ECON3049: ECONOMETRICS I

DUMMY VARIABLES
Marlon Tracey
Summer 2012
 Outline
2

 Nature and use of dummy variables

 Creation of dummy variables
 Dummy as explanatory variable

 Dummy variable trap

 Intercept dummy

 Slope dummy

 Dummy as dependent variable

 Linear probability model

Reading: Gujurati Chapt. 9 & 15; Wooldridge Chapt. 7

 Nature and Use of Dummy Variables
3

 If a variable is qualitative or categorical, standard regression can still

be used to analyze the effect of such variables. This is done by using
dummy variables.
 Dummy variables are variables that take on only two values 0 and 1,
where 0 means the absence of a quality and 1 means the presence of
the quality.
 A dummy variable can be used as an explanatory variable:
 to capture the effects of qualitative characteristics of individuals (e.g. race, religion,
gender, level of education), firms (e.g. private /public, large/medium/small/micro),
countries (e.g. OECD or otherwise, developed, developing, undeveloped,
underdeveloped) etc.;
 to examine before and after situations (pre-liberalization/post-liberalization; pre-
recovery/post-recovery) or

 To test whether a regression function is different for one group versus another group.

 A dummy variable can also be used as a dependent variable to predict

“choices”, “options” or qualitative outcomes.
 Creation of dummy variables
4

 Convert a variable with two categories (such as “distance from home” or

gender) to a dummy variable (D) as per the following examples:
1 if live near 1 if male
Dn   Dm  
0 if live far 0 if female

0 if live near 0 if male

Df   Df  
1 if live far 1 if female
 Convert a variable with three or more categories (such as quality of
degree) to dummy variables by creating for each category a separate
dummy variable as follows:
1 First Class 1 Upper Second Class
D1   D2  
0 Otherwise 0 Otherwise

1 Lower Second Class 1 Pass

D3   D4  
0 Otherwise 0 Otherwise
 Dummy as explanatory variable
5

 When dummy variables are used as explanatory variables

in a regression, one must be careful to avoid the dummy
variable trap.
 A dummy explanatory variable can either be an intercept
dummy and/or a slope dummy:
 An intercept dummy is used when the researcher theorizes the
dummy variable to have a level effect on the dependent variable,
that is, the level of the dependent variable is higher for one
category versus another.
 A slope dummy is used when the researcher theorizes the dummy
variable to change the marginal effect of an explanatory variable
on the dependent variable. It changes the magnitude and/or
direction of the relationship between an explanatory variable and
a dependent variable.
 Dummy variable trap
6

 In regression analysis, all possible dummies for a qualitative

explanatory variable should not be included in the equation.
 Doing so creates a dummy variable trap, which is due to perfect
multicollinearity.
 To avoid the dummy variable trap:
 If the variable has M categories, only M-1 dummies should be entered
in the regression equation.
 Or we can drop the intercept from the model and include all the
dummies.

 The final category not represented in the regression is referred to as

the reference category.
 The coefficients of the other dummy variables are interpreted
relative to the reference category.
 Intercept dummy
7

 Consider a regression model with one non-categorical (edu) and one

dummy (Df = 1 if female) as follows:
wage = 0 + 1edu + 0Df + u
 This can be interpreted as an intercept shift
 If Df = 0, then wage = 0 + 1edu + u
 If Df = 1, then wage = (0 + 0) + 1edu + u

 In this case, the group that is left out is “Males” and is represented by Df =
0. This group is the reference group.

 Note that the dummy Df when equal to 1, shifts the intercept from 0 to
(0 + 0). It is therefore called an intercept dummy.

 The effect of the intercept dummy is interpreted as follows: the difference

between wage if Df =1 and wage if Df =0 is 0 , ceteris paribus. More
intuitively, the gender wage differential is 0 .
 Slope dummy
8

 Consider a regression model with one non-categorical (edu) and one

dummy (Df = 1 if female) as follows:
wage = 0 + 1edu+ 1Df *edu+ u
 This can be interpreted as an intercept shift
 If Df = 0, then wage = 0 + 1edu + u
 If Df = 1, then wage = 0 + (1 +1) edu + u

 In this case, the group that is left out is “Males” and is represented by Df =
0. This group is the reference group.

 Note that the dummy Df when equal to 1, shifts the slope from 1 to (1 +
1). It is therefore called an slope dummy.
 The effect of the slope dummy is interpreted as follows: the difference
between the effect of education on wage if Df = 1 and the effect if Df = 0
is 1 , ceteris paribus. More intuitively, the difference between the marginal
return to education for females and males is 1 .
 Dummy as dependent variable
9

 In many cases, a researcher may be interested in the set of

variables that predict certain binary “choices” or “options”.
For example, the researcher might be interested in the factors
that determine the following:

 Decision to work or go to school

 Voting (vote or not vote)
 Marital Status (married or not)
 Decision to Pass or fail a course.

 Can OLS regression analysis still be conducted when the

dependent variable is binary?
 Linear Probability Model I
10

 Consider the simple linear regression model as follows:

Y   0  1 X  u
 Where the population regression function (PRF) is
E (Y | X )   0   1 X  E (u | X )  0

 Since Y is a dummy (or binary) variable at each value of X,

the E (Y | X )  p , where p is the probability that y = 1.
Therefore, in this case, the PRF is referred to as a linear
probability model.
 When the PRF is estimated by OLS on a particular sample,
the estimated linear probability equation is given by:
pˆ  ˆ 0  ˆ1 X
 Linear Probability Model II
11

 However, the linear probability model violates several

assumptions of our standard regression model:
 Linearity assumption is no longer realistic.
 Since X has a linear effect β1 on y, it is possible that pˆ  0 or pˆ  1 but
probability must lie in the interval [0,1].

 Y and therefore, u is non-normal. However, in large sample this

is not so much of a problem.

 u is heteroskedastic, since V ( u )  p (1  p ) , where p depends

on X. However, this problem can be corrected using Weighted
Least Squares.