# Evaluating Marketing Communication Strategies: Attribute Importance Versus Transfer Discrepancy

ABSTRACT - There are three parts to this paper. Part one demonstrates that a relatively unknown measure of attribute importance (the difference between two conditional intentions) yield importance weights which are equivalent to attribute importance weights derived from conjoint analysis. Part two compares a simplistic communication model implied by attribute importance (i.e., communicate only those attributes which are important) with the recently developed transfer discrepancy communication model. Part three develops more realistic dynamic communication model based upon attribute importance weights. This paper concludes that the transfer discrepancy model is the superior model.

##### Citation:

*
Jeffrey E. Danes, James G. Johnston, Philippe Cattin, and John E. Hunter (1980) ,"Evaluating Marketing Communication Strategies: Attribute Importance Versus Transfer Discrepancy", in NA - Advances in Consumer Research Volume 07, eds. Jerry C. Olson, Ann Abor, MI : Association for Consumer Research, Pages: 561-565.
*

There are three parts to this paper. Part one demonstrates that a relatively unknown measure of attribute importance (the difference between two conditional intentions) yield importance weights which are equivalent to attribute importance weights derived from conjoint analysis. Part two compares a simplistic communication model implied by attribute importance (i.e., communicate only those attributes which are important) with the recently developed transfer discrepancy communication model. Part three develops more realistic dynamic communication model based upon attribute importance weights. This paper concludes that the transfer discrepancy model is the superior model.

INTRODUCTION

A number of market researchers have asserted that measures of attribute importance could be used for the selection of messages for marketing communication strategy (e.g., Green & Wind 1975 p. 112; Rao and Craig 1975 p. 327). Recently, however, Danes and Hunter (1979) developed a multimessage communication model designed specifically for selecting messages which change purchase intentions in a desired direction. The central task of the present paper is to compare two communication models based upon attribute importance weights with the transfer discrepancy communication model presented by Danes and Hunter (1979).

The first part of this paper demonstrates that the difference between two conditional intentions yield attribute importance weights which are equivalent to attribute importance weights derived from conjoint analysis. The second part of this paper compares a simplistic communication model implied by attribute importance weights with the recently developed transfer discrepancy model developed by Danes and Hunter (1979). The third part of this paper derives a more complex and realistic, dynamic attribute importance communication model; it then shows the transfer discrepancy model to be the superior predictor of change in purchase intention.

Part I: Validating Differential Conditional Intention

Differential conditional intention is a relatively new measure for estimating attribute importance. Derived from research by Wyer and Goldberg (1970) and Wyer (1974; 1975). Jaccard and King (1977) have termed it "psychological relevance" and used it to estimate the importance of attributes related to tobacco smoking and to voting intentions. This attribute importance measure is simply the difference between two conditional intentions:

Attribute Importance = P[i/b]-P[i/b']

Where, P[i/b] is the consumer's purchase intention given that a object is associated with a specific attribute; P[i/b'] is the consumer's purchase intention given that an object (brand/product) is __not__ associated with a specific attribute. Below are two example questions used to estimate P[i/b] and P[i/b']:

Given that Myrtle Beach South Carolina offered GOLF, would you vacation at Myrtle Beach?

Likely _:_:_:_:_:_:_:_:_:_:_:_:_:_ Unlikely

Improbable _:_:_:_:_:_:_:_:_:_:_:_:_:_ Probable

I Would _:_:_:_:_:_:_:_:_:_:_:_:_:_ I Would Not

Given that Myrtle Beach South Carolina DID NOT offer GOLF, would you vacation at Myrtle Beach?

Likely _:_:_:_:_:_:_:_:_:_:_:_:_:_ Unlikely

Improbable _:_:_:_:_:_:_:_:_:_:_:_:_:_ Probable

I Would _:_:_:_:_:_:_:_:_:_:_:_:_:_ I Would Not

Consumers simply check the corresponding semantic differential type-items; each item is scored from zero (unlikely, improbable), to ten (likely, probable), averaged and then divided by ten to convert the resulting scale to the units of subjective probability (see Wyer 1974; Danes & Hunter 1979). Danes and Hunter (1979) have found the above three bipolar items to be correlationally equivalent: The reliability of the scales exceed .90.

Since conjoint analysis is one of the most popular methods used for the estimation of attribute importance in marketing research, it was chosen as the criterion measure for validation. Given two levels of one attribute (x) where one level represents the presence of the attribute and the other it's negation (i.e., lack of presence), then attribute importance as defined by conjoint analysis is:

Attribute Importance = U(x) - U(x')

Where U(x) is the utility value for the presence of the attribute and U(x') is the utility value for the absence of the attribute.

Currently, the term "conjoint analysis" refers to a family of decompositional methods which estimate the structure of consumer's "preferences" given overall judgments to a set of alternatives, prespecified in terms of attribute levels. Following Green and Srinivasan (1978, p. 105) the conjoint procedure, in this study, may be described as the following concatenation of steps: (1) part-worth function "preference" model, (2) tradeoff questions for data collection, (3) verbal (written) description of attribute "bundles," (4) semantic differential-type rating scales with purchase intention as the dependent measure, these scales followed the forced choice tradeoff questions, and (5) ordinary least squares regression to estimate the part-worth, utility weights.

Each respondent was given 32 paired descriptions of Myrtle Beach (the object); for each of the paired descriptions, the respondent was asked to choose the most preferred description. Based upon the description they chose, the respondent was then asked to rate the degree to which they were inclined to vacation at Myrtle Beach. Two example "rated" paired comparisons appear below:

Sunny & Warm | | Cloudy & Cool | |

No Golf | | Golf | |

Given your choice, would you vacation at Myrtle Beach?

Likely _:_:_:_:_:_:_:_:_:_:_:_:_:_ Unlikely

Improbable _:_:_:_:_:_:_:_:_:_:_:_:_:_ Probable

I Would _:_:_:_:_:_:_:_:_:_:_:_:_:_ I Would Not

Sunny & Warm | | Cloudy & Cool | |

Golf | | No Golf | |

Given your choice, would you vacation at Myrtle Beach?

Likely _:_:_:_:_:_:_:_:_:_:_:_:_:_ Unlikely

Improbable _:_:_:_:_:_:_:_:_:_:_:_:_:_ Probable

I Would _:_:_:_:_:_:_:_:_:_:_:_:_:_ I Would Not

Each of the 32 "rated" paired comparisons produced one observation for the regression. The attribute levels in the paired comparisons correspond to a dummy code of +1 or -1; where the sign simply coded the level of attribute. The attributes not present received a zero. The rated intention to vacation at Myrtle Beach was then multiplied by the sign (1 or -1) of the attribute level; this was done to keep track of the level chosen. For the first example question given above, if the respondent chose "Cloudy & Cool and Golf" over "Sunny & Warm and No Golf," the rated intention to vacation was multiplied by -1 (see Figure 1). All ratings were averaged and scaled so that the highest possible score equaled 1.0 and the lowest equaled .00; i.e., trans- formed into units of subjective probability.

ORTHOGONAL REGRESSOR MATRIX USED FOR THE ESTIMATION OF PART-WORTH FUNCTIONS

The eight two-level attributes used in the study were:

X1 Inexpensive/Expensive Lodging

X2 Water Skiing & Sailing/No Water Skiing or Sailing

X3 Disco Dancing/No Disco Dancing

X4 Golf/No Golf

X5 Sunny & Warm/Cloudy & Cool

X6 Clean Beaches/Dirty Beaches

X7 Singles Bars/No Singles Bars

X8 Peace & Quiet/No Peace or Quiet

Data were collected from 49 young adults attending a large eastern university; after expressing willingness to participate each respondent was rewarded for his/her time and effort. The aggregate importance weights derived from conjoint analysis and from the conditional probability differences appear in Table 1. The aggregate differential conditional intention attribute importance estimates reported in Table 1 represent individual values averaged over the 49 respondents. The conjoint attribute importance estimates are unstandardized, part-worth functions (regression coefficients) obtained from ratings averaged over the 49 respondents.

ATTRIBUTE IMPORTANCE ESTIMATES DERIVED FROM CONDITIONAL INTENTIONS AND CONJOINT ANALYSIS

The rank-order correlation between the two aggregate measures of attribute importance is .952; the product-moment correlation is .968. The stability of the product-moment correlation was assessed using Tukey's Jackknife upon Fisher's Z transformations (Mosteller & Tukey 1969; Wainer & Thissen 1975). The Jackknife Z equals 2.049 which, when translated back to correlation, equals .956. The standard error for the Jackknifed Z equals .360; this produces a 95% confidence interval (with df = 7; critical t = 1.895) of .857 to .959. The results demonstrate that conditional probability differences yield attribute importance weights which are equivalent to attribute importance weights derived from conjoint analysis.

PART II: COMPARISON OF SIMPLE ATTRIBUTE IMPORTANCE AND TRANSFER DISCREPANCY MODELS

One of the major goals of marketing communication is to encourage consumers to purchase (use) certain products (services). And, one of the major roles of the marketing/advertising manager is to guide the selection of communication content. The successful selection of persuasive message content (i.e., words, symbols, signs, etc) is of critical importance to the financial success of communication efforts. The problem of selecting persuasive message content for communication is a problem encountered by many advertising/marketing managers. In essence, the problem is this: Given a number of product/ service attributes (physical or perceptual), which or what set of these attributes ought to be communicated in a promotional campaign? Which attributes ought to be associated with the object (brand/product) so that potential consumers will be encouraged to purchase/use the product/service?

A Simple Attribute Importance Communication Model

A number of market researchers have asserted that attribute importance weights could be used for the design of marketing communication strategy (e.g., Green & Wind 1975; Rao & Craig 1975). Specific procedures or guidelines for this purpose are noticeably absent in each of these articles. The implication, however, is one of simply __emphasizing__ the __important attributes__ within a given marketing communication. To translate this implication into symbolic form, let A_{j} denote the relative attribute importance of the jth attribute (j = 1,2,..., n). Given the traditional "additive" model, we may write the following difference equation in which AP[i] represents a change in purchase intention and n is the number of attributes used in a marketing communication:

The term "object" is used in reference to a brand name, product, or service; B is a scale dependent parameter which may be estimated using ordinary least squares regression. If we estimate A_{j} as the difference between two conditional intentions, we may write equation (1) as:

The Transfer Discrepancy Communication Model

Danes and Hunter (1979) have offered a different strategy for forecasting changes in purchase intentions as a function of message content, a theory of transfer discrepancy. In the transfer discrepancy model, there are two separate discrepancy functions. First, there is message-belief discrepancy (d), i.e., the distance between the belief value communicated in the message (P[m]) and the pre-message belief held by the receiver P[b]. Second, there is "cognitive" discrepancy, i.e., the distance between the conditional intention P[i/b] or P[i/b'] and the marginal intention P[i]. Before communication, cognitive discrepancy is presumably consistent with belief. However, after the reception of a belief-changing message, the new belief and the prior cognitive discrepancy are no longer consistent. Cognitive consistency is restored upon the change in P[i]. With positive belief change, change is in the direction of P[i/b]; with negative belief change, change is in the direction of P[i/b'].

Where m is a "power" parameter which depends upon communication variables such as source credibility, evidence, etc. i.e., the "persuasiveness" of the communication. P[m] is 1.0 if the communication associates an attribute with an object and P[m] is .O0 if the message disassociates an attribute with an object.

If we denote the difference between P[m] and F[b] as d, and denote the difference between P[i/b] (or P[i/b']) and P[i] as R, equation 3 may be expressed as:

DP[i] = a d R (4)

Given the traditional "additive" model, we may write an equation which predicts changes in purchase intention as a function of more than one message:

We assume, however, that the "power" of communication uniformly applies to each of the messages within the communication, i.e.:

a_{1} = a_{2} = ... = a_{j} = a

Thus, the "power" parameter a is constant for each of the belief-attacking (associative/disassociative) messages within the communication. Hence, equation 5 may be expressed as:

We are now left with the product of two constants: the scale dependent parameter 8 and the "power" parameter a. Since the product of two constants is itself just another constant, the transfer discrepancy model may be written in the following form:

The basic differences between the attribute importance and transfer discrepancy models is that the attribute importance model is primarily a computational procedure which does not provide an explicit theoretical link between communication and change in purchase intention. It simply summarizes a communication strategy implied by Green and Wind (1975) and Rao and Craig (1975). The transfer discrepancy model, derived from Wyer's (1974) probabilistic theory of cognitive consistency (Danes & Hunter, 1979) is a dynamic model which explicitly links communication to change in purchase intention; hence, it is expected to he the superior predictor of change in purchase intention.

Empirical Evaluation

To compare the simple attribute importance with the transfer discrepancy communication model, the data bank used in the Danes and Hunter (1979) study was employed. In that study 108 individuals were given a pre-test in which purchase intention, belief and differential conditional intentions were measured. The 108 individuals were randomly divided into two groups. One group (n = 63) received and "associative" communication in which the Fiat 131 was associated with three attributes: affordable, roomy, and safe. The other group (n = 45) received a "disassociative" communication in which the Fiat 131 was disassociated with five attributes: economy, exterior paint quality, four-doors, radio, and rear-window defroster. Following communication of the experimental messages, post-test data was collected. The scales for the post-test were identical to those of the pre-test. A detailed description of these scales and of the experimental procedures used for data collection appear in Danes and Hunter (1979).

To assess the degree to which each model predicts change in purchase intention, three separate regression analyses were performed on each of the two experimental groups. This analysis consisted of estimating the parameters of the three equations presented below.

The first equation is the simple attribute importance model with attribute importance defined as the difference two conditional intentions as defined in equation 2. The second equation is the transfer discrepancy model; the third equation is a combination of the two; hence, the coefficients are partial slopes.

The estimated parameters and related information for the first two equations appear in Table 2.

SLOPES AND BETA WEIGHTS FOR THE TRANSFER DISCREPANCY AND SIMPLE ATTRIBUTE IMPORTANCE MODELS

The slope t for the transfer discrepancy model is .336 for the associative messages and .135 for the disassociative messages; both of these slopes are significant. The standardized slopes (beta weights) equal .532 for the associative messages and .548 for the disassociative messages. The beta weights indicate that as Sd_{j}R_{j} increases (or decreases) one standard deviation, we can expect approximately a one-half standard deviation increase (or decrease) in purchase intention change. The slope b" SIZE="2 for the attribute importance model is -.013 (beta = -.037) for the associative messages and is .034 (beta = .221)for the disassociative messages. Both of these coefficients are not significant.

The third equation given above, provides a simultaneous comparison of the degree to which each of the two models predict change in purchase intention. The resulting coefficients reported in Table 3 are partial slopes; they estimate the degree to which one model predicts change in intention change when the other is "held constant."

For the associative messages, the partial slope for the transfer discrepancy model is .387 (partial beta = .613) which is significant. The partial slope for the attribute importance model is -.084 (partial beta -.242) which is significant; however, it is in the wrong direction.

For the disassociative messages, the partial slope for the transfer discrepancy model is .130 (partial beta = .528) which is significant. The partial slope for the attribute importance model is only .022 (partial beta = .145) and is not significant.

The results of these analyses indicate the transfer discrepancy model to be the superior predictor of change in purchase intention. The simplistic attribute importance model did not fare well as a predictor of change in purchase intention. The problem with the simple attribute importance model is that it is strictly a static, computational procedure lacking an explicit theoretical connection between communication and change in purchase intention. Could a dynamic attribute importance model work better?

PART III: COMPARISON OF DYNAMIC ATTRIBUTE IMPORTANCE AND TRANSFER DISCREPANCY MODELS

The failure of the simple attribute importance model shows that a straightforward selection of important attributes does not yield messages which produce change in purchase intention. One criticism of this model is that it does not take into account prior belief, i.e., the initial perceived association between the object (brand/product) and attribute. If we incorporate belief P[b], we get the classic preference model:

And, if we assume that purchase intention is proportional to preference, we get:

The resulting model is static and if it is used to select communication content, the resulting communications should serve only to __reinforce__ (not change) pre-existing purchase intentions. However, if we change belief DP[b] we obtain:

Subtracting equation 12 from equation 14 gives change in purchase intention as a function of belief change:

The degree to which belief changes, however, may be predicted from the message-belief discrepancy model (Danes, Hunter, & Woelfel, 1978; Danes & Hunter 1979)

DP[b] = a (P[m] - P[b] (16)

0 __<__ a __<__ 1

where a is the "power" parameter discussed above; P[m] is 1.0 if the message associates the object (brand/ product) with an attribute or P[m] is .00 if the message disassociates the object with an attribute. This model simply states that beliefs change in the direction of messages, and that the change obtained is proportional to the "power" or persuasiveness of communication, a.

Substituting the right-hand side of equation 16 for DP[b] in equation 15 gives.

Given that the "power" of communication uniformly applies to each of the messages within the communication, i.e.,

a_{1} = a_{2} = ... = a_{j} = a

equation 17 may be written as:

However, since the product of two constants is itself a constant, we may write ab= l.

Now, if we estimate attribute importance as a difference between two conditional intentions:

"_{j} = P[i/b]_{j} - P[i/b']_{j} (20)

we arrive at the following dynamic attribute importance model:

Although the dynamic attribute importance model is derived from different theoretical premises, this model is __identical__ to the "consistency relevance" model proposed and tested by Danes and Hunter (1979). In their research, equation 21 was compared with the transfer discrepancy model and the transfer discrepancy model was shown to be the superior predictor of change in purchase intention.

SUMMARY

We first demonstrated that attribute importances as measured by the difference between two conditional intentions yields weights equivalent to attribute importance weights derived from conjoint analysis. Using this new measure, we then compared a simple attribute importance communication model with the transfer discrepancy model; we found the transfer discrepancy model to be the superior predictor of change in purchase intention. We then derived a dynamic attribute importance model and illustrated that the dynamic attribute importance model is identical to the "consistency relevance" model derived by Danes and Hunter (1979). Of the three models: simple attribute importance, dynamic attribute importance, and transfer discrepancy, it was concluded that the transfer discrepancy model is the superior model.

REFERENCES

Danes, J. and Hunter, J. (1979), "On the Development of a Multimessage Communication Model," in D. Montgomery and D. Wittink (eds.), __Marketing Measurement and Analysis__, Marketing Science Institute, __In Press__.

Danes, J.; Hunter, J. and Woelfel, J. (1978) "Mass Communication and Belief Change: A Test of Three Mathematical Models," __Human Communication Research__, 4, 243-252.

Green, P. and Wind, Y. (1975), "New Way to Measure Consumers' Judgements," __Harvard Business Review__, 53, 107-117.

Green, P. and Srinvisan, V. (1978), "Conjoint Analysis in Consumer Research: Issues and Outlook," __Journal of Consumer Research__, 5, 317-328.

Jaccard, J. and King, G. (1977), "The Relation Between Behavioral Intentions and Beliefs: A Probabilistic Model," __Human Communication Research__, 3, 326-334.

Mosteller, F. and Tukey, J. (1968), "Data Analysis: Including Statistics," in Lindzey and Aronson (eds.), __The Handbook of Social Psychology__ (2nd ed., Vol. II). Reading, Mass.: Addison-Wesley.

Rao, V. and Craig, C. (1975), "Applications of Conjoint Measurement to Communication Research," __Communication Research__, 2, 317-328.

Wainer, H. and Thissen, D. (1975), "When Jackknifing Fails (Or Does It?)," __Psychometrica__, 40, 113-114.

Wyer, R. and Goldberg, L. (1970), "A Probabilistic Analysis of the Relationship Among Beliefs and Attitudes,'' __Psychological Review__, 77, 100-120.

Wyer, R. (1974), __Cognitive Organization and Change: An Information Processing Approach__. Potomac, Md.: Erlbaum Press.

Wyer, R. (1975), "Functional Measurement Methodology Applied to a Subjective Probability Model of Cognitive Functioning." __Journal of Personality and Social Psychology__, 31, 94-100.

----------------------------------------

##### Authors

Jeffrey E. Danes, Virginia Polytechnic Institute and State University

James G. Johnston, MARCOM, Inc.

Philippe Cattin, University of Connecticut

John E. Hunter, Michigan State University

##### Volume

NA - Advances in Consumer Research Volume 07 | 1980

##### Share Proceeding

## Featured papers

See More#### Featured

### Promoting Pi Day: Consumer Inferences about Special Day-Themed Promotions

Daniel M. Zane, University of Miami, USA

Kelly Haws, Vanderbilt University, USA

Rebecca Walker Reczek, Ohio State University, USA

#### Featured

### Gaze Reflects Loss Aversion

Feng Sheng, University of Pennsylvania, USA

Arjun Ramakrishnan, University of Pennsylvania, USA

Darsol Seok, University of Pennsylvania, USA

Puti Cen, University of Pennsylvania, USA

Michael Platt, University of Pennsylvania, USA

#### Featured

### Narrow Algorithmic Searches Perpetuate Bias

Eugina (Wai Lun) Leung, Erasmus University Rotterdam, The Netherlands

Oleg Urminsky, University of Chicago, USA