Sie sind auf Seite 1von 28

0

Abstract

The paper presents a new approach to analyzing data from laddering interviews based on

the latent class model. An à priori enumeration of all potential ladders, stemming from

given sets of attributes, consequences, and values, forms the rows of a data matrix, while

individual respondents evoked ladders, form the columns. When this data matrix is sub-

mitted to a latent class analysis guidelines are provided for determining the number of

evoked chains and their content. Hence, the approach may be seen as quantitative, model

based analogue to the LadderMap program that is the usual standard tool employed in

doing means-end chain research.


1

A new approach to analyse data from laddering interviews

by

Carsten Stig Poulsen


The MAPP Centre
AArhus School of Business
Haslegaardsvej 10
DK-8210 AArhus V
Denmark
csp@asb.dk

Hans Jorn Juhl


Institute of Information Processing
AArhus School of Business
Fuglesangsallé 4 10
DK-8210 AArhus V
Denmark

Klaus G. Grunert
The MAPP Centre
AArhus School of Business
Haslegaardsvej 10
DK-8210 AArhus V
Denmark
2

A new approach to analyse data from laddering interviews

Introduction

In analysing what motivates consumers to choose one option instead of another in a

choice situation, the means-end approach has gained a lot of popularity in recent years.

Rooted in work of Kelly (1955) and developed as a tool for consumer behaviour research

mainly by Gutman (Gutman 1982, 1991; Reynolds and Gutman 1984, 1988; Reynolds,

Gutman and Fiedler, 1985), the basic tenet of this approach is that consumers become

motivated to purchase goods and services because of the self-relevant consequences the

purchase and consumption of these goods and services will have for them, or put another

way, how the purchase and use of these items will help them attain goals at various levels

of abstraction, including life values which motivate behaviour across a wide range of

situations (Schwartz, 1992). The means-end approach thus assumes that human behaviour

is goal-directed, and that the way in which knowledge about products and services is

subjectively linked to the attainment of goals provides motivation for the choice of prod-

ucts and services.

The means-end approach is intuitively appealing, and has won acceptance both in aca-

demic research and in practice. However, there has also been a growing frustration in

academic circles about the lack of progress both at the theoretical and at the methodo-

logical level. At the theoretical level, it has been criticized that the means-end approach is

underconceptualised, since it only provides a loose model of consumer cognitive struc-


3

ture, but no theory on how this structure is supposed to be related to behaviour, especially

in choice situations, or how it develops, based on external stimulation (Grunert, Sørensen

et al., 1995). At the methodological level, the means-end approach has been twinned with

the laddering technique, a qualitative interview technique subsequently subjected to some

crude quantification and resulting in the so-called hierarchical value maps. This technique

has been employed in numerous studies, but has also been subject to considerable criti-

cism due to its many ad-hoc solutions and lack of underlying statistical theory (Grunert

and Grunert, 1995). If the means-end approach is to strive in the future as a means of

academic inquiry, it is important that progress is made both in terms of theory and

method development.

A few promising attempts have recently been made. Brunsø, Scholderer et al. (2001)

have attempted to make theoretical progress by integrating the means-end approach into a

broader cognitive theory encompassing both top-down and bottom-up processes. Valette-

Florence and collaborators have experimented with new data analysis techniques

(Aurifeille and Valette-Florence, 1995; Roehrich and Valette-Florence, 1991; Valette-

Florence, 1998; Valette-Florence and Aurifeille, 1992; Valette-Florence and Rapacchi

1989, 1990, 1991). Ter Hofstede, Steenkamp and Wedel have worked with means to

make the technique amenable to measurement in surveys with large samples (ter

Hofstede, Audenaert et al., 1998; ter Hofstede, Steenkamp and Wedel, 1999). In the pres-

ent paper, we contribute to this stream by proposing a new technique to analyse laddering

data, which can replace the current ad-hoc solutions employed to derive hierarchical

value maps.
4

Collection of laddering data

In a widely cited article, Reynolds and Gutman (1988) describe the laddering method in

the following way:

“Laddering refers to an in-depth, one-to-one interviewing technique used to develop an understanding of

how consumers translate the attributes of products into meaningful associations with respect to self, fol-

lowing Means-End Theory (Gutman, 1982). Laddering involves a tailored interviewing format using pri-

marily a series of directed probes, typified by the “What is important to you?” question, with the express

goal of determining sets of linkages between the key perceptual elements across the range of attributes (A),

consequences (C) and values (V). These association networks, or ladders, referred to as perceptual orienta-

tions, represent combinations of elements that serve as the basis for distinguishing between and among

products in a given product class…”

Laddering can be executed in a number of different ways. One major distinction is ac-

cording to the way the attributes used in the hierarchical probes are generated . Various

methods of elicitation are in use, including triadic sorting, free elicitation, and reasoned

rank-ordering, all of which seem to have advantages and disadvantages (Bech-Larsen and

Nielsen, 1999). A second major distinction is between ‘hard’ and ‘soft’ laddering

(Botschen and Thelen, 1998; Grunert and Grunert, 1995). In ‘hard’ laddering, respon-

dents are forced through the A-C-V structure ladder by ladder, and the personal inter-

viewing can be substituted by paper-and-pencil methods (Walker and Olson, 1991) or the

use of a computer. ‘Soft’ laddering resembles more a depth interview, in which the un-

derlying means-end structure may not even be obvious to the respondent and is uncov-

ered later by the analyst in the coding process.


5

Laddering is a complex technique, and can be subject to various kinds of biases depend-

ing on how the interview is conducted. However, it seems that the technique is surpris-

ingly robust with regard to factors like the respondent’s verbal ability or situational de-

terminants in the interview (Bech-Larsen, Nielsen et al., 1996; Sørensen, Grunert et al.,

1996).

Analysis of laddering data

Laddering interviews result typically in a number of ladders per respondent, where each

ladder consists, in the respondent’s idiosyncratic terms, of a set of attributes, conse-

quences and values. These ladders are then subjected to a content analysis procedure, re-

placing the idiosyncratic answers with a number of summary codes. The purpose of the

summary codes is of course to obtain categories mentioned more than once across re-

spondents but also to avoid too much loss in information content. The transformation of

individual answers into summary codes is a problematic process discussed intensively in

Grunert and Grunert (1995).

The typical next step in the analysis of laddering data is to construct an implication ma-

trix in an attempt to quantify the results of the qualitative method (Reynolds and Gutman,

1988). The implication matrix is a square table with a number of rows/columns equal to

the sum of attributes, consequences and values. The table displays the number of times

each element leads to each other element. The relations may be either direct or indirect.

A direct relation exists if the relation is between adjacent elements in a ladder. If a conse-
6

quence mediates the relation between an attribute and a value then the relation between

the attribute and the value is indirect. Based upon the implication matrix a hierarchical

value map (HVM) of means-end chains is constructed, representing the cognitive struc-

ture at the aggregate level. The most typical approach to constructing a HVM is to try to

map all relations above several different cut-off levels while avoiding redundant relations

and favouring longer chains over shorter ones, i.e., if there is a chain A-C-V, there cannot

at the same time be a direct link A-V. The HVMs derived for several different cut-off

levels are then evaluated with the aim of selecting the most interpretable one. The choice

of cut-off level is quite arbitrary and hence quite problematic. If a high cut-off level is

chosen then the HVM only communicates self-evident results and if a low cut-off level is

chosen nearly no simplification is achieved compared to reading all the individual state-

ments. The LadderMap software, Reynolds and Gengler (1987), is commonly used to de-

rive implication matrices and construct HVMs.

Previously proposed alternative approaches

A number of alternative proposals for analysing laddering data have been proposed,

though none of them has found wide acceptance.

Valette-Florence and Rapacchi (1991) present a correspondence analysis and clustering

approach with the individual items (attributes, consequences and values) as units of

analysis. The main drawback of this approach is that links between items are not taken

into account. Firstly it may lead to clusters without any items at a basic means-end level
7

and to clusters with too many items. Secondly there is no information about how to link

the items included in a given cluster so the number of ladders represented by each cluster

cannot be determined. As a consequence of these problems Aurefeille and Valette-

Florence (1995) proposed to use the ladders as units of analysis instead. Their method

retains those chains among all possible chains in HVM which are closest, within a so-

called semantic space, to the individual ladders elicited from the consumers. The ap-

proach presented in this paper has ladders as units of analysis, too. Hence we will con-

centrate on a presentation of the ideas in Aurefeille and Valette-Florence (1995) and

compare our approach to theirs.

Aurefeille and Valette-Florence (1995) define a concept called relative dominance of a

means-end chain. The dominance of a chain is dependent upon two criteria – a frequency

criterion and a representativeness criterion. The frequency criterion refers to the number

of ladders from individual consumers that are represented by the chain. The representa-

tiveness criterion refers to the degree to which a chain accurately represents the underly-

ing set of ladders. The approach treats collected ladders as indicators of consumer’s la-

tent (non-observable) means-end structure. A multidimensional representation of col-

lected ladders produces a semantic space where the importance of all possible relations

between items, including those not specified in the interviews, can be evaluated through

the calculation of Euclidean distances between items. The representativeness of a chain is

then defined as the maximum distance between any of its items and any items of a set of

evoked ladders.

To sum up, the steps in their procedure are:


8

• Perform non-metric MDS on a similarity matrix between items with Lerman’s coeffi-

cient (Lerman,1988) as similarity measure and calculated across respondents

• Generate a list of potential chains

• Calculate the semantic difference between these chains and the ladders, using item

coordinates in the semantic space.

• Assign the ladders to a fixed number of potential chains (p) so that the semantic dif-

ferences between these chains and their assigned ladders are minimized

• Determine the dominant chains by testing successive values of p until an “optimum”

is found by making a kind of “scree-test” as known from factor analysis.

In our opinion the approach by Aurefeille and Valette-Florence (1995) have some prob-

lems. Firstly there is no formal model underlying the entire sequence of analyses and

hence it is hard to evaluate goodness.-of-fit and impossible to do inferential testing. Sec-

ondly the steps require several arbitrary choices by the researcher and hence it is difficult

for other researchers to interpret and replicate the results. In contrast, we propose a fairly

straigthforward approach based upon a well-known statistical model that can be esti-

mated and tested with standard software.

An alternative approach: Latent class analysis

Data structure

Let A1 … AK ; C1 …CL ; V1 …VM denote the summary codes representing K attributes, L

consequences and M values. Each (potential) ladder is a combination of an attribute

ak ∈ { A1 ,… , AK } , a consequence cl ∈ {C1 ,… , CL } , and a value vm ∈ {V1 ,… , VM } , and a


9

complete enumeration of the sample space of ladders is feasible. Normally the sum of at-

tributes, consequences and values is between 30 and 50 (Reynolds and Gutman, 1988).

This space forms the rows of the data matrix to be analysed. The columns of the matrix

consist of the respondents R1 ,… , RN . Finally, a typical cell ( , n ) contains an indicator

variable that takes on the value 1 if respondent Rn evokes ladder and 0 otherwise.

Table 1. Lay-out of the laddering data as preparation to latent class analysis.

In table 1 we assume ladders to be of the following 3 forms: type 1 consisting of ladders

with one attribute and one consequence, type 2 consisting of ladders with one conse-

quence and one value and finally type 3 contains complete ladders with one attribute, one

consequence and one value. If a ladder consists of two attributes ( A1 , A2 ) and a conse-

quence (C1 ) we treat it as two ladders ( A1 , C1 ) and ( A2 , C1 ) . Other special cases are han-

dled similarly. This is necessary in order to keep the number of possible ladders at a rea-

sonable level but has the drawback that ladders consisting of many concepts will be rep-

resented by many chains and thus receive more weight in the estimation.

Latent class analysis

Latent class analysis (LCA) has seen increased popularity in recent years as a model

based approach to clustering, cf. Wedel and Kamakura (2000). Typically, individuals are

grouped together based on a set of nominally or ordinally scaled measures. In terms of

data layout the individuals constitute the rows and the measures the columns in the data

matrix. As presented above in the present case the rows of the data matrix are made up of

a complete enumeration of all possible ladders, according to the restrictions imposed.


10

Hence, the focus of clustering is on the ladders rather than on the individuals who evoke

them. This is in accordance with the qualitative nature of laddering studies where the

number of respondents typically is fairly small (10-100), while the number of possible

ladders is large. As we shall indicate, however, it will still be possible to analyze respon-

dent heterogeneity within the modeling framework.

Applying the LCA to the data matrix will provide us with a set of classes or groups of

ladders which are homogeneous in the sense that ladders within a cluster have similar

probability of being evoked across respondents. In addition, one of the defining proper-

ties of LCA, conditional independence between (manifest) variables given the latent class

variable, implies that independence across respondents – an assumption that may be rea-

sonable due to the design of the study - is reestablished by the model. This means that the

joint probability for a ladder to be evoked by two or more respondents within a class can

be evaluated by multiplying the individual probabilities.

The number of classes of the model indicates how many groups of ladders that can be re-

liably detected. However, the determination of the proper number of classes in a LC

model is well-known to be problematic and heuristics based on some information criteria

(AIC, BIC, CAIC, etc.) are usually employed. Class sizes can be interpreted as the num-

ber of individual ladders that goes into the class.

Although the LCA by clustering works on the rows (ladders) of the data matrix, at the

same time it provides structural information on the columns (respondents). Within a class

the response profile gives the individual probabilities of evoking a member of that cluster
11

of ladders. Consequently, homogeneity in terms of respondents implies that these prob-

abilities are the same for all respondents. Such an assumption can (in principle) be tested

by imposing this restriction on the LC model.

Example: Wine cooler data.

To illustrate the approach we have reanalyzed the hypothetical wine cooler data from the

original article by Reynolds and Guttman (1988). The data involves 67 respondents and

the list of attributes, consequences, and values displayed in table 2.

Table 2. List of attributes, consequences, and values. Wine cooler example.

The data matrix contains 496 rows (potential ladders) and 67 columns (respondents). To

estimate the LC model we used Latent GOLD, Vermundt and Jagdison (2000). In order

to avoid local optima we estimated the model for a specific numer of classes with differ-

ent seeds and based on the AIC, BIC, and CAIC criteria and interpretability a solution

with two classes were selected, cf. figure 1.

Figure 1. Information theoretic criteria for selecting the number of classes.

To illustrate the output the first 10 respondents are presented in table 3. We note that the

first (and largest) class comprises the 85% of all the potential ladders that are not evoked

by any respondent. The second class consists of ladders that have more than a negligible

probability of being evoked by one or more respondents. It contains approx. 75 ladders.


12

Two tables are basic to the interpretation of the latent class structure. In table 3.a the class

sizes give the (unconditional) probability of observing a member of any given class,

while the class profile provides the conditional probability for each respondent to evoke a

ladder within a class. The ProbMeans in table 3.b describe the tendency for each respon-

dent to evoke ladders of each of the two types. Comparing these row-conditional prob-

abilities across respondents reveals some variation. This indicates that different ladders

may be evoked by different respondents, and these differences may subsequently be re-

lated to background characteristics.

Table 3. The estimated latent structure in extenso. Wine cooler example.

Since the potential ladders forming the rows of the data matrix is an enumeration of the

combinations of elements from the sets of attributes, consequences, and values, each lad-

der can be described by three corresponding variables A, C, and V. In order to pinpoint

the specific ladders belonging to each class, each ladder is assigned to the class s with the

highest ‘recruitment’ probability and then the cross-classification A×C×V is analyzed,

conditional on the class. For the sake of parsimony and stability of the estimated pa-

rameters within the class, we shall assume conditional independence between variables A

and V, given C. This assumption has received empirical support, cf. ter Hofstede, F., A.

Audenaert, et al. (1998). Using standard results from probability theory, the joint prob-

ability of evoking a specific chain within class s with levels ( klm ) on the three vari-

ables A, C , and V may be calculated as follows:


13

p Ak ClVm ( ∈ s ) = p Ak ( ∈ s ) ⋅ pCl | Ak ( ∈ s ) ⋅ pVm |Cl ( ∈ s )

Since class1 contains the potential ladders with high probability of not being evoked,

only class s =2 is of interest here. Table 4.a-c presents the results in tabular form.

Table 4a. The marginal frequency of the evoked attributes within latent ladder class
2, p Ak ( ∈ 2 ) .Wine cooler example.

Table 4b. The conditional probabilities of evoking the consequences, given the attributes
within latent ladder class 2, pCl | Ak ( ∈ 2 ) .Wine cooler example within latent ladder

Table 4c. The conditional probabilities of evoking the values, given the consequences
within latent ladder class 2, pVm |Cl ( ∈ 2 ) .Wine cooler example.

The way to read the tables should be fairly straightforward. The initial probabilities of

evoking an attribute are provided in table 3a, followed by the transition probability of

evoking a consequence (value), given a specific evoked attribute (consequence) in table

3b (c). Hence, this may be seen as a probabilistic formulation of the hierarchical value

map. E.g. the attribute ‘crisp’ is evoked with a (marginal) probability of 0,21, it is related

to consequences ‘refreshing’, ‘thirst-quenching’, ‘reward’, and ‘impress others’ with

equal probability of 0,25, while e.g. ‘refreshing’ is linked to values ‘accomplishment’

(0,40) and ‘belonging’ (0,60), ‘impress others’ is linked to ‘belonging’ (0,57) and ‘self-

esteem’ (0,43), etc. In this way the entire map can be detected. Still building on the as-

sumption of conditional independence between A(ttributes) and V(alues), given

C(onsequences), we can use standard results of probability theory to compute the prob-

ability of being evoked for each potential ladder within a class. Having done that a Top-

10 of potential ladders can be studied and represented in a hierarchical value map


14

(HVM), cf. table 5.

Table 5. Top-10 on the list of potential ladders within class 2. Wine cooler example.

Supplementary, in the spirit of LadderMap, we can draw a map, representing the Top-10

of potential ladders and indicate the strength of the associations by the transition prob-

abilities. Introducing an (arbitrary) cut-off value in terms of transition probabilities is

analogous to the cut-off in LadderMap, but note that this can be done after the analysis,

not as in LadderMap as part of the input.

Figure 2. Pictorial representation of the Top-10 of potential ladders of class 2. Wine


cooler example.

Discussion

We have presented a new approach for analysing data from laddering interviews, con-

sisting of the following 10 steps:

1. Define summary codes A = {a1 ,… aK } , C = {c1 ,… cL } , and V = {v1 ,… vM }

2. Define the sample space of ladders: A × C × V .

3. Impose any prior defined structure on the possible transitions

4. Perform a complete enumeration of the remaining elements (potential ladders).

5. Define the indicator matrix with potential ladders as rows and respondents as col-

umns, where 1 means that a specific ladder is evoked by a given respondent.

6. Perform LCA on the indicator matrix and determine the number of classes. One class

is likely to contain the ladders with negligible probability of being evoked.


15

7. Assign ladders to latent classes, using modal posterior probabilities.

8. Assuming conditional independence between A and V, given C, analyse within each

class X the frequency A | X , and the cross-classifications C | AX and V | CX , ig-

noring the class consisting of ladders with low probability of being evoked.

9. Compute the joint probability of being evoked within each class, p Ai C jVk ( ∈ s ) and

order the ladders accordingly to produce a list of Top-n.

10. Display the results graphically in a hierarchical value map, using the transitional

probabilities as measures of association.

The procedure seem to offer several advantages over previous efforts of analysing lad-

dering data. First, it is fairly simple to perform, given the results from the coding process.

Second, it clearly takes the ladders as units of analysis, but allows respondents to enter

via the latent class probabilities and concomitant variables. Third, as a statistical model it

is well-defined. Even if data are sparse as is usually the case with qualitative studies, the

fact that respondents are treated as variables rather than observations in the analysis elivi-

ates the problem considerably. Fourth, with today’s readily available software packages,

including several programs for doing latent class analysis, the computations involved pre-

sent no problem to the researcher. Fifth, the approach offer a quantification of the HVM,

based on probability theory rather than counts post hoc. Especially noteworthy is the fact,

that the Top-n list may include ladders not evoked in the data. Sixth, having the option of

detecting more than one class of (evoked) ladders is similar to the problem of “simple

structure” in factor analysis. The transition probabilities correspond to factor loadings,

and the interpretation is facilitated if these probabilities are close to 0 or 1. Increasing the
16

number of classes has the effect of simplifying the structure. (Note, however, that no ar-

bitrary ‘rotation’ criterion is needed here.)

The approach may easily be extended to include several stages in the means-end hierar-

chy. E.g. attributes may be classified as concrete and abstract, the consequences as func-

tional and psychosocial, and the values as instrumental and terminal, (Gutman, 1982).

Clearly, our approach share some conceptual similarities with Aurefeille and Valette-

Florence (1995). The idea of a latent means-end structure, reflected by the evoked (mani-

fest) ladders, is common to both procedures. The class size parameter in our model is

similar to the concept of ‘frequency’, and the posterior recriutment probabilities for each

evoked ladder, used to assign ladders to classes, may be seen as expressing ‘representa-

tiveness’. The main difference as we see it, is that these idea are implemented in a well-

defined, statistical model and given a precise meaning within that framework.

Still, additional work is needed. Apart from well-known problems inherent in the LCA

(determining the number of classes, evaluating ‘goodness-of-fit’ in sparse data sets) an

especially useful extension of the approach would model respondents and ladders in a

more integrated way by treating the ladders as repeated measurements of the respondents

and working simultaneously with two corresponding levels of latent class models. Ex-

actly how this may be done awaits further research.


17

References

Aurifeille, J.-M. and P. Valette-Florence.1995. “Determination of the dominant means-

end chains: constrained clustering approach.” International Journal of Research in Mar-

keting 12 (3): 267-278.

Bech-Larsen, T. and N. A. Nielsen. 1999. “A comparison of five elicitation techniques

for elicitation of attributes of low involvement products.” Journal of Economic Psychol-

ogy 20: 315-341.

Bech-Larsen, T., N. A. Nielsen, et al. 1996. Means-end chains for low involvement food

products - a study of Danish consumers' cognitions regarding different applications of

vegetable oil. MAPP working paper. Aarhus, The Aarhus School of Business.

Botschen, Günther and Eva Thelen. 1998. Hard versus soft laddering: Implications for

appropriate use.In: Balderjahn, Ingo, Claudia Mennicken, and Eric Vernette, editors, New

developments and approaches in consumer behaviour research. Stuttgart: Schäffer-

Poeschl Verlag.

Brunsø, K., Scholderer, J. and Grunert, K. G. 2001. Closing the gap between values and

behavior – A means-end theory of lifestyle. Paper to be presented at the 4th International

Research Seminar on Marketing Communications and Consumer Behavior, La Londe les

Maures, June 5-8, 2001.

Grunert, K. G. and S. C. Grunert. 1995. “Measuring subjective meaning structures by the

laddering method: Theoretical considerations and methodological problems.” Interna-

tional Journal of Research in Marketing 12: 209-225.

Grunert, K. G., E. Sørensen, et al. 1995. Analysing food choice from a means-end per-

spective. European advances in consumer research. F. Hansen. Provo, UT, Association

for Consumer Research. 2: 366-371.


18

Gutman, J. 1982. “A means-end chain model based on consumer categorization proc-

esses.” Journal of Marketing 46 (2): 60-72.

Gutman, J. 1991. “Exploring the nature of linkages between consequences and values.”

Journal of Business Research 22 (2): 143-149.

Kelly, G. A. 1955. The psychology of personal constructs. New York, NY, Norton.

Lerman,I.C. 1988. Comparing relational variables according to likelihood of the links

classification method. In Recent developments in clustering and data analysis. New York,

Academic Press.

Reynolds, T. J. and J. Gutman. 1984. Laddering: Extending the repertory grid methodol-

ogy to construct attribute-consequence-value hierarchies. In: R. E. Pitts and A. G. Wood-

side (eds.): Personal values and consumer psychology. Lexington, MA, Lexington Books.

Reynolds, T. J. and J. Gutman. 1988. “Laddering theory, methods, analysis, and inter-

pretation.” Journal of Advertising Research 18 (1): 11-31.

Reynolds, T. J., J. Gutman, J. A. Fiedler. 1985. Understanding consumers' cognitive

structures: the relationship of levels of abstraction to judgements of psychological dis-

tance and preference. In: L. F. Alwitt and A. A. Mitchell (eds.): Psychological processes

and advertising effects. Hillsdale, NJ, Lawrence Erlbaum Associates: 261-272.

Reynolds, T. J. and Gengler, C. E. 1987. Laddermap. Dallas, University of Texas and

Clarkson University, Potsdam.

Roehrich, Gilles and Pierre Valette-Florence. 1991. A weighted cluster-based analysis of

direct and indirect connections in means-end chains: An application to lingerie retail. Pa-

per presented at Workshop on values and lifestyle research in marketing, Brussels.

Schwartz, S. H. 1992. Universals in the content and structure of values: Theoretical ad-

vances and empirical tests in 20 countries. In: M. P. Zanna (ed.): Advances in experi-
19

mental social psychology. San Diego, CA, Academic Press. 25: 1-65.

Sørensen, E., K. G. Grunert, et al. 1996. The impact of product experience, product in-

volvement and verbal processing style on consumers' cognitive structure with regard to

fresh fish. MAPP working paper no. 42. Aarhus, The Aarhus School of Business.

ter Hofstede, F., A. Audenaert, et al. 1998. “An investigation into the association pattern

technique as a quantitative approach to measuring means-end chains.” International Jour-

nal of Research in Marketing 15 (1): 37-50.

ter Hofstede, F., J.-B. E. M. Steenkamp, and M. Wedel 1999. “International market seg-

mentation based on consumer-product relations.” Journal of Marketing Research 36 (Feb-

ruary): 1-17.

Valette-Florence, P. 1998. “A causal analysis of means-end hierarchies in a cross-cultural

context: Methodological refinements.” Journal of Business Research 42 (2): 161-166.

Valette-Florence, P. and J.-M. Aurifeille. 1992. A chain constrained clustering approach

in means-end analysis: An empirical illustration. Proceedings of the Annual Conference

of the European Marketing Academy. K. G. Grunert and D. Fuglede. Aarhus, The Aarhus

School of Business: 49-60.

Valette-Florence, P. and B. Rapacchi 1989. Value analysis by linear representation and

appropriate positioning: an extension of laddering methodology. In: . G. J. Avlonitis, N.

K. Papavasilou and A. G. Kouremenos (eds.): Marketing thought and practice in the

1990's: Proceedings of the Annual Conference of the European Marketing Academy Ath-

ens, The Athens School of Economics and Business. 2: 969-989.

Valette-Florence, P. and B. Rapacchi. 1990. A cross-cultural means-end chain analysis of

perfume purchases. In: N. E. Synodinos, C. E. Keown, K. G. Grunert and T. E. Muller.

(eds.): Proceedings of the third symposium on cross-cultural consumer and business


20

studies. Honolulu, HI, University of Hawaii: 161-172.

Valette-Florence, P. and B. Rapacchi (1991). “Improving means-end chain analysis using

graph theory and correspondence analysis.” Journal of Advertising Research 31 (1): 30-

45.

Vermundt, J. K. and J. Magdison. 2000. Latent Gold - User's Guide. Belmont, MA, USA:

Statistical Innovations Inc.

Walker, B. A. and J. C. Olson (1991). “Means-end chains: Connecting products with

self.” Journal of Business Research 22 (2): 111-119.

Wedel, Michel and Wagner A. Kamakura. 2000. Market segmentation: Conceptual and

Methodological Foundations. 2 ed: Kluwer Academic Publishers.


21

Table 1. Lay-out of the laddering data as preparation to latent class analysis.


Ladder no. A C V R1 R2 … Rn …. RN
1 A1 C1 - 1 0 … 0 … 0
2 A2 C1 - 0 1 … 1 … 0

K⋅L AK CL - 1 1 … 0 …. 1
K⋅L+1 - C1 V1 0 1 … 0 …. 0
K⋅L+2 - C1 V1 0 0 … 0 …. 1

K⋅L⋅+L⋅M - CL VM 1 0 … 1 …. 1
K⋅L⋅+L⋅M+1 A1 C1 V1 0 1 … 1 … 0
K⋅L⋅+L⋅M+2 A2 C1 V1 1 0 … 0 … 1
… 1
Ak Cl Vm 1 0 … 1 … 0

K⋅L⋅+L⋅M+K⋅L⋅M AK CL VM 0 1 … 1 … 0
22

Table 2. List of attributes, consequences, and values. Wine cooler example.


Attributes: carbonation bottle shape
crisp less alcohol
expensive smaller
label
Consequences: quality avoid negatives
filling avoid waste
refreshing reward
consume less sophisticated
thirst-quenching impress others
more feminine socialize
Values accomplishment belonging
family self-esteem
23

Figure 1. Information theoretic criteria for selecting the number of classes.

4800
4600
4400
4200
4000
3800
3600
3400
3200
3000
1 class 2 classes 3 classes

BIC AIC CAIC


24

Table 3. The estimated latent structure in extenso. Wine cooler example.

a. Profile b. ProbMeans

Col.% Class1 Class2 Row % Class1 Class2


Class Size 0,86 0,14 Overall Prob. 0,86 0,14
Respondent Respondent
R01 R01
Not evoke 1,00 0,87 Not evoke 0,88 0,12
Evoke 0,00 0,13 Evoke 0,00 1,00
R02 R02
Not evoke 1,00 0,97 Not evoke 0,87 0,13
Evoke 0,00 0,03 Evoke 0,00 1,00
R03 R03
Not evoke 1,00 0,90 Not evoke 0,87 0,13
Evoke 0,00 0,10 Evoke 0,21 0,79
R04 R04
Not evoke 1,00 1,00 Not evoke 0,86 0,14
Evoke 0,00 0,00 Evoke 1,00 0,00
R05 R05
Not evoke 1,00 1,00 Not evoke 0,86 0,14
Evoke 0,00 0,00 Evoke 1,00 0,00
R06 R06
Not evoke 1,00 0,82 Not evoke 0,88 0,12
Evoke 0,00 0,18 Evoke 0,00 1,00
R07 R07
Not evoke 1,00 0,91 Not evoke 0,87 0,13
Evoke 0,00 0,09 Evoke 0,00 1,00
R08 R08
Not evoke 1,00 0,96 Not evoke 0,87 0,13
Evoke 0,00 0,04 Evoke 0,00 1,00
R09 R09
Not evoke 1,00 0,87 Not evoke 0,88 0,12
Evoke 0,00 0,13 Evoke 0,00 1,00
R10 R10
Not evoke 1,00 0,97 Not evoke 0,87 0,13
Evoke 0,00 0,03 Evoke 0,00 1,00
25

Table 4a. The marginal frequency of the evoked attributes within latent ladder class 2,
p Ak ( ∈ 2 ) .Wine cooler example.
Latent class * Attributes Crosstabulation

Attributes
carbonation crisp expensive label bottle shape Total
Latent 2,00 Count 14 8 10 3 4 39
class % within Latent class 35,9% 20,5% 25,6% 7,7% 10,3% 100,0%

Table 4b. The conditional probabilities of evoking the consequences, given the attributes

within latent ladder class 2, pCl | Ak ( ∈ 2 ) .Wine cooler example.

Attributes * Consequences * Latent class Crosstabulation

Consequences
thirst-que more impress
Latent class quality refreshing nching feminine reward sophisticated others Total
2,00 Attributes carbonation Count 3 4 4 3 14
% within Attributes 21,4% 28,6% 28,6% 21,4% 100,0%
crisp Count 2 2 2 2 8
% within Attributes 25,0% 25,0% 25,0% 25,0% 100,0%
expensive Count 2 3 2 3 10
% within Attributes 20,0% 30,0% 20,0% 30,0% 100,0%
label Count 1 2 3
% within Attributes 33,3% 66,7% 100,0%
bottle shape Count 2 2 4
% within Attributes 50,0% 50,0% 100,0%

Table 4c. The conditional probabilities of evoking the values, given the consequences within

latent ladder class 2, pVm |Cl ( ∈ 2 ) .Wine cooler example.

Consequences * Values * Latent class Crosstabulation

Values
accompli
Latent class shment belonging self-esteem Total
2,00 Consequences quality Count 2 1 3
% within Consequences 66,7% 33,3% 100,0%
refreshing Count 2 3 5
% within Consequences 40,0% 60,0% 100,0%
thirst-quenching Count 2 3 2 7
% within Consequences 28,6% 42,9% 28,6% 100,0%
more feminine Count 1 2 3
% within Consequences 33,3% 66,7% 100,0%
reward Count 2 4 3 9
% within Consequences 22,2% 44,4% 33,3% 100,0%
sophisticated Count 1 4 5
% within Consequences 20,0% 80,0% 100,0%
impress others Count 4 3 7
% within Consequences 57,1% 42,9% 100,0%
26

Table 5. Top-10 on the list of potential ladders within class


2. Wine cooler example.
Ladder Prob. of being evoked
carbonation→impress others→self-esteem 0,0615
expensive→impress others→self-esteem 0,0614
carbonation→refreshing→belonging 0,0461
carbonation→reward→belonging 0,0456
carbonation→thirst-quenching→belonging 0,0440
carbonation→impress others→belonging 0,0439
expensive→impress others→belonging 0,0439
crisp→impress others→self-esteem* 0,0410
bottle shape→more feminine→self-esteem 0,0344
carbonation→reward→self-esteem 0,0342
*
not evoked in the data set.
27

Figure 2. Pictorial representation of the Top-10 of potential ladders of class 2. Wine


cooler example.
accomplishment belonging self-esteem

thirst- impress more


sophisticated quality refreshing reward
quenching others feminine

label carbonation crisp expensive bottle shape


8% 36% 21% 26% 10%

Das könnte Ihnen auch gefallen