Consumer Attitude Metrics for Guiding Marketing Mix Decisionscosts of each marketing instrument can determine the respective investment appeal of each of these actions. After a description

This article was downloaded by: [128.97.27.20] On: 24 January 2016, At: 10:35Publisher: Institute for Operations Research and the Management Sciences (INFORMS)INFORMS is located in Maryland, USA

Marketing Science

Publication details, including instructions for authors and subscription information:http://pubsonline.informs.org

Consumer Attitude Metrics for Guiding Marketing MixDecisionsDominique M. Hanssens, Koen H. Pauwels, Shuba Srinivasan, Marc Vanhuele, Gokhan Yildirim

To cite this article:Dominique M. Hanssens, Koen H. Pauwels, Shuba Srinivasan, Marc Vanhuele, Gokhan Yildirim (2014) Consumer Attitude Metricsfor Guiding Marketing Mix Decisions. Marketing Science 33(4):534-550. http://dx.doi.org/10.1287/mksc.2013.0841

Full terms and conditions of use: http://pubsonline.informs.org/page/terms-and-conditions

This article may be used only for the purposes of research, teaching, and/or private study. Commercial useor systematic downloading (by robots or other automatic processes) is prohibited without explicit Publisherapproval, unless otherwise noted. For more information, contact [email protected].

The Publisher does not warrant or guarantee the article’s accuracy, completeness, merchantability, fitnessfor a particular purpose, or non-infringement. Descriptions of, or references to, products or publications, orinclusion of an advertisement in this article, neither constitutes nor implies a guarantee, endorsement, orsupport of claims made of that product, publication, or service.

Copyright © 2014, INFORMS

Please scroll down for article—it is on subsequent pages

INFORMS is the largest professional society in the world for professionals in the fields of operations research, managementscience, and analytics.For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org

http://pubsonline.informs.org

http://dx.doi.org/10.1287/mksc.2013.0841

http://pubsonline.informs.org/page/terms-and-conditions

http://www.informs.org

Vol. 33, No. 4, July–August 2014, pp. 534–550ISSN 0732-2399 (print) � ISSN 1526-548X (online) http://dx.doi.org/10.1287/mksc.2013.0841

© 2014 INFORMS

Consumer Attitude Metrics for GuidingMarketing Mix Decisions

Dominique M. HanssensAnderson School of Management, University of California, Los Angeles, Los Angeles, California 90095,

[email protected]

Koen H. PauwelsÖzyegin University, Istanbul, 34794 Turkey, [email protected]

Shuba SrinivasanBoston University School of Management, Boston University, Boston, Massachusetts 02215, [email protected]

Marc VanhueleHEC Paris, 78351 Jouy-en-Josas, France, [email protected]

Gokhan YildirimManagement School, Lancaster University, Bailrigg, Lancaster LA1 4YX, United Kingdom,

[email protected]

Marketing managers often use consumer attitude metrics such as awareness, consideration, and preference asperformance indicators because they represent their brand’s health and are readily connected to marketing

activity. However, this does not mean that financially focused executives know how such metrics translateinto sales performance, which would allow them to make beneficial marketing mix decisions. We propose fourcriteria—potential, responsiveness, stickiness, and sales conversion—that determine the connection betweenmarketing actions, attitudinal metrics, and sales outcomes.

We test our approach with a rich data set of four-weekly marketing actions, attitude metrics, and salesfor several consumer brands in four categories over a seven-year period. The results quantify how marketingactions affect sales performance through their differential impact on attitudinal metrics, as captured by ourproposed criteria. We find that marketing–attitude and attitude–sales relationships are predominantly stable overtime but differ substantially across brands and product categories. We also establish that combining marketingand attitudinal metrics criteria improves the prediction of brand sales performance, often substantially so.Based on these insights, we provide specific recommendations on improving the marketing mix for differentbrands, and we validate them in a holdout sample. For managers and researchers alike, our criteria offer averifiable explanation for differences in marketing elasticities and an actionable connection between marketingand financial performance metrics.

Keywords : consumer attitude metrics; responsiveness; potential; stickiness; sales conversion; hierarchical linearmodel; cross-effects model; empirical generalizations; dynamic programming model; optimal marketingresource allocation

History : Received: August 14, 2012; accepted: December 10, 2013; Preyas Desai served as the editor-in-chiefand Donald Lehmann served as associate editor for this article. Published online in Articles in AdvanceMarch 7, 2014.

IntroductionBrand managers are urged to compete for the“hearts and minds” of consumers and often col-lect brand health indicators such as awareness, lik-ing, and consideration to this end. These indicatorshelp understand the state of mind of consumers andhow marketing affects it. More bottom-line-orientedmanagers, in contrast, typically assess marketingeffectiveness at the observable transaction level, withmeasures such as “advertising elasticity” and “returnon sales.” This practice may satisfy managers focused

on financial returns (including the chief financial offi-cer (CFO)), but it leaves the deeper reasons for mar-keting success or failure unexplored. Inasfar as thesereasons change, past sales impact of marketing maynot be the best predictor of its future sales impact.

Marketers work under the assumption that brandhealth indicators are predictive of later marketing andbottom-line performance but have little guidance onhow a better understanding of this connection can betranslated into improved decisions on the marketingmix. How actionable is it, for instance, to know that

534

Dow

nloa

ded

from

info

rms.

org

by [

128.

97.2

7.20

] on

24

Janu

ary

2016

, at 1

0:35

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.

Hanssens et al.: Consumer Attitude Metrics for Guiding Marketing Mix DecisionsMarketing Science 33(4), pp. 534–550, © 2014 INFORMS 535

brand consideration stands at 70% while brand lik-ing stands at 40%? Conventional wisdom (e.g., Kotlerand Keller 2012) suggests investing in the “weakestlink,” i.e., the metric with the most remaining poten-tial. However, brand liking may have hit its glassceiling at 40%, whereas momentum in considerationmay still be possible. In addition, consideration couldbe more responsive to marketing actions than brandliking, and any gains in brand liking may be short-lived because of fickle consumers or tough competi-tors; gains in consideration could be longer-lasting.As to the end result, consideration gains may con-vert into sales at a higher or lower rate than likinggains do. To complicate matters, marketing–attitudeand attitude–sales relationships may be generic to thecategory or specific to the brand, indicating competi-tive (dis)advantage. Finally, these relationships couldchange over time, obscuring their value and necessi-tating their dynamic evaluation to guide future mar-keting mix allocations.

In sum, it is no small task for financial and market-ing managers alike to use consumer attitude informa-tion to guide their marketing strategies and actions.Yet such guidance is important because managersare charged to allocate marketing resources that pro-vide noticeable and long-lasting improvements in theirbrands’ business performance. Our objective is there-fore to provide concrete directions on how the effec-tiveness of marketing mix actions, and therefore theallocation of resources, can be improved by exam-ining attitude metrics. More specifically, we proposetheory-based criteria on these metrics that identifyconditions under which they should be targets ofmarketing action. By applying these criteria, man-agers with access to the relevant information on thecosts of each marketing instrument can determinethe respective investment appeal of each of theseactions.

After a description of our contributions, we beginby proposing four theory-based criteria for the anal-ysis of attitude metrics and show how they can beoperationalized. In the empirical section, we describethe data set and demonstrate how the relevant param-eters can be estimated. Next, we apply our relevancecriteria, first for a diagnostic analysis and then for aforward-looking analysis using a dynamic program-ming (DP) analytical model. We conclude with a clas-sification of brands based on the role that attitudemetrics play in the connection between marketingactions and sales.

ContributionsOur research contributes to the marketing litera-ture in four ways (a comparative tabular sum-mary with previous work is provided in WebAppendix C1, available as supplemental material at

http://dx.doi.org/10.1287/mksc.2013.0841). First, itprovides an empirically testable framework on theconditions under which consumer attitude move-ments result in sales movements. Traditionally, mar-keting mix models almost exclusively focus on theresponse of sales to marketing expenditures in orderto derive normative implications for marketing bud-get setting. This is not sufficient for the brand man-ager interested in quantifying the linkage betweena firm’s marketing actions, consumer attitude met-rics, and the brand’s market performance, as con-ceptualized in the brand value chain (Keller andLehmann 2006). Srinivasan et al. (2010) introducedmind-set metrics into standard sales response modelsand demonstrated that these metrics indeed improvesales response models and are advance indicators oflater sales results. Stahl et al. (2012) made a similardemonstration with customer lifetime value.

Second, our research objective is normative, aim-ing to use the informational value of these mind-setvariables for improving marketing decision making.As such, our work is fundamentally different from themind-set metric article by Srinivasan et al. (2010).The objective of this article was to demonstrate thatthe explanatory power of a market response modelcan be increased by adding mind-set metrics. Vec-tor autoregressive (VAR) time-series model estimatedthere is descriptive and does not lend itself to norma-tive inferences such as deriving optimal advertisingspending for a brand. In a VAR model, the marketing(and other) variables are jointly endogenous and theirimpact is measured as shock effects, both short termand long term. This approach is useful for descriptiveand forecasting purposes but not for optimal policyinferences (see Sims 1986 for a detailed discussion inan economic policy context). In the spirit of theoryand practice in marketing, our paper builds on thesedescriptive inferences to (1) describe criteria for mon-itoring the evolution of brand mind-set metrics and(2) make brand-specific marketing recommendationsthat are tested in a holdout sample. Because theirobjective is fundamentally different, Srinivasan et al.do not assess the forecasting accuracy of their modelsin a holdout sample.

Fischer et al. (2011a) proposed a hierarchical deci-sion model with mind-set metrics and conversionrates to guide marketing resource allocations in abusiness-to-business (B2B) setting. Their approach is,unfortunately, not applicable for managers in a typi-cal business-to-consumer (B2C) context for three rea-sons. First, it requires individual-level data, with ahigh collection cost (at regular intervals with broadmarket coverage) for B2C manufacturers. As a result,B2C companies rely on syndicated panel data thatare only available at an aggregate level. Second, theirmodel assumes a fixed hierarchy-of-effects sequence.

Dow

nloa

ded

from

info

rms.

org

by [

128.

97.2

7.20

] on

24

Janu

ary

2016

, at 1

0:35

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.

Hanssens et al.: Consumer Attitude Metrics for Guiding Marketing Mix Decisions536 Marketing Science 33(4), pp. 534–550, © 2014 INFORMS

Although this is appropriate for their specific B2Bapplication, where customers follow a standardizedpurchasing process, such fixed sequences have beeninvalidated in a B2C context (Batra and Vanhonacker1988). Third, for their estimations, they develop anad hoc likelihood function without a closed-formsolution. Instead, we use more general econometricestimation techniques. Based on our estimates, wedevelop a general dynamic programming model thatincludes sales response (with assumed profit mar-gins) as the outcome variable and the mind-set andmarketing mix as input variables. We obtain optimalmarketing resource allocation outcomes and demon-strate their optimality by comparing them to actualbehavior. In these outcomes, we separate marketingeffectiveness in a “transaction route” and a “mind-setroute.”

A third contribution of the paper is the conceptu-alization of criteria on attitude metrics that identifyconditions under which they should be targets of mar-keting action. We offer theoretical foundations anddelineate four key criteria—potential, responsiveness,stickiness, and sales conversion—that help us deter-mine and understand the connection between market-ing actions, attitudinal metrics, and sales outcomes.By applying these criteria, we can determine the mar-keting investment appeal of each marketing instru-ment. For example, if the sales conversion is the samefor two brands but one of them obtains sales conver-sion with less advertising (i.e., it has higher respon-siveness), it is possible for that brand to obtain acompetitive advantage. For managers and researchersalike, our criteria, more generally, offer a verifiableexplanation for changing marketing elasticities and anactionable connection between marketing and finan-cial performance metrics.

Finally, our mixed-effects response models of thelink between marketing actions, attitudes, and salesare well suited for a decision support focus. In partic-ular, cross-effects models establish the extent to whichthe four criteria connecting attitudes to behavior varyover time and across brands. They also indicate whatmatters more: brand or time variation. In addition,longitudinal hierarchical linear models (HLMs) examinehow marketing–attitude and attitude–sales relationsvary by brand. Using our approach, we demonstratesuperior results, in terms of both forecast accuracyand business performance evolution, from using acombined transaction and mind-set approach com-pared with using only attitudinal (mind-set) or mar-keting mix (transaction) models.

Our work is the first to provide criteria for thedecomposition of sales effects through attitudinalmetrics and to offer mind-set-specific guidelines forimproving marketing mix decisions.

Operationalizing the Criteria forAttitude MetricsOur conceptual framework, displayed in Figure 1,contrasts marketing effects that occur throughchanges in attitudinal metrics with those that occurwithout such changes. We denote the former as themind-set route and the latter as the transaction routein Figure 1. We do not propose that purchases canoccur without the customers’ minds or hearts beinginvolved (e.g., one needs to be aware of a brandat least right before buying it), but instead that cus-tomers may simply react to a marketing stimuluswithout changing their mind or heart (e.g., the brandwas in the consideration set before and remains inthe consideration set after a stimulus-induced pur-chase). Our framework therefore accounts for bothgenerally accepted channels of marketing influence:through building the consumer attitudes that consti-tute the brand’s health and/or through leveraging thebrand’s existing health.

To move from an analysis of attitude metrics to rec-ommendations for marketing mix decisions, we haveto identify the managerially relevant attitude metrics.Market research firms provide many possible survey-based consumer attitudes metrics. However, not allof those can be expected to be relevant for marketingplanning for a given brand at a given time. We musttherefore provide specific relevance criteria for thesemetrics. We propose that relevant attitude metricshave potential for growth, are sticky and resistantto competitive erosion, respond to marketing stimuli,and convert into sales.

Potential as a driver of marketing impact has longbeen appreciated and used, especially in the con-text of market potential (e.g., Fourt and Woodlock1960). The central premise is that of diminishingreturns; i.e., the larger the remaining distance to themaximum or ceiling, the higher the impact poten-tial.1 Fourt and Woodlock (1960) applied this prin-ciple to new product penetration forecasting andfound that penetration evolves as a constant frac-tion of the remaining distance to the ceiling. Thusif awareness affects new product trial, then, all elseequal, marketing spending aimed at awareness build-ing will have more impact potential if the beginningawareness is 20% as opposed to 70%. Not account-ing for potential ignores diminishing return effects,resulting in possible overspending with consequently

1 As noted by Haley and Case (1979), over a large range of atti-tude scores, their brand share effect can display increasing returns,at least in self-stated consumer interviews (there are no observedsales or market share data in their study). However, our study con-siders only the leading brands in each category, so we do not haveany observations at the lower end of the attitude scales. Therefore,we do not expect to observe increasing returns to attitude changes.

Dow

nloa

ded

from

info

rms.

org

by [

128.

97.2

7.20

] on

24

Janu

ary

2016

, at 1

0:35

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


Figure 1 Conceptual Framework

lower returns. It can also result in missed opportu-nities on metrics with high potential. The theoret-ical foundation for “potential” comes from marketsize considerations, where the “market” is definedin attitude space. For example, if 100 individualscomprise a regional market, and 20 of those haveprior awareness of brand A, then any marketing cam-paign aimed at improving brand awareness has thepotential of affecting 80 prospects. As the attitudescore increases and approaches 100%, the “untappedmarket size” shrinks, and holding constant the qual-ity of the marketing effort, its anticipated impact willlessen accordingly.

Potential (POTt) is operationalized as the remainingdistance to the maximum, preferably expressed as aratio in light of the multiplicative nature of marketresponse. For example, if maximum awareness (MAX)is 100% and previous awareness At−1 is 30%, then

POTt = 6MAX −At−17/MAX = 0070 (1)

Most consumer attitude metrics are expressed in per-cent (MAX = 100%) or in Likert scales (e.g., 1 to 7,where MAX = 7), both of which readily accommodateour proposed definition of potential.Stickiness refers to the staying power as a result of

the inertia or lock-in of a change in the attitudi-nal metric, in the presence of competitive market-ing. It captures more than the carryover effect of

marketing, as in the decay rate of advertising inthe Koyck model (Hanssens et al. 2001). Stickinessis a property of the attitudinal metric: if the met-ric changes for any reason, how fast does it revertback to its mean? The reasons for such changes mayinclude marketing, competitive marketing, externalshocks, etc. For example, if consumer memory for thebrands in a category is long-lasting, it will take lit-tle or no reminder advertising for a brand to sus-tain a recently gained increase in brand awareness.Similarly, if consumers in a category exhibit stronghabits and routinely choose among the same subset offour brands, then the consideration metric for any ofthese four brands may be sticky. Overall, if a market-ing effort increases a brand’s score on a sticky atti-tudinal metric, then all else equal, that effort is morelikely to have higher returns. The theoretical foun-dation for “stickiness” comes from memory theoryand habit formation theory (Bagozzi and Silk 1983).The results of experiments in social cognition (e.g.,Lingle and Ostrom 1979) and consumer behavior (e.g.,Kardes 1986) suggest that consumers who processbrand attribute information to make evaluative judg-ments will base subsequent brand evaluations primar-ily on the recalled judgment rather than the originalfactual information. In the popular associate networkmodel of the brain, information is “encoded in long-term memory as a pattern of linkages between con-cept nodes” (Burke and Srull 1988, p. 56). However,

Dow

nloa

ded

from

info

rms.

org

by [

128.

97.2

7.20

] on

24

Janu

ary

2016

, at 1

0:35

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


brand-related memories may fade over time and as aresult of additional learning (e.g., Baumgartner et al.1983, Belch 1982, Bettman 1979, Schank 1982). Man-agers need to know the extent of stickiness in themetrics they use; not accounting for stickiness mayresult in myopic decision making and possibly waste-ful marketing spending.Responsiveness refers to marketing’s ability to “move

the needle” on the attitude metric. In this context,different marketing actions will likely have differentresponsiveness. For example, advertising is known tobe better at inducing trial purchases than repeat pur-chases (Deighton et al. 1994), so an awareness metricmay be more responsive to it than a preference met-ric. The theoretical foundation for “responsiveness”comes from utility theory. If a commercial message oran offer provides a net addition to a consumer’s per-ceived utility from purchasing the associated brand,then a response effect is expected. Note that “net”addition here refers to a comparison of marginal util-ity versus marginal cost. Marginal cost could be pricerelated (e.g., the willingness to pay more for an adver-tised brand) or habit related (e.g., the inherent costassociated with taking a risk and switching to a pre-viously unknown brand).

Responsiveness is operationalized as the short-termresponse of the attitude metric with respect to a mar-keting stimulus. Furthermore, stickiness is the carry-over magnitude in the same response equation. Wepropose to use a well-established, robust responsefunction to estimate responsiveness. The standardmultiplicative response model has the advantage ofproducing elasticities as responsiveness metrics:

At = cA�t−1X

�11t X

�22t X

�33t e

ut 1 (2)

where A is the attitudinal metric of awareness asan example and Xi (i = 11213) are marketing instru-ments. Not only does a multiplicative response modelprovide readily interpretable results, it has also beenshown to outperform more complex specificationsin forecasting product trials for consumer packagedgoods (CPG) (e.g., Hardie et al. 1998). Stickiness ismeasured through the carryover parameter �. Forexample, if � = 006, this means that 60% of any changein At is carried over to the next period. We expect� to be less than 1 because of memory decay effectsthat are well documented in psychology (Baddeleyet al. 2009).

Note that responsiveness may be related to poten-tial as follows: the closer the attitude metric is toits ceiling value, the more difficult it will be toregister further increases through marketing. Thatphenomenon is readily incorporated in (2) by express-ing the dependent variable as an odds ratio (e.g.,Johansson 1979):

A′

t =At/4MAX −At5= cA′�t−1X

�11t X

�22t X

�33t e

ut 1 (3)

where the response parameters �i now indicate eithera concave (�i < 1) or an S-shaped (�i > 1) responsecurve. The resulting response elasticity �i is now con-tingent on the attitude metric’s potential as follows:

�i = �i × POTt0 (4)

For example, in an awareness-to-advertising relation-ship with a response elasticity 0.2 at zero initialawareness, the response elasticity will decline to 002×

006 = 0012 when awareness reaches 40%.Sales conversion indicates that changes in an atti-

tudinal metric convert into sales performance. Salesconversion can be expected to vary in different stagesof the purchase funnel; e.g., the lower the funnelstage, the higher the sales conversion. This follows ingeneral from the hierarchy-of-effects model (even ifthe exact hierarchy may not be fixed; see Batra andVanhonacker 1988). For example, a 10% increase inadvertising awareness may increase sales by only 3%,whereas a 10% increase in brand liking may increasesales by 6% (Srinivasan et al. 2010). Not account-ing for sales conversion runs the risk of silo mar-keting practice, i.e., attitude metrics are viewed asthe ultimate performance indicator for marketing, butfinancial executives have no evidence of marketing’simpact on cash flows.

The theoretical foundation for “conversion” comesfrom attitude behavior theory. Although consumersmay have awareness of and attitudes toward severalbrands in a category, they will not necessarily pur-chase all these brands. Since the behavior (purchas-ing) is costly, but the attitude is not, attitude behaviorconversion will only occur when an attitude changeis sufficiently strong. For example, a consumer mayreport increased ad awareness, consideration, and/orliking for a brand (e.g., because of a new ad cam-paign) but fail to purchase it in the supermarket for ahost of reasons, including a high price or habit inertiafor the previously purchased brand.

Conversion rates are typically well below unity;for example, Jamieson and Bass (1989) reported ratiosof actual versus stated consumer trial in 10 prod-uct categories ranging from 0.009 to 0.896, averagingaround 0.5. When historical data are available, conver-sion metrics may be estimated from a “funnel” model,with metrics such as awareness and preference or lik-ing. Logically, one might expect that there is a hierar-chy among mind-set metrics and that awareness, forinstance, leads to consideration. Research on the hier-archy of effects, however, shows that evidence on theexact sequence of effects is mixed (Franses and Vriens2004, Vakratsas and Ambler 1999). As such, we do notwant to impose a hierarchy of effects, because there islittle support for such fixed unidirectional hierarchies(e.g., Batra and Vanhonacker 1988, Norris et al. 2012).

Dow

nloa

ded

from

info

rms.

org

by [

128.

97.2

7.20

] on

24

Janu

ary

2016

, at 1

0:35

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


Instead, we allow for a multiplicative funnel modelthat can be applied across conditions. For example,with intermediate attitudinal metrics awareness (A),consideration (C), and liking (L), a multiplicative fun-nel model for sales revenue (S) would be

St = cS�t−1A

�1t−1C

�2t−1L

�3t−1e

ut 0 (5)

Conversion models such as (5) can be tested eitherwith longitudinal or with mixed cross-sectional time-series data.2

How do the proposed criteria relate to traditionalnotions of short-term and long-term marketing elas-ticity? Short-term marketing-sales elasticity is a combina-tion of the marketing responsiveness and the potentialand sales conversion of each metric. Our decomposi-tion allows managers to assess whether, for instance,low short-term elasticity is due to low marketingresponsiveness versus low inherent potential versuslow sales conversion of a metric. Stickiness corre-sponds to the carryover of marketing effects, so addingthis to the other criteria constitutes long-term market-ing elasticity. As a special case, permanent marketing-sales effects (Dekimpe and Hanssens 1999) arise whena marketing action succeeds in increasing a sales-converting metric that has a stickiness of 1. Finally,our decomposition across metrics allows managers toassess whether a given marketing sales elasticity isdriven by the mind-set route through awareness, con-sideration, or liking.

In conclusion, marketing may influence consumerattitudes, and this, in turn, may improve the brand’sbusiness performance. The degree to which this willoccur depends on the nature of the category (forexample, low versus high consumer involvement)and on the potential, stickiness, responsiveness, andconversion of the attitude metrics. Each of these cri-teria is supported by either consumer behavior ormarket response theory. By combining these scores, abrand may obtain an a priori indication of how effec-tive different marketing campaigns are likely to be.In what follows, we apply our framework for differ-ent brands in multiple categories varying in consumerinvolvement level.

Product Categories, Data, andModelingThe data come from a brand performance trackerdeveloped by Kantar Worldpanel, which reports the

2 To avoid correlated errors stemming from the use of mind-setvariables both as dependent variables (Equation (3)) and as predic-tors, we used the first lagged mind-set variables as predictors in the“sales conversion” model (Equation (5)). This approach is intuitive,as attitudes logically precede choices; i.e., consumers enter the pur-chase occasion with a preexisting attitude, then combine that withthe prevailing marketing mix conditions (prices, promotion, etc.)and make a decision.

marketing mix, consumer attitude metrics (based on8,000 households in France), and performance metricsacross brands in each category on a four-week basis.

For the period between January 1999 and May 2006,we analyze data for the six major brands in eachof these four categories: bottled juice, bottled water,cereal, and shampoo. The broad nature of our data setallows us to investigate whether the extent to whichattitude metrics affect sales varies across brands andproducts. Specifically, as a first validation of ourmodel, we verify whether sales conversion from atti-tudes into purchase behavior differs between higherversus lower involvement purchase situations withinthe studied fast-moving consumer goods. Nelson(1970) developed an economic perspective classifyinga brand purchase decision as either low involvement,where trial is sufficient, or high involvement, whereinformation search and conviction are required priorto purchase. When product involvement is high, abrand needs to change consumers’ hearts and mindsto overcome consumers’ reluctance to change theirpurchase behavior (Bauer 1967, Peter and Tarpey1975). In such cases, we expect movements in atti-tudinal metrics to be strongly associated with sales(i.e., there is sales conversion). In contrast, when prod-uct involvement is low, consumers may choose abrand simply because it is available or promoted,without having fundamentally changed their opin-ion about it. This low involvement path is compatiblewith Ehrenberg’s (1974) awareness-trial-reinforcementmodel. In such cases, we expect low sales conversion.

Marketing actions may have a direct impact onsales without affecting the attitudinal metrics; thisis called the transaction route in Figure 1. In ourdata set, involvement is measured at the categorylevel through several items, including “product cat-egory X is important; you have to be careful whenchoosing a product.” The results show that shampoo(37.8%) is more involving than the food and bever-age categories juice (29.8%), cereal (28.4%), and bot-tled water (28.2%),3 which have minimal variation.The focal brand performance measure is sales vol-ume4 aggregated across all product forms of eachbrand (in milliliters or grams). The marketing mixdata include average price paid, value-weighted dis-tribution coverage, promotion, and total spending onadvertising media.

3 We show the percentge agreeing that a buyer has to pay closeattention to the product chosen.4 Although the actual measure of brand performance is purchases,as registered by consumers, and not sales, as registered by stores,we use the word “sales” in the remainder of the paper. Futureresearch should include actual market-level sales data as a depen-dent variable, particularly if the emphasis is on resource allocation.

Dow

nloa

ded

from

info

rms.

org

by [

128.

97.2

7.20

] on

24

Janu

ary

2016

, at 1

0:35

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


After a discussion with the data provider, weselected the following three measures from the avail-able attitudinal metrics: advertising awareness,5 inclu-sion in the consideration set, and brand liking. Thisselection aimed at covering the three main stages ofthe purchase funnel. The first two measures refer tothe cognitive status of a brand in the consumer’smind, whereas brand liking obviously refers to theaffect status. Two other available measures were notincluded for lack of variation, aided brand aware-ness (which exhibited ceiling effects) and collinearity(intention to purchase correlated highly with the con-sideration set, and the data provider considered thelatter to be managerially more useful).

For advertising awareness, survey respondentsindicated, in a list of all brands present on the market,those for which they “remember having seen or heardadvertising in the past two months.” Our measuregives the percentage of respondents who were aware.For the consideration set, respondents were asked toindicate “the brands that you would consider buying”from a list of all brands in the market. We use thepercentage of respondents who consider buying asthe relevant measure. Liking is measured on a seven-point scale (from “like enormously” to “not at all”),and the measure we use is the average rating.6 Moredetails on these data sources are described in Srini-vasan et al. (2010).

With a time sample of more than seven years, thepresence of different players with different strategiesin different product categories, and wide coverageof the marketing mix as well as consumer attitudi-nal metrics, these data are uniquely suited to addressour research questions. The country of investigation isFrance, which is more homogeneous than large mul-ticultural markets such as the United States in termsof consumer behavior and retail industry structure.

5 Whereas awareness typically means “brand awareness” in mar-keting theory, recent empirical studies (Lautman and Pauwels 2009,Pauwels et al. 2013, Srinivasan et al. 2010) have shown that adver-tising awareness is a key driver of sales across different indus-tries (e.g., drugs, food, drinks, health, beauty) and countries (e.g.,the United States, the United Kingdom, France, Brazil). Intuitively,advertising awareness is important because consumers are exposedto hundreds of ads daily and can recall only a fraction of them(Burke and Srull 1988). Moreover, previous studies found thatadvertising awareness also increased with factors other than adver-tising (Table 6 in Srinivasan et al. 2010). Thus, advertising aware-ness may be a proxy for brand salience (Tulving and Pearlstone1966). Because our data do not contain measures for brand salience,we leave its possible connection with advertising awareness as apromising area for future research.6 The asymmetry in the brand liking versus brand considerationand ad awareness scales is typical for commercially available atti-tude measurement. However, it may partially explain high averagescores for brand liking (as the typical respondent answers the ques-tion for each brand) compared with ad awareness and considera-tion (because the respondent needs to put the effort into clickingon the brand from a list of all other brands in the category).

Econometric ModelingOur empirical setting covers multiple brands in fourdifferent categories over time. Thus we face somecritical questions about the stability and specificityof the relationships we seek to estimate. In particu-lar, we need to test whether attitude stickiness andsales conversion are stable over time or are idiosyn-cratic to certain time periods. In addition, we needto establish whether different brands experience dif-ferent marketing–attitude response effects or if theeffects are generic to the product category. These dis-tinctions are not only econometrically important, theyalso have different strategic implications. For exam-ple, if the attitude-to-sales conversion parameters,including competitive actions, are found to be simi-lar across brands, then no single brand can claim acompetitive sales advantage from lifting an attitudemetric, though a brand can achieve an advantage if itcan lift an attitude metric efficiently.

Our response models focus on brand-level responseto marketing and attitudinal variables. Naturally,these actions and attitudes take place in a competi-tive environment. Our parameter estimates implicitlycontrol for competitive effects because the dependentvariables are real-world observations. Thus, brandsales lifts are implicitly bounded by total categorydemand and consumer loyalty for competing brands.We choose not to include formal competitive mod-els (e.g., equations reflecting individual competitivebehavior and/or market share models) because thesectors under study (shampoo, cereal, fruit juice,and bottled water) are mature, with several impor-tant brands competing in each time period. There isalways at least one brand promoting or advertising;thus competition is stable at the category level. Fur-thermore, research in similar markets (see Steenkampet al. 2005 for a study on more than 1,200 Euro-pean CPG brands) has shown that the predominantcompetitive reaction in advertising or promotion isno reaction at all. We leave the formal inclusion ofbrand-level competition as an area for future research.Table 1 contains an overview of the econometricsmodels used in estimation.

These models are a combination of attitude andsales response and are estimated as mixed-effectsmodels. This allows us to combine fixed and ran-dom effects to separate and investigate how eachlevel affects the attitude criteria (see Web Appendix Afor a technical explanation and model specificationchoices). First, cross-effects (CRE) models allow ran-dom effects to vary both by brand and over time(Baltagi 2005). A typical operationalization is withina cross-sectional or dynamic panel where the cross-sectional dimension (brand, market, etc.) is crossedwith the dynamic factor “time.” In this study’s con-text, CRE models enable us to establish the extent

Dow

nloa

ded

from

info

rms.

org

by [

128.

97.2

7.20

] on

24

Janu

ary

2016

, at 1

0:35

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


Table 1 Overview of Metrics and Models

Metrics/models Equation Model Table

Potential metric Equation (1) POTt = 6MAX −At−17/MAX Table 8

Stickiness metric Equation (3) A′

t = At/4MAX −At 5= cA′�

t−1X�11t X

�22t X

�33t e

ut Table 8

Responsiveness model Equation (3) A′

t = At/4MAX −At 5= cA′�

t−1X�11t X

�22t X

�33t e

ut Table 4

Conversion model (mind-set route) Equation (5) St = cS�t−1A

�1t−1C

�2t−1L

�3t−1e

ut Table 3

Marketing mix model (transactions route) Equation (2) applied to sales St = cS�

t−1X�11t X

�22t X

�33t e

ut Table 5

Transactions + Consumer attitude modela Equation (6) St = cS�

t−1X�41t X

�52t X

�63t A

�4t−1C

�5t−1L

�6t−1e

ut Table 6

aEquation (6) combines Equations (2) and (5) and has both marketing mix and attitudinal metrics.

to which the four criteria on translation of attitudesto behavior vary over time and across brands. Sec-ond, longitudinal hierarchical linear models (HLMs)enable us to investigate how marketing–attitude andattitude–sales relations vary by brand. We can there-fore assess whether higher involvement scenariosimply higher responsiveness, stickiness, and salesconversion of attitude metrics. Moreover, the longi-tudinal HLM separates the variance of an outcomevariable into “among” and “within” variances, whichincreases the precision of estimates (Rabe-Heskethand Skrondal 2005).

We estimate the CRE and longitudinal HLMs onlogistic-transformed (in Equation (3) in Table 1) orlog-transformed (all other equations in Table 1) data.To select the final empirical model, we performseveral tests based on our HLM and CRE modelformulations (see Equations (A1) and (A4) in WebAppendix A). We discuss the results of these testsin the ensuing section in detail. Because of the largenumber of equations and parameters that were esti-mated, we present only a few illustrative tables andgraphs. A full set of econometric results may be foundin the Web appendix.

Estimation ResultsModel TestingThe CRE and HLM estimations across 24 brands infour categories allow us to make several generaliza-tions on the four criteria that govern the attitude-to-sales relationship. For each of the HLMs and CREmodels, we test whether a mixed-effects specifica-tion with both fixed and random effects is superiorto a conventional regression with fixed effects only.Tables 2–8 report the main results.

As Tables 3–6 show, the likelihood ratio (LR) testresults are significant for all models, justifying the useof HLMs and CRE models.7 We also compare (i) thevarying intercept model and (ii) the varying intercept

7 The LR test results for CRE models are available from the authorsupon request.

and varying slope model. The information criteria(Akaike’s information criteria and Bayesian informa-tion criteria) support the latter.8 To obtain the brand-level effects, we combine fixed and random effects atthe brand level. As a diagnostic check, we performnormal Q–Q plots for the standardized residuals.We find no violation of the normality assumption.9

Mediation Test. To test our overall framework(summarized in Figure 1), we conduct a formal medi-ation analysis, using the Sobel–Goodman mediationtest (Sobel 1982) to determine whether a mediator(e.g., attitudinal metric) carries the influence of anindependent variable (e.g., marketing action) to thedependent variable, sales. Full mediation would indi-cate that the attitudinal metrics benchmark model(without marketing mix) is sufficient to predict sales,as the “transaction” route of marketing influence isfully subsumed in the “mind-set” route. On the otherhand, no mediation would indicate that the marketingmix benchmark model (without attitudinal metrics) isappropriate. Finally, partial mediation would suggestthat the full model with both marketing mix and atti-tudinal metrics is superior because it acknowledgesboth transactions and mind-set routes of influence.The Sobel–Goodman tests obtained using the HLMestimation results revealed evidence of partial media-tion (see Web Appendix C2 for results), leading us toconclude in favor of the full model with both market-ing mix and attitudinal metrics, as shown in Figure 1.

Endogeneity Test. The model we use is subject toa potential endogeneity issue as a result of laggedregressors and the brand-level random-effects term.To deal with this, we use the dynamic panel modeldeveloped by Arellano and Bond (1991), using a gen-eralized method of moments (GMM) estimator thatrelies on constructing moment conditions based on

8 The information criteria statistics are available from the authorsupon request.9 The normal Q–Q plots are available from the authors uponrequest.

Dow

nloa

ded

from

info

rms.

org

by [

128.

97.2

7.20

] on

24

Janu

ary

2016

, at 1

0:35

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


the lagged variables.10 The results indicate that thedynamic panel estimates are similar to those of lon-gitudinal HLMs in terms of sign and significance.11

We conclude that accounting for endogeneity doesnot impact our substantive findings, and we reportthe HLM results. These estimates offer superior prop-erties over dynamic panel estimation, including theability to detect to what extent the groups (brands)vary in the intercept and/or slope parameters andthe ability to obtain estimates for the category levelbrand level simultaneously. Furthermore, the varianceof the dependent variable is split into “between” and“within” variances, which increases the precision ofestimates.

Generalizations About Attitudes andTheir Sales ConversionWe organize our findings around the four criteria ofstickiness, conversion, marketing responsiveness, andpotential.

Sales Conversion Is Predominantly Stable OverTime. The CRE model results reveal that brandvariation is more important than time variation inattitude-to-sales conversion models. Table 2 reportsthe percentage variation due to brands and time forall these models. Except for the cereal category, weobserve that the variation in estimates is more brandspecific than time specific. This result highlights thebenefit of strong consumer attitudes favoring a brandand resulting in sales conversion.12

Brand-Specific Attitude Responsiveness to Mar-keting Dominates. Turning to the effects of market-ing actions on attitude metrics, Table 2 shows theyare also more brand specific than time specific. Atti-tude responsiveness is typically specific to the brandand stable over time. This is especially pronouncedin the shampoo category: for each attitude metric, thevast majority (63%–80%) of responsiveness variationis due to brands.

Combining all results from the CRE models, wefind that attitude criteria are predominantly stableover time but vary substantially across brands withinthe same category.13 Therefore, we proceed by nesting

10 Their approach is as follows: first, take first differences of bothsides and eliminate the group effects (ui), then look for instrumen-tal variables (lagged variables), and finally, use GMM to estimatethe model. We implemented these steps and compared the resultsof the dynamic panel GMM estimations with the results of longi-tudinal HLMs.11 These are available from the authors upon request.12 The statistical results and the time-series plots highlighting thesefindings from the CRE models are available from the authors uponrequest.13 These variations across brands may be due to several factors suchas the quality of the product experience, past market spending,

the time variation within the brand variation in lon-gitudinal HLMs to investigate the magnitude of themarketing–attitude and attitude–sales relationships,with a view to understanding the nature of compet-itive brand advantage. The longitudinal HLM resultsfor sales conversion and responsiveness are shown inTables 3 and 4, respectively. We observe differencesacross brands but also note general patterns regard-ing attitude criteria, which can be grouped in two keysets of findings.

Sales Conversion vs. Stickiness for Liking vs.Awareness and Consideration. Table 3 shows theliking-to-sales conversion elasticities for each cate-gory. The median across all brands is 0.549, imply-ing that sales move approximately with the square rootof liking. This affect conversion is more than threetimes the cognitive conversion of consideration forall categories. However, as shown in Table 4, con-sumer liking has two less desirable characteristics forbrands. First, it is less sticky than the cognitive atti-tude metric of awareness. Across the four categories,the median level of stickiness for awareness is 0.499,whereas that for consideration is 0.238 and for lik-ing is 0.234. Second, as shown in Table 4, brand lik-ing is responsive to advertising in only two of thefour categories, whereas consideration is responsiveto advertising in all but one category. Advertisingmoves the needle on all three mind-set metrics ofawareness, consideration, and liking; 9 of the 12coefficients are statistically significant. Price movesthe needle on awareness, consideration, and liking,though only 4 of the 12 responsiveness coefficientsare statistically significant. Promotion influences con-sideration only for the shampoo category. Liking hashigh sales conversion ranging from 0.403 to 0.926,whereas consideration has lower sales conversion:consideration conversions range from 0 to 0.193 (seeTable 3). Note that the highest conversions to salesfrom both consideration (0.193) and liking (0.926) arein the shampoo category, which is higher in consumerinvolvement than the others.

These results may be explained by consumer behav-ior theory. Feldman and Lynch’s (1988) accessibility–diagnosticity framework predicts that a given pieceof information “will be used as an input to a sub-sequent response if the former is accessible and if

and market share. Our data only allow us to check the correla-tion of the latter variable with the attitude criteria across brands.The only substantial correlation is that between market share andadvertising awareness conversion (0.29). Thus, large brands tendto have a higher sales conversion of ad awareness. A plausiblerationale is that, once they notice a brand’s advertising, consumershave an easier time remembering, finding, and buying larger versussmaller brands. This is consistent with the double jeopardy effect(Ehrenberg et al. 1990).

Dow

nloa

ded

from

info

rms.

org

by [

128.

97.2

7.20

] on

24

Janu

ary

2016

, at 1

0:35

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


Table 2 Variation Across Brands and Time in Attitude Responsiveness to Marketing and Sales Conversion: CRE

Category Awareness responsiveness (%) Consideration responsiveness (%) Liking responsiveness (%) Sales conversion (%)

ShampooBrand variation 62048 70086 80022 15014Time variation 10052 3075 4053 6047

Bottled waterBrand variation 44007 76076 64009 71071Time variation 12045 4029 4004 6030

JuiceBrand variation 56061 58015 54089 22093Time variation 5076 8015 5006 4074

CerealBrand variation 32056 53012 30082 6035Time variation 15025 4053 25049 17081

Note. From the cross-effects model output in the Web appendix, read as follows: “Of the total variation in the awareness responsiveness model in the shampoocategory, 62.48% is due to brands, 10.52% due to time, the remainder (27.00%) is residual variation.”

Table 3 Maximum Likelihood Fixed Effects Estimates of Sales Conversion in Longitudinal HLM

Shampoo Bottled water Juice Cereal

Coefficient SE p > �z� Coefficient SE p > �z� Coefficient SE p > �z� Coefficient SE p > �z�

Fixed effectsConstant −10722∗ 00414 00000 00273 00384 00477 −10271∗ 00344 00000 −00617∗ 00258 00017AR(1) 00569∗ 00035 00000 00678∗ 00031 00000 00711∗ 00030 00000 00743∗ 00028 00000Awareness (t − 1) 00011 00056 00845 00013 00029 00667 00055 00032 00080 00006 00032 00844Consideration (t − 1) 00193∗ 00085 00023 −00017 00059 00773 00193∗ 00061 00002 00145∗ 00052 00002Liking (t − 1) 00926∗ 00414 00000 00403∗ 00191 00035 00641∗ 00201 00001 00457∗ 00152 00003

LR test �24 = 35064, p > �2 = 0000 �2

4 = 33040, p > �2 = 0000 �24 = 36096, p > �2 = 0000 �2

4 = 14061, p > �2 = 0001

Notes. The dependent variable is sales. t − 1 indicates one lag is taken. Because of space limitations, random effects estimates are in Table C3 in the Webappendix.

∗Statistically significant effects at p < 0005.

it is perceived to be more diagnostic than other acces-sible inputs” (p. 431). The greater the accessibilityand diagnosticity of an input for a judgment relativeto alternate inputs, the greater the likelihood that itwill be used (Simmons et al. 1993). In high involve-ment categories, such as shampoo, attitudinal changesin consideration make the consumer’s brand expe-rience diagnostic and accessible resulting in highersales conversion. Purchases of low involvement prod-ucts, on the other hand, are not preceded by signifi-cant attitude change, particularly as it pertains to thecognitive attitudinal metrics of awareness and con-sideration. This shows the limitation of relying onlyon attitudinal response for making marketing impactinferences. Even when marketing succeeds in lift-ing an attitudinal metric, it does not imply that thisspecific attitude metric, in turn, converts into sales.Accounting for the full chain reaction of events allowsfor an actionable connection between marketing andfinancial performance metrics.

Attitude Potential Is Higher for Cognitive thanfor Affect Metrics. The cognitive metrics of aware-ness and consideration have higher potential of 73%

and 72%, respectively, whereas the potential for theaffect metric of liking averages 19% (see Table 8). Forinstance, some consumers who are not considering abrand may well have tried it and not liked it, result-ing in higher potential for consideration relative toliking. This suggests that, all else equal, brands havehigher opportunity to make progress on cognitivemetrics. Thus consumer satisfaction (“liking”) runshigh across brands, indicating high product quality,and consequently, the marketing challenges for indi-vidual brands have more to do with their progress inthe cognitive metrics.

Assessing Managerial Relevance

Prediction Test. Given that additional costs areinvolved in the collection of attitudinal data, man-agers will want to ensure that these data improve theaccuracy of sales forecasts, conditional on their mar-keting plans. We assess these improvements by com-paring conditional forecast results for the monthlyobservations of periods 85–96, where the brand’s mar-keting mix decisions for those periods are known

Dow

nloa

ded

from

info

rms.

org

by [

128.

97.2

7.20

] on

24

Janu

ary

2016

, at 1

0:35

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


Table 4 Maximum Likelihood Fixed Effects Estimates of Attitude Responsiveness in Longitudinal HLM

Model 1 (DV = Awareness) Model 2 (DV = Consideration) Model 3 (DV = Liking)

Coefficient SE p > �z� Coefficient SE p > �z� Coefficient SE p > �z�

ShampooFixed effects

Constant −00575∗ 00112 00000 −10055∗ 00132 00000 10132∗ 00179 00000AR(1) 00431∗ 00036 00000 00228∗ 00040 00000 00194∗ 00042 00000Price −00269∗ 00114 00018 −00086 00087 00323 −00071 00100 00481Promotion 00028 00016 00075 00036∗ 00017 00031 00007 00021 00747Advertising 00010∗ 00004 00009 00002∗ 00001 00009 00001 00001 00225

LR test �25 = 110082, p > �2 = 0000 �2

5 = 174010, p > �2 = 0000 �25 = 197043, p > �2 = 0000

Bottled waterFixed effects

Constant −10050∗ 00160 00000 −00881∗ 00408 00031 00623∗ 00226 00006AR(1) 00567∗ 00031 00000 00247∗ 00040 00000 00274∗ 00040 00000Price −00331∗ 00137 00016 −00275 00266 00302 −00552∗ 00160 00001Promotion 00033 00021 00111 −00002 00017 00889 00010 00026 00697Advertising 00013∗ 00002 00000 00003∗ 00001 00003 00004∗ 00002 00029

LR test �25 = 62044, p > �2 = 0000 �2

5 = 175082, p > �2 = 0000 �25 = 149041, p > �2 = 0000

JuiceFixed effects

Constant −00503∗ 00113 00000 −00816∗ 00288 00005 00932∗ 00228 00000AR(1) 00667∗ 00027 00000 00382∗ 00037 00000 00559∗ 00033 00000Price −00093 00096 00332 00053 00163 00743 −00108 00179 00546Promotion 00014 00037 00702 00049 00048 00307 00001 00056 00996Advertising 00009∗ 00002 00000 00002∗ 00001 00054 00004∗ 00002 00028

LR test �25 = 80027, p > �2 = 0000 �2

5 = 133018, p > �2 = 0000 �25 = 83002, p > �2 = 0000

CerealFixed effects

Constant −00673∗ 00197 00001 −00621∗ 00129 00000 10388∗ 00144 00000AR(1) 00297∗ 00038 00000 00067 00042 00110 00109∗ 00041 00008Price 00700 00491 00154 −10094∗ 00276 00000 00185 00233 00427Promotion 00067 00037 00072 00043 00023 00064 −00021 00030 00489Advertising 00008∗ 00003 00004 00000 00003 00889 00001 00004 00749

LR test �25 = 140005, p > �2 = 0000 �2

5 = 208007, p > �2 = 0000 �25 = 126072, p > �2 = 0000

Notes. Because of space limitations, random effects estimates are in Table C4 in the Web appendix. DV, dependent variable.∗Statistically significant effects at p < 0005.

(i.e., planned) at the end of period 84. The benchmarkforecasts are obtained from the marketing mix mod-els (without attitudinal metrics) reported in Table 5,as well as from the attitudinal metrics model (with-out marketing mix) reported in Table 3. The com-parison forecasts are obtained from full models withboth marketing mix and attitudinal metrics reportedin Table 6. These models thus allow marketing actionsto have both transaction and mind-set route effects onsales.

We proceed with comparisons that are based onone-step ahead and multistep forecasts, i.e., projec-tions up to 12 periods ahead. Although the one-step forecasts are expected to be more accurate, themultistep predictions are more realistic and strategi-cally valuable in a 12-month marketing planning sce-nario. Table 7 shows the comparative results, with afocus on prediction accuracy, as measured by meanabsolute percentage error (MAPE). Importantly, the

sales predictions made by the combined “marketingmix and attitudinal metrics” models outperform thebenchmark forecasts obtained using the model withonly attitudinal metrics or the model with only mar-keting mix in all cases. The combined model offerssizeable improvements in prediction: across categoriesand forecast horizons, the average MAPE for theattitude model is 15.7%, for the marketing mix-onlymodel is 17.7%, and for the combined model is 12.0%.As can be expected, the sales prediction improve-ments for one-step forecasts are lower because theseare more accurate in the benchmark models.

Marketing Mix Scenarios Test. The brand speci-ficity of results suggests that individual brands faceunique circumstances that should govern their mar-keting moves. Therefore, in theory, we could performformal optimization of marketing mix spending bybrand and by period, as done, for example, by Fischeret al. (2011a) in the global express delivery sector.

Dow

nloa

ded

from

info

rms.

org

by [

128.

97.2

7.20

] on

24

Janu

ary

2016

, at 1

0:35

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


Table 5 Maximum Likelihood Fixed Effects Estimates of Marketing Mix Models (Transaction Route) in Longitudinal HLM



Fixed effectsConstant 00310∗ 00121 00011 00611∗ 00158 00000 00276∗ 00093 00003 00901∗ 00244 00000AR(1) 00530∗ 00036 00000 00717∗ 00030 00000 00765∗ 00026 00000 00478∗ 00034 00000Price −00253∗ 00106 00017 −00446∗ 00085 00000 −00183∗ 00071 00010 −00018 00374 00961Promotion 00113∗ 00018 00000 00034∗ 00016 00032 00094∗ 00022 00000 00094∗ 00015 00000Advertising 00005∗ 00001 00001 00003 00002 00139 00004∗ 00001 00000 00009∗ 00004 00020

LR test �24 = 38075, p > �2 = 0000 �2

4 = 61024, p > �2 = 0000 �24 = 21079, p > �2 = 0000 �2

4 = 89043, p > �2 = 0000

Notes. The dependent variable is sales. Because of space limitations, random effects estimates are in Table C5 in the Web appendix.∗Statistically significant effects at p < 0005.

Table 6 Maximum Likelihood Fixed Effects Estimates of Transactions+Consumer Attitude Models in Longitudinal HLM



Fixed effectsConstant −10443∗ 00402 00000 −00026 00343 00939 −10460∗ 00280 00000 −00610∗ 00309 00048AR(1) 00420∗ 00036 00000 00658∗ 00031 00000 00658∗ 00029 00000 00503∗ 00032 00000Price −00366∗ 00096 00000 −00443∗ 00111 00000 −00202∗ 00050 00000 −00209 00266 00433Promotion 00127∗ 00016 00000 00038∗ 00019 00048 00098∗ 00020 00000 00088∗ 00015 00000Advertising 00005∗ 00001 00000 00004∗ 00002 00040 00004∗ 00001 00000 00009∗ 00003 00003Awareness (t − 1) 00035 00054 00510 00080 00050 00108 00086∗ 00024 00000 00033 00041 00417Consideration (t − 1) 00166∗ 00079 00050 00009 00058 00876 00156∗ 00056 00006 00168∗ 00069 00014Liking (t − 1) 00897∗ 00261 00001 00317 00189 00092 00674∗ 00176 00000 00426∗ 00138 00002

LR test �27 = 36013, p > �2 = 0000 �2

7 = 47089, p > �2 = 0000 �27 = 22092, p > �2 = 0001 �2

7 = 82010, p > �2 = 0000

Notes. The dependent variable is sales. t − 1 indicates one lag is taken. Because of space limitations, random effects estimates are in Table C6 in the Webappendix.

∗Statistically significant effects at p < 0005.

Such an exercise requires the use of brand-specificcost and profit margins as well as a clear under-standing of each brand’s business objective (for exam-ple, share gains versus profit maximization). Absentsuch financial and strategic information in the currentapplication, we will provide diagnostic informationfor several brands, based on a simulation of different

Table 7 Predictive Performance (MAPE) for the Combined Model vs.the Consumer Attitude and Marketing Mix Models (HoldoutSample: Periods 85–96)

Forecast Consumer Marketing Combinedsolution Category attitude model (%) mix model (%) model (%)

One-step Shampoo 27063 29028 26037ahead Bottled water 6062 5005 3031

Juice 9033 9092 9002Cereal 5086 7054 3081

Multistep Shampoo 33013 37077 26050ahead Bottled water 16067 15069 10079

Juice 16098 23032 9005Cereal 9074 12084 7024

Notes. One-step ahead forecasts update each consecutive period, whereasmultistep forecasts predict 1–12 periods ahead without updating. MAPEdenotes the mean absolute percentage error over the 12-month forecastperiod.

spending scenarios. Using our framework, we diag-nose the brands at the beginning of the holdoutperiod and offer recommendations for changes in themarketing mix, i.e., should the brand’s preexistingmarketing support be increased, maintained, or cut,with the goal of increasing sales. Then we comparetheir business outcomes in function of their actualmarketing spending decisions. For each category, wechoose the two top selling brands: SA and SB inshampoo, WA and WB in bottled water, JA and JBin juice, and CA and CB in cereal. Leaving periods85–96 of our data as a holdout sample, we summa-rize the brands’ market positions in time period 84.As shown in Table 8, we estimate individual brand-level response models for these focal brands andexamine the shifts in marketing spending that thesebrands experienced in periods 85–96 to draw conclu-sions regarding marketing mix decisions.

As an example of brand diagnostics, shampoobrand SA has ample room for mind-set expansionacross the board: awareness is 27%, consideration17%, and liking 71% (5 of 7); potential for aware-ness is therefore 73%, consideration 83%, and liking29%. As a result, the brand’s prospects in attitudi-nal space are high, especially when compared with its

Dow

nloa

ded

from

info

rms.

org

by [

128.

97.2

7.20

] on

24

Janu

ary

2016

, at 1

0:35

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


Table8

Diag

nosticsforthe

TwoTopBran

dsin

Each

Catego

ry,T

=84

Awareness

SASB

WA

WB

JAJB

CACB

Pote

ntia

l(%

)=73

7579

6484

6264

86

Stic

kine

ss=

0.35

0.33

0.48

0.48

0.56

0.57

0.30

0.30

Resp

onsi

vene

ssCo

eff.

SEp>

�z�

Coef

f.SE

p>

�z�

Coef

f.SE

p>

�z�

Coef

f.SE

p>

�z�

Coef

f.SE

p>

�z�

Coef

f.SE

p>

�z�

Coef

f.SE

p>

�z�

Coef

f.SE

p>

�z�

Advertising

0014

100

056

0001

200

095

0007

400

202

0021

300

088

0001

50019

400

088

0002

700

007

0004

400

879

0002

100

044

0062

40051

300

147

0000

12023

100

222

0000

0Prom

otion

0002

700

020

0016

500

035

0002

200

117

0003

400

008

0000

000

034

0002

200

125

0000

700

039

0086

70013

000

040

0000

10010

400

038

0000

60014

800

054

0000

6Price

−0025

900

114

0002

3−0027

300

114

0001

7−0033

200

019

0000

0−0032

800

137

0001

6−

0014

600

105

0018

1−

0013

700

105

0019

51019

100

241

0000

000

700

0049

100

154

Salesconversion

0003

000

059

0061

10009

100

010

0000

000

052

0004

800

278

0007

200

039

0006

200

048

0003

600

179

0003

100

036

0039

00003

000

008

0000

0−

0000

100

034

0097

8

Consideration

SASB

WA

WB

JAJB

CACB

Pote

ntia

l(%

)=83

7150

7069

8173

80

Stic

kine

ss=

0.16

0.16

0.24

0.24

0.27

0.27

0.03

0.03

Resp

onsi

vene

ssCo

eff.

SEp>

�z�

Coef

f.SE

p>

�z�

Coef

f.SE

p>

�z�

Coef

f.SE

p>

�z�

Coef

f.SE

p>

�z�

Coef

f.SE

p>

�z�

Coef

f.SE

p>

�z�

Coef

f.SE

p>

�z�

Advertising

0000

200

015

0091

500

004

0001

500

806

1007

900

219

0000

00078

800

227

0000

100

146

0020

000

465

1023

700

298

0000

00043

400

121

0000

00072

100

176

0000

0Prom

otion

0007

300

020

0000

000

044

0002

700

112

0000

100

025

0098

000

016

0002

600

541

0005

600

077

0046

90014

500

053

0000

60006

700

027

0001

30009

200

036

0001

1Price

−00

086

0008

700

323

−00

086

0008

700

323

−00

275

0026

600

302

−00

275

0026

600

302

0019

300

287

0050

300

198

0028

700

490

1037

000

076

0000

01009

400

276

0000

0Sa

lesconversion

0019

800

089

0002

70028

800

010

0000

000

001

0006

800

988

0006

500

013

0000

00021

400

063

0000

10019

200

063

0000

20018

700

006

0000

00013

600

051

0000

8

Liking

SASB

WA

WB

JAJB

CACB

Pote

ntia

l(%

)=29

112

229

2726

24

Stic

kine

ss=

0.14

0.14

0.25

0.25

0.36

0.36

0.09

0.09

Resp

onsi

vene

ssCo

eff.

SEp>

�z�

Coef

f.SE

p>

�z�

Coef

f.SE

p>

�z�

Coef

f.SE

p>

�z�

Coef

f.SE

p>

�z�

Coef

f.SE

p>

�z�

Coef

f.SE

p>

�z�

Coef

f.SE

p>

�z�

Advertising

0000

100

001

0003

500

001

0002

100

953

0000

900

123

0094

300

022

0009

700

819

0001

900

002

0000

00000

800

002

0000

00062

000

147

0000

000

218

0019

100

256

Prom

otion

−00

014

0006

400

827

−00

014

0003

700

702

−00

018

0002

700

495

−00

017

0002

700

524

0002

600

059

0065

5−

0003

700

059

0053

000

065

0004

200

122

−00

010

0004

500

819

Price

−00

078

0010

000

434

−00

078

0010

000

435

−0058

200

160

0000

0−0058

200

160

0000

0−

0021

800

209

0029

8−

0021

800

209

0029

800

185

0023

300

427

0018

300

233

0043

1Sa

lesconversion

0075

600

286

0000

10075

800

082

0000

000

254

0021

500

236

0025

100

047

0000

00045

800

213

0003

20045

700

213

0003

20046

700

028

0000

00045

600

162

0000

5

Note

.Bol

din

dica

tes

stat

istic

ally

sign

ifica

ntef

fect

satp<

0005

.

Dow

nloa

ded

from

info

rms.

org

by [

128.

97.2

7.20

] on

24

Janu

ary

2016

, at 1

0:35

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


competitor, shampoo brand SB. By contrast, the areaswhere bottled water brand WA has more potentialthan its competitor WB are less marketing actionable.For example, WA has more advertising awarenesspotential and less liking potential than WB. However,the latter matters much more in this low involvementcategory. Thus, any marketing effort that stimulatesattitude metrics other than liking is likely to have onlynegligible demand effects.

Two brands implemented major shifts in their mar-keting allocations after T = 84. Shampoo brand SAincreased its advertising spending by 50% and keptits prices the same. In contrast, water brand WAcut its advertising spend by 42% while also keep-ing prices the same. What are the consequences ofthese brands’ strategic actions? We make directionalsales forecasts up to 12 months later, based on theirattitude criteria shown in Table 8. As an illustra-tive example, for shampoo brand SA, there is a highresponsiveness of attitudinal metrics to advertising.Specifically, an increase in advertising moves the nee-dle on both the attitudinal metrics of awareness andliking. The awareness metric has an ample potentialof 73%, whereas the liking metric has a potential of29%. Finally, liking converts to sales resulting in aforecasted increase in sales, which we denote with a“↑” in Table 9.

Similar calculations through the chain of eventsfrom marketing actions → attitudinal metrics →

sales conversion are performed for each of the fourbrands in the analyses. In Table 9, we offer model-based recommendations on changes in marketingmix decisions for advertising, i.e., increasing (“↑”),decreasing (“↓”), or maintaining (“−−”) with a viewto increasing brand sales. As Table 9 shows, brandsthat followed a different course from the model-based recommendations on marketing mix decisions(as depicted in the column “Agreement with model-based recommendation on spend”) performed worsein terms of actual sales outcomes compared withbrands that followed a course consistent with model-based recommendations. Thus, diagnosing attitudi-nal brand metrics can help directionally predict theimpact of different marketing mix decisions on sales.

Table 9 Illustrations of Model-Based Marketing Recommendations and Sales Outcomes

Sales outcomeAdvertising spend Agreement with model-based Forecast conditional on

Brand Recommend Actual recommendations on spend agreement with recommendation Actual

SA ↑ ↑ Yes ↑ ↑

SB ↓ ↓ Yes ↑ −−

WA ↑ −− No ↑ ↓

WB ↑ −− No ↑ ↓

Note. ↑ denotes an increase, ↓ denotes a decrease, and −− denotes no change.

Resource Allocation ImplicationsWe explore two additional ways in which our resultsare helpful for guiding marketing decision making.First, at any point in time, a manager can gaugethe overall attractiveness of investing in marketingactions that move the needle on different attitudinalmetrics. As an illustration, consider the diagnostics(at time = 84) for shampoo brand SA versus juicebrand JB (in Table 8). Brand SA can make progressin its liking through advertising, which affects itspreference score; this, in turn, converts into highersales. By contrast, brand JB needs to use promo-tions to increase consideration that converts into sales.Numerous comparative scenarios can be derived fromTable 9, but these are limited to top-line inferences;i.e., there are no cost and profit implications of thesescenarios.

Second, we conduct a more formal analysis of opti-mal marketing mix spending using dynamic pro-gramming. To illustrate how to make marketing mixdecisions by taking into account a mind-set metric,we pick two different shampoo brands, SC and SD,as they have similar sales levels but varying levels ofawareness; shampoo brand SC has higher awarenessthan brand SD. We assume that both brands have thesame 10% growth targets in terms of sales and aware-ness over the last 12 periods. Our goal is to obtain theoptimal marketing mix path for both brands and val-idate our results using out-of-sample data from t = 85to t = 96.

We address the optimal marketing path by consid-ering two main steps (Gupta and Steenburgh 2008).In step 1, we estimate sales response and attituderesponse (awareness) models using time-series econo-metrics. In step 2, we find the optimal marketingpath by using the estimated parameters in the opti-mization part. Specifically, we follow a DP approachused for solving multistage optimization problems(Rust 2006). The central idea of the DP is to maxi-mize (minimize) the sum of today’s reward (loss) andthe discounted expected reward (loss) from the futureperiods (Bellman 1957). The details of these two stepsare described in Web Appendix B.

Dow

nloa

ded

from

info

rms.

org

by [

128.

97.2

7.20

] on

24

Janu

ary

2016

, at 1

0:35

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


Figure 2 Optimal Price and Advertising Policies for Brands SCand SD

84 86 88 90 92 94 960

2

4

6

8

10Optimal advertising policy

Time period

Adv

ertis

ing

84 86 88 90 92 94 960

1

2

3Optimal price policy

Time period

Pric

e

Brand SCBrand SD

Figure 2 displays the optimal marketing mix pathover 12 periods to achieve the targeted sales (+10%)and targeted awareness (+10%). In addition, we com-pute (X∗

p1 t × St − X∗a1 t), where t is the time index;

X∗p and X∗

a are the optimal price and advertisingpolicies, respectively; and S is the observed saleslevel. The cost of increasing revenue performance isthrough increased advertising or lowering price, orboth. Note that despite the similar sales starting posi-tion and target, our model recommends different mar-keting policies for each brand.

Figure 3 shows the optimum sales revenuesachieved upon implementing the optimal price andadvertising actions from t = 85 to t = 96. Both brands

Figure 3 Optimal Revenues for Brands SC and SD

substantially increase projected sales revenues: 40%for brand SC and 34% for brand SD.

Managerial Implications andConclusionsWe argued in our introduction that the CFO’s needsfor financial accountability of marketing may well bemet by traditional marketing mix models on transac-tions data. However, the chief marketing officer alsoneeds to understand the consumer behavior reasons whymarketing does or does not affect business perfor-mance. Our paper has demonstrated that the objec-tives of both stakeholders can be met by recognizingthe unique properties of attitudinal metrics and theirrelationship to sales performance. In particular, thesemeasures have potential, stickiness, and responsive-ness to marketing that can be assessed from the data.Furthermore, the relevance of these metrics may beassessed by their conversion into sales performance,which provides the critical accountability link withthe CFO’s needs. By applying our approach, man-agers can develop actionable guidelines on how toapply closed-loop learning on the attitude metrics(e.g., “if one observes metrics with the following val-ues/characteristics, then this marketing action will bemost effective”).

Different product categories and brands withinthem vary significantly in the magnitude of the fourproposed criteria, and these differences form the basisfor formulating marketing mix strategies that aremore likely to succeed. Table 10 provides an overviewof four corner cases. The estimates reported in Table 8allow a classification of the brands into the four cells.

First, if a brand has low sales conversion from con-sumer attitudinal metrics, and low responsiveness tomarketing, we label that scenario a transactions effect atbest. For the eight brands with three attitudinal met-rics in Table 8, only one can be classified into this cell.Thus, for most brands, marketing mix strategies resultin sales conversion through the “mind-set effect”; i.e.,at least one attitudinal metric/marketing mix combina-tion is sales relevant for all of these brand scenarios,lending strong support to our current approach.

Turning to the second case, if a brand has lowconversion to sales from consumer attitudinal met-rics but high responsiveness to marketing, we label

Table 10 Strategic Importance of Attitudinal Metrics

Sales conversionResponsivenessof attitudeto marketing Low High

Low Transactions effect at best Ineffective marketing leverHigh Ineffective marketing focus Long-term effect potential

Dow

nloa

ded

from

info

rms.

org

by [

128.

97.2

7.20

] on

24

Janu

ary

2016

, at 1

0:35

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


that scenario an ineffective marketing focus. For exam-ple, brands that invest substantially in considerationset-enhancing advertising may fail to see a substan-tial sales lift. A case in point is water brand WA withrespect to advertising. Increases in advertising gen-erate awareness and consideration lifts that do notconvert to sales. Hence, for water brand WA, advertis-ing represents an ineffective marketing focus that mayplease managers focused on awareness and consider-ation metrics but not managers focused on increasingthe top line.

Third, if the attitudinal metric has high sales con-version but does not respond well to increased mar-keting spending, that would result in an ineffectivemarketing lever scenario. This is the situation thatshampoo brand SB finds itself in with regard toadvertising. Consideration and liking have high salesconversion for SB, but they do not respond well tomarketing spending. Managers can use such insightsto motivate a detailed analysis of the reasons, whichmay include the wrong message, the wrong execu-tion, the wrong communication channel, the wrongtiming, etc. In contrast, increases in shampoo brandSA’s advertising generate liking that converts intosales. Hence, advertising is an “effective marketinglever” for shampoo brand SA to generate sales con-version from a lift to liking.

Finally, if the attitude metric has high sales conver-sion, and there is high responsiveness to marketing,we label that as a situation with long-term poten-tial. For example, cereal brand CA has sales conver-sion from awareness and consideration, which havea high responsiveness to all marketing actions. Thisoffers an opportunity to allocate marketing resourcesto move the needle on the consumer attitudinal met-ric of awareness and consideration and eventually toa long-term sales lift.

Our research opens up several avenues for futurework. One area is to examine alternative functionalforms on the relationship between attitudes and sales,with an assessment of the relative performance of log-log models, log-linear models, as well as other formsof nonlinearity. Moreover, the effectiveness of market-ing actions (e.g., promotions and advertising) couldbe modeled as functions of awareness, consideration,and liking. In addition, we model potential as theremaining distance to the maximum for its ease ofelasticity definition, but future research should com-pare alternative operationalizations (e.g., square rootof the remaining distance to the maximum) as wellas alternative models and estimation techniques (e.g.,Bayesian vector autoregressive models).

Future research should also explore category com-parisons with even higher levels of consumer involve-ment, such as durables and high-value services,possibly using observations at different time intervals

(e.g., weekly, monthly, quarterly) and including dataon competitors. Comparisons between brands couldalso be made in a more systematic way on the basisof their strategic orientation—for instance, in termsof their degree of differentiation. Degree of differen-tiation and other factors such as market share maybe important drivers of the cross-brand differencesin mind-set metric criteria. A company interested inbrand equity may, in addition, want to examine towhich extent its brand can command a revenue pre-mium compared with other brands and private labels(Ailawadi et al. 2003). The revenue premium can itselfbe modeled as a function of mind-set metrics. More-over, data on the profits gained from better decisionswould enable managers to weigh them against thecost of collecting attitudinal metrics, thus providing areturn on investment measure for such data.

If individual-level attitude metrics are available,these could be used in more granular response modelspecifications. Indeed, the lack of attitudinal metricsthat match the transactional records is a limitationof our approach. Attitudinal tracking data are typ-ically survey based, which is costly and subject tosampling error. The digital age offers new opportu-nities in this regard. Instead of surveying consumers,one can observe how they express themselves on theInternet via searches, chat rooms, social media, socialnetwork sites, blogs, product reviews, and similaronline word-of-mouth forums. Some preliminary evi-dence suggests that “Internet-derived consumer opin-ions” are predictive of subsequent behavior (e.g., Shinet al. 2013). Future research should examine whichInternet-derived attitudinal metrics are the most rel-evant and investigate the extent to which measure-ment error in online versus off-line metrics may mat-ter. These metrics could then be substituted for thesurvey-based measures that were used in this paper.

Supplemental MaterialSupplemental material to this paper is available at http://dx.doi.org/10.1287/mksc.2013.0841.

AcknowledgmentsThis paper was originally intended to appear in the specialissue on Theory and Practice in Marketing Conference Spe-cial Section, but the review process was not concluded intime. The authors are listed in alphabetical order. The authorsthank Kantar Worldpanel France for providing the data usedin this paper. They also thank participants at the Market-ing Science Conference, the Marketing Dynamics Confer-ence, and the Theory+Practice Conference for their valuablesuggestions.

ReferencesAilawadi KL, Lehmann DR, Neslin SA (2003) Revenue premium as

an outcome measure of brand equity. J. Marketing 67(4):1–17.

Dow

nloa

ded

from

info

rms.

org

by [

128.

97.2

7.20

] on

24

Janu

ary

2016

, at 1

0:35

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


Arellano M, Bond S (1991) Some tests of specification for paneldata: Monte Carlo evidence and an application to employmentequations. Rev. Econom. Stud. 58(2):277–297.

Baddeley A, Eysenck MW, Anderson MC (2009) Memory (Psychol-ogy Press, London).

Bagozzi RP, Silk AJ (1983) Recall, recognition, and the measurementof memory for print advertisements. Marketing Sci. 2(2):95–134.

Baltagi BH (2005) Econometric Analysis of Panel Data, 3rd ed. (JohnWiley & Sons, London).

Batra R, Vanhonacker WR (1988) Falsifying laboratory resultsthrough fields tests: A time-series methodology and someresults. J. Bus. Res. 16(4):281–300.

Bauer RA (1967) Consumer behavior as risk taking. Cox DF, ed. RiskTaking and Information Handling in Consumer Behavior (HarvardUniversity Press, Boston), 23–33.

Baumgartner MH, Leippe MR, Ronis DL, Greenwald AG (1983)In search of reliable persuasion effects: II. Associative interfer-ence and persistence of persuasion in a message-dense envi-ronment. J. Personality Soc. Psych. 45(3):524–537.

Belch GE (1982) The effects of television commercial repetition oncognitive response and message acceptance. J. Consumer Res.9(1):56–65.

Bellman R (1957) Dynamic Programming (Princeton University Press,Princeton, NJ).

Bettman JR (1979) Memory factors in consumer choice: A review.J. Marketing 43(2):37–53.

Burke RR, Srull TK (1988) Competitive interference and consumermemory for advertising. J. Consumer Res. 15(1):55–68.

Deighton J, Henderson CM, Neslin SA (1994) The effects of adver-tising on brand switching and repeat purchasing. J. MarketingRes. 31(1):28–43.

Dekimpe M, Hanssens DM (1999) Sustained spending and persis-tent response: A new look at long-term marketing profitability.J. Marketing Res. 36(4):397–412.

Ehrenberg ASC (1974) Repetitive advertising and the consumer.J. Advertising Res. 14(2):25–34.

Ehrenberg ASC, Goodhardt GJ, Barwise TP (1990) Double jeopardyrevisited. J. Marketing 54(3):82–91.

Feldman JM, Lynch JG Jr (1988) Self-generated validity and othereffects of measurement on belief, attitude, intention and behav-ior. J. Appl. Psych. 73(3):421–435.

Fischer M, Giehl W, Freundt T (2011a) Managing global brandinvestments at DHL. Interfaces 41(1):35–50.

Fischer M, Albers S, Wagner N, Frie M (2011b) Dynamic marketingbudget allocation across countries, products, and marketingactivities. Marketing Sci. 30(4):568–585.

Fourt LA, Woodlock JW (1960) Early prediction of market successfor new grocery products. J. Marketing 25(2):31–38.

Franses PH, Vriens M (2004) Advertising effects on awareness, con-sideration and brand choice using tracking data. ERIM ReportSeries ERS-2004-028-MKT, Erasmus University Rotterdam,Rotterdam, The Netherlands.

Gupta S, Steenburgh TJ (2008) Allocating marketing resources.HBS Working Paper 08-069, Harvard Business School, Boston.http://www.hbs.edu/faculty/Pages/item.aspx?num=32002.

Haley R, Case PB (1979) Testing thirteen attitude scales for agree-ment and brand discrimination. J. Marketing 43(4):20–32.

Hanssens DM, Parsons LJ, Schultz RL (2001) Market Response Mod-els: Econometric and Time Series Analysis, 2nd ed. (Kluwer Aca-demic Publishers, Norwell, MA).

Hardie BGS, Fader PS, Wisniewski M (1998) An empirical com-parison of new product trial forecasting models. J. Forecasting17:209–229.

Jamieson LF, Bass FM (1989) Adjusting stated intention measuresto predict trial purchase of new products: A comparison ofmodels and methods. J. Marketing Res. 26(3):336–345.

Johansson JK (1979) Advertising and the S-curve: A new approach.J. Marketing Res. 16(3):346–354.

Kardes FR (1986) Effects of initial product judgments on subsequentmemory-based judgments. J. Consumer Res. 13(1):1–11.

Keller KL, Lehmann DR (2006) Brands and branding: Research find-ings and future priorities. Marketing Sci. 25(6):740–759.

Kotler P, Keller KL (2012) Marketing Management, 14th ed. (PearsonPrentice-Hall, Upper Saddle River, NJ).

Lautman M, Pauwels K (2009) What is important? Identifying met-rics that matter. J. Advertising Res. 49(3):339–359.

Lingle JH, Ostrom TM (1979) Retrieval selectivity in memory-basedimpression judgments. J. Personality Soc. Psych. 37(2):180–194.

Nelson P (1970) Information and consumer behavior. J. PoliticalEconom. 78(2):311–329.

Norris IB, Peters K, Naik PA (2012) Discovering how advertisinggrows sales and builds brands. J. Marketing Res. 49(6):793–806.

Pauwels K, Erguncu S, Yildirim G (2013) Winning hearts, mindsand sales: How marketing communication enters the purchaseprocess in emerging and mature markets. Internat. J. Res. Mar-keting 30(1):57–68.

Peter JP, Tarpey LX Sr (1975) A comparative analysis of three con-sumer decision strategies. J. Consumer Res. 2(1):29–37.

Rabe-Hesketh S, Skrondal A (2005) Multilevel and Longitudinal Mod-eling Using Stata (Stata Press, College Station, TX).

Rust J (2006) Dynamic programming. Entry for consideration by theNew Palgrave Dictionary of Economics. Accessed September 1,2013, http://gemini.econ.umd.edu/jrust/research/papers/dp.pdf.

Schank RC (1982) Dynamic Memory: A Theory of Reminding in Com-puters and People (Cambridge University Press, New York).

Shin HS, Hanssens DM, Kim KI, Choe JA (2013) Positive vs. neg-ative e-sentiment and the market performance of high-techproducts. MSI Working Paper 13-123, University of California,Los Angeles, Los Angeles.

Simmons CJ, Bickart BA, Lynch JG Jr (1993) Capturing and creat-ing public opinion in survey research. J. Consumer Res. 20(2):316–329.

Sims CA (1986) Are forecasting models usable for policy analysis?Federal Reserve Bank of Minneapolis Quart. Rev. 10(1):2–6.

Sobel ME (1982) Asymptotic confidence intervals for indirect effectsin structural equation models. Sociol. Methodol. 13:290–312.

Srinivasan S, Vanhuele M, Pauwels K (2010) Mindset metrics inmarket response models: An integrative approach. J. MarketingRes. 67(4):672–684.

Stahl F, Heitmann M, Lehmann DR, Neslin SA (2012) The impactof brand equity on customer acquisition, retention, and profitmargin. J. Marketing 76(4):44–63.

Steenkamp J-B, Nijs VR, Hanssens DM, Dekimpe MG (2005) Com-petitive reactions and the cross-sales effects of advertising andpromotion. Marketing Sci. 24(1):35–54.

Tulving E, Pearlstone Z (1966) Availability versus accessibility ofinformation in memory for words. J. Verbal Learn. Verbal Behav.5(4):381–391.

Vakratsas D, Ambler T (1999) How advertising works: What do wereally know? J. Marketing 63(1):26–43.

Dow

nloa

ded

from

info

rms.

org

by [

128.

97.2

7.20

] on

24

Janu

ary

2016

, at 1

0:35

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.

Documents

Consumer Attitude Metrics for Guiding Marketing Mix Decisionscosts of each marketing instrument can determine the respective investment appeal of each of these actions. After a description