LECTURE 7:

MORE ON FUNCTIONAL FORMS

Jan Zouhar Introductory Econometrics
 What transforms do we use, and when?
2

 we already know that linear regression can be used to describe non-

linear relationships (we’ve been using logs routinely, after all)
 there is a plethora of functional transforms one can think of, but
practitioners mostly restrict themselves to the following four

transform formula description

Units Only used as a matter of convenience (to make
x/1000
change results easier to read).
Changes interpreted on a relative scale. May
Logs log(x) help reduce the effect of outliers (CEO salary
example).
Allows for a u-shaped or inverted-u-shaped
Squares x2
relationship (as in age vs wage).
Effect of x1 depends on the level of x2
Interactions x1 · x 2
and vice versa.

Introductory Econometrics Jan Zouhar

 More on the use of logarithms
3

 remember we used the following approximation:

change in log(y) ≈ relative change in y
 relative changes are a bit tricky: if my wage increases by 50% next
month, and decreases by 50% the following month, the total effect is a
drop of 25%
wage × 1.5 × 0.5 = 0.75wage
 consider a country where the average wage is 100 for men and 125 for
women; then
 women earn by 25% more than men
 men earn less by 20% less than women
 in other words, the base category (men or women) matters
 as we know, in regressions it does not (see next slide); is there anything
wrong?

Introductory Econometrics Jan Zouhar

 OLS estimates
Dependent variable: l_wage

(1) (2)

const 0.4317** 0.08352 Intercept has

(0.1045) (0.1011) changed, why?
educ 0.08584** 0.08584**
(0.007183) (0.007183) Coefficients on
exper 0.009691** 0.009691** other variables
(0.001433) (0.001433) unaffected by the
smsa 0.1592** 0.1592** base category
(0.04241) (0.04241)

female -0.3482** Different base

(0.03722) categories, only
male 0.3482** the sign has
(0.03722) changed
n 526 526
R-squared 0.3696 0.3696
lnL -292.1 -292.1

4 Introductory Econometrics Jan Zouhar

 More on the use of logarithms (cont’d)
5

 βfemale model (1) equals −βmale in model (2)

 interpreting this the usual way,
 women earn by 35% less than men
 men earn less by 35% more than women
 but: 0.65 × 1.35 = 0.88 ≠ 1
 in fact, there is no inconsistence, all of this is due to our approximate
interpretation of the logarithm, which only works for small changes (in
the log, or small relative changes)
 Exact interpretation: if e.g. log(wage )  0  1educ  2 female  u ,
exponentiating both sides, and writing down for men and women yields
women: wage  exp( 0  1educ  2  u )
men: wage  exp( 0  1educ  u )
 wage for men = exp(β2) × wage for women

Introductory Econometrics Jan Zouhar

 More on the use of logarithms (cont’d)
6

 interpreting the results in our previous Gretl output:

 exp(0.35) = 1.42, men earn by 42% more than women
 exp(−0.35) = 0.70, women earn by 30% less than men
 note that this solves the apparent inconsistency, as 1.42 × 0.7 = 1; or, in
general,

exp(  female )  exp( male )  exp(  female  male )

 exp(  male  male )
 exp(0)
1

 to conclude, the exact relative change in y due to a unit change in xj is

y / y  exp(  j )  1, or
%y  100[exp(  j )  1]

Introductory Econometrics Jan Zouhar

 Interactions
7

Example: returns to schooling for men and women

 one might argue that in many countries, returns to schooling differ for
men and women
 What do you think is the case in your country?
 Are there any objective reasons why women should be rewarded
more/less for their education that men?
 How do we formulate a model that allows the effect of education to
vary with gender?
 as the title suggests, the answer to the last question lies in interactions
wage  0  1educ  2 female  u (1)
wage  0  1educ  2 female  3 female  educ  u (2)
wage
 It is easily seen that the effect of additional year of education, , is
educ
 β1 in equation (1)
 β1 + β3female in equation (2)

Introductory Econometrics Jan Zouhar

 Interactions (cont’d)
8

wage  0  1educ  2 female  u

Men: E wage = β0 + β1educ

expected wage

β1
Women: E wage = (β0 + β2) + β1educ
β1

β0 1

β0 + β2
educ

Introductory Econometrics Jan Zouhar

 Interactions (cont’d)
9

wage  0  1educ  2 female  3 female  educ  u

Men: E wage = β0 + β1educ

expected wage

β1
Women: E wage = (β0 + β2) + (β1+ β3)educ

β1 + β3
β0
1
β0 + β2
educ

Introductory Econometrics Jan Zouhar

 Squares
10

 allow for a changing sign of the relationship

 note that while logarithms are a non-linear transform, they do not allow
the relationship to change sign (log is strictly increasing)
 many nonlinear functions allow this, but the quadratic is the simplest
one → hardly ever we use anything beyond that

unemployment
probability expected wage
“inverted u” shape
“u” shape

turning point age turning point age

Introductory Econometrics Jan Zouhar

 Squares (cont’d)
11

Example
 wage vs. work experience
 we estimate
wage  0  1exper  2exper 2  u

 In Gretl: first we need to create a new variable containing squared

experience (Add → Squares of selected variables)
 the estimated equation (using Wooldridge’s wage1 data) is:

^wage = 3.73 + 0.298*exper - 0.00613*sq_exper

(0.346)(0.0410) (0.000903)

n = 526, R-squared = 0.093

(standard errors in parentheses)

 Quizz: is this a u or an inverted-u curve? Where is the turning point?

Introductory Econometrics Jan Zouhar

 Squares (cont’d)
12

 a plot may help answer these questions

(Graphs → Fitted, Actual plot → Against exper)

 but the turning point will not be guessed accurately from the plot, and
the plot looks ugly if we include control variables
Introductory Econometrics Jan Zouhar
 Where exactly is the turning point?
13

 use first-order conditions for a maximum/minimum of a function

 differentiate the equation wage  0  1exper  2exper2  u
with respect to exper and set equal to zero:

wage
 1  22exper  0
exper

1
 so the turning point is: exper  
22

 our estimate of the turning point (based on the estimated equation) is

coefficient on the linear term

estimated turning point  
2  coefficient on the squared term

0.298
 in our example, this is exper    24.3 years
2  0.00613 

Introductory Econometrics Jan Zouhar

 More on squares
14

 u or inverted-u shape? Determined by the sign of the coefficient on the

squared term (positive → u; negative → inverted u)
 partial effect of experience:

wage wage
  1  22exper, so wage   1  22exper  exper
exper exper

 in particular, the change in wage brought about by a unit increase in

experience (Δexper = 1) is 1  22exper
 now wait, we used to log the wage in most regressions
 fortunately, log is an increasing function, log(wage) increases whenever
wage does, so our turning point formulas work even for

log(wage )  0  1exper  2exper 2  u

 partial effect:  log(wage )   1  22exper  exper , so

%wage  100  1  22exper  exper
Introductory Econometrics Jan Zouhar
 Tests for functional form misspecification
15

 even if we restrict ourselves to squares, logs, and interactions, there’s

many different functional forms we can produce with given variables;
how do we choose?
 the most important criteria are: (i) our research question and the
underlying economic theory, and (ii) the desired interpretation of the
parameters (see Slide 2 of this presentation)
 even though some statistical tests have been developed to detect
functional form misspecification, we should use them sparingly: they can
lead to overspecified (= overly complicated) models that do not interpret
easily
Using F-tests
 it is straightforward to check for the omission of squares and
interactions in a particular model using an F-test
 just add squares and/or interactions of the regressors and use the F-test
for joint significance

Introductory Econometrics Jan Zouhar

 Tests for functional form misspecification (cont’d)
16

Ramsey’s RESET test

 a popular test for general functional form misspecification
 procedure:
1. First, use OLS to estimate your equation, say
y  0  1x1    k x k  u.

2. Save the fitted values, ŷ.

3. Estimate the equation

y  0  1x1    k x k  1 yˆ 2   2 yˆ 3  u

and use the F-test for joint significance of ŷ2 and ŷ3.
 note that ŷ2 and ŷ3 are themselves functions of cubes, squares, and
interactions of the x’s, but using ŷ2 and ŷ3 instead of all possible
interactions and squares saves up on degrees of freedom dramatically

Introductory Econometrics Jan Zouhar

 LECTURE 7:
MORE ON FUNCTIONAL FORMS
Jan Zouhar Introductory Econometrics