KAPIL BAWA, SRINI S. SRINIVASAN, and RAJENDRA K. SRIVASTAVA'The measurement of consumers' coupon proneness and the predictionof their redemption behavior is important to the evaiuation of marketers'couponing programs. Although considerable attention has been paid inthe couponing literature to the identification of factors that influence couponusage behavior, relatively little work has been done to develop modelsthat can heip managers predict consumer response to specific couponsand design effective coupon promotions. The authors propose a model ofcoupon redemption that extends previous models of coupon usage byconsidering the joint effects of coupon attractiveness and coupon prone-ness on redemption, and does not require explicit measurement of thesevariables. Empirical application of the model shows that it correctly pre-dicts redemption intentions for nearly 90% of consumers in a holdoutsample and substantially outperforms a logit model that includes tradi-tional measures of coupon proneness, coupon characteristics, and demo-graphics. The proposed model also provides insights into consumerresponse to coupons that are not provided by the logit model. Overall, themodel shows considerable promise as an aid to managers in designingcoupon promotions and developing precision targeting strategies.Coupon Attractiveness and CouponProneness: A Framework for ModelingCoupon RedemptionWith 292 billion coupons distributed in 1995 in the Unit-ed States and approximately 6 hillion coupons redeemed fora total savings of $4 billion (NCH Promotional Services1996). coupons continue to be among tbe most importantpromotional vehicles used today. From a managerial per-spective, predicting consumers' coupon redemption behav-ior is essential to tbe evaluation ol' couponing strategies andthe identification of target segments for coupon promotions.Althougb considerable attention bas been paid in thecouponing literature to the identification of factors tbat in-fluence coupon usage bebavior, relatively little work basbeen done to develop models tbat can belp managers predictconsumer responses to specific coupons and design effectivecoupon promotions.Previous researcb on coupon redemption and usage ex-amines it from two perspectives. One stream of researcb isdevoted to understanding the factors that motivate con-sumers to use coupons and identifying the characteristics ofconsumers wbo are \"coupon prone\" (e.g.. Bawa and Shoe-maker 1987a: Levedabl 1988; Narasimhan 1984; Teel,Williams, and Bearden 1980). Althougb mucb has beenlearned about coupon usage bebavior. one limitation of stud-ies in this area is that they bave defmed coupon pronenesson the basis of observed coupon use without considering thecharacteristics of the coupons available to eacb consumer.Arguably, a person's coupon usage behavior will depend notonly on his or her inherent coupon proneness or desire to usecoupons, but also on the attractiveness of tbe coupons en-countered. For example, a consumer may be inclined to usecoupons but exhibit low coupon usage if he or she fails tofind coupons that are sufficiently attractive {i.e., couponswitb bigb face values or for a preferred brand). Thus, failureto include coupon attractiveness as a predictor of coupon us-age can lead to an inaccurate assessment of coupon prone-ness and an inability to predict bow the consumer would re-spond to coupons with different sets of characteristics.A second stream of researcb focuses on modeling couponredemption rate as a function of tbe cbaracteristics of theJournal oJ Marketing ResearchVol. XXXtV (November 1997). 517-525* Kupil Bawa is A.sstKiaie Professor of Murkcling. Faculty of Manage-ment. McGill Universily. Srini S. Srinivasan is Assisiani Professor of Mar-keting. College of Business and Administration. Drcxcl Universily. Rajen-dra K. Srivastava is Sam Barshop Professor of Marki;ting. University ofTexas al Austin. The order of authorship is alphabetical and rellects equalcontributions by all authors. The authors thank Mark Aipert, University ofTexas at Austin, for supporting this study and the JMR editor and fouranonymous JMR reviewers fur their comments and suggestions.517518JOURNAL OF MARKETING RESEARCH, NOVEMBER 1997Other researcbers have examined the impact of couponcharacteristics on redemption rates. Reibstein and Traver(1982) found that bigher face value coupons and in-packcoupons are associated witb higher redemption rates. Wardand Davis (1978) also reported higher redemption rates forbigber face value coupons and direct mail coupons. Sboe-maker and Tibrewala (1985) and Bawa and Sboemaker(1987b) found that redemption rates increase with face val-ue and that tbe consumers wbo are most likely to redeemcoupons are those wbo are most likely to buy the brand inthe first place. Neslin and Clarke (1987) und Krishna andShoemaker (1992) also reached similar conclusions.These studies do not consider coupon proneness as a pre-dictor of redemption bebavior but nevertbeless provide im-portant insigbts into the factors tbat determine coupon at-tractiveness. Taken togetber, tbese studies suggest tbatcoupon attractiveness varies with (1) tbe coupon face value,(2) the type of coupon or delivery vehicle (e.g., free-stand-ing insert |FSI| or direct mail), and (3) wbetber the couponis for a preferred brand. To tbe extent that tbe coupon has ahigher face value, requires less effort to obtain or use, and isfor a preferred brand, the coupon is likely to be perceived asmore attractive and hence more likely to be redeemed, ce-teris par thus.Prior research also shows that coupon redemption bebav-ior varies witb the product category. Sucb variation mayarise because of category characteristics, sucb as averageprice level, purchase frequency, coupon availability, andbrand loyalty (Bawa and Sboemaker 1987a; Webster 1965).Blaltberg and Neslin (1990) note tbat category redemptionrates vary widely and thus managers must study couponproneness at tbe product category level before formulatingpromotional strategy. We address this issue by estimatingour model separately for each category and obtaining cate-gory-level measures of coupon proneness. Two product andtwo service categories are considered in the analysis.Our review of tbe literature suggests tbat tbere is a needfor developing more general models of redemption bebaviorthat include botb coupon proneness and coupon attractive-ness as predictors. In the next section we present sucb amodel tbat is based on IRT. Tbe model estimates latent orunobserved values of coupon proneness and coupon attrac-tiveness, unlike tbe multi-item measures based on classicalmeasurement tbeory tbat bave been used traditionally. Tbisis advantageous not only because of the difficulty of mea-suring coupon proneness independently of coupon attrac-tiveness, a.s noted previously, but also because it makes tbemodel easier to implement as model estimation requires da-ta only on consumer response to tbe coupon (i.e., redemp-tion intentions or bebavior). For a more comprehensive dis-cussion of tbe relative advantages of IRT and classical mea-surement tbeory measures, we refer the reader to Hamblctonand Swaminatban (1985).A MODEL OF COUPON REDEMPTIONThe basic premise of the model is that consumers have anunobserved tendency io use coupons (termed coitpoti prone-ness). whicb interacts with the intrinsic attractiveness of thecoupons encountered to determine their redemption behav-ior. The response to a given coupon is assumed to varyamong consumers because of variations in coupon prone-ness if coupon attractiveness is constant; similarly, varia-coupon promotion (e.g., Bawa and Shoemaker 1987b; Reib-stein and Traver 1982; Ward and Davis 1978). Althoughthese models have shown .several coupon characteristics,such as face value, to be related to the redemption rate, theaggregate-level nature of the analysis has typically preclud-ed consideration of such individual-level characteristics ascoupon proneness. This makes it difficult to predict howconsumers with different levels of coupon proneness will re-spond to a specific coupon promotion.The foregoing discussion suggests tbat there is a need fordeveloping models of coupon redemption that can capture theinteraction between coupon proneness and coupon attractive-ness and help managers predict how a given consumer willrespond to a specific coupon. The objective of this study is topropose such a model. The mode! is drawn from tbe Item Re-sponse Tbeoretic (IRT) literature and has been used widely ineducational measurement and testing. Used in tbe context ofcoupon promotions, the model enables us to estimate couponattractiveness and coupon proneness as unobserved variablestbat are inferred from consumers' redemption intentions for aset of coupon stimuli, which obviates the need for explicitmeasurement of coupon proneness and coupon attractive-ness. Predictive testing of the model with our data indicatesthat it has a predictive accuracy of nearly 90% in holdoutsamples and significantly outperforms a logit model with tra-ditional measures of coupon proneness, coupon characteris-tics, and demographics.We seek to contribute to the literature on couponing intwo principal ways. First, we extend previous models ofcoupon usage by considering tbe joint effects of coupon at-tractiveness and coupon proneness on redemption and pro-vide a methodological framework that does not require ex-plicit measurement of tbese variahles. Second, we providemanagers with an effective tt>ol for predicting consumer re-sponse to coupons, which can be used for a variety of appli-cations ranging from design of coupon promotions to devel-opment of precision targeting strategies. Given the low re-demption rates and unprofitable nature of coupons in gener-al, the model appears to bave the potential for making an itii-portant contribution to tbe practice of couponing.PRIOR RESEARCHMany researchers have sought to identify tbe cbaracteris-tics of coupon-prone or deal-prone consumers (e.g., Bawaand Shoemaker 1987a; Levedabl 1988; Narasimban 1984;Teel, Williams, and Bearden 1980; Webster 1965) and typi-cally bave measured coupon proneness in terms of the con-sumer's observed (or self-reported) coupon redemption be-bavior. Although tbese studies have contributed much tt) ourunderstanding of coupon usage behavior, tbe measures ofcoupon proneness in tbese studies do not consider tbe cbar-acteristics of tbe coupons encountered and tbus bave limit-ed usefulness as predictors of coupon usage behavior. Be-cause different consumers may be exposed to coupons witbdifferent cbaracteristics. a consumer's observed redemptionbebavior witb respect to specific coupons does not neces-sarily retlect his or bcr underlying coupon proneness ifcoupon attractiveness is not taken into account. Similarly,the observed response to a given coupon among a group ofconsumers does not necessarily reflect the coupon's inher-ent attractiveness if the consumers' coupon proneness is notconsidered.Modeling Coupon Redemptionlions in response across different coupons for a given con-sumer are assumed to be a function of coupon attractivenessif coupon proneness is constant. We use the well-knowntwo-parameter model derived from IRT (Bimbaum 1968;Hambleton and Swaminathan 1985) that has been used inthe context of attitude scaling in the marketing literature(e.g., Balasubramanian and Kamakura 1989). The model es-timates the unobserved coupon proneness and coupon at-tractiveness from stated redemption intentions.' In this mod-el, the probability of a consumer j redeeming a coupon i isrepresented as a logistic funetion of the consumer's couponproneness and coupon characteristics such that- b,)519Figure 1COUPON CHARACTERISTIC CURVES(I)1 + exp[ai(9j -b,)whereP,j = probability that consumer j intends to redeemcoupon i,9j = the (unobserved) coupon proneness of con.sumer j,bj = the (unobserved) attractiveness of coupon i, andaj = the (unobserved) ability of coupon i to discriminateamong consumers with different levels of couponproneness.This model places consumers and coupons in the sameunobserved continuum. Consumers are represented by 9j,which can be interpreted as coupon proneness. The largerthe 6.. the higher the likelihood of redemption for given ajand b,. The coupon \"position parameter\" b, is inversely re-lated to the unobserved coupon attractiveness of coupon i,because a lower bj yields a higher redemption probability fora given 9j. This is illustrated in Figure I, which shows plotsof the function in Equation I lor each of three coupon stim-uli. We term these plots cimpon characteristic cun'es, anal-ogous to \"item characteristic curves\" in the IRT literature.Note that the characteristic curve for the $1 coupon is to theleft of (and therefore represents higher attractiveness than)the characteristic curves for the 75- and 4()-cent coupons.Hence, more attractive coupons—those with smaller valuesof bj—are likely to be redeemed even by consumers withlow levels of coupon proneness. But less attractive couponswith high b,'s will be redeemed only by highly coupon-prone consumers (i.e., consumers with large latent 9j's). Onthe basis of past research, we expect that coupons with high-er face values, those that are easier to use (e.g.. FSI or on-pack coupons, which require less effort to use than mail-incoupons), and those for the buyer's favorite brands will bemore attractive and thus have lower values of bj.The parameter a,, which represents the ability of coupon ito discriminate among consumers with different levels ofcoupon proneness, is defined as the tiiaximum slope of thefunction in Equation 1. The maximum slope of the functionoccurs at the midpoint of the curve shown in Figure 1, whena consumer's coupon proneness 9j is numerically equal to bj.From Equation 1 it can be observed that this occurs when'The mcHlet can be applied to eilher intentions dala or actual redempiiunbehavior. We use intentions data here because our resources did not permitus to conduct a large-scale Held experiment on coupon usage where IheLouptin stimuli eould be maniptilated as needed. We discuss subsequentlythe limitiitioiis due lo the use of intentions data.Coupon Proneness (61the probability of redemption, Pjj, is .50. When the discrim-ination (or slope) parameter aj is equal to zero, the couponcharacteristic curve is represented by a flat, straight line, ir-respective of the coupon proneness of consumers. The high-er the value of aj, the steeper the slope of the curve and thebetter the ability of the coupon to discriminate among con-sumers who have coupon proneness levels above and belowthe midpoint of the curve. For a coupon with a high aj val-ue, a small change in Sj in the vicinity of the midpoint (i.e.,when 8, is close to bj) will result in a large variation in thelikelihood of coupon redemption. Hence, the slope parame-ter aj can be said to capture the discriminating ability of thecoupon.For example, it is easy to see in Figure 1 that though the40-cent coupon is the least attractive (has the largest valuefor bj) and is likely to be redeemed only by consumers withhigh levels of coupon proneness, it has the highest discrim-inatory power, because small changes in 9j around b^ resultin a large change in the probability of redemption from al-most zero to one. Similarly, the 75-cent coupon has the leastdiscriminating ability because its characteristic curve is lessresponsive to changes in coupon proneness.Note that the model only requires data on the dependentvariable (redemption intentions or behavior) to estimatecoupon proneness and coupon attractiveness. In otherwords, as long as a manager has household-level data onwhether a set of coupons was redeemed, he or she can esti-mate coupon attractiveness and coupon proneness. If in ad-dition, information on the characteristics of each coupon(e.g.. face value) is available, the relationship between thesecharacteristics and model parameters also can be analyzed,as we demonstrate subsequently.-METHODOLOGYData for this study were obtained from a survey of gro-cery shoppers in a Southwestern city. Respondents werecontacted at two stores of a major grocery chain in the city.-In some respects the iRT approach to the analysis of coupon preferencesmay appear similar to a tull-prntlle eonjoint analysis. Both approaches arebased on lhe measurement of responses to lull priiflle stimuli and estimalemodel parameters using decomptisitional methods. In conirust to conjoinlanalysis, however. IRT docs nol require stimulus attributes and levels to beknown in advance and accounis lor heterogeneity tbrimgh tbe B, parameter.520JOURNAL OF MARKETING RESEARCH. NOVEMBER 1997scores (e.g., proportion of positive responses to thecoupons). The distrihution of respondents is then divided in-to fractiles or groups whose members are likely to have sim-ilar levels of coupon proneness, and the proportion of eachfractile or group responding to (i.e.. intending to redeem)each coupon is determined for purposes of formulating alog-likelihood function. Parameter estimates of a, and bj arethen obtained through maximum likelihood procedures.Given the individual responses to coupons, the couponproneness level 9j is then reestimated on the basis of the es-timates of aj and bj, and the iterative process is continued un-til it converges.The model was estimated by the maximum likelihoodmethod implemented in MULTILOG (Thissen 1991). LetUjj 0 = '• •••' ^) denote consumer j's intention to redeemcoupon i (i = I, ..., n), with U,j = I if he or she intends to re-deem it, and 0 otherwise. Let u be the Nn dimensional vec-tor of responses of the N consumers for the n coupons. Thelikelihood of observing the response vector u, given thecoupon parameter vectors a, b and the individual parametervector 6, is(2)Customers waiting in the checkout line were selected ran-domly and requested to participate in the survey. They weregiven u copy of the survey instrument, a prepaid return en-velope, and one dollar as a token of appreciation. Of fivehundred questionnaires distributed, three hundred forty-fivecompleted questionnaires were received, which representeda response rate of 69%. The final sample consisted of 78%women, had an average age of 39 years, and an averagehousehold size of 2.8 people, and the mean number of hoursworked per week was 37.Research InstrumentEarly versions of the questionnaire were pretested on overone hundred students and randomly selected grocery shop-pers. In the final version of the questionnaire we measuredrespondents' redemption intentions with respect to couponsin four categories: two grocery product categories—coffeeand detergent—and two service categories—beautysalon/barber shop and oil change for automobiles. For eachcategory, we constructed a set of coupon profiles by sys-tematically varying the type of coupon, the coupon face vai-ue, and preference for the couponed brand. Respondents' re-demption intention for each coupon was measured as a di-chotomous variable by asking whether they would redeemthe coupon; each respondent evaluated all coupon profiles.Profiles were constructed on the basis of three types ofcoupons {FSI, on-pack, and mail-in), three face values (40cents, 75 cents, and $1 in the case of FSI and on-packcoupons, and 75 cents, $1, and $1.50 in the case of mail-incoupons),-'' and two brands (\"most frequently purchasedbrand\" and \"occasionally purchased brand\"). The most fre-quently purchased and occasionally purchased brands in thequestionnaire referred to the respondent's favorite brand anda brand occasionally purchased by the respondent in the cat-egory.* We also obtained other information related tocoupon usage. A multi-item explicit measure of couponproneness was obtained using the .scale developed by Licht-enstein, Netemeyer, and Burton {1990).Mode! EstimationThe basic problem in estimating the parameters of themodel is to find consistent and sufficient estimators forcoupon proneness of consumers (9i) and parameters that de-fine the coupon characteristic curves (aj and b|). An iterativeapproach is used to estimate and refine the two sets of para-meters. Starting values of 6j are obtained through summated-'Free-standing insert coupons were described in ihc i|iie;.tioniiaire asmanufacturer coupons printed in the Sundiiy newspaper ihat ciitild he usedin any store carrying Ihe briind. Mail-in coupons were described as thoseihal customers must mail to the manufaeturer wiih a proof of purchase. On-pack coupons were described as those that appeared on the oulside of thepackage and eould he used for a subsequent ptirchaso. For mail-in coupons,a $l..^n face value was selected fcir testing instead of the 40-eenl coupon,because pretests showed that y 40-cent lace value for mail-in coupons wasnut realistic and was unlikely lo generate any response given mailing cosls.^For the two service categories, the profiles were constructed differentlybecause lhe eoupiining environment is differeiil for services. For example,on-pack coupons are noi applicable for services. Similarly, inaii-in couponsare not applicable because a discounl on the curreni purchase is typicallygiven by lhe service provider at the time of purchase, if at all. Coupon pro-files for services were therefore based tm lhe type of service provider(provider normally patronized and occasionally patroni/ed) and the facevalue of the eoupon. For the beauty salon service, coupon face value wasdescribed in lhe form of a percentage off the regular price.where Pjj is the probability that consumer j intends to re-deem coupon i, as defmed in Equation I, and Q|j = 1 - Pjj.The log-likelihood is maximized with respect to the para-meters a, b, and 6. To eliminate indeterminacy in the mod-el, the 0j parameters in each category are standardized withzero mean and unit standard deviation across the sample.Hambleton and Swaminathan (1985) discuss details of theprocedure.RESULTS AND DISCUSSIONThe redemption intentions data obtained from the surveywere u.sed to estimate the modei for the four categories. Thecoupon parameters b; and a, for coffee and detergent areshown in Table I. For ease of exposition, face values in thetable are reported as low, medium, and high. These corre-spond to face values of 40 cents, 75 cents, and $1, respec-tively, for FSI and on-pack coupons, and 75 cents, $1, and$ 1.50, respectively, for mail-in coupons.Because bj is inversely related to coupon attractiveness, itshould decrease with face value and preference for thecouponed brand and increase with effort required to use thecoupon. Table I, Part A, confirms these expectations.Coupons with high face values have the lowest bj values,whereas coupons with low face values have the highest b,values. Mail-in coupons, which require more effort to re-deem than FSI or on-pack coupons, have substantially high-er b, values than FSI and on-pack coupons. Finally, couponsfor the favorite brand have lower values of b, than couponsfor occasionally purchased brands after controlling forcoupon type and face value.\"^ These results are intuitively ap-^An anonymous reviewer poinied out that the impact of the \"most fre-quently purchased brand\"\" versus \"occasionally purchased brand\" manipu-lation may vary across categories wiih the level of brand loyalty in the eat-egory. We measured brand loyalty for the coffee and detergent categoriesand found category diflerence.s (o be small {wiih means of 5.1 and 4.9 on aseven-poinl scale lor the two categories, respeclively), whieh implies tbalany differences in lhe impact of the manipulation across these categoriesmay nol be signillcant in the aggregate.Modeling Coupon RedemptionTABLE 1521PARAMETER ESTIMATES BY COUPON TYPE. CATEGORY, PREFERENCE FOR COUPONED BRAND. AND COUPON FACE VALUE(COFFEE AND DETERGENT CATEGORIES)aPart A: Estimates ofh,ParametersFavorite BrandLowFaceValue-.79-.69-.71-.682.301.45MediumFaceVatue-1.31-1.04-1.09-.98HighFaceValue-6.20-2.92-!.49-1.72.63.32Low-FaceVatue.20.15.25.112.301.52Occasionally Purchased BrandMediumFaceCategoryI'SIOn-PackMail-InDetergeniCoffeeDetergentCoffeeDetergentCoffeeVii/ue-.57-.50HighFaceValue-1.45-1.36-.94-1,10.82.58-.39-.401.19.89.9!.68Part B: Estimates of a,ParametersFavorite BrandLowFaceVatue1.592.552.222,951.332.07MediumFaceValueOccasionally Purchased BrandHighFaceValue.541.383.582.561.321.35LowFaceValue1.422.131.642.452.494.91Mediumfacel,9()8.702.533.892.473.95H'KhFaceValue1.681.832,532.18i.8O1.60CategoryFSIDetergeniCoffeeDelergentCoffeeDetergeniCoffee2.123.922.954.811.661.85On-PackMail-In^Low, mediumand high face values correspond to 40 cents, 75 cents, and $1. respectively, for FSI and on-pack coupons, and 75 cents. $1. and $1.50 formail-in coupons.pealing and suggest that the model and the parameter esti-mates have face validity.Table I. Part B, shows that low-value eoupons tend tohave higher values of a, compared with high-value coupons,which implies that coupons with low face values providemore infonnation ahout the coupon proneness of the con-sumers they attract. This is understandable hecause re-deemers of low-value coupons tend to be highly coupon-prone consumers and thus a relatively homogeneous group.However, the relationship between face value and the aj pa-rameter does not appear to be uniformly monotonic, becausesome medium-value coupons have higher values of aj thanthe corresponding low-value or high-value coupons. Theseparticular coupons also tend to have bj values ranging from-1.30 to -.40; that is, they tend to attract consumers with Gjvalues of -1,30 or higher. The implication is that the medi-um-value coupons used in this study may have been at thethreshold of acceptability for such consumers, so they dis-criminated well between consumers who were above andbelow this level.^Table 2 shows the a^ and b; estimates for the beauty salonand oil change services. In general, the fmdings for thesecategories are similar to those for coffee and detergent.''We are grateful to an anonymous reviewer for suggesting thisexplanation.TABLE 2PARAMETER ESTIMATES BY CATEGORY PREFERENCE FOR SERVICE PROVIDER. AND COUPON FACE VALUE(OIL CHANGE AND BEAUTY SALON CATEGORIES)aService ProviderNormatlx PatronizedLnwSer\\'ice ProviderOccasionally PatronizedHighFaceVatue-1.85-2.043.043.13LowParameterh.CategoryOil ChangeBeauty SalonOil ChangeBeauty SalonFaceValue-.47Mi'diiiiiiFaceValue-.86-1.093.453.14FaceValue1.461.392.984.62MediumFaceValue.26.337.188.48HighFaceVahif-.50-.443.967.06-.941.8S1.50^Low. medium,and high lace values correspond to $2, $4, and $6, respeetively, for the oil change eategory. and 10%. 20%, and 30% discount for the beautysalon eategory.522JOURNAL OF MARKETING RESEARCH, NOVEMBER 1997The a, values appear to be largely unrelated to couponcharacteristics. The positive coefficient for the face value(75 cents) variahle implies that, as was noted previously,coupons with medium face values tend to discriminate het-ter between different coupon proneness levels in our sample.The negative coefficient for the category dummy indicatesthat detergent coupons provide less infonnation about thecoupon proneness of consumers {relative to coffee), whichsuggests that they attract a broader cross section of con-sumers with varying levels of coupon proneness. Althoughan investigation of this issue is beyond the scope of ourstudy, we note that our sample reported substantially greateruse of detergent compared to coffee, whieh may account forthis finding.These findings suggest that managers can maximize the\"draw\" of their coupon promotions by selecting the appro-priate categories and brands to promote and hy choosing theoptimal combination of coupon characteristics. Coupons formore heavily used categories tend to attract a broader crosssection of consumers with varying levels of coupon prone-ness. Coupons for large-share brands and with higher facevalues tend to enjoy higher redemption rates. However, thereappears to be a threshold effect for coupon face value;among our suhjects, $1 and $1.50 coupons were seen asmore attractive than 40-cent coupons, whereas 75-centcoupons were not. Furthermore, delivery vehicles and re-demption requirements intluence coupon attractiveness. Forexample, this study found FSI eoupons to be the most at-tractive and mail-in coupons to be the least.Predictive Validity anci Model TestingBecause the IRT-hased model estimates latent or unob-served variables, it is not directly comparable to existingmodels that use direct measures of the predictor variahles.But the usefulness of the proposed model can be assessed hycomparing its predictive ability with that of more tradition-al models that are based on directly measured predictors. Weelected to use as a benchmark a logit, or logistic regression,model (Maddala 1983) hecause it has a similar functionalform. With the iogit model, the prohability of redemption isgiven by(3)Coupon attractiveness increases for higher face values andthe consumer's usual service provider. As before, mediumface value coupons appear to discriminate best betweenvarying levels of eoupon proneness. An Interesting aspect ofthe data here is that the a, values for these services are gen-erally higher than for coffee and detergent. This suggeststhat response to coupons for these services tends to be moresensitive to small changes in coupon proneness. Alternative-ly, it may he that coupons for services are more informativebecause they discriminate between users and nonusers ofthese services.To gain further insights into the relationship hetweencoupon attractiveness and coupon characteristics, ordinaryleast squares regression models were estimated with the pa-rameters aj and bj as dependent variables and coupon char-acteristics and category dummies as predictors. The analysiswas conducted for the 36 coupon profiles in the coffee anddetergent categories (the hcauty salon and oil change cate-gories were not included hecause the coupon profiles weredefined differently). Table 3 shows the coefficient estimatesfor the regression models. As can be observed from thetable, h, is related significantly to several coupon character-istics. The b, parameter decreases with coupon face valueand preference for the couponed brand and is higher formail-in coupons relative to FSI and on-pack coupons. Fur-ther, there is no significant difference in b, values hetween40-cent and 75 cent coupons, but coupons with face valuesof $1 and higher are significantly more attractive. These re-sults are consistent with our previous observations and indi-cate that coupons with higher face values and those for fa-vorite brands are more attractive, whereas mail-in couponsare less attractive compared with FSI and on-pack coupons.The small positive coefficient for on-pack coupons indicatesthat they are marginally less attractive than FSI coupons.The insignificant coefficient for category implies that thereis no difference between coffee and detergent coupons in b;values.TABLE 3COUPON PARAMETERS AS A FUNCTION OF CATEGORY, FACEVALUE, COUPON TYPE, AND PREFERENCE FOR THECOUPONED BRANDSDependent VariahleIndependent VariahleFace Value (75 cents)^Face Value ($1)'^Face Value ($1.50)^Calegory-'Favorite Brand*!Mail-In Coupon'^^On-Pack Coupon ^Model ft-'l>.-.58-1.81***-2.01***-.04-.81***3.18***.62*.80a.1.47**.If-.25-L07**-.52-.16exp(p'X,.38.4S\"Based on parameter estitnate.s for cotTee and detergent categories.•\"Dummy variable.'^Ecjuals I tor detergent cDupuns, 0 {ithei wise,tor coupons lor lhe favorite brand. 0 otherwise.tor mail in coup<)ns. 0 otherwise,'Bquals I Tor on-pack coupons. 0 otherwise.*;)<.IO.**/;<.05.**V<.01.where P;J is the probability that consumer j intends to re-deem coupon i, X, are characteristics of coupon i, Zj arecharacteristics of consumer j. and (3 and yare vectors of co-efficients. Coupon characteristics Xj included in the modelwere face value, coupon type, and preference for thecouponed brand, and consumer characteristics Zj includedwere gender, age, income, hours worked per week, house-hold size, and Lichtenstein, Netemeyer, and Burton's {1990)multi-item measure of coupon proneness.For the analysis of predictive validity, we began by ran-domly splitting the total sample into estimation and holdoutsamples of equal size. We assessed predictive validity for thelogit model in the usual way hy applying the coefficient es-timates from the estimation sample to the predictor variablesin the holdout sample. For the IRT-based model, we used adifferent procedure hecause there are no explicitly measuredpredictors in this model. The procedure was based on theModeling Coupon RedemptionTABLE 4PERCENTAGE OF REDEMPTION INTENTIONS CORRECTLY PREDICTED AND MAD OF PREDICTIONS BY TYPE OF MODELIRT based ModelE.stimciiion SampleCategoryCoffeeDetergeniOil ChangeBeauty SalonAverage. Hil Rale\"MAD^.1790Holdout SampleHit Rate\"908.1899489523ModelEstimation SampleHit Rate\"1177787677MAD>'.32.30.29.29.30Holdout SampleHit Rale\"75MAD>'MAl)^.19.22.15.1 1.17m919692.20.14.10.1576737775.32.30.31.30.31•\"Percenlage of redemption iniemions correctly predicted.absolute deviation between predicted response probability and stated intenlion (see text for details).sample independence property of the a; and b, parameters ofthe model, which states that the a; and b, parameter esti-mates are independent of the tested sample and can be u.sedto predict response for any consumer whose 9j is known(Hambleton and Swaminathan 1985). The IRT model's pre-dictive validity therefore can be tested by estimating thecoupon parameters on one sample and applying them (usingEiquation 1) to a different sample for which the 6j\"s have al-ready been estimated. Accordingly, we estimated the modelseparately for the estimation and holdout samples. Estimatesof a, and b, from the estimation sample were combined withthe 6. estimates from the holdout sample to predict the re-sponse probability for each coupon for each consumer in theholdout sample. This response probability was then com-pared with the consumer's stated redemption intention forthe coupon.^Two criteria were used to evaluate predictive validity: (I)the \"hit rate\" in predicting redemption intentions and (2) themean absolute deviation (MAD) between the predicted re-sponse prohability and the stated redemption intention. Thehit rate was calculated as the percentage of time the modelpredicted a redemption probability of more than .50 and theconsumer's stated intention was to redeem the coupon. TheMAD was calculated by comparing the predicted responseprobability with unity (if the consumer intended to redeemthe coupon) or zero (if lhe consumer did not intend to re-deem the coupon). The results, averaged across all couponprotlles, are shown in Table 4 for the estimation and holdoutsamples.As can he observed from Table 4, the IRT-based modelpertbrms remarkably well with an average hit rate of 89%correct classifications and MAD of .17 on the holdout sam-ples and slightly better results for the estimation samples. Itis also interesting to note thai the IRT predictions for the twoservice categories are better on average than those for thetwo product categories. For example, the IRT based model isabie toclassify correctly 94% of consumers for beauty saloncoupons. This might be accounted for by the relatively highvalues ofaj for coupons for services. Consumer response tosuch coupons by definition would be more polarized (i.e..the characteristic curve would be steeper) and therefore eas-ier to predict.^Thc use ol\" thf 6| eslimates from ihe holdout sample is necessary lor pur-poses ot\" prediction because ol' the lack ol\" explicitly measurcil predictorvariables in the IRT model. In etTcci. the procedure tests the prediclivevalidity of lhe a, and b, parameters but not ul 6,.In contrast, the logit model averages a hit rate of 75% anda MAD of .31 in both the holdout and the estimation sam-ples. The lower predictive ability of the logit suggests that itmay be difficult to capture consunier heterogeneity incoupon proneness using demographics and explicit mea-sures of coupon proneness such as Lichtenstein, Netemeyer,and Burton's scale. This result appears consistent with pre-vious research that has found demographics and psycho-graphics to be poor predictors of behavior. The result alsosuggests that not only must deal-proneness measures be do-main-specific (Lichtenstein, Netemeyer, and Burton 1995).but they also must be category-specific when the objective isto predict consumer response to specific deals.** Couponproneness measures that are not category-specific have lowpredictive power and may be particularly poor in explainingcoupon usage in categories (such as services) that are nottraditionally associated with couponing.In summary, our results show that the IRT-based approachhas high predictive validity both in absolute terms and incomparison to a more traditional approach that is based onexplicitly measured variables. An additional benefit of ourapproach is that examination of the a, and b; parameter val-ues can yield insights into consumer response to couponsthat would not be provided by a traditional logit-type mod-el. With respect to implementation, the IRT-based approachhas an advantage, because estimation of the model does notrequire data on predictor variables but only on coupon ex-posure and redemption. However, if data on predictor vari-ables are available, they can be included in the analysis ofmodel parameters as shown in Table 3.CONCLUSIONS AND MANAGERIAL IMPLICATIONSAs the importance of couponing has grown, so have con-cerns about the effectiveness of coupon promotions (Bawa1996). Here, we suggest a framework ft)r modeling couponredemption that makes it possible to evaluate the relativemerits of different coupon promotions and examine howconsumer response to coupons varies by coupon character-istics. Empirical application of our model and methodologyshows that important insights into consumer response tocoupons can be obtained and that consumer response can bepredicted correctly for nearly 90% of the sample. In con-'*Suppon for this interpretation was provided by predictive testing of abaseline logit model that did not include the coupon proneness measure.The hil rate for this liaseline model was the same as that for the full model,which implies that lhe measure did not add predictive power lo the model.524JOURNAL OF MARKETING RESEARCH, NOVEMBER 1997mailings can estimate the model for each product categoryto measure the coupon proneness of shoppers in the catego-ry. Information about customers' coupon proneness then canbe used to deliver lower-value coupons to the more coupon-prone shoppers. Such a precision targeting strategy can heipto maximize coupon redemption rates while minimizingpromotion costs.Incremental sales assessment. Many researchers havehighlighted the importance of measuring incremental salesin the evaluation of coupon effectiveness (e.g.. Bawa andShoemaker 1989). Our model provides a method for assess-ing a coupon's potential incremental sales by evaluating thedrawing power of the coupon among occasional and loyalbuyers of the brand of interest. This can be determined bycomparing the response probabilities when the coupon is forthe consumer's favorite brand versus when it is for an occa-sionally purchased brand, evaluated at the average level ofcoupon proneness (i.e., 6j = 0).\"' For example, if for a givenset of coupon characteristics (e.g., a 40-cent mail-incoupon), the response probability (at 8j = 0) is .80 when thecoupon is for the favorite brand and .40 when the coupon isfor an occasional brand, the modei would predict that 33%of redemptions wouid be made by occasionai buyers.\" Bycomparing this with predictions for other coupon types, thepotential incremental sales from each coupon relative to oth-er coupons can be assessed.The IRT-based approach to modeling coupon redemptionshows considerable promise as a framework for understand-ing consumer behavior and designing and evaluating couponpromotions. Given the unprofitable nature of most couponpromotions today, the model can be of great value to man-agers as a basis for improving coupon profitabiiity.REFERENCESBagozzi. Richard R. Hans Baumgartner, and Youjae Yi (1992),\"Appraisal Processes in the Enactment of Intentions to UseCoupons,'\" Rsvchology <& Mitrkeling. 9 (November/December),469-86.Balasubramanian. Siva K. and Wagner A. Katnakura (1989). \"Mea-suring Consumer Altiludes Toward the Marketplace With Tai-lored Interviews,\" Journal of Marketing Research. 26 (August),311 -26.Bawa, Kapil (1996), \"Intluences on Consumer Response to DirectMail Coupons: An Integrative Review,\" Psychology ami Market-ing. 13 (March), 129-56.lra.sl, a traditional logit approach that is based on explicitlymeasured variables is able to predict consumer response15% of the time.A hmitation of our study is that it does not consider suchcoupon characteristics as coupon expiration date (Inman andMcAlister 1994), advertising copy, and layout. These werenot manipulated in our study so as not to impose an unrea-sonable burden on our respondents. Further research isneeded to study the role of these characteristics. Anotherlimitation is that our analysis was conducted on redemptionintentions rather than redemption behavior. Although thetwo are likely to be highly correlated, projecting from inten-tions to behavior must be done with caution. To the extentthat consumers misplace or forget to use coupons they have,observed redemption rates would be lower than predictedfrom intentions. Other factors also may mediate the rela-tionship between redemption intentions and behavior(Bagozzi. Baumgartner. and Yi 1992). Research is needed todetermine the correlation between redemption intention andbehavior and to identify the individual-level and category-level characteristics that influence this correlation.The item response theoretic approach described here ispotentially a valuable tool that can help managers under-stand the impact of their coupon promotions and facilitatethe design of more effective and profitable promotions. Di-rect marketers and retailers with customer databases that in-clude data on coupon drops and redemptions (such as su-permarket \"clubs\") may find this approach particularly use-ful.*^ The model can be estimated by obtaining either inten-tions data from a survey or behavioral data from householdsto which coupons are delivered by direct mail (e.g., couponexperiments on scanner panels could yield appropriate be-havioral data). When the model parameters are estimated,the model can be implemented in a variety of ways to im-prove the effectiveness of coupon promotions, as we discusssubsequently.Coupon design. Perhaps the most natural use of the mod-el would be to evaluate the impact of coupon characteristics,such as face value and delivery vehicle, on coupon attrac-tiveness in order to design coupons that maximize consumerresponse. This type of analysis is most suitable when a full-profile approach is used for the design of coupon stimuli, sothat the number of stimuli is sufficiently large relative to thenumber of coupon characteristics. We noted previously howexamination of a, and bj parameters can yield insights intohow consumers respond to coupons. For example, our em-pirical test of the model suggests that a 75-cent coupon isnot significantly more attractive than a 40-cent coupon,which implies that the additional cost of offering a 75-centcoupon would not be worthwhile. Sensitivity analyses alsocan be conducted to investigate the effects of coupon char-acteristics on response probabilities by changing the level ofone characteristic while holding the others constant and not-ing the impact on the dependent variable. For example, theimpact of increasing face value from 40 cents to one dollaror of using a shorter expiration period could be evaluated.while keeping other characteristics constant.Precision targeting. Direct marketers with informationabout consumers\" redemption histories for past coupon'^We are grateful to an anonymous reviewer lor sujrgesiing thisapplication.'\"AUernatively, a researcher could analyze response to a specific brand'scoupons by conducting the analysis among differeni prelerenue segments(e.g.. \"loyal\" buyers, for whom the brand ol' interest i.s (he favourilc brand,und \"'itfcasinnar' buyers of tbe brand).' 'Let p(Rk)) dcniilc ihc probability of redemplion given that the brand isan occasional brand, and p(Rlf) the probability of redemption given Ihat thebrand is a favorite hraiiJ. t,et pU'l and p(f( denole the numbers DI' occa-sional and loyal buyers of the brand as a proportion i>r those who receivethe coupon. Assume these are the only segments in the markcl and p(o) -p(0. The posterior probahihty Ihat Ihe hrand is an occasional brand, giventhat ihe coupon is redeemed, is given by p(i)IR) = p(Rlu) x p(o) / fp(Rln) xp(o) + p(Rin X pd\")]. In lhe example, p(Rlo) - .4. p(Rlf) = .8. so that p(olR)= .4/(.4 + .8) = .3.1. Nnle that the actual redemption rate may be lower if themodel is estimated using intentions data. We assume for simplicity that p(o)= pd) and that consumer preferences can be represented by characterizingbrands as \"favorite\" or \"occasional.\" However, the analysis can be extendedeasily lo accoinmodalc other scenarios.525JOURNAL OF MARKETING RESEARCH, NOVEMBER 1997and(1995). \"Assessing the Dotnain-Specificity of Deal Proneness: A Field Study,\" Journal of Con-sumer Research. 22 (December), 314-26.Maddala, G. S. (1983), Limited-Dependent and Qualitative Vari-ables in Econometrics. New York: Cambridge University Press.Narasimhan, Chakravarthi (1984). \"A Price Discrimination Theoryof Coupon Usage.\" Marketing Science. 3 (Spring), 128-46.NCH Promotional Services (1996). Worldwide Coupon Distribu-tion and Redemplion Trends, Vol. XXIX, Chicago: NCH Pro-motional Services.Neslin, Seott A. and Dan-al G. Clarke (1987). \"Relating the BrandUse Profile of Coupon Redeemers to Brand and Coupon Char-acteristics,\" Journal of Advertising Research. 27(February/March), 23-32.Reibstein, David J. and Phlllis A. Traver (1982), \"Factors AffectingCoupon Redemption Rates.\" Joumal of Marketing, 46 (Fall),102-13.Shoemaker, Robert W. and Vikas Tibrewala (1985), \"RelatingCoupon Redemption Rates to Past Purchasing of the Brand,\"Journal of Advertising Research. 25 (October/November),40-47.Teel, Jesse E., Robert H. Williams, and William O. Bearden (1980),\"Correlates of Consumer Susceptibility to Coupons in New Gro-cery Product Introductions,\" Journal of Advertising. 9 (3),31-35.Thissen, David (199\\), MULTILOG: Using Item Response Theory.Chicago: Scientific Software.Ward, Ronald W. and James E. Davis (1978). \"A Pooled Cross-Section Time Series Model of Coupon Promotions,\" AmericanJournal of Agricultural Economics. 60 (August). 393-401.Webster, Frederick E., Jr. (1965), \"The 'Deal-Prone' Consumer,\"Journal of Marketing Research, 2 (May), 186-89.and Robert W. Shoemaker (1987a), \"The Coupon-ProneConsumer: Some Findings Bused on Purchase Behaviour AcrossProduct Classes.\" Journal of Marketing, 51 (October), 99-1 iO.and (1987b), \"The Effects of a Direct Mail Couponon Brand Choice Behavior,\" Journa! of Marketing Research, 24(November), 370-76.and (1989), \"Analyzing Incremental Sales From aDirect Mail Coupon Promotion.\" Journal of Marketing. 53 (Ju-ly), 66-78.Bimbaum, A. (1968), \"Some Latent Trait Models and Their Use inInferring an Examinee's Ability.\" in Statisticat Theories of Men-tal Test Scores, F.M. Lord and M.R. Novick, eds. Reading. MA:Addison-Wesley.Blattberg, Robert C, and Scott A. Neslin (1990), Sales Promotion.Englewood Cliffs, NJ: Prentice Hall.Hambleton, R. K. and H. Swaminathan (1985), Item Response The-ory Principles and Applications. Boston: Kluwer Nijhoff.Inman. J. Jeffrey and Leigh McAlister (1994). \"Do Coupon Expi-ration Dates Affect Consumer Behavior?\" Journal of MarketingResearch. 31 (August), 423-28.Krishna, Aradhna and Robert W. Shoemaker (1992), \"Estimatingthe Effect.s of Higher Coupon Face Values on the Timing of Re-demptions, the Mix of Coupon Redeemers, and Purchase Quan-tity.\" Psychology & Marketing. 9 (November/December),453-67.Levedahl, J. William (1988), \"Coupon Redeemers: Are They Bet-ter Shoppers?\" Journal of Consumer Affairs, 22 (Winter).264-83.Lichtenstein. Donald R., Richard G. Netemeyer, and Scot Burton(1990), \"Distinguishing Coupon Proneness From Value Con-sciousness: An Acquisition-Transaction Utility Theory Perspec-tive,\" Journal of Marketing, 54 (July), 54—67.The School of Business at the University of Hong Kong, Hong Kong's leading university, invites applications forappointment as Professor: Chair tif Marketing (Ref: RF-97/98-18), tenable from as soon as possible.Hong Kong is a major international business centre and this appointment offers special opportunities to contribute to thefields of international and/or services marketing. The School teaches at the doctoral, MBA and BBA levels.The University wouid prefer to make a fixed term appoinlment of preferably not less than three academic years, butconsideration may also be given to appointment on superajmuable terms. Tlie University reserves the right not to fill theChair or to fill the Chair by invitation or to make an appointment at a lower level.Annual salary [attracting 15% (taxable) terminal gratuity] will be within the professorial range, whieh is not less thanHK$1M (approx. US$129,870; US Dollar equivalent as at September 24, 1997) with starting salary depending onqualifications ami experience.At current rates, salaries tax will not exceed 15% of gross income. Leave, medical/dental benefits, an allowance forchildren's education in Hong Kong, and housing at a charge of a percentage of salary, currently 7.5%, are provided. Pleasesubmit current curriculum vitae, a letter of interest and the names of three referees to the Appointments Unit, Registry,Tlie University of Hong Kong, Hong Kong.Review of applications will begin immediately and continue until the position is filled by a qualified candidate,.^plications on hand prior to October 15, 1997 will receive priority consideration. Further particulars and applicationforms can be ohtained on WWW at http:/Avv™.liku.hk; or from the y^pointments Unit (Fax (852) 2540 6735/2559 2058;E-mail: APPTUN1T@REG.HKII.HK).Professor: Chair of Marketing