M PRA
Munich Personal RePEc Archive
Understanding Advertising Adstock
Transformations
Joseph Joy
Independent
15. May 2006
Online at [Link]
MPRA Paper No. 7683, posted 12. March 2008 01:45 UTC
UNDERSTANDING ADVERTISING ADSTOCK TRANSFORMATIONS
JOY V. JOSEPH1
1 The author is a Vice-President with Citigroup, NY, USA.
1
ABSTRACT
Advertising effectiveness and Return on Investment (ROI) are typically measured
through econometric models that measure the impact of varying levels of
advertising Gross Ratings Points (GRPs) on sales or on purchase decision and
choice. TV advertising has both dynamic and diminishing returns effects on
sales, different models capture these dynamic and nonlinear effects differently.
This paper focuses on reviewing the econometric rationale behind the
popularized Adstock transformation model that allows the inclusion of lagged and
non-linear effects in linear models based on aggregate data.
Keywords: Advertising, Adstock Model, Non-linear transformation, Marketing-
Mix
2
1. Introduction
This paper is intended as a review of existing models of Television Adstock
transformations that enable the inclusion of dynamic and nonlinear effects of
Television advertising within linear sales response models.
Television advertising is one of the largest investments for consumer
marketing and companies invest a lot of effort in measuring the impact and
ROI of TV advertising. This is typically done using either individual response
(e.g. Discrete Choice models) aggregate response (e.g. Marketing-mix
models). For the purposes of this paper, we will restrict our inquiry to
aggregate response models. These models are linear in parameters but can
account for non-linearity through variable transformations.
It is well known that TV advertising has both dynamic and diminishing returns
effects on sales. Television advertising has an effect extending several
periods after the original exposure, which is generally referred to as
advertising carry-over or ‘Adstock’ (Broadbent, 1979).
The underlying theory of Adstock is that exposure to Television Advertising
builds awareness in consumer markets, resulting in sales. Each new
exposure to advertising increases awareness to a new level and this
awareness will be higher if there have been recent exposures and lower if
there have not been. This is the decay effect of Adstock and this decay
eventually reduces awareness to its base level, unless or until this decay is
reduced by new exposures.
1.1. Decay Effect
This decay effect can be mathematically modelled and is usually
expressed in terms of the ‘half-life’ of the advertising. A ‘two-week half-life’
means that it takes two weeks for the awareness of an advertising to
decay to half its present level. Every Ad copy is assumed to have a unique
half-life. Some academic studies have suggested half-life range around 7-
12 weeks (Leone 1995), while industry practitioners typically report half-
lives between 2-5 weeks, with the average for Fast Moving Consumer
Goods (FMCG) Brands at 2.5 weeks.
Adstock half-life can be estimated through a distributed lag model
response with lags of the TV Gross Ratings Point (GRP) variable, using
Least Squares, or from the lag parameter in the Adstock formulation
(geometric lag), or indirectly using a ‘t-ratio’ method by recursively testing
different values for the decay parameter through an iterative process with
sales panel data or awareness/image tracking data against the
corresponding advertising schedules (Fry, Broadbent, Dixon 1999).
3
1.2. Diminishing Returns Effect
Advertising can also have diminishing returns to scale or in other words
the relationship between advertising and demand can be nonlinear. For
example, the effect of 200 GRPs of advertising in a week on demand for a
brand maybe less than twice that achieved with 100 GRPs of advertising.
Typically, each incremental amount of advertising causes a progressively
lesser effect on demand increase. This is a result of advertising saturation.
The usual approach to account for saturation is to transform the
advertising variable to a non-linear scale for example log or negative
exponential transformations. It is this transformed variable that is used in
the sales response models.
Tt can be transformed to an appropriate nonlinear form like the logistic or
negative exponential distribution, depending upon the type of diminishing
returns or ‘saturation’ effect the response function is believed to follow.
For example if advertising awareness followed a logarithmic distribution,
then in a linear sales response model we would have:
ST = βLog (TT) + εT (1)
Where ST is sales at time T, TT is the level of advertising GRPs at time T
and εT is the random error component. In this case for 100 GRPs and
assuming β =1, the sales effect of advertising would be 4.6 units and for
200 GRPs the sales effect would be 5.3 units. Therefore, for a 100%
increase in advertising we would only have a 15% increase in sales.
Advertising typically has a lower elasticity than other elements of the
marketing-mix. This is considered acceptable by Brand Managers since
advertising is also believed to have a long-term positive effect on Brand
Equity, which is usually not captured by most econometric models.
Several versions of Adstock transformation are applied in the industry, we
will examine some popular models in the following sections.
2. Adstock Models
2.1. Simple Decay-Effect Model
Below is a simple formulation of the basic Adstock model of Broadbent
(1979):
At = Tt + λAt-1 t=1,…., n (2)
Where At is the Adstock at time t, Tt is the value of the advertising variable at
time t and λ is the ‘decay’ or lag weight parameter. Inclusion of the At-1 term,
4
imparts an infinite lag structure to this model, with the effect of the first
Adstock term, approaching 0, as t tends to ∞.
This is a simple decay model, because it captures only the dynamic effect of
advertising, not the diminishing returns effect.
This model is also approximately equivalent to an infinite distributed lag
model as shown below:
From (1) we have, At = Tt + λAt-1
Recursively substituting and expanding we have,
At = Tt + λTt-1 + λ2Tt-2 +… λnTt-n (3)
Since λ is normally less than 1, λn will tend to 0 in limit as n tends to ∞.
Therefore, this infinite polynomial distributed lag can be approximated by
imposing a finite lag structure within an Almon distributed lag model (Almon
1965).
2.2. Log Decay Model
The Log Decay model applies a straightforward logarithmic distribution to the
advertising variable
At = LogTt + λAt-1 (4)
This is a relatively inflexible non-linear specification of the Adstock model, as
it doesn’t allow for varying saturation levels.
2.3. Negative Exponential Decay Model
The below formulation applies a negative exponential distribution to the basic
Adstock formula, using two parameters, the ‘decay’ or lag weight parameter λ
and the learning rate or saturation parameter ν.
At = 1-e (-ν Tt) + λAt-1 (5)
This model is comparatively more flexible as different values of the parameter
ν can be empirically tested in a response model to correctly measure the level
of current advertising saturation.
2.4. Logistic (S-Curve) Decay Model
Using a logistic distribution instead of negative exponential will impart an S-
shape to the Adstock variable, implying an inflexion point or ‘threshold’ level
5
of GRPs before diminishing returns set in. Below this threshold, the logistic
function imparts exponential returns.
At = 1/ (1+e (-ν Tt)) + λAt-1 (6)
As in the negative exponential model, the parameter ν can be used to model
different saturation levels.
3. Half-Life Estimation
Advertising half-life, η is calculated in the same manner as estimating decay
half-life for radioactive substances:
Assume that in time-period t+n, At would have decayed to At/2. Therefore
using equation (1) and assuming that no new advertising is present (so the
first term in the equation equals zero),
At+1 = λAt (7)
In time period t+n, where n is the value of the half-life η, we have,
At+n = At/2 (8)
Therefore from (3) and (4), and recursively substituting,
At/2 = λnAt (9)
And,
λn=1/2 (10)
Finally taking n as the value of the Half-Life η, we have,
η = Log (0.5)/Log (λ) (11)
Different values for λ can be empirically tested in an econometric model to
estimate the half-life for an advertising program.
For example, a λ of 0.25 implies a half-life of 2.4 weeks.
4. Conclusion
Adstock transformation is an efficient and effective technique to incorporate
nonlinear and dynamic advertising effects in sales response models. The
alternative option of building a dynamic nonlinear model is both
computationally expensive and complex to estimate.
6
The industry popularity of Adstock transformations is due to its relatively
simple nature and direct applicability to measuring Advertising Saturation. It
helps Brand Managers determine if more or less investment is needed to
make advertising more effective and yield a better ROI. One point that should
be kept in mind is that the minimum threshold ROI acceptable is typically a
break-even ROI, since in the short-term advertising pays for itself and there is
a longer-term benefit over and above the short-term return in the form of
Brand Equity improvement.
Measuring the Advertising Half-Life enables Brand Managers to also
efficiently space advertising schedules to maximize the effect of each
advertising exposure. Depending on whether the half-life of a copy is long or
short exposures can be efficiently spaced to maximize sales response.
References
Almon, S. (1965) “The distributed lag between capital appropriations and
expenditures”, Econometrica, 33, 178-196.
Broadbent, S. (1979) “One Way TV Advertisements Work”, Journal of the
Market Research Society Vol. 23 no.3.
Dubé, J., Hitsch, G. J., Manchanda, P. (2004) “An Empirical Model of
Advertising Dynamics”, Quantitative Marketing and Economics, 3 (2005) 2
Pg. 107-144
Fry, T.R.L., Broadbent, S. and Dixon, J.M. (2000), “Estimating Advertising
Half-life and the Data Interval Bias”, Journal of Targeting, Measurement &
Analysis in Marketing, 8, 314-334.
Leone, R.P. (1995) “Generalizing what is known about temporal aggregation
and advertising carry-over”, Marketing Science, 14, G141-G150.
