reciprocal spillover effects

8/8/2019 reciprocal spillover effects

1/11

Subramanian Balachander & Sanjoy Ghose

Reciprocal Spillover Effects: AStrategic Benefit of BrandExtensionsA commonly advanced rationale for the proliferation of brand extensions is companies' motivation to leverage thequity in established brands, thereby developing profitable products relatively easily. A more interesting strategiargument for brand extensions that has been advanced is that extensions would favorably affect the image of thparent brand and thereby influence its choice. In this research, the authors investigate the existence of such recprocal spillover effects emanating from the advertising of a brand extension. The authors use scanner panel datand study spillover effects of advertising on brand choice. They develop implications for brand and product linmanagement.

I n recent years, there has been a plethora of line and brandextensions. Between 1977 and 1984, 40% of the 120 to175 new brands introduced each year in supermarketswere extensions (Aaker 1990). One commonly advancedrationale for this proliferation of extensions is companies'motivation to leverage the equity in established brands anddevelop profitable products relatively easily (Morein 1975).A second motivation for extensions is to affect the image ofthe umbrella brand favorably, which thereby influencessales in other categories. Aaker (1996) offers several exam-ples in which existing products obtain this reciprocal bene-fit from brand extensions. He suggests that Gallo opted toattach its name with a "jug-wine" reputation to its line ofupscale, corked wines, Ernest and Julio Gallo Varietals, toimprove the quality perception of its low-end wine product.In another example, Contadina, which was perceived as astrong canned-foods brand with an authentic Italian her-itage, was revitalized by its entry into fresh refrigerated pas-tas and sauces.

If introduction of a brand extension can produce such(reciprocal) spillover benefits to existing products, it can beexpected that advertising of the brand extension will alsohave a positive spillover effect on sales of existing products.For example, Aaker (1996) suggests that advertising of thebrand extension Hidden Valley Honey Dijon Ranch saladdressing made the advertising for the Hidden Valley brandgroup more effective. Such advertising spillover effectswould also have implications for allocation of advertisingbudget among the new brand extension and existing prod-ucts with the same brand name. However, it is surprisingSubramanian Balachander is Assistant Professor of Marketing, KrannertGraduate School of Management, Purdue University. Sanjoy Ghose isProfessor of Marketing, School of Business Administration, University ofWisconsin-Milwaukee. The authors acknowledge the helpful comments ofDennis Gensch, Jugal Ghorai, PK. Kannan, Manohar Kalwani, VithalaRao, and Brian Ratchford and a seminar at Carnegie Mellon University forsparking this inquiry. Part of this research was conducted when the sec-ond author was a Senior Visiting Fellow at the National University ofSingapore.

that there is little documentation of the existence of advetising spillover effects in academic literature in spite of thavailability of high-quality, single-source scanner panedata. In the only field research known to us, Sullivan (1990studies the effect on the demand of used Jaguar m odels ariing from the introduction of a new Jaguar m odel. Her resulcan be interpreted as somewhat mixed. Although she founthat the event of introduction of the new model increasedemand for used Jaguar models (positive spillover), advetising of the new Jaguar model depressed demand for useJaguars (negative spillover). If the event of product introduction is considered an "information shock" and advertising represents a more steady flow of information, similaeffects for both forms of information should be expecteHowever, Sullivan does not find such a result. A possibreason is that Sullivan does not separate the negative substtution effect that arises from a brand extension, depressinsales of existing products from a likely positive spilloveeffect. Thus, unlike the spillover effect of information shoccaused by a product introd uction, spillover effects fromongoing advertising may not be strong enough to ovewhelm a substitution effect.

In this article, we separate the substitution and spilloveffects by estimating demand at the disaggregate level usinACNielsen's single-source scanner panel data. We find thathe advertising of brand extensions produces significant reiprocal spillover that favorably affects the choice of the paent brand. Th us, we find empirical support for the anecdotevidence presented previously. Our separation of spilloveand substitution effects may enable us to find a positivspillover effect of advertising, in contrast to Sullivan (1990who finds a negative spillover effect for Jaguar advertising

'Category differences (yogurt is a category we analyze) probbly account for some of the differences between Sullivan's (199results and ours. In a category such as yogurt, a consumer can puchase both parent (low-fat yogurt) and extension (non-fat yogurtas a result of complementarity or variety-seeking behavior. In sua case, positive spillover effects may be enhanced. We thank anonymous reviewer for suggesting the role of categordifferences.


2/11

In contrast to our use of disaggregate data, Sullivan usesaggregate data from published sources. Tellis and Weiss(1995) suggest that aggregation may lead to bias iti estimat-ing the effects of advertisitig. Our use of disaggregate dataalso provides a strong test for the existence of advertisingspillover effects because advertising effects have been diffi-cult to establish with such data (Tellis and Weiss 1995). Iti alaboratory study, Morrin (1999) finds that advertising of abrand extension facilitates recall of the parent brand. How-ever, our research focuses on brand ch oice and examines theadvertising effect of both a brand and its extension in a fieldsetting.In the remainder of this article, we present the concep-tual background that pertains to spillover effects. We thendevelop the empirical models and present the analysis andresults. Finally, we discuss the managerial implications ofour findings and limitations of our study and offer a briefconclusion.

Conceptual BackgroundSpillover Advertising EffectAdvertising spillover becomes relevant when a brand nameis used on two or more products that are separately adver-tised. Consider two products, A and B, that carry anumbrella brand name in common (e.g., Yoplait yogurt andYoplait non-fat yogurt). We conceptualize a spillover effectas the impact of Product A's (B's) advertising on the utilityto the consumer of Product B (A). In general, the spillovereffects between products may not be symmetric. In particu-lar, we distinguish between the spillover effects from adver-tising of the parent (the product that originally used thebrand name) and those from advertising of the child orextension (the line or brand extension). We refer to the for-mer effect as the forward spillover effect and the latter effectas the reciprocal spillover effect. The definition of parent asthe product that originally used the brand name is similar tothe definition of "core" brand in Keller and Aaker's (1992)study. In a m ore theoretical sense, we consider the parent asthe product most closely associated with the umbrella brandname in the consumer's mind (Farquhar, Herr, and Fazio1990; Morrin 1999). This perspective is similar to the con-cept of a flagship product used by John, Loken, and Joiner(1998) or "instance d om inance" used by Herr, Farquhar, andFazio (1996).

We offer several commercial examples in which a posi-tive reciprocal spillover effect is anticipated. There are twomain theoretical reasons to expect such a positive reciprocalspillover effect. First, a positive spillover eflect would beconsistent with the existence of economies of information inadvertising when an umbrella or "range " brand is applied todifferent products (Aaker 1996; Morein 1975). Indeed, in anempirical study. Smith (1992) finds that advertising expen-ditures for umbrella-branded products are lower, which isconsistent with economies of information. As Aaker (1996,p. 295) notes, such economies are realized because "thefixed cost of maintaining a brand name can be spread acrossdifferent busin esses." The implication of this rationale is thatumbrella-branded products benefit one another with theiradvertising because of positive spillover effects, resulting in

less advertising expenditure for each product. In a similarvein, Morein (1975) suggests that economies of informationare realized because an advertised product produces a "haloeffect" that increases sales of other umbrella-branded prod-ucts. Thus, we expect both reciprocal and forward spillovereffects to be positive because of economies of information.Although the economies of information argument isintuitively reasonable, the specific mechanism throughwhich advertising spillover or halo occurs is not clear. Wern-erfelt's (1988) analysis of umbrella branding suggests onesuch mech anism. U sing a signaling mo del, Wernerfelt theo-rizes that umbrella-branded products are perceived to offerhigher quality because profits from other umbrella-brandedproducts act as a "performance bond" for the quality of anyof the umbrella-branded products. In other words, if a low-quality product is offered with an umbrella brand name, itleads consumers to conclude that all other products with thesame brand name are also of low quality, which thus threat-ens the profits from these other products. Therefore, a firmwould optimally extend an established brand name only tohigh-quality products, thus rendering consumers' percep-

tions (that umbrella-branded products are of high quality)accurate. Wernerfelt's analysis offers a mechanism by whichadvertising can spill over and enhance sales for otherumbrella-branded products.2 Essentially, the advertising ofother products with the same brand name makes consumersaware of the performance bond at stake for the firm, therebyincreasing quality perceptions of unadvertised products andenhancing their sales. In summary, the theory of economiesof information suggests positive reciprocal and forwardspillover effects.Second, Anderson (1983) offers a consumer memory-based explanation for the existence of reciprocal spillovereffects with the associative network theory. Brand associa-

tions in consumer memory are a key component of brandequity and brand-related effects (Aaker 1996). The associ-ated network theory has been particularly useful in analyz-ing the effect of brand associations (see, e.g., Herr, Far-quhar, and Fazio 1996; John, Loken, and Joiner 1998; Keller1993; Morrin 1999). This theory conceptualizes knowledgeabout a brand as being a network of nodes (or concepts) con-nected by links, which represent associations between theconcepts. M oreover, the strength of a link is a measure of theassociation strength between the concepts. Thus, the brand(e.g., Yop lait), the paren t (e.g., Yoplait [regular] yo gurt), andthe brand or line extension (e.g., Yoplait non-fat yogurt), aswell as beliefs about the brand, are conceptualized as nodesin a knowledge network, and the links between the nodesvary in strength. A consumer retrieves a particular piece ofknowledge from memory when the corresponding node isactivated above a threshold level, through priming by exter-nal cues such as advertising or by "spreading" activationfrom other linked nodes. The extent of spreading activationto a new node increases with the strength of the link between2ln experimental studies, Dacin and Smith (1994) find that con-sumers' evaluation of the quality of an umbrella-branded productincreases with the number of products using the same brand name.Because several products using the same brand name imply a big-ger performance bond, the result appears to support Wernerfelt's

(1988) analysis. See also Erdcm (1998).


3/11

the new node and the previously activated node. A strongerlink facilitates the spreading activation to the new nodeabove the threshold to be retrieved from memory.Because we expect the parent to be strongly associatedwith the brand in consumers' memories, the link betweenthe parent node and the brand node is likely to be strong(Farquhar, Herr, and Fazio 1990). When exposure to adver-tising of the brand extension activates the brand node, theactivation will likely spread to the parent because of thestrength of the link between the two nodes. The resultingretrieval of the parent will produce a positive spillovereffect. In this manner, exposure to advertising of Yoplaitnon-fat yogurt may activate the parent category, regularyogurt, through activation of the Yoplait name because theYoplait node is strongly linked to the regular yogurt node.Thus, we form the following hypothesis for the reciprocalspillover effect:HI: Both the economies of information theory and the associ-ated network theory suggest that the reciprocal spillovereffect, that is, the spillover effect of exposure to advertis-ing of a child on the revealed preference for a parent, is

positive.Consider the forward spillover effect of parent's adver-tising on the choice of the child. As previously discussed,arguments based on economies of information favor a posi-tive forward spillover effect. Moreover, such arguments donot suggest any asymmetry in the magnitudes of forwardand reciprocal spillover effects. In particular, Wernerfelt(1988) argues that both the parent and child are perceived tobe of high quality as a result of umbrella branding, whichsuggests that the forward and reciprocal effects are equal inmagnitude. In contrast, although the associated network the-ory favors a positive forward spillover effect, it sugge sts that

the forward spillover effect is weaker than the reciprocalspillover effect.^ In particular, exposure to the parent'sadvertising will activate the brand node, whence activationwill spread to the child, with potential for a positivespillover effect. However, because the child is newer, thelink between brand name and child is likely to be weakerthan the link between brand name and parent, especially inthe early stage s of introduction of the child. Activation of thechild node may not always exceed the threshold levelneeded for retrieval from memory. Therefore, exposure toparent's advertising is less likely to activate the child in con-sum ers' mem ories as compared with the effect of the child'sadvertising on activation of the parent. Thus, the associatednetwork theory suggests that the forward spillover effect islikely to be positive but weaker than the reciprocal spillovereffect.

H2: Both the information economies theory and the associatednetwork theory suggest that the forward spillover effect,that is, the spillover effect of exposure to advertising of aparent on the revealed preference for a child, is positive.H3: If the theory of information economies holds, both forwardand reciprocal spillover effects will be equal in magnitude.

H4: If the associated network theory of consumer memorholds, the forward spillover effect will be weaker than threciprocal spillover effect.

Empirical ModelModel SpecificationsWe use the multinomial logit model to study the impact oadvertising and other marketing variables on consumechoice.** The logit mod el h as been extensively applied imarketing literature (e.g., Guadagni and Little 1983Kanetkar, Weinberg, and Weiss 1992; Krishnamurthi anRaj 1991; Tellis 1988). With this model, the probability thahousehold i chooses brand j on choice occasion t is given b(1)

where J is the number of products, and Ujj, is the (revealedindirect utility of household i for product j on choice occasion t. We use two formulations of Ujj, here. In the first formulation, we use brand loyalty to capture unobserved heerogeneity in preferences across househ olds as do Gu adagnand Little (1983) and Kanetkar, Weinberg, and Weis(1992). Thus, Uijt is given by(2)

In this equation, Oj is the intercept for brand j , and Xjjis the value of the explanatory variable k for household i anproduct j on choice occasion t. The parameter Pi^ is thunknown coefficient of explanatory variable k that is to bestimated, and ejj, is the random error that follows aextreme value distribution. The hou sehold's brand loyalty one of the m explanatory variables in Equation 2 and is computed from household purchase history as described bGuadagni and Little (1983). Thus, brand loyalty Ljj, fohousehold i toward brand j at purchase occasion t is given b(3) L , , = a L . , , - Klwhere Yy, _ 1 = 1 if brand j is purchased at purchase occsion t - 1 and 0 otherwise, and a is the smoothing constaFollowing Gupta (1988), we use a = .8.

Recent research suggests that the previous brand loyaltmeasure, which has traditionally captured household heterogeneity in logit models, biases parameter estimates becausof correlation of the measure with the error term (e.g., Chintagun ta, Iain, and Vilcassim 1991 ; Gon ul and Srinivasa1993). Thus, we use a second formulation of utility that catures heterogeneity in intrinsic preference across householdusing the latent class approach proposed by Kamakura anRussell (1989). Given that recent research by AilawadGed enk , and Neslin (1999) finds little difference in the estmated response elasticities across different methods o

''We thank an anonymous reviewer for pointing out the differentpredictions of the two theories regarding the relative magnitudes ofthe forward and reciprocal spillover effects.focus on brand choice because, using coffee data, Gup(1988) finds that brand-switching accounts for 84% of the overasales increase due to promotions.


4/11

incorporating heterogeneity, the Kamakura-Russell (K-R)approach has the advantage of being computationally lessburdensome. Thus, we use a second formulation.(4 )

k = l

In this equation, Ujsjt is the indirect utility to householdi for product j on choice occasion t, and household i belongsto latent household segment s, s = 1, 2, ..., M. The parame-ter ajj is the intrinsic preference of household segment s forbrand j and replaces the brand loyalty measure in this model.Other terms in Equation 4 have similar interpretations tothose in Equation 3. We determine the parameters and theproportion of the latent segments fj using a maximum like-lihood procedu re. The number of segm ents, M, is selected tominimize Schwarz's (1978) Bayesian information criterion(BIC).DataWe estimated the previous models using the ACNielsenscanner panel data for two product-markets: yogurt in theSpringfield, Mo., market and powdered detergents in theSioux Falls, S.Dak., market. Although household purchasedata in this data set are available between Janu ary 1986 andAugust 1988, the data used for estimation pertain to theperiod between September 1987 and August 1988, duringwhich household exposure to advertising was recorded.However, we used purchase data before the estimationperiod to calibrate loyalty of samp le households.

Choosing brand or line extensions for analysis involvessome judgment. At the extreme, each item with a distinctUniversal Product Code can be considered a brand or lineextension. Thus, we used prior knowledge to identify signif-icant extensions (from a consumer perspective) on which tofocus. For example, in yogurts, prior knowledge suggeststhat extension of a brand into the non-fat yogurt subcategoryis important. As another criterion, we considered an exten-sion significant if it was advertised separately. After weidentified significant extensions, we considered a purchaseof any U niversal P roduct Code for eaeh brand a purchase ofthat brand. We ignored brand sizes for reasons similar tothose offered by Kanetkar, Weinberg, and Weiss (199 2).

First, advertising focuses on brands and not package sizes.Second, package size decisions are not likely to bepurchase-to-purchase decisions according to the literature(;Blattberg, Eppen, and Lieberman 1981; Krishnamurthi andRaj 1988; Tellis 1988 ).For purposes of manageability, we restricted the study tobrands with share of purchases exceeding 5% and theirextensions. The resulting set yielded nine and six brands,

respectively, in the yogurt and powdered detergent cate-gories. For yogurt, spillover advertising came from lineextensions w ithin the category. Thu s, there was advertisingfrom line extensions in the Dannon and the Yoplait familiesof brands. However, for powdered detergents, spilloveradvertising came from liquid detergents. In this case, therewas advertising from brand extensions in the liquid deter-gent category for the Cheer, Surf, and Tide brands. Table 1identifies the selected brands and provides descriptive sta-tistics, advertising exposures, and introduction dates foryogurts. Table 2 provides the same information for detergentbrands. From households that purchased only the selectedbrands, we removed light-user households, which we defineas those that made less than eight purchases in the entireperiod for which purchase data were available or less thanthree purchases in the estimation period. After removingpurchases used for loyalty calibration, in the yogurt categorywe were left with 157 households that accounted for 1674purchases in the estimation period. The corresponding num-bers in the powdered detergents category were 163 house-holds and 1336 purchases.Explanatory Variables

Own advertising: This variable captures the effect of abrand's own advertising on its utility to the household. Sim-ilar to Tellis and Weiss (1995), we formulated advertising asa stock variable that is an exponentially weighted average ofpast exposures and current exposu res. The current exposure,Ajj t of household i to product j's own advertising at pur-chase occasion t, was measured by the number of televisionadvertisements that a household was exposed to in the timeperiod between the previous purchase and the current pur-chase occasion. This definition of current household adver-tising exposure is similar to that used by Kanetkar, Wein-berg, and Weiss (1992). Also, similar to Kanetkar, Weinberg,

TABLE 1Yogurt Brands: Descriptive StatisticsYogurtProductDannon low-fat (prior to 1981)**Dannon non-fat (1987)Dannon fresh flavors (around 1985)Dannon mini-pack (1985)Nordica low-fatWells-Bunny low-fatWeight Watcher'sYoplait non-fat (1986)Yoplait (1976-77)

Share(Percentage)11.771.973.58.1218.518.1811.419.8034.65

Price perOunce(Dollars).084.074.081.133.065.053.076.106.100

Display*0000 .0305.0203.0024.0018.0024

Feature*.0230 .0210 .220.056.030.019.035

Number ofAdvertisingExposures1540366826634429 6121

"Proportion of purchase occasions in estimation-period data."N um be rs in parentheses indicate estimated introduction dates of the brands.


5/11

TABLE 2Detergent Brands: Descriptive StatisticsDetergentBrandBold powderCheer powder (1950)Oxydol powderPurex powderSurf powder (before 1985)Tide powder (1946)

Share(Percentage)10.462.04

8.948.7610.1559.64

Price perOunce(Dollars).052.049.053.030.047.040

Display*.0082.0239.0052.0449.0284.0861

Feature*.0067.0247.0060.0254.0419.1871

Number ofAdvertisingExposures

8211195172

'Proportion of purchase occasions in estimation-period data.Notes: Number of advertising exposures for Cheer iiquid, Surf liquid, and Tide liquid is 43, 39, and 28, respectively. Numbers in parentheses inthe first column indicate estimated introduction dates of the brands. Estimated introduction dates for Cheer liquid. Surf liquid, and Tideliquid were 1986, 1988, and 1984, respectively.

and Weiss, we considered a household to have been exposedto an advertisement if the television set was tuned to theadvertisement for more than half the advertisement's dura-tion. Let ASjj t be the stock of own-advertising exposures forproduct j for household i at purchase occasion t. As in Tellisand W eiss's (1995) study, ASy [ is given by

The element ASjjt _ i is the advertising stock for productj , house hold i, and purchase occasion t - 1. The param eter Xis a smoothing constant to be determined.Spillover advertising:We examined the order of entry ofproducts that share the same brand name and designated theproduct that originally used the brand name as the parent.We inferred introduction dates of products on the basis ofthe earliest date the prod uct occurred in the data set. In cases

in which the products were available from the beginning ofthe data collection period, we consulted trade magazines toascertain the order of introduction of products. Tables 1 and2 give estimated introduction dates of umbrella-brandedproducts.In all cases, the parents identified in this manner appearto be the product that is most closely associated with thebrand name, consistent with our theoretical perspective. Therelatively high market shares of the parent in all cases alsosupports this view. For example, in the Dannon brand fam-ily, Dannon low-fat yogurt was deemed the parent, and Dan-non mini-pack, Dannon non-fat, and Dannon fresh flavors asits children. In the other case of multiple products with a

common brand name in the yogurt data, Yoplait yogurt wasconsidered the parent of Yoplait non-fat yogurt. The remain-ing three products in yogurt data are stand-alone productswith no parent or children. In the powdered detergent data,none of the six brands had significant extensions in the samecategory; however, some benefited from spillover advertis-ing of branded relatives in the liquid detergent category. Inthis case, all powder brands were considered parents. Forexample. Tide Liquid is the child of Tide Powder (theparent).In product j's utility function, a parent spillover adver-tising variable measured stock of exposures of household i

to advertising of the parent of product j at purchase occasiont. Similarly, a child spillover exposure variable in product j ' s

utility function measured stock of advertising exposures oproduct j's children for household i at purchase occasion tIn both cases, the stock variable was an exponentiallyweighted average of past stock and current exposures as inEquation 5 with the same smoothing constant X as was usedfor own advertising.

Although the spillover advertising variables capturespillover effects in our model, the own-advertising variablecaptures the substitution effect from a brand's advertisingThus, we separately estimate the substitution and spilloveeffects by including both advertising variables in our modeand using household scanner-panel data. In contrast, Sullivan (1990) could not distinguish between these two effectin her measurement of spillover effects: Her data show theaggregate impact of both effects.Consumer sales promotion: We employed two separat

0/1 (absence/presence) variables. They were features and instore displays.List price: As in Tellis and Weiss's (1995) study, thivariable represents the price before coupons for the producat purchase occasion t and is expressed in dollars per ounceCoupon value: The data provide information on thvalue of coupons redeemed with purchase, but they have noinformation on coupon availability to households. As dGonul and Srinivasan (19 93), we assume that if a m anufacturer's coupon was redeemed for a brand in a particulaweek, it is available to all households during that week. Fostore coupons, we assume that if a store coupon wa

redeemed for a brand in a week, that coupon was availableto all households shopping that store during that week.Brand loyalty: When we estimated the model given byEquation 2, we used brand loyalty given by Equation 3 as anexplanatory variable. We used the first five purchases of ahousehold or all of a household's purchase before the estimation period (whichever was larger) for calibrating the loyalty measure for each household.

Analysis and ResultsAs do Tellis and Weiss (1995), we estimated different logimodels using values of the advertising carryover parameteX that varied from 0 to I in increments of. I. We found tha


6/11

a value of .5 for X provided the best fit on the basis of theBIC for both yogurt and detergent. Thus, we set X equal to.5 throughout our analysis.''Yogurt DataIn Table 3, we present the estimation results for a sequenceof models leading up to the full model, M2, containingbrand loyalty, display, feature, list price, coupon value, ownadvertising, spillover advertising from children (child-advertising), spillover advertising from parent (parent-advertising), and the brand dummies (Guadagni and Little1983).6 List price, coupon value, brand loyalty, display, andfeature are strongly significant in all models and have theexpected signs. Consistent with the scanner literature onadvertising effects (e.g., Tellis and Weiss 1995), own adver-tising is not significant in the different models. The coeffi-cient of child-advertising is relevant to us in the context ofH|. In both Ml and M2, the coefficient of child-advertisingis positive and strongly significant {p < .006). Thus, theresults suggest the existence of a significant and positivereciprocal spillover effect.

As can be seen from M2, the coefficient of parent-advertising is not statistically significant, thus H2 is not sup-5A long stream of literature suggests that the effect of advertis-ing repetition produces a greater response (e.g., attitude, purchaseintention, sales) among consumers loyal or familiar with the adver-tised brand than among consumers who are not (Calder and Stern-thai 1980; Raj 1982; Sawyer 1973; Tellis 1988). In our analysis, wetested for such effects by including loyalty x advertising variablesfor both own advertising and spillover advertising. None of theinteractions was significant, so we present the results without theseinteractions. Another stream of literature suggests that advertisinghas an indirect effect on utility through its effects on price-sensi-

tivity (e.g., Kanetkar, Weinberg, and Weiss 1992; Krishnamurthiand Raj 1985). Similar to Kanetkar, Weinberg, and Weiss (1992),we m odeled this effect using a list price x own-advertising interac-tion term. We found that the interaction was not significant andhave chosen to present the results without these interaction terms.*We omit the brand intercept parameters in the presentation ofall results to conserve space.

ported. This result, combined with the finding of a signifi-cant, positive reciprocal spillover effect, offers some supportfor the associated network theory hypothesis, H4. Althougha one-tailed significance test of the difference in the coeffi-cients of parent-advertising and child-advertising in M2(using the estimated variance-covariance matrix for the vec-tor of coefficients) rejects the null hypothesis of no differ-ence in the coefficients (p < .06), a two-tailed test margin-ally fails to reject the null hypothesis (p = .11). Given thedirectional prediction of H4, we believe that a one-sided testis more appropriate in this case. Overall, given the lack ofsignificance of parent-advertising in comparison with thestrong significance of child-advertising, and given the previ-ous results of the one-sided test, we conclude that our resultsweakly favor the associated network theory hypothesis, H4,over the information economies hypothesis, H3.7 The p2values in Table 3 suggest a good fit of the models. We per-formed predictive validation on a holdout sample as a fur-ther test of a model with spillover effects. In Table 3, thelikelihood ratio test rejects model MO in favor of model Ml(X^ < .01 ), though we fail to reject model M l in favor ofmodel M2 . Thus, m odel M1 seems to offer the best fit to thedata. To test further the appropriateness of this model, weused mo dels MO and M1 calibrated on a randomly selectedestimation sample of 110 households to make market sharepredictions in the remaining separate holdout sample of 47households. (The results from the estimation sample ofhouseholds were similar to those of the overall sample:Mod el M l was favored over model MO and child-

^This conclusion is not critical to our central finding that recip-rocal spillover effects exist. If the null hypothesis, H3, is insteaddeemed to be supported by virtue of the two-sided test, we wouldinfer that the forward spillover effect is of comparable magnitudeto the reciprocal spillover effect, thereby supporting the informa-tion of economies hypothesis. Thus, the outcome of the compari-son between parent-advertising and child-advertising is mainly oftheoretical interest.

TABLE 3Brand Loyalty ModelsYogurt BrandsParameter Estimates (t-Ratio)VariablesBrand loyaltyDisplayFeatureList priceCoupon valueOwn advertisingChild-advertisingParent-advertisingLL(n = 1674)p2*X2 ( -2L L) "BIC

Model4.70271.0813.5863-31.59154.978.1515

-1366.9.5302733.81418.9

MO(38.750)(3.478)(3.698)(-6.342)(8.876)(1.178)

Mode l4.72011.1118.5773-31.97855.451.1625.3971

-1363.1.5312726.21418.7

M l(38.678)(3.571)(3.641)(-6.408)(8.938)(1.147)(2.996)

Model4.71601.1090.5782-31.91255.388.1852.3775- .3698-1362.7.5312725.41422.1

M2(38.645)(3.565)(3.648)(-6.397)(8.929)(1.305)(2.752)(-.782)

*p2 values are with respect to model with brand intercepts only."Li ke lih oo d ratio test rejects model MO in favor of model M l (p < .01), where as mod el M l is not rejected in favor of model M2.Notes: LL = log-likelihood.


7/11

advertising was significant.)* Marketing literature has cus-tomarily characterized the accuracy of market share predic-tions using measures such as root mean square error(RMSE) and mean absolute deviation (MAD) (e.g., Mont-gomery 1997). The RMSE penalizes larger deviations inpredictions more than MAD does. The RMSE between pre-dicted and actual market shares in the holdout sample dete-riorates by 15.2% to a value of .811 with model MO, from avalue of .704 obtained with M l. The M AD between pre-dicted and actual market shares in the holdout sample dete-riorates by 15.03% to a value of .680 with model MO, froma value of .591 obtained w ith M l. A chi-square test of thedifference in market share predictions between the twomo dels MO and M l rejects the hypo thesis that there is nodifference between the two predictions (x^ with 8 degrees offreedom [d.f.] = 25.157, p < .005), showing that theimprovement in prediction is statistically significant. Wealso compared disaggregate predictions of mod els MO andM l in the validation samp le at the individual choice levelusing hit rates as suggested by Gensch (1987). To computethe hit rates, we defined the predicted choice on each pur-chase occasion as the alternative with the highest predictedprobability based on the model (MO or Ml ). We then com-pared the predicted choice with the actual choice for eachpurchase occasion to obtain a hit rate (percentage of choicescorrectly predicted). Computed in this manner, the hit rateImproved from 73.3 for model MO to 74.0 for m odel M1. Wefound the imp rovement in disaggregate prediction of m odelM1 over MO to be statistically significant using the Krishnantest of strength of prediction at the individual choice level(j? < .007) (for additional details on the Krishnan test, seeGensch 1987). Thus, the predictive results further supportthe model with reciprocal spillover effect. M l. U sing stan-dard tests from the literature, we eliminated alternativeexplanations for our results such as collinearity, influentialobservations, or o utliers.

The K-R model estimated on yogurt data yielded fivepreference segments on the basis of minimization of theBIC. We present the results of the full model here. The esti-mated parameters in Table 4 reveal that child-advertising ispositive and significant (p = .03), whereas parent-advertising is not significant. In contrast to the brand loyaltymodels, own advertising is significant (p = .04) in the K-R*ln Tables 3 and 5, we present the estimation results using datafrom all households to maintain comparability with the K-R mod-

els, in which the additional data were helpful, considering the largenumber of parameters estimated with those models.TABLE 4K-R Heterogeneity ModelYogurt Brands

VariablesDisplayFeatureList priceCoupon valueOwn advertisingChild-advertisingParent-advertising

Parameters.9215.4656-40.809960.2923.2478.3120

- .2840

t-Ratio3.2293.432-8.0699.8832.0142.152- .671

model. Because the K-R model avoids potential bias inparameter estimates that are possible with the brand loyaltymodel, it appears that own-advertising effects may be sig-nificant. Overall, it is reassuring that the results on spilloveradvertising from the K- R model are consistent with those othe brand loyalty model.Detergent DataResults of the brand loyalty model are presented in Table 5The parent-advertising effect was not estimated with thisdata because we modeled the choice of powdered detergentalone (there were too few observations of liquid detergenchoices for us to model their purchase). Child-advertising isagain positive and significant in this category (p < .002)whereas own advertising is not significant. The likelihoodratio test rejects model MO in favor of model Ml (x^< .005)Similar to the procedure for yogurt, we used m odels MO andMl calibrated on a randomly chosen estimation sample of100 households to make market share predictions in theremaining separate holdout sample of 63 households. (Theresults from the estimation sample of households were similar to those of the overall sample: Mod el M l was favoredover model MO and child-advertising was significant.) TheRMSE between predicted and actual purchase shares in theholdout sample deteriorates by 12.2% to a value of .852 withmodel MO, from a value of .759 with model M l. Th e MA Dbetween predicted and actual market shares in the holdousample deteriorates by 10.7% to a value of .613 w ith m odeMO, from a value of .554 obtained with M l. A chi-squaretest of the difference in mark et share predictions betweenthe two models MO and M l rejects the hypothesis that thereis no difference between the two prediction s (x^ with 5 d.f. =13.376, p < .025), showing that the prediction improvemenis statistically significant. Disaggregate predictions fodetergents, as measured by the hit rate, improved from 83.2for model MO to 84.1 for model M l. We found the improvement in disaggregate prediction of model M I over MO to bstatistically significant using the Krishnan test of strength ofprediction at the individual choice level (p < .001). Thus, thprediction results affirm the model with (reciprocal

TABLE 5Brand Loyalty Mod elsDetergent BrandsParameter Estimates (t-Ratio)VariablesBrand loyaltyDisplayFeatureList priceCoupon valueOwn advertisingChild-advertisingLL(n = 1336)X2 ( -2LL)"BIC

iUlodel4.8871.6629.9115-92.652153.148.1133

-551.5.6921103591.1

MO(26.583)(6.104)(3.561)(-8.802)(10.217)

(.568)

Model4.9271.6463.9452-92.540155.385- .07611.7804-547.4.6941094.8590.6

M l(26.486(6.034(3.686(-8.759)(15.103(-343(3.137

*p2 values are with respect to model with brand intercepts only."Lik elih oo d ratio test rejects model MO in favor of model M l (p


8/11

spillover effects as being the most consistent with the data.Again, using standard tests we eliminated alternative expla-nations for our results such as collinearity, influential obser-vations, or outliers.The K-R model estimated on these data yielded sevenpreference segments by means of the BIC. Parameter esti-mates of the K -R model are presented in Table 6. Consistentwith results of the brand loyalty model, child-advertising isonce again positive and significant (j? = .03), whereas ownadvertising is not significant.

DiscussionManagerial ImplicationsWe estimated spillover effects in two different product cate-gories and geographic markets, using two different ways ofrepresenting unobserved heterogeneity. Our results providestrong and consistent support to the hypothesis of a positivespillover effect from advertising of a child on choice of aparent brand (reciprocal spillover effect). Thus, exposure toadvertising of Yoplait non-fat yogurt, for example, had apositive effect on the choice prob ability of its parent, Yoplaityogurt, by househ olds. However, we did not find evidence tosupport the existence of forward spillover effects (i.e.,advertising of a parent increasing the choice probability of achild). This result is consistent with the prediction of asso-ciated network theory that forward spillover effect would beweaker (H4). Statistical testing weakly supports this hypoth-esis while contradicting the expectation of symmetricspillover effects based on information economies (H3). Theprincipal theoretical rationale for a weaker forward spillovereffect was that parent's advertising might not evoke the childin the consumer's mind as much as the child's advertisingwould evoke the parent.Consistent with previous studies (Kanetkar, Weinberg,and W eiss 1992; Tellis 1988; Tellis and Weiss 1995), theeffect of own advertising is weak or nonexistent. In particu-lar, own advertising has a positive, significant effect foryog urt in some mo dels, but has no significant effect in deter-gents. The significance of reciprocal spillover effects whenown-advertising effects are weak or nonexistent can beascribed to both the newness of child-advertising and theflagship nature of the parents that benefit from the spillover.Because of their flagship position for the umbrella brand,parents can gain considerably, if not the most, from the newsvalue and interest generated by the advertising of the child(Aaker 1996). Indeed, the newness and freshness of child-advertising may make such advertising more effective in

TABLE 6K-R Heterogeneity ModelDetergent BrandsVariables Parameters t-RatioDisplayFeatureList priceCoupon valueOwn advertisingChild-advertising

1.4875.9402-136.0419166.0217.09591.4658

5.1793.553-12.16117.020.4962.089

garnering attention for the parent than the parent's ownadvertising. In summary, spillover effect from a child maybe of equal or greater impo rtance than a paren t's own adver-tising. A key managerial implication of this finding concernsthe allocation of advertising spending between the parentand its children. To the extent that the parent benefits fromthe advertising of its children and to the extent that suchspillover advertising may be more effective in increasing thechoice share of the parent, less advertising money may beallocated to the parent.

The productivity or effect size of an advertising expo-sure can be best captured by the brand-choice elasticities,which measure the increase in choice probability (purchaseshare) that results from increase in exposure. We follow themethod used for computing choice elasticity for featureadvertising in the literature (Chintagunta 1993; Chintagunta,Iain , and Vilcassim 1991) by calculating advertising (own orspillover) elasticity as the fractional relative change in theprobability of purchase due to an advertising exposure. Asan illustration of the effect size comparison, in Table 7 wepresent the relative productivity of a parent brand's ownadvertising and spillover advertising of a child, for the twoparent yogurt brands, D annon low-fat and Yoplait. We com-puted the elasticities using parameters from the brand loy-alty model.5* Althou gh the ow n-adv ertising effect was no tsignificant in the brand loyalty model, we have chosen topresent the point elasticity for own advertising for compari-son. For interested readers, we also provide the 95% confi-dence intervals for the choice elasticities.

As an example, the choice elasticity for Dannon low-fat's own advertising in Table 7 should be interpreted as fol-lows: One exposure to Dannon low-fat advertising results inan average increase of 5.7% in its choice probability. Thetable shows that spillover advertising from its children hasan impact on purchase share of Dannon low-fat that is morethan twice the impact of Dannon low-fat's own advertising.In the case of Yoplait, spillover advertising from Yoplaitnon-fat has nearly twice the impact on purchase share ofYoplait, as does Yoplait's own advertising. Therefore, opti-

'Ailaw adi, Gedenk, and Neslin (1999) find that different hetero-geneity formulations cause little change in elasticity estimates.

TABLE 7Productivity Comparison of Own and SpilloverAdvertising for Yogurts

Choice Elasticity ofParent Brand

Parent BrandParent'sOw nAdvertising

SpilloverChild-AdvertisingDannon low-fatYoplait

.057[-.046, .171].041[-.027, .093]

.144[.043, .255].079[.017, .144]Notes: 95% confidence intervals are in brackets.


9/11

mal allocation of advertising spending may favor the lineextensions in the case of both Dannon and Yoplait.'^Analogous productivity comparisons for detergentadvertising (Table 8) yield conclusions similar to that forYoplait advertising. The main difference between the deter-gent and the yogurt results is the somewhat higher magni-tude of the reciprocal spillover elasticities for detergents. Webelieve that this outcome may be due to the presence offewer brands in the detergent category, which makes adver-tising expo sures m ore productive. However, consistent withthe yogurt results, the reciprocal spillover has a greaterimpact on the purchase share of parents than the respectiveparent's own advertising, for all three extended brands ofpowdered detergents. The apparent managerial implicationis that these brands are better off devoting relatively moreadvertising spending to the newer brand extensions. M orrin(1999) reaches the opposite conclusion that shifting ofadvertising funds to extensions may hurt the parent. How-ever, she studies the effect of brand extension advertising onrecall and recognition of the parent rather than its choice. Itis possible that advertising of the child may change beliefsabout the parent and thus influence its overall evaluation andchoice, without affecting recall or recognition. For example,advertising of the extension may increase perceptions ofparent quality (Dacin and Smith 1994; Wernerfelt 1988) orparent innovativeness. In such a case, a strong reciprocalspillover effect on a parent's evaluation may more than com-pensate for a weaker spillover effect on recall of the parent(as Morrin finds), resulting in a overall spillover elasticitythat is greater than the parent's own-advertising elasticity (aswe find). The different conclusions between our studies mayalso be due to the attention to advertising that was forced onsubjects in Morrin's laboratory study, while it is possiblethat advertising of newer brand extensions command greaterattention from the audience in real-world settings (Pieters,Rosbergen, and Wedel 1999).

The existence of beneficial spillover effects from brandextensions suggests an additional strategic benefit of intro-'OA more formal analysis of the implication of spillover effectsfor advertising allocation, using an analytical model, can beobtained directly from the authors on request.

TABLE 8Productivity Compa rison of Own and SpiiioverAdvertising for Detergents

Parent BrandCheerSurfTide

Choice Elasticity ofParent BrandParent'sOwnAdvertising

n.s.n.s.n.s.

SpilloverChild-Advertising.600[.160, 1.273].893[.236, 1.956].145[.056, .229]

Notes: 95% confidence intervals are in brackets, n.s. = not slgnifi-canfly different from zero.

ducing line or brand extensions (Aaker 1990). This strategibenefit may be distinguished from and may supplement thcommonly advanced rationale for line or brand extensionswhich is to satisfy new market segments profitably by leveraging existing brand equity. Proliferation of brand extensions may have another strategic benefit, which is to crowthe product space and deter entry (Schmalensee 1978).The information in Tables 7 and 8, when combined witthe introduction dates from Tables 1 and 2, yields the following interesting observation: The superiority of reciprocaspillover advertising in comparison with a brand's owadvertising is greatest for recently introduced brand extensions and, at least within a given product category, tends tdiminish more or less monotonically with age of the extension. This conclusion is consistent with the intuitive expectation that with passage of time, the child's advertising maprovide less by way of new information that may influencthe choice of the parent. Also, as time passes, the child magain a stronger presence in the consumer's mind and be leslikely to evoke the parent with its advertising. The associated network theory predicts that such inhibition of the par

ent in consumers' memories will result from a strengtheninof the link between the brand and child nodes. A caveat tour previous conclusion regarding the relation between thspillover effect and age of the child is that it comes fromcross-brand comparison and not a within-brand comparisonFurther research might examine this issue with more extensive longitudinal data. Overall, it is important to note thatpositive and statistically significant reciprocal spilloveeffect was observed over a wide variety of relative introduction times of the parent and child in the two categories.Methodological Advantages and Limitations ofthe StudyThe advantage offered by the high quality of scanner panedata is well known. However, the household-level data alsohelp mitigate the problem of reverse causality faced withadvertising studies using aggregate data. This problemarises because managers might set advertising budgets as percentage of sales or market share. As Tellis (1988 , p. 142points out, "it is unlikely that managers set advertisingexposures in expectation of purchases at such disaggregatlevels" (individual household levels). Tellis also indicatethat between-brand analysis with such data provides greatepower and efficiency than a separate analysis for each brandAs with typical econometric studies, the conclusionfrom our study should be considered tentative subject treplication by other studies using different models omethodologies or by replication in other product-marketsFurthermore, our results may be sensitive to the level oadvertising by competing brands observed in the categorieand markets we examined." Although we used a variety otests to verify that a few influential observations did noaffect the pattern of our findings, it would nevertheless bworthwhile to replicate our analysis with categories havina greater number of advertising exposures. Future studie

"We thank an anonymous reviewer for pointing out thilimitation.


10/11

could also explore how advertising content affects spillovereffects.

ConclusionsIn this article, we study the reciprocal spillover effect ofadvertising of a line or brand extension on the choice of theparent brand. Using scanner panel data on two product cat-egories and in two geographical markets, we find evidence

for a significant recip rocal spillover effect. Indeed , we findthat such spillover advertising can increase the choice prob -ability of the parent more than is possible with the parent'sown advertising. Our results suggest that firms should favorthe line or brand extension with a greater allocation of theadvertising budget than otherwise. These results also indi-cate a new strategic benefit from line or brand extensionswhereby a firm introducing the extensions can expect posi-tive reciprocal spillover effects for the parent brand.

REFERENCESAaker, David A. (1990), "Brand Extensions: The Good, the Bad,and the Ugly," Sloan Management Review, 31 (Summer), 47-5 6.(1996), Building Strong Brands. New York: The FreePress.Ailawadi, Kusum L., Karen Gedenk, and Scott A. Neslin (1999),"Heterogeneity and Purchase Event Feedback in Choice Mod-

els: An Empirical Analysis with Implications for Model Build-ing," International Journal of Research in Marketing, 16 (Sep-tember), 177-98.Anderson, John R. (1983), "A Spreading Activation Theory ofMemory," Journal of Verbal Learning an d Verbal Behavior, 22(3), 261-95.Blattberg, Robert D ., Gary D . Eppen, and Joshua Lieberman (1981),"A Theoretical and Empirical Evaluation of Price Deals for Con-sumer Nondurables," Journal of Marketing, 45 (Winter), 116-29.Calder, Bobby J. and Brian Stemthal (1980), 'Television Commer-cial Wearout: An Information Processing View," Journal ofMarketing Research, 17 (May), 173-86.Chintagunta, Pradeep K. (1993), "Investigating Purchase Inci-dence, Brand Choice and Purchase Quantity Decisions ofHouseholds," Marketing Science, 12 (2), 18 4-208., Dipak C. Jain, and Naufel J. Vilcassim (1991), "Investi-gating Heterogeneity in Brand Preferences in Logit Models forPanel Data," Journal of Marketing Research, 28 (November),417-28.Dacin, Peter A. and Daniel C. Smith (1994), 'The Effect of BrandPortfolio Characteristics on Consumer Evaluations of BrandExtensions," Journal of Marketing Research, 31 (May), 229-42.Erdem, Tulin (1998), "An Empirical Analysis of Umbrella Brand-ing," Journal of Marketing Research, 35 (August), 339-5 1.Farquhar, Peter H., Paul M. Herr, and Russell H. Fazio (1990), "ARelational Model for Category Extensions of Brands," inAdvances in Consumer Research, Vol. 17, Marvin E. Goldberg,Gerald Gorn, and Richard W. Pollay, eds. Provo, UT: Associa-tion for Consumer Research, 856-60.Gensch, Dennis H. (1987), "Empirical Evidence Supporting theUse of Multiple Choice Models in Analyzing a Population,"Journal of Marketing Research, 24 (May), 197 -207.Gonul, Fusun F. and Kannan Srinivasan (1993), "Modeling Multi-ple Sources of Heterogeneity in Multinomial Logit Models:Methodological and Managerial Issues, Marketing Science, 12(3), 213-29 .Guadagni, Peter M. and John D.C. Little (1983), "A Logit Modelof Brand Choice Calibrated on Scanner Data," Marketing Sci-ence, 2 (Summer), 203-38.Gupta, Sunil (19 88), "Impact of Sales Promotions on W hen, What,and How Much to Buy," Journal of Marketing Research, 25(November), 342-55.Herr, Paul M., Peter H. Farquhar, and Russell H. Fazio (1996),"Impact of Dominance and Relatedness on Brand Extensions,"Journal of Consumer Psychology, 5 (2), 135-59.John, Deborah Roedder, Barbara Loken, and Christopher Joiner(1998), 'The Negative Impact of Extensions: Can Flagship Prod-ucts Be Diluted?" Journal of Marketing, 62 (January), 19-32.Kamakura, Wagner A. and Gary J. Russell (1989), "A ProbabilisticChoice Model for Market Segmentation and Elasticity Struc-ture," Journal of Marketing Research, 26 (November), 379-90.

Kanetkar, Vinay, Charles B. Weinberg, and Doyle L. Weiss (1992),"Price Sensitivity and Television Advertising Exposures: SomeEmpirical Findings," Marketing Science, 11 (4), 359-71.Keller, Kevin (1993), "Conceptualizing, Measuring, and ManagingCustomer-Based Brand Equity," Journal of Marketing, 57 (Jan-uary), 1-22.and David A. Aaker (1992), 'The Effects of SequentialIntroduction of Brand Extensions," Journal of MarketingResearch, 29 (February), 35-50.Krishnamurthi, Lakshman and S.P Raj (1985), "The Effect ofAdvertising on Consumer Price Sensitivity," Marketing Sci-ence,! (,1), 1-20.and (1988), "A Model of Brand Choice and Pur-chase Quantity Price Sensitivity," Marketing Science, 7 (1),1-20.and (1991 ), "An Empirical Analysis of the Rela-tionship Between Brand Loyalty and Consumer Price Elastic-ity," Marketing Science, 10 (Spring), 172-8 3.Montgomery, Alan L. (1997), "Creating Micro-Marketing PricingStrategies Using Supermarket Scanner Data, Marketing Sci-ence, 16(4), 315-37.Morein, Joseph A. (1975), "Shift from Brand to Product Line Mar-keting," Harvard Business Review, 53 (September-October),56-64.Morrin, Maureen (1999), 'The Impact of Brand Extensions on Par-ent Brand Memory Structures and Retrieval Processes," Journalof Marketing Research, 36 (November), 517-2 5.Pieters, Rik, Edward Rosbergen, and M ichel Wedel (1999), "VisualAttention to Repeated Print Advertising: A Test of Scanpath The-ory," Journal of Marketing Research, 36 (November), 424-38.

Raj, S.P (1982), "The Effects of Advertising on High and LowLoyalty Consumer Segments," Journal of Consumer Research,9 ( 0 , 7 7 - 8 9 .Sawyer, Alan (1973), "The Effects of Repetition of Refutationaland Supportive Advertising Appeals," Journal of MarketingResearch, 10 (February), 23-33.Schmalensee, Richard (1978), "Entry Deterrence in the Ready-to-Eat Breakfast Cereal Industry," Bell Journal of Economics, 9(2), 305-27.Schwarz, Gideon (1978), "Estimating the Dimension of a Model,"The Annals of Statistics, 6 (2), 461-64.

Smith, Daniel C. (1992), "Brand Extension and Advertising Effi-ciency: What Can and Cannot Be Expected" Journal of Adver-tising Research, 32 (November-December), 11-20.Sullivan, Mary (1990), "Measuring Image Spillovers in Umbrella-branded Products," Journal of Business, 63 (3), 309-29.Tellis, Gerard J. (1988), "Advertising Exposure, Loyalty, andBrand Purchase: A Two-Stage Model of Choice," Journal ofMarketing Research, 25 (May), 13444.and Doyle L. Weiss (1995), "Does TV Advertising ReallyAffect Sales? The Role of Measures, Models, and Data Aggre-gation," Journal of Advertising, 24 (3), 1-12.Wernerfelt, Birger (1988), "Umbrella Branding as a Signal of NewProduct Quality: An Example of Signalling by Posting a Bond,"Rand Journal of Economics, 19 (Autumn), 458-66.


11/11

reciprocal spillover effects

Documents

Transcript of reciprocal spillover effects