Nonlinear regression

32 downloads 341926 Views 103KB Size Report
q Nonlinear regression models q Weight loss data q What to do? q Delta method q Nonlinear ... p. 2/11. Today's class s Nonlinear regression. s Examples in R.
Statistics 203: Introduction to Regression and Analysis of Variance

Nonlinear regression Jonathan Taylor

- p. 1/11

Today’s class

● Today’s class



● Nonlinear regression models ● Weight loss data ● What to do?



Nonlinear regression. Examples in R.

● Delta method ● Nonlinear regression ● Nonlinear regression: details ● Iteration & Distribution ● Confidence intervals ● Weight loss data

- p. 2/11

Nonlinear regression models

● Today’s class



● Nonlinear regression models

We have usually assumed regression is of the form

● Weight loss data ● What to do? ● Delta method

Y i = β0 +

● Nonlinear regression ● Nonlinear regression: details

p−1 X

βj Xij + εi .

j=1

● Iteration & Distribution ● Confidence intervals ● Weight loss data



Or, the regression function f (x, β) = β0 +

p−1 X

β j xj

j=1



is linear in beta. Many real-life phenomena can be parameterized by non-linear regression functions. Example: ◆ Radioactive decay: half-life is a non-linear parameter f (t, θ) = C · 2−t/θ . - p. 3/11

Weight

0

50

100

Weight loss data

180 170 160 150 140 130 120 110

Days

150

200

250

- p. 4/11

What to do?

● Today’s class



● Nonlinear regression models ● Weight loss data ● What to do? ● Delta method



● Nonlinear regression ● Nonlinear regression: details ● Iteration & Distribution ● Confidence intervals

In some cases, it is possible to linearize to get the original model. Suppose Yi = C · 2−Xi /θ εi , then

● Weight loss data

log(Yi ) = C 0 − Xi /θ + ε∗i .



If ε∗i have approximately the same variance, than it looks like original model. What if Yi = C · 2−Xi /θ + εi where the εi ’s have the same variance?

- p. 5/11

Delta method

● Today’s class



● Nonlinear regression models

The delta method tells us that

● Weight loss data

Var(f (Yi )) ' f 0 (E(Yi ))2 Var(Yi ).

● What to do? ● Delta method ● Nonlinear regression ● Nonlinear regression: details



● Iteration & Distribution ● Confidence intervals

In this case Var(log(Yi )) '

● Weight loss data

■ ■

1

2 σ . /θ

C 2 2−2Xi Could use weighted least-squares if θ was known. One possibility: IRLS. bj , j ≥ 0, regress log(Yi ) onto Xi ◆ Starting at some initial θ solve for βb1j+1 , and set 1 j+1 b θ0 = − j+1 βb1

- p. 6/11

Nonlinear regression

● Today’s class



● Nonlinear regression models

Alternatively,

● Weight loss data ● What to do?

b = argmin b θ) (C,

● Delta method ● Nonlinear regression

(C,θ)

● Nonlinear regression: details ● Iteration & Distribution

n X

(Yi − C2−Xi /θ )2 .

i=1

● Confidence intervals ● Weight loss data



In general, given a possibly non-linear regression function f (x, θ), θ ∈ Rp : we can try to solve θb = argmin θ



n X

(Yi − f (xi , θ))2

i=1

where xi is the i-th row of the design matrix. Techniques: usually use a steepest descent approach instead of Newton-Raphson.

- p. 7/11

Nonlinear regression: details

● Today’s class ● Nonlinear regression models



p X ∂f 0 0 0 b f (xi , θ) ' f (xi , θ ) + (θ − θ ) = η i i i (θ). ∂θ j (xi ,θ)=(xi ,θb0 ) j=1

● Weight loss data ● What to do? ● Delta method ● Nonlinear regression ● Nonlinear regression: details ● Iteration & Distribution ● Confidence intervals ● Weight loss data

Given an initial value θb0



In matrix form η 0 (θ) = ω 0 + Z 0 θ where

∂f 0 Zij = ∂θj

(x,θ)=(xi ,θb0 ) p X

ωi0 = f (xi , θ0 ) −

∂ηi = ∂θj

θ=θb0

0 θi0 Zij

j=1



The vector η 0 (θ) is our approximation to the regression surface, we want to choose θ to get as close to Y as possible. Project onto “tangent space.” - p. 8/11

Iteration & Distribution

● Today’s class ● Nonlinear regression models



● Weight loss data ● What to do?

Want to choose θb1 as follows θb1 = argmin

● Delta method ● Nonlinear regression

θ

● Nonlinear regression: details ● Iteration & Distribution

n X

(Yi − ωi0 − Z 0 θ)2

i=1

● Confidence intervals ● Weight loss data





Solution Given θbj , set

θb1 = Z 0 t Z 0

−1

θbj+1 = Z j t Z j



Repeat until convergence.



b Approximate distribution of θ: where

Z 0 t (Y − ω 0 ).

−1

Z j t (Y − ω j ).

b −1 ), bt Z) θb ∼ N (θ, σ b 2 (Z

n X b 2 /(n − p). (Yi − f (xi , θ)) σ b2 =

- p. 9/11

Confidence intervals

● Today’s class



● Nonlinear regression models ● Weight loss data ● What to do? ● Delta method ● Nonlinear regression ● Nonlinear regression: details



● Iteration & Distribution ● Confidence intervals ● Weight loss data



If θ is restricted, say θ ≥ 0 the asymptotic confidence intervals may be inaccurate (may overlap into negative numbers). library(MASS) provides another method to obtain confidence intervals based on “inverting” an F -test. Basic idea, confidence interval for θ1 : for each fixed value θ1,0 we could compute the “extra sum of squares” between the unrestricted model and the model with θ1 fixed at θ1,0 . b SSEθ1 =θ1,0 (θb2:p ) − SSE(θ) ∼ F1,n−p F (θ1,0 ) = 2 σ b





at least approximately under H0 : θ1 = θ1,0 . Or, q T (θ1,0 ) = sign(θb − θb1,0 ) F (θ1,0 ) ∼ tn−p . Confidence interval:

{θ1 : −tn−p,1−α/2 < T (θ1 ) < tn−p,1−α/2 }.

- p. 10/11

150 140 130 120 110

Weight

160

170

180

Weight loss data

0

50

100

150 Days

200

250 - p. 11/11