How to Work with Dummy Independent Variables

Chapter 8 is devoted to dummy (independent) variables. This How To answers common

questions on working with and interpreting dummy variables.
Questions:
1) How to include dummy variables in a regression?
2) How to interpret a coefficient on a dummy variable?
3) How to test hypotheses with dummy variables and interaction terms?
4) How to create a double-log functional form with dummy variables?
5) How to interpret a coefficient on a dummy variable with a log dependent variable?

HoToUseDummyXVariables.doc

Page 1 of 7

1) How to include dummy variables in a regression?

Example:
You want to include Region of the United States in your earnings function regression.
You obtain the variable GMREG from the Current Population Survey (CPS), and it has
four possible values that the codebook maps to a region like this:
GMREG
Region
1
Northeast
2
Midwest
3
South
4
West

The data are sitting in an Excel file column like this:

GMREG
3
3
1
4
2
4
3

Obviously, you CANNOT use GMREG directly in a regression.

To incorporate region as a dummy variable, follow these steps:
1) Create Number of Categories 1 new variables. (In this example, 4 1 = 3 new
variables)

In a new column, enter Northeast as the label for the variable.

Use an IF statement to create a 1 is GMREG is 1; otherwise a 0.

=IF(B2=1,1,0)
(where B2 has the value of GMREG for that observation)

Repeat for Midwest and South.

2) Include the Number of Categories 1 variables in the regression.

The choice of which category to leave out (in this example, West) is totally arbitrary and
has no effect on the final results. The actual coefficients of the regression equation do, of
course, depend on the category left out (called the base case), but because you interpret a
dummy variable coefficient relative to the base case, the predicted values end up the
same. See Section 8.2 for more.

HoToUseDummyXVariables.doc

Page 2 of 7

2) How to interpret a coefficient on a dummy variable?

For a single dummy variable without an interaction term, the value of the coefficient tells
you the change in the value of the dependent variable compared with the base case.
Example:
Predicted Wage = 10.2 + 1.3Northeast 0.9Midwest 1.7South
Consider cases:

Person is from the Midwest. Then this person is predicted to make 10.2 0.9 =
9.3 (dollars per hour)since the values for Northeast and South are zero.

Southerners make $1.70 per hour less than Westerners.

Example:
Predicted Wage = 12.7 + 2.1Education 1.8Female

Holding Education constant, Female make $1.80 per hour less than Males (the
base case).

See Sections 8.3 and 8.4 for more.

If you have an interaction term, the situation get more complicated.
Example from Female.xls:
Predicted Wage = 0.076 + 0.88Education 4.28Female + 0.16Female*Education
The Female coefficient is an intercept shifter and the interaction term affects the slope on
Education. Here is a graph that explains what is going on:

Average Hourly Wage

Predicted Wage from Regression of Wage on Education, Female, and FemaleEducation (Interaction Term)
0
1 F-M Differential
8 $ 7.12 $ 4.13 $
(2.99)
Male and Female Regression Lines
9 $ 8.00 $ 5.17 $
(2.83)
$18
10 $ 8.88 $ 6.22 $
(2.67)
$16
$14
11 $ 9.76 $ 7.26 $
(2.51)
$12
11.5 $ 10.20 $ 7.78 $
(2.43)
$10
12 $ 10.65 $ 8.30 $
(2.35)
Male
$8
$6
13 $ 11.53 $ 9.34 $
(2.19)
Female
$4
14 $ 12.41 $ 10.38 $
(2.02)
$2
16 $ 14.17 $ 12.47 $
(1.70)
$18 $ 15.93 $ 14.55 $
(1.38)
8
9 10 11 12 13 14 15 16 17
Education

Figure 8.5.2: Graphical Interpretation of the Interaction Term

HoToUseDummyXVariables.doc

Page 3 of 7

18

A very useful approach to interpretation is to create a table and compute the predicted Y
for given values of the included Xs.
See Section 8.5 for more.
NOTE: Sometimes, when you add an interaction term, the coefficient on the dummy
variable becomes counter intuitively signed. For example, adding Black*Education
might make the Black dummy variable coefficient positive. But the negative sign on the
Black*Education must also be included when figuring out the effect of being black on
predicted wage, and you will then obtain the expected result that blacks make less than
non-blacks.
3) How to test hypotheses with dummy variables and interaction terms?
The F-test is the way to do this.
Obtain the SSR for the restricted and unrestricted models, compute the F-statistic
(properly adjusting for the degrees of freedom in numerator and denominator), then find
the P-value.
Example using data and results from Female.xls (in the Chapter 8\ExcelFiles folder):
In a study of the determinants of wages, we want to see whether being female matters
after controlling for education. We think that being female affects both the slope and
intercept of the relationship between wages and education. Here are results from the
unrestricted regression:
Predicted Wage = 0.076 + 0.88Education 4.28Female + 0.16Female*Education
SSR = 419141
n = 8546
The corresponding restricted regression results are:
Predicted Wage = -1.53 + 0.92 Education
SSR = 430216.
n = 8546
(This regression is not in Female.xls, but it is very easy to run.)
For a test that there is no difference between male and female wages after controlling for
education, the null hypothesis says that the coefficients on Female AND
Female*Education together equal zero.
The test statistic is

HoToUseDummyXVariables.doc

Page 4 of 7

F stat

430216 41941

419141
8542
5441

49
111 .

The P-value is very small. We would reject the null hypothesis that being female doesnt
matter for wage determination, after controlling for education.
See Section 17.7 for more.
4) How to create a double-log functional form with dummy variables?
You cannot take the natural log of a dummy variable because ln(0) is undefined.
Thus, you cannot create a completely double-log specification when you have dummy
independent variables.
What is usually done is to take the natural log of the Y and continuous X variables,
leaving the dummy variables untransformed.
Example (based on Duquette (1999)):
Own Units (or Levels) Specification:
CharitableContributions i 0 1 Price i 2 Disposable Incomei 3 Married i t

All variables are continuous except Married. Price measures the after-tax cost of a
contribution. For example if for every dollar I give, my tax bill falls by 25 cents, the
price is 0.75.
To make this a log-transformed model we take the natural logs of the continuous
variables and leave the dummy variables alone.
Log Specification:
ln CharitableContributions i 0 1ln Price i 2 ln Disposable Income i 3 Married i t

The b1 and b2 estimates could be interpreted as price and income elasticities respectively.
The estimate of b3 would be roughly the approximate percentage change in charitable
contirbuitons for married tax payers versus unmarried tax payers. See the answer to the
next question to learn how to precisely interpret the coefficient estimate for Married.

HoToUseDummyXVariables.doc

Page 5 of 7

5) How to interpret a coefficient on a dummy variable with a log dependent

variable?
The coefficient on a dummy variable with a log-transformed Y variable is interpreted as
the percentage change in Y associated with having the dummy variable characteristic
relative to the omitted category, with all other included X variables held fixed.
Example (Duquette (1999, p. 201)):
Predicted ln Charitable Giving = - 1.3 lnPrice + 0.91 lnIncome +Married
Approximate Interpretation:
Predicted Charitable Giving is approximately 46 percent higher in for married tax
payers, holding constant price of giving and income.
Exact Interpretation:
Actually, it is possible to obtain a more exact interpretation, and because the coefficient
on the dummy variable is large, it matters. What we want is the percentage change in
Charitable Giving between married and single taxpayers. This may be done as follows.
Let PCG denote the predicted value of CharitableGiving (in the levels, not in natural
logarithms) under a particular set of circumstances. We may then compute the percentage
change in the predicted values between Married = 1 and Married = 0, with all other
included Xs held constant, as

PCGMarried 1 PCGMarried 0
PCG Married 0
We can express the percentage difference in predicted values as
PCG Married 1
1
PCG Married 0

Now what we actually obtain from the log-specified regression equation is:
ln( PCG Married 1 ) ln( PCG Married 0 )
ln PCG Married 1 PCG Married 1
b3
0.46.

By the properties of logs, it then follows we can take the anti-log to obtain
PCG Married 1
exp b3 1.58.
PCG Married 1

Then the exact computation of the percentage change is

HoToUseDummyXVariables.doc

Page 6 of 7

 PCG Married 1 PCG Married 0

PCG Married 1
1 exp(b3 ) 1 exp(0.46) 1 1.58 1 0.58
PCG Married 0
PCG Married 0
Or, 58 percent. That is, married taxpayers, on average, make charitable contributions
58% percent higher than unmarried studies holding income and price constant.

The smaller the coefficient value, the closer the approximation will be to the exact
computation because x gets closer to exp(x) 1 the smaller the x.
Reference:
Duquette, Christopher M. (1999) Is Charitable Givign by Nonitemizers Responsive to
Tax Incentives? New Evidence. National Tax Journal 52(2): 195-206.

HoToUseDummyXVariables.doc

Page 7 of 7