STEVEN BELLMAN, GERALD L. LOHSE,AND ERIC J. JOHNSON

Predictors ̂Online Buying Behavior

What personal characteristics predict whether ornot people buy on the Net? Look for a ''wired" lifestyle and

time starvation, not demographics.

CONSUMERS WORLDWIDE CAN SHOP ONLINE 24 HOURS A DAY, SEVEN DAYS A WEEK, 365 DAYS

a year. Some market sectors, including insurance, fmancial services, computer hardware and

software, travel, books, music, video, flowers, and automobiles, are experiencing rapid growth

in online sales [7]. For example, in Jan. 1999, Dell Computer Corp. was selling an average of

$14 million of equipment online per day [3], and Amazon.com has become the third largest

bookseller in the U.S., despite being in business only since 1995 [6]. With projections that the

Internet will generate consumer and business-to-business sales in excess of $294 billion by 2002

[2], online retailing raises many questions about how to market on the Net.

Shopping there is far from universal, even forthose already online for other tasks. To help iden-tify the factors influencing online shopping, theWharton Forum on Electronic Commerce{ecom.wharton.upenn.edu) sponsors an ongoingresearch project begun in 1997 called the WhartonVirtual Test Market {WVTM). It seeks to under-stand Web consumer demographics, attitudesabout online shopping, and predictors of onlinebuying behavior. Ihe major difference between the1998 survey we conducted of WVTM membersand other surveys is that its focus was on buying onthe Net, trying not just to describe, but to model,

3 2 December I 999/Vol 42. No. I 2 COMMUNICATIONS Or TME ACM

the relationship between individual Net users' per-.sonal characteristics and their buying activity. TheNX'VTM is a panel, most of whose members agreedto participate in further research. Although the1998 survey established a baseline, we expect futureresearch to examine differences in shopping behav-ior over time.

The Wharton Forum on Electronic Commercebegan recruiting a panel of Web users from aroundthe world in October 1997. As part oi the participantregistration process, 10,180 people completed a sur-vey asking 62 questions about online behavior andattitudes about Internet communication and privacy

issues, as well as rotitine demographic questions. Peo-ple were attracted to the survey site (wvtm.whar-ton.upenn.edu) through an online banner advertisingcampaign targeting specific segments of Web usersworldwide, links provided by members' sites, andword of mouth. Survey respondents self-selectedthemselves as panel members, and their answers tosurvey questions were self-reported. Our analysis ofthe response data is a snapshotof the factors associated withbuying online and the amountsof money online buyers arewilling to spend there.

Self-selection and self-reporting also represent twolimitations of the survey. Weaddressed self-selection bymatching the survey sampledemographics to known popu-lation data. While the panel isnot a random sample, we tar-geted member recruitment tomatch the Media Metrixpanel, which is a random sam-ple of 10.000 PC-owning U.S.households used for projectingaudience estimates of Websites, and a U.S. Census sam-ple (collected in 1990),according to such demo-graphic information as gender, age, income, andeducation level. The size and demographics of theWVTM compare favorably with other samples ofWeb users recruited online, such as the regularGeorgia Tech Graphics, Visualization &: Usability(GVU) Center surveys of the Internet population(www.gvu.gatech.edu/ gvu/user_surveys/) [5, 9] andrandom sampling, such as the Media Metrix panel(wwTv.media metrix.com/).

Self-reported data is subject to the fallibility ofpersonal memories, idiosyncratic scale use, and evendeliberate alteration by way of social desirabilitybiases. We took steps to ensure that all respondentsanswered the WVTM questionnaire only once by,for example, using cookie technology to recordwhether or not they had already completed and sub-mitted their responses, minimizing the possibility ofcollecting multiple entries from the same person. Allrespondents also had to provide email addresses thatwere valid at the time of completion. Still, addi-tional research is needed to compare self-reportedbehavior against actual Web use.

WVTM Characteristics and theOnline PopulationFigure 1 compares the percentage of males andfemales from the U.S. participating in the WVTMwith the percentages in other stirveys and to the per-centage in the overall U.S. population according cothe most recent U.S. Census (1990). However, allthese Internet surveys agree that the online popula-

tion is relatively younger, more educated, wealthier,and has fewer African-Americans than the overallU.S. population, although the gaps are graduallyclosing. The median age of the WVTMs U.S.members is 29, or just younger than the overall pop-ulation median (30 to 34 years). The median edu-cational attainment of these U.S. members is "somecollege," compared to the population median level("high-school graduate"). The median householdincome of the U.S. members and the ninth GVUsurvey (April-May 1998) is $35,000 to $49,999. arange that includes the U.S. population median($35,225 in 1990), though lower than the medianfor the Media Metrix panel (Julyl997-June 1998)($50,000 to $74,999). Among the WVTMs U.S.members, 85% are white, a percentage not much dif-ferent from that in the overall U.S. population(82%) and less than the percentage of whites in theGVU sample (88%). However, African-Americansconstitute only 3% of the WVTM population andonly 2.3% of the GVU sample, compared to 12% inthe overall U.S. population. These comparisons

COMMUNICATIONS OF THE ACM December 1999/Vol. 42, No 12 33

Figure I. Distribution of males and femalesfrom the U.S. in the WVTM survey (Nov. 1997-

May 1998). the 1990 U.S. Census, the MediaMetrix panel (July 1997-June 1998), and the ninth

GVU survey (Apr. -May 1998).

56% 59%51%

46% 44%41%

Males Females

WVTM U.S. Census I Media Metrix GVU

show that the opinions of WVTM members closelymatched those of the overall U.S. online community,as shown through the comparison with the MediaMetrix panel and with the overall U.S. population.

The WVTM panel includes participants from 82countries; see Table 1 tor a breakdown by continent.Having a worldwide panel allows us to compare theonline behavior and attitudes of U.S. Web users withtheir counterparts in Europe, Asia, and elsewhere. Inmany countries, the Internet was introduced onlyrecently, and the growth rate of Net use has beenmuch slower, making it difficult to gauge the repre-sentativeness of a sample. But when grouped byregion, the responses of WVTM members may indi-cate differences in Web use and attitudes shared bythe Web users in similar commercial environmentsand cultures around the world.

Online BehaviorAlthough survey respondents report connecting tothe Web more frequently at the office and atschool, more of their Web-use hours are at home,with 21% reporting spending more than 20 hoursper week on the Web browsing from home. Most ofthese people connect from home through 28.8 and33.6 KB/sec modems, and from work or school viacable and direct connections {such as Ethernet). Themost common regular use for the Internet {morethan once per week) is for work—at home (52.3%)and at work (37.8%). The Internet is also used reg-ularly at home to read news (19.1%) and for enter-tainment (10.8%). Less regularly (less than once aweek), the Internet is put to more varied uses.

including looking for computer software and share-ware (6.7%), banking (6.4%), and personal financeand investment information (3.8%). Only 2.9% ofWVTM members say they look for product infor-mation, using the Internet less than once a week.Only 3.1% say they use the Internet for shopping.

Who Buys Online?The survey asked members of the WVTM samplewhether they had ever bought anything online.Roughly 4,368 (42.9%) said they have never boughtanything online. We analyzed the factors that pre-dicted actual purchases using logistic regression [8].We took advantage of the large sample size by split-ting the WVTM randomly into two separate halvesto produce separate calibration and holdout sam-ples. The calibration sample used a stepwise proce-dure to add candidate variables to the logisticregression equation. The final regression for the cal-ibration sample was then used to predict whethermembers of the holdout sample would or would notbuy anything online. This process was repeated,using the original holdout sample to calibrate theregression, then cross-validating this regressionequation in the original calibration sample. For ahighly significant 66% of members of the WVTM,the resulting equation correctly predicted whetheror not they bought online. Figure 2 shows the vari-ables that were significant in both runs of this dou-

Table 1. Continents represented in WVTM survey.

ContinentAfricaAsiaAustralasiaEuropeRest of North AmericaSouth AmeriaU.S.

N28

172199415562

288.749

Percent

0.3%1.7%2.0%4.1 %5.5%0.3%

86.2%

ble cross-validation procedure, arranged indecreasing order of importance.

A Wired LifestyleLooking for product information on the Internet isthe most important predictor of online buyingbehavior. More interesting, perhaps, are the resultsfor the other predictors of online buying. They showthat a typical online buyer has a "wired" lifestyle.Such people have been on the Internet for years, not

34 December 1999/Vol 42. No, 12 COMMUNICATIONS OF THE ACM

Looking for productinformation on theInternet is the mostimportant predictorof online buyingbehavior.just a few months. They receive a large number ofemail messages every day, they work on the Internetin their offices every week, and they agree that theInternet and other developments in communicationtechnology have improved their productivity atwork. Just as they use the Internet for most of theirother activities {such as readingthe news at home), these peoplenaturally turn to the Internet tosearch for product informationand in many cases to huy prod-ucts and services.

Another influence on a per-son's decision to shop online isthe amount of discretionary timethey have. As the total number ofhours worked by members of ahousehold increases, the less timethere is to search for and buyproducts and services in the tradi-tional way (by, say, visiting brick-and-mortar shops). Moreimportant, this effect is evenstronger if one's spouse alsoworks. Dual-income householdsseek new ways to fmd informa-tion and buy things that are fasterand more convenient. In the past,they may have used catalogs; nowthey take advantage of e-com-merce sites on the Web.

We asked members of theWV'i'M who bought thingsonline how many online transac-tions they had made during thepast six months and the value of their most recentonline transactions. The median number of transac-tions completed during that time was two in the U.S.and Asia. But in Europe, the median was only onetransaction during the preceding six months. The

median amount spent on the last transaction was $30in the U.S., Asia, and Europe—or about the same asthe average purchase from Amazon.com ($35) [4], butmuch less than other more conservative estimates of theaverage online transaction amount. For example,BizRate.com, which collects consumers' ratings of

Figure 2. Factors that predict buying vs. not buyingonline, in decreasing order of influence.

Standardized Estimates0.05 O.I 0.15 0.2 0.25

Look atproduct info

Months online

Number ofdaily emails

Work online atwork every week

Read news online athome every week

Total householdworking hours

Click on banners

Agree that Internetimproves productivity

online businesses {www.bizrate.com). estimated in late1997 that the average online purchase for first-timeshoppers was $109.

Figure 3 shows the distribution of the WVTMannual purchase value. The smaller the number of

COMMUNICATIONS OF THE ACM December 1999/Vol 42, No 12 3S

We can projecttotal onrinespending couldincrease to$23.7 billionin 2000.transactions, the more likely a smaller amount ofmoney is involved as well. That is, the variance inannual value increases in proportion to an increase inannual value. The median amount spent onlineannually in 1997-1998 by the U.S. members of thesample was $200. European members spent $240,and Asian members spent $160. New users in theWVTM sample (those on the Internet two years orless) tend to spend less online per year compared withexperienced users ($163 vs. $290).

These estimates are again conservative when com-pared to the $600 that members ot the GVU sampleclaimed to have spent online in 1997 [5]. Assumingthere are 79 million people online in the U.S. andCanada, of whom 20 million buy online [1], the totalvalue of online buying annually in the U.S. and

Figure 3. Annual purchase value of U.S.members of the WVTM.

$l-$3

$3-$7

$7-$20

$20-$55

$55-$l50

$I5O-$4OO

$400-$ 1.000

$l.000-$3,000

$3.000-$8,000

$8,000-$20,000

>$20,000

Canada was approximately $4 billion. From this data,we can project that total online spending couldincrease to $23.7 billion in 2000, or about the sameas Emarketer's $26 billion prediction tor 2002 [2],but higher than the $7 billion projected for 2000 byForrester Research, and much less than the $115 bil-lion estimate by Morgan Stanley [11].

We also used another regression model to exam-ine factors ptedicting annual spending online. First,we isolated the subsample of online buyers from theWVTM and again split this subsample in half ran-domly, using double cross-validation. The fmairegression equation could explain only 11% of thevariance in annual value, but this percentage is sig-nificant given a sample oF this size. The variablesincluded in the final equation are listed in Figure 4,

in decreasing order of importance;the length of each bar indicates thecontribution one unit of each vari-able added ro the annual value of amembet of the WVTM.

Two influences again seem tobe working in these variables—the wired lifestyle and time star-vation. The wired lifestylecharacterizes users who like to bethe first to use the latest commu-nication technologies, spendmore hours online than most,and turn to the Internet to emailtheir orders even when shoppingfrom print catalogs. Time starva-tion is reflected in the negativecoefficient for mailing in ordersfrom print catalogs. For the peo-ple who spend more moneyannually online, regular physicalmail is just too slow. Time-starved people, as well as thoseliving a wired lifestyle, look torproduct intormation on the

25% 30% 35%

36 December 1999/Vol 42. No 12 COMMUNICATIONS OF THE ACM

Figure 4. Influence (in $U.S. of annual spending) for factorspredicting annual online spending by WVTM members.

ber of hours worked by the house-hold), so we were unable to test theinfluence of such variables outsidethe WVTM. Their probableimportance can be gauged byexamining the proportion of vari-ance they explain in the WVTMcompared to the variables sharedwith the WVTM and the GVU.Indicators of time starvation (order-ing from online catalogs and fromprint catalogs online) contribute ahighly significant 8.5% of totalexplained variance in online trans-actions in the WVTM.

Predicting the Lack ofOnline Buying1 he first of several surprises inthese results is that demographicsalone do not seem to influencewhether or not people buy online,nor the amount of money theyspend there. Demographics havesome influence on whether or nota person is online in the first placecompared with the rest of theoverall national population; theyare. for example, more likely to hewhite and more highly educated.However, once people are online,whether they buy there and howmuch they spend has more to do

Internet, and a lot of what they see, they buy. with whether they like being online and whether theTo show that these influences are also predictive of time they have for buying things elsewhere is lim-

online buying behavior by people who are not mem- ired. However, even these general lifestyle character-bers of the WVTM, we ran similar regressions using istics explain only a small proportion of why peopledata from the ninth GVU survey (collected April-May buy online and the amount of money they spend1998) [5]. Again, the wired lifestyle variables—num- there.ber of months on the Internet, hours online per week,hours per week working online, searching for productinformation online, and the attitude that email is

Order from catalogsusing the Internet

Use Internet at officeregularly for work

Search for productinfo online

Order fromcatalogs

Like being first to usenew technologies

Number ofyears online

Agree that Internetimproves productivity

Hours per week online

Number of daily emails

Not ordering by mail

$0.00 $0.50 $1.00 $1.50Change in Annual Value

$2.00

indispensable—predict buying and not buying for77% of this sample. Because the entire GVU ques-tionnaire was not compulsory, we maximized rhe avail-

In the GVU data, demographics do show a slightinfluence: The higher a person's income, education,and age, the more likely that person will buy online,and the higher a persons income, the more onlinetransactions that person is likely to make. But thisinfluence is barely significant; demographics alone

able sample size by predicting the number of online predict 1.2% of decisions to buy or not buy and onlyi (TV 6 ) h h htransactions (TV = 7,670), rather than the amount

spent online (A^= 1,141). (In the WVTM, number oftransactions and amount spent were highly correlated,and the same variables predict both outcomes.) Thesame wired lifestyle variables explain 28% of annualonline spending (note that "use of online news"replaces "agree that email is indispensable"). The GVUsurvey did not ask about time starvation (such as num-

0.3% of the variance in the number of purchasesmade by online buyers. These results match Bndingsfrom consumer behavior studies in other media inwhich demographics and lifestyle variables explainonly a small percentage of people's choice behavior[10]. The most important information for predictingshopping habits—online and offline—are measures ofpast behavior, not demographics.

COMMUNICATIONS OF THE ACM December I 99VVol. 42. No. I 2 37

Another surprise is the unimportance of privacyissues as predictors of buying vs. not huying onlineand the amount of money spent there. When thepeople who responded to the ninth GVU survey(April-May 1998) were asked why they boughtthings online, 39.1% said they were concerned aboutthe security of credit card information [5]. However,in the same survey, only 1.9% of respondentsreported having had a bad experience with buyingonline, and 76.2% had ordered something online,though ordering something online does not necessar-ily involve transmitting credit card information.

Security and privacy issues are important to allWVTM members. For example, they indicated theirconcern about the monitoring of online behavior andthe exchange of information by third parties-—<on-cern shared equally in the U.S., Europe, and Asia. Forexample, when deciding to send personal informationto a particular site, 49% of the U.S. WVTM mem-bets, 47% of the European members, and 46% of theAsian members said that whether or not a site usesencryption software was an important consideration.In any case, such concerns did not affect tbe decisionto shop online and are not significant predictors ofeither the decision to shop online or the amount ofmoney spent there. Thus, while they are concerned,WVTM members indicated that security and privacyconcerns are increasingly less important predictors ofshopping behavior.

ConclusionThe prototypical Web consumer leads a wiredlifestyle and is time starved. So it seems that Webconsumers shop online or use online services to savetime. This result suggests several implications for thedesign of online shopping environments:

• Sites should make it more convenient to buystandard or repeat-purchase items (such as theone-click-to-purchase approach at Amazon.comand at 1-800-flowers, www.1800flowers.com);

• Customization should provide the informationneeded to make a purchase decision; and

• The checkout process should be easy for the con-sumer [6].

A more global conclusion might well be that theseconsumers are much more like catalog shoppersthan Kmart shoppers, that is, they seem to value theWeb's time savings over its cost savings. While thisconsumer attitude may change over time, conve-nience, rather than cost savings, may be a key bene-fit offered by successful online stotes.

A final issue, which deserves to be examined in fur-

ther research, concerns the future of e-commerce ingeneral. Our results show that people who spendmore money online have a more wired lifestyle, arc onthe Net more, and receive more email compared toother hiternet users. As Net use diffuses throughoutthe worldwide population, this profile is becomingincreasingly common. But can we also expect a paral-lel increase in online shopping? The notion thatonline consumer commerce is a "following," ratherthan a "leading," use of home computers is bestanswered through "longitudinal" surveys oi the mem-bers of the WVTM; we'll be doing just that in ourfuture research. B

REFERENCES1. Brocrsma M. Ca-ching! Suddenly Web shopping is a growth biwinew.

ZDNei News (Aug. 24. 1998); sec www.zdnei.tom/zdnn/stories/ztlnn_smgraph_display/0,34436,213l 140,00.html.

2. Deck, S. Study sees growth in online shopping. Cimijwtetti'orid (May 21,1998); see www.computcrworld.com/homc/onlinc9697.nsf/all/98O52!studylF956.

3- Dell Computer Corp. Fi.ua/ 1999 in Rn'icu. Annual Report, RoundRock. Tex; .see www.dell.com/corporate/acceM/aiinu.il tepnrt.'i/report99/Oell_1999_FiscalJn_Rcview.pdf.

4. Jimnarkar, S. Study looks at Amazon's future. Ndi'.i.aim {Aug. 20,1998); secwww.news.com/News/Itcm/Textonly/0,25,25516.00.html?ptv.

5. Kehoe. C, Pitkow, J., and Rogers J. GVlJ\s Ninth WWW iher Suri'eyRrpiiri. Office ot leclinology Licensing, Georgia Tech Research Corp.,Atlanta, 1998.

6. Lohse, G. and Spillcr, I'. Electronic shopping: How do customer inter-faces produce sales on the Internet? Cammun. ACM 41. 7 (]\i\y 1998),81-87.

7. Meeker, M. and Pearson, S. The hiternel Refiii/ins Repm-r. Morgan Stan-ley, Dean Witter, Discover & Co., New York (May 28, 1997); secwww.ms.com/insight/misc/inctrctail.htmi).

8. Neter, |. (Ed.), Kutner, M,, and Nachtshcim, C. ApplieJLinear liej>rei-sion Muiltls. in/El/. Irwm, Chicago, 1996.

9. Pitkow, J. and Kehoc, C. Emerging trends in the WWW user popula-tion. Commiin. ACM. 39. 6 (June 1996), 106-108.

10. Rich, S. and Subhash J. Social class and life cycle as predictors of shop-ping behavior./ Mkrg. Res. 5. 1 (Feb. 1968), 41-49.

11. Secretariat for Electronic Commerce. The Emerging Digital Economy.Report, U.S. Department of Commerce, Washington, D.C. (Apr.1998); see www.ecommetce.gov/emerging.htm.

STEVEN BELLMAN (stevenb^markedng,wharton.upenn.edu) is aresearch fellow in the Wharton Forum on Electronic Commerce at theUniversity of Pennsylvania in Philadelphia.GERALD L . LOHSE (lohsdAvharton.upenn.edu) is the researchdirector of the Wharton Forum on Electronic Commerce at theUniversity of Pennsylvania in Philadelphia.ERIC J. J O H N S O N (ejj3@columbia.edu) is a professor of marketingin the Columbia School of Business at Columbia University in NewYork.

Permission lo make iiij;ital or hard copies of all or part of this work for personal or class-nx)m use is granted without tec provided that copies are not made or distribiitptl forprofir or commertia! atlviiniagc and thai topics bear ihis notici- and the full ciiation onthe first page. To i-opy otherwise, to republish, to post on servers or to redistribute tolists, requires prior specific permission and/or t fee.

@ 1999 ACM 0002-0782/99/1200 $3.00

38 December 1999/Vol 42. No 12 COMMUNICATIONS OF THE ACM