Disclaimer: This is a machine generated PDF of selected content from our databases.

This functionality is provided solely for your

convenience and is in no way intended to replace original scanned PDF. Neither Cengage Learning nor its licensors make any
representations or warranties with respect to the machine generated PDF. The PDF is automatically generated "AS IS" and "AS
AVAILABLE" and are not retained in our systems. CENGAGE LEARNING AND ITS LICENSORS SPECIFICALLY DISCLAIM ANY
AND ALL EXPRESS OR IMPLIED WARRANTIES, INCLUDING WITHOUT LIMITATION, ANY WARRANTIES FOR AVAILABILITY,
ACCURACY, TIMELINESS, COMPLETENESS, NON-INFRINGEMENT, MERCHANTABILITY OR FITNESS FOR A PARTICULAR
PURPOSE. Your use of the machine generated PDF is subject to all use restrictions contained in The Cengage Learning
Subscription and License Agreement and/or the Gale General OneFile Terms and Conditions and by using the machine generated
PDF functionality you agree to forgo any and all claims against Cengage Learning or its licensors for your use of the machine
generated PDF functionality and any output derived therefrom.

Adstock modelling for the long term

Authors: Simon Broadbent and Tim Fry
Date: Oct. 1995
From: Journal of the Market Research Society(Vol. 37, Issue 4)
Publisher: Sage Publications, Inc.
Document Type: Article
Length: 8,259 words

Abstract:
There are two procedures for modelling the long-term impact of advertising. One is the application of adstocks with a long half life.
The other makes use of a new expression known as the cumulative adstock which is always in dynamic balance with negative forces.
The result of this is a floating base which may either fluctuate or remain static. Case histories are presented.

Full Text:
The paper suggests two ways to model the long term effects of advertising. One simply uses adstocks with a long half life. The other
uses a new expression: cumulative adstock is in dynamic balance with a negative trend which represents hostile forces. This
produces a 'floating base', which may move up or down or be stationary. Eleven case histories are summarised, all adequately fitted
by one of these systems. A practical strategy for the analyst is recommended.

Introduction

Adstock modelling (Barnard & Smith 1989; Broadbent 1993; Institute of Practitioners in Advertising (1981-95)), with a short half life, is
a well-established way of representing the size of those advertising effects which last a few weeks. The equation for the objective
variable includes a constant and an adstock term as well as the effects of other factors; in its linear form (a multiplicative equivalent is
in Appendix 2)

B + d.[Sads.sub.t]

The industry continues its attempts to represent longer term and indirect effects (Baker 1990). A family of models which does so
successfully should have these characteristics:

* cover a wide variety of situations,

* be easily understood by non-technical managers and `sensible', that is, consistent with their own mental model of the market-place,
which may of course change as a result of the modelling,

* and be fitted by the regression techniques generally available (Ordinary Least Squares).

This paper suggest a family of models to meet these needs. We do not see why a single model should cover all situations, but we
hope that a small set of models will cover most of them.

These models are not about `steps' or `jumps', when a dramatic re-positioning rapidly shifts a brand to a higher level -- or when a
competitive launch does the reverse. These cases can be dealt with in obvious ways. We are tackling the cases where there are
trends and irregular movements in sales or other behaviour which we think may be linked to different amounts of recent advertising.

The models

Since adstock stands for the current `pressure' from previous advertising, one obvious way to show a long term effect is to use
adstock with a long half life(*). In our applications a short half life is generally six or eight weeks and is defined as ten or less. A long
half life can be for half lives like six months or a year or two. Estimating the half life is discussed in Appendix 1.

Thus the `tease' in our equation is not just a constant ([B.sub.t] = B) but a constant plus a term which moves up and down depending
on the long term advertising history ([B.sub.t] = B + e.[Lads.sub.t]). Short term adstock and other factors can be added to this. We call
this the long term adstock model.
There may be a negative (and assumed linear) trend term among the other factors, representing for example the constant hostile
actions of other brands. With rather constant advertising weight, such a trend term will in due course outweigh advertising and the
brand will sink. But there may be situations in which constant advertising is able to counteract this trend and keep the brand afloat. In
such a case we may think of the base [B.sub.t] being changed in each period by a constant amount, which is the period's contribution
to the negative trend, say c, but also increased by an amount proportional to current long term adstock, say e. [Lads.sub.t].

Thus [B.sub.t] = [B.sub.t-1] c + e. [Lads.sub.t].

This recursive formula can be turned into a normal form by starting at the original base B at t = 1, and adding the successive
movements. We finish at time t with

[B.sub.t] = B + t.c + e. [Sigma] [Lads.sub.t]

where the long term adstocks have been cumulated from time t = 1 to time t.

Short term adstock and other factors can be added to this. We call this the cumulated long term adstock model; it differs
fundamentally from the long term adstock model.

At first sight [Sigma] [Lads.sub.t] is an unusual term. It seems to be a double-cumulation -- once because each term in the advertising
history is represented in [Lads.sub.t] and again because of the [Sigma]. But the terms have been multiplied by a small number before
being added in adstock, and on average [Lads.sub.t] is only as large as the average advertising weight. The point of the continually-
increasing [Sigma] term is that it may be balanced by the growth of t.c . They in fact cancel out when

-c / e = average [Lads.sub.t].

This is the condition that the base appears to be constant: it is floating but bobbing up and down at some average level. Or, there
may be irregular but overall growth because the copy is effective enough and the weight sufficient to cause overall increase (or vice
versa). Hence, we also call this the floating base model.

Note that even when modelling concludes that a constant base is a reasonable fit (long term effects are non-significant), this may well
be because the brand is actually in the equilibrium just described. The correct conclusion in such a case is not that advertising
definitely has no long term effect, but that our observations may have been made in a period when average [Lads.sub.t] are equal or
close to -c /e. This on its own does not allow us to estimate -c or e, but guessing -c (`without advertising we might decline at 10% a
year') allows us to calculate a possible e. In this intuitively acceptable but non-rigorous way we may show alternative explanations of
history; we may also make a guess at the effect of changes in the advertising budget.

The concept of a floating base has an important consequence when we come to the calculation of price and other elasticities. Prices
often change while other aspects of the brand are changing at the same time. We advertise to support a price rise, for example; a
competitor introduces a benefit and raises his price (so our relative price falls) and because our quality is now relatively less good,
our base falls too.

The simple procedure often adopted in order to estimate price elasticity is that we regress our sales share on our relative price. We
may also allow for some other factors, but typically we assume the base is fixed, though this assumption is seldom spelled out. It is
easy to show that when this assumption is broken -- and our argument for a floating base implies that it often is broken -- we get a
biased estimate for price elasticity. If the base is increasing, and price has generally been rising, then our estimate of price elasticity
is (numerically) smaller than it should be; it is even possible for a true negative to be calculated as a positive price elasticity (an
anomaly which is occasionally observed but rarely explained). The other combinations (of base and of price movement) produce the
complementary biases. Thus we argue that, even for purposes outside the estimation of ad effects, a floating base should be one of
the hypotheses examined.

These two models -- in their multiplicative form as well as additive -- are the `family of models' we propose. In Appendix 2 we write
them out in full and discuss how to fit them. For completeness, the formula for adstock, derived from the advertising history, is given
in Appendix 1. Later in the paper, we show the result of fitting these models in real cases.

In practice, advertising may not have detectable effects at all, so we are ready to find these advertising terms non-significant. Also,
the presence or absence of a trend term depends on the market-place circumstances. Nor is it clear in advance which of the four
sorts of model will be appropriate -- long term or cumulated long term, additive or multiplicative. Then there are various ways of fitting
the models -- directly or using differences, which we have tested and which are discussed in Appendix 2. The purpose of the real
cases analysed below is to suggest the strategy for the analyst in deciding which sorts of fits to try.

We have attempted, but do not recommend in this paper, modelling the indirect effects of advertising. For example, with sufficient
amounts of the right support, are we able to charge a better price? Or are promotions more effective? Or do we find we have a lower
price elasticity when we raise relative price? It is difficult to separate these sorts of benefit from a straight gain in share or volume.

Consumer Brand Equity (Broadbent 1992) has been suggested as a statistic which quantifies the `value' of a brand, separate from
temporary accidents of price and distribution. Is the floating base the same measure? It is not, though there are similarities. The base
here may be calculated with other factors taken into account than just price and distribution (which are used in Consumer Brand
Equity). The elasticities in brand modelling are specific to the brand rather than averages for the category. In both approaches we
attempt to put a number on the underlying worth or value of the brand, as shown by consumers' behaviour and distinct from short
term movements. Sometimes it is well worth making the distinction between total sales, base or equity, and immediate reactions to
marketing activities.

Example of floating base adstock modelling

The first of our examples is now examined in some detail, both as an example of the method of fitting and of the type of explanation
discussed with the client. An outline of the market-place situation is given in Appendix 3; this is Case A.

The data consisted of 119 4-week periods of volume sales share in the appropriate category (the objective variable), with the
following explainers (others having been tried but found wanting):

-- the brand's price, relative to the category average, -- two sorts of promotion which are expected to have more than the normal price
effect; in total these ran for 15 periods and applied to varying proportions of the turnover, -- a number of adstocks with different short
half lives

These were all successful candidates for short term effects.

In addition there was a figure for All-Commodity Volume sales (ACV), since the share of this brand varied with the feel-good factor in
the economy. This number moves slowly, and it is more natural to count it in the long-term factors.

Then there were the possible components of the base, in the sense used in this paper:

-- a number of adstocks with different and longer half lives. -- the cumulative adstocks calculated from these, -- a trend term.

We did not know in advance whether the dynamic balance between trend and [Lads.sub.t] discussed above would prove to be an
explainer worth consideration, nor what half life might be appropriate. It might also be that longer half life adstock was a better model.

Thus the array for analysis consisted of 119 rows with one column for the objective variable, four short term and four long term
adstocks, four cumulative adstocks and five other terms -- 18 columns in all (originally more, with explainers later rejected).

Ordinary least squares was used in both stepwise and normal regressions, to try out a large number of different explanations. The
solution proposed is shown in Appendix 3.

The different fits for the cumulative adstocks with different half lives were very close. The choice of half life can hardly be made on
statistical grounds. Here as elsewhere, fitting is an art which depends not only on the data in front of us, but on other evidence and on
common sense. To suggest a three-year half life seems less reasonable in this case than one year, but the point is arguable. The fits
with long term adstocks were however clearly worse than the floating base model.

A referee has asked, with reference to Case D in Appendix 3, how it is possible to estimate a long half life such as the 52 weeks
there, when there are only two or three years data available. The answer is that we are not estimating half life directly, but the fade
parameter F. We are not waiting for the effect to be over, or even half over, but we are looking for the reduction in the effect between
each of the periods available. For 52 weeks half life and 4-week data, F is 0.946. In case D, we get an R-squared of 0.986 with this
decay, and t = 33.2 for the adstock coefficient; but R-squared is 0.971 and t = 23.1 when F = 0.891, which corresponds to a 26 week
half life. These seem sufficient grounds for choosing the former.

Three charts here indicate how the results can be presented to the client. The base is made up of two variable components shown in
figure 2. The floating base consisted of the cumulative adstock minus trend, each current value being multiplied by the relevant
coefficient. This difference resulted in growth for the first 80 periods, then in sharp decline. This was because the advertising
investment fell in real terms. The difference between the high and the low of the floating base is seven category share points --
commercially very important. The second component is the gentle movement of ACV.

There is a presentation point here. This index has an average of 115 and the coefficient is 0.27; their product is +31. Relative price
averages 102 and its coefficient is -0.40; their product is -41. The intercept of the regression is 53. To present the base we could add
to the intercept the effects of trend, cumulative adstocks and ACV. Then the short term price effects could be presented as relative
price times its coefficient. A large `tease' then has a large `price effect' subtracted from it. This is the way such numbers are usually
calculated but they give the wrong impression. Instead we use an ACV variable which is the observed value minus the minimum ACV
value observed, and add the corresponding difference to the base. In effect we are regressing with the value of the variable above its
minimum, rather than with the whole index. To correct this, a constant corresponding to the minimum price effect is added. Thus the
`tease' has a sensible-looking value, and the price effects are shown as lost sales from this base when the price is above the
minimum.

Figure 3 shows this result: the new base trends first up and then down. Price effects are mainly lost sales below this, especially in the
last three years; but the promotions show as intermittent steps up, sometimes with overall results above the base. Commonsense
can accept this view of `base' i.e. what sales would be if price were at its minimum and there were no special promotions.

Finally, figure 4 shows the resulting calculated fit again, this time compared with actual sales. Clearly the model captures the main
movements; the it-squared which is 0.87 (for both actual and adjusted) and especially the t-values are very satisfactory. There are
discrepancies which need further study, for example by describing competition in more detail, or looking at other promotions, or by
changing the assumption of a common creative effectiveness over all advertising (the short term effect of the last burst could well be
higher). There are two outliers which should be investigated. But here we do not go into these further refinements. We can now
comment on the advertising weight and efficiency required to control the main movement. The relevant terms are:
trend = -0.54 per 4-weeks

cumulative long term adstock effect = 0.0063 * [Sigma] [Ads.sub.52]

where the average [Ads.sub.52] over the 119 periods is 87, but has been as low as 47 and as high as 120. Thus in every period the
floating base would drop by 0.54 share points; but it is sustained by an advertising effect which averages 0.55 (i.e. 0.0063 * 87) and
has in fact varied between 0.30 and 0.76. Sales may drop in each period by 0.24 share points (i.e. -0.54 + 0.30) or rise by 0.22
points, depending on the current adstock which shows the recent history of adspend. The indications are that with present copy the
brand will be stable at 86 * 13 = 1,118 TVRs annually(0.54/0.0063 = 86). At the lower adspend of 47 * 13 = 610 TVRs annually, the
sales efficiency has to be improved by 85% (87 /47 = 1.85) for sales stability.

Without advertising support the estimated annual loss in share points is 13 * 0.54 = 7; this is an annual percentage loss of 15% of
share, since share itself averages 46% and 100 * 7/46 = 15.

At this point we consider the statistical tests often reported with such econometric analyses. The Shazam procedure used offers a
large battery of such tests, but we give only a choice of those which we believe really matter. The reason for being selective is that
we come to marketing time series with prior information which is not assumed in pure statistics. We know for example that
consumers' behaviour has a certain momentum: what they buy this month is more likely (than a simple random probability assumes)
to be what they bought last month. Also, during periods in which we (or a competitor) have a temporary advantage, and this is not
reflected in the explanatory variables (for example a new competitive brand establishes itself on the shelf), then residuals are going to
be positive (or negative) for a few continuous periods. This clearly happens here between periods 65 and 77, for example, when
actual sales are above the fit, and between periods 25 and 37, when they are below. In other words, successive observations are not
independent. The Durbin-Watson statistic reflects this: it is 1.1. Serial correlation is evident at lags of one or two periods. In this
regard, the model is mix-specified. We do not report Durbin's h test, as this is relevant only when the model includes a lagged
dependent variable (e.g. a lagged share is on the right-hand side).

Thus serial correlation would in some circles be considered a problem here. But what is the result of this `problem'? It is to reduce the
standard errors of the parameter estimators: the t values shown in Appendix 3 are too high. But what is more important is that the
estimates themselves are still unbiased, consistent and normally distributed. When we get a t value of 6.6, as we do for cumulative
adstock, this hardly matters: the effect is still rather well determined. When we get 2.4, as we do for one of the promotional
techniques, it is possible that this effect does not really pass the traditional 95% significance level. Does this matter? We think not.
The task is to make the best estimate we can of the size of each effect, knowing in advance from previous experience what we
expect. It is not to reject any estimate which has a calculated probability (on inappropriate assumptions) above 5% of having the
wrong sign. Thus we would happily use the estimated coefficient for this type of promotion in future decisions, remembering only that
it is not particularly well determined.

Some of the other standard tests are passed by the model for Case A (the output is available on request). The residuals are
homoscedastic (seven heteroscedasticity tests were passed) and normally distributed (goodness of fit and Jarque-Bera tests),
without evidence of ARCH; the coefficient estimates are stable; there is no evidence of structural change or structural break. We do
not use the Chow test as we do not have any post-prediction data. There is some evidence of mix-specification (Reset tests) which is
related to the serial correlation discussed above.

Eleven examples

In Appendix 3, eleven cases are outlined. Their main results are now summarised. The examples cover a wide range of situations:
sales of different packaged goods in the UK (figure 1 indicates that one brand is growing, another is in decline, one goes first one way
then the other, and a fourth is stable) and road safety behaviour in Australia. It is the effect of television advertising, measured by
TVRs or TARPs, which is studied in all these cases.

In every case (and there were no rejected cases) a long term effect of advertising was measured. Eight of these were with floating
base modelling (i.e. a negative trend and a positive cumulative long term adstock variable -- both statistically significant) and three
used the long term adstock model. A short half life adstock variable was also significant in eight cases -- seven of the floating base
cases and one with long term adstock.

We can use advertising elasticity for the short term to decide whether the advertising effect is economic, since the investment can be
compared with a return which is over within a few months. The costs and the benefit are both reasonably well-determined. With the
seven straightforward short term adstocks we found an average elasticity of 0.12; this is usually part only of the total advertising
benefit.

Planning ahead with the floating base model, elasticity is no longer an easy way to assess the effectiveness of our average adspend
or ratings, nor whether this is economic. We have to consider the effect over more than a (financial) year. Thus we require forecasts
of category movements, fixed and variable costs, discount rates and so on, in order to determine the current net worth of our
decision. We have a method for this, called Continuous Budgeting, developed with Arthur Andersen, similar to the spreadsheet
system previously described (Broadbent 1990) but not discussed here.

We nevertheless give elasticities below for the five commercial cases with cumulative adstocks and find the range 0.06 to 1.00, with
the average 0.48. This is well above that for the short term, consistent with the often-made comment that the longer term benefits of
advertising -- even on volume alone -- are usually greater than those for the short term.
Countering the cumulative or long term adstocks are the trend terms. Here elasticity is meaningless, but we can calculate the
equivalent percentage decline in brand shares in the absence of advertising, in the way shown in the example above. The range is 0
to 36%, with the average decline 16%. This figure is high, and if typical it shows the tough competition facing advertised packaged
goods today. They have often lifted themselves through communications but have not reached really stable positions. Without
continued support they may fall steeply.

Of the seven commercial cases, price relative to the category had a significant effect in six and other promotions in three. The range
of price elasticities was between -0.6 and -3.4, with an average of-1.6, close to figures previously reported.

Some form of distribution was significant in three cases. Seasonality was a significant factor in five cases.

Overall, our requirements for a family of models has been met. The fits are acceptable to commonsense, the sizes of effects are
reasonable, and an easily avail- able method of regression is sufficient.

Discussion

If these cases are typical, the strategy suggested for the analyst is as follows (in addition to normal procedures):

-- Create adstocks with a range of short half lives, for example 2, 4, 6, 8 and 10 weeks (as explained in Appendix 1). -- Create
adstocks and cumulative adstocks with a range of longer half lives, for example 26, 39, 52, 78 and 104 weeks. -- Include a trend term
with the other explainers, for which obvious candidates are seasonality, relative price, distribution, periods of exceptional promotions,
competitive activities. -- Analyse this data set and also one of the logs of these variables. If the results are not seriously different,
choose the additive model. -- Start with the presumption that trend, a cumulative adstock (say with 52 weeks half life) and a short half
life adstock (say 6 weeks) are likely to be major explainers. -- Alternatively, a long half life adstock (with or without trend) and a short
one. -- Stepwise regression is a good way to explore the structure, but reject variables with counter-intuitive signs (negative
coefficients for adstock, for example) or very small t values. Explore alternative half lives round the suggested values.

The estimation of long term direct advertising effects is not as hard as has often been said.

APPENDIX 1 Adstock and half lives

To calculate adstocks [a.sub.1], [a.sub.2], ... for a series of TV ratings [r.sub.1], [r.sub.2], ... we need two further items. First, the
average TVRs before this series began (actually, the previous adstock at period 0); call this P. Second, the fade parameter F which
determines how fast the effect fades and which states that, without any added advertising,

[a.sub.1] = F*[a.sub.t-1].

F is related to the half life of the advertising or HL, in the units of period length, by HL = log{(1+F)/4} / log{4}. We use the convention
`first period counts half' which allows for the fact that for the period in which advertising appears, its exposure may be either before or
after the time of purchase.

We define C = (1-F)/(1 +F), [U.sub.0] = C * P, [V.sub.0] = 2 * F * P/(1+F) and then

[a.sub.1], [a.sub.2], ... are calculated recursively from

[U.sub.1] = C * [a.sub.i]

[V.sub.i] = 2 * F * [U.sub.i-1] + F * [V.sub.i-1],

[a.sub.i] = [U.sub.i] + [V.sub.i]

To choose the fade parameter or its equivalent the half life there are two methods. We may use Marquardt or a similar estimation
process in which this parameter is simply one more of those calculated during the fit. These procedures are not Ordinary Least
Squares and are not everywhere available. Nor do we see the overall result of choosing different half lives -- we get a single estimate.

Hence it is recommended to use a range of half lives, such as those named above, calculate adstocks and fit using each of them. We
do not do a single regression, we do several. In any case we have to do many regressions to understand the results of using different
sets of explainers and then apply experience and commonsense to each fit. This part of the work is more an art than a mechanistic
procedure.

Here is an example of such a procedure in a real but much simpler case than any quoted here. A very short term response was
expected and a trend term had to be allowed for. Six regressions were run with the following goodness of fit:

HL, weeks: 0 1 2 3 4 5

R-squared: 0.72 0.75 0.71 0.69 0.68 0.67

Using adstock was a little better than using TVRs in the same week; one-week half life was a little better than two weeks; longer half
lives were not recommended.
APPENDIX 2 Model specification.

In what follows the variable to be modelled is [S.sub.i], which may be brand share, sales volume, rate of sale or serious casualty
crashes. [P.sub.t] is a measure of (relative) price, [A.sub.t] is a long term (long half life) adstock variable, t is a time trend and
[O.sub.t] represents other variables (e.g. short term adstocks, consumer promotions, seasonality etc.). Coefficients are e, f, g; c is the
(negative) change in the base each period, or trend. The [Delta] operator defines first differences ([Delta][S.sub.t], = [S.sub.t] -
[S.sub.t-1]) and In defines the natural logarithm.

LINEAR SPECIFICATIONS

The underlying linear specification of the floating base model is:

[S.sub.t] = [B.sub.t] + [fP.sub.t] + [gO.sub.t] +...,

where the base in period t is given by:

[B.sub.t] = [B.sub.t-1] + c + [eA.sub.t].

Then [B.sub.1] = [B.sub.0] + c + [eA.sub.1]

[B.sub.2] = [B.sub.1] + c + [eA.sub.2] = [B.sub.0] +2c + e([A.sub.1] +

[A.sub.2])

and so on until [B.sub.t] = [B.sub.0] + ct + e([A.sub.1] + [A.sub.2] + ...

+ [A.sub.t]).

Two estimable versions of this model are considered. The first, termed LIN, comes from substituting the above in the equation for
[S.sub.t]. On re-arranging, this gives the estimable equation for the linear cumulated long term model:

[S.sub.t] = [B.sub.0] + ct + e[summation]n = [1.sup.t] [A.sub.n] +

[fP.sub.t] + [gO.sub.t] +..., to which we append an (assumed) `well behaved' error term and estimate by OLS.

The second version of the linear model, termed LINDIFF, is obtained by lagging [S.sub.t] and forming the difference, [S.sub.t] -
[S.sub.t-1], which after re-arrangement yields the estimable form for the linear long term model:

[S.sub.t] - [S.sub.t-1] = c + [eA.sub.t] = f ([P.sub.t] - [P.sub.t-1]) +

g([O.sub.t] - [O.sub.t-1]) + ...,

Again a `well behaved' error term is appended and estimation carried out using OLS.

We should note that calculation (estimation) of the base, [B.sub.t], is straightforward in LIN but is slightly more involved in LINDIFF.
This gives us an a priori preference for using LIN.

MULTIPLICATIVE SPECIFICATIONS

The underlying multiplicative specification of the floating base model is:

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] where the base in period t is given by:

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]

Again two estimable versions of this model are considered and these are obtained in an analogous manner to the linear case
discussed above. The first, termed LOG, comes from exploiting the recursion for [B.sub.t]:

1n([B.sub.t]) - 1n([B.sub.t-1]) = 1n(c) + e1n([A.sub.t]),

which yields an estimating form for the multiplicative cumulated long term model:

1n([S.sub.t]) = 1n(B) + 1n(c)t+e[summation]1n([A.sub.t]) + f1n([P.sub.t])

+ g 1n ([O.sub.t]) + ...,

A `well behaved' error term is appended to this estimating equation and estimation proceeds using OLS.

The second version of the multiplicative model, termed LOGDIFF, is obtained by lagging [S.sub.t], taking logarithms and forming the
difference,
1n([S.sub.t]) - 1n([S.sub.t-1]) = 1n([S.sub.t]/[S.sub.t-1]),

which after re-arrangement yields the estimable form for the multiplicative long term model:

1n([S.sub.t]/[S.sub.t-1]) =

1n(c)+e1n([A.sub.t])+f1n([P.sub.t]/[P.sub.t-1])+g1n([O.sub.t]/[O.sub.t-1])+...

Again a `well behaved' error term is appended and estimation carried out using OLS.

We should note that calculation (estimation) of the base, [B.sub.t], is straightforward in LOG but is slightly more involved in LOGDIFF.
This gives us an a priori preference for using the LOG form.

TREATMENT OF ZEROS

Of the variables appearing in the models, it is possible that some of them will have periods for which we observe a zero. This will
occur for the [O.sub.t] variables, such as `dummy variables' relating to particular events, consumer promotions and distribution (or
facings) for new products/sizes. It may also occur for certain adstock variables -- either long or short term. In the linear specifications
the occurrence of these observations will not cause problems for the estimation of the models. However, the interpretation of a
difference variable ([O.sub.t] - [O.sub.t-1],) is difficult when [O.sub.t] is a dummy variable -- indeed it may be argued that such a
difference variable does not make sense.

The situation for the multiplicative specification is more difficult as the logarithms required are not defined in this case. For the
adstock variables the solution used is to add 1 to each observation, i.e. (1 + [A.sub.t]), which when [A.sub.t] = 0 yields 1n(1) = 0 in the
models. This solution seems appropriate when there are only a few zeros in the data set, which is typically the case for the long term
(long half life) variable [A.sub.t]. It is possible that it may not be a good solution for any short term (short half life) variables which
enter the model.

To avoid the problem of undefined logarithms in the case of zero observations in [O.sub.t] variables these should enter the
multiplicative specification for [S.sub.t] as exp([gO.sub.t]). Thus, in the LOG model they will enter in the `level' form as + [gO.sub.t]
and in the LOGDIFF form as differences + g([O.sub.t] - [O.sub.t-1]).

STOCHASTIC SPECIFICATION

In the discussion above it is assumed that it is possible to append a `well behaved' error (stochastic) term to each of the estimating
equations (LIN, LINDIFF, LOG and LOGDIFF). If this is the case then the use of OLS estimation will be appropriate. Indeed for the
estimation results reported in this paper the estimation was carried out using OLS. We now consider the applicability of the `well
behaved' error assumption to each of the four estimating forms and where it may not be applicable suggest how we might see
evidence of this in the estimation results.

For the LIN form the assumption is probably acceptable. In this case the estimating form for [S.sub.t] becomes:

[S.sub.t] = B + ct + [eCA.sub.t] + [fP.sub.t] + [gO.sub.t] +... + [u.sub.t],

which corresponds to a model for [S.sub.t] given by:

[S.sub.t] = [B.sub.t] + [fp.sub.t] + [gO.sub.t] +... + [u.sub.t],

However, the appearance of the cumulative adstock terms in this estimating form is likely to result in positive autocorrelation (low
Durbin-Watson (DW) statistics) if the model is mix-specified in any way. In particular, if the model has an excluded variable which is
highly correlated with the cumulative adstock term.

On the other hand, to use OLS in the LINDIFF form requires the estimating form to be:

[S.sub.t] - [S.sub.t-1] = c + [eA.sub.t] + f([P.sub.t] - [P.sub.t-1]) +

g([O.sub.t] - [O.sub.t-1]) +...+ [u.sub.t.].

This is inconsistent with the model for [S.sub.t] (= [B.sub.t] + [fP.sub.t] + [gO.sub.t] +...+ [u.sub.t]) as this model would yield the
estimating form:

[S.sub.t] - [S.sub.t-1] = c + [eA.sub.t] + f([P.sub.t] - [P.sub.t-1]) +

g([O.sub.t] - [O.sub.t-1])+...+ [e.sub.t].

where [e.sub.t] = [u.sub.t] - [u.sub.t-1], a moving average error. Thus if we estimate LINDIFF by OLS it is possible that the error term
may not tee `well behaved' and our estimation is then invalid. Evidence of this problem will be reflected in bad DW statistics (high
values, say 2[is greater than or equal to] 3).
To use OLS for the LOG form the arguments are similar to those for the LIN form. If we write [S.sub.t] as:

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]

then our estimating form becomes:

1n([S.sub.t]) = 1n(B) + 1n(c)t + e[summation]1n([A.sub.t]) +

f1n([P.sub.t]) + g1n([O.sub.t])+...+ [u.sub.t]..

This model can be estimated by OLS and the assumption that [u.sub.t] is `well behaved' is probably acceptable. However, the
occurrence of the cumulative adstock term in this estimating form is likely to result in positive autocorrelation if the model is mis-
specified in any way.

On the other hand, to use OLS in the LOGDIFF form requires the estimating form to be:

1n([S.sub.t]/[S.sub.t-1]) = 1n(c) + e1n([A.sub.t]) +

f1n([P.sub.t]/[P.sub.t-1]) + g1n([O.sub.t]/[O.sub.t-1])+...[u.sub.t]

This is inconsistent with the model for [S.sub.t] = [B.sub.t][P.sub.t]f[O.sub.t]g exp([u.sub.t]), as this model would yield the estimating
form:

1n([S.sub.t]/[S.sub.t-1]) = 1n(c) + e1n([A.sub.t]) +

f1n([P.sub.t]/[P.sub.t-1]) + g1n([O.sub.t]/[O.sub.t-1])+...+[e.sub.t]

where [e.sub.t] = [u.sub.t] - [u.sub.t-1], a moving average error. Thus if we estimate LOGDIFF by OLS it is possible that the error term
may not tee `well behaved' end our estimation is then invalid. Evidence of this problem will be reflected in bad DW statistics (high
values, say [is greater than or equal to] 3). We should note that the above discussion again potentially gives us a priori reasoning to
prefer the LIN and LOG forms of the floating base model.

It is clear from the discussion above that when modelling with this family of models inspection of the value of the Durbin-Watson
statistic is important. In the DIFF forms it can indicate whether a moving average process is present and in the LIN and LOG forms, it
can indicate whether the model is mis-specified in any way, such as the serial correlation discussed above in the example, which
does not affect estimator bias. This should not imply that the analyst should ignore the other diagnostics commonly produced by
econometric/statistical packages. Indeed since the desirable properties of the OLS estimators (e.g. unbiasedness, consistency)
depend upon the underlying assumption of a `well behaved' error term, analysts should be encouraged to subject their models to a
range of diagnostic tests (e.g. heteroscedasticity tests, CUSUM tests).

JMRS `FULL' SPECIFICATION

In the original Broadbent (1990) paper the `full' floating base model specification was given by:

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]

[B.sub.t] = [cB.sub.t-1] (1+[A.sub.t]).

This may be viewed as a hybrid of the linear and multiplicative specifications considered here. To obtain estimating forms we proceed
as in the multiplicative case and find:

FULL

1n([S.sub.t]) = 1n(B) + 1n(c)t + [summation]1n(1 + [eA.sub.t]) +

f1n([P.sub.t]) + g 1n([O.sub.t])+...+ [u.sub.t]...

FULLDIFF

1n([S.sub.t]/[S.sub.t-1]) = 1n(c) + 1n(1 + [eA.sub.t]) + f

1n([P.sub.t]/[P.sub.t-1]) + g 1n([O.sub.t]/[O.sub.t-1])+...+ [u.sub.t].

Again it seems plausible to argue that the error term in FULL is `well behaved', but that the error term in FULLDIFF may not be `well
behaved' (it is possibly of moving average form). Also the recovery of estimates of the base, [B.sub.t] will be more difficult in
FULLDIFF. Thus, attention should be centred upon the FULL form.

Unfortunately, since FULL is not linear in parameters, it cannot be estimated using OLS. It should be possible to estimate the model
in a purpose written computer program using non-linear least squares (e.g. using a modified Marquardt algorithm), but not in standard
packages such as those we used (Limdep and Shazam). It should be noted that the same is true of the FULLDIFF form.
Given that our family of models is likely to cover the majority of cases that the analyst might encounter we feel that the additional
complexity, both of specification and of estimation, of the `full' specification should not be used unless there are compelling reasons
to support its use. That is the LIN and LOG forms discussed in this paper should be sufficient for most (if not all) cases that we will
encounter in practice. Furthermore, both the LIN and LOG forms are linear regression models, can be estimated by OLS, and are
relatively straightforward to interpret. The choice of one over another should be done on a priori reasoning. For example, we may
prefer the multiplicative (and hence the LOG form) if we wish to use a constant elasticity model.

APPENDIX 3 The case histories

In this section we describe very briefly the market-place circumstances of our eleven examples. We show the variables used in the
fits, with their t values (a one-sided hypothesis is appropriate) and elasticities, as well as the number of observations and R-squareds.

The seven packaged goods or commercial examples covered a variety of sales movements (see figure 1), from the sharp growth of
Brand D, still in launch, to the decline of Brand C. They include the stability of Brand K and the growth followed by collapse of Brand
A. The average shares were as small as 2.6% (Brand J) and as large as 82% (Brand C), with shares like 5%, 13% and 46% in
between. There were between two and nine years of data; the period lengths were 4-weeks or a month.

The commercial examples all had positive elasticities for advertising. The purpose of Road Safety advertisements is to reduce
accidents, and here elasticities were negative.

The examples fall into three groups within which the order is by sample size: Floating base, commercial -- four examples Floating
base, Road Safety -- four examples Long term adstock, commercial -- three examples

In the commercial examples there was little difference between the additive and multiplicative or log fits and neither was consistently
better. Nor is there a prior case for the interactivity implied by the multiplicative model. The additive model is easier to explain to non-
technical people and is the one used here. In the Road Safety examples the multiplicative method gave better fits in all cases and is
the one used. Logs of all the variables were therefore taken, but this is shown below only for the adstock terms.

Case A For the data used in this example we are indebted to Louise Cook of BMP DDP Needham. The product is a brand leader for
which there is a long run of data. It has recently been attacked by very low price discounters so we first looked at its share in the
`normal price' sector, in which we get a good fit to sales. However the more interesting and convincing fit is in the total category. Its
steady growth for the first six years is followed by decline as the price cutters took their toll. At the cusp, the economy of the country
suffered from no-growth: the brand does better, like many brand leaders, when times are good, shown here by the total sales of all
commodities in the research source. The creative work altered relatively little over this time. In recent years its advertising weight has
declined as the manufacturer did not realise sufficiently what benefit he was really getting: the model shown here was then not
available. Modelling has been remarkably efficient in sorting out these simultaneous factors: more detail was given in the technical
explanation above. Two of the promotional techniques (called here C1 and C2) were particularly successful.

n = 119 [summation][ADs.sub.52] t = 6.6 elas = 1.02 R-sq = 0.87 Trend -5.6 decline p.a. -15%

Relative price -4.7 elas = -0.90

All commodity volume 6.7 0.70

C1 2.4

C2 3.1

Case B This brand was in decline shortly after launch, with reducing advertising support. Adstocks worked in two ways: by
counteracting decline and with short term blips. Retailers however influenced sales considerably through the percentage of brand
facings or stocks held `forward'. C1 and C2 are again two sorts of price promotion, by added value packs.

n = 54 [summation][ADs.sub.52] t = 2.6 elas = 0.57 R-sq = 0.91 [ADs.sub.6] 3.1 0.19

Trend -2.8 decline p.a. -36%

Forward stocks 11.5 elas = 0.99

Seasonality 2.0 0.52

C1 4.9 0.037

C2 1.4 0.003

Case C This brand is also category leader and is faced with low-priced competition. It had two bursts of advertising during the period
studied, which had both short and long term effects. Nevertheless there is a natural erosion heightened when competitive prices are
very low.

n = 32 [summation][ADs.sub.26] t = 4.0 elas = 0.40 R-sq = 0.86 [ADs.sub.10] 4.2 0.04

Trend -4.6 decline p.a. -14%

Relative price 8.9 elas = -1.7

Case D This was a launch of a high-priced quality product. It was supported by a variety of promotions, most of which contributed to
sales in proportion to the price reduction offered. In two cases (called again C1 and C2) there were benefits above the price
reduction. Short term adstock was a significant explainer but the main contributor was cumulated adstock. At this stage in the brand's
growth, trend was not significant. [summation][LADs.sub.t] correlated highly with distribution, but since rate of sale could also be
correlated with adstock we have used only one of these explainers.

n = 38 [summation][ADs.sub.52] t = 33.2 elas = 1.00 R-sq = 0.98 [ADs.sub.10] 9.1 0.15

RelPr -11.9 -1.6

C1 -5.8 -0.02

C2 -3.4 -0.01

THE ROAD SAFETY EXAMPLES

These examples concern the analysis of road safety measures, including advertising, in the state of Victoria, Australia. Two
advertising campaigns are analysed. One is a Drink Drive campaign and the other a Speed Kills campaign. The campaigns are
designed to reduce accidents involving drink drivers and speeding drivers respectively. The response variable modelled is the
number of serious casualty crashes. That is the number of crashes involving either a fatality or injury requiring admission to
hospital.132 monthly observations for the period January 1983 to December 1993 are used to estimate the models.

To match the response variable and the road safety campaigns we partition the data set by region of the state and by time of day.
That is, the data on serious casualty crashes (SCCs) are partitioned into those occurring in the Melbourne statistical division and
those occurring in the rest of Victoria. A further partition is into those SCCS occurring in 'High Alcohol Hours' and those occurring in
`Low Alcohol Hours'. The `Low Alcohol Hours' are Monday-Thursday 6am to 6pm, Friday 6am to 4pm, Saturday 8am to 2pm and
Sunday 10am to 4pm and are periods in which the percentage of drivers killed or admitted to hospital with a blood alcohol content
exceeding the legal limit of 0.05% is below 4%. The `High Alcohol Hours' are the converse of these times and are periods in which
38% of drivers killed or admitted to hospital have a blood alcohol content above 0.05%.

This partitioning leads to four data sets for analysis: high alcohol hours in Melbourne, high alcohol hours in the rest of Victoria, low
alcohol hours in Melbourne and low alcohol hours in the rest of Victoria. The advantages of this partitioning process are that it allows
for the identification of regional differences in results (e.g. differing impacts of the advertising) and for a closer matching of the road
safety advertising campaigns with the response variable. That is, the Drink Drive campaign is aimed at reducing serious casualty
crashes which occur predominately in high alcohol hours and the Speed Kills campaign is aimed at reducing serious casualty crashes
which occur predominately in low alcohol hours.

E: Road safety: High alcohol hours, Melbourne

n = 120 [summation][LnAds.sub.52] t = -2.9 elas = -0.056 R-sq = 0.83 Trend 4.5 0.24

[LnAds.sub.5] -4.7 -0.036

Random breath tests -2.9 -0.012

Seasonality 7.1 0.12

Unemployment -2.0 -0.16

Alcohol sales 2.5 0.20

F: Road Safety: High alcohol hours, Rest of Victoria

n = 120 [summation][LnAds.sub.52] t = -5.2 elas = -0.063 R-sq = 0.66 Trend 3.0 0.13

[LnAds.sub.5] -2.9 -0.025

Seasonality 6.0 0.87

Alcohol sales 2.1 0.27

G: Road Safety: Low alcohol hours, Melbourne

n = 120 [summation][LnAds.sub.52] t = -2.4 elas = -0.019 R-sq = 0.78 Trend 8.3 0.26

[LnAds.sub.5] -5.8 -0.045

Random breath tests -2.9 -0.012

Seasonality 9.0 0.98

H: Road Safety: Low alcohol hours, Rest of Victoria

n = 120 [summation][LnAds.sub.52] t = -4.3 elas = -0.038 R-sq = 0.68 Trend 4.7 0.17

[LnAds.sub.5] -3.4 -0.03

Seasonality 10.3 1.02

Speed cameras -1.5 -0.03

Case I Various models of sales share were constructed, with very different results depending on the sector or category used for
comparisons. For example, price had little effect when the brand is compared with the category, but a strong effect when sector
comparisons are made. This is because the sector has recently been growing with the addition of cheaper competitors, but these
hardly affect category averages.

Short term adstocks were correlated with volume and category sales share, but may have helped competitors so are not significant in
sector share.

The sector model is reported because of its higher it-squared: in the category this reached 0.90. `Facings' or the percentage of
facings in the category on the shelf, show the support which the retailers give or withhold, critical in this category.

n =91 [Ads.sub.104] t = 1.6 elas = 0.06 R-sq = 0.97 RelPr -10.5 -3.4

Trend -2.9 decline p.a. -15%

Our Facings 1.7 0.10

Comp Facings -6.1 0.07

OL Facings -3.2 0.04

Case J A relaunch in the first year of our data produced growth. This was followed by stability then some decline. However the last
seven periods showed a small step up. Advertising supports a higher price than example A, but this time with clear long term and
short term components.

n = 55 [Ads.sub.52] t = 7.8 elas = 0.10 R-sq = 0.74 [Ads.sub.6] 4.0 0.04

RelPr -5.1 -0.73

Step 4.8 0.01

Case K This brand has been steady for years in a stable category. The advertising objective is to sustain a profitable brand at a price
premium. The price is indeed high and the elasticity is satisfactorily low. Introduction of a variant has increased overall facings and
been of some benefit to the parent despite cannibalisation. Cumulated adstock and trend are both non-significant. The only adstock
term entering the fit has half life midway between short and long. R-squared here is much lower than in other examples, which is a
result of the stability of the general sales level -- there is simply less overall variation to explain: the t values are satisfactory.

n = 55 [Ads.sub.26] t = 3.2 elas = 0.22 R-sq = 0.32 RelPr -4.3 -0.59

Variant Dis 2.9 0.29

References

BAKER, C. (1990). The longer and broader effects of advertising. Institute of Practitioners in Advertising.

BARNARD, N. & SMITH, G. (1989). Advertising and modelling. Institute of Practitioners in Advertising. A general description of
adstock modelling, though with a slightly different definition.

BROADBENT, S. (1990). Modelling beyond the blip. Journal of the Market Research Society, 32, 1, pp 61-102

BROADBENT, S. (1992). Using data better. Admap, January, pp 48-54.

BROADBENT, S. (1993). Advertising effects: more than short term. Journal of the Market Research Society, 35, 1, pp 37-49.

INSTITUTE OF PRACTITIONERS IN ADVERTISING (1981-1995) Advertising Works series.

LODISH, L. M. & LUBETKIN, B. (1992). General truths? Nine key findings from IRI test data. Admap, February, pp 9-15.

(*) This has been reported several times since the Dettol paper in Advertising Works (1981). Winalot Prime in Advertising Works
(1988) is an example. A longer half life is claimed as an advantage for PG Tips over Tetley, Advertising Works (1991), and for Andrex
compared with average, Advertising Works (1993). Examples of major sales effects longer than a year after the campaign are shown
in US experiments reported privately by Abraham, M. and Lodish, L. M. (1989) for IRI, confusingly also under the title Advertising
Works. See Lodish, L. M. and Lubetkin, B. (1992).

----------

Please note: Illustration(s) are not available due to copyright restrictions.

Copyright: COPYRIGHT 1995 Sage Publications, Inc.

http://www.sagepub.com.proxy-ub.rug.nl
Source Citation (MLA 8th Edition)
Broadbent, Simon, and Tim Fry. "Adstock modelling for the long term." Journal of the Market Research Society, vol. 37, no. 4, Oct.
1995, p. 385+. Gale General OneFile, https://link-gale-com.proxy-
ub.rug.nl/apps/doc/A17852279/ITOF?u=groning&sid=ITOF&xid=348bb798. Accessed 7 Apr. 2020.
Gale Document Number: GALE|A17852279