7 Nonlinear

Moving Beyond Linearity
The truth is never linear!
1 / 23

Or almost never!
1 / 23

Or almost never!
But often the linearity assumption is good enough.
1 / 23

Or almost never!
But often the linearity assumption is good enough.
When its not . . .
polynomials,
step functions,
splines,
local regression, and
generalized additive models
offer a lot of flexibility, without losing the ease and

interpretability of linear models.
1 / 23
Polynomial Regression
yi = 0 + 1 xi + 2 x2i + 3 x3i + . . . + d xdi + i
| || | || | | || | || || | || | || || | || || | || | || | | | | | || | || |
0.10
0.05
Pr(Wage>250 | Age)
200
150
100
0.00
50
Wage
0.15
250
300
0.20
Degree4 Polynomial
20
30
40
50
Age
60
70
80
||| || || ||| ||| ||| ||| ||| ||| ||||| ||||| ||| ||| ||||| ||||| ||||| ||||| ||| ||||| ||| ||| ||| ||| ||||| ||| ||| ||||| ||| ||| ||| ||| ||| ||| ||| ||| ||||| ||| || ||| ||| ||| ||| ||| ||||| ||| ||| || || || || ||| || || || || | | || || |
20
30
40
50
60
70
||
80
Age
2 / 23
Details
Create new variables X1 = X, X2 = X 2 , etc and then treat
as multiple linear regression.
3 / 23
Details
Not really interested in the coefficients; more interested in
the fitted function values at any value x0 :
f(x0 ) = 0 + 1 x0 + 2 x20 + 3 x30 + 4 x40 .
3 / 23
Details
f(x0 ) = 0 + 1 x0 + 2 x20 + 3 x30 + 4 x40 .

Since f(x0 ) is a linear function of the ` , can get a simple
expression for pointwise-variances Var[f(x0 )] at any

value x0 . In the figure we have computed the fit and
pointwise standard errors on a grid of values for x0 . We
show f(x0 ) 2 se[f(x0 )].
3 / 23
Details
f(x0 ) = 0 + 1 x0 + 2 x20 + 3 x30 + 4 x40 .

Since f(x0 ) is a linear function of the ` , can get a simple
expression for pointwise-variances Var[f(x0 )] at any

value x0 . In the figure we have computed the fit and
pointwise standard errors on a grid of values for x0 . We
show f(x0 ) 2 se[f(x0 )].
We either fix the degree d at some reasonably low value,
else use cross-validation to choose d.
3 / 23
Details continued
Logistic regression follows naturally. For example, in figure
we model
Pr(yi > 250|xi ) =
exp(0 + 1 xi + 2 x2i + . . . + d xdi )

.
1 + exp(0 + 1 xi + 2 x2i + . . . + d xdi )
To get confidence intervals, compute upper and lower
bounds on on the logit scale, and then invert to get on

probability scale.
4 / 23
Details continued
we model
Pr(yi > 250|xi ) =
exp(0 + 1 xi + 2 x2i + . . . + d xdi )

.
1 + exp(0 + 1 xi + 2 x2i + . . . + d xdi )

probability scale.
Can do separately on several variablesjust stack the
variables into one matrix, and separate out the pieces

afterwards (see GAMs later).
4 / 23
Details continued
we model
Pr(yi > 250|xi ) =
exp(0 + 1 xi + 2 x2i + . . . + d xdi )

.
1 + exp(0 + 1 xi + 2 x2i + . . . + d xdi )

probability scale.
Can do separately on several variablesjust stack the
variables into one matrix, and separate out the pieces

afterwards (see GAMs later).
Caveat: polynomials have notorious tail behavior very
bad for extrapolation.
Can fit using y poly(x, degree = 3) in formula.

4 / 23
Step Functions
Another way of creating transformations of a variable cut
the variable into distinct regions.
C1 (X) = I(X < 35),
C2 (X) = I(35 X < 50), . . . , C3 (X) = I(X 65)
| || | | | | || || || | | || || | || | | || | | | | | | | | || | | | |
0.10
Pr(Wage>250 | Age)
0.05
200
150
100
0.00
50
Wage
0.15
250
300
0.20
Piecewise Constant
20
30
40
50
Age
60
70
80
|| ||| || || ||| || ||| ||| ||| |||| ||||| || ||| ||| ||| ||| ||||| ||| ||| |||||||||| |||||||||| ||| ||| ||||| ||||| ||| ||| ||||| ||||| ||| ||||| ||| ||| ||| ||| ||| ||| ||| ||| ||| ||| ||| || || || || || || || || | || | || || | || |
20
30
40
50
Age
60
70
||
80
5 / 23
Step functions continued

Easy to work with. Creates a series of dummy variables
representing each group.
6 / 23

Useful way of creating interactions that are easy to

interpret. For example, interaction effect of Year and Age:
I(Year < 2005) Age,
I(Year 2005) Age
would allow for different linear functions in each age

category.
6 / 23


I(Year < 2005) Age,
I(Year 2005) Age

category.
In R: I(year < 2005) or cut(age, c(18, 25, 40, 65, 90)).
6 / 23


I(Year < 2005) Age,
I(Year 2005) Age

category.
In R: I(year < 2005) or cut(age, c(18, 25, 40, 65, 90)).
Choice of cutpoints or knots can be problematic. For
creating nonlinearities, smoother alternatives such as

splines are available.
6 / 23
Piecewise Polynomials
Instead of a single polynomial in X over its whole domain,
we can rather use different polynomials in regions defined

by knots. E.g. (see figure)
(
01 + 11 xi + 21 x2i + 31 x3i + i if xi < c;
yi =
02 + 12 xi + 22 x2i + 32 x3i + i if xi c.
Better to add constraints to the polynomials, e.g.
continuity.
Splines have the maximum amount of continuity.
7 / 23
200
40
50
60
70
20
30
40
50
Age
Age
Cubic Spline
Linear Spline
60
70
60
70
200
150
100
50
50
100
150
Wage
200
250
30
250
20
Wage
150
Wage
50
100
150
50
100
Wage
200
250
Continuous Piecewise Cubic
250
Piecewise Cubic
20
30
40
50
Age
60
70
20
30
40
50
Age
8 / 23
Linear Splines
A linear spline with knots at k , k = 1, . . . , K is a piecewise
linear polynomial continuous at each knot.
We can represent this model as
yi = 0 + 1 b1 (xi ) + 2 b2 (xi ) + + K+3 bK+3 (xi ) + i ,
where the bk are basis functions.
9 / 23
Linear Splines
A linear spline with knots at k , k = 1, . . . , K is a piecewise
linear polynomial continuous at each knot.
We can represent this model as
yi = 0 + 1 b1 (xi ) + 2 b2 (xi ) + + K+3 bK+3 (xi ) + i ,
where the bk are basis functions.
b1 (xi ) = xi
bk+1 (xi ) = (xi k )+ ,
k = 1, . . . , K
Here the ()+ means positive part; i.e.

xi k if xi > k
(xi k )+ =
0 otherwise
9 / 23
0.9
0.7
0.3
0.5
f(x)
0.0
0.2
0.4
0.6
0.8
1.0
0.6
0.8
1.0
0.2
0.0
b(x)
0.4
0.0
0.2
0.4
x
10 / 23
Cubic Splines
A cubic spline with knots at k , k = 1, . . . , K is a piecewise
cubic polynomial with continuous derivatives up to order 2 at
each knot.
Again we can represent this model with truncated power basis
functions
yi = 0 + 1 b1 (xi ) + 2 b2 (xi ) + + K+3 bK+3 (xi ) + i ,
b1 (xi ) = xi
b2 (xi ) = x2i
b3 (xi ) = x3i
bk+3 (xi ) = (xi k )3+ ,
where
(xi
k )3+

=
k = 1, . . . , K
(xi k )3 if xi > k
0 otherwise
11 / 23
1.6
1.4
1.0
1.2
f(x)
0.0
0.2
0.4
0.6
0.8
1.0
0.6
0.8
1.0
0.2
0.0
b(x)
0.4
0.0
0.2
0.4
x
12 / 23
Natural Cubic Splines

A natural cubic spline extrapolates linearly beyond the
boundary knots. This adds 4 = 2 2 extra constraints, and
allows us to put more internal knots for the same degrees of
freedom as a regular cubic spline.
150
100
50
Wage
200
250
Natural Cubic Spline

Cubic Spline
20
30
40
50
60
70
13 / 23
Fitting splines in R is easy: bs(x, ...) for any degree splines,

and ns(x, ...) for natural cubic splines, in package splines.
| || | | || | || || || || | || | | || | | || | || || | || | | || || || | | ||
0.10
Pr(Wage>250 | Age)
0.05
200
150
100
0.00
50
Wage
0.15
250
300
0.20
20
30
40
50
Age
60
70
80
|| || || || || ||| || ||| ||| ||||| ||||| ||| ||| ||| ||||| ||||| ||| ||||| ||| ||| ||| ||| ||| ||||| ||| ||| ||||| ||| ||||| ||||| ||| ||| ||||| ||| ||| ||| ||| ||| ||| ||| ||| ||| ||||| || ||| || || || || || || || || || | || | | | |
20
30
40
50
60
70
||
80
Age
14 / 23
Knot placement
One strategy is to decide K, the number of knots, and then
place them at appropriate quantiles of the observed X.

A cubic spline with K knots has K + 4 parameters or
degrees of freedom.
A natural spline with K knots has K degrees of freedom.
200
150
100
Comparison of a
degree-14 polynomial and a natural
cubic spline, each
with 15df.
50
Wage
250
300

Polynomial
20
30
40
50
Age
60
70
80
15 / 23
Knot placement
One strategy is to decide K, the number of knots, and then
place them at appropriate quantiles of the observed X.

A cubic spline with K knots has K + 4 parameters or
degrees of freedom.
A natural spline with K knots has K degrees of freedom.
200
150
100
Comparison of a
degree-14 polynomial and a natural
cubic spline, each
with 15df.
ns(age, df=14)
poly(age, deg=14)
50
Wage
250
300

Polynomial
20
30
40
50
Age
60
70
80
15 / 23
Smoothing Splines
This section is a little bit mathematical
Consider this criterion for fitting a smooth function g(x) to
some data:
Z
n
X
(yi g(xi ))2 + g 00 (t)2 dt
minimize
gS
i=1
16 / 23
Smoothing Splines
some data:
Z
n
X
(yi g(xi ))2 + g 00 (t)2 dt
minimize
gS
i=1
The first term is RSS, and tries to make g(x) match the
data at each xi .
16 / 23
Smoothing Splines
some data:
Z
n
X
(yi g(xi ))2 + g 00 (t)2 dt
minimize
gS
i=1
data at each xi .
The second term is a roughness penalty and controls how
wiggly g(x) is. It is modulated by the tuning parameter
0.
16 / 23
Smoothing Splines
some data:
Z
n
X
(yi g(xi ))2 + g 00 (t)2 dt
minimize
gS
i=1
data at each xi .
0.
The smaller , the more wiggly the function, eventually
interpolating yi when = 0.
16 / 23
Smoothing Splines
some data:
Z
n
X
(yi g(xi ))2 + g 00 (t)2 dt
minimize
gS
i=1
data at each xi .
0.
The smaller , the more wiggly the function, eventually
interpolating yi when = 0.
As , the function g(x) becomes linear.
16 / 23
Smoothing Splines continued

The solution is a natural cubic spline, with a knot at every
unique value of xi . The roughness penalty still controls the
roughness via .
17 / 23

roughness via .
Some details
Smoothing splines avoid the knot-selection issue, leaving a
single to be chosen.
17 / 23

roughness via .
Some details
The algorithmic details are too complex to describe here.
In R, the function smooth.spline() will fit a smoothing
spline.
17 / 23

roughness via .
Some details
The algorithmic details are too complex to describe here.
In R, the function smooth.spline() will fit a smoothing
spline.
The vector of n fitted values can be written as g
= S y,
where S is a n n matrix (determined by the xi and ).
The effective degrees of freedom are given by
df =
n
X
i=1
{S }ii .
17 / 23
Smoothing Splines continued choosing
We can specify df rather than !
In R: smooth.spline(age, wage, df = 10)
18 / 23
Smoothing Splines continued choosing
We can specify df rather than !
In R: smooth.spline(age, wage, df = 10)

The leave-one-out (LOO) cross-validated error is given by
RSScv () =
n
X
i=1
(yi
(i)
g (xi ))2

n
X
yi g (xi ) 2
i=1
1 {S }ii
In R: smooth.spline(age, wage)
18 / 23
Smoothing Spline
200
50 100
0
Wage
300
16 Degrees of Freedom
6.8 Degrees of Freedom (LOOCV)
20
30
40
50
60
70
80
Age
19 / 23
Local Regression
O
0.0
0.2
0.4
0.6
0.8
1.5
1.0
0.5
0.0
0.5
OO
O
O OO
O
O
O
O
O
O
O
OOO
O
OO O
O
O
O O
O
O
O O OO O
O
O O
OOO
O OO
O
OO O
O OO OOO O
O
O
O
OO
O OO O
OO OO
O O
O
O
O
O
O OO
O
O O
O
O O
O
OO
O
OO O O
O
OO
OO O O
O O
1.0
OO
O
O OO
O
O
O
O
O
O
O
OOO
O
OO O
O
O
O O
O
O
O O OO O
O
O O
OOO
O OO
O
OO O
O OO OOO O
O
O
O
OO
O OO O
OO OO
O O
O
O
O
O
O OO
O
O O
O
O O
O
OO
O
OO O O
O
OO
OO O O
O O
1.0
0.5
0.0
0.5
1.0
1.5
Local Regression
1.0
O
0.0
0.2
0.4
0.6
0.8
1.0
With a sliding weight function, we fit separate linear fits over

the range of X by weighted least squares.
See text for more details, and loess() function in R.
20 / 23
Generalized Additive Models

Allows for flexible nonlinearities in several variables, but retains
the additive structure of linear models.
yi = 0 + f1 (xi1 ) + f2 (xi2 ) + + fp (xip ) + i .
HS
<Coll
Coll
>Coll
2005
2007
year
2009
20
10
10
20
40
30
50
2003
10
f2 (age)
30
20
10
0
10
20
30
f1 (year)
f3 (education)
20
30
10
30
40
20
<HS
20
30
40
50
age
60
70
80
education
21 / 23
GAM details
Can fit a GAM simply using, e.g. natural splines:
lm(wage ns(year, df = 5) + ns(age, df = 5) + education)
22 / 23
GAM details

Coefficients not that interesting; fitted functions are. The
previous plot was produced using plot.gam.
22 / 23
GAM details

Can mix terms some linear, some nonlinear and use
anova() to compare models.
22 / 23
GAM details

Can use smoothing splines or local regression as well:
gam(wage s(year, df = 5) + lo(age, span = .5) + education)
22 / 23
GAM details

Can use smoothing splines or local regression as well:
gam(wage s(year, df = 5) + lo(age, span = .5) + education)

GAMs are additive, although low-order interactions can be
included in a natural way using, e.g. bivariate smoothers or

interactions of the form ns(age,df=5):ns(year,df=5).
22 / 23
GAMs for classification

log
p(X)
1 p(X)

= 0 + f1 (X1 ) + f2 (X2 ) + + fp (Xp ).
<Coll
Coll
>Coll
f2 (age)
0
2
f1 (year)
f3 (education)
HS
2003
2005
2007
year
2009
20
30
40
50
age
60
70
80
education
gam(I(wage > 250) year + s(age, df = 5) + education, family = binomial)
23 / 23

7 Nonlinear

Uploaded by

Copyright:

Available Formats

7 Nonlinear

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

7 Nonlinear

Uploaded by

Copyright:

Available Formats

Moving Beyond Linearity

The truth is never linear!

Moving Beyond Linearity

Moving Beyond Linearity

Moving Beyond Linearity

local regression, and

generalized additive models

offer a lot of flexibility, without losing the ease and

as multiple linear regression.

as multiple linear regression.

Not really interested in the coefficients; more interested in

the fitted function values at any value x0 :

f(x0 ) = 0 + 1 x0 + 2 x20 + 3 x30 + 4 x40 .

as multiple linear regression.

Not really interested in the coefficients; more interested in

the fitted function values at any value x0 :

f(x0 ) = 0 + 1 x0 + 2 x20 + 3 x30 + 4 x40 .

expression for pointwise-variances Var[f(x0 )] at any

as multiple linear regression.

Not really interested in the coefficients; more interested in

the fitted function values at any value x0 :

f(x0 ) = 0 + 1 x0 + 2 x20 + 3 x30 + 4 x40 .

expression for pointwise-variances Var[f(x0 )] at any

We either fix the degree d at some reasonably low value,

else use cross-validation to choose d.

Pr(yi > 250|xi ) =

exp(0 + 1 xi + 2 x2i + . . . + d xdi )

To get confidence intervals, compute upper and lower

bounds on on the logit scale, and then invert to get on

Pr(yi > 250|xi ) =

exp(0 + 1 xi + 2 x2i + . . . + d xdi )

To get confidence intervals, compute upper and lower

bounds on on the logit scale, and then invert to get on

Can do separately on several variablesjust stack the

variables into one matrix, and separate out the pieces

Pr(yi > 250|xi ) =

exp(0 + 1 xi + 2 x2i + . . . + d xdi )

To get confidence intervals, compute upper and lower

bounds on on the logit scale, and then invert to get on

Can do separately on several variablesjust stack the

variables into one matrix, and separate out the pieces

Caveat: polynomials have notorious tail behavior very

bad for extrapolation.

Can fit using y poly(x, degree = 3) in formula.

C2 (X) = I(35 X < 50), . . . , C3 (X) = I(X 65)

Step functions continued

representing each group.

Step functions continued

representing each group.

Useful way of creating interactions that are easy to

I(Year < 2005) Age,

I(Year 2005) Age

would allow for different linear functions in each age

Step functions continued

representing each group.

Useful way of creating interactions that are easy to

I(Year < 2005) Age,

I(Year 2005) Age

would allow for different linear functions in each age