Example

MOOC Econometrics log(Wage)i =

Lecture 2.2 on Multiple Regression: β1 + β2 Femalei + β3 Agei + β4 Educi + β5 Parttimei + εi

Representation
Christiaan Heij ‘Wage’: yearly wage (index, median = 100)
‘Female’: gender dummy (1 for females, 0 for males)
‘Age’: age (in years)
‘Educ’: education (4 levels, from 1 for low to 4 for high)
‘Parttime’: part-time job dummy (1 if work on 3 or less days
per week, 0 if more than 3 days per week)

Lecture 2.2, Slide 2 of 13, Erasmus School of Economics

Notation Matrix notation

Write yi = xi0 β + εi for 500 observations in database:

yi = log(Wage)i
   0   
x1i = 1 x2i = Femalei x3i = Agei y1 x1 ε1
 y2   x 0   ε2 
  2 
 ..  =  ..  β +  .. 
  
x4i = Educi x5i = Parttimei  .   .   . 
y500 0
x500 ε500
Let xi be (5 × 1) vector with components (x1i , . . . , x5i ).
Let β be (5 × 1) vector with components (β1 , . . . , β5 ). Let y : (500 × 1) vector with components yi
X : (500 × 5) matrix with rows xi0
Then wage equation can be written as
ε: (500 × 1) vector with components εi
yi = 5j=1 βj xji + εi = xi0 β + εi
P

Then wage equation for 500 observations becomes:

Symbol 0 (prime): transposition (see Building Blocks).
y = Xβ + ε
Lecture 2.2, Slide 3 of 13, Erasmus School of Economics Lecture 2.2, Slide 4 of 13, Erasmus School of Economics
Multiple regression model Multiple regression model

Let database contain n observations for all variables.

Model with k explanatory factors:
Pk As before, let
yi = β1 + β2 x2i + . . . + βk xki + εi = j=1 xji βj + εi
y : (n × 1) vector with components yi
(with x1i = 1).
X : (n × k) matrix with elements xji
yi is dependent or explained variable, β: (k × 1) vector with components βj
x1i , . . . , xki are regressor variables or explanatory factors. ε: (n × 1) vector with components εi

First ‘explanatory’ factor is the constant x1i = 1. Then model can be written as
y = Xβ + ε

Lecture 2.2, Slide 5 of 13, Erasmus School of Economics Lecture 2.2, Slide 6 of 13, Erasmus School of Economics

Set of linear equations Test answers

y = Xβ + ε y = X β where X is (n × k) with rank(X ) = r .

→ X β is ‘explained’ part of y → Always r ≤ k and r ≤ n.
→ ε is ‘unexplained’ part of y → If r = n = k: y = X β has unique solution.
→ If r = n < k: y = X β has multiple solutions.
X explains much of y if y ≈ X β for some choice of β.
→ If r < n: y = X β has (in general) no solution.
y = X β is set of n equations in k unknown parameters β.
(Nearly always) n > k.
Test
We assume r = k < n.
Let X be (n × k) matrix with rank(X ) = r .
What is the number of solutions of the equations y = X β? So y = X β has (in general) no exact solution.

Lecture 2.2, Slide 7 of 13, Erasmus School of Economics Lecture 2.2, Slide 8 of 13, Erasmus School of Economics
Interpretation of model coefficients Decomposition of total effect

Total effect if factors are mutually dependent (and x1i = 1):

dy ∂y Pk ∂y ∂xh Pk ∂xh
dxj = ∂xj + h=2,h6=j ∂xh ∂xj = βj + h=2,h6=j βh ∂xj
Model: yi = β1 + β2 x2i + . . . + βk xki + εi .
Pk
What happens to y if xj increases by one unit while all other Indirect effects xj → xh → y combined: h=2,h6=j βh ∂xh
∂xj
x-variables xh (with h 6= j) remain fixed?
So: Total effect = Partial effect + Indirect effect
∂y
Partial effect: ∂xj = βj (if xh remains fixed for all h 6= j).
Example if part-time jobs more common for higher education:
Only possible as thought-experiment, called the ‘ceteris paribus’
assumption. Direct: Educ ↑ ⇒ Wage ↑
Indirect: Educ ↑ ⇒ Parttime ↑ ⇒ Wage ↓
Total: Sum of ↑ and ↓ effect, need effect sizes

Lecture 2.2, Slide 9 of 13, Erasmus School of Economics Lecture 2.2, Slide 10 of 13, Erasmus School of Economics

Testing for model restrictions Test answers (βj = 0 ⇒ xj no effect on y ?)

Factor xj in model if (relevant) effect on y .

Yes, in sense that xj has no partial effect
Test for single factor j: Test H0 : βj = 0 against H1 : βj 6= 0. (assuming all other explanatory factors remain fixed).

Test for two factors j and h: Test H0 : βj = βh = 0 against No, in sense that xj may have indirect effect
H1 : βj 6= 0 and/or βh 6= 0.
(via other factors xj → xh → y ).
General: Test H0 : Rβ = r against H1 : Rβ 6= r
→ R is given (g × k) matrix with rank(R) = g Example: log(Wage)i = β1 + β2 Educi + β3 Parttimei + εi
→ r is given (g × 1) vector
If β2 = 0 and β3 6= 0, then higher education still has indirect effect on
Test wage if having part-time job is related to education level.
If βj = 0, does this mean that xj has no effect on y ?
Motivate your answer.
Lecture 2.2, Slide 11 of 13, Erasmus School of Economics Lecture 2.2, Slide 12 of 13, Erasmus School of Economics
TRAINING EXERCISE 2.2

Train yourself by making the training exercise (see the website).

After making this exercise, check your answers by studying the

webcast solution (also available on the website).

Lecture 2.2, Slide 13 of 13, Erasmus School of Economics