribus, the addition of intermediate levels will lead respondents to distribute the exterior attribute levels more toward the extremes of the internal continuum. Method of Parameter Estimation A second potential source of the number-of-levels effect is the method of parameter estimation. In their seminal paper, Currim, Weinberg, and Wittink (1981) mentioned the method of parameter estimation as a possible source of the number-of-levels effect. For the 2x3 trade-off matrix they used to obtain their empirical results, they estimated the parameters based on permissible configurations of preferences using ordinary least squares. They derived the importance weights for two extreme cases: a dominant two-level attribute and a dominant three-level attribute. They found that the two-level

PAGE 18

8 attributes importance weights varied between 0.20 (i.e., when the three-attribute level attribute was dominant) and 0.60 (i.e., when the two-level attribute was dominant). Analogously, the three-level attributes importance weights ranged from 0.40 (i.e., when the two-level attribute was dominant) to 0.80 (i.e., when the three-level attribute was dominant). In other words, while the range of possible derived importance weights was not affected by the number of levels, its size was affected by it. Currim et al. (1981) used this analytical evidence as support for their hypothesis that the method of parameter estimation is a source of the number-of-levels effect. Subsequent studies have provided additional support for the method of parameter estimation account of the number-of-levels effect. Wittink et al. (1989) showed analytically why the number-of-levels effect occurs on rank-order preferences on a multitude of full factorial and fractional factorial designs. Assuming an additive, main-effect model, they calculated raw attribute importance weights for any number of attributes and any attribute levels using least-squares analysis. Based on the raw attribute importance weights, they estimated maximum relative weights for each attribute, and the differences and ratios between the attributes within a design. They found that maximum relative weights increase with the number of levels and concluded that the method of parameter estimation is a source of the number-of-levels effect. Wittink, Krishnamurthi, and Nutter (1982) gathered participants preference data for hypothetical summer jobs, defined on four attributes, of which two attributes alternated between two and four attribute levels, depending on the condition. Two data collection methods were used: a fractional factorial profile evaluation and a trade-off matrix. Wittink et al. (1982) estimated the parameters using a non-metric method (LINMAP, Srinivasan and Shocker

PAGE 19

9 1973) and found that the derived importance weights vary with the number of levels on which an attribute is defined. They concluded that the number-of-levels effect is at least partially accounted for by the method of parameter estimation. Perceptual Processes Since the initial identification of the number-of-levels effect, researchers have acknowledged that the effect could be a consequence of perceptual processes (Currim et al. 1981). More specifically, it has been proposed that respondents may consciously or unconsciously assign more weight to attributes with more levels. Several studies have included one or more tests for this weighting explanation, but have failed to find support for the hypothesis (Steenkamp and Wittink 1994; Verlegh et al. 2002; Wittink et al. 1992b). For example, Verlegh et al. (2002) hypothesized that increased attention should affect the stated importance of a set of attributes and paired comparisons varying one attributes exterior levels at a time. However, no number-of-levels effect was observed on these two measures. Attentional Contrast Several of the proposed accounts of the number-of-levels effect can partially explain the effect. However, no single explanation can wholely account for the number-of-levels effect. I propose that attentional contrast may be an additional source of the number-of-levels effect. The proposed attentional contrast explanation is based on two observations. First, the occurrence of the number-of-levels effect depends on the preference elicitation method. Decompositional methods ask respondents to provide multiattribute judgments for the full-profile descriptions of alternatives and then derive individual attribute level values from the judgments. The number-of-levels effect is typically found using

PAGE 20

10 decompositional methods. Compositional methods ask respondents to assign values to each level of an attribute. The values are combined to construct an overall evaluation of an alternative. The number-of-levels effect is typically not found using compositional methods. There are three studies that have used a combination of decompositional and compositional methods. Wittink et al. (1992a) showed that the magnitude of the number-of-levels effect was greater for a full-profile procedure than for an adaptive conjoint analysis. Whereas full-profile procedure is an entirely decompositional method, adaptive conjoint analysis is a combination of a decompositional and compositional procedure. No number-of-levels effect was observed in the compositional data. Verlegh et al. (2002) used a decompositional and compositional procedure and found evidence for a number-of-levels effect in the compositional data. Yet, the compositional data were collected after the decompositional data, suggesting the decompositional procedure may have biased the values provided in the compositional procedure. Steenkamp and Wittink (1994) also found minimal evidence for a number-of-levels effect on stated attribute importances that were gathered after the decompositional data. Johnson (1992) found evidence for a number-of-levels effect using only a compositional method. He asked participants to assess attribute values directly. However, the values of two attributes were assessed at the same time. More specifically, participants were told that a base TV with 17 inch screen and monophonic sound was worth $200. Then, they were asked to evaluate what other options were worth: 19 inch screen, 21 inch screen and excellent stereo sound in the 3+2 condition and 21 inch screen, good stereo sound and excellent stereo sound in the 2+3 condition. Hence, participants

PAGE 21

11 could interpret this task as full-profile evaluation. For example, when evaluating the 19 inch screen, it is likely that they imagined a monophonic sound TV. Second, the designs used in studies on the number-of-levels effect invariably compare conditions in which the number of levels of more than one attribute is varied at once (i.e., varying the number of levels of attribute 1 is confounded with varying the number of attribute levels of attribute 2). For example, consider an experiment in which participants respond to a 2x4 design (i.e., attribute 1 has two levels and attribute 2 has four levels) or a 4x2 design. If one finds that the derived importance weight on attribute 1 is larger in the 4x2 condition than the 2x4 condition, then the weighting difference may be due to the increased number of levels in attribute 1 or to the accompanying decrease in the number of levels in attribute 2. In other words, the weighting effect that has been interpreted as a number-of-levels effect may not (solely) be caused by the increased number of levels in that attribute. It may also be caused by the reduced number of levels of other attributes. It is important to note that all published demonstrations of the number-of-levels effect have this characteristic built into their manipulations. An overview of the designs can be found in Table 1. Based on these two observations, I will now propose a new account for the number-of-levels effect: the attentional contrast account. The attentional contrast account depends on the idea that people direct attention away from attribute levels they have seen more often and towards attribute levels they have seen less often. For example, in a 4x2 design, the first attribute has four levels and each of these levels will be experienced twice in a full-factorial design. The second attribute has two levels and each of these levels will be experienced four times in a full-factorial design. Thus, the levels of the first attribute will

PAGE 22

12 be relatively more novel than the levels of the second attribute. The attentional contrast account proposed that attention will contrast away from the more common levels (i.e., attribute 2) and toward the less common levels (i.e., attribute 1). Thus, the attribute with more levels will be perceived as more important. The attentional contrast hypothesis has its foundation in the categorical learning literature. The most successful theories in categorization and learning incorporate some notion of selective attention (Nosofsky 1986, 1991; Sutherland and Mackintosh 1971). More specifically, consumers increasingly pay attention to attributes and attribute levels that are important, and decreasingly pay attention to irrelevant attributes, irrelevant attribute levels, or both (Goldstone and Steyvers 2001). For example, Goldstone and Steyvers (2001) showed that selective attention can be directed toward arbitrary dimensions created by randomly pairing faces. They selected four photographs of bald heads and generated two dimensions by morphing between the photographs. Each of the sixteen created faces varied in the proportion of each of the four faces. Based on category relevance, both sensitization of relevant dimensions and desensitization of irrelevant dimensions were observed. Moreover, Goldstone and Steyvers (2001) showed that once a participant had learned to direct attention toward or away from a dimension, this learning extended to different stimuli. Kersten, Goldstone, and Schaffert (1998) provide additional evidence that differential selective attention mechanisms operate on attributes and attribute levels. They show that within a given trial, a mechanism of attentional contrast directs attention away from attribute levels that have been associated with an outcome. In a series of experiments, participants learned verbs that were defined on two attributes: path and

PAGE 23

13 manner. Each verb was always accompanied by a particular value of one or both of these attributes. One group of participants learned conjunctive verbs (i.e., verbs that could be distinguished on path and manner). Another group first learned path verbs (i.e., verbs that could only be distinguished on the basis of path). When this second group learned new conjunctive verbs (i.e., verbs that could be distinguished based on manner as well as the same values of path), they attended more to level of manner than the participants who first learned conjunctive verbs. In other words, since the path values had been associated with an outcome (i.e., a verb), attentional contrast was directed toward the new manner values within subsequent trials. The attentional contrast account is consistent with the observation that the occurrence of the number-of-levels effect depends on whether the method is compositional or decompositional. In a compositional method, each level of an attribute is repeated only once and it is repeated out of context, so no attentional contrast can operate. In a decompositional method, attribute levels are repeated and this repetition occurs in the context of other attribute levels that have been more or less repeated, so attentional contrast can operate. The attentional contrast account also makes specific predictions in the event the levels of attributes are unconfounded in an experimental design. For example, imagine three designs are investigated: a 4x2, a 2x4, and a 4x4. The derived importance weight of attribute one (two) should be greater (smaller) in the 4x2 than the 2x4 studies, as has been observed in prior studies. However, the attentional contrast account predicts that the derived importance weight of attribute one should be greater in the 4x2 design than the 4x4 design because the number of levels of attribute two are more frequent in the 4x2

PAGE 24

14 condition. The attentional contrast account also predicts that the derived importance weight of attribute two should be equal in the 4x2 and the 4x4 conditions because the number of levels of attribute one are equivalent. Similar predictions can be made when comparing the 2x4 and the 4x4 conditions. The predictions of the attentional contrast hypothesis are not consistent with the predictions that might be expected if one were simply to count the levels of an attribute. For example, if the predictions were based solely on the number of levels of an attribute, the derived importance weight of attribute one should be the same in the 4x2 condition and the 4x4 condition because the number of levels of attribute one are equivalent. The derived importance weight of attribute two should be less in the 4x2 condition than the 4x4 condition because the number of levels of attribute two is less in the 4x2 condition. Thus, an explanation based solely on attribute levels makes a set of qualitatively different predictions than the attentional contrast hypothesis. In a series of four experiments, I will test this attentional contrast account for the number-of-levels effect. In experiments 1 and 1A, I will extend the designs that have typically been used to examine the number-of-levels effect to test an attentional contrast account. In experiment 2, I will manipulate attentional contrast directly to examine its potential impact when there are no differences in number of levels. In experiment 3, I create attentional contrast on an attribute with a lower number of levels.

PAGE 25

15 Table 1. Number-of-levels manipulations used in published studies Authors Number-Of-Levels Manipulation Currim, Weinberg and Wittink (1981) number-of-levels manipulation between attributes (2 or 3 levels) Johnson (1992) 3+2 vs. 2+3 (used lists) Orme (1998) number-of-levels manipulation between attributes (2 to 5 levels) Steenkamp and Wittink (1994) 2x4 vs. 4x2 Verlegh, Schifferstein and Wittink (2002) 4x2x2x2x3x2 vs. 2x4x2x2x3x2 Wittink, Huber, Fiedler and Miller (1992a) 3x4x2x2x4x2x2x2x2 vs. 3x2x4x4x2x2x2x2x2 Wittink, Huber, Zandan and Johnson (1992b) 2x2x4x4 vs. 4x4x2x2 Wittink, Krishnamurthi and Nutter (1982) 4x4x2x2 vs. 4x2x4x2 Wittink, Krishnamurthi and Reibstein (1989) 3 vs. 5 levels

PAGE 26

CHAPTER 3 EXPERIMENT 1: DEMONSTRATION OF THE NUMBER-OF-LEVELS EFFECT WITH EXPANDED DESIGN The objective of the first experiment was to test the attentional contrast account of the number-of-levels effect by expanding the experimental design that has typically been used to examine the bias. More specifically, the number of levels of only one attribute was varied across conditions. The attentional contrast hypothesis (Kersten et al. 1998), predicted that the number-of-levels effect would be sensitive to the relative novelty of each attribute level. Method Design The design was a 3 cell between-subjects design (4x2, 2x4, and 4x4 attribute levels) with three between-subject product replicates (refrigerators, LCD televisions, and digital cameras). Each offer was defined on two attributes. Participants were asked to provide their reservation price for each product offer. This dependent variable should not be sensitive to the uniform distribution problem associated with a fixed scale (Verlegh et al. 2002) or suffer from a lack of metricity (Steenkamp and Wittink 1994). The design of experiment 1 is presented in Figure 1. Stimuli The product categories were refrigerators, LCD televisions, and digital cameras. Each product offer was defined on two attributes for which participants were known to have monotonic preferences. The attributes were the following: annual energy cost and 16

PAGE 27

17 capacity (refrigerators), screen size and horizontal viewing angle (LCD televisions), imaging sensor resolution and image storage capacity (digital cameras). Each attribute had two or four possible attribute levels (see Table 2 for an overview of the attribute levels). These attribute levels were either exterior (i.e., the two values in a 2-level attribute or the two extreme values in a 4-level attribute) or interior (the two inserted levels in a 4-level attribute). Exterior offers consisted of the exterior levels of the two attributes (e.g., [1,1], [1,4], [2,1], [2,4] in a 2x4 design). Intermediate offers consisted of the interior levels of the two attributes (e.g., [2,2], [2,3], [3,2], [3,3] in a 4x4 design), or of an exterior level of one attribute and an interior level of a second attribute (e.g., [1,2], [1,3], [2,2], [2,3] in a 2x4 design). Procedure Participants were seated at personal computers. Participants read instructions stating that they would be stating reservation prices for products described on two dimensions. Participants were asked to assume that the products were identical on all non-defined attributes and were informed of the average price of the product. Then, the product offers were presented. First, the participants stated their reservation price for the intermediate offers. These four stimuli were presented in random order. Next, participants stated their reservation price for the exterior offers. Again, the stimuli were presented in random order. Thus, each participant provided reservation price judgments for eight stimuli with the last four being the identical exterior stimuli. Results The analysis was conducted using the mean difference scores of the exterior offers. For example, attribute ones differences score in a 2x4 design is the difference between the average of [1,1] and [1,4] and the average of [2,1] and [2,4]. Similarly, attribute twos

PAGE 28

18 differences score in a 2x4 design is the difference between the average of [1,1] and [2,1] and the average of [1,4] and [2,4]. The four exterior offers were identical across conditions, so these difference scores could be compared. The results of participants whose ratings on the exterior offers were non-monotonic were removed from the dataset. Table 3 presents the mean difference scores per attribute for each condition. Refrigerators Data of 91 participants were retained for analysis. The analysis for each attribute was performed separately. The first analysis was on attribute one (annual energy cost). The mean difference scores of the 2x4 condition (X = $78.97) and the 4x2 condition (X = 144.24) were significantly different (F(1, 88) = 10.38, p < .01). This replicates the number-of-levels effect using a design that is common in the academic literature. The mean difference scores of the 2x4 condition (X = $78.97) and the 4x4 condition (X = $63.84) were not significantly different (F(1,88) = 0.63, p > .05). This result is inconsistent with the hypothesis that the number-of-levels effect is sensitive only to the number of levels of an attribute. The mean difference scores of the 4x2 condition (X = $144.22) and the 4x4 condition (X = $63.84) were significantly different (F(1, 88) = 17.16, p < .01). This result is consistent with the attentional contrast hypothesis. The second analysis was on attribute two (capacity). The mean difference scores of the 2x4 condition (X = $122.55) and the 4x2 condition (X = $120.17) were not significantly different (F(1, 88) = 0.01, p > .05). The mean difference scores of the 2x4 condition (X = $122.55) and the 4x4 condition (X = $100.93) were not significantly different (F(1, 88) = 0.96, p > .05). The mean difference scores of the 4x2 condition (X = $120.17) and the 4x4 condition (X = $100.93) were not significantly different (F(1, 88) = 0.73, p > .05)

PAGE 29

19 LCD Televisions Data of 90 participants were retained for analysis. The first analysis was on attribute one (screen size). The mean difference scores of the 2x4 condition (X = $176.20) and the 4x2 condition (X = $257.37) were significantly different (F(1, 87) = 3.98, p < .05). The mean difference scores of the 2x4 condition (X = $176.20) and the 4x4 condition (X = $196.82) were not significantly different (F(1, 87) = 0.26, p > .05). The mean difference scores of the 4x2 condition (X = $257.37) and the 4x4 condition (X = $196.82) were not significantly different (F(1, 87) = 2.21, p > .05). These results are not supportive of the number of levels hypothesis or the attentional contrast hypothesis. The second analysis was on attribute two (horizontal viewing angle). The mean difference scores of the 2x4 condition (X = $140.64) and the 4x2 condition (X = $130.37) were not significantly different (F(1, 87) = 0.26, p > .05). The mean difference scores of the 2x4 condition (X = $140.64) and the 4x4 condition (X = 122.82) were not significantly different (F(1, 87) = 0.78, p > .05). The mean difference scores of the 4x2 condition (X = $130.37) and the 4x4 condition (X = $122.82) were not significantly different (F(1, 87) = 0.14, p > .05). Digital Cameras Data of 117 participants were retained for analysis. The first analysis was on attribute one (imaging sensor resolution). The mean difference scores of the 2x4 condition (X = $134.90) and the 4x2 condition (X = $169.31) were significantly different (F(1, 114) = 3.18, p = .08). The mean difference scores of the 2x4 condition (X = $134.90) and the 4x4 condition (X = $131.49) were not significantly different (F(1, 114) = 0.03, p > .05). The mean difference scores of the 4x2 condition (X = $169.31) and the

PAGE 30

20 4x4 condition (X = $131.49) were significantly different (F(1, 114) = 4.07, p < .05). These results are supportive of the attentional contrast hypothesis. The second analysis was on attribute two (image storage capacity). The mean difference scores of the 2x4 condition (X = $110.42) and the 4x2 condition (X = $106.81) were not significantly different (F(1, 114) = 0.07, p > .05). The mean difference scores of the 2x4 condition (X = $110.42) and the 4x4 condition (X = $95.86) were not significantly different (F(1, 114) = 1.15, p > .05). The mean difference scores of the 4x2 condition (X = $106.81) and the 4x4 condition (X = $95.86) were not significantly different (F(1, 114) = 0.63, p > .05). Discussion The data from experiment 1 replicate the number-of-levels effect that has been reported in the literature and provide some new insights as to the potential source of this effect. First, all three replicates showed a number-of-levels effect on attribute one when the analysis was restricted to the 2x4 and 4x2 conditions. This type of design is common in the number-of-levels literature. Second, when the 4x4 condition was also included in the analysis, the mean difference score for attribute one was similar to the difference score in the 2x4 condition, but less than the difference score in the 4x2 conditions. Although this pattern of results held for only the refrigerator and digital camera replicates, and only for attribute one, the result is consistent with the attentional contrast hypothesis. The most perplexing result in experiment 1 is the apparent attentional contrast effect on attribute one and the lack of the attentional contrast effect on attribute 2. One possible reason for this finding is that attentional contrast only operates on attributes that are perceived to be important (i.e., worth attending to). In other words, an attribute must

PAGE 31

21 be perceived to be important if its attribute levels are to exhibit an attentional contrast based number-of-levels effect. The implication is that attribute two could also exhibit a number-of-levels effect if it were to be perceived as more important. Experiment 1A will examine this hypothesis. Experiment 1A The goal of experiment 1A was to assess whether the differential effects observed on attributes one and two in experiment 1 could be a consequence of a (perceived) difference in importance between the two attributes. In Experiment 1A, the (relative) importance of the second attribute was increased in two ways. First, the range of the attribute twos levels was increased and the range of attribute ones levels was decreased. Second, participants were asked to elaborate on product usage. This second strategy was adopted from Huffman (1997), who showed that elaboration on product usage increases the weight of usage-related attributes whereas elaboration on product purchase increases the weight of purchase-related attributes. Of the three replicates, only the refrigerator category had a purchase-related attribute (annual energy cost) and a usage-related attribute (capacity). Given the failure to find a number-of-levels effect on the usage-related attribute in this category, participants were asked to elaborate on usage. Method The design was identical to experiment 1 with the limitation that only one of the replicates (refrigerators) was investigated. The stimulus were adjusted so that the range of attribute two (capacity) was doubled and the range of attribute one (annual energy cost) was reduced. The attribute levels for capacity were 13, 17, 21, and 25 cubic feet. The attribute levels for annual energy cost were $90, $85, $80, and $75.

PAGE 32

22 The procedure was identical to experiment 1 except for one change. Prior to being exposed to the eight product offers, participants were asked to elaborate on product usage. Participants were told to Think back to the times you have used a refrigerator. Picture, in your minds eye, the situations in which you have used it. Below, please describe your thoughts and feelings about using the refrigerator. What was important to you when you were using the refrigerator? Results Table 4 presents the mean difference scores per attribute for each condition. Data of 120 participants were retained for analysis. The first analysis was on attribute one (annual energy cost). T the mean difference scores of the 2x4 condition (X = $74.25) and the 4x2 condition (X = $98.37) were significantly different (F(1, 117) = 4.09, p < .05). The mean difference scores of the 2x4 condition (X = $74.25) and the 4x4 condition (X = $71.30) were not significantly different (F(1, 117) = 0.05, p > .05). The mean difference scores of the 4x2 condition (X = $98.37) and the 4x4 condition (X = $71.30) were significantly different (F(1, 117) = 5.70, p < .05). These results are supportive of the attentional contrast hypothesis. The second analysis was on attribute two (capacity). The mean difference scores of the 2x4 condition (X = $265.13) and the 4x2 condition (X = $180.93) were significantly different (F(1, 117) = 7.10, p < .05). The mean difference scores of the 4x2 condition (X = $180.93) and the 4x4 condition (X = $220.72) were not significantly different (F(1, 117) = 1.75, p > .05). The mean difference scores of the 2x4 condition (X = $265.13) and the 4x4 condition (X = $220.72) were not significantly different (F(1, 117) = 1.76, p > .05). These results are supportive of the attentional contrast hypothesis.

PAGE 33

23 Discussion Experiment 1A shows that a number-of-levels effect can be produced on both attributes in a two-factor conjoint design, provided the participant perceives each attribute as important. When the importance of a second attribute was increased by increasing its range and by getting consumers to elaborate on situations in which the attribute might be important, a number-of-levels effect was obtained. In addition, the mean difference score for attribute one in the 4x4 condition was similar to the difference score in the 2x4 condition, but less than the difference score in the 4x2 conditions. Similarly, the mean difference score for attribute two in the 4x4 condition was similar to the difference score in the 2x4 condition, but less than the difference score in the 2x4 conditions. These results are consistent with the attentional contrast hypothesis. Experiments 1 and 1A show that the number-of-levels effect does not occur because of increasing levels on the focal attribute, but because of the decreasing levels on the accompanying attribute. Moreover, the effect depends on the (perceived) importance of an attribute. These results are consistent with an attentional contrast account, which predicts that within subsequent offers, attentional contrast is produced on the attribute that has more levels, as its levels are relatively more novel than the attribute levels of the attribute with the lower number of levels. However, experiments 1 and 1A were designed to manipulate the number of levels and are, therefore, not the most direct test of the attentional contrast account. While the conditions that were compared differed on attentional contrast on the focal attributes, they also differed on the number of levels of the accompanying attribute. Experiment 2 provides a more direct test of the attentional contrast account by showing that it can

PAGE 34

24 create a weighting difference while the number of levels on both focal and accompanying attributes is kept constant.

PAGE 35

25 A Attribute 2 1 2 3 4 1 2 3 Attribute 1 4 B Attribute 2 1 2 3 4 1 2 3 Attribute 1 4 C Attribute 2 1 2 3 4 1 2 3 Attribute 1 4 Note.Exterior offers are colored in black, intermediate offers are shaded in grey, and interior offers are hatched. Figure 1. Design for experiments 1 and 1A. A) 4x2 Condition. B) 4x2 Condition. C) 4x4 Condition

PAGE 36

26 Table 2. Attribute levels from experiment 1 Attribute level Product category Attribute 1 2 3 4 Refrigerators annual energy cost ($n) 100 90 80 70 capacity (n cubic feet) 16 18 20 22 LCD televisions screen size (n inches) 20 24 28 32 horizontal viewing angle (n off-center) 45 55 65 75 Digital cameras imaging sensor resolution (n MegaPixels) 3.2 4.4 5.6 6.8 image storage capacity (up to n MB) 128 256 384 512 Table 3. Mean difference scores and simple effects for experiment 1 Mean difference score F values Attribute 1 4x2 2x4 4x4 4x2 vs 2x4 4x2 vs 4x4 2x4 vs 4x4 Annual energy cost (Refrigerators) 144.24 78.97 63.84 10.38 a 17.16 a 0.63 Screen size (LCD televisions) 257.37 176.20 196.82 3.98 b 2.21 0.26 Imaging sensor resolution (Digital cameras) 169.31 134.90 131.49 3.18 c 4.07 b 0.03 Attribute 2 4x2 2x4 4x4 4x2 vs 2x4 4x2 vs 4x4 2x4 vs 4x4 Capacity (Refrigerators) 120.17 122.55 100.93 0.01 0.73 0.96 Horizontal viewing angle (LCD televisions) 130.37 140.64 122.82 0.26 0.14 0.78 Image storage capacity (Digital cameras) 106.81 110.42 95.86 0.07 0.63 1.15 a implies significance at 99% level of confidence b implies significance at 95% level of confidence c implies significance at 90% level of confidence

PAGE 37

27 Table 4. Mean difference scores and simple effects for experiment 1A Mean difference score F values Attribute 1 4x2 2x4 4x4 4x2 vs 2x4 4x2 vs 4x4 2x4 vs 4x4 Annual energy cost (Refrigerators) 98.37 74.25 71.30 4.09 b 5.70 b 0.05 Attribute 2 4x2 2x4 4x4 4x2 vs 2x4 4x2 vs 4x4 2x4 vs 4x4 Capacity (Refrigerators) 180.93 265.13 220.72 7.10 a 1.75 1.76 a implies significance at 99% level of confidence b implies significance at 95% level of confidence

PAGE 38

CHAPTER 4 EXPERIMENT 2: TEST OF ATTENTIONAL CONTRAST ACCOUNT WITH NO DIFFERENCE IN NUMBER OF LEVELS The goal of experiment 2 was to examine if attentional contrast can create a weighting effect when there is no difference in number of levels. In experiments 1 and 1A, we found evidence for an attentional contrast account by varying the number of levels of only one attribute at a time. However, attentional contrast was confounded with the number of levels manipulation. By manipulating attentional contrast in a context in which there is no difference in number of levels, we want to obtain more direct evidence that the attentional contrast by itself can account for the weighting difference. In addition, the three replicates and procedure from experiment 1 were used in the experiment. It was felt that demonstrating a robust attentional contrast effect on one attribute was superior to demonstrating an attentional contrast effect on two attributes of a single replicate. Method Design and Stimuli The design was a 3 cell between-subjects design (contrast on attribute 1, contrast on attribute 2, balanced contrast) with three between-subject replicates (refrigerator, LCD televisions, digital cameras). In all three experimental conditions, each attribute was defined on 4 levels (see Figure 2). The cells differed on the attribute on which attentional contrast was created. In the contrast on attribute 1 condition, attentional contrast was created on attribute 1. This contrast was created via the selection of offers that were initially rated by the participant. The intermediate offers were [2,1], [2,2], [2,3], [2,4], 28

PAGE 39

29 [3,1], [3,2], [3,3], and [3,4]. Thus, when the participant saw the exterior offers of [1,1], [1,4], [4,1], and [4,4], the values of the first attribute were novel and the values of the second attribute were familiar. In the contrast on attribute 2 condition, the intermediate offers were [1,2], [1,3], [2,2], [2,3], [3,2], [3,3], [4,2],and[4,3]. Thus, when the participant saw the exterior offers of [1,1], [1,4], [4,1], and [4,4], the values of the second attribute were novel and the values of the first attribute were familiar. In the balanced contrast condition, the intermediate offers were [2,2], [2,3], [3,2], and [3,3]. Thus, when the participant saw the exterior offers of [1,1], [1,4], [4,1], and [4,4], the values of both attributes were on average equally novel, hence there was no contrast. Note that this last condition is particularly important because it will show that the number-of-levels effect is not an issue of novelty (i.e., seeing levels that have not been seen before), but an issue of attentional contrast (i.e., seeing levels that are novel relative to the levels seen on the other attribute). In all other respects, the procedure was identical to the one used in experiment 1. Predictions The attentional contrast hypothesis makes predictions for attributes one and two. With respect to attribute one, the attentional contrast hypothesis predicts that the mean difference score should be greater in the contrast on attribute 1 condition than in the contrast on attribute 2 condition or the balanced contrast condition. With respect to attribute two, the attentional contrast hypothesis also predicts that the mean difference score should be greater in the contrast on attribute 2 condition than in the contrast on attribute 1 condition or the balanced contrast condition. Given the results of experiments 1 and 1A, it was anticipated that the attentional contrast effect would be obtained on attribute 1, but not attribute 2.

PAGE 40

30 Results Table 5 presents the mean difference scores per attribute in each condition. Refrigerators Data of 104 participants were retained for analysis. The analysis for each attribute was performed separately. The first analysis was on attribute one (annual energy cost). The mean difference scores of the contrast on attribute 1 condition (X = $99.37) and the contrast on attribute 2 condition (X = $68.60) were significantly different (F(1, 101) = 4.96, p < .05). The mean difference scores of the contrast on attribute 1 condition (X = $99.37) and the balanced contrast condition (X = $70.04) were significantly different (F(1, 101) = 4.87, p < .05). The second analysis was on attribute two (capacity). The mean difference scores of the contrast on attribute 2 condition (X = $94.02) and the contrast on attribute 1 condition (X = $126.58) were not significantly different (F(1, 101) = 2.65, p > .05). The mean difference scores of the contrast on attribute 2 condition (X = $94.02) and the balanced contrast condition (X = $99.79) were not significantly different (F(1, 101) = 0.08, p > .05). LCD televisions Data of 111 participants were retained for analysis. The first analysis was on attribute one (screen size). The mean difference scores of the contrast on attribute 1 condition (X = $286.04) and the contrast on attribute 2 condition (X = $228.72) were not significantly different (F(1, 108) = 2.13, p > .05). The mean difference scores of the contrast on attribute 1 condition (X = $286.04) and the balanced contrast condition (X = $192.99) were significantly different (F(1, 108) = 5.27, p < .05). The second analysis was on attribute two (horizontal viewing angle). The mean difference scores of the contrast on

PAGE 41

31 attribute 2 condition (X = $133.74) and the contrast on attribute 1 condition (X = $146.52) were not significantly different (F(1, 108) = 0.26, p > .05). The mean difference scores of the contrast on attribute 2 condition (X = $133.74) and the balanced contrast condition (X = $101.04) were not significantly different (F(1, 108) = 1.75, p > .05). Digital cameras Data of 133 participants were retained for analysis. The first analysis was on attribute one (imaging sensor resolution). The mean difference scores of the contrast on attribute 1 condition (X = $190.12) and the contrast on attribute 2 condition (X = $143.15) were significantly different (F(1, 130) = 5.78, p < .05). The mean difference scores of the contrast on attribute 1 condition (X = $190.12) and the balanced contrast condition (X = $144.37) were significantly different (F(1, 130) = 6.14, p < .05). The second analysis was on attribute two (image storage capacity). The mean difference scores of the contrast on attribute 2 condition (X = $128.43) and the contrast on attribute 1 condition (X =$114.12) were not significantly different (F(1, 130) = 0.89, p > .05). The mean difference scores of the contrast on attribute 2 condition (X = $128.43) and the balanced contrast condition (X = $119.46) were not significantly different (F(1, 130) = 0.34, p > .05). Discussion Experiment 2 manipulated attentional contrast directly by varying the relative novelty of the attribute levels while keeping the number of levels of the attributes constant. The results showed that attentional contrast can create a weighting difference when there is no difference in number of levels. While this effect was consistently produced on attribute 1, the results on the second attribute were non-significant. Since the stimuli used in this experiment were identical to the ones used in experiment 1, we

PAGE 42

32 attribute this lack of significance to the fact that attribute 2 was perceived to be less important (see experiment 1A).

PAGE 43

33 A Attribute 2 1 2 3 4 1 2 3 Attribute 1 4 B Attribute 2 1 2 3 4 1 2 3 Attribute 1 4 C Attribute 2 1 2 3 4 1 2 3 Attribute 1 4 Note.Exterior offers are colored in black, intermediate offers are shaded in grey. The hatched offers are presented twice. Figure 2. Design for experiment 2. A) Contrast on Attribute 1 Condition. B) Contrast on Attribute 2 Condition. C) Balanced Contrast Condition.

PAGE 44

34 Table 5. Mean difference scores and simple effects for experiment 2 Mean difference score F values Attribute 1 contrast on attribute 1 contrast on attribute 2 balanced contrast contrast on attribute 1 vs contrast on attribute 2 contrast on attribute 1 vs balanced contrast Annual energy cost (Refrigerators) 99.37 68.60 70.04 4.96 b 4.87 b Screen size (LCD televisions) 286.04 228.72 192.99 2.13 5.27 b Imaging sensor resolution (Digital cameras) 190.12 143.15 144.37 5.78 b 6.14 a Attribute 2 contrast on attribute 1 contrast on attribute 2 balanced contrast contrast on attribute 1 vs contrast on attribute 2 contrast on attribute 2 vs balanced contrast Capacity (Refrigerators) 126.58 94.02 99.79 2.65 0.08 Horizontal viewing angle (LCD televisions) 146.52 133.74 101.04 0.26 1.75 Image storage capacity (Digital cameras) 114.12 128.43 119.46 0.89 0.34 a implies significance at 99% level of confidence b implies significance at 95% level of confidence

PAGE 45

CHAPTER 5 EXPERIMENT 3: TEST OF ATTENTIONAL CONTRAST ACCOUNT WITH REVERSED NUMBER OF LEVELS The goal of experiment 3 was to show that the number-of-levels effect can be reversed when attentional contrast is created on the attribute with the lower number of levels. In experiments 1 and 1A, attentional contrast was shown to create weighting differences but the number of levels of the accompanying attribute was varied, but not the number of levels of the focal attribute. In experiment 2, attentional contrast created weighting differences when there was no difference in number of attribute levels. In experiment 3, the test of the attentional contrast account is made more rigorous. More explicitly, I test whether the number-of-levels effect can be obtained on the attribute with the lower number of levels. Method Design and Stimuli The design was a 3 cell between-subjects design (4x6, 6x4, and 6x6 attribute levels) with three between-subject product replicates (refrigerators, LCD televisions, digital cameras). Each attribute had four or six possible attribute levels. For an overview of the attribute levels, see Table 6. The intermediate offers were [3,1], [3,2], [3,3], [3,4], [3,5], [3,6], [4,1], [4,2], [4,3], [4,4], [4,5], and [4,6] in the 4x6 condition, [3,1], [4,1], [3,2], [4,2], [3,3], [4,3], [3,4], [4,4], [3,5], [4,5], [3,6], and [4,6] in the 6x4 condition, and [1,3], [1,4], [2,2], [2,5], [3,1], [3,6], [4,1], [4,6], [5,2], [5,5], [6,3], and [6,4] in the 6x6 condition. The four exterior offers were [1,1], [1,6], [6,1], and [6,6]. The design of 35

PAGE 46

36 experiment 3 is presented in Figure 3. In the 4x6 condition, the levels of attribute 1 were more novel than the attribute levels of attribute 2 at the time the participants evaluated the exterior offers. In other words, attentional contrast was created on attribute 1. In the 6x4 condition, the levels of attribute 2 were more novel than the attribute levels of attribute 1 at the time the participants evaluated the exterior offers. In other words, attentional contrast was created on attribute 2. In the 6x6 condition, the attribute levels of attribute 1 and attribute 2 were on average equally novel when participants evaluated the exterior offers. In sum, attentional contrast was created on the four-level attributes. If the number-of-levels effect depends only on the number of levels, the four-level attributes will have lower derived importance weights than the six-level attributes. If the number-of-levels effect depends on attentional contrast, then the four-level attributes will have higher derived importance weights. Predictions The attentional contrast hypothesis makes predictions for attributes one and two. With respect to attribute one, the attentional contrast hypothesis predicts that the mean difference score should be greater in the 4x6 condition than in the 6x4 condition or the 6x6 condition. With respect to attribute 2, the attentional contrast hypothesis predicts that the mean difference score should be greater in the 6x4 condition than in the 4x6 condition or the 6x6 condition. Given the results of experiments 1, 1A, and 2, it was anticipated that the attentional contrast effect would be obtained on attribute 1, but not on attribute 2. Results Table 7 presents the mean difference scores per attribute for each condition.

PAGE 47

37 Refrigerators Data of 87 participants were retained for analysis. The analysis for each attribute was performed separately. The first analysis was on attribute one (annual energy cost). The mean difference scores of the 4x6 condition (X = $119.89) and the 6x4 condition (X = $77.59) were significantly different (F(1, 84) = 7.52, p < .05). The mean difference scores of the 4x6 condition (X = $119.89) and the 6x6 condition (X = $76.75) were significantly different (F(1, 88) = 6.72, p < .05). The second analysis was on attribute two (capacity). The mean difference scores of the 6x4 condition (X = $110.09) and the 4x6 condition (X = $132.15) were not significantly different (F(1, 84) = 1.34, p > .05). The mean difference scores of the 6x4 condition (X = $110.09) and the 6x6 condition (X = $94.92) were not significantly different (F(1, 84) = 0.55, p > .05). LCD televisions Data of 96 participants were retained for analysis. The analysis for each attribute was performed separately. The first analysis was on attribute one (screen size). The mean difference scores of the 4x6 condition (X = $246.95) and the 6x4 condition (X = $201.55) were not significantly different (F(1, 93) = 1.07, p > .05). The mean difference scores of the 4x6 condition (X = $246.95) and the 6x6 condition (X = $228.70) were not significantly different (F(1, 93) = 0.15, p > .05). The second analysis was on attribute two (horizontal viewing angle). The mean difference scores of the 6x4 condition (X = $138.15) and the 4x6 condition (X = $129.30) were not significantly different (F(1, 93) = 0.16, p > .05). The mean difference scores of the 6x4 condition (X = $138.15) and the 6x6 condition (X = $133.70) were not significantly different (F(1, 93) = 0.05, p > .05).

PAGE 48

38 Digital cameras Data of 123 participants were retained for analysis. The analysis for each attribute was performed separately. The first analysis was on attribute one (imaging sensor resolution). The mean difference scores of the 4x6 condition (X = $175.06) and the 6x4 condition (X = $169.20) were not significantly different (F(1, 120) = 0.09, p > .05). The mean difference scores of the 4x6 condition (X = $175.06) and the 6x6 condition (X = $153.29) were significantly different (F(1, 120) = 1.17, p > .05). The second analysis was on attribute two (image storage capacity). The mean difference scores of the 6x4 condition (X = $157.11) and the 4x6 condition (X = $126.92) were not significantly different (F(1, 120) = 2.50, p > .05). The mean difference scores of the 6x4 condition (X = $157.11) and the 6x6 condition (X = $111.00) were significantly different (F(1, 120) = 5.54, p < .05). Discussion The results of experiment 3 point in the direction of an attentional contrast explanation. On the first attribute, the means are in the direction predicted by the attentional contrast account, though the results are significant in only one of three product categories (refrigerators). On attribute 2, the results of the refrigerators category differs from the predictions of both number-of-levels effect and attentional contrast. The means of the LCD televisions category are in the direction predicted by attentional contrast, though non-significant. The results for the digital cameras category are in the direction predicted by attentional contrast. Since this experiment used the same exterior levels as experiments 1 and 2, the partial lack of significance on attribute 2 may be attributed to the perceived lack of importance of attribute 2 (see experiment 1A).

PAGE 49

39 A Attribute 2 1 2 3 4 5 6 1 2 3 4 5 Attribute 1 6 B Attribute 2 1 2 3 4 5 6 1 2 3 4 5 Attribute 1 6 C Attribute 2 1 2 3 4 5 6 1 2 3 4 5 Attribute 1 6 Note.Exterior offers are colored in black, intermediate offers are shaded in grey. Figure 3. Design for experiment 3. A) 4x6 Condition. B) 6x4 Condition. C) 6x6 Condition.

PAGE 50

40 Table 6. Attribute levels from experiment 3 Attribute level Product category Attribute 1 2 3 4 5 6 Refrigerators annual energy cost ($n) 100 95 90 80 75 70 capacity (n cubic feet) 16 17 18 20 21 22 LCD televisions screen size (n inches) 20 22 24 28 30 32 horizontal viewing angle (n off-center) 45 50 55 65 70 75 Digital cameras imaging sensor resolution (n MegaPixels) 3.2 3.8 4.4 5.6 6.2 6.8 image storage capacity (up to n MB) 128 192 256 384 448 512 Note.The underlined attribute levels are the levels that are added when an attribute has 6 levels. Table 7. Mean difference scores and simple effects for experiment 3 Mean difference score F values Attribute 1 4x6 6x4 6x6 4x6 vs 6x4 4x6 vs 6x6 Annual energy cost (Refrigerators) 119.89 77.59 76.75 7.52 a 6.72 a Screen size (LCD televisions) 246.95 201.55 228.70 1.07 0.15 Imaging sensor resolution (Digital cameras) 175.06 169.20 153.29 0.09 1.17 Attribute 2 4x6 6x4 6x6 4x6 vs 6x4 6x4 vs 6x6 Capacity (Refrigerators) 132.15 110.09 94.92 1.34 0.55 Horizontal viewing angle (LCD televisions) 129.30 138.15 133.70 0.16 0.05 Image storage capacity (Digital cameras) 126.92 157.11 111.00 2.50 5.54 b a implies significance at 99% level of confidence b implies significance at 95% level of confidence

PAGE 51

CHAPTER 6 GENERAL DISCUSSION Experiment 1 provides evidence that the weighting effect as observed in published studies may not be a number-of-levels effect, but that it may be due to attentional contrast away from the less novel attribute level within each trial. Experiment 2 shows that attentional contrast can create a weighting effect when there is no difference in the number of levels. Experiment 3 shows that attentional contrast can overrule the number-of-levels effect, i.e., that it can result in greater derived importance weights on the attribute with the lower number of levels. In the three experiments, results were consistent on attribute 1, but did not reach significance on attribute 2. In experiment 1A, I showed that this lack of significance was due to the perceived relative unimportance of the second attribute. In experiment 1A, the distinction between attribute level importance and attribute importance is highlighted. Simonson (1993) shows that consumers often do not possess well-formed beliefs about attribute importance. Instead, they construct their preferences during judgment and choice (Payne, Bettman, and Johnson 1992). First, the experiments in this dissertation show that attribute importance is influenced by the attribute levels that are presented. Second, experiment 1A shows that this influence only occurs when the attribute is perceived to be important. While in experiments 1, 2 and 3 attentional contrast only created a weighting effect on attribute 1, increasing the accessibility of consumers product experiences in experiment 1A resulted in a weighting effect on attribute 2. In the number-of-levels effect literature, price was often used as one of the attributes. Since this 41

PAGE 52

42 dissertation uses price as a dependent variable, one of the product attributes that are usually perceived as being inherently important was not eligible, which may have reduced the overall perceived importance of the attributes used in this dissertation compared to attributes used in earlier studies. This effect may be enhanced by the fact that the choice of attributes was restricted to attributes that have numerical values and for which evaluations are known to be monotonically distributed. The attentional contrast explanation is consistent with all the results that have been published in the number-of-levels effect literature. First, all empirical studies (except Johnson 1992) used a sequential evaluation procedure. Attentional contrast can only operate when attribute levels are repeated and when this repetition occurs in the context of other attribute levels that have been more or less repeated. Second, attentional contrast is consistent with the observation that the number-of-levels effect was not produced in the compositional measures that were added before the sequential evaluation procedure (Wittink et al. 1992a) and that only minimal effects were produced when the compositional measures were added after the sequential evaluation procedure (Steenkamp and Wittink 1994; Verlegh et al. 2002). Finally, the reason why attentional contrast was not detected in earlier studies is that they invariably compared conditions in which the number of levels of more than one attribute was compared at once (see Table 1). While the main contribution of this dissertation is the proposition of a new account for a repeatedly published effect, two additional contributions can be mentioned. First, until now, the process of attentional contrast had only been documented in a category learning task. Based on the shared characteristic of sequential trials between the

PAGE 53

43 classification learning and conjoint or multiattribute tasks, I showed that attentional contrast also occurs in an evaluation (versus categorization) task. Second, in previous studies (Kersten et al. 1998), attentional contrast was documented using pictorial values. I generalized its occurrence to the use of metric values. While I conclude that attentional contrast is at least partly responsible for the weighting effect that was observed in studies examining the number-of-levels effect, the data do not provide information as to how attentional contrast creates a weighting difference. One possibility is that participants attend longer to the relatively novel attribute levels, and that attribute levels that are attended to for longer durations receive more weight. However, it is possible that it is not mere attention which results in the weighting difference. It is likely that participants elaborate more on the relatively novel attribute levels. For example, when asked to evaluate an LCD television with a horizontal viewing angle which is 60 degrees off-center, participants may imagine themselves in their apartments and try to assess mentally from which area they would have a decent view of the screen. Aside from the question how attentional contrast works, one may ask when it works. First, the impact of the attentional contrast effect is likely to be dependent on the assumptions with regards to the total pool of attention. In the experiments presented in this dissertation, students evaluated the trials at their own pace. In other words, the pool of attention was variable. This means that increased attention to one attribute is not necessarily accompanied by decreases in attention to a second attribute. In contrast, in a model in which the total pool of attention is fixed, increased attention to a first attribute results in decreases in attention to the second attribute. A model with a fixed pool of

PAGE 54

44 attention is more closely operationalized by a computer-paced experiment. Using a computer-paced experiment may enhance the chances that attentional contrast creates a weighting effect on both attributes. Second, as shown in experiment 1A, the occurrence of the weighting effect depends on the importance of the attribute. The prior level of attention to an attribute may influence the impact of changes in attention towards the attribute levels. Attributes that normally receive a low level of attention may show small reductions in attention as a result of attentional contrast away from the attribute levels. Analogously, attributes that tend to receive high levels of attention may show small increases in attention as a result of attentional contrast towards the attribute levels. Therefore, attentional contrast is likely to have the greatest impact when attributes are moderately important. Finally, a main focus of many number-of-level effect studies was to find a way to eliminate the bias. Based on the account offered in this dissertation, one should try to keep the relative novelty of attribute levels in subsequent trials constant. Since this is impossible when attributes are defined on unequal numbers of levels, one may prefer to use a data collection method that does not involve subsequent trials. For example, the use of a trade-off matrix may be preferred in case of an unequal number of levels.

PAGE 55

CHAPTER 7 LIMITATIONS AND FUTURE RESEARCH While this dissertation takes an important first step towards proposing a new account for the number-of-levels effect, there are several issues that remain unexamined and that could be extended to potentially useful future research. There are several characteristics of the experiments that limit the generalizability of the findings. First, only one dependent measure was used in all studies. Participants stated their reservation price for product offers. While the results of these studies supported an attentional contrast explanation, the number-of-levels effect has been identified using a variety of dependent measures. To generalize the results of this dissertation, it would be useful to examine whether attentional contrast can also explain the results of studies that use other dependent measures such as paired comparisons and rating scales. Second, no compositional measures were collected in this dissertation. While there is considerable evidence from the literature that the number-of-levels effect is not produced on compositional measures, it could be useful to gather additional evidence that shows that the number-of-levels effect is inherent to the sequential evaluation procedure. Moreover, it could be useful to examine if the weighting effect caused by attentional contrast carries over to a posteriori collected compositional measures. A final extension of this dissertation has the potential of being useful. This research shows that attentional contrast does occur not only during category learning tasks, but also during multiattribute judgment tasks. The role of attentional contrast during other 45

PAGE 56

46 phenomena that have been identified in the multiattribute judgment literature (e.g., frequency effect) would be worth examining.

PAGE 57

LIST OF REFERENCES Currim, Imran S., Charles B. Weinberg, and Dick R. Wittink (1981), The Design of Subscription Programs for a Performing Arts Series, Journal of Consumer Research, 8 (June), 67-75. Goldstone, Robert L. and Mark Steyvers (2001), The Sensitization and Differentiation of Dimensions During Category Learning, Journal of Experimental Psychology: General, 130 (March), 116-139. Green, Paul E. and V. Srinivasan (1978), Conjoint Analysis in Consumer Behavior: Issues and Outlook, Journal of Consumer Research, 5 (September), 103-123. Green, Paul E. and V. Srinivasan (1990), Conjoint Analysis in Marketing Research: New Developments and Directions, Journal of Marketing, 54 (October), 3-19. Huffman, Cynthia (1997), Elaboration on Experience: Effects on Attribute Importance, Psychology & Marketing, 14 (August), 451-474. Johnson, Richard M. (1992), Comment on Attribute Level Effects, Second Annual Advanced Research Techniques Forum Proceedings, Rene Mora (ed.), Chicago: American Marketing Association, 62-64. Kersten, Alan W., Robert L. Goldstone, and Alexandra Schaffert (1998), Two Competing Attentional Mechanisms in Category Learning, Journal of Experimental Psychology: Learning, Memory, and Cognition, 24 (6), 1437-1458. Nosofsky, Robert M. (1986), Attention, Similarity, and the Identification-Categorization Relationship, Journal of Experimental Psychology: General, 115 (March), 39-57. Nosofsky, Robert M. (1991), Tests of an Exemplar Model for Relating Perceptual Classification and Recognition Memory, Journal of Experimental Psychology: Human Perception and Performance, 17 (February), 3-27. Orme, Bryan K. (1998), Reducing the Number-of-Attribute-Levels Effect in ACA with Optimal Weighting, 1998 Sawtooth Software Conference Proceedings, Ketchum, ID: Sawtooth Software. Payne, John W., James R. Bettman, and Eric J. Johnson (1992), Behavioral Decision Research: A Constructive Processing Perspective, Annual Review of Psychology, 43, 87-131. 47

PAGE 58

48 Simonson, Itamar (1993), Get Closer to Your Customers by Understanding How They Make Choices, California Management Review, 35 (4), 68-84. Srinivasan, V. and Allan D. Shocker (1973), Estimating the Weights for Multiple Attributes in a Composite Criterion Using Pairwise Judgments, Psychometrika, 38 (December), 473-493. Steenkamp, Jan-Benedict E.M. and Dick R. Wittink (1994), The Metric Quality of Full-Profile Judgments and the Number-of-Attribute-Levels Effect in Conjoint Analysis, International Journal of Research in Marketing, 11 (June), 275-286. Sutherland, N.S. and N.J. Mackintosh (1971). Mechanisms of Animal Discrimination Learning. New York: Academic Press. Verlegh, Peeter W.J., Hendrik N.J. Schifferstein, and Dick R. Wittink (2002), Range and Number-of-Levels Effects in Derived and Stated Measures of Attribute Importance, Marketing Letters, 13 (February), 41-52. Wittink, Dick R., Joel Huber, John A. Fiedler, and Richard L. Miller (1992a), Attribute Level Effects in Conjoint Revisited: ACA versus Full Profile, Second Annual Advanced Research Techniques Forum Proceedings, Rene Mora (ed.), Chicago: American Marketing Association, 51-61. Wittink, Dick, Joel Huber, Peter Zandan, and Richard M. Johnson (1992b), The Number of Levels Effect in Conjoint: Where Does It Come From and Can It Be Eliminated? 1992 Sawtooth Software Conference Proceedings, Ketchum, ID: Sawtooth Software. Wittink, Dick R., Lakshman Krishnamurthi, and Julia B. Nutter (1982), Comparing Derived Importance Weights Across Attributes, Journal of Consumer Research, 8 (March), 471-474. Wittink, Dick R., Lakshman Krishnamurthi, and David J. Reibstein (1989), The Effects of Differences in the Number of Attribute Levels on Conjoint Results, Marketing Letters, 1, 113-123.

PAGE 59

BIOGRAPHICAL SKETCH Els De Wilde was born in Gent, Belgium, on July 18, 1975. After attending high school at the Onze-Lieve-Vrouw-Presentatie in Sint-Niklaas, Belgium, she attended the Katholieke Universiteit Leuven in Louvain, Belgium. At this university, she obtained the degrees of Kandidaat in de Psychologie (1996), graduating with distinction, and Licentiaat in de Psychologie (1999), for which she graduated with great distinction. As part of her studies for the Licentiaat degree, she spent five months at the Universidad Complutense de Madrid, Spain. She wrote a thesis on categorization under the guidance of Professor Gert Storms. She spent seven months on a research internship, split between the Laboratory of Experimental Psychology at the Catholic University of Louvain, where she worked with Karl Verfaillie, and the marketing research company TNS Dimarso in Brussels, Belgium, where she was a trainee at the Fast Moving Consumer Goods unit. These research experiences were instrumental in shaping her interest in pursuing doctoral studies in the area of marketing, which she started at the University of Florida in August 1999. During her four years in the Ph.D. program at the Department of Marketing, she collaborated with Professors Chris Janiszewski and Alan Cooke. In 2001, Els taught an undergraduate course in consumer behavior at the University of Florida. Since Summer 2003, Els has been working as an Assistant Professor of Marketing at HEC Montral, Canada. 49