Functional form

Quantitative Methods in Economics

Functional form

Maximilian Kasy

Harvard University, fall 2016

1 / 15
Functional form

Roadmap, Part I

1. Linear predictors and least squares regression

2. Conditional expectations
3. Some functional forms for linear regression
4. Regression with controls and residual regression
5. Panel data and generalized least squares

2 / 15
Functional form

Takeaways for these slides

Functional forms:
I Quadratic: decreasing or increasing returns
I Interactions: returns vary with covariates
I Discrete regressors, dummy variables, and saturated regressions
I Polynomial
I Linear in logarithms: elasticities
I Justification via Mincer model

3 / 15
Functional form

I Quadratic polynomial:
I Y = earnings, Z = experience; X1 = Z , X2 = Z 2

E ∗ (Y | 1, X1 , X2 ) = β0 + β1 X1 + β2 X2

I Evaluating this at Z = c gives β0 + β1 c + β2 c 2 .

I Interactions:
I Z1 = experience, Z2 = education; X1 = Z1 , X2 = Z2 , X3 = Z1 · Z2

E ∗ (Y | 1, X1 , X2 , X3 ) = β0 + β1 X1 + β2 X2 + β3 X3

I Evaluating this at Z1 = c, Z2 = d gives β0 + β1 c + β2 d + β3 c · d.

4 / 15
Functional form

Questions for you

I Interpret these two functional forms.

I What happens as education is increased?
I How does that depend on the education we start with?
I How does that depend on experience?

5 / 15
Functional form

I Recall conditional expectation: Solution to

min E [Y − g (Z )]2
g

is the regression function:

r (z ) = E (Y | Z = z ).

I Orthogonality Conditions: Consider any function h(·). Define

U = Y − r (Z ).

I Then U ⊥ h(Z ), i.e. E [Uh(Z )] = 0, and in particular

E ∗ (Y |r (Z ), h(Z )) = β1 r (Z ) + β2 h(Z ) = r (Z ).

I Put differently: If E (Y |X = x ) is linear in x, then

E (Y |X = x ) = E ∗ (Y |X = x ).
6 / 15
Functional form

Discrete regressors
I Assume
Z1 ∈ {λ1 , . . . , λJ }, Z2 ∈ {δ1 , . . . , δK }.
I Dummy Variables:
(
1, ifZ1 = λj , Z2 = δk
Xjk =
0, otherwise
= 1(Z1 = λj , Z2 = δk ).

I Claim: E (Y | Z1 , Z2 ) = E ∗ (Y | X11 , . . . , XJ1 , . . . , X1K , . . . , XJK )

Questions for you

Prove this.

7 / 15
Functional form

Solution:
I Any function g (Z1 , Z2 ) can be written as

J K
g (Z1 , Z2 ) = ∑ ∑ γjk Xjk
j =1 k =1

with γjk = g (λj , δk ).

I Thus
J K
E (Y | Z1 , Z2 ) = ∑ ∑ βjk Xjk ,
j =1 k =1

where
βjk = E (Y | Z1 = λj , Z2 = δk ).
I Since E (Y |X = x ) is linear in x, we get

E (Y | Z1 , Z2 ) = E (Y |X ) = E ∗ (Y |X ).

8 / 15
Functional form

Sample Analog
I Data: (yi , zi1 , zi2 ), i = 1, . . . , n.
I Dummy Variables: xi ,jk = 1(zi1 = λj , zi2 = δk ),
   
y1 x1,jk
 ..   .. 
y =  . , xjk =  .  .
yn xn,jk
I Least Squares:
n
 J K
2
min ∑ yi − ∑ ∑ bjk xi ,jk
b
i =1 j =1 k =1

I This gives:
∑ni=1 yi xi ,jk
bjk = = ȳ | λj , δk .
∑ni=1 xi ,jk
9 / 15
Functional form

Polynomial regressors

I Assume

∼ β0 + β1 s + β2 s2 + β3 t · s + β4 t + β5 t 2 .
E (Y |Z1 = s, Z2 = t ) =

I Example: Jacob Mincer, Schooling, Experience and Earnings,

1974, Table 5.1; 1 in 1000 sample, 1960 census; 1959 annual
earnings; n = 31093;
I y = log(earnings), s = years of schooling, t = years of work
experience;

ŷ = 4.87 + .255s − .0029s2 − .0043t · s + .148t − .0018t 2 .

10 / 15
Functional form

Predictive Effect

I Returns to college:

E (Y |Z1 = 16, Z2 = t ) − E (Y |Z1 = 12, Z2 = t )

∼ β1 · 4 + β2 (162 − 122 ) + β3 · 4 · t .
=

I Returns to high school:

E (Y |Z1 = 12, Z2 = t ) − E (Y |Z1 = 8, Z2 = t )

∼ β1 · 4 + β2 (122 − 82 ) + β3 · 4 · t .
=

11 / 15
Functional form

Plugging in the estimates

Experience Returns to college Returns to high school

0 .70 .79
10 .52 .62
20 .35 .44

Questions for you

Verify this.

12 / 15
Functional form

From predicting log W to predicting W

I Suppose E (log W | Z ) is a linear function:

E (log W | Z ) = β0 + β1 Z .
I Define U = log W − β0 − β1 Z , so E (U | Z ) = 0.
I

W = β0 + β1 Z + U
⇒ W = exp(β0 + β1 Z ) · exp(U ).
I If U and Z are independent, E [exp(U ) | Z ] = E [exp(U )] and
E (W | Z = d ) ∼ β1 (d − c ) + 1,
= exp[β1 (d − c )] =
E (W | Z = c )
 
E (W | Z = d ) ∼ 100β1 (d − c ).
100 −1 =
E (W | Z = c )
13 / 15
Functional form

Mincer model
I Compound Interest (4 = fraction of one year):

$1 → $(1 + r 4) → $(1 + r 4)2 → $(1 + r 4)3 → . . .

I Annual Return (1 + r 4)1/4 .
f (1 + x ) − f (1) log(1 + x )
lim = f 0 (1); lim =1
x →0 x x →0 x
log(1 + r 4)
lim log[(1 + r 4)1/4 ] = r · lim =r
4→0 4→0 r4
I Thus
lim (1 + r 4)1/4 = exp(r ) ≈ 1 + r
4→0
I Works for small r :

exp(.06) = 1.062; exp(.35) = 1.42; exp(.70) = 2.01.

14 / 15
Functional form

I PV (S ) = present value at t = 0 of earning 0 while in school for an

additional S years and then earning W (S ) for a very long time:
Z ∞
PV (S ) = W (S ) exp(−rt ) dt = W (S ) · exp(−rS )/r .
S
I Returns such that students are indifferent about dropping out:

PV (S ) = PV (0) ⇒ W (S ) · exp(−rS )/r = W (0)/r ,

I Thus:
log(W (S )) = log(W (0)) + rS .
I Linear Predictor:

E ∗ (Y | 1, S ) = γ0 + (r + γ1 )S ,

where Y = log(W (S )), A = log(W (0)), and

E ∗ (A | 1, S ) = γ0 + γ1 S .
15 / 15