Вы находитесь на странице: 1из 37

The Effects of Banner Advertising on Consumer Inter-purchase

Times and Expenditures in Digital Environments

Puneet Manchanda
Jean-Pierre Dubé
Khim Yong Goh
Pradeep K. Chintagunta

January 2002
This version: June 2002

Puneet Manchanda is Kilts Center Fellow and Assistant Professor of Marketing, Jean-Pierre Dubé
is Assistant Professor of Marketing, Khim Yong Goh is a doctoral student and Pradeep K.
Chintagunta is Robert Law Professor of Marketing at the Graduate School of Business,
University of Chicago. The authors would like to thank France Leclerc, Stijn M. J. van Osselaer,
Wendy Moe and Young-Bean Song for feedback; John Mracek, Steve Findley, Nancy Niu and
Susan Gertzis for help on data issues; two anonymous firms for providing the data and the Kilts
Center for Marketing at the Graduate School of Business, University of Chicago, for research
support. Dubé would also like to acknowledge research support from the Beatrice Foods Faculty
Fund at the University of Chicago. All correspondence may be addressed to the first author at the
University of Chicago, Graduate School of Business, 1101, East 58th Street, Chicago, IL 60637
or via e-mail at puneet.manchanda@gsb.uchicago.edu.

One of the major advances of the digital economy is the facilitation of building and managing
individual customer relationships. A fundamental aspect of this relationship is customer retention
(i.e., repeat purchasing). In this study, we focus on the use of banner advertising for customer
retention and shopping expenditure. Using a rich database consisting of repeat customer
purchases at a website along with individual advertising exposure, we measure the impact of
banner advertising on purchase behavior. In measuring the effectiveness of banner advertising,
we explicitly account for the weight, timing, diversity of creatives and the number of sites on
which individual consumers are exposed to banner advertising.

Our analysis is based on a database consisting of repeat customer purchases at a website along
with individual advertising exposures. We specify an individual-level joint purchase timing and
expenditure model as a function of advertising exposure. The purchase timing model investigates
the role of advertising in inducing repeat purchases by allowing for temporal separation between
exposure and visits. The purchase expenditure model investigates the effect of advertising on
consumer spending during purchase visits. We find substantial evidence of temporal separation
between advertising exposure and subsequent purchase behavior, i.e., exposure to advertising
leads to purchases in the future. Specifically, we find that advertising weight, copy, and timing
affect consumers’ decision to revisit websites and make purchases. Past expenditure seems to be
the best predictor of current expenditure on a shopping trip. We find evidence of considerable
heterogeneity across consumers in response to advertising. Our results suggest that there are
three segments of consumers in the data. These segments are differentially affected by the
frequency and recency of banner advertising as well as monetary value of their past purchases.
For practitioners, our findings suggest that instantaneous measures such as click-through rates
provide an inadequate measure of advertising effectiveness. In terms of advertising research in
general, we provide some unique evidence that advertising affects purchase behavior of repeat

Keywords: Advertising Response, Digital Environments, Banner Advertising, E-commerce,

Internet Retailing, Hazard Models, Latent Class Models

1. Introduction

One of the major advances of the digital economy is the facilitation of building and managing
individual customer relationships. A fundamental aspect of this relationship is customer retention
(i.e., repeat purchasing). The electronic mediation of interactions between consumers and firms
vastly expands the sources and quality of data available to firms, and thus the ability to manage
this relationship. These expansive on-line databases enable a plethora of new targeted-marketing
stimuli, including customized promotions/advertising, prices, product recommendations, and
product assortments. Despite the quick adoption of on-line marketing by practitioners,
remarkably little is known about the size of the effects and the potential payoffs of such efforts.
Though there are some empirical studies investigating the role of pricing (see Smith et al. 1999
for an excellent review) and product assortment (e.g., Haubl and Trifts 2000), the role of
advertising in digital media has not been investigated in detail. In the current analysis, we focus
on the use of banner advertising for one aspect of customer retention. In particular, using a rich
database consisting of repeat-customer purchases at a website along with individual advertising
exposures, we measure the impact of banner advertising on accelerating consumers’ repeat
purchases at that website. Additionally, we measure the effects of such advertising on
consumers’ expenditures at that website.
The Internet is beginning to emerge as a viable advertising medium (Silk et al. 2001). On-
line advertising has become an important component of the Internet economy and the advertising
industry in general. The latest estimate of total industry expenditure on digital media in 2000 is
$8.2 billion (Internet Advertising Bureau). These numbers compare favorably with expenditure
on more established media such as Outdoor1 - $1.8 billion in 2000 - and Cable TV - $11.2 billion
in 2000 (BusinessWeek On-line 2001b). In addition, contrary to popular belief, advertising on
the Internet has not declined dramatically in 2001 – total expenditure in the first three quarters of
2001 is estimated at $5.6 billion (Internet Advertising Bureau). In fact, Forester Research
predicts that total expenditure on digital media will grow rapidly to about $42 billion by 2005 as
the number of consumers in digital environments grows worldwide (BusinessWeek On-line
2001b). While several forms of advertising in digital environments have emerged, industry
reports indicate that the majority of digital advertisements are banner advertisements (Cho et al.

This term refers to advertising media such as billboards, hoardings etc.

2001, On-line Media Planning Report 2001, IAB 1999, Hofacker and Murphy 1998). A banner
advertisement is a section of on-line advertising space that is generally 480 x 60 pixels in size. It
typically consists of a combination of graphic and textual content and contains a link to the
advertiser’s website via a click-through URL (Uniform Resource Locator), which acts as a web
address. Given the magnitude of the Internet advertising sector, and banner advertisements in
particular, measuring the effectiveness of banner advertising is likely to be of considerable
interest to both academics and practitioners.
Interestingly, the effectiveness of banner advertisements has been debated since the early
stages of Internet commerce.1 Websites hosting on-line ads have been pushing for traditional
“exposure” based metrics, such as “impressions” served, to allow them to charge for each banner
exposure. However, difficulties in measuring on-line impressions precisely have caused much
dissatisfaction amongst managers resulting in reluctance to commit funds to banner advertising
(Hoffman and Novak 2000). Moreover, advertisers, who prefer to pay based on the performance
of their ads, feel that impressions generally overstate advertising effectiveness. Instead,
advertisers have been pushing for heuristic metrics of performance such as “click-through”,
which indicates when a web surfer clicks through to the advertiser’s URL via the banner.
However, the effectiveness of click-through as a valid measure is also being called into question
(Briggs 2001, BusinessWeek On-line 2001a, Song 2001). The fact that typical click-through
rates are quite small in magnitude, 0.5% on average (Sherman and Deighton 2001, Dahlen 2001,
Warren 2001), has led practitioners to believe that banners are ineffective. Moreover, click-
through is a measure of a visit to the website. Since there is a lot of evidence that a very small
proportion of visits translate into final purchase (Moe and Fader 2001), click-through may be too
imprecise for measuring the effectiveness of banners served to the mass market. These studies
therefore, underscore the importance of investigating the impact of banner advertising on actual
purchase behavior.
From an academic point of view, click-through may also be a poor measure of advertising.
Prior research in marketing has shown that the effects of advertising are not necessarily
immediate (IAB 1997, Smith and Swinyard 1982, Vakratsas and Ambler 1999). In other words, a
temporal separation may exist between exposure and action. Click-through, being an

Note that electronic commerce via private value-added networks (VANs) has existed since the mid 1980s. Our
focus is specifically on Internet commerce, which only emerged after the commercialization of the World Wide Web
during the mid 1990s.

instantaneous metric (Chatterjee et al. 1998), does not reflect this separation and may
subsequently understate the effectiveness of advertising. The temporal separation could occur as
exposure results in immediate cognitive (e.g., memory) and/or attitudinal (e.g., attitude) effects
that influence behavior after some time. Several recent industry and academic studies have
documented attitudinal changes in response to banner advertising. The IAB On-line Advertising
Effectiveness Study (1997) showed that exposure to banner advertising increased advertisement
awareness, brand awareness, and purchase intention. Engage, Adknowledge’s On-line
Advertising Report (2001) showed that consumers who have been exposed to banner advertising
are more likely to interact with the advertisers’ web sites. Gallagher et al. (2001) found that
consumer response in terms of awareness and purchase intention is similar for an identical
advertisement regardless of whether consumers are exposed to it in a digital environment or a
conventional print environment. Finally, Dahlen (2001) showed that increased exposures result
in higher brand awareness and attitude for inexperienced users. These findings highlight the
importance of incorporating the possibility of temporal separation between advertising exposure
and response into a model that measures advertising effectiveness.
Taken together, the previously described studies provide the basis for analyzing the effects of
banner advertisements. First, we need to look at the effects of these ads on actual consumer
purchase behavior rather than on intermediate measures. Second, the analysis should account for
the possible temporal separation between advertising exposure and purchase behavior.
Even if banner ads are successful in persuading a consumer to make a first time purchase at
an e-commerce website, the long-term viability of the site depends critically on the ability to
“retain” the consumer. In other words, repeat visit and purchase behavior of consumers at the
website are of critical importance. Accordingly, our focus in this paper is on that specific aspect
of customer retention i.e., the ability of banner ads to influence repeat purchases by accelerating
such purchases in time. We also investigate whether banner ads influence the level of consumer
expenditures at the time of purchase. We formulate a model of individual behavior as a function
of advertising exposure. This model comprises two sub-models. The first sub-model, a purchase
timing model, investigates the role of advertising in accelerating repeat purchase behavior by
allowing for temporal separation between exposure and purchases. The second sub-model, an
expenditure model, investigates the role of advertising on consumer expenditure during purchase
visits. We also allow for various facets of advertising, such as weight, timing, creative, and the

number of sites on which the consumer was exposed to advertising, to affect timing and
expenditure behavior differentially. Our proposed model also controls for unobserved individual
differences by specifying a non-parametric distribution of support points for the vector of
estimated parameters. This approach enables us potentially to identify the sizes of sub-groups in
the population that may be influenced by banner ads to a lesser or greater extent. In terms of the
broader area of research on the effects of (any type of) advertising on individual consumers, our
work adds to the studies that have investigated the effects of advertising on purchase timing
behavior. We find substantial evidence of temporal separation between advertising and
subsequent purchase behavior i.e., exposure to advertising leads to purchases in the future.
Specifically, we find that advertising weight, copy, and timing affect consumers’ decision to visit
websites and make purchases. In addition, we also find that environmental factors and browsing
behavior have an effect on site visits. We find evidence of considerable heterogeneity across
consumers in response to advertising. Our results suggest that there are three segments of
shoppers – loyal but infrequent shoppers, impulse shoppers and frequent shoppers – in the data.
These segments are differentially affected by the frequency and recency of banner advertising as
well as monetary value of their past purchases. We also find that click-through is a poor measure
of advertising effectiveness as it accounts for a very small proportion of overall purchases.
Finally, contrary to the findings in earlier studies (e.g., Deighton et al. 1994), we find that
advertising does have an effect on repeat purchases. We also discuss how the estimated
parameters provide guidelines to managers on the use of banner advertising to stimulate response
and build extended customer relationships.
The structure of the paper is as follows. We briefly discuss prior work in this area in Section
2. We give an overview of the data in Section 3. Section 4 contains the details of the models we
estimate. We discuss the results and the managerial implications of our findings in Section 5.
Section 6 concludes the paper with a discussion of the limitations of the present study and
provides directions for future research.

2. Literature Review

The role of advertising in digital environments is an emerging area of research.2 We provide an

overview of academic research in Table 1. Most of the academic (see studies by Dahlen, Cho et.
al. and Gallagher et. al. in Table 1) and industry research on advertising in digital environments
has focused on measuring changes in brand awareness, brand attitudes, and purchase intentions
as a function of exposure (as against the effects of banner advertising on actual purchase
behavior). This is usually done via field surveys or laboratory experiments using individual (or
cookie) level data.
In contrast to studies using experimental data, Sherman and Deighton (2001) describe the
process of serving banner advertisements and collecting response data in detail. They also report
the results of an experiment carried out by a web advertising agency and an on-line merchant that
showed that targeting advertising to specific customers and websites increases response rates and
drives down the average cost-per-action (due to confidentiality restrictions, they report only
broad, aggregate level findings).
The studies in Table 1 deal with data from digital environments. An important study that
investigates the effect of television advertising on repeat purchases using individual level data is
by Deighton et al. (1994). This study derives a model that is based on theoretical findings in the
literature and estimates it on single source data from the ketchup, liquid detergent and powder
detergent markets. Using results from prior studies (e.g., Raj 1982), the study predicts that
advertising will have an influence on repeat purchases due to a “framing” effect. In other words,
advertising enhances repeat purchase probabilities by influencing what consumers learn from
product usage experience. However, the study finds that advertising influences brand choice but
has no effect on repeat purchases (for individual brands). The authors postulate that this
unexpected result may be due to the nature of the product categories (mature), data (collinearity)
and/or the model (the logit formulation).
To summarize, studies that have examined the effects of conventional (television) advertising
on repeat purchases have found little or no effect (Deighton et. al. 1994). Banner advertising
research on the other hand has largely been limited to their influence on attitudes rather than on

For research on the effects of advertising in conventional media, we refer the reader to two excellent review papers
– Lodish et al. (1995) and Vakratsas and Ambler (1999).

behavior. Additionally, there has been no research on the influence of these ads on repeat
purchasing behavior.3 In our paper, we address this gap in the literature by modeling data that
reflects the purchase behavior of repeat customers as a function of banner advertising exposure.
Findings from industry research (Businessweek Online 2001a, 2001b, Tran 2001, Song 2001,
IAB 1997, DoubleClick Press Release 2001, Warren 2001, Briggs 2001) show that banner
advertising has attitudinal effects and that click-through is a poor measure of advertising
response. These findings are generally consistent with the findings of the academic research
discussed earlier. Interestingly, in addition to the attitudinal effects of banner advertising, we find
a few studies that provide some informal evidence of behavioral effects as well. In this paper, we
use a formal model to investigate these behavioral effects for repeat customers.
Other recent modeling research in marketing has focused on describing browsing behavior
(or clickstream) data (see Table 1). These studies typically use activity data from web log files
and/or surveys and therefore do not capture the effects of marketing instruments on sales.
Two studies investigate different aspects of within-site activity. Chatterjee et al. (1998) use a
multinomial logit model to predict click-through on banner advertisements. They find that the
probability of clicking through on a given visit is unaffected by exposures on prior visits. In
other words, click-through measures an instantaneous consumer response. They also find that the
propensity to click follows a non-monotonic relationship with the number of exposures. Bucklin
and Sismeiro (2001) examine web site “stickiness” using page choice and visit duration (given
page choice) using a joint probit-hazard model. They find evidence of stickiness i.e., as
consumers look at more pages on the web site, they spend more time on each page. They also
find some evidence of “learning” behavior whence consumers choose fewer pages on repeat
visits but spend the same amount of time per page. Importantly, they find that browsing behavior
varies across consumers.
Bhatnagar and Ghose (2002) use survey data to investigate duration and frequency of search
on the Internet. They model duration and frequency using a joint discrete hazard-ordered probit
formulation. They find that consumers search for longer durations when they seek price
comparison data, availability of product/services or information from vendors. Consumers spend

There are very few studies that have documented the effects on advertising on purchase behavior of repeat
customers. For more details on the few other studies that have investigated this question directly or indirectly, we
refer the reader to the Deighton et al. (1994) paper mentioned above.

the maximum time searching for information from vendors, followed by availability of
product/services, and price comparisons.
Finally, Moe and Fader (2001) investigate conversion rates on websites by modeling the
proportion of visits that result in a purchase.4 They find that, for a site with very high conversion
rates,5 most on-line buyers are not directed buyers. Also, as these buyers make more visits to the
web site, the impact of these visits on purchase behavior diminishes. In addition, they find that
purchasing thresholds increase as a function of discounted purchasing experiences, implying that
a consumer is likely to repeat-purchase soon after a transaction occurs. Conversion probabilities
are also found to be decreasing over time, as a function of past visit experiences and as a result of
increased purchasing experiences. They also find evidence of significant heterogeneity in
consumers’ conversion propensities. These findings are partially consistent with industry
findings that conversion rates are higher for returning customers relative to first-time visitors, for
reputable and older web sites, and during the holiday season (CNET News.com 2001).
The above studies show that there exists a growing interest in modeling consumer behavior
in digital environments. The sophisticated techniques used in the studies described above have
uncovered interesting and useful patterns in the data, e.g., evidence for unobserved
heterogeneity. However, these studies have typically focused on activity data such as browsing
or click-through and thus have not been able to quantify the effect of marketing instruments on
sales. Our research builds on these studies by applying a formal model of purchase timing and
expenditure (with unobserved heterogeneity) to data that reflect consumer behavior from
advertising exposure to actual purchases.

Since they do not have actual purchase data, they infer a purchase if the consumer was logged at the Order
Confirmation page.
Note that their data reflect a conversion rate of 14%. These data come from Amazon.com, which is known to have
much higher conversion rates (~ 15%) relative to the rest of the industry. Industry studies show that rates of 2%
to5% are fairly typical while rates close to 10% are considered remarkable (CNET News.com 2001).

3. Data

The data come from an Internet-only firm engaged in selling healthcare and beauty products as
well as non-prescription drugs to consumers.6 The data span a period of three months during the
third quarter of 2000, specifically from June 11th to September 16th. The data are available at
the individual cookie level. As mentioned earlier, most datasets used to investigate on-line
environments usually comprise of browsing behavior only. Our data are unique in that we have
individual level stimulus (advertising) and response (purchase amount). The data are contained
in three databases - labeled CAMPAIGN, TRACER and PURCHASE. These are described in
brief below.
• The CAMPAIGN database comprises the on-line advertisement banner exposure and
click-through response originating from promotional campaigns that were run on web
sites. The data fields in the CAMPAIGN database consist of consumer data - a unique
cookie identifier identifying the individual computer,7 an indicator variable denoting
consumer response to the banner advertisement (view or click),8 and the date and time of
banner view or click; and advertising data – the portal or alliance site’s web page where
the banner advertisement view or click occurred, and a unique key identifying the
specific banner advertisement. Note that though we have a unique identifier for each the
site on which the banner advertisement was served, we do not know the specific identity
of each site. In terms of the advertising message contained in the various creatives, we
know that the majority of the messages were of the brand-building type for the website
(i.e., the message consisted of the name of the web site and a line describing the benefits
of purchasing from the website). A limitation of the data is that we do not have
information on the specific message in each banner (even though we have an indicator
that tells us that one creative was different from another).

The data were processed and made available to us by the firm that was responsible for serving the advertisements
for the firm in question. Due to the nature of the data sharing agreement between us and the firms, we are unable to
reveal the name of either firm.
We use the term consumer and cookie identifier interchangeably for the sake of exposition. However, as
mentioned earlier, our data only allow us to identify a unique computer and not a unique consumer.
Note that since we are working with behavioral data, we are unable to control for the exact nature of exposure. In
other words, we are making the assumption that if the consumer was on a specific page and the banner appeared on
that same page, s/he actually viewed the advertisement. This assumption is consistent with prior research that has
tried to document the effect of advertising exposure on sales for individual consumers (see detailed discussion in
Deighton et al. 1994, p. 34).

• The TRACER database contains the date and time of the purchase transaction for each
unique cookie identifier.
• The data fields in the PURCHASE transactions database consist of the total dollar value
of the transaction (we do not have information on the prices and quantities of each
purchase occasion) and a unique key that links these values with the unique cookie
identifier in the TRACER database.

The CAMPAIGN database contains records of the company’s advertising on portal and
alliance web sites such as Yahoo!, AOL, Women.com, iVillage.com, Healthcentral.com, and
E*Trade. These sites accounted for over 80% of all advertising activity by the firm during this
period. Advertising activity typically consisted of a specific creative that operated over several
weeks. This creative was delivered to websites in the form of a digital graphic, generally referred
to as a GIF. These GIFs were of the usual size for banner advertisements (480 x 60 pixels).9
New GIFs were typically released at the beginning of a calendar week i.e., on Sunday and/or
Monday, reflecting media buying patterns. During the period covered by our data, there were one
hundred total GIFs spread over fifteen major sites. However, the majority of exposures came
from a small number of GIFs – seven GIFs accounted for about 55% of all exposures.
We merged the PURCHASE database and the TRACER database using the unique merge
keys in both databases. We then deleted transactions with blank cookies, repeat transactions
(identical transactions at identical times), transaction with erroneous entries (e.g., with negative
expenditure) and then matched the remaining transactions with the campaign dates. Since we are
interested in repeat purchase behavior in response to advertising exposure, we extracted the
transactions with corresponding cookie entries in the campaign database and with transaction
dates later than or equal to the last recorded banner advertisement exposure or click.10 We then

The majority of the firm’s banner advertising used standard banner sizes. A very small proportion (< 5%) of
advertising consisted of short banners (392 x 72 pixels) and vertical banners (120 x 240 pixels). Our data does not
provide information on the size of a specific banner advertisement in the data.
The firm provided us click-through information only if it resulted in a purchase. This was done to minimize the
data processing effort and the size of the resulting database. The firm’s advertising agency confirmed that the mean
click-through rate during this period was between 0.25%-0.5% (in our sample, the click-through rate is close to
0.25%). This is consistent with the rates documented in other studies discussed earlier (though this rate is somewhat
lower). In addition, it seems clear from the data that click-through only purchases are an order of magnitude smaller
than purchases driven by banner advertising (1134 purchases versus 15754 purchases across all purchasers). These
data, combined with feedback from the firm’s executives, lead us to conclude that click-through is not an important
path to purchase for customers of this website.

retained only those customers (cookies) for whom we observe at least two purchases (we need
two purchases to compute an inter-purchase time). This resulted in a sample of 2192 unique
cookies.11 The mean number of purchase occasions in the data for these cookies is 3.23 with a
standard deviation of 0.80. The mean inter-purchase days for these customers were 36.29 with a
standard deviation of 28.28 (see Figure 1 for the distribution of inter-purchase times). This
distribution seems reasonable given that the firm sells health and beauty products. The mean
expenditure was $29.42 with a standard deviation of $37.86. In our model (details provided in
the next section), we use the log expenditure as the dependent variable – the mean log
expenditure was 3.10 with a standard deviation of 0.78 (see Figure 2 for the distribution of log

4. Modeling the Customer-Website Relationship

We investigate two major decisions made by the repeat purchasers in our data – when to visit the
website and how much to spend. These are modeled as a function of the frequency at which these
consumers were exposed to advertising, the temporal lag between the visit and the last seen
advertisement, and previous spending patterns. The two decisions are captured via a hazard
model of inter-purchase times (when to visit) and a flexible quantity model (expenditure). To
capture variability in individual choices, we impose a flexible mixing distribution on the
estimated parameters.

4.1 Model Formulation

The 2192 repeat purchasers were selected (using the at least two purchase occasions criterion) from a total of
14370 purchasers on whom we had purchase data (after combining the various databases). Though this may not
seem to be a very large proportion, note that most modeling studies using data from digital environments have
reported that the number of retained cookies is very small (both in proportions and absolute levels) once purchase or
exposure criteria are used. For example, Moe and Fader (2001) report that only 851 out of 10000 panelists made at
least one purchase on Amazon.com over an eight month period. The average number of purchases by these
consumers in this time period was 1.8. Similarly, Chatterjee et al. (1998), are able to retain only 1056 out of 21,783
registered panelists given at least one exposure over an eight month period. This is true even in purchase incidence
studies that use scanner data, e.g., Bucklin and Lattin (1991) report that heavy buyers in the saltine cracker category
made purchases on 612 trips out of a total of 9,999 trips over a two-year period and Bucklin and Gupta (1992) report
2275 purchases in 30,966 trips in the liquid detergent category over a two-year period.

As noted above, our model formulation jointly specifies the time to purchase at the site in
question along with the expenditure incurred when the purchase occurs. We use a hazard
function specification (Jain and Vilcassim 1991, Kalbfleisch and Prentice 1980) to model the
inter-purchase times and a regression model for the expenditures (Blattberg and Neslin 1990,
Neslin et al. 1985). Earlier modeling research using customer browsing data has found evidence
of heterogeneity (Moe and Fader 2001, Bucklin and Sismeiro 2001). We therefore account for
heterogeneity using a latent class approach (Kamakura and Russell 1989). Specifically,
households are allowed to vary in the parameters of the timing and expenditure models. Note that
we are aware of only one other study that has estimated a purchase timing-purchase expenditure
model - Neslin et al. (1985). However, their model was lacking on two dimensions. First, they
use a regression framework to model inter-purchase times. As has been shown in subsequent
literature, this can lead to biased estimates (Helsen and Schmittlein 1993). Second, they did not
account for unobserved heterogeneity. Our model addresses both these shortcomings.

4.1.1 Specification of the baseline hazard for the purchase-timing model

Consistent with the previous literature on hazard functions, the marginal hi(ti) depends upon the
baseline hazard, effects of covariates and unobserved heterogeneity. Research on univariate
hazard functions suggests that the baseline hazard function must be sufficiently flexible to allow
for potential non-monotonicity (see Jain and Vilcassim 1991). Several possible approaches
include the Box-Cox formulation (Flint and Heckman 1982), the log-logistic formulation
(Chintagunta 1998, Kalbfleisch and Prentice 1980), the inverse Gaussian formulation
(Kalbfleisch and Prentice 1980) and the semiparametric method (Meyer 1990). We found that
the log-logistic formulation outperformed the other specifications based on the criterion of model
fit. Hence, we choose the log-logistic distribution for inter-purchase times. The baseline hazard
hi(ti) for consumer i who makes a purchase at the site after time ti has elapsed from the previous
purchase is given as follows:
ω -1
ωi ρiωi ti i
hi (ti ) =
1 + ( ρi ti )ωi

where the parameters of the log-logistic distribution are ωi, ρi > 0. We define ωi = exp(γ0i) and

ρi = exp(γ1i).

4.1.2 Effects of covariates

Two approaches have been suggested in the marketing literature to account for the effects of
covariates on the purchase timing decision. The first approach is the proportional hazards
approach. Here it is assumed the effects of covariates increase or decrease the level of the hazard
in a proportional manner without influencing the shape of the hazard function. Let ψi(Xit) > 0
denote the function that incorporates the effects of covariates for consumer i at time t. Then the
hazard function is simply the product of the baseline hazard and the effects of covariates.

λi(ti) = hi(ti) * ψi(Xit)

To illustrate, suppose ψi(Xit) increases from 1 to 2, then the hazard (proportionately) doubles.
However, the proportional hazards model does not allow the marketing variables to influence the
shape of the hazard function. To alleviate this restrictive feature, researchers have used the
accelerated failure time model, where the covariates directly influence the purchase times. In the
case of the log-logistic model described above, the effects of covariates will enter via ωi or ρi
(allowing for covariates to affect both parameters leads to identification problems). We tried both
specifications and found the model fit was improved by reparameterizing ρi as (ψi(Xit)*ρi ).
Hence, the specification of the hazard function that we use in the paper is as follows:

ω -1
ω (ψ ( X ) ρ )ωi t i
hi (ti ) = i i it i i ωi
1 + (ψi ( X it ) ρi ti )

For tractability, we specify the function ψi(Xit) > 0 as ψi(Xit) = exp(Xit γi). This specification
is consistent with the previous literature (e.g., Chintagunta 1998).

4.1.3 Specification of the regression equation for expenditure

When a consumer makes a purchase at the site, we observe the total expenditures incurred by
that consumer on that purchase occasion. We relate the (natural logarithm of) total expenditure Y
incurred by consumer i who makes a purchase at time ti to a set of covariates Zit as follows.

Y = Zit βi + εit

In the above expression we assume that εit has a normal distribution with zero mean and
unknown variance σ2. Such a specification is standard in the literature (Blattberg and Neslin

4.1.4 Unobserved heterogeneity

Jain and Vilcassim (1991) use a single parameter operationalization of unobserved heterogeneity.
Further, they find that it may not be appropriate to impose a specific parametric form on the
distribution of such heterogeneity across households. Consequently, they specify a discrete
distribution for the unobserved heterogeneity and then estimate the support points and
probability masses for the distribution from the data. While we follow the discrete distribution
approach, we also allow for the effects of the covariates to vary across consumers. Additionally,
we allow the parameters of the expenditure equation to vary across consumers following a
discrete distribution. Let Θi = { γ1i , γ0i ,γi , βi } be the vector of unknown parameters across the

purchase timing and expenditure decisions. Since the distribution of this vector across consumers
is unknown, we approximate it with a discrete distribution consisting of a finite number of

supports, Q (i.e., { γ1iq , γ0iq ,γiq , βiq }, q=1, 2,..., Q) and their associated probability masses πq, q = 1,

2, …, Q. We turn next to the estimation procedure.

4.2 Estimation Procedure

The likelihood function for consumer i, conditional on the parameters Θi is given by:
Li Θi = ∏ { gi * f i (ti )} ri {Si (ti )} ri 
δ 1-δ

r =1
 


Ri is the number of spells for consumer i;

gi is the density function for the univariate normal regression model;

fi(ti) is the density function corresponding to the log-logistic hazard;

Si(ti) is the survivor function corresponding to the log-logistic hazard;

1 if rith spell ends in a purchase

δr = 
0 otherwise.

The term involving Si (ti ) accounts for right censoring of the data, as the end of the data
collection period usually does not coincide with a purchase for all consumers. Also note that the
expenditure component occurs only for completed spells when consumers make a purchase at the
The unconditional likelihood for household i is obtained by integrating over the distribution
G(Θ) of the random variable Θ. Doing so, we obtain

Li = ∫ Li Θ dG (Θ)

While a specific multivariate parametric form may be assumed for the distribution of Θ, we
follow instead the procedure suggested by Heckman and Singer (1984) and implemented by Jain
and Vilcassim (1991). In this approach, no specific parametric assumptions are made for G(•).
Instead, the distribution of Θ is determined empirically from the data by approximating it by a

discrete number, Q of supports Θq, q = 1, 2, …, Q and their associated probabilities πq, q = 1, 2,

…, Q. Consequently, the consumer’s unconditional likelihood above can be written as

Li = ∑ Li Θq π ( Θq )

Standard maximum likelihood methods can be used to estimate the model parameters.

4.3 Model Specification

The temporal sequence of events for a typical repeat purchaser (see Section 3) in our data is
illustrated in Figure 3. The data reflect an individual consumer (cookie)'s purchases (as well as

the final "no purchase" which is captured through the survivor function included in the
likelihood). In between purchases (or the final "no purchase"), this consumer is exposed to a
variety of banner ads on different sites and at different times. Thus there is within-consumer
variation in the effect of advertising on purchase timing. Once we use all the repeat purchasers,
we have cross-sectional (across-consumer) variation in the effects of advertising on purchase
timing. These two sources of variation allow us to measure the effects of advertising on purchase
timing (and purchase expenditure).
We postulate that the decision of when to purchase and how much to purchase will be
affected by advertising factors (weight and quality). For repeat customers, we postulate that
banner advertisements act as reminders and/or brand builders via framing and may lead to higher
purchase intentions (Deighton et al. 1994, Cho et al. 2001).12 We also need to control for
observed differences across consumers in terms of prior purchase behavior and browsing
We construct three sets of variables to model the timing and quantity decision. The first set
of variables reflects advertising exposure:
• VIEWLAST represents the total number of advertising exposures between last purchase
occasion and current purchase occasion for each customer,
• VIEWLAG represents the lag time (in days) from most recent ad exposure until purchase,
• FEATURE is an indicator variable that is set to 1 if the consumer is exposed to banner
advertising on the day on purchase; 0 otherwise,
• ADNUM represents the number of creatives (GIFS) that the consumer was exposed to
between the last purchase and current purchase occasions.

Prior research has shown that repeated exposures to an advertisement prevent the early decay
of advertising effects (Pechmann and Stewart 1988, Cacioppo and Petty 1985). We therefore
expect that increased exposure to advertising (VIEWLAST) should lead to lower inter-purchase
times (higher hazard). In addition, as we expect the effects of advertising to diminish over time,
the longer the lag from last exposure (VIEWLAG), the higher the inter-purchase times (lower
hazard). In terms of the variety of creative execution, prior research has indicated that response

As mentioned earlier, almost all of the banner advertisements advertised the site and the benefits of shopping at
the site.

to different creatives can be quite different (Lodish et al. 1995). It has also been shown that recall
is enhanced if consumers are exposed to different creatives in the same campaign (Rao and
Burnkrant 1991). However, in our case, while all the creatives essentially advertise the website,
they are not part of a single campaign. We therefore have no prediction regarding the effect of
ADNUM on inter-purchase times. Finally, a special case of the lag between last view and
purchase is when both occurred on the same day (11% of purchase observations in the data). We
use the FEATURE dummy to capture this and expect that the hazard will be higher if FEATURE
is one. For expenditure, since people who shop on the same day as they are exposed are likely to
be impulse (or promotion sensitive) shoppers, we expect a negative sign on the coefficient for
FEATURE in the expenditure equation. We do not have any prior hypotheses for the other three
The second set of variables reflects the individual consumer’s purchase behavior:
• LEXPLAG represents the log lagged (last period) purchase expenditure,
• WEDDAY is an indicator variable which is set to 1 if the day of purchase is a
Wednesday; 0 otherwise.

We expect the hazard to be lower if the expenditure was higher on the last purchase occasion
(LEXPLAG). For expenditure, we expect that lagged expenditure will control for cross-sectional
heterogeneity (this is similar to the quantity loyalty variable in Tellis 1988, except that we use
dollar value and not units) i.e., we expect that the sign will be positive.13 Also, preliminary
analysis showed that a higher proportion of sales took place on a Wednesday so we used a
dummy variable to indicate purchase on a Wednesday and expect the hazard to be higher on that
We use the number of sites (SITENUM) on which each consumer was exposed to advertising
as a proxy for that consumer’s browsing behavior.14 SITENUM represents the number of unique
web sites on which the consumer was exposed to advertising between purchase occasions.15

It could be argued that lagged expenditure is a proxy for product inventory. However, we think that this is
unlikely since the firm sold products in many different product categories at different price points.
Note that the firm and its advertising agency were not targeting advertising to individual cookies based on prior
browsing behavior. Thus the use of this variable is not susceptible to the simultaneity bias.
Another measure that we could have inferred from the data is the total number of pages on which the consumer
was exposed to advertising on these sites. However, when we computed the number of pages, we found that on 90%

Prior research has shown that viewing a series of advertisements leads to higher recall and
more positive attitudes (Pechmann and Stewart 1988, Zielske and Henry 1980). We therefore
expect that the hazard of purchase is higher for consumers exposed to advertising on many
different websites (SITENUM). We do not have any hypothesis for the effect of prior browsing
behavior on expenditure. We summarize the expected signs for each of the variables in our
specification in the table below:

Variable Hazard Exp


Table 2 provides the descriptive statistics for the covariates.

Insert Table 2 about here.

5. Results

5.1 Model Estimates

In this section, we describe our metric for choosing the number of segments,16 the in-sample
overall model fit and then discuss the segment level parameters in detail.
We found evidence of considerable heterogeneity in the data. We used the Bayesian
Information Criterion (BIC) to choose the number of segments that provided the best fit. The
BIC for a one, two, three and four segment solution was 6687.54, 7310.59, 7609.64 and 7604.69
respectively. This indicated that a model with three segments provided the best fit relative to two
or four segments. Note that the BIC (as a function of the number of segments) follows an

of the viewing occasions, consumer viewed the advertising on one page on the site. The correlation between number
of sites and the number of pages is 0.96.
We use the labels “segment” and “support” interchangeably throughout the paper.

inverted U shape in this case since the dependent variables are not constrained. In contrast, the
BIC follows a U shape in discrete choice models where the dependent variable is constrained
between 0 and 1.
In terms of absolute in-sample fit, we compute the U 2 statistic using the Akaike Information
Criterion (Ben-Akiva and Lerman 1985, p. 91) and we find it to be 0.40. This compares fairly
well with U 2 statistics reported for brand choice models and is better than those reported for
purchase incidence models.17 Thus, it seems that model fits the data reasonably well. In addition,
this implies that there is no evidence that consumers visit the site at fixed intervals i.e., the
covariates have a significant effect resulting in a modification of the baseline hazard.18 The signs
of the covariates are mostly as expected (see Table 3 for a summary).

Insert Tables 3-4 about here.

The overall pattern of the results indicates that advertising weight, timing, creatives and the
number of sites on which this advertising appears affects inter-purchase times. These effects are
mostly as predicted. Advertising weight (the number of advertisements exposed to between
purchases) has a positive effect on the hazard, leading to lower inter-purchase times. The recency
of the last exposure (time since last exposure) has a negative effect on the hazard, leading to
higher inter-purchase times. The effect of creative is somewhat surprising, in that exposure to a
higher number of creatives has a negative effect on the hazard, leading to higher inter-purchase
times. Finally, exposure to banners advertisements at a higher number of websites has a positive
effect on the hazard, leading to lower inter-purchase times.
We now discuss the results for each segment in greater detail. As can be see from Table 4,
Segment 1 and 3 are about the same size while Segment 2 is very small.19 Segment 1 is the
largest segment (mass 0.501) and may be loosely described as comprising of “loyal but
infrequent” shoppers. The scale parameter that characterizes the hazard, ω1 = exp(γ01), is greater

17 2
Bucklin and Lattin (1991) report a U of 0.21 for a combined purchase incidence and brand choice model while
Bucklin and Gupta (1992) report a U of 0.38 for the brand choice component and 0.21 for purchase incidence.
As is usual in this literature (e.g., Jain and Vilcassim 1991), we interpret the baseline hazard as capturing
regularity in inter-purchase times net of covariates. Thus, the time at which the hazard peaks after a purchase is the
underlying inter-purchase time.
Even though Segment 2 is small, there is a large jump in the BIC once that segment is added. In addition, this
segment is distinctly identified regardless of the overall number of segments (when more than one segment is

than one resulting in a non-monotonic hazard function (see Figure 4). The baseline hazard
(without covariates) peaks at 63 days. This implies that, in the absence of covariate effects,
consumers tend to make purchases every 63 days. In terms of the covariates, we find that the
hazard is affected positively by the number of exposures, expenditure on the previous shopping
trip, Wednesday and the number of sites on which the consumer was exposed to advertising.
The hazard also increases on the day of exposure. The time since last exposure had a negative
effect on the purchase hazard rate. The diversity of the creatives has a negative effect on the
hazard. These effects are as expected. An examination of the expenditure parameters shows that
advertising plays no role in affecting the amount spent on the website. As mentioned earlier, this
segment seems to comprise of loyal customers whose last trip expenditure is a good predictor of
current expenditure.
Segment 2 is a small but unique segment and may be described as a segment of “impulse”
shoppers. The shape parameter for the baseline hazard for this segment is less than one resulting
in a monotonic (declining) hazard function (see Figure 4). The hazard for this segment is affected
positively by the number of exposures, the Wednesday effect and the number of sites that
consumers were exposed to the advertising. In contrast to Segment 1, there was no significant
increase in hazard on the day of exposure. Also, in contrast with Segment 1, advertising affects
the expenditure behavior. We find that higher exposures, time since last exposure both lead to
lower expenditures. The expenditure is also lower on the day of exposure. In terms of the effect
of browsing behavior on expenditure, there is a positive effect of Wednesdays and a negative
effect of different sites on which consumers were exposed.
Finally, the shape parameter for Segment 3 is also greater than one resulting in a non-
monotonic hazard. The baseline hazard peaks at twelve days and thus this segment may be
described as comprising of “regular and frequent” shoppers. The hazard of purchase is lower for
this segment the longer the time since last exposure. It is also lower for a higher number of
creatives but is higher for Wednesdays and for exposure on a higher number of sites.
Expenditure for this segment is affected positively by the time since last exposure, the number of
creatives and the amount spent on the last visit. What sets this segment apart from Segment 1 is
the frequency with which consumers in this segment make purchases. While consumers in
Segment 1 are less frequent shoppers (regularity of purchases is 63 days), consumers in this
segment tend to be more frequent shoppers.

In summary, we find that, in our data, advertising weight, copy and timing affect consumers’
decision to visit websites and make purchases. In addition, we also find that environmental
factors and browsing behavior have an effect on purchase visits. In terms of purchase
expenditure, we find that past expenditure is a good indicator of expenditure on a given visit. Our
findings on the effect of advertising on purchase expenditure are mixed and it does not seem
possible to draw unequivocal conclusions. We also find that these effects apply differentially to
different sets of consumers.

5.2 Managerial Implications

Our findings above have several implications for managers. First, we do find evidence of
temporal differences between exposure and action for a majority of the consumers in our sample.
This result implies that managers can correctly evaluate the effectiveness of advertising
campaigns only if they account for these temporal differences. This finding also argues against
the use of instantaneous metrics such as click-through to measure advertising effectiveness.
Second, our findings also provide some insight into the nature of advertising response. These
insights are based on the estimated elasticities (see Table 5). We used the estimated parameters
to derive the elasticities. For the effect on purchase timing, we computed the hazard using the
given data. We then increased the value of the covariates by 1% and measured the subsequent
change in the hazard (Gupta 1991, Jain and Vilcassim 1991). We then computed the elasticity for
each covariate. For the expenditure, we computed the elasticity at each observation using the
analytical expression and report the means.
In general, the elasticities are all less than one. In terms of the advertising effects, we find
that the time since last exposure and the number of creatives have a much larger effect size on
purchase timing relative to the number of exposures. This result has implications for advertising
copy and ad timing. In terms of copy, exposing the same consumer to several (unrelated)
creatives may be less beneficial than a single creative (given the assumption that all the creatives
are of the same quality). This finding complements the findings cited earlier on the effects of
exposure to many creatives within the same campaign. In terms of the timing of advertising, it
may be more beneficial to expose consumers to a series of evenly spaced ads relative to massed
exposures. Given that we find a strong same day purchase effect given exposure, advertisers

could potentially use banner advertisements to smooth out sales and run special promotions.
Finally, there seem to strong positive benefits of ensuring that customers are exposed to the same
advertisement across many websites.
In terms of the expenditure elasticities, the largest effect size is from lagged expenditure.
This implies that managers need to ensure that once consumers reach the site, their shopping
baskets should be as large as possible (in dollar terms) e.g., they could offer “point of checkout”
promotions to increase basket expenditures. Advertising across many sites however has a small
negative effect on expenditure. This result is in contrast to the finding that exposure on more
sites leads to shorter inter-purchase times (see above). Thus, managers need to trade off the
benefits of media plans that expose consumers to advertising on many sites with the negative
impact on expenditure. Finally, shopping expenditures seem positively affected when consumers
are exposed to more creatives. One possible explanation for this phenomenon is that different
items could be promoted in each creative. However, exposing a consumer segment to many
creatives leads to higher inter-purchase times. Thus, managers should optimize the number of
creatives by considering the marginal effect on both inter-purchase times and shopping
Insert Table 5 about here.

Finally, we find that purchase behavior in digital environments is characterized by

differences among consumers. As can be seen from Table 3, the effects of advertising and other
variables differ across the three segments. One interesting way to isolate the marginal effects of
the independent variables on the mean inter-purchase times would be look at the total hazard
with covariates (computed at the mean level of all covariates) and then set each covariate to zero
and re-compute the hazard and contrast the day at which the hazard is maximal for each scenario.
This method may be useful for managers since it provides a sense of the relative impact on mean
inter-purchase times in days.
Using this method, we find that the change in the number of advertisements from zero to
the mean level allows the hazard to peak a day earlier for Segment 1, increases the hazard by
49% for Segment 2 and causes no change for Segment 3.20 The change in peak hazard due to

For Segment 2, the hazard is always at its maximum on the first day. We therefore report the percentage change in
hazard on day 1 for this segment in this analysis.

change in time since last exposure is a delay of twenty days for Segment 1 and a delay of
eighteen days for Segment 3 (Segment 2 is not affected by this change). Lagged expenditure
causes a reduction in peak hazard by ten days for Segment 1 and no change for the other two
segments. Finally, viewing the advertising on different websites caused a reduction in peak
hazard of eight days for Segment 1, an increase in the hazard of 92% for Segment 2 and
reduction in peak hazard by eleven days.
Given these findings, managers may find it useful to classify the segments on the variables
under the RFM framework - recency of last exposure, frequency of exposures and lagged
monetary value.21 Segment 1 (the largest) is affected by all three variables. Thus, building a
relationship with this segment requires optimizing over all three variables. In contrast, Segment 2
is affected by only by recency. Finally, Segment 3 is affected by recency and monetary value.
Thus, segment level targeting plans could be designed and implemented.

6. Conclusion

Our research fits into a small, but fast growing subfield of empirical research dedicated to
measuring how the Internet provides new marketing opportunities in areas such areas as pricing,
product assortment decisions and advertising. This paper is the first attempt, to the best of our
knowledge, to model the effects of advertising in digital environments on purchase behavior. We
use a unique dataset to investigate the effect of banner advertising on inter-purchase times and
shopping expenditures. Our results show evidence of temporal separation between advertising
and purchase behavior. We speculate that the temporal separation exists because advertising acts
as a brand building tool and/or a reminder. We also find that measures of instantaneous behavior
such as click-through may be a poor measure of advertising effectiveness.
Specifically, we find that advertising weight, copy and timing affect consumers’ decision to
visit websites and make purchases. The time since last exposure seems to have the largest effect
on consumers’ inter-purchase times while the effect of advertising weight is small.
Complementary to previous research, we find that the greater the number of (unrelated) creatives
The RFM framework is usually used in the direct mail industry – these firms use behavioral data on the recency of
purchase, frequency of purchase and last purchase monetary value to describe consumer segments. In our case, the
analogous variables are the recency of banner advertising, the frequency of banner advertising and the last purchase
monetary value.

to which a consumer is exposed, the longer is the inter-purchase time. However, inter-purchase
times are lower if consumers are exposed to the firm’s advertising on many different websites.
For shopping expenditure, we find that past expenditure seems to be the best predictor of current
expenditure on a shopping trip. However, the effects of advertising, as expected, are not
unequivocal on the amount purchased during a shopping trip.
We also find evidence of considerable heterogeneity across consumers in response to
advertising. Our results suggest that there are three segments of consumers in the data. These
segments are differentially affected by the frequency and recency of banner advertising as well
as monetary value of their past purchases.
Finally, in terms of the broader area of research on the effects of (any type of) advertising,
we provide somewhat unique evidence that advertising does affect the purchase behavior of
current (repeat) customers.
We would also like to note some limitations of our research. These limitations arise primarily
from the lack of information in our data. First, we note that our results may not apply to
customers who have not purchased items at least twice at this website. Second, we do not have
any demographic information on the cookies. This information may have been useful in
generating segment membership a priori. Third, our results would have been richer if we had
information on the actual message contained in each advertisement and the identity of the
referral sites. Fourth, we do not have any knowledge of the other marketing variables such as
price and promotion during consumers’ site visits. These limitations may be addressed in future
research by running formal field experiments (as in Lodish et al. 1995) or by obtaining richer
datasets that provide natural variation on these dimensions.


Ben-Akiva, M. and S. Lerman (1985), Discrete Choice Analysis, MIT Press, Cambridge, MA.

Bhatnagar, A. and S. Ghose (2002), “An analysis of frequency and duration of search on the
Internet,” Journal of Business, forthcoming.

Blattberg, R. C. and S. A. Neslin (1990), Sales Promotion – Concepts, Methods, and Strategies,
Prentice Hall.

Briggs, R. (2001), “Abolish Clickthrough now!”, from < http://www.intelliquest.com/

resources/whitepapers/abolishclick.pdf >.

Bucklin, R. E. and J. M. Lattin (1991), “A two-stage model of purchase incidence and brand
choice,” Marketing Science, 10(4), 24-39.

Bucklin, R. E. and S. Gupta (1992), “Brand choice, purchase incidence and segmentation: An
integrated modeling approach,” Journal of Marketing Research, 34(5), 201-215.

Bucklin, R. E. and C. Sismeiro (2001), “A model of web site browsing behavior estimated on
clickstream data,” Working Paper, The Anderson School, University of California, Los

BusinessWeek On-line (2001a), “A chat with a true believer,” Special Report: Internet
Advertising, July 12, 2001.

BusinessWeek On-line (2001b), “On-line advertising: It’s just the beginning,” Special Report:
Internet Advertising, July 12, 2001.

Cacioppo, J. T. and R. E. Petty (1985), “Central and peripheral routes to persuasion: The role of
message repetition,” in Psychological Processes and Advertising Effects, A. Mitchell and
L. A. Alwitt, eds., Erlbaum, Hilsdale, NJ.

Chatterjee, P., D. L. Hoffman, and T. P. Novak (1998), “Modeling the clickstream: Implications
for web-based advertising efforts,” Project 2000 Vanderbilt Working Paper, Owen
Graduate School of Management, Vanderbilt University.

Chintagunta, P. K. (1998), “Inertia and variety seeking in a model of brand purchase timing,”
Marketing Science, 17(3), 253-270.

Cho, C., J. Lee, and M. Tharp (2001), “Different forced-exposure levels to banner
advertisements,” Journal of Advertising Research, 41(4), 45-56.

CNET News.com (2001), “Getting Clicks with Casual Customers”, June 4, 2001, from <
http://news.cnet.com/news/0-1007-201-6162000 0.html?tag=st.ne.1007.saslnk. saseml >.

Dahlen, M. (2001), “Banner advertisements through a new lens,” Journal of Advertising

Research, 41(4), 23-30.

Deighton, J., C. M. Henderson and S. A. Neslin (1994), “The effects of advertising on brand
switching and repeat purchasing,” Journal of Marketing Research, 31(2), 28-43.

DoubleClick (1996), “Research findings: Banner burnout”, from < http://www.

doubleclick.com/learning_center/research_findings/banner_burnout.htm >.

DoubleClick Press Release (2001), “Diameter’s campaign evaluation illustrates direct correlation
between exposure to on-line ads and brand impact,” July 30, 2001.

Engage Adknowledge On-line Advertising Report (2000), Third Quarter, from < http://
www.engage.com/company/oar_docs/oar_3rdqtr00.pdf >.

Flint, C. and J. Heckman (1982), “Models for the analysis of labor force dynamics,” in Advances
in Econometrics, Vol. 1, G. Rhodes and R. Basmann, eds., JAI Press Inc., Greenwich,
CT, 35-95.

Gallagher, K., K. D. Foster, and J. Parsons (2001), “The medium is not the message: Advertising
effectiveness and content evaluation in print and on the Web,” Journal of Advertising
Research, 41(4), 57-70.

Gupta, S., (1991), “Stochastic models of inter-purchase time with time dependent covariates,”
Journal of Marketing Research, 28(1), 1-15.

Haubl, G. and V. Trifts (2000), “Consumer decision making in on-line shopping environments:
The effects of interactive decision aids,” Marketing Science, 19(1), 4-21.

Heckman, J. and B. Singer (1984), “A method for minimizing the impact of distributional
assumptions in econometric models for duration data,” Econometrica, 52(2), 271-320.

Helsen, K. and D. C. Schmittlein (1993), “Analyzing Duration Times in Marketing: Evidence for
the Effectiveness of Hazard Rate Models,” Marketing Science, 11(4), 395-414.

Hofacker, C.F. and J. Murphy (1998), “World Wide Web banner ad copy testing,” European
Journal of Marketing, 32(7/8), 703-712.

Hoffman, D. and T. Novak (2000), “When exposure-based advertising stops making sense (and
what CDNOW did about it),” Working Paper, Owen Graduate School of Management,
Vanderbilt University.

IAB (1997), IAB On-line Advertising Effectiveness Study, from < http://www.
intelliquest.com/resources/reports/mbi_report02.asp >.

IAB (1999), IAB Internet Advertising Revenue Report, 1999 Second Quarter Results, from <
http://www.iab.netreports/PwC_IAB1999Q2.PDF >.

Jain, D.C. and N. J. Vilcassim (1991), “Investigating household purchase timing decisions: A
conditional hazard function approach,” Marketing Science, 10(1), 1-23.

Kalbfleisch, J. D. and R. L. Prentice (1980), The Statistical Analysis of Failure Time Data, John
Wiley & Sons, New York, NY.

Kamakura, W. A. and G. J. Russell (1989), “A probabilistic choice model for market

segmentation and elasticity structure,” Journal of Marketing Research, 26(11), 379-90.

Lodish, L. M., M. Abraham, S. Kalmenson, J. Livelsberger, B. Lubetkin, B. Richardson, M. E.

Stevens (1995), “How TV-advertising works - A meta-analysis of 389 real-world split
cable TV-advertising experiments,” Journal of Marketing Research, 32(2), 125-139.

Meyer, B. (1990), “Unemployment Insurance and Unemployment Spells,” Econometrica, 58(4),


Moe, W. W. and P. S. Fader (2001), “Dynamic conversion behavior at e-commerce sites,”

Working Paper, McCombs School of Business, University of Texas at Austin.

Neslin, S. A., C. Henderson and J. Quelch (1985), “Consumer promotions and the acceleration of
product purchases,” Marketing Science, 4(2), 147-165.

On-line Media Planning Report (2001), “The State of On-line Advertising”, February 2001, from
< http://www.adrelevance.com/intelligence/report20.pdf >.

Pechmann, C. and D. W. Stewart (1988), “Advertising repetition: A critical review of wearin and
wearout”, in Current Issues and Research in Advertising, J. H. Leigh and C. R. Martin,
Jr., eds., 11(1,2), Michigan Business School, University of Michigan.

Raj, S. P. (1982), “The effects of advertising on high and low loyalty consumer segments,”
Journal of Consumer Research, 9(June), 77-89.

Rao, U. H. and R. E. Burnkrant (1991), “Effects of repeating varied ad executions on brand name
memory,” Journal Of Marketing Research, 28(4), 406-417.

Seetharaman, S. and P. K. Chintagunta (2001), “The proportional hazard model for purchase
timing: A comparison of alternative specifications,” Working Paper, John M. Olin School
of Business, Washington University at St. Louis.

Sherman, L. and J. Deighton (2001), “Banner advertising: Measuring effectiveness and
optimizing placement,” Journal of Interactive Marketing, 15(2), 60-64.

Silk, A. J. , L. R. Klein, and E. R. Bernt (2001), “The emerging position of the Internet as an
advertising medium”, Netnomics, 3, 129-148.

Smith, M. D., J. Bailey, and E. Brynjolfsson (1999), “Understanding digital markets: Review and
assessment,” in Understanding the Digital Economy, Erik Brynjolfsson and Brian Kahin,
eds., MIT Press, Cambridge, MA.

Smith, R. E. and W. R. Swinyard (1982), “Information response models: An integrated

approach,” Journal of Marketing, 46(1), 81-93.

Song, Y. (2001), “Proof that on-line advertising works,” Atlas Institute, from < http://
www.atlasdmt.com/insights/dm.asp >.

Tellis, G. J. (1988), “Advertising exposure, loyalty and brand purchase: A two-stage model of
choice,” Journal of Marketing Research, 25(2), 134-144.

Tran, K. T. L. (2001), “A dictionary of on-line-marketing jargon,” Wall Street Journal, April 23,

Vakratsas, D. and T. Ambler (1999), “How advertising works: What do we really know,”
Journal of Marketing, 63(1), 26-43.

Warren, S. (2001), “Winners and Losers,” Wall Street Journal, April 23, 2001.

Zielske, H. and W. Henry (1980), “Remembering and forgetting television advertisements,”

Journal of Advertising Research, 20(2), 7-13.

Table 1: Overview of Research on On-line Advertising and Consumer Behavior
Product Dependent
Study Research Issue Type of Data
Category Variables
Our study Evaluate impact of Healthcare and Market data: Purchase timing
banner advertising beauty products - Advertising Purchase
on consumer on-line Non-prescription exposure expenditure
purchase behavior drugs - Actual purchase
Dahlen (2001) Study impact of Insurance product Experimental data: Brand awareness
brand familiarity Travel services - Laboratory Brand attitude
and Internet user Detergent experimentation Click-through rate
experience on Ice cream
banner ad Coffee
effectiveness Automobile parts
Cho, Lee and Evaluate effects of Consumer brands Experimental data: Brand awareness
Tharp (2001) different levels of Retail, financial, - Laboratory Attitude towards ad
forced exposure to travel services experimentation and brand
banner ads on Click-through rate
consumer response Purchase intention
Gallagher, Evaluate Tourism service Experimental data: Brand recall and
Foster and effectiveness of Coffee - Laboratory recognition
Parsons (2001) web-based ads, Arts and crafts experimentation Attitude towards ad
compared to print- and brand
based ads
Sherman and Evaluate efficacy of Healthcare products Market data: Web browsing
Deighton target on-line - Cookie data frequency
(2001) advertising - Advertising site Purchase incidence
response Conversion rate
Bhatnagar and Study relationship Information Survey data: Search duration
Ghose (2002) between on-line - On-line survey Search frequency
search patterns and questionnaire Type of information
type of information
Chatterjee, Model consumer High-technology, Browsing data: Click-through
Hoffman and click-through on durable goods - Click-stream logs response
Novak (1998) banner ads
Bucklin and Model web site Automobiles Browsing data: Web page choice
Sismeiro “stickiness” and - Click-stream logs Visit duration
(2001) consumer learning
Moe and Fader Model consumer Books Browsing data: Purchase incidence
(2001) rates of visit-to- - Click-stream logs (proxied)
purchase conversion Conversion

Table 2: Descriptive Statistics

Variable Mean Std. Dev.

IPDAYS 36.29 28.28
LEXPEND 3.10 0.78
VIEWLAST 7.10 52.42
VIEWLAG 32.20 28.44
ADNUM 1.65 2.39
FEATURE 0.08 0.27
LEXPLAG 3.03 0.81
WEDDAY 0.12 0.33
SITENUM 0.74 0.90
PAGENUM 0.83 1.03

Table 3: Expected and Estimated Effects

Expected Support 1 Support 2 Support 3

Haz Exp Haz Exp Haz Exp Haz Exp
VIEWLAST + ? + ns + – ns ns
VIEWLAG – ? – ns ns – – +
ADNUM ? ? - ns ns ns – +
FEATURE + – + ns ns - + ns
LEXPLAG – + + + ns ns ns +
WEDDAY + ? + ns + + + ns
SITENUM + ? + ns + – + ns
Haz: Hazard function
Exp: Expenditure
ns: Not significant at p=0.10
+: Positive and significant at p=0.10
–: Negative and significant at p=0.10

Table 4: Model Estimation Results

Support 1 Support 2 Support 3

Mean SE Mean SE Mean SE
Purchase Timing
Gamma_0 1.282 0.041 -0.183 0.078 0.528 0.024
Gamma_1 2.975 0.063 2.426 0.644 4.291 0.124
VIEWLAST 0.100 0.051 8.921 2.960 -0.083 0.049
VIEWLAG -0.884 0.079 0.736 0.694 -2.917 0.133
ADNUM -0.086 0.009 -0.136 0.188 -0.283 0.022
FEATURE 0.105 0.055 -0.610 0.682 0.562 0.088
LEXPLAG 0.040 0.017 0.245 0.156 0.002 0.033
WEDDAY 0.081 0.033 1.003 0.536 0.536 0.078
SITENUM 0.124 0.020 1.248 0.334 0.402 0.037


INTERCEPT 2.179 0.091 3.673 0.318 1.097 0.096

VIEWLAST -0.051 0.059 -7.920 1.134 0.023 0.037
VIEWLAG 0.032 0.109 -2.160 0.355 0.318 0.102
ADNUM 0.008 0.013 0.111 0.110 0.027 0.014
FEATURE -0.040 0.073 -1.052 0.308 -0.118 0.063
LEXPLAG 0.349 0.026 -0.083 0.070 0.610 0.028
WEDDAY -0.002 0.059 1.464 0.220 -0.068 0.051
SITENUM -0.029 0.029 -0.940 0.166 -0.006 0.028

Size 0.501 0.031 0.468

Note: Estimates in bold are significant at p=0.10.

Table 5: Elasticities of Covariates

Variable Support 1 Support 2 Support 3 Mean

Purchase Timing
VIEWLAST 0.010 0.019 -0.004 0.007
VIEWLAG -0.519 0.040 -0.619 -0.486
ADNUM -0.477 -0.054 -0.485 -0.425
FEATURE 0.318 -0.155 0.546 0.424
LEXPLAG 0.316 0.189 0.005 0.100
WEDDAY 0.239 0.287 0.534 0.446
SITENUM 0.397 0.285 0.364 0.365
VIEWLAST -0.001 -0.117 0.000 -0.004
VIEWLAG 0.002 -0.166 0.024 0.008
ADNUM 0.006 0.085 0.021 0.015
FEATURE -0.002 -0.045 -0.005 -0.005
LEXPLAG 0.353 -0.085 0.617 0.463
WEDDAY 0.000 0.113 -0.005 0.001
SITENUM -0.013 -0.423 -0.002 -0.021
Note: Elasticities in bold are based on significant (p=0.10) parameter estimates.

Figure 1: Distribution of Inter-Purchase Times





0 10 20 30 40 50 60 70 80 90 98

Figure 2: Distribution of Log-Expenditure




0 .5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6 6.5
Log expenditure

Figure 3: Time Sequence of Events for a Typical Consumer

Ad exposure Ad exposure
Ad exposure Ad exposure

Second purchase Third purchase End of data

First purchase
Exposure and purchase
sequence repeats

Time axis

Figure 4: Plots of Baseline Hazard Function for Each Support

Support 1








Support 2




Support 3