Chapter 4: Estimating Demand

ESTIMATING DEMAND USING MARKETING RESEARCH TECHNIQUES

Three different marketing research methods that can be used in analyzing demand.
1. Consumer Surveys
 Consumer surveys involve questioning a sample of consumers to determine such factors as their
purchase intent, willingness to pay, sensitivity to price changes, and awareness of advertising
campaigns.
 Consumer surveys can provide a great deal of useful information to a firm
 Consumer expectations about future business and credit conditions may provide significant insights
into the consumers’ propensity to purchase many items, especially
durable goods.
2. Consumer Focus Group
 Another means of recording consumer responses to changes in factors affecting demand is through the
use of panel data on consumer focus groups
 focus groups A market research technique employing close observation of discussion among target consumers.
 In these situations, for example, experimental groups of consumers are given a small amount of money
with which to buy certain items. The experimenter can observe the impact on actual purchases as
price, prices of competing goods, and other variables are recorded. Then the group of consumers is
closely observed discussing the choices they made and why. Of course the costs of setting up and
running such a clinic are substantial, and the participants may suspect that the experimenter is
interested in sensitivity to prices and may respond more than otherwise would be the case.
3. Market Experiments in Test Stores
 market experiment, which examines the way consumers behave in controlled purchase environments.
 A test store may vary one or more of the determinants of demand, such as price or advertising or the
presence of an NFL or NBA logo on a sweatshirt, and observe the impact on quantity demanded.
 This approach may be especially useful in developing measures of the promotion elasticity of demand
for a product.
STATISTICAL ESTIMATION OF THE DEMAND FUNCTION
 Effective decision making eventually requires the quantitative measurement of economic relationships.
 Econometrics is a collection of statistical techniques available for estimating such relationships.
 The principal econometric techniques used in measuring demand relationships are regression and
correlation analysis.
Specification of the Model
 Normally the specific functional form of the regression relation to be estimated is chosen to depict the
true demand relationships as closely as possible.
 Graphing such relationships often will tell whether a linear equation is most appropriate or whether
logarithmic, exponential, or other transformations are more appropriate.
Linear Model
Q = α + β1A + β2P + β3M + e
Multiplicative Exponential Model
 This model is also popular because of its ease of estimation. For instance, Equation 4.5 may be
transformed into a simple linear relationship in logarithms (adding an error term) as follows:
log Q = log α + β1log A + β2log P + β3log M + e
 Demand functions in the multiplicative exponential form possess the convenient feature that the
elasticities are constant over the range of data used in estimating the parameters and are equal to the
estimated values of the respective parameters
 multiplicative exponential demand functions have constant price and other elasticities. This property
contrasts sharply with the elasticity of a linear demand function that changes continuously over the
entire price or income range of the demand curve.
A SIMPLE LINEAR REGRESSION MODEL
 The analysis in this section is limited to the simplest case of one independent and one dependent
variable, where the form of the relationship between the two variables is linear:
Y = α + βX + e [4.10]
 X is used to represent the independent variable and Y the dependent variable
Assumptions Underlying the Simple Linear Regression Model
1. Assumption 1 The value of the dependent variable Y is postulated to be a random variable, which is
dependent on deterministic (i.e., nonrandom) values of the independent variable X
2. Assumption 2 A theoretical straight-line relationship (see Figure 4.1) exists between X and the
expected value of Y for each of the possible values of X. This theoretical regression line
E(Y|X) = α + βX [4.11]
has a slope of β and an intercept of α. The regression coefficients α and β constitute population
parameters whose values are unknown, and we desire to estimate them.
3. Assumption 3 Associated with each value of X is a probability distribution, p(y|x), of the possible
values of the random variable Y. When X is set equal to some value xi, the value of Y that is observed
will be drawn from the p(y|xi) probability distribution and will not necessarily lie on the theoretical
regression line. As illustrated in Figure 4.2, some values of y|x i are more likely than others, and the
mean E(y|xi) lies on the theoretical regression line. If ei is defined as the deviation of the observed yi
value from its theoretical value y0 i, then
yi = yi + εi
yi = α + βxi + εi
or, in general, the linear regression relation becomes (as illustrated in Figure 4.3)
Y = α + βX + e [4.13]
where e is a zero mean stochastic disturbance (or error) term.
The disturbance term (ei) is assumed to be an independent random variable [that is, E(eiej) = 0 for I ≠ j]
that is normally distributed with an expected value equal to zero [that is, E(e i) = 0] and with a constant
variance equal to σ2ε [that is, Eðε2i Þ = σ2ε for all i]. Together, Assumptions 1 and 3 imply that the N(0,
σ2ε ) disturbance term is expected to be uncorrelated with the independent variables in the regression
model.
Estimating the Population Regression Coefficients
 Once the regression model is specified, the unknown values of the population regression coefficients α
and β are estimated by using the n pairs of sample observations (x1, y1), (x2, y2),…, (xn, yn).
 This process involves finding a sample regression line that best fits the sample of observations the
analyst has gathered.
 The sample estimates of α and β can be designated by a and b, respectively. The estimated or
predicted value of Y, ŷi, for a given value of X (see Figure 4.4) is
ŷi = a + bxi
 Letting ei be the deviation of the observed yi value from the estimated value ŷi, then
yi = ŷi + ei
= a + bxi + ei
or, in general, the sample regression equation becomes
Y = a + bX + e
 Although there are several methods for determining the values of a and b (that is, finding the
regression equation that provides the best fit to the series of observations), the best known and most
widely used is the method of least squares. The objective of least-squares analysis is to find values of a
and b that minimize the sum of the squares of the ei deviations. (By squaring the errors, positive and
negative errors cumulate without canceling each other. From Equation 4.15, the value of e i is given by
ei = yi − a – bxi
Squaring this term and summing over all n pairs of sample observations, one obtains

Using calculus, the values of a and b that minimize this sum of squared deviations expression are given
by

where x and y are the arithmetic means of X and Y, respectively (that is, x = Σx=n and y = Σy=n) and
where the summations range over all the observations (i = 1, 2,…, n).
USING THE REGRESSION EQUATION TO MAKE PREDICTIONS
A regression equation can be used to make predictions concerning the value of Y, given any particular value of
X. This is done by substituting the particular value of X, namely, xp, into the sample regression equation
ŷ = a + bxp
where, as you recall, ŷ is the hypothesized expected value for the dependent variable from the probability
distribution p(Y|X).6
 A measure of the accuracy of estimation with the regression equation can be obtained by calculating
the standard deviation of the errors of prediction (also known as the standard error of the estimate).
 standard error of the estimate The standard deviation of the error term in a linear regression model.
 The error term ei was defined earlier in Equation 4.17 to be the difference between the observed and
predicted values of the dependent variable. The standard deviation of the e i term is based on the
summed squared error (SSE) Σe2i normalized by the number of observations minus two:

 If the observations are tightly clustered about the regression line, the value of s e will be small and
prediction errors will tend to be small. Conversely, if the deviations e i between the observed and
predicted values of Y are fairly large, both se and the prediction errors will be large.
 The standard error of the estimate (se) can be used to construct prediction intervals for y.8 An
approximate 95 percent prediction interval is equal to9
ŷ ± 2se
Inferences about the Population Regression Coefficients
 For repeated samples of size n, the sample estimates of α and β—that is, a and b—will tend to vary
from sample to sample.
 In addition to prediction, often one of the purposes of regression analysis is testing whether the slope
parameter β is equal to some particular value β0.
 One standard hypothesis is to test whether β is equal to zero.10
 In such a test the concern is with determining whether X has a significant effect on Y.
 If β is either zero or close to zero, then the independent variable X will be of no practical benefit in
predicting or explaining the value of the dependent variable Y.
 When β = 0, a one-unit change in X causes Y to change by zero units, and hence X has no effect on Y.
 To test hypotheses about the value of β, the sampling distribution of the statistic b must be known. 11
 It can be shown that b has a t-distribution with n − 2 degrees of freedom.12,13
 The mean of this distribution is equal to the true underlying regression coefficient β, and an estimate
of the standard deviation can be calculated as
 For repeated samples of size n, the sample estimates of α and β—that is, a and b—will tend to vary
from sample to sample.
 In addition to prediction, often one of the purposes of regression analysis is testing whether the slope
parameter β is equal to some particular value β0.
and the decision is to reject the null hypothesis, if t is either less than −t0.25,n−2 or greater than +t0.25,n−2 where
the t0.25,n−2 value is obtained from the t-distribution (with n – 2 degrees of freedom) in Table 2 (Appendix B).15
 Business applications of hypothesis testing are well advised to keep the level of significance small (i.e.,
no larger than 1 percent or 5 percent). O
 ne cannot justify building a marketing plan around advertising and retail displays incurring millions of
dollars of promotional expense unless the demand estimation yields a very high degree of confidence
that promotional expenditures actually “drive” sales (i.e., β ≠ 0).
 There are simply too many other potentially more effective ways to spend marketing dollars.
 In the Sherwin-Williams Company example, suppose that we want to test (at the k = 0.05 level of
significance) whether promotional expenditures are a useful variable in predicting paint sales. In effect,
we wish to perform a statistical test to determine whether the sample value—that is, b = 0.433962—is
significantly different from zero.
 The null and alternative hypotheses are
H0: β = 0 ðno relationship between X and YÞ
Ha: β ≠ 0 ðlinear relationship between X and YÞ
14The level of significance (k) used in testing hypotheses indicates the probability of making an incorrect decision with the
decision rule—i.e., rejecting the null hypothesis when it is true. For example, with H 0: β ≥ 0, setting k equal to .05 (i.e., 5
percent) indicates that there is one chance in 20 that we will conclude that an effect exists when no effect is present—i.e., a
5 percent chance of “false positive” outcomes.
Because there were 10 observations in the sample used to compute the regression equation,
the sample statistic b will have a t-distribution with 8(= n − 2) degrees of freedom.
From the t-distribution (Table 2 of Appendix B), we obtain a value of 2.306 for
t.025,8.
 Therefore, the decision rule is to reject H0—in other words, to conclude that β ≠
0 and that a statistically significant relationship exists between promotional expenditures
and paint sales—if the calculated value of t is either less than −2.306 or greater
than +2.306.

Correlation Coefficient

 The value of the correlation coefficient (r) ranges from +1 for two variables with perfect positive
correlation to −1 for two variables with perfect negative correlation.
 In Figure 4.6, Panels (a) and (b) illustrate two variables that exhibit perfect positive and negative
correlation, respectively.
 A positive correlation coefficient indicates that high values of one variable tend to be associated with
high values of the other variable, whereas a negative correlation coefficient indicates just the opposite
—high values of one variable tend to be associated with low values of the other variable.
 Very few, if any, relationships between economic variables exhibit perfect correlation.
 Figure 4.6 Panel (c) illustrates zero correlation—no discernible relationship exists between the
observed values of the two variables.
The Analysis of Variance
 Another convenient measure of the overall “fit” of the regression model to the sample of observations
is the r-squared.
 We begin by examining a typical observation (yi) in Figure 4.7. Suppose we want to predict the value of
Y for a value of X equal to xi.
 While ignoring the regression line for the moment, what error is incurred if we use the average value
of Y (that is, y) as the best estimate of Y?
 The graph shows that the error involved, labeled the “total deviation,” is the difference between the
observed value (yi) and y.
 Suppose we now use the sample regression line to estimate Y.
 The best estimate of Y, given X = xi, is ŷi.
 As a result of using the regression line to estimate Y, the estimation error has been reduced to the
difference between the observed value (yi) and ŷi. In the graph, the total deviation (yi − y) has been
partitioned into two parts—the unexplained portion of the total deviation (y i − ŷi) and that portion of
the total deviation explained by the regression line (ŷi − y); that is,

 coefficient of determination A measure of the proportion of total variation in the dependent variable that is
explained by the independent variable(s).
MULTIPLE LINEAR REGRESSION MODEL
 A linear relationship containing two or more independent variables is known as a multiple linear
regression model
 In the (completely) general multiple linear regression model, the dependent variable Y is hypothesized
to be a function of m independent variables X1, X2,…, Xm, and to be of the form

Use of Computer Programs

 Using matrix algebra, procedures similar to those explained for the simple linear regression model can
be employed for calculating the estimated regression coefficients (α’s and β’s).
 A variety of computer programs can be used to perform these procedures.
 The output of these programs is fairly standardized to include the estimated regression coefficients, t-
statistics of the individual coefficients, R2, analysis of variance, and F-test of overall significance.

Using the Regression Model to Make Forecasts

 As in the simple linear regression model, the multiple linear regression model can be used to make
point or interval predictions.
 Point forecasts can be made by substituting the particular values of the independent variables into the
estimated regression equation.
 In the Sherwin-Williams example, suppose we are interested in estimating sales in a sales region where
promotional expenditures are $185,000 (i.e., A = 185), selling price is $15.00 (P), and disposable
income per household is $19,500 (i.e., M = 19.5).
 Substituting these values into Equation 4.31 yields
ŷ = 310.245 + .008(185) − 12.202(15.00) + 2.677(19.5) = 180.897 gallons
  Whether to include one, two, or all three independent variables in predicting ŷ depends on the mean
prediction error (e.g., here 185,000 − 180,897 = 4,103) in this and subsequent out-of-sample forecasts.
 The standard error of the estimate (se) from the output in Figure 4.8 can be used to construct
prediction intervals for Y.
 An approximate 95 percent prediction interval is equal to
ŷ ± 2se
 For a sales region with the characteristics cited in the previous paragraph (i.e., A = 185, P = $15.00, and
M = 19.5), an approximate 95 percent prediction interval for paint sales is equal to
180.897 ± 2(17.417)
or from 146,063 to 215,731 gallons.
Inferences about the Population Regression Coefficients
 Most regression programs test whether each of the independent variables (Xs) is statistically significant
in explaining the dependent variable (Y). This tests the null hypothesis: H 0: βi = 0
against the alternative hypothesis: Ha: βi ≠ 0
 The decision rule is to reject the null hypothesis of no relationship between paint sales (Y) and each of
the independent variables at the .05 significance level, if the respective t-value for each variable is less
than −t.025,6 = −2.447 or greater than t.025,6 = +2.447. As shown in Figure 4.8, only the calculated t-value
for the P variable is less than −2.447.
 Hence, we can conclude that only selling price (P) is statistically significant (at the .05 level) in
explaining paint sales.
 This inference might determine that marketing plans for this type of paint should focus on price and
not on the effects of promotional expenditures or the disposable income of the target households.

 Empirical estimates of the demand relationships nare essential if the firm is to achieve its goal of
shareholder wealth maximization.
 Without good estimates of the demand function facing a firm, it is impossible for that firm to make
profit maximizing price and output decisions.
 Consumer surveys involve questioning a sample of consumers to determine such factors as their
willingness to buy, their sensitivity to price changes or levels, and their awareness of promotional
campaigns.
 Focus groups make use of carefully directed discussion among groups of consumers. The results maybe
influenced by significant experimental bias.
 Market experiments observe consumer behavior in real-market situations.
 By varying product characteristics, price, advertising, or other factors in some markets but not in
others, the effects of these variables on demand can be determined.
 Market experiments are very expensive.
 Statistical techniques are often found to be of great value and relatively inexpensive as a means to
make empirical demand function estimates.
 Regression analysis is often used to estimate statistically the demand function for a good or service.
 The linear model and the multiplicative exponential model are the two most commonly used functional
relationships in demand studies.
 In a linear demand model, the coefficient of each independent variable provides an estimate of the
change in quantity demanded associated with a one-unit change in the given independent variable,
holding constant all other variables.
 This marginal impact is constant at all points on the demand curve.
 The elasticity of a linear demand model with respect to each independent variable (e.g., price elasticity
and income elasticity) is not constant, but instead varies over the entire range of the demand curve.
 In a multiplicative exponential demand model, the marginal impact of each independent variable on
quantity demanded is not constant, but instead varies over the entire range of the demand curve.
 However, the elasticity of a multiplicative exponential demand model with respect to each
independent variable is constant and is equal to the estimated value of the respective parameter.
 The objective of regression analysis is to develop a functional relationship between the dependent and
independent (explanatory) variable(s).
 Once a functional relationship (that is, regression equation) is developed, the equation can be used to
make forecasts or predictions concerning the value of the dependent variable.
 The least-squares technique is used to estimate the regression coefficients. Least-squares minimizes
the sum of the squares of the differences between the observed and estimated values of the
dependent variable over the sample of observations.
 The t-test is used to test the hypothesis that a specific independent variable is useful in explaining
variation in the dependent variable.
 The F-test is used to test the hypothesis that all the independent variables (X 1, X2, . . . , Xm) in the
regression equation explain a significant proportion of the variation in the dependent variable.
 The coefficient of determination (r2) measures the proportion of the variation in the dependent
variable that is explained by the regression equation (that is, the entire set of independent variables).
 The presence of association does not necessarily imply causation.
 Statistical tests can only establish whether or not an association exists between variables.
 The existence of a cause-and-effect economic relationship should be inferred from economic
reasoning.
Exercises