On the econometrics of the Koyck model∗

Philip Hans Franses and Rutger van Oest

Econometric Institute, Erasmus University Rotterdam
P.O. Box 1738, NL-3000 DR, Rotterdam, The Netherlands

Econometric Institute Report 2004-07

Abstract
The geometric distributed lag model, after application of the so-called Koyck
transformation, is often used to establish the dynamic link between sales and
advertising. This year, the Koyck model celebrates its 50th anniversary.
In this paper we focus on the econometrics of this popular model, and we
show that this seemingly simple model is a little more complicated than we
always tend to think. First, the Koyck transformation entails a parameter
restriction, which should not be overlooked for efficiency reasons. Second, the
t-statistic for the parameter for direct advertising effects has a non-standard
distribution. We provide solutions to these two issues.
For the monthly Lydia Pinkham data, it is shown that various practical
decisions lead to very different conclusions.

Key words and phrases: Koyck model, sales-advertising relationship

∗
We thank Dennis Fok, Richard Paap and Marco Vriens for helpful discussions, and Lotje
Kruithof for excellent research assistance. The address for correspondence is Econometric Institute,
Erasmus University Rotterdam, P.O. Box 1738, NL-3000 DR Rotterdam, email: [email protected]
or [email protected]. The computer programs used in this paper are available from the second
author.

1
1 Introduction
The geometric distributed lag model is often used to investigate the current and
carryover effect of advertising on sales. This model makes current sales a function of
current and past advertising levels, where the lag coefficients have a geometrically
decaying pattern. As this model involves an infinite number of lagged variables, one
often considers the so-called Koyck transformation (Koyck, 1954). In many studies
the resultant model is hence called the Koyck model. Leendert Marinus Koyck (1918-
1962) was a Dutch economist who studied and worked at the Netherlands School of
Economics, which is now called the Erasmus University Rotterdam.
In this research note we will discuss the basic Koyck model, and illustrate that
this model is less straightforward to analyze than is usually assumed or suggested.
We will provide a discussion of possible solutions. Next, we will show for the well-
known monthly Lydia Pinkham data that various approaches lead to different con-
clusions, thereby emphasizing the relevance of the proper methods.

2 The Koyck model

Consider the variables sales St and advertising At , and assume that the link between
these two variables is given by

St = µ + β(At + λAt−1 + λ2 At−2 + ...) + εt , (1)

where εt is an uncorrelated error variable. The parameter λ is usually called the

retention rate. The current effect of advertising is β, whereas the carryover effect is
β
equal to 1−λ
. To provide some intuition for the discussion below, note that when
β = 0, the parameter λ disappears from the model, and cannot be retrieved. From a
marketing perspective this is no problem as the model assumes that the current and
carryover effects of advertising are both a function of β. However, from a statistical
perspective, matters become a little complicated.
As this model contains a very large number of variables, it is common practice
to apply the Koyck transformation. This entails subtracting λSt−1 from (1), to get

St = µ + βAt + λSt−1 + εt − λεt−1 . (2)

1
In time series jargon, this model is called an ARMAX model, see Franses (1991)
for more details on ARMAX models. The autoregressive [AR] part concerns St−1 ,
the moving average part [MA] concerns εt−1 and the explanatory variables part [X]
concerns At . Note that the parameter λ appears twice, and hence that, except for
the intercept, there are only two parameters to estimate, while the model effectively
contains three explanatory variables. Additionally, note that when β = 0, the model
contains the lag polynomial 1 − λL, with L the lag operator, on both sides, which
gets cancelled, that is, when β = 0, the model reduces to St = µ + εt .

Parameter estimation

There are several approaches that one can follow to estimate the parameters in the
resulting Koyck model in (2). Of course, the appropriate estimation method here
is the maximum likelihood method, which imposes that the AR and MA parameter
are the same. The (conditional) log-likelihood function is given by

T −1³ ´ X T
2 2 ε2t
ln L(µ, β, λ, σ ) = − ln(2π) + ln(σ ) − 2
, (3)
2 t=2
2σ

where T denotes the number of observations, and the {εt } are recursively defined as

ε1 = 0,

εt = St − µ − βAt − λSt−1 + λεt−1 , t = 2, . . . , T. (4)

This approach is also described in Hamilton (1994, p. 132) for general ARMA
models. Asymptotic standard errors are obtained by taking the square roots of the
diagonal elements of the estimated covariance matrix, which in turn can be computed
as minus the inverse of the Hessian of (3) evaluated for the optimal parameter
values. Numerical techniques, such as the BFGS algorithm or the Newton-Raphson
algorithm, have to be used to get the maximum likelihood parameter estimates.
It is tempting though to decide to all the way neglect the MA part, so that the
model parameters can be estimated using the method of ordinary least squares. This
is obviously not a sound approach, as εt−1 and St−1 are not uncorrelated, and thereby
one of the basic premises of regression theory gets violated. Simulation results in
Table 1 suggest that the resultant downward bias of λ̂ can be substantial.

2
 One can also decide to estimate the parameters in an unrestricted ARMAX
model, that is,
St = µ + βAt + λ1 St−1 + εt − λ2 εt−1 . (5)

It is possible to get estimators for λ1 and for λ2 which are consistent estimators
for λ. In practice, however, it is most likely that the corresponding estimates take
different values, and the question appears which one should take. Also, (5) can-
not be transformed back to a model like (1), and hence is less interesting from a
theoretical perspective. Moreover, correctly imposing that λ1 = λ2 yields a more
efficient estimator. The simulation results in Table 1 suggest that estimating λ1 in
an unrestricted ARMAX model leads to almost no bias.

Testing advertising effects

A second issue of concern for the Koyck model is a test for the significance of the
advertising effects. Indeed, one may want to examine whether β is equal to zero. This
is not trivial as under the null hypothesis of interest, that is, β = 0, the parameter
λ disappears from the model, see (1) and (2). This is what is known as the Davies
(1977) problem, and it seriously complicates statistical analysis. The issue is that
the usually considered t-statistic depends on λ for which it is not clear which value to
take. An appealing approach might seem to simply set λ at its maximum likelihood
value. However, this would make the test (and its critical values) dependent on the
data, and the asymptotic distribution would not be standard normal.
Recent solutions to the Davies problem are provided by Andrews and Ploberger
(1994) and Hansen (1996), see also Carrasco (2002). The main idea is that one
constructs a new test statistic based on the entire distribution of the original test
statistic over a range of values of the unidentified parameter λ. In the Koyck model,
a sensible range for λ would be the interval [ 0, 1). One possibility, involving the
entire distribution over λ, would be to consider the class of “sup test statistics”,
which corresponds to the highest value of the original test statistic within the range
for λ. This approach is advocated by Davies (1977), see also Hansen (1996) and
Carrasco (2002). Alternatively, one can consider the class of “ave test statistics”,
based on the average value of the original test statistic. This approach is put forward

3
by Andrews and Ploberger (1994), and is further investigated by Hansen (1996). In
each case, the asymptotic distribution of the resulting test statistic is not standard
normal, so that its distribution has to be simulated.
In this paper we consider the “ave” and “sup” versions of the absolute t statistic
|tβ | and the Wald statistic t2β , where tβ is obtained by taking the ratio of the maxi-
mum likelihood estimate of β and its estimated asymptotic standard error. So, we
focus on four test statistics, that is, ave absolute t, ave Wald, sup absolute t, and
sup Wald. Although the two sup tests are equivalent, we include them both in order
to achieve symmetry. Table 2 contains the simulated (asymptotic) critical values for
the four tests at confidence levels of 80%, 90%, 95%, and 99%. In order to obtain
these critical values, we ran 40000 simulations for T = 1000 observations. In each
simulation, advertising data were drawn from a standard normal distribution. Next,
sales data were generated from the Koyck model under the null hypothesis β = 0,
that is, we set β = 0, µ = 0, and we assumed variance 0.25, like in Table 1. For
each simulation, the four test statistics were computed, and their values were stored.
Finally, the four resulting samples were ordered in an ascending way, so that the
quantiles became available. In each simulation, the sup absolute t and sup Wald
statistics were obtained via a grid search over λ from 0 to 0.999 with step size 0.001.
Before we turn to an application of these tests to real-life data, it seems wise to
see which of these tests performs best in practice. For that purpose, we ran another
set of simulations, and the results are given in Table 3. Clearly, the power of the
supremum tests is not very high, and hence we would recommend the use of the
average tests.

3 An illustration
One might now be tempted to think (or hope) that the above considerations would
not have a substantial impact on empirical analysis. Unfortunately, they do, as can
already be illustrated for the illustrious monthly Lydia Pinkham data, which have
been used in many marketing studies.

4
 If one decides to neglect the MA part of the Koyck model, that is,

St = µ + βAt + λSt−1 + vt , (6)

and replaces the intercept by twelve monthly dummies to correct for seasonal effects,
then β gets estimated at a value of 0.360, with standard error 0.118, and λ is
estimated equal to 0.370, with standard error 0.125.
Unrestricted estimation of the Koyck model, that is, (5) with monthly dummies,
renders a β of 0.335 (0.101), a λ1 of 0.690 with standard error 0.119, and a λ2 of
0.561 with error 0.178. Clearly, these two λ parameters are rather different. Notice
that the t−ratios for β in the above two models are 3.039 and 3.307, respectively.
The maximum likelihood estimates of β and λ for (2), again including monthly
dummies, are found to be 0.339 (0.098) and 0.703 (0.125) respectively, where the
reported standard errors are asymptotic standard errors. Consistent with the simu-
lation evidence from Table 1, the estimates of λ1 in the unrestricted model and λ in
the restricted Koyck model are approximately equal (0.690 and 0.703). Furthermore,
it can be seen that ignoring the MA part of the Koyck model indeed results in serious
underestimation of the retention rate (0.370 versus 0.703), as suggested by Table 1.
To put it differently, neglecting the MA component would result in a 90% duration
interval of 1.3 months, whereas appropriate maximum likelihood estimation would
result in a much longer 90% duration interval of 5.5 months1 .
The values of the four (average and supremum) statistics for testing β = 0 are
4.61, 22.06, 5.55 and 30.79, respectively. By comparing these realized values with the
critical values in Table 2, we conclude that the two “ave tests” indicate a significant
advertising effect at a 1% significance level, whereas the two “sup tests” fail to reject
the null hypothesis at a 20% level. These contradictory results confirm our findings
in Table 3 which suggests that the supremum tests do not have much power.
Finally, for illustrative purposes, Figure 1 shows the underlying distributions of
the absolute t statistic |tβ | and the Wald statistic t2β over the different values of λ.

1
Clarke (1976, p346) defines the (100 × α)% duration interval as the time period τα during
which (100 × α)% of the expected cumulative advertising effect has taken place. It can be shown
that τα = ln(1−α)
ln(λ) − 1.

5
For both statistics, the supremum corresponds to λ = 0.663, which is quite close to
the maximum likelihood estimate λ̂ = 0.703.

4 Conclusion
The Koyck model is often applied in marketing practice, but it is more complicated
to analyze than one would perhaps think. Proper parameter estimation requires
imposing parameter restrictions in the estimation routine. And, proper inference
on the advertising effects requires new test statistics with non-standard asymptotic
distributions, as the retention parameter disappears under the null hypothesis of
no effect of advertising. In this paper we showed how these tools work. For the
monthly Lydia Pinkham data we showed that various tests can lead to contrasting
conclusions.

6
Table 1: Estimating the retention parameter λ in a Koyck
model. In case (A) the MA term is neglected. In case (B) the
model parameters are estimated using nonlinear least squares
with unrestricted parameters λ1 and λ2 , where only λ1 is re-
ported. The simulation results are based on 1000 replications.
The cells contain the mean value of the 1000 estimates and
the associated standard deviation. Advertising data are drawn
from a standard normal distribution. Next, sales data are gen-
erated for µ = 0, β = 1 and an error process having variance
0.25.

n = 50 n = 500 n = 5000

λ = 0.5

A 0.412 (0.062) 0.420 (0.018) 0.420 (0.012)

B 0.502 (0.053) 0.500 (0.014) 0.500 (0.010)

λ = 0.8

A 0.705 (0.053) 0.732 (0.013) 0.734 (0.004)

B 0.797 (0.025) 0.800 (0.005) 0.800 (0.002)

λ = 0.9

A 0.825 (0.046) 0.855 (0.010) 0.859 (0.003)

B 0.899 (0.014) 0.900 (0.002) 0.900 (0.001)

λ = 0.95

A 0.879 (0.051) 0.923 (0.008) 0.927 (0.002)

B 0.948 (0.012) 0.950 (0.008) 0.950 (0.000)

7
Table 2: Critical values of various tests for the hypothesis that β = 0 in the Koyck
model. Number of replications is 40000. The sample size is 1000. The grid for λ runs
from 0.000 to 0.999 with step size 0.001. Data are generated similar to those in Table
1.

Confidence level Ave absolute t Ave Wald Sup absolute t Sup Wald

80 % 1.21 1.67 6.18 38.13

90 % 1.53 2.56 7.47 55.86
95 % 1.80 3.50 8.41 70.65
99 % 2.34 5.79 9.80 95.99

8
Table 3: Empirical power of various tests for the hypothesis that β = 0
in the Koyck model. Number of replications is 1000 for each value of β.
The sample size is 1000. The critical value is set at the 5% level. Data are
generated similar to those in Table 1.

β Ave absolute t Ave Wald Sup absolute t Sup Wald

0.00 5.90 5.70 5.20 5.20

0.02 29.80 29.40 5.30 5.30
0.04 79.30 78.70 4.50 4.50
0.06 98.20 98.20 5.30 5.30
0.08 100.00 100.00 7.40 7.40
0.10 100.00 100.00 16.80 16.80

9
 absolute t × lambda
5

0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

30 Wald × lambda

20

10

0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

Figure 1: Testing for the significance of advertising effects in the Koyck model
applied to the monthly Lydia Pinkham data. The values of the absolute t statistic
and the Wald statistic are shown for different values of the retention parameter λ.

10
References
Andrews, D.W.K. and W. Ploberger (1994), Optimal tests when a nuisance param-
eter is present only under the alternative, Econometrica, 62, 1383-1414.

Carrasco, M. (2002), Misspecified structural change, threshold, and Markov-switching

Models, Journal of Econometrics, 109, 239 - 273.

Clarke D.G. (1976), Econometric measurement of the duration of advertising ef-

fect on sales, Journal of Marketing Research, 8 (November), 345 - 357.

Davies, R.B. (1987), Hypothesis testing when a nuisance parameter is present only
under the alternative, Biometrika, 64, 247-254.

Franses, P.H. (1991), Primary demand for beer in The Netherlands: An applica-
tion of ARMAX model specification, Journal of Marketing Research, 28, 240-245.

Hamilton, J.D. (1994), Time Series Analysis, Princeton University Press.

Hansen, B.E. (1996), Inference when a nuisance parameter is not identified under
the null hypothesis, Econometrica, 64, 413-430.

Koyck, L.M. (1954), Distributed Lags and Investment Analysis, Amsterdam: North-
Holland.

11