6 ESTIMATION OF DEMAND

FUNCTIONS
CHAPTER OUTLINE

Chapter objectives
Introduction
Estimating demand functions
Interviews and survey methods
Questionnaires
Consumer experiments
Market studies
Statistical analysis
Pitfalls using regression analysis
The economic veri¢cation of regression models
Statistical veri¢cation of regression models
Econometric veri¢cation of regression estimates
Case Study 6.1 The demand for alcoholic drink
Chapter summary
Review questions
References and further reading

CHAPTER OBJECTIVES

This chapter aims to explore the empirical estimation of demand functions.

At the end of the chapter you should have an understanding of the main
empirical techniques to gain information about consumer behaviour and
should be able to:
t Describe the main methods of data collection.
t Discuss the advantages and disadvantages of using survey and question-
naire techniques.
t Understand the nature of regression analysis.
t Distinguish between rime series and cross-sectional data.
t Interpret the statistical coe⁄cients and explain their economic
signi¢cance.
t Understand the main statistical tests used to verify the signi¢cance of
estimated regression.
t Explain the advantages and disadvantages of regression analysis.
104 PART II g KNOWING THE MARKET

INTRODUCTION

Knowledge of the market demand function and of the key factors in£uencing future
changes in demand is important for the management of the ¢rm, not only for setting
prices but also for planning production capacity and the choice of goods or services to
produce. The purpose of this chapter is to discuss aspects of the empirical estimation of
demand functions including:

g Methods of data collection.

g Interview and survey analysis.
g Regression methods for analysing the data and obtaining demand functions.
g Statistical veri¢cation of resulting functions.
g Case study of alcoholic drinks.

ESTIMATING DEMAND FUNCTIONS

The task of estimating a formal statistical demand function for a single product is su⁄-
ciently arduous and costly for few ¢rms to be willing to devote the necessary resources
to the task when demands for more than a few items must be estimated. Many ¢rms
rely on traditional behavioural rules of thumb to gain some insights into the shape of
their demand curves. Such rules are based on past experience, data collection and
experiments and generally work well when the conditions of demand are relatively
constant. Managers may make educated guesses based on a summing up of the
situation when compared with experience of similar situations in the past or they can
engage in more formal statistical methods. The hunch, or educated guess, method is
what managers do most of the time. From this process an implicit demand curve or
function is postulated and a guess is made of the likely quantity demanded for a
narrow range of prices. Such an informal approach may be the only feasible method
for many ¢rms for both existing and new products. The latter present a particular
problem because the hunch has to be made without any current or historical data
being available.
Many ¢rms, however, prefer to be better informed about the nature of their demand
function, so that they can answer the ‘‘what if ’’ questions that many businessmen
ponder. The usual ‘‘what if ’’ questions relate to the consequences of altering one of
the variables thought to be important in in£uencing demand: the response of
consumers to changes in prices, advertising or in the case of consumer durables credit
terms. To do this, more formal statistical modelling of the demand function is required.
This process requires choices to be made about:

g The key variables to be included in the demand function.

g The likely mathematical relationship between dependent and independent
variables.
g The method of collecting information for each of the independent variables.
 CHAPTER 6 g ESTIMATION OF DEMAND FUNCTIONS 105

This chapter will now examine, in a non-technical manner, survey and statistical
methods to estimate demand functions.

INTERVIEWS AND SURVEY METHODS

The most obvious way to try to identify the relationships important in a demand
function is simply to ask actual or potential buyers. Thus, you could ask a group of
buyers how they might react to price changes, product re-speci¢cation and cheaper
credit. Collating the results of the study should then give some indication to the ¢rm of
the likely consequences of changing one or more of the key variables. This kind of
information can be collected by:

g Selecting a sample of existing buyers and asking each person a series of questions.
g Selecting a random sample of people and asking each person a series of questions.
g Gathering together a group of buyers (nowadays known as a focus group) for
discussion and questioning.

In all of these approaches the sample of people to be asked is important. On some

occasions it may be appropriate to have a random sample of the population as a
whole, on others a random sample of existing buyers might be appropriate. Yet again,
it may be that only a subgroup of buyers is required: for example, a ¢rm may be
interested in the leisure drinking habits of 18 to 25-year-olds, while another might be
interested in the holidaying preferences of the over-65s.
The use of questionnaires and interviews is a common procedure administered on
behalf of ¢rms by specialist market research companies. In many high streets on any
day of the week shoppers and passers-by are asked for their responses to a given set of
questions. The results are then used to o¡er information and advice to managers to
enable them to make more informed decisions. However, for whatever purpose a
survey is used, its validity is always questioned on a number of grounds.

Shortcomings

1 The ¢rst relates to the group of people questioned and whether they were
appropriate for the purpose. Clearly, those participating should represent the
target group as a whole. If they do not, then the sample is biased and the results
may not be meaningful.
2 A second problem relates to the response rate. A questionnaire sent to a randomly
selected group of buyers may not be so random when the returns are received. A
low response rate may mean that the data collected are not representative of the
group as a whole: for example, a postal survey may bring responses from people
who either have the time for or enjoy ¢lling in questionnaires.
3 The third relates to the answers given by respondents. At the time of the question-
naire the respondent may or may not tell the truth. Even if they think they might
106 PART II g KNOWING THE MARKET

respond to a price cut at the time of the survey, they may not do so at the time of the
actual price change. In a similar way, consumers asked to classify themselves into
income or social groups may either overestimate or underestimate their actual
income or social class.
4 A fourth relates to face-to-face interview, as the answers may be in£uenced by the
interviewer. The attitude and personality of the interviewer may in£uence the
answers of respondents who may be unwilling to give answers which may be
truthful but which they perceive the interviewer does not wish to receive or which
might make them feel uncomfortable.
5 A ¢fth problem may relate to the questions asked. If the questions are not simple
and precise they may be open to misunderstanding and misinterpretation by
respondents.
6 A sixth problem relates to respondents who may be asked about aspects of a product
or market that they do not have su⁄cient knowledge to be able to answer.

Baumol (1965, p. 212) describes interviews as a, ‘‘dangerous and unreliable procedure.

People just have not thought out in advance what they would do in these hypothetical
situations, and their snap judgements thrown up at the request of an interviewer
cannot inspire a great deal of con¢dence.’’

QUESTIONNAIRES

Much work has been done by statisticians and practitioners to overcome many of these
problems and to make surveys and interviews an e⁄cient method of collecting
information. Questionnaires must be constructed carefully to encourage respondents
to give truthful answers and for answers to be checked one against another to ensure
consistency. Problems may arise with words having multiple meanings, with
questions that can be misinterpreted, with multiple answers that do not allow the
respondent to re£ect fully their opinions or preferences and with the order of questions
which may guide the respondent to particular answers.
The derivation of a demand curve using hypothetical data is illustrated in Table
6.1. Assume that 1,000 people are asked whether they would purchase a new product
at a variety of prices: each is asked to assess their willingness using a 5-point scale
ranging from 1, ‘‘no’’, to 5, ‘‘de¢nitely yes’’. The 5-point scale is as follows:

1 100% chance of saying no, 0% chance of saying yes.

2 75% chance of saying no, 25% chance of saying yes.
3 50% chance of saying no, 50% chance of saying yes.
4 25% chance of saying no, 75% chance of saying yes.
5 100% chance of saying yes.

From the data in Table 6.1 we can ¢nd the anticipated quantity demanded at each price
 CHAPTER 6 g ESTIMATION OF DEMAND FUNCTIONS 107

Table 6.1 Responses to survey question

Price Chance of a number of people buying the product Quantity demanded Percentage of
(») sample buying
0% 25% 50% 75% 100%

10 425 225 175 125 50 287.5 28.75

9 375 175 200 125 125 362.5 36.25
8 250 150 250 150 200 475 47.5
7 150 100 250 225 275 593.75 59.37
6 75 25 275 275 350 700 70.0
5 25 0 200 325 450 793.75 79.37

Source Author

level. At a price of »10, for example, the quantity demanded (DQ ) is the sum of the
anticipated volume of sales to each group of respondents given their responses, or:
DQ ¼ ð425
0:0Þ þ 225
0:25 þ 175
0:5 þ 125
0:75 þ 50
1:0 ¼ 287:5
The anticipated values for the other prices can be calculated in a similar way. The
results show that demand increases as the price falls. This information can be plotted
in a price quantity diagram to form an anticipated demand curve for the samples
shown in Figure 6.1. If the sample is random and typical of a larger group of
customers, a demand curve for the larger group can be inferred. The information
could also be plotted as a buyer response curve with the proportion buying on one axis
and price on the other, as shown in Figure 6.2. The six data points shown in Figure

12

10

8
Price

0
0 100 200 300 400 500 600 700 800 900
Anticipated quantity demanded
Figure 6.1 Anticipated demand curves
108 PART II g KNOWING THE MARKET

90

80

70

Proportion buying
60

50

40

30

20

10

0
0 2 4 6 8 10 12
Price
Figure 6.2 Buyer response curve

6.1 indicate an inverse relationship between price and quantity demanded. A simple
regression line can be estimated between quantity demanded and price. The result
obtained is as follows:
Q ¼ 1315:7  103.9P R 2 ¼ 0:995
ð31:5Þ

where the t statistic is in parentheses. This linear regression would give a quantity
intercept of 1,315.7 and a price intercept of 12.6. The estimation of linear regression
relationships is discussed later in this chapter. The equation can be used to predict
demand at any price. For example, at a price of »5 the predicted quantity demanded is
796 compared with the survey estimate of 783.75. For a price of »3, not included in
the original survey, the model predicts that the quantity demanded would be 1,004.

CONSUMER EXPERIMENTS

Another way of trying to discover the response of consumers to changes in price,

advertising or product speci¢cation is to invite a group of people to a consumer clinic,
or laboratory, and simulate situations in which their behaviour can be observed.
Di¡erent groups may be shown varying price structures or di¡erent product con¢gura-
tions in comparison with existing products. Participants may, for example, be asked to
spend ‘‘play’’ money in a shopping environment. They are asked to visit the shop and
make purchases with di¡erent sets of prices in operation at each visit.
The results of such arti¢cial experiments have to be assessed carefully because they
may not re£ect what the respondent would actually do in a real situation. They may
just play the game, providing the organizers with the answers they are expecting,
 CHAPTER 6 g ESTIMATION OF DEMAND FUNCTIONS 109

rather than their own true views. Nevertheless, they may provide useful information,
particularly about product characteristics and the combinations of characteristics that
consumers prefer.
The Gabor^Granger Test is used to test the potential of new products by comparing
a new product with an existing one: Half the group are shown the new product and
asked whether they would buy it at various prices on a random price list. They are
then shown the existing product. The other half are shown the original product, ¢rst,
and the new product, second. The objective is not only to gain some idea of the
acceptance of the product by consumers but also to eliminate bias by showing the
products to the two groups in a di¡erent order.

MARKET STUDIES

Market studies involve testing real products in real markets with real people. For
example, a ¢rm might select a region of the UK with its own regional or local
commercial radio and television station to test-market a new chocolate bar or washing
powder. If the new product sells su⁄ciently well against a competitor’s and is seen as
indicating consumer acceptance and satisfaction, the producer may then decide to
launch the product in other regions or to go nationwide. An existing product might be
promoted in one area at a lower price, backed by advertising to again check consumer
reactions.

STATISTICAL ANALYSIS

The traditional economic approach to estimating a demand function is for the ¢rm to
use statistical methods, using data collected either by the ¢rm or other outside sources,
such as industry associations or government agencies. Historic data can be of two
types: ¢rst, time series data for sales and other variables over a period of time
measured for a discrete time interval, such as monthly, quarterly or yearly; second,
cross-sectional data, such as expenditure by di¡erent income groups on a product at
the same point in time. Statistical procedures are applied to these data to look
for meaningful relationships, the most commonly used methodology being linear
regression analysis.
When modelling the relationship between quantity demanded and independent
variables, the analysis we undertook in Chapter 5 enables us to hypothesize what the
expected sign of the coe⁄cient for each independent variable should be: for example,
the sign of the coe⁄cient for the product’s own price should normally be negative; for
a substitute product and income the coe⁄cients should normally be positive.
Assuming data have been collected for the key independent variables, the task then is
to obtain from the available data the best ¢t, or statistically most acceptable, equation
that explains the quantity demanded.
110 PART II g KNOWING THE MARKET

The typical form of such a linear demand equation for cars using more than one
independent variable would be:
Qc ¼ a þ b1 Pc þ b2 Py þ b3 Y þ b4 A þ bn Xn
where Qc ¼ quantity demanded, Pc ¼ price of the product, Py ¼ the price of other
products, Y ¼ income, A ¼ advertising expenditure and Xn ¼ all the other variables
that might be included in the model.
If the equation for the demand for cars is recast logarithmically, then we have:
log Qc ¼ a þ bi log Pc þ b2 log Py þ b3 log Y þ b4 log A þ bn log Xn
The advantage of this procedure is that the estimated coe⁄cients of the demand
function are the various elasticities of demand: that is, b1 ¼ the own price elasticity of
demand, b2 ¼ the cross price elasticity of demand, b3 ¼ the income elasticity of
demand, b4 ¼ the elasticity of advertising and bn ¼ the elasticity with respect to that
variable. To estimate a demand function using regression analysis, data have ¢rst to
be collected for each of the variables to be included in the model: that is, for quantity
demanded, prices, income, advertising and any other variable considered worthy of
inclusion.
The simplest estimating procedure is linear regression analysis. Linear regressions
can be estimated using various computer software programs including spreadsheets
and speci¢c statistical packages for economists and social scientists, such as SPSS (see
Judge 2000; Whigham 2001). The resulting output is an equation together with
statistical inference statistics, which provides the means for assessing the statistical
signi¢cance of the estimated coe⁄cients of the included variables. A simple example of
this procedure is illustrated in Chapter 21, where a demand function is estimated for
the UK bus market using time series data. The case study of alcoholic drinks in this
chapter estimates demand functions for beer, wine and spirits.

PITFALLS USING REGRESSION ANALYSIS

Speci¢cation errors

A speci¢cation error arises when one or more important determinants of demand are
omitted from the model or when the wrong functional form was speci¢ed to estimate
the function: for example, if a linear rather than a non-linear relationship is speci¢ed.
The results of mis-speci¢cation show in a low value for R 2 , while the omission of
important variables leads to variables not having the expected signs.

Identi¢cation problems

The identi¢cation problem occurs because of the simultaneous change between one
variable included in the model and one not included: for example, there may a simul-
 CHAPTER 6 g ESTIMATION OF DEMAND FUNCTIONS 111

P3 C
Price

P2 B

P1
A

D3
D2
D1

O Q1 Q2 Q3
Quantity
Figure 6.3 Identi¢cation of demand curves

taneous relationship between quantity demanded and consumer income, which was
not included in the model.
The problem is illustrated in Figure 6.3, where cross-sectional data collection yields
the three combinations of quantity and price, labelled A to C representing price^
quantity combinations P1 Q1 , P2 Q2 and P3 Q3 . These points joined together show that a
fall in price leads to an increase in the quantity demanded. This relationship suggests
an upward rather than a downward-sloping demand curve, which is not normally
expected. The problem may arise because points A, B and C do not lie on a single
demand curve but on separate demand curves D1 , D2 , D3 , each demand curve being
associated with di¡erent levels of income and/or preferences. The true position of each
demand curve cannot be identi¢ed. The identi¢cation problem occurs, therefore,
because of the simultaneous change between price, included as an explanatory
variable, and income, which is not included. If an upward-sloping demand curve is
estimated from the data collected but the identi¢cation problem is not identi¢ed, then
the measured price elasticity would have a positive rather than a negative sign,
indicating speci¢cation problems.

THE ECONOMIC VERIFICATION OF REGRESSION MODELS

Statistics are known to generate apparent strong but spurious statistical relationships.
The ¢rst things an economist should check when looking at the output of a regression
calculation are the signs and magnitudes of the estimated variables. Economic analysis
112 PART II g KNOWING THE MARKET

suggests that for a normal good the own price elasticity of demand should have a
negative sign, the cross-price elasticity of demand should have a negative sign and the
income elasticity of demand should be positive. While contrary signs are not
necessarily wrong they indicate that one should proceed with caution. Likewise, if the
magnitude of the estimated variables is outside the expected range, then again one
should proceed with caution. In Table 6.2 there are a number of results contrary to
expectations: for example, the own price elasticity for beer has a positive sign. Signi¢-
cance tests show this coe⁄cient to be signi¢cant at the 5% level. The coe⁄cient for the
price of other goods is also negative, contrary to expectations, but signi¢cance tests
show that its value is not signi¢cantly di¡erent from zero.

STATISTICAL VERIFICATION OF REGRESSION MODELS

The second stage is to examine the statistical indicators as measured by the estimating
procedure to see whether the model is statistically signi¢cant and successfully explains
variations in the dependent variable or quantity demanded. These tests deal with the
overall explanatory power of the model, as well as the role of each independent
variable.

Correlation coe⁄cient

The overall explanatory power of a regression model is measured by R 2 or adjusted R 2 .

R 2 measures the goodness of ¢t, or the amount of variation explained by the
independent variables included in the model. The adjusted R 2 takes into account
degrees of freedom, because otherwise the value of R 2 can be improved by adding
more independent variables. Thus, an R 2 of 0.9 indicates that independent variables
account for 90% of the variation, or changes, in the dependent variable. It also
indicates that 10% of variation is unaccounted for ^ explained by missing variables.
As a general rule the closer the value of R 2 is to 1 the better the ¢t, while the closer it
is to 0 the poorer the ¢t. A good overall ¢t, or high R 2 , is not the end of testing,
because it may be the result of problems with independent variables. The degrees of
freedom are measured by deducting the number of independent variables from the
number of observations: for example, if there are 25 observations and 4 variables,
then the degrees of freedom would be 21. In Table 6.2 the adjusted R 2 ’s are all highly
signi¢cant for all three products, showing that, initially, the model has been success-
fully constructed.

F-test

The F-test also assesses the overall validity of the regression model. It is used to see
whether there is a signi¢cant relationship between the dependent variable and the
 CHAPTER 6 g ESTIMATION OF DEMAND FUNCTIONS 113

group of independent variables. The test itself is based either on accepting or rejecting
the null hypothesis that there is no signi¢cant statistical relationship between the
dependent and independent variables as a group. The test proceeds by comparing the
F-value estimated when measuring the regression relationship and the benchmark
value obtained from F distribution statistical tables. The benchmark values are a
function of the degrees of freedom of the denominator, the degrees of freedom for the
numerator and the probability of being wrong. Thus, a 5% probability combined with
20 degrees of freedom for the denominator and 20 degrees of freedom for the
numerator gives a benchmark value for F of 2.12; whereas with 120 degrees of
freedom for both denominator and numerator a 5% probability gives a benchmark
value of 1.35.
If the estimated F-value is greater than the benchmark value, then the null
hypothesis can be rejected and, obversely, it can be claimed that there is a signi¢cant
relationship between the two variables.
The degree of freedom for the numerator is the number of independent variables
(excluding the constant term), while the degrees of freedom for the denominator is
given by the total number of independent variables including the constant and
subtracting them from the number of observations. Thus, a regression function with 3
independent variables and 28 observations would have 3 degrees of freedom for the
denominator and 24 degrees of freedom for the numerator, giving a benchmark value
for F of 8.64.

t-test

The t-test is used to establish whether there is a signi¢cant statistical relationship

between an independent and a dependent variable. It is based on the hypothesis that
there is no signi¢cant statistical relationship between the dependent and the
independent variable; this is known as the null hypothesis. The t-test tries to show
that the null hypothesis can be proved incorrect and that there is a signi¢cant relation-
ship between the dependent and independent variables.
The test proceeds by comparing the t-value estimated when measuring the
regression relationship and the benchmark value obtained from statistical tables. The
benchmark values are a function of the degrees of freedom and the probability of being
wrong. Thus, a 5% probability combined with 20 degrees of freedom produces a
benchmark value for t of 1.81. If the estimated t-value is greater than the benchmark
value, then the null hypothesis can be rejected and, obversely, it can be claimed that
there is a signi¢cant relationship between the two variables. Thus, as a general rule
the benchmark value for t is taken to be around 2, and if the estimated t-value is
greater than 2, then that independent variable is assumed to play a signi¢cant role in
the model.
In Table 6.2 the signi¢cant t-tests that indicate that the null hypothesis can be
rejected are marked by a superscript ‘‘a’’; these show that the real price and real
income are signi¢cant explanatory variables for all three products. The price of other
goods is signi¢cant for spirits and wine, while advertising is only signi¢cant for beer.
114 PART II g KNOWING THE MARKET

Standard error of estimate

The standard error of estimate is used to check whether the relationship between an
independent and dependent variable is signi¢cant. It measures the degree of dispersion
of the data around the estimated value for the variable. It can then be used to indicate
the degree of con¢dence that the value of the variable will fall within the measured
limits. The general rules for using the standard error are that there is:

g A 68% probability that actual values will fall within plus or minus one standard
errors of its estimated value.
g A 95% probability that actual values will fall within plus or minus two standard
errors of its estimated value.
g A 99.7% probability that actual values will fall within plus or minus three standard
errors of its estimated value.

ECONOMETRIC VERIFICATION OF THE REGRESSION ESTIMATES

Initially, it was suggested that the ¢rst step in checking the overall validity of a
regression model was to look at the value of R 2 . However, economic modelling is beset
with di⁄culties that can in£ate the value of R 2 , because of problems associated with re-
lationships between the independent variables, as well as a particular problem
associated with the use of time series known as autocorrelation.

Multi-collinearity

Multi-collinearity is said to exist when the independent variables within an estimated

model are correlated to each other. Problems arise when two independent variables
present the same information in a di¡erent way: for example, if social class and income
are included in a demand model for di¡erent types of alcoholic drink, then social class
may duplicate the information provided by the income variable, in so far as social
class is associated with income levels. High levels of multi-collinearity will have an
e¡ect on the estimated coe⁄cients.
Detecting multi-collinearity is usually achieved by investigating the pattern of
correlation coe⁄cients between the independent variables and by examining the R 2
and t-statistics: for example, the model may achieve a high R 2 , but the t-statistics may
indicate that a number of variables, expected to be signi¢cant, are in fact insigni¢cant.
If multi-collinearity is identi¢ed as a problem, then the solution may be to drop the less
signi¢cant of the highly correlated variables or to introduce time lags for some of the
variables and to re-estimate the model.

Autocorrelation

Autocorrelation may be found where the error terms within a regression are serially
correlated. It is a problem because it can lead to either overestimating or under-
estimating the unexplained variation in the dependent variable. The consequences
 CHAPTER 6 g ESTIMATION OF DEMAND FUNCTIONS 115

may be to accept a model as a good ¢t when the result is dependent on autocorrelation.

The test for autocorrelation relies on a comparison between the Durbin^Watson
Statistic and an upper and lower value derived from statistical tables. The Durbin^
Watson Test checks to see whether null hypothesis holds and that there is no autocorre-
lation present in the model. For the 5% level of signi¢cance with 25 data observations
and 3 independent variables, the lower limit is 1.12 and the upper limit is 1.66. If the
estimated Durbin^Watson Statistic is less than 1.12, then the null hypothesis is
rejected; while if the estimate Durbin^Watson Statistic is greater than 1.66, then the
null hypothesis that there is no autocorrelation present in the model is accepted.
Du¡y (1983) uses the Durbin^Watson Statistic to demonstrate that serial
correlation is not a problem and that the overall model works well (see Table 6.2).
Thus, if the estimated model passes the various statistical tests outlined, then the
estimated model may be regarded statistically as a model that ¢ts the data in a statistic-
ally acceptable way. The economic value of the model depends on whether the correct
explanatory variables have been included.

Case Study 6.1 The demand for beer, wine and spirits
To illustrate the use made of regression analysis by economists to estimate demand
functions, reference will be made of studies that have estimated demand functions for
alcoholic drinks. Duffy (1983) estimated demand functions for beer, spirits and wine
using quarterly data for the years 1963 to 1978. His aim was ‘‘to obtain reasonably
reliable estimates of the quantitative importance of the various factors which influence
the demand for alcoholic drink’’ (pp. 126–127). The demand equations were derived
using different methods, but here only the log-linear results using ordinary least squares
which were found to have the greatest explanatory power are reported: these are found in
Table 6.2.

Table 6.2 Estimates of log-linear demand functions for alcoholic drink

in the UK

Variable Beer Spirits Wine

Constant 31418a 2:2399

a
2:6390
a
ð11:2225Þ ð4:9596Þ ð4:3719Þ

Real price of good 0:2376a 1:1802

a
0:6385
a
ð1:7141Þ ð4:8437Þ ð1:7227Þ

Real price of all other goods 0:1530 0:9827a 0:6714a

ð1:0895Þ ð5:3567Þ ð2:1715Þ

Real income 0:8018a 1:6677a 2:5045a

ð6:7752Þ ð8:9848Þ ð11:6745Þ

Real per capita advertising 0:0742 0:0142 0:0865

ð2:6327Þ A ð0:3770Þ ð1:3869Þ

Adjusted R 2 0.950 0.975 0.963

Durbin–Watson Statistic 1.716 2.048 2.109
Standard error (10 3 ) 0.5068 0.4376 0.3070

Note t-ratios in parentheses

a
Statistically significant at the 5% level
Source Parts of table 1 from Duffy (1983).
116 PART II g KNOWING THE MARKET

In Table 6.2 the following information is reported for the log-linear demand functions for
each product:
g Constant term.
g Real price of the good.
g Real price of all other goods.
g Real income.
g The coefficient of determination, or adjusted R 2 .
g Coefficients for each independent variable.
g t-ratios for each variable.
g The Durbin–Watson Statistic.
The significant results found by Duffy (1983) include the following:
g Changes in real income are significant for all three products and measured income
elasticities are positive: for beer it is less than 1 (0.8), for spirits (1.6) and wine (2.8) it
is greater than 1.
g Own price elasticities are negative for wines and spirits (the expected sign), but less
than 1 for wine (0:6) and greater than 1 for spirits (1:18): for beer price elasticity is
positive rather than negative though the measured elasticity of (0.2) is not signifi-
cantly different from 0.
g The elasticity for advertising is positive and significant, but small for beer (0.07) and
negative and insignificant for wine (0.01) and spirits (0.08). The results show that
advertising has a very small impact on the total sales of beer, wine and spirits.
g Statistically, the adjusted R 2 indicates that the models have significant explanatory
power, while the Durbin-Watson Statistic indicates there were no problems with
autocorrelation as explained above.
The studies by Duffy and others (see Brewster 1997, pp. 153–154) show that beer tends to
have a very low price elasticity, a low cross elasticity of demand and a low but positive
income elasticity of demand.
More sophisticated models have been developed using demand systems to estimate
the elasticities: Duffy (1987) found the own price estimate of the elasticity of demand for
beer, using data from 1975 to 1983, to be 0:36, for income to be þ0:71 and for advertising
þ0:05.

CHAPTER SUMMARY

In this chapter we brie£y reviewed the methods used to obtain information about the
characteristics of the demand function for a ¢rm’s products. In doing this we analysed:

g Surveys and questionnaires.

g Consumer and market experiments.
g Regression.

None of the methods is entirely satisfactory: survey methods have drawbacks relating to
the questions asked and the veracity of the answers and statistical methods also su¡er
from information, estimation and interpretation problems. Despite the shortcomings
 CHAPTER 6 g ESTIMATION OF DEMAND FUNCTIONS 117

identi¢ed, it is imperative for the ¢rm to discover the nature of the demand functions for
its products and the variables in£uencing demand.
Knowing the size of the elasticities for price, income and advertising can shape not
only the pricing and sales strategies of ¢rms but also the long-term growth of sales.

REVIEW QUESTIONS

1 In what circumstances should a company employ survey methods to obtain more

information on the demand for its product and the relative merits of its product
compared with those of rivals?
2 What are the advantages and disadvantages to a company of using surveys and
questionnaires to estimate demand functions?
3 What are the advantages and disadvantages of using regression analysis to
estimate a demand function?
4 What is the identi¢cation problem?
5 The estimated log-linear regression, where PX ¼ the price of the product, PO ¼ the
price of another product and Y ¼ real income, is as follows:
QX ¼ 450  1:53PX þ 0:87PO þ 2:36Y R 2 ¼ 0:91
where the t-ratios are in parentheses.
^ Explain the meaning and signi¢cance of the coe⁄cients for each variable. Are the
signs in line with economic theory?
^ If the initial values of the variables are PX ¼ »100, PO ¼ »120 and Y ¼ »1,000,
then calculate the quantity demanded.
^ Calculate the impact on sales of a plus or minus change of 10 in the value of each
of the independent variables on the quantity demanded.
5 What are the di⁄culties that a ¢rm faces in making estimates of the potential
demand for a new product? What methods might it use to gauge potential
consumer reaction?
6 If you were to estimate demand functions for beer, wine and spirits now, in what
ways would you expect the results to be di¡erent from those of Du¡y (1983). In
your answer consider the own price elasticity of demand, the income elasticity of
demand and cross-price elasticity.
7 You are asked to estimate a demand function for electricity. What information
would you need to collect to estimate such a function?

REFERENCES AND FURTHER READING

Baumol, W.J. (1965) On empirical determination of demand relationships. Economic Theory and
Operations Analysis (2nd edn, Chapter 10). Prentice Hall, Englewood Cli¡s, NJ.
Brewster, D. (1997) Demand and revenue analysis. Business Economics (Chapter 7). Dryden Press,
London.
Du¡y, M. (1983) The demand for alcoholic drink in the UK, 1963^78. Applied Economics, 15,
125^140.
118 PART II g KNOWING THE MARKET

Du¡y, M. (1987) Advertising and the inter-product distribution of demand: A Rotterdam model
approach. European Economic Review, 31, 1051^1070.
Green, P.E. and D.S. Tull (1988) Research for Marketing Decisions (5th edn). Prentice Hall,
Englewood Cli¡s, NJ.
Gri⁄ths, A. and S. Wall (1996) Market demand. Intermediate Micro-economics (Chapter 2).
Longman, London.
Hill, S. (1989) Demand theory and estimation. Managerial Economics (Chapter 5). Macmillan,
Basingstoke, UK.
Judge, G. (2000) Computing Skills for Economists. John Wiley & Sons, Chichester, UK.
Luck, D.J., H.G. Wales and D.A. Taylor (1987) Marketing Research (7th edn, Chapters 9 and 10).
Prentice Hall, Englewood Cli¡s, NJ.
Whigham, D. (2001) Demand analysis. Managerial Economics Using Excel (Chapter 5). Thompson
Learning, London.