Content uploaded by Richard Briesch
Author content
All content in this area was uploaded by Richard Briesch on Jan 16, 2017
Content may be subject to copyright.
1630$.1)$3-(5,0*(4($3&+
Vol. XLVIII (June 2011), 457 –471
*Raj Sethuraman is Professor and Chair of Marketing and Marilyn and
Leo Corrigan Endowed Professor (e-mail: rsethura@ cox. smu. edu), and
Richard A. Briesch is Associate Professor of Marketing and Marilyn and
Leo Corrigan Endowed Professor (e-mail: rbriesch@ cox.smu.edu), Edwin
L. Cox School of Business, Southern Methodist University. Gerard J. Tellis
is Neely Chair of American Enterprise, Professor of Marketing, Manage-
ment, and Organization, and Director of Center for Global Innovation,
Marshall School of Business, University of Southern California (e-mail:
tellis@marshall. usc.edu). The authors thank Don Lehmann, Mike Hanssens,
the participants at the Advertising Generalization Conference at the Whar-
ton School at the University of Pennsylvania, and participants in the Mar-
keting Science conference in Ann Arbor. The authors also thank Ranga
Venkatesan and Prerit Souda for assistance in data collection and analysis.
This study benefited from a grant from Don Murray to the USC Marshall
Center for Global Innovation. Greg Allenby served as associate editor for
this article.
' ()*'#$'' )""(0<3 '''(
)740CB7=@A2=<3C2B0;4B00<0:GA8A=5 A7=@BB4@;0<3:=<6
B4@; 38@42BB=2=<AC;4@ 1@0<3 03D4@B8A8<6 4:0AB828B84A 4AB8;0B43 8<
ABC384A>C1:8A74314BE44<0<3)74ABC3G58<3AA4D4@0:<4E
4;>8@820:64<4@0:8H0B8=<A01=CB03D4@B8A8<64:0AB828BG)74;=AB8;>=@B0<B
0@40A5=::=EA)740D4@064A7=@BB4@;03D4@B8A8<64:0AB828BG8AE7827
8AAC1AB0<B80::G:=E4@B70<B74>@8=@;4B00<0:GB82;40<=5B74@470A
144< 0 342:8<4 8< B74 03D4@B8A8<6 4:0AB828BG =D4@ B8;4 0<3 03D4@B8A8<6
4:0AB828BG 8A 78674@ 5=@ 3C@01:4 6==3A B70< <=<3C@01:4 6==3A 8<
B7440@:GAB064B70<B74;0BC@4AB064=5B74:8542G2:45=@G40@:G30B0
B70<?C0@B4@:G30B00<3E74<03D4@B8A8<68A;40AC@438<6@=AA@0B8<6
>=8<BAB70<;=<4B0@GB4@;A)74;40<:=<6B4@;03D4@B8A8<64:0AB828BG8A
E7827 8A ;C27 :=E4@ B70< B74 8;>:843 ;40< 8< B74 >@8=@ ;4B0
0<0:GA8A #0<G =5B74@4AC:BA5=@A7=@BB4@; 4:0AB828BG 7=:35=@:=<6
B4@; 4:0AB828BG E8B7 A=;4 <=B01:4 4F24>B8=<A )74 0CB7=@A 38A2CAA B74
8;>:820B8=<A=5B74A458<38<6A
Keywords03D4@B8A8<64:0AB828BG;4B00<0:GA8A4;>8@820:64<4@0:8H0B8=<
>@=;=B8=<;0@94B8<6;8F
=E,4::=4A3D4@B8A8<6,=@9
4<4@0:8H0B8=<A5@=;#4B0<0:GA8A=5
@0<33D4@B8A8<6:0AB828B84A
© 2011, American Marketing Association
ISSN: 0022-2437 (print), 1547-7193 (electronic) 457
Advertising is one of the most important elements of the
marketing mix. Controversy rages over whether firms are
getting adequate returns on their advertising expenditures
(Aaker and Carman 1982; Tellis 2004). One key element in
this controversy is how effective advertising is in generat-
ing sales. The effectiveness of advertising is often measured
in terms of advertising elasticity, or the percentage increase
in sales or market share for a 1% increase in advertising.
Obtaining generalizable estimates of advertising elasticity
and identifying factors that influence advertising elasticity
can further the field’s understanding of the effectiveness of
advertising.
Assmus, Farley, and Lehmann (1984) provide the first
empirical generalizations on advertising elasticity. In par-
ticular, these authors perform a meta-analysis of 128 esti-
mates of advertising elasticity from 16 studies published
between 1962 and 1981 and provide useful generalizations
on advertising effect. More than 25 years have passed since
that publication. This period (1984–2008) has witnessed
significant changes on many fronts that may have an impact
on the measurement and effectiveness of advertising. First,
the marketing environment has changed as a result of
greater competition, globalization, the advent of the Inter-
net, and the ability of the consumer to opt out of television
commercials through devices such as TiVo. Second, the data
and methods for estimating advertising elasticity are
increasing in sophistication with the use of disaggregate
scanner data and the application of New Empirical Indus-
trial Organization econometric models. Therefore, it seems
prudent to update the empirical generalizations on advertis-
ing elasticity by including data from studies published since
1981.
This study conducts a meta-analysis of 751 short-term
brand-level direct-to-consumer advertising elasticities and
402 long-term advertising elasticities from 56 studies pub-
lished between 1960 and 2008. Our study disconfirms a few
of Assmus, Farley, and Lehmann’s (1984) findings, vali-
dates some of the previous findings, and uncovers several
new empirical generalizations and insights, which we sum-
marize in the conclusion of this article.
In this regard, our research is similar in spirit to other
follow-up meta-analytic studies in recent times. For exam-
ple, Bijmolt, Van Heerde, and Pieters (2005) update the
meta-analysis of price elasticity conducted previously by
Tellis (1988). Hu, Lodish, and Krieger (2007) provide a par-
tial update to Lodish et al.’s (1995) meta-analysis related to
television advertising experiments. Our study can also be
viewed as a meta-analytic complement to Vakratsas and
Ambler’s (1999) broad review of advertising literature.
They develop a taxonomy, review 250 studies, and provide
insights into how advertising works. Our study performs a
meta-analysis of econometric estimates of advertising elas-
ticity and provides insights into whether advertising works,
the magnitude of the effect, and the factors that influence
elasticity. In the process, the study adds to Hanssens’s
(2009) list of empirical generalizations about marketing’s
impact.
Our study complements Fischer and Albers’s (2010)
recent article. Both studies offer insights into the effect of
marketing mix on sales. However, whereas Fischer and
Albers provide an excellent analysis of the effect of market-
ing efforts (detailing, journal advertising, and consumer
advertising) on primary demand (category expansion) in
pharmaceutical products, our analysis focuses on the effect
of consumer advertising on selective demand (competitive
brand sales) across a wide range of consumer products,
including pharmaceuticals.
We organize the remainder of the article as follows:
Next, we describe the data. Then, we describe the meta-
analysis procedure and present the empirical findings. Fol-
lowing this, we discuss the results and their implications. In
the final section, we summarize the results in the form of
empirical generalizations and provide some limitations and
future research directions.
This section describes the compilation of the database
used in the meta-analysis. The data consist of observations
on advertising elasticity (dependent variable) and the poten-
tial influencing factors of advertising elasticity (indepen-
dent variables).
'7(35,4,0*.$45,&,5:
For this meta-analysis, we selected studies that provide
estimates of brand-level, short- or long-term consumer
advertising elasticity from econometric models using mar-
ket data. Thus, our meta-analysis excludes (1) category
advertising effects, (2) effects based on experimental or
other noneconometric designs, and (3) business-to-business
(B2B) advertising. The following paragraphs explain each
of these choices.
First, category-level advertising elasticity measures the
increase in category sales (primary demand) for a 1%
change in total category advertising. In general, these
effects are of interest to economists and public policy mak-
ers who investigate whether advertising expands category
demand in products such as milk, alcohol, and cigarettes
(e.g., Gallet 2007). Fischer and Albers (2010) provide a
recent comprehensive analysis of the primary demand
effects of marketing efforts in the pharmaceutical industry.
In contrast, our perspective is that of the brand managers,
who are interested in the extent to which advertising of their
brands affects their own brand’s sales (selective demand).
Second, following the scope of Assmus, Farley, and
Lehmann’s (1984) meta-analysis, we restrict our analysis to
econometric estimates. Lodish and others conduct numer-
ous television advertising experiments and meta-analyze
those results (e.g., Hu, Lodish, and Krieger 2007). How-
ever, their focus is mainly on whether advertising produces
a significant impact on sales in controlled experiments and
not in natural market scenarios, which is the purpose of our
study.
Third, consistent with Assmus, Farley, and Lehmann’s
(1984) study, we focus on consumer advertising only. Stud-
ies that provide advertising elasticity in the B2B context are
few and, in general, pertain to journal advertising to physi-
cians (B2B in the pharmaceutical industry). Fischer and
Albers (2010) do a thorough job of analyzing this industry.
We adopted the following procedure for compiling the
studies. We began our literature review with Assmus, Far-
ley, and Lehmann (1984) as the base. Then, we used the
Social Science Citation Index to identify 132 publications
that reference the 1984 meta-analysis. Next, we used key-
word searches (e.g., “advertising elasticity”, “advertising
response”, “sales response”) in online search engines such
as Google Scholar, ABI/INFORM, and Lexis Nexis to iden-
tify articles that discuss the subject area. We also reviewed
the reference lists in all of the previously mentioned studies.
We considered studies that provide econometric estimates
of advertising elasticity. Although most studies directly
report advertising elasticity, in some cases we had to compute
the elasticity according to available data or with inputs from
the studies’ authors. The process yielded 751 short-term
elasticities from 56 publications. We provide details of these
studies in Web Appendix A (http://www.marketingpower.
com/jmrjune11).
Advertising can affect sales both in the short (current
period) and long (current and future periods) run. We define
“long-term advertising elasticity” as the percentage change
in a brand’s current and future period sales for a 1% change
in the brand’s current advertising.1Some studies directly
provide estimates of long-term advertising elasticity. Others
provide estimates of short-term elasticity and carryover
coefficient (coefficient of the lagged dependent variable)
from the Koyck model. We calculate long-term elasticity as
[short-term advertising elasticity/(1 – carryover coefficient)]
(Clarke 1976). A few researchers measure advertising as ad
stock, defined as a weighted combination of current and
past advertising based on an exponential smoothing coeffi-
cient. In this case, the estimate of advertising elasticity is
%*'$"%#'!)$'(' *$
1The combined advertising effect of current and future period is a lso
referred to as cumulative advertising effect.
=E,4::=4A3D4@B8A8<6,=@9
the long-term elasticity. The short-term elasticity is [long-
term elasticity ¥(1 – smoothing coefficient)] (Danaher,
Bonfrer, and Dhar 2008). We obtained 402 long-term adver-
tising elasticities from the 38 studies listed in Web Appen-
dix A (http://www.marketingpower.com/jmrjune11).
0).6(0&,0*$&5134
In addition to advertising elasticity, we collected data on
22 variables that could potentially influence elasticity and
classified them into the following six factors:
• ,/($0'3(&(44,10)$&5134: Median year of data and duration
of recession during the estimation period;
•31'6&5$0' *(1*3$2+,& )$&5134: Product type, product life
cycle, and geographic area;
•$5$ &+$3$&5(3,45,&4: Temporal interval, data aggregation,
dependent measure, advertising measure, and advertising type;
•/,55(' 7$3,$%.(4: Omission of lag-dependent variable, lag
advertising, lag price, price, quality, promotion, and distribution;
•1'(.&+$3$&5(3,45,&4: Functional form, estimation method,
incorporating endogeneity, and incorporating heterogeneity;
and
•5+(3&+$3$&5(3,45,&4: Published versus unpublished work.
Many of these variables correspond to those in Assmus, Far-
ley, and Lehmann’s (1984) original meta-analysis. How-
ever, availability of new data permits us to investigate sev-
eral new variables, such as the time trend, presence of a
recession, additional product types (service goods and phar-
maceuticals), additional continents (Asia and Australia),
and additional method factors (incorporation of endogene-
ity and heterogeneity). Table 1 provides the levels of the
independent variables and a detailed description of the
expected relationship and the operationalizations of the
variables.
!
!0,7$3,$5(0$.:4,4
First, we performed a univariate analysis to obtain an
estimate of the mean short-term advertising elasticity, which
we then compared with the corresponding value in Assmus,
Farley, and Lehmann (1984). We also analyzed the median
and distribution of advertising elasticity.
(5$0$.:5,&1'(.)13+135 (3/.$45,&,5:
The purpose of this meta-analysis is to identify the poten-
tial influencing factors of advertising elasticity. Prior meta-
analytic studies (e.g., Assmus, Farley, and Lehmann 1984;
Tellis 1988) have used ordinary least squares (OLS) regres-
sion of the following form:
(1) Asj = Xsjb+ esj,
where Asj is the jth advertising elasticity from sth study; Xsj
are characteristics of the market, study design, and model
that influence the elasticity (listed in Table 2, Column 3); bs
are the meta-analytic parameters of interest, and esj are the
error terms, initially assumed to be independently and iden-
tically distributed N(0, s2). To account for within-study
error correlations, Bijmolt and Pieters (2001) point out that
unique study-specific characteristics not captured in the
independent variables would appear in the error structure.
This would result in nonzero correlations leading to nonzero
error covariance (within-study), violating the assumptions of
OLS regression. Therefore, they suggest using a hierarchi-
cal linear model estimated with iterative generalized least
squares to allow for a within-study, block nonzero variance–
covariance matrix. Specifically, they suggest a model of the
following form:
(2) Asj = Xsjb+ zs+ esj,
where zsis the unobserved study-specific effect and is
assumed to be distributed with mean zero and standard
deviation sz. Following Bijmolt, Van Heerde, and Pieters
(2005), we use Model 2 in our analysis for identifying
potential influencing factors.
.5(30$5((5$0$.:5,&1'(.451 (451%6450(44
There are several issues regarding the estimation of
Model 2. First, the advertising elasticities themselves are
not true parameters but are estimated with error. We
accounted for this uncertainty surrounding the true advertis-
ing elasticity using the following procedure: (1) We com-
piled the estimated advertising elasticity Asj, the jth adver-
tising elasticity from sth study along with its standard error
Bsj; (2) we assumed that the true parameters lie in the nor-
mal distribution N(Asj, Bsj); (3) we drew A˜
sjobservation for
each s, j from the corresponding normal distribution; (4) we
estimated Model 2 with A˜
sj instead of Asj to get bk, where
bkis the vector of regression parameters from the kth itera-
tion. Then, we repeated Steps 1–4 for 500 iterations and
averaged bkacross all 500 iterations to get an average esti-
mate that takes into account the uncertainty surrounding the
advertising elasticity.2
A second major issue is collinearity. Several method
variables are likely to be correlated. We assessed the extent
of multicollinearity through traditional measures such as
bivariate correlation, variance inflation factor, condition
index, and proportion of variance explained (Belsley, Kuh,
and Welsch 1980). However, because most of the method
factors are discrete dummy variables, correlation measures
often tend to be lower. Therefore, we used cross-tab analy-
sis of pairs of variables and the corresponding chi-square
measure to detect deviation from independence of two dis-
crete variables. Using these multiple measures, we identi-
fied pairs of variables with potential problems of collinear-
ity. We excluded one of those variables from Model 2 and
tested for robustness by inspecting the significance of the
other included variable.
(451)'',5,10$."$3,$%.(4
We carried out several analyses to gain additional
insights. First, we explored the following four interaction
effects3:
•(&(44,10¥231'6&55:2(: A recession may affect advertising
elasticity of high-priced durable goods more than low-priced
food products;
2We thank the associate editor for suggesting this method to account for
uncertainty in elasticity estimates.
3Numerous other interaction effects are possible, but incorporating all
possible interactions would contribute to collinearity and compromise the
stability of the model. We included those interactions either for which we
had some prior knowledge based on theory or intuition (e.g., product life
cycle ¥dependent measure) or that were otherwise of managerial interest
(e.g., recession ¥product type).
%*'$"%#'!)$'(' *$
6/%(3 "$3,$%.((7(. 92(&5('(.$5,104+,28,5+'7(35,4,0*.$45,&,5: 2(3$5,10$.,;$5,1018$5$%5$,0('
,/( 3(0'$0'(&(44,10
1 Time trend:
Median year of
data
Advertising elasticity should decrease over time because of
increased competition, ad clutter, the advent of the Internet as an
alternate information source, and the consumer’s ability to opt
out of television commercials through devices such as TiVo.
We used median year of the estimation period to detect time
trend. For example, if the study used data from 1982 to 1990
to estimate advertising elasticity, that advertising elasticity
observation is said to come from its median year 1986.
2 Recession:
Months of
recession
During recessionary times, consumers should become more
price conscious and tend to ignore generally image-based
advertising. Therefore, advertising elasticity should be smaller
during recessionary times.
Recession is defined as two quarters of negative gross
domestic product growth. Data for the United States and other
countries were obtained from the National Bureau of Economic
Research and the Organisation for Economic Co-operation and
Development websites. When recession data are not identified,
we substituted U.S. data. We measured recession variable as
the number of months that the economy is in recession as a
proportion of total months in the estimation period.
31'6&5$0'(1*3$2+,&$&5134
3 Product type:
•Pharmaceutical
•Durable
•Food
•Nonfood
•Service
No prior expectations. We obtained category type information directly from the
individual studies.
4 Product life cycle:
•Growth
•Mature
Advertising either provides information or persuades
consumers about the advertised brand, both of which are more
relevant in the early stage of the life cycle, when consumers
know little about the brand or have not formed preferences.
Thus, advertising elasticity will be higher for products in the
early stage than in the mature stage of the life cycle.
We obtained product life-cycle information directly from the
individual studies. When the authors called the studied
product “established,” we classified it as a mature product.
5 Region:
•America
•Europe
•Other
Europe may have underadvertising due to regulation, short
history of advertising, or culture, while the United States may
have optimal or overadvertising because of the long history of
advertising, intense competition, and advertising wars. If this
reasoning is correct, advertising elasticity is likely to be
higher in Europe than in the United States.
Region information is based on the continent in which the data
for estimation of advertising elasticity are obtained. Most data
for the North American continent are from the United States.
Many studies that estimate data from Europe state the specific
country (e.g., France, Germany). However, we did not analyze
at the country level, because there are inadequate country-
level observations to obtain credible regression estimates.
$5$+$3$&5(3,45,&4
6 Dependent
measure:
•Absolute (sales)
•Relative (share)
Absolute sales captures both competitive gains and gains due
to primary market expansion. Relative sales captures only
competitive gains. Because advertising can increase primary
demand, we expect advertising elasticity to be higher when
sales are recorded in absolute terms.
We obtained data on dependent measure directly from the
estimation model that produced the elasticity. If the model
has unit or dollar sales as dependent variable, we considered
it absolute. If the dependent variable is market share, we
classified it as relative.
7 Temporal interval:
•Weekly
•Quarterly
•Yearly
In general, advertising has a carryover impact. Therefore, the
greater the level of temporal aggregation, the more likely it is
to capture the sales resulting from the carryover advertising.
Thus, we expect current-period advertising elasticity to be
larger when data are more aggregated (yearly or quarterly)
than less aggregated (weekly or daily) if the model does not
capture the carryover effect.
We obtained data interval information directly from the
individual studies. We combined some levels because of
paucity of data:
•Weekly: daily, weekly, monthly
•Quarterly: bimonthly, quarterly
•Yearly: annual
8 Data aggregation:
•Firm
•Panel
Estimation of advertising elasticity using firm-level data
implicitly aggregates consumer-level information in a linear
way. Estimation of advertising elasticity with panel data uses
MLE that combines information in a nonlinear way.
Therefore, the results may be different, but the direction of
change is unknown.
Most of the data are aggregated across consumers at the
brand level. These aggregated data are called firm-level data.
The panel data type consists of individual consumer–level
choice data with corresponding individual advertising data
(exposures).
9 Advertising
measure:
•Monetary
•GRP
•Relative
Depending on what competitors in the market are doing,
changes in absolute advertising may not necessarily reflect
the same changes in relative advertising. If competitors tend
to match a target firm’s advertising, large changes in absolute
advertising will translate into small changes in relative
advertising. As a result, elasticity estimated with absolute
advertising should be smaller than those estimated with
relative advertising. The reverse should hold if competitors do
not match the advertising.
A brand’s advertising may be measured in absolute
(monetary values or GRP) or relative (the brand’s share of
advertising in the market) terms. We obtained data on
advertising measure directly from the individual studies.
When the researchers used absolute advertising, information
is provided as to whether advertising is measured in
monetary units or GRP.
10 Advertising type:
•Print
•Television
•Aggregate
It is not clear whether consumers are more responsive to print
or television advertising. However, because aggregate
advertising is a combination of print and television
advertisements, we expect advertising elasticity from aggregate
advertising to be between print and television advertising.
We obtained these data directly from the type of advertising
described in individual studies. Advertising is aggregate if
the advertising is a combination of more than one type of
advertising (e.g., print, television, billboards). When nothing
is mentioned, we classified it as aggregate advertising.
)01:4
)%'($"*$$+')($"())(-&)'")%$(&($%&')%$"/)%$(
=E,4::=4A3D4@B8A8<6,=@9
•31'6&5.,)(&:&.(¥231'6&55:2(: Advertising might be more
important in the early stage of durable products because they
are less easily known by trial;
•31'6&5.,)(&:&.(¥'(2(0'(05/($463(: For growth products,
advertising may have greater influence on sales than share
because of the higher potential for category sales growth,
whereas for mature products advertising may influence market
share more than absolute sales because firms are competing for
a share of fixed total market; and
•$5$,05(37$.¥1/,44,10 1).$**('4$.(4: When data are more
aggregate (yearly), omitting lagged sales (carryover effect)
may not affect advertising elasticity as much as when data are
less aggregate (weekly).
Second, we investigate whether some brand characteris-
tics influence advertising elasticity. We collected data on
brand market share, brand advertising share, and relative
price for 212 of 751 observations. However, Sethuraman,
Srinivasan, and Kim (1999) state that the elasticity measure
is related to brand share and advertising share by definition.
For example, in the linear model, (absolute) advertising
effect is measured as d(market share)/d(advertising). To
6/%(3 "$3,$%.((7(. 92(&5('(.$5,104+,28,5+'7(35,4,0*.$45,&,5: 2(3$5,10$.,;$5,1018$5$%5$,0('
/,55('"$3,$%.(4
11 Lag dependent
variable:
•Omitted
•Included
Lagged sales are likely to be correlated positively with
current-period sales and current-period advertising (because
current advertising is often set as a proportion of past sales).
Therefore, we expect omission of lagged sales to bias
advertising elasticity positively.
We obtained these data directly from the model from which
the advertising elasticity is estimated.
12 Lag advertising:
•Omitted
•Included
Lagged advertising is likely to be correlated positively with
current-period sales and current-period advertising. Therefore,
we expect omission of lagged advertising to bias the
advertising elasticity measure positively.
We obtained these data directly from the model from which
the advertising elasticity is estimated.
13 Lag price:
•Omitted
•Included
Lagged price is likely to be correlated negatively with
current-period sales and positively with current-period
advertising. Therefore, omission of lagged price should
negatively bias the advertising elasticity.
See preceding item.
14 Price:
•Omitted
•Included
Price is likely to be correlated negatively with current-period
sales and positively with current-period advertising.
Therefore, we expect omission of price to bias the advertising
elasticity measure negatively.
See preceding item.
15 Quality:
•Omitted
•Included
We were unable to predict the sign of correlation between
quality and sales and, thus, the direction of the effect.
See preceding item.
16 Promotion:
•Omitted
•Included
We were unable to predict the sign of correlation between
promotion and advertising.
See preceding item.
17 Distribution:
•Omitted
•Included
Distribution is likely to be correlated positively with current-
period sales and positively with current-period advertising.
Therefore, we expect omission of distribution to bias the
advertising elasticity measure positively.
See preceding item.
1'(.+$3$&5(3,45,&4
18 Functional form:
•Double log
•Linear, Share
•Other
No prior expectations. We obtained these data directly from the model from which
the advertising elasticity is estimated.
19 Estimation method:
•OLS
•GLS
•MLE
•Other
No prior expectations. We obtained these data directly from the model from which
the advertising elasticity is estimated.
20 Endogeneity:
•Omitted
•Included
We had no theoretical reasoning, but Villas-Boas and Winer
(1999) find that omitting endogeneity in price elasticity
estimation biases the estimate toward zero.
A model incorporates endogeneity if it treats advertising as a
dependent variable endogenously determined within the
model structure.
21 Heterogeneity:
•Omitted
•Included
No prior expectations. A model that allows for differences in advertising response
parameters across households or segments in the sample is
deemed to incorporate heterogeneity. Both discrete and
continuous heterogeneity are included in this measure.
5+(3$&5134
22 Study type:
•Published
•Unpublished
Because of the general bias toward publishing articles that
produce significant effects, advertising elasticity in published
articles should be higher than advertising elasticity from
unpublished works.
We obtained these data directly from the studies.
)01:4
%$)$*
Notes: OLS = ordinary least squares, GLS = generalized least squares, and MLE = maximum likelihood estimation.
%*'$"%#'!)$'(' *$
)01:4
)%'($"*$$+')($"())(#%"''((%$%$)(()$'''%'(
+135 (3/'7(35,4,0*.$45,&,5: 10* (3/'7(35,4,0*.$45,&,5:
6/%(3 "$3,$%.( (7(. 92(&5(',*0 $,0))(&50.: $,005(3$&5,10))(&5 $,0))(&50.: $,005(3$&5,10))(&5
0 Intercept All obs. .07 (.12) –.08 (.17) .06 (.2) –.26 (.21)
,/( 3(0'$0'(&(44,10
1 Time Trend Year of data – –.004 (.002)** –.004 (.001)*** –.005 (.002)** –.007 (.003)***
2 Recession Months of recession – .01 (.06) .02 (.06) .42 (.10)*** .34 (.09)***
31'6&5$0'(1*3$2+,&$&5134
3 Product type Drug ? .19 (.09)** .16 (.14) .14 (.11) .29 (.13)***
Durable ? .29 (.09)*** .26 (.06)*** .27 (.08)*** .48 (.1)***
Food ? .03 (.03) –.15 (.06)** .08 (.05) –.13 (.1)
Service ? .10 (.07) –.04 (.11) –.10 (.08) –.08 (.08)
Nonfood Base 0 0 0 0
4 Product life cycle Mature – –.08 (.05)** –.08 (.06)* .08 (.06) .15 (.12)
Growth Base 0 0 0 0
5 Region (continent) Europe + .09 (.05)** .16 (.06)*** .16 (.05)*** .34 (.08)***
Other ? .05 (.05) .06 (.05) .09 (.06) .08 (.05)
North America Base 0 0 0 0
$5$+$3$&5(3,45,&4
6 Dependent measure Absolute + –.03 (.02) .08 (.07) –.12 (.04)*** .20 (.10)**
Relative Base 0 0 0 0
7 Temporal Interval Weekly – .04 (.03) .05 (.03) .05 (.06) .04 (.07)
Yearly + .08 (.04)** .09 (.03)*** .10 (.11) –.03 (.18)
Quarterly Base 0 0 0 0
8 Data aggregation Firm ? –.20 (.05)*** –.24 (.06)*** .07 (.08) –.03 (.06)
Panel Base 0 0 0 0
9 Advertising measure Relative ? .07 (.03)** .07 (.03)*** –.11 (.07)* –.10 (.07)
GRP ? .16 (.08)** .16 (.09)** .16 (.06)*** .30 (.09)***
Monetary Base 0 0 0 0
10 Advertising type Television ? .19 (.09)** .21 (.09)** –.17 (.05)*** –.18 (.05)***
Aggregate ? .11 (.06)* .14 (.05)** –.03 (.05) –.01 (.05)
Print Base 0 0 0 0
/,55('"$3,$%.(4
11 Lag dependent variable Omitted + .10 (.07)* .09 (.06)* — —
Included Base 0 0 — —
12 Lag advertising Omitted + .002 (.03) .01 (.03) .14 (.04)*** .13 (.03)***
Included Base 0 0 0 0
13 Lag price Omitted – .01 (.05) .03 (.05) .04 (.08) .17 (.12)
Included Base 0 0 0 0
14 Price Omitted – –.01 (.03) –.01 (.03) .03 (.05) .02 (.07)
Included Base 0 0 0 —
15 Quality Omitted ? –.02 (.07) –.04 (.06) –.15 (.08)* –.11 (.07)
Included Base 0 0 0 0
16 Promotion Omitted ? –.03 (.08) –.01 (.08) .13 (.08)* .14 (.07)**
Included Base 0 0 0 0
17 Distribution Omitted + .11 (.04)*** .11 (.04)*** .10 (.05)** .11 (.06)**
Included Base 0 0 0 0
=E,4::=4A3D4@B8A8<6,=@9
)01:4
%$)$*
+135 (3/'7(35,4,0*.$45,&,5: 10* (3/'7(35,4,0*.$45,&,5:
6/%(3 "$3,$%.( (7(. 92(&5(',*0 $,0))(&50.: $,005(3$&5,10))(&5 $,0))(&50.: $,005(3$&5,10 ))(&5
1'(.+$3$&5(3,45,&4
18 Functional form Double log ? .14 (.07)** .14 (.08)* .16 (.07)*** –.07 (.09)
Linear ? .26 (.10)*** .24 (.09)*** .07 (.10) .12 (.10)
other ? .23 (.09)*** .22 (.08)*** –.28 (.09)*** –.13 (.10)
share Base 0 0 0 0
19 Estimation method GLS ? –.02 (.03) –.04 (.03) –.17 (.05)*** –.18 (.04)***
MLE ? .05 (.05) .02 (.05) .12 (.08) .18 (.09)**
OLS ? .004 (.02) –.01 (.02) –.10 (.06)* –.12 (.07)*
Other Base 0 0 0 0
20 Endogeneity Omitted – –.16 (.08)** –.16 (.08)** –.01 (.07) .04 (.07)
Included Base 0 0 0 0
21 Heterogeneity Omitted ? –.06 (.06) –.04 (.07) .07 (.09) .09 (.10)
Included Base 0 0 0 0
5+(3+$3$&5(3,45,&4
22 Study type Published – .09 (.10) .07 (.12) –.02 (.17) –.16 (.16)
Working Base 0 0 0 0
05(3$&5,10))(&54
1 Mature life cycle ¥durable ? — –.34 (.13)*** — –.39 (.11)***
2 Mature life cycle ¥food ? — –.10 (.18) — .21 (.14)
3 Mature life cycle ¥nonfood ? — .12 (.12) — —
4 Life cycle ¥absolute sales – — –.13 (.09)* — –.40 (.14)***
*2< .10.
**2< .05.
***2< .01.
Notes: OLS = ordinary least squares, GLS = generalized least squares, and MLE = maximum likelihood estimation. We used a one-tailed test if expected sign is unambiguous ( + or –) and a two-tailed test if
expected sign is ambiguous (?). Expected sign: + = positive relationship (compared with base level); – = negative relationship; ? = ambiguous relationship; and — = coefficient not included or not estimable. We
rounded all coefficient estimates to two decimals.
convert into (percent) elasticity measure, we multiplied the
effect by mean brand advertising and divided by brand mar-
ket share. Because market share is in the denominator and
advertising is in the numerator, by definition, advertising
elasticity tends to be larger for brands with low market share
and high advertising share. Therefore, we only test whether
advertising elasticity is higher for high-priced brands by
reestimating Model 2 with brand relative price included.
Third, we estimate the following logit model to identify
the conditions when the advertising elasticity is signifi-
cantly greater than zero (2< .05, one-tailed test):
where Vsj= Xsjg+ esj, where Xsjis the set of influencing
variables listed in Table 2, and gis the coefficient vector
measuring the influence of the variable on the statistical sig-
nificance of advertising elasticity.
0$.:4,41)10* (3/'7(35,4,0*.$45,&,5:
As with short-term elasticity, we first performed univari-
ate analysis to obtain an estimate of the mean advertising
elasticity and compared it with Assmus, Farley, and
Lehmann’s (1984) value. Next, we obtained insights into
the median and distribution of long-term advertising elastic-
ity. Then, we estimated Model 2 to identify factors that
influence long-term elasticity. We did not conduct robust-
ness checks, because we did not have sufficient data and
because standard errors are not available for long-term elas-
ticity.
!
!0,7$3,$5(0$.:4,4
Figure 1 presents the distribution of short-term advertis-
ing elasticity. There are 751short-term brand-level advertis-
ing elasticities with magnitudes ranging from –.35 to 1.80.
More than 40% of the elasticities are between 0 and .05.
Approximately 7% of advertising elasticities are negative,
() ( ) exp( )
exp( ) ,3 0 1
PA issignificantly V
V
sj
sj
sj
>=
+
though in general, we expected advertising elasticity to be
positive. In the spirit of meta-analysis, we retained the nega-
tive elasticities because the meta-analytic model reveals
whether any method or environmental variable is responsi-
ble for such negative estimates.
The mean short-term advertising elasticity across the 751
observations is .12, which is substantially lower than Ass-
mus, Farley, and Lehmann’s (1984) mean of .22 from 128
observations. We attribute the difference to (1) reduction in
advertising elasticities over time; (2) Assmus, Farley, and
Lehmann’s inclusion of 32 product-level elasticities, which
are in general higher than the brand-level elasticities; and
(3) Assmus, Farley, and Lehmann’s omission of Lambin’s
(1976) estimates, which are, in general, below the previ-
ously stated mean of .22. Our estimate is closer to the mean
advertising elasticity of .104 that Hu, Lodish, and Krieger
(2007, Table 1) report from 210 real-world television adver-
tising tests. Our estimate is also similar to that of Sethura-
man and Tellis (1991), who find that the mean brand-level
advertising elasticity across a wide range of categories is
.11.
The median short-term advertising elasticity is even
lower at .05, but it is closer to the uncorrected mean short-
term elasticity of .04 in Fischer and Albers’s (2010, Table
W1) recent comprehensive study on pharmaceuticals. The
standard errors (or t-values) are reported for 437 of the 751
observations. Advertising elasticities are significantly
greater than zero at the 95% confidence level in 57% of the
cases.
7(37,(81)(5$0$.:5,&(46.54
Table 2 (Column 5) presents the results for the main
meta-analytic Model 2 for short-term advertising elasticity.
The model explains 37% of the variance in advertising elas-
ticity, which is comparable to Assmus, Farley, and Lehmann’s
meta-analysis (36%). Coefficients corresponding to 12 of
the 22 independent variables are statistically significant at
least at 2< .10: year of data, product type, product life
cycle, region, temporal interval, level of data aggregation,
%*'$"%#'!)$'(' *$
86C@4
()'*)%$%(%'))'#$$"%$)'#$+')($"())(
Percent of Observations
Advertising Elasticity
I I I I I I
(7=@BB4@;
"=<6B4@;
=E,4::=4A3D4@B8A8<6,=@9
measure of advertising, advertising type, omission of lag
sales, omission of distribution, functional form, and omis-
sion of endogeneity.
We accounted for uncertainty in advertising elasticity by
compiling the 437 observations in which information on
uncertainty (standard error) is available. We drew 500 ran-
dom data sets using the method described in the “Proce-
dure” section, estimated Model 2, and computed the aver-
age of the estimates for each variable (reported in Web
Appendix B at http://www.marketingpower.com/jmrjune11).
All significant variables in the main regression model
except advertising measure are also significant in this new
model. In addition, recession is positive and significant.
We assessed the extent of collinearity by inspecting the
variance inflation factors, condition indexes, bivariate cor-
relations, and cross-tabulations. All but one of the variance
inflation factors are lower than 5, and all condition indexes
are less than 20. Some correlations are high (> 5); therefore,
we deleted one variable at a time and inspected the robust-
ness of other results. We also estimated stepwise regression.
Almost all the original results are robust, except in two
cases: The variable recession is positive and significant in a
few alternate models, and omission of endogeneity is not
statistically significant in a few models. (Details of these
results are available on request.)
We included each interaction effect mentioned in the pro-
cedure section one at a time, and we retained them if they
were significant. Two effects—product life cycle ¥product
type and product life cycle ¥dependent variable—are sta-
tistically significant. The percentage of explained variance
increases from 37% in the model with only main effects to
40% in the model incorporating interaction effects. Table 2
presents the regression results (Column 6).
To ascertain whether a brand’s relative price affects
advertising elasticity, we used 212 observations for which
information on price was available and estimated Model 2.
The coefficient of relative price is positive (.03, SE = .05)
but not statistically significant.
Which factors influence the statistical significance of
advertising elasticity? There are 437 observations for which
information on statistical significance (t-values) was avail-
able. In 249 of the observations (57%), advertising elasticity
is significantly greater than zero. Web Appendix B (http://
www.marketingpower.com/jmrjune11) presents results of
the Logit Model 3. Because of a lack of data, we did not
include interaction effects. The results reveal fewer signifi-
cant coefficients than in the original model. Advertising
elasticity is more likely to be significantly greater than 0 in
Europe than the United States, for panel data than aggregate
firm data, and for television than print advertising. Some
coefficients are marginally significant. A possible explana-
tion for the lack of significance in this model is as follows:
A significant advertising elasticity, whether .1 or .5, is taken
as 1. A nonsignificant coefficient, whether .001 or .05, is
deemed to be 0. This recoding absorbs meaningful variation
in the data, which can result in fewer significant estimates
than in the original model, with magnitude of advertising
elasticity as the dependent variable. In the next section, we
present the results on short-term advertising elasticity for
significant variables.
,/( 3(0'$0'(&(44,10
(',$0:($31)'$5$. Given the increased competition in
consumer products, the advent of the Internet as an alternate
information source, and the ability of consumers to opt out
of television commercials, we would expect consumers to be
less responsive to advertising in more recent times than in
the past. Consistent with our expectations, the corresponding
regression coefficient of time (median year of data) is nega-
tive. This result is robust across models. (We also tested a
quadratic effect of time; the coefficient was nonsignificant.)
Assmus, Farley, and Lehmann’s (1984) meta-analysis
uses pre-1980 data, whereas ours includes post-1980 data.
Therefore, we compared advertising elasticity for pre-1980
data (1940–1979) with that for post-1980 data (1980–2004).
Instead of treating year of data as a continuous variable, we
included a dummy variable to indicate pre- and post-1980
periods. The regression coefficient for after 1980 is –.11 (SE
= .06), indicating a significant decline between the two time
periods. The mean advertising elasticity is .13 before 1980
(n = 463) and .10 after 1980 (n = 288).
Temporal differences in advertising elasticity can occur
because of differences in consumer response to advertising
over time or differences in market characteristics and
research methods. To explore the impact of market/method
factors on the temporal differences in predicted advertising
elasticity, we followed Bijmolt, Van Heerde, and Pieters’s
(2005, p. 151) approach and computed the contribution of
various factors to the difference. Table 3 presents the five
key contributing factors that influence difference in the
advertising elasticity between pre-and post-1980 periods.4
)01:4
!.%$)'*)%'()%$($(%'))'#+')($"()).&'$&%()
1'(.$,0))(&5 (3&(05$*(,03162 (3&(05$*(,03162 1053,%65,1051
"$3,$%.( (7(. 1()),&,(05 )5(3 ()13( '7(35,4,0*.$45,&,5:
Aggregation Firm –.195 46.5 93.3 .09
Ad type Television .193 59.7 21.4 .07
Region Europe .086 1.1 48.4 –.04
Data interval Yearly .083 3.8 36.3 –.03
Life cycle Mature –.083 87.9 57.5 –.03
Notes: Sample interpretation and illustration for aggregation: The regression coefficient from main effects Model 2 for data aggregation (firm vs. panel) is
–.195; that is, advertising elasticity estimated from firm data is .195 lower than that from panel data. Panel data are rarely used, and firm-level data are pre-
dominantly used before 1980: About 93% of short-term advertising elasticity observations before 1980 are from firm data, whereas 47% of the observations
after 1980 use firm data. This difference in data representation results in an increase in mean predicted advertising elasticity of .09 in the post-1980 period
compared with the pre-1980 period (computed as –.195 ¥[46.5 – 93.3]/100 = .09, rounded to two decimals).
4We thank one of the anonymous reviewers for suggesting this analysis.
Firm-level aggregate data constitute the primary database
before 1980 (93% use); in contrast, because of the advent of
scanner data, post-1980 databases use panel data and firm
data is substantially reduced (47% use). This difference in
database results in a .09 increase in advertising elasticity
after 1980 compared with before 1980. Television advertis-
ing is used more in the estimation after 1980 (60%) than
before 1980 (21%), resulting in an increase in advertising
elasticity of .07 after 1980. Europe is grossly underrepre-
sented after 1980 (1%) compared with before 1980 (48%).
Because advertising elasticity in Europe is higher than in
the United States, this difference causes a reduction in post-
1980 predicted advertising elasticity by .04.
Early researchers tend to use more temporally aggregate
(yearly) data, while later researchers use less yearly and
more weekly data. This difference causes advertising elas-
ticity to decrease by .03 after 1980 compared with before
1980. Because markets in general have matured over time,
approximately 88% of products studied after 1980 are
mature products, compared with 58% before 1980. Because
mature products have lower advertising elasticity, there is a
.03 reduction in advertising elasticity after 1980.
In summary, changes in estimates of advertising elastic-
ity over time can be attributed to changes in market and
method characteristics. The observed negative regression
coefficient for year of data suggests that the effect persists
even after accounting for these factors. Why might advertis-
ing response be lower in recent times? Researchers point to
advertising clutter and competitive advertising, both of
which result in reduced recall and evaluation of the brand
being advertised. For example, Kent (1995) documents that
the average number of network advertisements per hour
tripled from 6 in the 1960s to 18 in the 1990s. Danaher,
Bonfrer, and Dhar (2008) report that advertising elasticity
declines in the presence of high competitive clutter.
(&(44,10. The current economic environment highlights
the need to understand how marketing strategies should be
modified in the face of recession. In particular, should
advertising budgets be curtailed or increased during reces-
sion (for a recent survey of relevant literature, see Tellis and
Tellis 2009)? If advertising elasticity is lower during reces-
sion, both the impact factor and budgetary considerations
would suggest a reduction in advertising. However, if reces-
sion advertising elasticity is equal to or higher than expan-
sion advertising elasticity, the decision is not clear. Our
meta-analysis reveals that advertising elasticity is not lower
during recessionary times. On the contrary, advertising elas-
ticity is higher during recession, though not significant.
31'6&5$0'(1*3$2+,&$&5134
31'6&55:2(. We tested for difference in advertising elas-
ticity among many types of product categories—pharma-
ceutical, food, nonfood, durable, and service goods (e.g.,
banks, movies). However, we did not have prior expecta-
tions for the relative magnitudes of these effects. Regression
coefficients (Table 2) and comparison of means reveal that
durable goods have the highest advertising elasticity, fol-
lowed by pharmaceuticals and service goods. Frequently
purchased food and nonfood products have the lowest
advertising elasticity. In general, nonfood, nondurable prod-
ucts tend to be low-involvement items, such as household
cleaners, whose purchase behavior advertising might not
significantly influence.
31'6&5.,)(&:&.(. Advertising either provides informa-
tion or persuades consumers about the advertised brand. In
either case, it is likely to be more relevant and useful in the
early stage of the product life cycle, when consumers know
little about the brand or have not formed preferences. Thus,
advertising elasticity will be higher for products in the early
stage of the life cycle than in the mature stage. Consistent
with this expectation, products in the growth stage of the
life cycle have a higher advertising elasticity (.16) than
products in the mature stage (.11), and the coefficient is sig-
nificant in the regression model (Table 2).
We also find an interaction between product life cycle
and product type, as Figure 2 shows. Declining advertising
elasticity from the growth stage to the mature stage of the
life cycle seems to be more prominent in durable goods,
moderate in food products, and nonsignificant in non-
durable, nonfood products.
(1*3$2+,&3(*,10. Europe may underadvertise because
of regulation, the short history of advertising in the region,
or culture. The United States may have optimal or over-
advertising because of the long history of advertising,
intense competition, and advertising clutter. If this reason-
ing is correct, advertising elasticity is likely to be higher in
Europe than the United States. This is indeed the case: We
find that Europe has a significantly higher mean advertising
elasticity (.17) than the United States (.11), and this effect
holds in the regression model after accounting for other fac-
tors (Table 2).
$5$+$3$&5(3,45,&4
(2(0'(05 /($463(. Sales can be measured in either
absolute (unit sales or dollar revenues or purchases) or rela-
tive (market share) terms. Absolute sales capture both
competitive gains and gains due to primary market expan-
sion. Relative sales capture only competitive gains. Because
advertising can increase primary demand, we expect adver-
%*'$"%#'!)$'(' *$
86C@4
#$(%'))'#+')($"())..&'%*)
"."$&'%*)).&
Predicted Mean and Elasticity
Product Life Cycle
@=EB7 #0BC@4
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
C@01:4
==3
$=<5==3
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
Notes: Pharmaceutical and service goods not reported due to small sam-
ple sizes (< 10).
=E,4::=4A3D4@B8A8<6,=@9
tising elasticity to be higher when sales are recorded in
absolute terms. Advertising elasticity is slightly higher for
(absolute) sales elasticity (.13) than for relative (share) elas-
ticity (.11), but the difference is not significant in the regres-
sion model.
However, we find a marginally significant (2< .10) inter-
action effect between product life cycle and dependent
measure in the regression model (Table 2). Figure 3 presents
the predicted means. In growth products, advertising elas-
ticity is higher when measured with sales as the dependent
variable than when share is the dependent variable. This
result is intuitive because the potential for an increase in pri-
mary demand (which is better captured in the sales model)
is higher in the growth stage of the life cycle.
(/213$.,05(37$.. In general, advertising has not only an
instantaneous impact on sales but also a carryover impact.
Therefore, the greater the level of temporal aggregation, the
more likely it captures the sales resulting from the carryover
advertising. Moreover, the greater the level of aggregation,
the greater is the bias caused by wrongly capturing the
carryover effect. Thus, we expect current period advertising
elasticity to be greater when data are more aggregate (yearly
or quarterly) than less aggregate (weekly or daily) if the
model does not fully and correctly capture the carryover
effect. Researchers typically estimate the carryover effect
with the Koyck model. However, using aggregate data leads
to a positive bias in this model’s estimation of carryover
effect (Tellis and Franses 2006).
It is noteworthy that we find a nonmonotonic relationship
between advertising and data interval; this effect holds
when lagged sales (carryover effect) are both included and
omitted. Advertising elasticity is lowest with quarterly data
and higher with weekly and yearly data.
$5$ $**3(*$5 ,10. Before the advent of scanner data,
researchers estimated advertising elasticity using predomi-
nantly firm-level aggregate data. Scanner data prompt the
use of panel data and the estimation of advertising response
at the individual level. Individual-level data seem to be a
more appropriate unit of analysis for measuring response to
advertising. Although the univariate means are not different
between the two groups (both approximately .12), after
accounting for other factors, advertising elasticities esti-
mated at the aggregate firm level are significantly lower
than those at the disaggregate consumer panel level. This
result is consistent with Christen et al.’s (1997) findings.
A plausible explanation for this effect is that aggregate
data are a linear combination of individual purchases,
whereas panel data estimation uses nonlinear combination
of purchases. Gupta et al. (1996, Tables 3 and 4) show that
the price elasticity from a linear approximation to a logit
model (without heterogeneity) is biased toward zero. Thus,
linear approximations using aggregate data may tend to
downwardly bias advertising elasticity.
'7(35,4,0*/($463(. A brand’s advertising may be mea -
sured in absolute (e.g., monetary value, gross rating points
[GRPs]) or relative (e.g., the brand’s share of all advertising
in the market) terms. Depending on what competitors in the
market are doing, changes in absolute advertising may not
necessarily reflect the same changes in relative advertising.
If competitors tend to match a target firm’s advertising,
large changes in absolute advertising will translate into
small changes in relative advertising. As a result, elasticities
estimated with absolute advertising will be smaller than
those estimated with relative advertising. The reverse holds
if competitors do not match a target firm’s advertising or if
they react in the opposite direction. Thus, from the differ-
ences in elasticity between relative and absolute advertising
measures, we can infer how competitors match a target
firm’s advertising strategy.
With respect to GRP and monetary (dollar) advertising
measures, we offer the following explanation: Advertisers
buy GRPs using dollars—one GRP being 1% of target audi-
ence (reach) given one exposure. Let a 1% increase in
advertising dollars increase GRPs by v% and sales by w%.
Then, by definition, dollar advertising elasticity = w, and
GRP advertising elasticity = w/v. It follows that, all else
being equal, dollar elasticity is greater than GRP elasticity
if v > 1, and GRP elasticity is greater than dollar elasticity if
v < 1.
Comparison of elasticities for absolute versus relative
advertising reveals mixed results. Advertising elasticity
with relative advertising is higher than advertising elasticity
with a dollar measure of absolute advertising but lower than
86C@4
#$(%'))'#+')($"())..&'%*)
"."$&$$)#(*'
Predicted Mean and Elasticity
Product Life Cycle
@=EB7 #0BC@4
(0:4A
(70@4
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
)01:4
%&)#"+')($%''$)+"*(%
+')($"()).
(3&(05
'7(35,4,0*
25,/$. 0&3($4(
'7(35,4,0*$4 (('('51
$4,4)13 '7(35,4,0* $(3&(05$*( $5&+
'7(35,4,0*.$45,&,5: .$45,&,5: 1)$.(4 3,&(65
3,13(5$0$.:4,4
Mean short run .22 8.5 5
633(05(5$0$.:4,4
Mean short run .12 4.6 30
Median short run .05 1.9 —a
aNot computable because the denominator in Equation 4 is zero or nega-
tive. The denominator being negative implies that the incremental profits
from an advertising increase are negative. Therefore, the firm should
reduce its advertising.
with a GRP measure of absolute advertising. However, in
absolute advertising, elasticity measured with GRP is higher
(.21) than advertising elasticity with monetary value (.09);
this difference is significant in the regression model (Table
2). As we stated previously, this finding indicates the possi-
bility of v < 1, on average. That is, firms may be operating
in a region in which a 1% increase in advertising dollars
yields a less than 1% increase in GRP. Further research can
assess whether this inference holds. We also investigated
whether the difference between dollar and GRP elasticity
was due to the correlation between advertising measure and
other model factors. We find that GRP elasticities were pri-
marily in nondurable and mature products. However, the
results on advertising measure did not change when we
excluded these variables from the model.
'7(35,4,0* :2(. Even though mean television advertis-
ing elasticity (.12) is not significantly greater than print
advertising elasticity (.11), after accounting for other factors,
we find print advertising has a lower short-term advertising
elasticity than television advertising in the regression analy-
sis (Table 2). One reason for this effect is that television
advertising, with its ability to arouse emotions, may be
more effective than print advertising, which relies primarily
on information appeals (Tellis, Chandy, and Thaivanich
2000).
/,55('"$3,$%.(4
/,4 4,10 1) .$**(' 4$.(4. The omission of a variable
biases advertising elasticity, if that omitted variable is cor-
related with both the dependent variable (sales) and the
included independent variable (advertising). The direction
of the bias is the product of the signs of the correlations of
the omitted variable with sales and advertising. Lagged
sales are likely to be correlated positively with both current-
period sales and current-period advertising (because current
advertising is often set as a proportion of past sales). There-
fore, we expect the omission of lagged sales to positively
bias advertising elasticity. Consistent with this expectation,
we find that, after accounting for other factors, the omission
of lagged sales significantly increases advertising elasticity.
Put another way, lagged sales picks up the carryover effect
of advertising. Omitting lagged sales ensures that the cur-
rent advertising picks up some of this carryover effect.
5+(31/,55('7$3,$%.(4. Among other variables, we find
that omission of distribution significantly increases advertis-
ing elasticity, perhaps because advertised brands are better dis-
tributed. Thus, distribution is positively related to both sales
and advertising, resulting in a positive omission bias. The
effects of the exclusion of other variables are nonsignifi-
cant.
1'(.+$3$&5(3,45,&4
60&5,10$.)13/$0'(45,/$5,10/(5+1'. Linear and dou-
ble log models tend to produce higher advertising elasticity
than share models. We find that, overall, functional form
has greater influence on short-term advertising elasticity
than estimation method.
0&1 3213$5 ,0* (0' 1*(0(, 5: $0' + (5(31*(0 (,5:. A more
recent trend in estimation of marketing mix is to incorpo-
rate endogeneity. We find that, in many models, omission of
endogeneity induces a negative bias in the estimates, which
is consistent with the belief that omitting endogeneity can
bias the estimates toward zero (Villas-Boas and Winer
1999). That is, advertising elasticity is lower when endo-
geneity is not incorporated.
Recent modelers have also incorporated heterogeneity by
allowing for differences in parameters across households in
the sample, through either random effects or Bayesian mod-
els (Allenby and Rossi 1999). However, we found that the
effect of omission of heterogeneity on advertising elasticity is
not significant; in other words, omitting heterogeneity does
not significantly alter the advertising elasticity estimates.
(46.541010* (3/'7(35,4,0*.$45,&,5:
Figure 1 presents the distribution of long-term advertis-
ing elasticities. There are 402 long-term brand-level adver-
tising elasticities. Their magnitudes range from –1.2 to 4.5.
More than 40% of the elasticities are between 0 and .1.
About 5% of advertising elasticities are negative. The mean
long-term advertising elasticity across the 402 observations
is .24, which is much lower than the mean of .41 in Assmus,
Farley, and Lehmann (1984).5The median long-term elas-
ticity is even lower than the mean, at .10.
Table 2 (columns 7 and 8) presents the meta-analytic
results for the long-term advertising. The variables in the
main effects model explain 29% of the variance in long-
term elasticity. The omission of lagged sales is not included
as an independent variable, because we use the coefficient
of lagged sales to estimate long-term elasticity (dependent
variable). Six variables that are statistically significant in
the short-term elasticity models are also significant in the
long-term model: year of data, product type, region, mea -
sure of advertising, omission of distribution, and functional
form.
Four variables significant in the short-term model are not
significant in the long-term elasticity model: product life
cycle, temporal interval, data aggregation, and omission of
endogeneity. One reason for the lack of significance is that
because long-term elasticity is computed using the formula
[short-term elasticity/(1 – carryover effect)], the influence
of a variable on long-term elasticity depends on its influ-
ence on both short-term elasticity and carryover effect.
These two effects acting together may enhance or dampen
the resultant coefficient. For example, mature products tend
to have smaller short-term advertising elasticity but may
have equal or smaller carryover effect compared with
growth products. Thus, the effect of product life cycle on
long-term elasticity may not be significant.
A surprising finding is that television advertising has
higher short-term elasticity but lower long-term elasticity
than print advertising. The higher long-term elasticity for
print advertising may be because information in print (espe-
cially magazines) remains in memory for a longer period than
television advertising. Some variables that are not signifi-
cant in the short-term model are significant in the long-term
model: Long-term advertising elasticity is higher during
recessions than in expansions; omitting lagged advertising
and promotion is also significant in the long-term elasticity
models.
%*'$"%#'!)$'(' *$
5The short-term advertising elasticity in Assmus, Farley, and Lehmann
(1984) is .22 and the mean carryover is .468, leading to mean long-term
advertising elasticity of .414 [= .22/(1 – .468)].
=E,4::=4A3D4@B8A8<6,=@9
/2.,&$5,104)13$0$*(34
'7(35,4,0* %6'*(5,0*. We find the mean short-term
advertising elasticity across all observations (1940–2004) is
.12, the median elasticity is .05, and elasticity is declining
over time. The finding that advertising elasticity is “small”
may upset many practitioners, especially those in the
agency business. This number seems even more troubling
compared with price elasticity, which meta-analyses suggest
is more than 20 times larger at –2.62 (Bijmolt, Van Heerde,
and Pieters 2005; Tellis 1988). However, comparing absolute
elasticity may miss some pertinent issues. First and most
important, price cuts affect revenues and profits immedi-
ately. Therefore, while a small price cut can greatly enhance
sales, it does not necessarily increase profits. Second, adver-
tising has the potential to support a higher price. Third,
price cuts can be selectively directed to only some con-
sumers to minimize harm to the bottom line. Fourth, price
cuts given to retailers might not be passed on to consumers.
Sethuraman and Tellis (1991) develop a model to integrate
these factors and draw managerial implications about how
advertising and price elasticities should affect managers’
trade-off between advertising and price discounting. In par-
ticular, they show that the advertising increase (DA) that
would yield the same profits as a given price cut (Dp) in the
short term can be computed using the following equation:
where p is the price, A is advertising, k is contribution to
price ratio, f is the fraction of consumers taking advantage
of the discount, g is retail pass-through of discount, S is dol-
lar sales, and epand eAare price and advertising elasticities,
respectively. The optimum advertising-to-sales ratio is
given by (A/S)* = (f/g)(eA/ep). Table 4 presents optimal
results using the preceding formulas for different values of
advertising elasticity with the following illustrative values:
ep(absolute) = 2.6 (Bijmolt, Van Heerde, and Pieters 2005),
f = .5, g = .5, k = .5, and A/S = .05 (Sethuraman and Tellis
1991).
Table 4 suggests a reduction in budgets allocated to con-
ventional advertising in keeping with the declining trend in
advertising elasticity. Alternatively, a firm can take steps to
increase advertising elasticity. Kent (1995) suggests creat-
ing unique messages, negotiating for noncompete coverage,
and more precise targeting of advertising as some ways to
overcome the harmful effects of advertising clutter and
increase consumers’ responsiveness to advertising.
'7(35,4,0*$0'3(&(44,10. The conventional belief is that
advertising should be reduced during recession because
sales are lower and consumers are more price sensitive and
less likely to be influenced by advertising in periods of
recession than expansion. First, our results reveal that nei-
ther short- nor long-term advertising elasticities are lower
during recession, measured as the percentage of the estima-
tion period under recession. Therefore, at a minimum, man-
agers need not reduce advertising in a recession because
they falsely believe that the sales impact of advertising will
be lower than in expansion periods.
() ,4∆
∆
AA
pp
kfg
kAs
p
A
=
−
−
ε
ε
Second, while the coefficient of recession for short-term
advertising elasticity is positive, though not statistically sig-
nificant, the coefficient for long-term elasticity is positive
and significant, suggesting that, in general, advertising elas-
ticity is higher during recession. Possible reasons for this
effect are that during recession compared with expansion,
(1) advertising clutter is lower due to cutbacks in advertis-
ing, (2) consumers pay more attention to ad messages to be
astute buyers, and (3) ad budgets are supported by higher
price and promotional incentives. One reason for the posi-
tive effect of long-term elasticity may be that the reduced
clutter and increased attention to advertising during reces-
sion may not translate into immediate purchases (short-
term), because of the tight economy at that time, but may
translate into purchases at a later point (long-term) when the
economy improves (for supporting evidence from other
studies, see Tellis and Tellis 2009).
10',5,104)$713,0*$'7(35,4,0*. Our finding of higher
advertising elasticity suggests that, all else being equal,
advertising should be higher for durable goods than non-
durable goods and for products in the early stage of the life
cycle than mature products. The higher advertising elastic-
ity in Europe than in the United States calls for a broader
understanding of whether there is overadvertising in the
United States (Aaker and Carman 1982) and underadvertis-
ing in Europe.
/2.,&$5,104)13(4($3&+(34
(/213$.,05(37$.. Advertising elasticity is significantly
different depending on if weekly, quarterly, or yearly data
are used. These significant differences underscore the need
for determining and using the “correct” data interval for
estimating advertising elasticity. The conventional view is
that the best data interval matches the interpurchase time for
the product category. Recently, Tellis and Franses (2006)
have shown that the optimal data interval that provides an
unbiased estimate of advertising elasticity is the unit expo-
sure time, defined as the largest calendar period such that
advertising occurs at most once and at the same time in that
period. This could be minutes or hours in television adver-
tising or days or weeks in print advertising. Furthermore,
these authors show that more temporally disaggregate data
does not bias the estimates. These findings suggest that, if
data are available, less aggregate daily or weekly data may
be better than quarterly or yearly data for obtaining unbi-
ased estimates of advertising elasticity.
0&13213$5,0*(0'1*(0(,5:. We find that in many models,
omitting endogeneity induces a negative bias in advertising
elasticity. One implication is that endogeneity should be
taken into account when appropriate for the model and con-
text; however, it should not be added as a “checklist” item
to the research (Shugan 2004). For example, there is some
question whether price is truly endogenous for panel mod-
els estimated at the daily level because retailers may not be
able to determine optimal prices and change them every day
(Erdem, Keane, and Sun 2008). Similar arguments could be
made for advertising response because it is unlikely that
brand manufacturers can detect and/or adjust their advertis-
ing levels quickly enough to adjust for weekly shifts in con-
sumer demand; they can detect and respond to shifts in
demand over longer periods of time (quarterly or yearly).
5+(3)$&5134. We find that the omission of distribution
has a positive effect on advertising elasticity, while func-
tional forms such as linear or double log may produce dif-
ferent elasticities. Researchers should be cognizant of these
differences and take the following steps when estimating
advertising response: (1) Include as many relevant covari-
ates (e.g., price, promotion, quality) as are available, and (2)
understand the right econometric approach for the problem
at hand or assess the sensitivity of their estimates of the
elasticity to estimation procedures.
104,*0,),&$057$3,$%.(4. Several variables are not sig-
nificant in the regression model. Does this mean that these
factors can be ignored while estimating advertising elastic-
ity? We note that the absence of evidence of effects in the
meta-analysis should not be taken as evidence of the
absence of an effect. The lack of significance could be due
to the lack of proper data, noise in the data, the aggregation
effect, or other procedural reasons. Our view is that the
“correct” data and procedure must be used when possible.
For example, the omission of heterogeneity does not signifi-
cantly influence advertising elasticity in our meta-analysis.
This does not mean researchers can safely ignore hetero-
geneity. In general, it is appropriate to account for it in
advertising response, especially when estimating with panel
data.
!
We meta-analyzed 751 brand-level short-term advertising
elasticities and 402 long-term advertising elasticities. Our
objective is to update Assmus, Farley, and Lehmann’s
(1984) meta-analysis and to add to Hanssens’s (2009)
inventory of empirical generalizations. We obtain several
useful generalizations, which we list in the following sec-
tion. Then, we present the limitations and future research
directions.
(:/2,3,&$.(0(3$.,;$5,104
The average short-term advertising elasticity across the
751 observations is .12, which is substantially lower than
Assmus, Farley, and Lehmann’s (1984) mean of .22 from
128 observations. The median advertising elasticity is even
lower at .05. The average long-term advertising elasticity
across the 402 observations is .24, which is lower than the
implied mean of .41 in Assmus, Farley, and Lehmann
(1984), from 128 observations. The median long-term
advertising elasticity is even lower at .10.
There is a decline over time in both short- and long-term
advertising elasticity. These results suggest a reduction in
conventional advertising if the firm was advertising opti-
mally in the past. On average, advertising elasticity does not
decrease during recession. If anything, there is a positive
relationship between months of recession and long-term
advertising elasticity. This result suggests that a firm does
not need to cut back on advertising in a recession because
its managers believe customers will not be responsive to
advertising. Advertising elasticity is higher for durable
goods than nondurable food and nonfood products. This
finding favors advertising for durable goods, all else being
equal.
Short-term advertising elasticity is higher for products in
the early stage of the life cycle than those in the mature
stage. This effect is especially prominent in durable goods.
This result supports focusing on advertising during the early
stage and price during subsequent stages of the life cycle,
especially for durable goods. In general, advertising elastic-
ity is higher in Europe than in North America, raising a
question of whether there is underadvertising in Europe and
overadvertising in the United States.
There is a nonmonotonic relationship between advertis-
ing elasticity and temporal interval. The elasticities esti-
mated from both weekly and yearly data are higher than
those from quarterly data. This result reinforces the need for
using the appropriate data interval as Tellis and Franses
(2006) derive.
In general, television advertising elasticity is higher than
print advertising elasticity in the short run, but print adver-
tising elasticity is higher than television advertising elastic-
ity in the long run. This finding calls for a careful consider-
ation of cost and effectiveness when allocating budgets
between the two media.
Advertising elasticity is lower when endogeneity in
advertising is not incorporated in the model. When appro-
priate, researchers should attempt to incorporate endogene-
ity using recently developed New Empirical Industrial
Organization models (Nevo 2001) or acknowledge that the
estimate may be lower because of the omission.
,/,5$5,104$0'635+(3(4($3&+
Our study has some limitations that are typical of most
meta-analytic research. First, while we have tried to be
exhaustive in our literature review, we may have overlooked
some publications that estimate advertising elasticity. Sec-
ond, in identifying the factors that influence advertising
elasticity, we are limited by the variables that are available
in the original studies. For example, we could not collect
data on all four stages of the life cycle or individual country
of origin, so we could not estimate influences of these
variables on advertising elasticity.
These limitations provide potential directions for further
research. On a more substantive level, researchers in the
future should try to analyze more growth products, durable
goods, industrial goods, and service goods. Further research
could also measure the effects of online advertising, incor-
porate them in meta-analyses, and perform a meta-analysis
of the duration of advertising—the period during which the
effect of advertising lasts.
REFERENCES
Aaker, David A. and James M. Carman (1982), “Are You Over
Advertising?” 1630$. 1) '7(35,4,0* (4($3&+, 22 (August/
September), 57–70.
Allenby, Greg M. and Peter E. Rossi (1999), “Marketing Models
of Consumer Heterogeneity,” 1630$. 1) &101/(53,&4, 89
(November), 57–78.
Assmus, Gert, John U. Farley, and Donald R. Lehmann (1984),
“How Advertising Affects Sales: Meta-Analysis of Econometric
Results,” 1630$.1)$3-(5,0*(4($3&+, 21 (February), 65–74.
Belsley, David A., Edwin Kuh, and Roy E. Welsch (1980), (*3(4
4,10,$*0145,&4. New York: John Wiley & Sons.
Bijmolt, Tammo H.A. and Rik G.M. Pieters (2001), “Meta-Analysis
in Marketing When Studies Contain Multiple Measurements,”
$3-(5,0*(55(34, 12 (May), 157–69.
———, Herald J. van Heerde, and Rik G.M. Pieters (2005), “New
Empirical Generalizations on Determinants of Price Elasticity,”
1630$.1)$3-(5,0*(4($3&+, 42 (May), 141–56.
%*'$"%#'!)$'(' *$
=E,4::=4A3D4@B8A8<6,=@9
Christen, Markus, Sachin Gupta, John C. Porter, Richard Staelin,
and Dick Wittink (1997), “Using Market-Level Data to Under-
stand Promotion Effects in a Nonlinear Model,” 1630$.1)$3
-(5,0*(4($3&+, 34 (August), 322–34.
Clarke, Darrel (1976), “Econometric Measurement of the Duration
of Advertising Effect on Sales,” 1630$.1)$3-(5,0*(4($3&+,
13 (November), 345–57.
Danaher, Peter, Andre Bonfrer, and Sanjay Dhar (2008), “The
Effect of Competitive Advertising Interference on Sales of
Packaged Goods,” 1630$.1)$3-(5,0*(4($3&+, 45 (April),
211–25.
Erdem, Tülin, Michael Keane, and Baohung Sun (2008), “The
Impact of Advertising on Consumer Price Sensitivity in Experi-
ence Goods Markets,” 6$05,5$5,7($3-(5,0*$0'&101/,&4,
6 (June), 139–76.
Fischer, Marc and Sonke Albers (2010), “Patient- or Physician-
Oriented Marketing: What Drives Primary Demand for Pre-
scription Drugs?” 1630$.1)$3-(5,0*(4($3&+, 47 (Febru-
ary), 103–121.
Gallet, Craig A. (2007), “The Demand for Alcohol: A Meta-Analysis
of Elasticities,” +( 6453$.,$0 1630$. 1) *3,&6.563$. $0'
(4163&(&101/,&4, 51 (June), 121–35.
Gupta, Sachin, Pradeep Chintagunta, Anil Kaul, and Dick R. Wit-
tink (1996), “Do Household Scanner Data Provide Representa-
tive Inferences from Brand Choices? A Comparison with Store
Data,” 1630$.1)$3-(5,0*(4($3&+, 33 (November), 383–98.
Hanssens, Dominique M. (2009), /2,3,&$. (0(3$.,;$5,104
%165$3-(5,0*/2$&5. Cambridge, MA: Marketing Science
Institute.
Hu, Ye, Leonard M. Lodish, and Abba M. Krieger (2007), “An
Analysis of Real World TV Advertising Tests: A 15-Year Update,”
1630$.1)'7(35,4,0*(4($3&+, 17 (September), 341–53.
Kent, Robert J. (1995), “Competitive Clutter in Network Televi-
sion Advertising,” 1630$.1)'7(35,4,0*(4($3&+, 33 (March/
April), 40–46.
Lambin, Jean L. (1976), '7(35, 4,0* 1/2(5 ,5,10 $ 0' $3- (5
10'6&5,0.,*121.:7(3 ,/(. Amsterdam: North-Holland.
Lodish, Leonard M., Magid M. Abraham, Stuart Kalmenson,
Jeanne Livelsberger, Beth Lubetkin, Bruce Richardson, and
Mary Ellen Stevens (1995), “How T.V. Advertising Works: A
Meta-Analysis of 389 Real World Split Cable T.V. Advertising
Experiments,” 1630$. 1) $3-(5,0* (4($3&+, 32 (April),
125–39.
Nevo, Aviv (2001), “Measuring Market Power in the Ready-to-Eat
Cereal Industry,” &101/(53,&$, 69 (March), 307–342.
Sethuraman, Raj, V. Srinivasan, and Doyle Kim (1999), “Asym-
metric and Neighborhood Cross-Price Effects: Some Empirical
Generalizations,” $3-(5,0*&,(0&(, 18 (Winter), 23–41.
——— and Gerard J. Tellis (1991), “An Analysis of the Tradeoff
Between Advertising and Price Discounting,” 1630$.1)$3
-(5,0*(4($3&+, 31 (May), 160–74.
Shugan, Steven (2004), “Endogeneity in Marketing Decision
Models,” $3-(5,0*&,(0&(, 23 (Winter), 1–3.
Tellis, Gerard J. (1988), “The Price Elasticity of Selective
Demand: A Meta-Analysis of Econometric Models of Sales,”
1630$.1)$3-(5,0*(4($3&+, 25 (November), 331–41.
——— (2004), ))(&5,7('7(35,4,0*18#+(0$0'#+:'7(3
5,4,0*#13-4. Thousand Oaks, CA: Sage Publications.
——— and Philip Hans Franses (2006),”Optimal Data Interval for
Estimating Advertising Response,” $3-(5,0* &,(0&(, 25
(May/June), 217–29.
———, Rajesh Chandy, and Pattana Thaivanich (2000), “Decom-
posing the Effects of Direct Advertising: Which Brand Works,
When, Where, and How Long?” 1630$. 1) $3-(5,0*
(4($3&+, 37 (February), 32–46.
——— and Kethan Tellis (2009), “Research on Advertising in
Recession: A Critical Review and Synthesis,” 1630$.1)'7(3
5,4,0*(4($3&+, 49 (September), 9–40.
Vakratsas, Demetrios and Tim Ambler (1999), “How Advertising
Works: What Do We Really Know?” 1630$.1)$3-(5,0*, 63
(January), 26–43.
Villas-Boas, J. Miguel and Russel S. Winer (1999), “Endogeneity
in Brand Choice Models,” $0$*(/(05&,(0&(, 45 (October),
1324–38.
Copyright of Journal of Marketing Research (JMR) is the property of American Marketing Association and its
content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's
express written permission. However, users may print, download, or email articles for individual use.
