Using the Juster Scale to Estimate the Demand-Price Relationship

BiographyMike Brennan and Don

Esslemont are Senior Lecturers,and Clarence U was a graduatestudent, in the Department ofMarketing, Massey University,New Zealand.

Acknowledgment:We would like to thank the

anonymous reviewers for theirvaluable and helpful comments,and the Massey UniversityResearch Fund for supporting theproject

Using the Juster Scaleto Estimate theDemand-PriceRelationshipMike Brennan, Don Esslemontand Clarence U

Palmerston North

NEW ZEALAND

Asia - Australia Marketing Journal Vol. 3, No.1· 27

AbstractThe Juster scale was used to

estimate the price-demand relationship for two branded products,Coca-Cola in cans, and Campbell'sRed and White Label soup. Foreach product predictions wereobtained of the number of itemsthat would be purchased, duringthe following four weeks, at eachof two hypothetical prices, oneabove and one below the currentretail price. Predictions wereobtained using two variations onthe Juster procedure: the multiplequestion method and the constantsum method. Actual numbers purchased were obtained in follow-upinterviews. Actual purchases ofCoca-Cola were almost exactly atthe predicted level, but purchasesof the soup brand were less than athird of those predicted. This is thefirst reported use of the JusterScale to predict purchases ofbrands as opposed to products, orto estimate price-demand relationships.

IntroductionForecasting consumer purchas

es has traditionally involved theuse of purchase intention scales.However, such scales are generallypoor predictors, in part because asignificant proportion of respondents who say they have no intention of buying, end up making apurchase. Juster (1966) demonstrated that significant improvements in predictive accuracy couldbe achieved by using purchaseprobabilities rather than purchaseintentions, and developed what hasbecome known as the Juster Scale.

Although various forms of theJuster Scale have been used (seeDay, Gan, Gendall & Esslemont1991, for a review), the standardform consists of an eleven pointnumerical scale, ranging from 0 to

10, each point associated with botha numerical probability and a verbal label (see Figure 1). Evidenceto date suggests that, overall, thestandard form provides more accurate predictions, and results inlower non-response, than otherscale forms (Juster 1966; Gendall,Esslemont & Day 1991).

The Juster Scale is used in conjunction with a question such as:"What are the chances that youpersonally will buy a (product) atsome time during the next (somany) months?" Each respondentthus estimates the probability thathe or she will, during the periodspecified, make the purchase inquestion. The mean of these probabilities is the estimate of the proportion of consumers who willmake the purchase.

Until recently, most of the studies using the Juster scale (Juster1966; Gabor & Granger 1972), orvariations of it (Clawson 1971;Pickering & Isherwood 1974;Isherwood & Pickering 1975),have used it to predict purchases ofcars and other durables, althoughClawson (1971) examined cars andlow-priced frequently purchasedservices. In these studies it wasshown that the Juster Scale canproduce quite accurate estimates ofcar purchase, but the accuracy ofestimates for other goods variedwidely.

More recently, experiments inNew Zealand have used the JusterScale to predict purchases ofdurables, services and fast-movingconsumer goods (Day, Gan,Gendall & Esslemont 1991). Of 21predictions, nine (43%) were within 20% of actual purchase rate,five of these within 10% 1. As inother studies, predictions of carpurchases were more accurate thanthose for other products.

In all of these studies, the Juster

28 • Asia - Australia Marketing Journal Vol. 3, No.1

Scale has been used to predict purchase rates, that is, the proportionof people likely to buy a particularproduct. But for many productsand services, some consumers arelikely to buy several items duringeven quite a short period. Heavyusers of goods such as butter,pasta, or cheese, and so on, maybuy varying amounts at a time, andmay do so several times during thechosen period. For such products itis not enough to predict the proportion of consumers who willbuy. It is necessary to predict themean purchase level, that is, theaverage number of units bought. Amethod of using the Juster scalefor this purpose was developed byHamilton Gibbs (Esslemont,Hamilton-Gibbs, & McGuinness1992).

In a ground-breaking study,they examined two ways of administering the standard Juster Scaleto obtain estimates of purchaselevels. One method, called themultiple question method, simplyasked for separate probabilities foreach level of purchase. The secondmethod, called the constant summethod, required respondents toallocate 10 tokens, each representing a purchase probability of 0.1,over different levels of purchase inproportion to the probability ofpurchase.

Each method allowed the calculation of the expected number ofitems to be bought by each respondent as the sum of the probabilitiesof buying each number, weightedby that number.

Thus

kE= LPnn

1

where E = expected number ofitems to be purchased by respondent

n =0,1,2, ...kpn = probability of buying

exactly n itemsk= largest value of n with

non-zero pnThe mean value of E for the

sample is an unbiased estimate ofthe population mean value.

They concluded that the constant sum method was superior tothe multiple question method. Itproduced more accurate predictions for five of the seven itemstested, generally gave much lowerpredictive error, and was apparently easier and less tedious forrespondents to use. In most casesthe predicted purchases were within about 10% of the actual level. Inall cases, the multiple questionmethod underestimated purchases,whereas the constant sum methodoverestimated purchases and, ingeneral, the constant sum methodproduced the more accurate predictions.

Evidence to date thus indicatesthat the Juster Scale gives moreaccurate predictions of purchasethan other forms of scale, and canbe used to obtain accurate estimates of purchase quantities aswell as purchase rates. However,all of these previous studies haveused the Juster Scale to estimatepurchase rates or purchase levelsof product classes. Furthermore,these estimates have typically beenobtained without reference toprice. Often, however, a marketeris interested in predicting purchaserates or purchase levels for individual brands, and for brandedproducts, price may be a criticalfactor in the purchase decision.

The present study extends previous work in two ways. First, itexamines the effectiveness of theJuster Scale for predicting purchases of branded items, and second, it uses the Juster Scale at dif-

ferent price points so that simpledemand curves can be constructed.

MethodThe purchase probability data

for this study was obtained fromthe 1991 Palmerston NorthHousehold Omnibus survey. Thisis an annual project conducted bythe Department of Marketing atMassey University. The interviewing is done by second year students, supervised by members ofstaff. The students receive severalhours training, spread over fourweeks. The interviewing is face-toface, and is conducted in the lastweek of April.

The survey covers householdswithin the Palmerston North cityboundary, and the sample is basedon clusters of four interviews (twowith males and two with females,aged 15 years or older) around randomly selected starting points.Substitutions are made for households where an interview isrefused or households where nocontact can be made with therespondent after three attempts.

At the end of the interview,respondents were asked for theirconsent to be re-interviewed,although no indication was givenas to the subject of the furtherresearch. Of the 427 respondentsto the survey, representing aresponse rate of 54%, 323 (76%)agreed to be re-interviewed. Fourweeks after the omnibus survey,interviews were successfullyobtained with 302 (94%) of theseindividuals. The re-interviewswere conducted by telephone, byprofessional interviewers.

ProcedureThe two products used in this

study were Campbell's Red andWhite Label canned soup andCoca-Cola in cans. For each product it was intended to estimate the


demand curve, that is, to obtain aprediction of how the mean purchase level would vary, ceterisparibus, if the price were to vary.This is equivalent to estimating themean value of E, the expectedlevel of purchases, at variousprices.

Before obtaining the purchaseprobability data for the two products, respondents were introducedto the Juster Scale, in the mannerused by Juster (1966), using theprospects of moving house as apractice exercise. The instructionswere as follows:

"We would like to know whatthe prospects are of you personallymoving to another residence sometime during the next four weeks."

HAND RESPONDENTSHOWCARD G (See Figure 1).UNDER NO CIRCUMSTANCESSHOULD YOU SUGGEST ANANSWER TO THE RESPONDENT. IF RESPONDENTSREPLIES WITH A WORDDESCRIPTION SUCH AS"GOOD POSSIBILITY", ASK"What number is that?". READOUT

"The answers you may give areon this card, arranged on a scale abit like a thermometer. If you arecertain or practically certain thatyou personally will move to another residence during the next fourweeks, choose the answer '10'. Ifyou think there is no chance oralmost no chance of moving, thebest answer would be '0'. If youare uncertain about the prospects,choose another answer as close to'0' or '10' as you think it shouldbe."

"What answer would you chosefor the prospect that you personally will move within the next 4weeks, in other words, between

now and the end of May?"Respondents were then asked

about the two branded products.For each respondent the expect

ed number of items that would bepurchased over the following fourweeks was elicited at each of twoprices: one above the current retailprice, the other rather below it.

Although there was some variation between outlets, at the time ofthe research the average retailprice of Campbell's Red andWhite label soup was $1.90. Theprices used in the questioning were$2.25 and $1.59.

The price of the canned CocaCola varied rather more, but theaverage was estimated to beapproximately $1.00. The pricesused were $1.20 and 90 cents.

Respondents were thereforeasked, using either the constantsum method or the multiple question method, their probability ofbuying various numbers of cansover the following four weeks.This period was chosen because itwas thought to be long enough toprovide an adequate level of variance in purchase levels while stillbeing short enough to permit reasonably accurate subsequent recallof actual purchases. A shorter period, for example seven days, oreven one day, would have providedmore accurate recall, but very fewrespondents would have made anypurchases during the period.

The instructions for obtainingthe purchase rate data, using CocaCola to illustrate, were as follows:

"The first product is Coca-Colain a can."

"Imagine the price is $1.20 acan. Taking everything intoaccount, what are the prospectsthat you personally will buy atleast one can of Coca-Cola sometime in the next four weeks;

between now and the end of May?The answer obtained was a

number from the Juster Scale,ranging from 0 (no chance, almostno chance: 1 in 100) to 10 ( certain, practically certain: 99 in 100).These replies correspond to estimates of the probability (rangingfrom 0 to 1) of buying any cans ofCoca-Cola during the followingfour weeks. The mean of theseindividual probabilities is the estimated proportion of the samplewho would buy during this period,if the price were actually $1.20.

Respondents were then asked,using either the constant sum ormultiple question method, theprobabilities of buying variousnumbers of cans during the following four weeks, and the whole procedure was repeated at the lowerpnce.

The proportion of the samplewho will actually buy the productis estimated by interpolatingbetween the proportions at the twoprices.

The Multiple QuestionMethodThe multiple question method

requires respondents to state theprobability of their household purchasing one unit of a given item,then two, and so on, until a number of units is reached for whichthe purchase probability is zero.With this method, it is possible,and very likely, that the sum of thereported probabilities of purchasing different amounts of an itemmay be greater than one. To compensate for this, the probabilitiesare subsequently adjusted so theydo sum to one.

The procedure for the multiplequestion method, again usingCoca-Cola to illustrate, was as follows:

"What are the prospects that


you personally would buy just onecan of Coca-Cola in the next fourweeks if the price was $1.20 acan?" RECORD RESPONSE.

"What would be your answerfor 2 cans, 3 cans ?

REPEAT FOR 3, 4, 5 CANS,AND SO ON, UNTIL THE PURCHASE PROBABILITY ISZERO. N.B. ALSO ASK THEPROBABILITY OF BUYING NOCANS. RECORD RESPONSES.

The above procedure wasrepeated for the second price, thenthe whole procedure repeated forthe other brand.

The Constant Sum MethodThe constant sum method

requires respondents to indicatethe probabilities of purchasing various quantities of an item, in sucha way as to force the sum of theprobabilities to one. This isachieved by providing the respondent with a board on which a gridis printed, and 10 counters. On thegrid, the rows represent differentquantities of items purchased,from 0 to 12, while each columnrepresents a point, from 1 to 10, onthe Juster Scale. Each counter represents a purchase probability of0.1.

For each item in turn, respondents are asked to place all 10counters on the grid to show theprobabilities associated with eachlevel of purchase. For example, ifthere is a 6 in 10 chance theywould purchase 3 units of an item,and a 4 in 10 chance they wouldpurchase 4, they would put 6 counters in the row representing 3 unitsand 4 counters in the row representing 4 units of that item.

The procedure and instructionsfor the constant sum method, usingCoca-Cola to illustrate, were asfollows:

"Now we would like you to

think about how many cans ofCoca-Cola you personally are likely to buy in the next four weeks ifthe price was $1.20 a can."

"Here are 10 counters; use 1counter to represent a 1 in 10chance, 2 counters to represent a 2in 10 chance, and so on"

"Please place the counters onthis board to show the prospects ofbuying different numbers of cansof Coca-Cola at this price."

"You should have at least onecounter on the maximum numberof cans you think there is even a 1in 10 chance of buying in the nextfour weeks."

As with the multiple questionmethod, this process was repeatedfor the second price of 90 cents,then for the other brand; ieCampbells soup at $2.25 and$1.59.

Demand CurvesThe responses were used to cal

culate the expected number ofitems, E, for each respondent ateach price. The mean value of Eatthe higher price (Ehigh) and.at thelower price (Elow) were calculatedfor the whole sample. These values(the predicted mean number ofcans to be bought at each price)were then plotted against theirrespective prices.

An estimated demand curvewas then constructed by drawing astraight line through these twopoints. There is of course not aprior reason for supposing that thecurve in this region will in fact bestraight, but the error introducedby any departure from linearitycannot be large in comparison toother potential errors.

The predicted mean number ofcans purchased during the fourweek period can be obtained byfinding the value of E corresponding to the average retail price. In

practice this was done by interpolation, rather than by reading fromthe graph.

Actual PurchasesFour weeks after their initial

interviews, the respondents werere-interviewed, by telephone. Theywere asked whether they personally had purchased either of the twoproducts, and if so, how much theyhad purchased in the previous fourweeks, since the last interview.

While relying on recall as a surrogate measure of actual purchase,we acknowledge that there is considerable evidence that recall dataobtained from single interviewstypically provides over-estimatesof actual purchases (Sudman &Bradburn 1974; Neter & Waksberg1964; Parfitt, 1974). However,there is also evidence that this'telescoping', as it is called, iseffectively reduced by using abounded recall procedure (Neter &Waksberg 1964), where the beginning of the recall period is bounded by a previous interview.Although our previous interviewgathered purchase probability datarather than recall data, we wouldexpect the effects to be similar,although this remains to be confirmed. The alternative of usingdiary panels was not feasible,because of time and cost constraints.

Results andDiscussionPredicting the proportion ofthe sample who will buyAlthough the respondents were

assigned to one of two samplegroups (constant sum method ormultiple question method), thequestions they were asked toobtain the purchase rate data wereidentical, and so their responseshave been combined. The estimat-


Table 1

Predicted and Actual Mean Purchase Rates (Proportion of Buyers)

Campbell's Soup

High Low Average]

($2.25) ($1.59) ($1.90)

Total Sample3

Predicted % 11.7 21.6 17.0Actual % 6.6Error %2 +158

Main Shopper4

Predicted % 13.6 25.2 19.8Actual % 8.6Error % +130

Non ShoppersPredicted % 7.0 15.1 11.3Actual % 3.8Error % +197

Coca Cola

High Low Average

($1.20) ($0.90) ($1.00)

26.4 34.0 31.533.1-5

16.6 23.1 21.025.2-17

38.6 47.1 44.343.8+1

Note. 1. Estimated average market price. The predicted mean purchase level at this price was obtained byinterpolation using the high and low price data.

2. Error of Prediction =[(predicted - actual)/actual ] x 1003. N =302 (includes 46 people who share shopping responsibilities)4. N =151 (excluding those who share shopping)5. N =105 (excluding those who share shopping)

ed and actual proportions of purchasers for the two products, basedon the responses of the total sample, and separately for the groceryshoppers and non-grocery shoppers in the sample, are shown inTable 1.

The predicted number of buyersfor Coca Cola was very accurate,particularly for the non-groceryshoppers, which was only 1%above the actual purchase rate. Incontrast, the predicted purchasesof Campbell's Red and White labelsoup were highly inaccurate forboth shoppers and non-shoppers,with the numbers of buyers beingonly a third or a half of those pre-

dicted.It may seem strange that the

Coca Cola predictions should bemore accurate for non-shoppersthan shoppers, but this makessense when one considers thatCoke is typically an individualpurchase. Campbell's Soup, on theother hand, is a household purchase, so it not surprising that forthis product the accuracy of prediction was greater for the groceryshoppers.

There are several possibleexplanations for the gross overestimation of the purchase rate forCampbell's Soup. One possibilityis that respondents may have

ignored the variety of Campbell'sSoup being used (Red and WhiteLabel), in spite of the repetitiveuse of the variety name, andinstead simply considered anyvariety of Campbell's Soup, oreven canned soup in general. In asubsequent omnibus survey of thesame population, we found thatapproximately 23% of the samplebought canned soup, and 4.6%specifically bought Campbell'sSoup. These figures suggest thatrespondents in the present studymay have included all brands ofcanned soup when providing purchase probabilities, but onlyincluded Campbell's soup, of


whatever variety, when reportingactual purchases. Alternatively,given that relatively few of therespondents are likely to be regularusers of Campbell's Red andWhite label, or even any brand ofCampbell's Soup, given the smallmarket share, it could be that somerespondents simply answered thequestion as if they wereCampbell's soup users. Of course,this is ,only speculation, but thepossibility of question misinterpretation when using the Juster Scalerequires further investigationbecause of the significant impact itcould have on the accuracy of predictions.

Other possible explanations canbe drawn from the literature pertaining to research on recall error.For example, on the basis of previ0us research (Sudman & Bradburn1974), one might expect the resultsto be affected by an interactionbetween the type of product beingconsidered, and the nature of thetask required to gather the purchases probability data. It is wellknown that recall of purchases typically results in substantial overestimation of actual purchase rates(Sudman & Bradburn 1974), particularly for less frequently purchased items (Cook 1987; Parfitt1974), and those attracting lowbrand loyalty (Parfitt 1974). In thisstudy, the errors in prediction cannot be attributed to telescoping ofthe recall data, since the effect ofthat would be in the oppositedirection to that observed. But perhaps telescoping did occur whenthe purchase probabilities wereobtained. This is what one wouldexpect if the predictions are basedon recall of past behaviour, whichis likely to suffer from telescoping.Due to the relatively infrequentand low level of purchases of thebrand, Campbell's Soup would be

particularly vulnerable to thiseffect.

It is also possible that the predictions could have suffered fromanother effect, called averaging(Cook 1987). This effect has beennoticed with recall of infrequentlyor irregularly purchased items.Averaging occurs when respondents regularise or average theirpurchases over a time period,thereby inflating the reported purchase rate. Averaging could haveoccurred if the predictions werebased on recall of previous purchases.

It should also be pointed outthat, for products with low purchase rates, the error of predictionis effectively inflated because, asthe number of purchasers decreases, the effect of any incorrect prediction is magnified. That is, thesize of the error of prediction tendsto be inversely proportional to thesize of the purchase rate. Thiseffect has been noted by Day (Day,Gan, Gendall & Esslemont 1991)and is apparent for Campbell'sSoup.

There are also other possibleexplanations for the results. Forexample, it is possible that sales ofcanned soup were affected by theweather during the study, and ordiscounting of other brands thatcould have occurred.Unfortunately, it was not possibleto control for or measure thesepossible influences, but clearlythese variables could and shouldbe controlled for in a more tightlydesigned experimental study.

While a number of possible reasons have been offered to accountfor the poor purchase rate predictions for Campbell's Soup, theseare mere speculation at this stage,and simply serve to highlight theneed for further investigation ofthe reliability of both recall and

purchase prediction data over dif~

ferent time-frames, with differenttypes of product, under differentcircumstances. The results suggestthat future research should takecare to obtain estimates from thepeople who are actually responsible for purchasing the productsunder investigation. In addition,previous research on accuracy ofrecall suggests that it may be sensible to take into account factorssuch as frequency of purchase, relative market share, brand loyaltyand respondents' previous experience of the product. Further investigation is also required into theextent to which telescoping andaveraging actually do affect boththe recall and purchase probabilitydata, and of alternative techniquesfor minimising these effects if present.

Predicting the number of cansboughtTo obtain estimates of purchase

levels, that is, of the quantity ofproduct purchased, one samplegroup was exposed to the multiplequestion method, while the otherwas exposed to the constant summethod. The predicted and actualmean purchase levels for the twomethods are shown in Table 2 forCampbell's Soup, and in Table 3for Coca Cola.

As with purchase rates, the estimates of purchase levels for CocaCola are very accurate, and, asfound by Esslemont et al. (1982),for the total sample, the multiplequestion method resulted in asmall underestimation, while theconstant sum method resulted in aslight overestimation, althoughthere was a slight under-estimationby non-shoppers.

In contrast, the estimates ofpurchase levels for Campbell'sSoup, particularly those of the


Table 2.

Predicted and Actual Mean Purchase Levels(Number of Cans Bought) for Campbell's Soup

Multiple Question Method

High Low Average '($2.25) ($1.59) ($1.90)

Constant Sum Method

High Low Average

($2.25)($1.59) ($1.90)Total SamplePredictedActualError2

N

Main ShopperPredictedActualErrorN

.418 .638

154

.550 .777

77

.535

.084+536

.670

.117+473

.337 .706

144

.312 .616

73

.533

.125+326

.473

.178+166

Non ShopperPredictedActualErrorN

.241 .494 .375.019

+187352

.275 .719

49

.510

.082+522

Note. "1. Estimated average retail price. The predicted mean purchase level at this price was obtained byinterpolation using the high and low price data.

2. Error of Prediction =[(predicted - actual)/actual ] x 100

non-shoppers, are very inaccurateand grossly over-represent actualquantities purchased. Of course,this is not surprising, given thelarge error in the purchase rate forCampbell's Soup. The predictionsfor Campbell's Soup are moreaccurate when the data from thegrocery shoppers only is used, buteven these over-estimate purchasesby 2.5 to 5.5 times.

One possible explanation forthe differing degrees of accuracymay lie in the extent to whichpotential buyers of the productsperceive them to be different fromsubstitutes. It is well known thatbuyers of cola drinks tend to be

loyal to their brand, but it could bethat buyers of soups do not makeclear distinctions between the different varieties of Campbell'ssoup, or even between differentbrands of soup. If this is so,respondents answering the Justerquestion about Campbell's Redand White label soup may havebeen thinking about soup in general. A reanalysis of the data provided some support for this hypothesis, but the findings could be spurious. Research using the methoddeveloped by Belson (1981) couldestablish the extent to which this isthe case. And of course, the effectsof telescoping and averaging dis-

cussed in relation to purchase ratescould equally apply to the purchase level estimates.

While both the multiple question method and the constant summethod produced fairly similarresults, the constant sum methodconsistently produced more accurate predictions of mean purchaselevels, a result consistent with thatreported by Esslemont et al.(1992).

Esslemont et al. suggested thatrespondents found the constantsum method easier to use, butfeedback from interviewers in thepresent study indicated that somerespondents, particularly among


Table 3. Predicted and Actual Mean Purchase Levels (Number of Cans Bought)for Coca-Cola

Total SamplePredicted I

ActualError2-

N

Main ShopperPredictedActualErrorN

Non ShopperPredictedActualErrorN

Multiple Question Method

High Low Average!

($1.20) ($0.90) ($1.00)

.632 2.124 1.9602.214-12

154

1.147 1.616 1.4602.000-27

77

2.395 2.950 2.7651.792

+5452

Constant Sum Method

High Low Average

($1.20) ($0.90) ($1.00)

.939 1.728 1.4651.451

+1144

.678 1.188 1.018.877+16

73

1.233 2.359 1.9842.102

-6

49

Note. 1. Estimated average retail price. The predicted mean purchase level at this price was obtained by interpolation using the high and low price data.

2. Error of Prediction = [(predicted - actua1)/actua1 ] x 100

the elderly, found the method quiteconfusing. This confusion waslargely overcome by providing ademonstration, and could almostcertainly be further reduced byusing more experienced interviewers. Even so, there is a clear needfor further investigation and development of the procedures forapplying the constant sum method.

In theory, the sum of the purchase probabilities that an individual assigns to alternative quantitiesof a product should equal one.This was mostly the case with theconstant sum method, since theprocedure is designed to force theprobabilities to sum to one. With

the multiple question method,however, this outcome was theexception rather than the rule. Theproportion of people whose purchase probabilities summed exactly to one for a particular productranged from 48% to 70% acrossthe six product-price combinations. For between 5% to 18% ofrespondents, the sum of the purchase probabilities was less thanone, while for between 15% to46%, the sum was greater thanone. In absolute terms, the sumsranged from .1 to 5.8 forCampbell's Soup, and from .1 to15.0 for Coca Cola.

To deal with this anomaly, each

respondent's purchase probabilitieswere weighted so that the probabilities did sum to one. However,the finding that people were quitehappy to assign high probabilitiesof purchasing different numbers ofitems raises questions about theirunderstanding of the concept of"odds" or "chances out of 10".This phenomenon might also helpexplain why the predictions produced by this method are lessaccurate than those produced bythe Constant Sum method. Furtherresearch is currently investigatingalternative ways of employing themultiple question method to minimise this problem, and to reduce


the demands it makes on respondents who have purchased largequantities of product.

A further point to note is thatthe estimates of purchase levelswere derived from demand curvesconstructed using data from justtwo price points and estimates ofthe "average" prices for the products. This procedure rests on theassumption that the demand curveover the price range under consideration can be approximated adequately by a straight line, andwhether this is so requires furtherinvestigation.

ConclusionsThe purpose of this study was

to examine the accuracy of theJuster Scale for predicting purchase rates (number of purchasers)and purchase levels (number ofpurchases), of branded products, atvarious prices. The accuracy ofprediction was very different forthe two products. The predictionsof both purchase rate and purchaselevel for Coca Cola were veryaccurate, while those for

Campbell's Red and White labelsoup were very inaccurate.

Although the accuracy of predictions was assessed for only twoproducts, the study has demonstrated that both methods can beused to accurately predict purchaselevels for at least some brandedfood products. While the constantsum method produced the moreaccurate predictions, the resultsfrom the two methods were comparable, and it would be prematureto ignore the multiple questionmethod, particularly since the useof the tokens in the constant summethod rules it out for telephonesurveys.

A major concern is that theaccuracy of predictions variesgreatly from product to product,regardless of which method isused. Further research is requiredto more fully define the situationsin which predictions of either purchase rate or purchase level can beaccurately made. There is someevidence that predictions are greatly affected by whether a product is

a regular or irregular purchase,whether the product has a relatively high or low market share, andwhether the respondent is the mainpurchaser of the product. Theeffects of each of these factorsrequires further investigation.

In spite of all of the limitationsto the results obtained from thissmall preliminary study, the methods used do hold promise. Whilethey have not yet been shown to bereliable methods for estimatingactual purchase rates or actual purchase levels, it is worth remembering that managers often do notrequire accurate estimates of eitherpurchase rates or purchase levels,but are confronted by the need tochose between alternative decisionoptions. An alternative use for themethods examined in this studycould be to provide data for ranking alternatives on the basis ofestimated sales volume or marketsize, and this possibility alsodeserves further investigation.


ReferencesBelson, W A (1981). Design and Understanding of Survey Questions. Gower, London.Clawson, C J (1971). How useful are 90-day purchase probabilities. Journal of Marketing, 35,43-47.Cook, W A (1987). Telescoping and memory's other tricks. Journal of Advertising Research, Feb/March,

27,RC5-RC8.Day, D, Gan B, Gendall, P & Esslemont D (1991). Predicting purchase behaviour. Marketing Bulletin, 2, 18-30.Esslemont D, Hamilton-Gibbs D, & McGuinness, D (1992). Estimating demand for frequently purchased items.

Marketing Bulletin, 3, 18-23.Gabor, A & Granger, C W J (1972\73). Ownership and acquisition of consumer durables: report on the

Nottingham Consumer Durables Project. European Journal of Marketing, 6, 234-248.Gendall, P, Esslemont, D, & Day, D (1991). A comparison of two versions of the Juster Scale using self-com

pletion questionnaires. Journal of the Market Research Society, 33(3), 257-263.Isherwood, B C & Pickering, J B (1975). Factors influencing individual purchases of motor cars in Great

Britain. Oxford Bulletin of Economics and Statistics, 37, 227-249.Juster, F T (1966). Consumer buying intentions and purchase probability. Occasional Paper 99, National Bureau

of Economic Research, Colombia University Press.Neter, J & Waksberg, J (1964). A study of response errors in expenditure data from household interviews.

Journal of the American StatisticalAssociation, 59, 18-55.Parfitt, J H (1974). A comparison of purchase recall with diary panel records. Journal of Advertising Research,

7, 16-31.Pickering, J F & Isherwood, B C (1974) Purchase probabilities and consumer durable buying behaviour. Journal

of the Market Research Society, 16(3),203-226.Sudman, S & Bradburn, N (1974). Response effects in surveys: A review and synthesis. Chicago: Aldine.


Documents

Using the Juster Scale to Estimate the Demand-Price Relationship