0% found this document useful (0 votes)

0 views

Week4 Nonlinear Models

Chapter 4 of 'Introduction to Econometrics' discusses nonlinear models and transformations of variables, emphasizing the distinction between linearity in variables and parameters. It introduces Ramsey's RESET test for functional misspecification, which helps identify potential nonlinearity in regression models. The chapter includes practical examples and data analysis to illustrate these concepts.

Uploaded by

meminatmaca55

Available Formats

Download as PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

0 views

Week4 Nonlinear Models

Uploaded by

meminatmaca55

Available Formats

Download as PDF, TXT or read online on Scribd

You are on page 1/ 222

Introduction to Econometrics

5th Edition
C. DOUGHERTY
Chapter heading

Chapter 4: Nonlinear Models and

Transformations of Variables

FALL 2024
Introduction to Econometrics
Chapter heading
LINEARITY and NONLINEARITY
LINEARITY AND NONLINEARITY

Linear in variables and parameters:

Y = 1 +  2 X 2 +  3 X 3 +  4 X 4 + u

This sequence introduces the topic of fitting nonlinear regression

models. First, we need a definition of linearity.
The model shown above is linear in two senses.
1. The right side is linear in variables because the variables are
included precisely as defined rather than as functions.
2. It is also linear in parameters since a different parameter
appears as a multiplicative factor in each term.

1
LINEARITY AND NONLINEARITY

Linear in variables and parameters:

Y = 1 +  2 X 2 +  3 X 3 +  4 X 4 + u

Linear in parameters, nonlinear in variables:

Y =  1 +  2 X 22 +  3 X 3 +  4 log X 4 + u

The second model above is linear in parameters, but nonlinear in variables.

4
LINEARITY AND NONLINEARITY

Linear in variables and parameters:

Y = 1 +  2 X 2 +  3 X 3 +  4 X 4 + u

Linear in parameters, nonlinear in variables:

Y =  1 +  2 X 22 +  3 X 3 +  4 log X 4 + u
Z 2 = X 22 , Z 3 = X 3 , Z 4 = log X 4

Such models present no problem at all. Define new variables as shown.

5
LINEARITY AND NONLINEARITY

Linear in variables and parameters:

Y = 1 +  2 X 2 +  3 X 3 +  4 X 4 + u

Linear in parameters, nonlinear in variables:

Y =  1 +  2 X 22 +  3 X 3 +  4 log X 4 + u
Z 2 = X 22 , Z 3 = X 3 , Z 4 = log X 4
Y = 1 +  2 Z 2 +  3 Z 3 +  4 Z4 + u

With these cosmetic transformations, we have made the model linear in both
variables and parameters.
6
LINEARITY AND NONLINEARITY

Linear in variables and parameters:

Y = 1 +  2 X 2 +  3 X 3 +  4 X 4 + u

Linear in parameters, nonlinear in variables:

Y =  1 +  2 X 22 +  3 X 3 +  4 log X 4 + u
Z 2 = X 22 , Z 3 = X 3 , Z 4 = log X 4
Y = 1 +  2 Z 2 +  3 Z 3 +  4 Z4 + u

Nonlinear in parameters:

Y = 1 +  2 X 2 +  3 X 3 +  2  3 X 4 + u

This model's parameters are nonlinear since the coefficient of X4 is the product of the
coefficients of X2 and X3. As we will see, some nonlinear models can be linearized by
appropriate transformations, but this is not one of them.
7
LINEARITY AND NONLINEARITY

Average annual percentage growth rates

Employment GDP Employment GDP

Australia 2.57 3.52 Korea 1.11 4.48

Austria 1.64 2.66 Luxembourg 1.34 4.55
Belgium 1.06 2.27 Mexico 1.88 3.36
Canada 1.90 2.57 Netherlands 0.51 2.37
Czech Republic 0.79 5.62 New Zealand 2.67 3.41
Denmark 0.58 2.02 Norway 1.36 2.49
Estonia 2.28 8.10 Poland 2.05 5.16
Finland 0.98 3.75 Portugal 0.13 1.04
France 0.69 2.00 Slovak Republic 2.08 7.04
Germany 0.84 1.67 Slovenia 1.60 4.82
Greece 1.55 4.32 Sweden 0.83 3.47
Hungary 0.28 3.31 Switzerland 0.90 2.54
Iceland 2.49 5.62 Turkey 1.30 6.90
Israel 3.29 4.79 United Kingdom 0.92 3.31
Italy 0.89 1.29 United States 1.36 2.88
Japan 0.31 1.85

We will start with an example of a simple model that can linearize a cosmetic
transformation. The table displays the average annual employment and GDP
growth rates for 31 OECD countries. 8
LINEARITY AND NONLINEARITY

2
e = 1 + +u
g
3
Employment growth rate

0
0 1 2 3 4 5 6 7 8 9
GDP growth rate

A plot of the data reveals that the relationship is clearly nonlinear. We will
consider various nonlinear specifications for the relationship in the course
of this chapter, starting with the hyperbolic model shown. 9
LINEARITY AND NONLINEARITY

2 1
e = 1 + +u z= e = 1 +  2 z + u
g g
3
Employment growth rate

0
0 1 2 3 4 5 6 7 8 9
GDP growth rate

This is nonlinear in g, but if we define z = 1/g, we can rewrite the model to be

linear in variables and parameters.
10
LINEARITY AND NONLINEARITY

Average annual percentage growth rates

e g z e g z

Australia 2.57 3.52 0.2841 Korea 1.11 4.48 0.2235

Austria 1.64 2.66 0.3757 Luxembourg 1.34 4.55 0.2199
Belgium 1.06 2.27 0.4401 Mexico 1.88 3.36 0.2976
Canada 1.90 2.57 0.3891 Netherlands 0.51 2.37 0.4221
Czech Republic 0.79 5.62 0.1781 New Zealand 2.67 3.41 0.2929
Denmark 0.58 2.02 0.4961 Norway 1.36 2.49 0.4013
Estonia 2.28 8.10 0.1234 Poland 2.05 5.16 0.1938
Finland 0.98 3.75 0.2664 Portugal 0.13 1.04 0.9603
France 0.69 2.00 0.5004 Slovak Republic 2.08 7.04 0.1420
Germany 0.84 1.67 0.5980 Slovenia 1.60 4.82 0.2075
Greece 1.55 4.32 0.2315 Sweden 0.83 3.47 0.2885
Hungary 0.28 3.31 0.3021 Switzerland 0.90 2.54 0.3941
Iceland 2.49 5.62 0.1779 Turkey 1.30 6.90 0.1449
Israel 3.29 4.79 0.2089 United Kingdom 0.92 3.31 0.3024
Italy 0.89 1.29 0.7723 United States 1.36 2.88 0.3476
Japan 0.31 1.85 0.5417

The data table displays the values of z, which are derived from the values of
g. In practice, you won’t need to do these calculations manually. Regression
applications typically include features that allow you to generate new
variables based on existing ones. 11
LINEARITY AND NONLINEARITY

. gen z = 1/g
. reg e z
----------------------------------------------------------------------------
Source | SS df MS Number of obs = 31
-----------+------------------------------ F( 1, 29) = 13.68
Model | 5.80515811 1 5.80515811 Prob > F = 0.0009
Residual | 12.3041069 29 .424279548 R-squared = 0.3206
-----------+------------------------------ Adj R-squared = 0.2971
Total | 18.109265 30 .603642167 Root MSE = .65137
----------------------------------------------------------------------------
e | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-----------+----------------------------------------------------------------
z | -2.356137 .6369707 -3.70 0.001 -3.658888 -1.053385
_cons | 2.17537 .249479 8.72 0.000 1.665128 2.685612
----------------------------------------------------------------------------

Here is the output for a regression of e on z.

12
LINEARITY AND NONLINEARITY

eˆ = 2.18 − 2.36 z
3 ------------------------
e | Coef.
-----------+------------
z | -2.356137
Employment growth rate

_cons | 2.17537
2
------------------------

0
0.0 0.2 0.4 0.6 0.8 1.0

z=1/g

-1

The figure shows the transformed data and the regression line for the
regression of e on z.
13
LINEARITY AND NONLINEARITY

2.36
eˆ = 2.18 − 2.36 z = 2.18 −
g
3
Employment growth rate

0
0 1 2 3 4 5 6 7 8 9

-1

-2

GDP growth rate

Substituting 1/g for z, we obtain the nonlinear relationship between e and g.

The figure shows this relationship plotted in the original diagram.
14
LINEARITY AND NONLINEARITY

2.36
eˆ = 2.18 − 2.36 z = 2.18 −
g
3
Employment growth rate

0
0 1 2 3 4 5 6 7 8 9

-1

-2

GDP growth rate

In this case, the relationship between e and g was nonlinear. In multiple

regression analysis, nonlinearity might be detected using the graphical
technique described in a previous slideshow. 15
Introduction to Econometrics
Chapter heading
RAMSEY’S RESET TEST OF
FUNCTIONAL
MISSPECIFICATION
RAMSEY’S RESET TEST OF FUNCTIONAL MISSPECIFICATION

k
Y = 1 +   j X j + u
j =2
k
Yˆ = ˆ1 +  ˆ j X j
j=2

Ramsey’s RESET test of functional misspecification is intended to

provide a simple indicator of evidence of nonlinearity. To implement it,
one runs the regression and saves the fitted values of the dependent
variable.
Since, by definition, the fitted values are a linear combination of the
explanatory variables, Y2 is a linear combination of the squares of the X
variables and their interactions.

1
RAMSEY’S RESET TEST OF FUNCTIONAL MISSPECIFICATION

k
Y = 1 +   j X j + u
j =2
k
Yˆ = ˆ1 +  ˆ j X j
j=2

Add Ŷ 2 to regression specification

^
If Y2 is added to the regression specification, it should pick up quadratic and
interactive nonlinearity, if present, without necessarily being highly correlated
with any of the X variables. 3
RAMSEY’S RESET TEST OF FUNCTIONAL MISSPECIFICATION

k
Y = 1 +   j X j + u
j =2
k
Yˆ = ˆ1 +  ˆ j X j
j=2

Add Ŷ 2 to regression specification

If the t statistic for the coefficient is significant, some kind of nonlinearity

may be present.
4
RAMSEY’S RESET TEST OF FUNCTIONAL MISSPECIFICATION

. reg EARNINGS S EXP

----------------------------------------------------------------------------
Source | SS df MS Number of obs = 500
-----------+------------------------------ F( 2, 497) = 35.24
Model | 8735.42401 2 4367.712 Prob > F = 0.0000
Residual | 61593.5422 497 123.930668 R-squared = 0.1242
-----------+------------------------------ Adj R-squared = 0.1207
Total | 70328.9662 499 140.939812 Root MSE = 11.132
----------------------------------------------------------------------------
EARNINGS | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-----------+----------------------------------------------------------------
S | 1.877563 .2237434 8.39 0.000 1.437964 2.317163
EXP | .9833436 .2098457 4.69 0.000 .5710495 1.395638
_cons | -14.66833 4.288375 -3.42 0.001 -23.09391 -6.242752
----------------------------------------------------------------------------
. predict FITTED
(option xb assumed; fitted values)
. gen FITTEDSQ = FITTED*FITTED

We will do this for a wage equation. Here is the output from a regression of
EARNINGS on S and EXP using EAWE Data Set 21. We save the fitted
values as FITTED and generate FITTEDSQ as the square.
5
RAMSEY’S RESET TEST OF FUNCTIONAL MISSPECIFICATION

. reg EARNINGS S EXP FITTEDSQ

----------------------------------------------------------------------------
Source | SS df MS Number of obs = 500
-----------+------------------------------ F( 3, 496) = 25.46
Model | 9386.33186 3 3128.77729 Prob > F = 0.0000
Residual | 60942.6344 496 122.868214 R-squared = 0.1335
-----------+------------------------------ Adj R-squared = 0.1282
Total | 70328.9662 499 140.939812 Root MSE = 11.085
----------------------------------------------------------------------------
EARNINGS | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-----------+----------------------------------------------------------------
S | -1.334163 1.413072 -0.94 0.346 -4.110507 1.442181
EXP | -.6441233 .7373115 -0.87 0.383 -2.092762 .8045155
FITTEDSQ | .0460798 .0200203 2.30 0.022 .0067447 .0854148
_cons | 25.09321 17.79509 1.41 0.159 -9.86984 60.05626
----------------------------------------------------------------------------

The coefficient of FITTEDSQ is significant at the 5 percent level, indicating

that adding quadratic terms may improve the model’s specification.
6
RAMSEY’S RESET TEST OF FUNCTIONAL MISSPECIFICATION

. reg EARNINGS S EXP FITTEDSQ

However, we also saw that a semilogarithmic specification was better. The

RESET test is intended to detect nonlinearity but not to be specific about
the most appropriate nonlinear model. 7
RAMSEY’S RESET TEST OF FUNCTIONAL MISSPECIFICATION

. reg EARNINGS S EXP FITTEDSQ

It may fail to detect some types of nonlinearity. However, it is very easy to

implement and requires only one degree of freedom.
8
Introduction to Econometrics
Chapter heading
ELASTICITIES AND
LOGARITHMIC MODELS
ELASTICITIES AND LOGARITHMIC MODELS

Definition:
The elasticity of Y with respect to
X is the proportional change in
Y per proportional change in X. A
Y

dY Y
elasticity =
dX X Y
dY dX X
= O0 X 52
Y X

This sequence defines elasticities and shows how to fit nonlinear models
with constant elasticities—first, the general definition of elasticity.
1
ELASTICITIES AND LOGARITHMIC MODELS

Definition:
The elasticity of Y with respect to
X is the proportional change in
Y per proportional change in X. A
Y

dY Y
elasticity =
dX X Y
dY dX X
= O0 X 52
Y X
slope of the tangent at A
=
slope of OA

Re-arranging the expression for the elasticity, we can obtain a graphical

interpretation.
2
ELASTICITIES AND LOGARITHMIC MODELS

Definition:
The elasticity of Y with respect to
X is the proportional change in
Y per proportional change in X. A
Y

dY Y
elasticity =
dX X Y
dY dX X
= O0 X 52
Y X
slope of the tangent at A
=
slope of OA

The elasticity at any point on the curve is the ratio of the slope of the
tangent at that point to the slope of the line joining the point to the origin.
3
ELASTICITIES AND LOGARITHMIC MODELS

Definition:
elasticity < 1
The elasticity of Y with respect to
X is the proportional change in
Y per proportional change in X. A
Y

dY Y
elasticity =
dX X Y
dY dX X
= O0 X 52
Y X
slope of the tangent at A
=
slope of OA

In this case, the tangent at A is clearly flatter than the line OA, so the
elasticity must be less than 1.
4
ELASTICITIES AND LOGARITHMIC MODELS

Definition:
elasticity > 1
The elasticity of Y with respect to
X is the proportional change in
Y per proportional change in X.
A
Y
dY Y
elasticity =
dX X
dY dX
= O0 X 52
Y X
slope of the tangent at A
=
slope of OA

In this case, the tangent at A is steeper than OA, and the elasticity is greater than 1.

5
ELASTICITIES AND LOGARITHMIC MODELS

Y = 1 +  2 X Y

dY dX
elasticity =
Y X A
slope of the tangent at A
=
slope of OA
2
=
( 1 +  2 X ) X
O Xx
2
=
( 1 X ) + 2

The elasticity will generally be different at different points on the function

relating Y to X.
6
ELASTICITIES AND LOGARITHMIC MODELS

Y = 1 +  2 X Y

dY dX
elasticity =
Y X A
slope of the tangent at A
=
slope of OA
2
=
( 1 +  2 X ) X
O Xx
2
=
( 1 X ) + 2

In the example above, Y is a linear function of X.

7
ELASTICITIES AND LOGARITHMIC MODELS

Y = 1 +  2 X Y

dY dX
elasticity =
Y X A
slope of the tangent at A
=
slope of OA
2
=
( 1 +  2 X ) X
O Xx
2
=
( 1 X ) + 2

The tangent at any point is coincidental with the line itself, so its slope is
always b2 in this case. The elasticity depends on the slope of the line
joining the point to the origin. 8
ELASTICITIES AND LOGARITHMIC MODELS

Y = 1 +  2 X Y

dY dX
elasticity = B
Y X A
slope of the tangent at A
=
slope of OA
2
=
( 1 +  2 X ) X
O Xx
2
=
( 1 X ) + 2

OB is flatter than OA, so the elasticity is greater at B than at A. (This ties in

with the mathematical expression: (1 / X) + 2 is smaller at B than at A,
assuming that 1 is positive.) 9
ELASTICITIES AND LOGARITHMIC MODELS

Y = 1 X 2

However, a function of the type shown above has the same elasticity for all
values of X.
10
ELASTICITIES AND LOGARITHMIC MODELS

Y = 1 X 2

dY
=  1  2 X  2 −1
dX

For the numerator of the elasticity expression, we need the derivative of Y

with respect to X.
11
ELASTICITIES AND LOGARITHMIC MODELS

Y = 1 X 2

dY
=  1  2 X  2 −1
dX

Y 1 X 2
= =  1 X  2 −1
X X

For the denominator, we need Y/X.

12
ELASTICITIES AND LOGARITHMIC MODELS

Y = 1 X 2

dY
=  1  2 X  2 −1
dX

Y 1 X 2
= =  1 X  2 −1
X X

d Y d X  1  2 X  2 −1
elasticity = =  2 −1 =  2
Y X 1 X

Hence, we obtain the expression of elasticity. This simplifies to 2 and is

therefore constant.
13
ELASTICITIES AND LOGARITHMIC MODELS

Y
Y = 1 X 2
 2 = 0.25

By way of illustration, the function will be plotted for a range of values of 2.
We will start with a very low value, 0.25.
14
ELASTICITIES AND LOGARITHMIC MODELS

Y
Y = 1 X 2
 2 = 0.50

We will increase 2 in steps of 0.25 and see how the shape of the function
changes.
15
ELASTICITIES AND LOGARITHMIC MODELS

Y
Y = 1 X 2
 2 = 0.75

2 = 0.75.

16
ELASTICITIES AND LOGARITHMIC MODELS

Y
Y = 1 X 2
 2 = 1.00

When 2 equals 1, the curve becomes a straight line through the origin.

17
ELASTICITIES AND LOGARITHMIC MODELS

Y
Y = 1 X 2
 2 = 1.25

2 = 1.25.

18
ELASTICITIES AND LOGARITHMIC MODELS

Y
Y = 1 X 2
 2 = 1.50

2 = 1.50.

19
ELASTICITIES AND LOGARITHMIC MODELS

Y
Y = 1 X 2
 2 = 1.75

2 = 1.75. Note that the curvature can be quite gentle over wide ranges of X.

20
ELASTICITIES AND LOGARITHMIC MODELS

Y
Y = 1 X 2
 2 = 1.75

This means that even if the true model is of the constant elasticity form, a
linear model may be a good approximation over a limited range.
21
ELASTICITIES AND LOGARITHMIC MODELS

Y = 1 X 2

log Y = log  1 X  2
= log  1 + log X  2
= log  1 +  2 log X

Fitting a constant elasticity function using a sample of observations is easy.

You can linearize the model by taking the logarithms of both sides.
22
ELASTICITIES AND LOGARITHMIC MODELS

Y = 1 X 2

log Y = log  1 X  2
= log  1 + log X  2
= log  1 +  2 log X

Y ' =  1' +  2 X ' where Y ' = log Y ,

X ' = log X
 1' = log  1

You thus obtain a linear relationship between Y' and X', as defined. All serious
regression applications allow you to generate logarithmic variables from
existing ones. 23
ELASTICITIES AND LOGARITHMIC MODELS

Y = 1 X 2

log Y = log  1 X  2
= log  1 + log X  2
= log  1 +  2 log X

Y ' =  1' +  2 X ' where Y ' = log Y ,

X ' = log X
 1' = log  1

The coefficient of X' will be a direct estimate of the elasticity, 2.

24
ELASTICITIES AND LOGARITHMIC MODELS

Y = 1 X 2

log Y = log  1 X  2
= log  1 + log X  2
= log  1 +  2 log X

Y ' =  1' +  2 X ' where Y ' = log Y ,

X ' = log X
 1' = log  1

The constant term will be an estimate of log 1. To obtain an estimate of 1,
calculate exp( ˆ1' ), where ̂ 1' is the estimate of  1' . (This assumes that you have
used natural logarithms, that is, logarithms to base e, to transform the model.)
25
ELASTICITIES AND LOGARITHMIC MODELS

FDHO

7000

6000

5000

4000

3000

2000

1000

0
0 10000 20000 30000 40000 50000 EXP

Here is a scatter diagram showing annual household expenditure on FDHO,

food eaten at home, and EXP, total annual household expenditure, both
measured in dollars, for 1995 for a sample of 869 households in the United
States (Consumer Expenditure Survey data). 26
ELASTICITIES AND LOGARITHMIC MODELS

. reg FDHO EXP

----------------------------------------------------------------------------
Source | SS df MS Number of obs = 6334
-----------+------------------------------ F( 1, 6332) = 3431.01
Model | 972602566 1 972602566 Prob > F = 0.0000
Residual | 1.7950e+09 6332 283474.003 R-squared = 0.3514
-----------+------------------------------ Adj R-squared = 0.3513
Total | 2.7676e+09 6333 437006.15 Root MSE = 532.42
----------------------------------------------------------------------------
FDHO | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-----------+----------------------------------------------------------------
EXP | .0627099 .0010706 58.57 0.000 .0606112 .0648086
_cons | 369.4418 10.65718 34.67 0.000 348.5501 390.3334
----------------------------------------------------------------------------

Here is a linear regression of FDHO on EXP. When using household data, it

is usual to relate types of consumer expenditures to total expenditures rather
than income. Household income data tend to be relatively erratic.
27
ELASTICITIES AND LOGARITHMIC MODELS

. reg FDHO EXP

The regression implies that, at the margin, 6.3 cents out of each dollar of
expenditure is spent on food at home. Does this seem plausible? Probably,
though a little low. 28
ELASTICITIES AND LOGARITHMIC MODELS

. reg FDHO EXP

It also suggests that $369 would be spent on food at home if total expenditure
were zero. This is impossible. It may be possible to interpret it as baseline
expenditure, but we must consider family size and composition.
29
ELASTICITIES AND LOGARITHMIC MODELS

FDHO

7000

6000

5000

4000

3000

2000

1000

0
0 10000 20000 30000 40000 50000 EXP

Here is the regression line plotted on the scatter diagram

30
ELASTICITIES AND LOGARITHMIC MODELS

LGFDHO
10

1
6 7 8 9 10 11 LGEXP

We will now fit a constant elasticity function using the same data. The scatter
diagram shows the FDHO logarithm plotted against the EXP logarithm.
31
ELASTICITIES AND LOGARITHMIC MODELS

. g LGFDHO = ln(FDHO)
. g LGEXP = ln(EXP)
. reg LGFDHO LGEXP
----------------------------------------------------------------------------
Source | SS df MS Number of obs = 6334
-----------+------------------------------ F( 1, 6332) = 4719.99
Model | 1642.9356 1 1642.9356 Prob > F = 0.0000
Residual | 2204.04385 6332 .348080204 R-squared = 0.4271
-----------+------------------------------ Adj R-squared = 0.4270
Total | 3846.97946 6333 .60744978 Root MSE = .58998
----------------------------------------------------------------------------
LGFDHO | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-----------+----------------------------------------------------------------
LGEXP | .6657858 .0096909 68.70 0.000 .6467883 .6847832
_cons | .7009498 .0843607 8.31 0.000 .5355741 .8663254
----------------------------------------------------------------------------

Here is the result of regressing LGFDHO on LGEXP. The first two commands
generate the logarithmic variables.
32
ELASTICITIES AND LOGARITHMIC MODELS

The estimate of the elasticity is 0.67. Does this seem plausible?

33
ELASTICITIES AND LOGARITHMIC MODELS

Yes, definitely. Food is a normal good, so its elasticity should be positive;

however, it is also a basic necessity. Expenditure on food should grow less
rapidly than overall expenditure, indicating that its elasticity is less than 1. 34
ELASTICITIES AND LOGARITHMIC MODELS

ˆ
LGFDHO ˆ = 2.02 EXP 0.666
= 0.701 + 0.666 LGEXP  FDHO

The intercept has no substantive meaning. To obtain an estimate of 1, we

calculate e0.701, which is 2.02.
35
ELASTICITIES AND LOGARITHMIC MODELS

LGFDHO
10

1
6 7 8 9 10 11 LGEXP

Here is the scatter diagram with the regression line plotted.

36
ELASTICITIES AND LOGARITHMIC MODELS

FDHO

7000

6000

5000

4000

3000

2000

1000

0
0 10000 20000 30000 40000 50000 EXP

Here is the regression line from the logarithmic regression plotted in the
original scatter diagram and the linear regression line for comparison.
37
ELASTICITIES AND LOGARITHMIC MODELS

FDHO

7000

6000

5000

4000

3000

2000

1000

0
0 10000 20000 30000 40000 50000 EXP

The logarithmic regression line gives a somewhat better fit, especially at low
expenditure levels.
38
ELASTICITIES AND LOGARITHMIC MODELS

FDHO

7000

6000

5000

4000

3000

2000

1000

0
0 10000 20000 30000 40000 50000 EXP

However, the difference in the fit is not dramatic. The main reason for
preferring the constant elasticity model is that it makes more sense
theoretically. It also has a technical advantage that we will discuss later
when we discuss heteroskedasticity. 39
Introduction to Econometrics
Chapter heading
SEMILOGARITHMIC MODELS
SEMILOGARITHMIC MODELS

Y =  1e  2 X

This sequence introduces the semilogarithmic model and shows how it may
be applied to an earnings function. The dependent variable is linear, but the
explanatory variables, multiplied by their coefficients, are exponents of e. 1
SEMILOGARITHMIC MODELS

Y =  1e  2 X

dY
=  1  2 e  2 X =  2Y
dX

The differential of Y with respect to X simplifies to 2Y.

2
SEMILOGARITHMIC MODELS

Y =  1e  2 X

dY
=  1  2 e  2 X =  2Y
dX

dY Y
= 2
dX

Hence, the proportional Y per unit change in X equals 2. It is, therefore,
independent of the value of X.
3
SEMILOGARITHMIC MODELS

Y =  1e  2 X

Y +  Y =  1 e  2 ( X + X )
=  1 e  2 X e  2 X
= Ye  2 X
 (  2 X )2

= Y  1 +  2 X + + ... 
 2 

Strictly speaking, this interpretation is valid only for small values of 2.
When 2 is not small, the interpretation may be a little more complex.
4
SEMILOGARITHMIC MODELS

Y =  1e  2 X

Y +  Y =  1 e  2 ( X + X )
=  1 e  2 X e  2 X
= Ye  2 X
 (  2 X )2

= Y  1 +  2 X + + ... 
 2 

Suppose that X increases by an amount of X and that, as a consequence Y

increases by an amount of Y.
5
SEMILOGARITHMIC MODELS

Y =  1e  2 X

Y +  Y =  1 e  2 ( X + X )
=  1 e  2 X e  2 X
= Ye  2 X
 (  2 X )2

= Y  1 +  2 X + + ... 
 2 

We can rewrite the right side of the equation as shown.

6
SEMILOGARITHMIC MODELS

Y =  1e  2 X

Y +  Y =  1 e  2 ( X + X )
=  1 e  2 X e  2 X
= Ye  2 X
 (  2 X )2

= Y  1 +  2 X + + ... 
 2 

We can simplify the right side of the equation as shown.

7
SEMILOGARITHMIC MODELS

Y =  1e  2 X

Y +  Y =  1 e  2 ( X + X )
=  1 e  2 X e  2 X
= Ye  2 X
 (  2 X )2

= Y  1 +  2 X + + ... 
 2 
Z2 Z3
e = 1+ Z +
Z
+ + ...
2! 3!

Now expand the exponential term using the standard expression for e to
some power.
8
SEMILOGARITHMIC MODELS

Y =  1e  2 X

Y +  Y =  1 e  2 ( X + X )
=  1 e  2 X e  2 X
= Ye  2 X
 (  2 X )2

= Y  1 +  2 X + + ... 
 2 

 (  2 X )
2

Y = Y   2 X + + ... 
 2 

Subtract Y from both sides.

9
SEMILOGARITHMIC MODELS

 (  2 X )2

Y = Y   2 X + + ... 
 2 

( 2 X )2 negligible

We now consider two cases: where 2 and X are so small that (2 X)2 is
negligible, and the alternative.
10
SEMILOGARITHMIC MODELS

 (  2 X )2

Y = Y   2 X + + ... 
 2 

( 2 X )2 negligible

Y = Y 2 X

Y / Y
= 2
X

If (2 X)2 is negligible, we obtain the same interpretation of 2 as we did

using the calculus, as expected.
11
SEMILOGARITHMIC MODELS

 (  2 X )2

Y = Y   2 X + + ... 
 2 

( 2 X )2 not negligible

Y / Y  22 X
= 2 + + ...
X 2
 22
= 2 + + ...
2

If (2 X)2 is not negligible, the proportional change in Y given a X change

in X has an extra term. (We are assuming that 2 and X are small enough
that terms with higher powers of X can be neglected.) 12
SEMILOGARITHMIC MODELS

 (  2 X )2

Y = Y   2 X + + ... 
 2 

( 2 X )2 not negligible

Y / Y  22 X
= 2 + + ...
X 2
 22
= 2 + + ... if X is one unit
2

Usually we talk about the effect of a one-unit change in X. If X = 1, the

proportional change in Y is as shown. The issue now becomes whether 2 is
so small that the second and subsequent terms can be neglected. 13
SEMILOGARITHMIC MODELS

Y =  1e  2 X

X = 0  Y =  1e 0 =  1

1 is the value of Y when X is equal to zero (note that e0 is equal to 1).

14
SEMILOGARITHMIC MODELS

Y =  1e  2 X

log Y = log  1e  2 X
= log  1 + log e  2 X
=  1' +  2 X log e
=  1' +  2 X

To fit a function of this type, you take logarithms of both sides. The right side
of the equation becomes a linear function of X (note that the logarithm of e, to
base e, is 1). Hence we can fit the model with a linear regression, regressing
log Y on X. 15
SEMILOGARITHMIC MODELS

. reg LGEARN S
----------------------------------------------------------------------------
Source | SS df MS Number of obs = 500
-----------+------------------------------ F( 1, 498) = 60.71
Model | 16.5822819 1 16.5822819 Prob > F = 0.0000
Residual | 136.016938 498 .273126381 R-squared = 0.1087
-----------+------------------------------ Adj R-squared = 0.1069
Total | 152.59922 499 .30581006 Root MSE = .52261
----------------------------------------------------------------------------
LGEARN | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-----------+----------------------------------------------------------------
S | .0664621 .0085297 7.79 0.000 .0497034 .0832207
_cons | 1.83624 .1289384 14.24 0.000 1.58291 2.089571
----------------------------------------------------------------------------

Here is the regression output from a wage equation regression using Data Set.
The estimate of 2 is 0.066. As an approximation, this implies that an extra
year of schooling increases hourly earnings by a proportion of 0.066. 16
SEMILOGARITHMIC MODELS

In everyday language, it is usually more natural to talk about percentages

rather than proportions, so we multiply the coefficient by 100. This implies
that an extra year of schooling increases hourly earnings by 6.6%. 17
SEMILOGARITHMIC MODELS

(  2 X )
2
not negligible

Y / Y  2
( 0.066 ) 2

If X is one unit, = 2 + 2
+ ... = 0.066 + = 0.068
X 2 2

If we consider that a year of schooling is not a marginal change and work out
the effect exactly, the proportional increase is 0.068, and the percentage
increase is 6.8%. 18
SEMILOGARITHMIC MODELS

(  2 X )
2
not negligible

Y / Y  2
( 0.066 ) 2

If X is one unit, = 2 + + ... = 0.066 +

2
= 0.068
X 2 2

In general, if a unit change in X is genuinely marginal, the estimate of 2 will be

small, and one can interpret it directly as an estimate of the proportional
change in Y per unit change in X. 19
SEMILOGARITHMIC MODELS

(  2 X )
2
not negligible

Y / Y  2
( 0.066 ) 2

If X is one unit, = 2 + + ... = 0.066 +

2
= 0.068
X 2 2

However, if a unit change in X is not small, the coefficient may be large, and the
second term might not be negligible. In the present case, a year of schooling is
not marginal, but even so, the refinement makes only a small difference.
20
SEMILOGARITHMIC MODELS

(  2 X )
2
not negligible

Y / Y  2
( 0.066 ) 2

If X is one unit, = 2 + + ... = 0.066 +

2
= 0.068
X 2 2

In general, when 2 is less than 0.1, working out the effect exactly can be of
little benefit.
21
SEMILOGARITHMIC MODELS

. reg LGEARN S
----------------------------------------------------------------------------
Source
Source | | SSSS dfdf MSMS Number
Number ofof obs
obs = = 540
500
-------------+------------------------------
-----------+------------------------------ F(F(1,1, 498)
538)
= = 60.71
140.05
Model
Model | |16.5822819
38.5643833 1 116.5822819
38.5643833 Prob
Prob > >
F F = =0.0000
0.0000
Residual
Residual | |136.016938
148.14326 498538.273126381
.275359219 R-squared
R-squared = =0.1087
0.2065
-------------+------------------------------
-----------+------------------------------ Adj
Adj R-squared
R-squared = =0.1069
0.2051
Total
Total | | 152.59922
186.707643 499539 .30581006
.34639637 Root
Root MSE
MSE = =.52261
.52475
----------------------------------------------------------------------------
------------------------------------------------------------------------------
LGEARN | Coef. Std. Err. t P>|t| [95% Conf. Interval]
LGEARN | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-----------+----------------------------------------------------------------
-------------+----------------------------------------------------------------
S | .0664621 .0085297 7.79 0.000 .0497034 .0832207
S
_cons | | .1096934
1.83624 .0092691
.1289384 11.83
14.24 0.000
0.000 .0914853
1.58291 .1279014
2.089571
_cons | 1.292241 .1287252 10.04 0.000 1.039376 1.545107
----------------------------------------------------------------------------
------------------------------------------------------------------------------

log ˆ1 = 1.836

ˆ1 = e1.836 = 6.27

The intercept in the regression is an estimate of log 1. From it, we obtain
an estimate of 1 equal to e1.836, which is 6.27.
22
SEMILOGARITHMIC MODELS

log ˆ1 = 1.836

ˆ1 = e1.836 = 6.27

This literally implies that a person with no schooling would earn $6.27 per hour.
However, it is dangerous to extrapolate so far from the range for which we have
data. 23
SEMILOGARITHMIC MODELS

Logarithm of hourly earnings

0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Years of schooling (highest grade completed)

Here is the scatter diagram with the semilogarithmic regression.

24
SEMILOGARITHMIC MODELS

120

100
Hourly earnings ($)

0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Years of schooling (highest grade completed)

Here is the semilogarithmic regression line plotted in a scatter diagram with

the untransformed data, with the linear regression shown for comparison.
25
SEMILOGARITHMIC MODELS

120

100
Hourly earnings ($)

0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Years of schooling (highest grade completed)

The fit of the regression lines is not much different, but the semilogarithmic
regression is more satisfactory in two respects.
26
SEMILOGARITHMIC MODELS

120

100
Hourly earnings ($)

0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Years of schooling (highest grade completed)

The linear specification predicts that hourly earnings will increase by a fixed
amount, $1.27, with each additional year of schooling. This is implausible for
high levels of education. The semi-logarithmic specification allows the
increment to increase with the level of education. 27
SEMILOGARITHMIC MODELS

120

100
Hourly earnings ($)

0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Years of schooling (highest grade completed)

Second, the linear specification predicts very low earnings for an individual
with no schooling. The semilogarithmic specification predicts hourly
earnings of $6.27, which at least is not obvious nonsense. 28
SUMMARY OF THE DIFFERENT NONLINEAR REGRESSION MODELS

93
SUMMARY OF THE INTERPRETATION
COEFFICIENTS OF DIFFERENT NONLINEAR
REGRESSION MODELS

X Y
X X+1 MODEL Y = f(X) Y = f(X+1) CHANGE CHANGE
100 101 Y=3+5X 503 508 1 units 5 units

100 101 LN(Y) = 2 + 0.08 LN(X) 10.68 10.69 1% 0.08%

100 101 LN(Y) = 0.2 + 0.04 X 66.7 69.4 1 units 4.08%

100 101 Y = 3 + 0.2 LN(X) 3.92 3.92 1% 0.002 units

100(exp(b)-1)% =
100(e0.04-1)% = 4.08%
Introduction to Econometrics
Chapter heading
THE DISTURBANCE TERM IN
LOGARITHMIC MODELS
THE DISTURBANCE TERM IN LOGARITHMIC MODELS

2
Y = 1 + +u
X
1
Z=
X

Y = 1 +  2 Z + u

Thus far, nothing has been said about the disturbance term in nonlinear
regression models.
1
THE DISTURBANCE TERM IN LOGARITHMIC MODELS

2
Y = 1 + +u
X
1
Z=
X

Y = 1 +  2 Z + u

For the regression results in a linearized model to have the desired properties,
the disturbance term in the transformed model should be additive and satisfy
the regression model conditions. 2
THE DISTURBANCE TERM IN LOGARITHMIC MODELS

2
Y = 1 + +u
X
1
Z=
X

Y = 1 +  2 Z + u

It should be normally distributed in the transformed model to perform the

usual tests.
3
THE DISTURBANCE TERM IN LOGARITHMIC MODELS

2
Y = 1 + +u
X
1
Z=
X

Y = 1 +  2 Z + u

In the case of the first example of a nonlinear model, there was no problem.
If the disturbance term had the required properties in the original model, it
would have them in the regression model. It has not been affected by the
transformation. 4
THE DISTURBANCE TERM IN LOGARITHMIC MODELS

Y = 1 X 2 e u = 1 X 2 v

log Y = log  1 +  2 log X + u

In the discussion of the logarithmic model, the disturbance term was

omitted altogether.
5
THE DISTURBANCE TERM IN LOGARITHMIC MODELS

Y = 1 X 2 e u = 1 X 2 v

log Y = log  1 +  2 log X + u

However, implicitly it was being assumed that there was an additive

disturbance term in the transformed model.
6
THE DISTURBANCE TERM IN LOGARITHMIC MODELS

Y = 1 X 2 e u = 1 X 2 v

log Y = log  1 +  2 log X + u

For this to be possible, the random component in the original model must be
a multiplicative term, eu.
7
THE DISTURBANCE TERM IN LOGARITHMIC MODELS

Y = 1 X 2 e u = 1 X 2 v

log Y = log  1 +  2 log X + u

We will denote this multiplicative term v.

8
THE DISTURBANCE TERM IN LOGARITHMIC MODELS

Y = 1 X 2 e u = 1 X 2 v

log Y = log  1 +  2 log X + u

When u is equal to 0, not modifying the value of log Y, v is equal to 1,

likewise not modifying the value of Y.
9
THE DISTURBANCE TERM IN LOGARITHMIC MODELS

Y = 1 X 2 e u = 1 X 2 v

log Y = log  1 +  2 log X + u

Positive values of u correspond to values of v greater than 1, the random factor

having a positive effect on Y and log Y. Likewise, negative values of u
correspond to values of v between 0 and 1, the random factor having a negative
impact on Y and log Y. 10
THE DISTURBANCE TERM IN LOGARITHMIC MODELS

f(v)
0.45

0.40 Y = 1 X 2 e u = 1 X 2 v
0.35
log Y = log  1 +  2 log X + u
0.30

0.25

0.20

0.15

0.10

0.05

0.00
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 v
16

Besides satisfying the regression model conditions, we need u to be

normally distributed if we are to perform t tests and F tests.
11
THE DISTURBANCE TERM IN LOGARITHMIC MODELS

f(v)
0.45

0.40 Y = 1 X 2 e u = 1 X 2 v
0.35
log Y = log  1 +  2 log X + u
0.30

0.25

0.20

0.15

0.10

0.05

0.00
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 v
16

This will be the case if v has a lognormal distribution, shown above.

12
THE DISTURBANCE TERM IN LOGARITHMIC MODELS

f(v)
0.45

0.40 Y = 1 X 2 e u = 1 X 2 v
0.35
log Y = log  1 +  2 log X + u
0.30

0.25

0.20

0.15

0.10

0.05

0.00
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 v
16

The mode of the distribution is located at v = 1, where u = 0.

13
THE DISTURBANCE TERM IN LOGARITHMIC MODELS

f(v)
0.45

0.40 Y =  1e  2 X e u =  1e  2 X v
0.35
log Y = log  1 +  2 X + u
0.30

0.25

0.20

0.15

0.10

0.05

0.00
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 v
16

The same multiplicative disturbance term is needed in the semilogarithmic

model.
14
THE DISTURBANCE TERM IN LOGARITHMIC MODELS

f(v)
0.45

0.40 Y =  1e  2 X e u =  1e  2 X v
0.35
log Y = log  1 +  2 X + u
0.30

0.25

0.20

0.15

0.10

0.05

0.00
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 v
16

Note that, with this distribution, one should expect a small proportion of
observations to be subject to large positive random effects.
15
THE DISTURBANCE TERM IN LOGARITHMIC MODELS

120

100
Hourly earnings ($)

0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Years of schooling (highest grade completed)

Here is the scatter diagram for earnings and schooling using Data Set 21.
You can see that there are several outliers, with the three most extreme
highlighted. 16
THE DISTURBANCE TERM IN LOGARITHMIC MODELS

Logarithm of hourly earnings

0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Years of schooling (highest grade completed)

Here is the scatter diagram for the semilogarithmic model with its regression
line. The same three observations remain outliers, but they do not appear to
be so extreme. 17
THE DISTURBANCE TERM IN LOGARITHMIC MODELS

160

140

120

100

0–3 –2 –1 0 1 2 3
-2.75 to -2.25 -2.25 to 1.75 -1.75 to -1.25 -1.25 to -0.75 -0.75 to -0.25 -0.25 to 0.25 0.25 to 0.75 0.75 to 1.25 1.25 to 1.75 1.75 to 2.25 2.25 to 2.75

Residuals (linear) Residuals (semilogarithmic)

The histogram above compares the distributions of the residuals from the linear
and semi-logarithmic regressions. To make them comparable, the distributions
have been standardized, that is, scaled so that their standard deviation is equal
18
to 1.
THE DISTURBANCE TERM IN LOGARITHMIC MODELS

160

140

120

100

0–3 –2 –1 0 1 2 3
-2.75 to -2.25 -2.25 to 1.75 -1.75 to -1.25 -1.25 to -0.75 -0.75 to -0.25 -0.25 to 0.25 0.25 to 0.75 0.75 to 1.25 1.25 to 1.75 1.75 to 2.25 2.25 to 2.75

Residuals (linear) Residuals (semilogarithmic)

It can be shown that if the disturbance term in a regression model has a

normal distribution, so will the residuals.
19
THE DISTURBANCE TERM IN LOGARITHMIC MODELS

160

140

120

100

0–3 –2 –1 0 1 2 3
-2.75 to -2.25 -2.25 to 1.75 -1.75 to -1.25 -1.25 to -0.75 -0.75 to -0.25 -0.25 to 0.25 0.25 to 0.75 0.75 to 1.25 1.25 to 1.75 1.75 to 2.25 2.25 to 2.75

Residuals (linear) Residuals (semilogarithmic)

Obviously, the residuals from the semilogarithmic regression are approximately

normal, but those from the linear regression are not. This is evidence that the
semi-logarithmic model is the better specification. 20
THE DISTURBANCE TERM IN LOGARITHMIC MODELS

Y = 1 X 2 + u

What would happen if the disturbance term in the logarithmic or semilogarithmic

model were additive rather than multiplicative?
21
THE DISTURBANCE TERM IN LOGARITHMIC MODELS

Y = 1 X 2 + u

log Y = log (  1 X  2 + u )

If this were the case, we could not linearize the model by taking logarithms.
There is no way of simplifying log (  1 X  + u ) . We should have to use some
2

nonlinear regression techniques. 22

Introduction to Econometrics
Chapter heading
COMPARING LINEAR AND
LOGARITHMIC
SPECIFICATIONS
COMPARING LINEAR AND LOGARITHMIC SPECIFICATIONS