Model Selection in Semiparametrics and Measurement Error Models

advertisement

Model Selection in Semiparametrics and Measurement Error Models

Raymond J. Carroll

Department of Statistics

Faculty of Nutrition and Toxicology

Texas A&M University http://stat.tamu.edu/~carroll

Outline

 Problem motivating the work

 Measurement error in nutritional epidemiology

 Power of model selection/averaging

 Difficulties with BIC

 Semiparametric formulation

 The Hjort and Claeskens asymptotics

 Major results on semiparametric efficiency

Co-authors, Nutritional Epidemiology

Victor Kipnis

National Cancer Institute

Laurence Freedman, Gertner

Institute (Israel) and National

Cancer Institute

Co-author, Semiparametrics

Gerda Claeskens, University of Leuvan

Papers

 Seemingly Unrelated Measurement Error

Models , Biometrics

Larry Freedman

, 2006, Victor Kipnis and

 Post-Model Selection Inference in

Semiparametric Models , with Gerda

Claeskens

 This paper has been under review for 6 months, at a journal that just asked me to review a paper in 6 weeks. 

Seemingly Unrelated Measurement

Error Models (SUMEM)

 Problem : Understand the properties of food frequency questionnaires (FFQ), the most used means of estimating nutrient intakes

 Issue : True intake cannot be measured

SUMEM

 Notation :

 Y = FFQ

 X = true intake ( Not Observable )

 Z = other covariates

 Goals : Estimate two important properties

 For pre-study sample sizes ρ

=corr (Y,X)

 post-study relative risk and power

λ

= attenuation = c ov(Y,X) var(Y)

The OPEN Study

 I will use data from the OPEN Study

 First large biomarker study for nutritional epidemiology

 Biomarkers available:

 Protein (urinary nitrogen)

 Energy/Calories (Doubly labeled water)

 Approximately 200 women in the study

SUMEM: Model for Protein

 Basic variance components model : ij

P

0

P P

1 i i

P P ij r = Normal(0,σ

2 )=person-specific bias i r,P ij

2

ε,P

 Error Model for Biomarker :

P P

W =X +U ij i

P ij

P

U = Normal(0,σ

2 ) ij u,P

SUMEM: Model for Protein

 More Data : Along with Protein, we also measure Energy (calories)

 Correlated :

 Energy = Protein + Fat + Carbohydrates + Alcohol

 Combine : We try to combine energy with protein

SUMEM: Model for Protein

 Variance components model with Energy :

P P P P

Y = β +β X ij 0 1 i

 β P

2

X i

E  r +ε i

P ij

E

Y = β + ij

E

0

β E

1

X i

P  β X

2 i

E  r ε i

E E

+ ij

Note how this model predicts FFQ protein using true energy

 Error Model for Biomarker :

P P

W =X +U ij i

P ij

E E

W =X +U ij i

E ij

SUMEM: More pain, no gain

 Variance components model with Energy :

P P P P

Y = β +β X ij 0 1 i

 β P

2

X i

E  r +ε i

P ij

E

Y = β + ij

E

0

β E

1

X i

P  β X

2 i

E  r ε i

E E

+ ij

Note how this model predicts FFQ protein using true energy

 No Gain : One can show that if you fit this measurement error model,

 Marginal protein fit unaffected, identical estimates of correlation and attenuation

SUMEM: More pain, no gain

 Variance components model with Energy :

P P P P

Y = β +β X ij 0 1 i

 β P

2

X i

E  r +ε i

P ij

E

Y = β + ij

E

0

β E

1

X i

P  β X

2 i

E  r ε i

E E

+ ij

Note how this model predicts FFQ protein using true energy

 Intuition : To predict FFQ Protein:

 If I tell you true protein intake, why should true energy intake matter?

SUMEM: More pain, more gain?

 Variance components model with Energy :

P P P P

Y = β +β X ij 0 1 i

 β P

2

X i

E  r +ε i

P ij

E

Y = β + ij

E

0

β E

1

X i

P  β X

2 i

E  r ε i

E E

+ ij

Full Model

 A reduced model :

P P P P

Y = β +β X ij 0 1 i

E

Y = β ij

E

0

 i

 β X

2 i

E  r +ε i

E ij

P ij Reduced Model

OPEN Biomarker Study Results

 Variance Reduction : As if the sample size increased by a factor of 3 (s.e. from Burnham

& Anderson)

Model

Correlation Full

ρ

Reduced

Estimate Std. err.

0.282

0.088

0.250

0.048

Attenuation Full

λ Reduced

0.129

0.114

0.041

0.024

OPEN: Model Selection

 AIC and Weights  2 ratio chi-squared statistic, full vs. reduced

 Let d be the difference in degrees of freedom

 The AIC weight to the reduced model is

ω

AIC,R

(

χ

2 /2-

d

)

-1

 We can do a weighted average of the full and reduced model estimates

OPEN: Model Selection

 BIC and Weights  2 ratio chi-squared statistic, full vs. reduced

 Let d be the difference in degrees of freedom

 The BIC weight to the reduced model is

ω

BIC,R

χ 2 d

 log(n

)/2)

-1

OPEN: Model Selection

 In OPEN : the AIC weight to the reduced model is

ω =0.967, ω =1.000

AIC,R BIC,R

 We can do a weighted average of the full and reduced model estimates

 In OPEN, this is obviously essentially the reduced model

OPEN: Bootstrapping

Bootstrap: resample the individuals

Note how the bootstrap distribution of the

AIC weights is nothing like what one would expect

OPEN: Bootstrapping

Note how the bootstrap distribution of the attenuation λ is much more variable than the Burnham and Anderson calculations suggest

OPEN: Bootstrapping

Note how the bootstrap distribution of the correlation ρ is much more variable than the Burnham and Anderson calculations suggest

OPEN Type Simulation

 Setup : We used OPEN parameters, but put in a full model

 n = 200, as in OPEN

 BIC is supposed to be a consistent model selector, hence should have weights near 0.

 AIC is not a consistent model selector

Full Model Simulation

The weights should be near zero , since the full model holds.

Note how BIC is an utter disaster : it almost always selects the reduced model

Full Model Simulation

Correlation : Fitting the reduced model here results in a

30% bias

This would mean a study would have only 55% of the sample size it needs to attain a desired power

Bias here is important!

Full Model Simulation

Attenuation : Fitting the reduced model here results in a

30% bias

Full Model Simulation

Correlation : Fitting the reduced model here results in a

25% bias

This would mean a study would have only 60% of the sample size it needs to attain a desired power

Bias here is important!

Full Model Simulation

For AIC, the Burnham and Anderson standard error estimates are OK, but coverage is somewhat low

For BIC , it is so badly biased that the coverage probabilities are

< 50% in all cases

Summary of Numerical Work

 BIC:

 seriously biased in the full model holds

 Seems to love the reduced model

 From the Biometrics article

Oracle Estimators

 BIC:

 This is an example of an oracle estimator

 Leeb and Poetscher state

Oracle Estimators

 Leeb and Poetscher go on to state

 This seems to be too good to be true, and it is

 It is a delusion to believe it carries statistical meaning

 It is remarkable that some of the lessons learned from Hodges’ counterexample seem not to have been received in the model selection literature

Asymptotic Framework

 I will next describe some local-alternative asymptotics for semiparametric models

 These go some way to understanding the problem with oracle estimators

 Semiparametric generalization of the work of

Hjort and Claeskens

Semiparametric Framework

 Likelihood Formulation

L

Y,X,

α

,

θ

true true

(Z )

 Example : Partially Linear Model

Y = X

T

α

true

+

θ

true

( Z) + ε

Semiparametric Framework

 Example : Heteroscedastic Linear Model

Y = X

T α

+ ε true var(ε) = e xp

θ true

(Z)

Semiparametric Framework

 Likelihood Formulation

L

Y,X,

α

,

θ

true true

(Z )

 Two models:

 Full model

α

(β,γ) full

=

 Reduced model

α β,0) red

=(

Semiparametric Framework

 Local misspecification assumption

α full

=(β ,γ=δ/ n)

 Intuitively: even an oracle will have to think a bit to distinguish between the full and reduced models

 Practically: the models are not extremely different

Semiparametric Framework

 Local misspecification assumption

α full

=(β ,γ=δ/ n)

 Hjort and Claeskens: parametric models

 We deal with semiparametric models

Profile Methods

 Likelihood Formulation

L

α θ

(Z )

  nonparametric regression to obtain

 Then maximize n 

L i=1

 i i

α

ˆθ

Z ) i

,

α

ˆ ( , )

Semiparametric Information Bound

 The semiparametric information is the semiparametric version of Fisher information

S (

) =cov

 d d

L

θ α

)

 Many references, good one is Murphy and van der Vaart, JASA, 2002

Semiparametric Information Bound

S

α

= (β,γ=δ/

 d d

L

 n )

θ

, )



 Standard Result the full model

 ( ) n

1/2

( full

 

  full

=

 d

Normal{ 0, S -1

(

)}

Semiparametric Information Bound

 Non-Standard Result : we have to compute the limit distribution of the reduced model estimate at the full (local) model

 ( )

 Easy using contiguity arguments (LeCam’s 1 st ,

2 nd and 3 rd Lemma)

Semiparametric Information Bound

S

α

= (β,γ=δ/

 d d

L

 n )

θ

, )



 Non-Standard Result : Note asymptotic bias n

1/2

( red

 

  red

 red

=

 d

Normal{

 red

, S -1



  red

S -1



 S



  d

Semiparametric Information Bound

 Model Averaging Result : Consider a model averaged estimator with weights for the reduced model depending on the profile likelihood, e.g.,

BIC, AIC, etc.

 Recall that α

= (β,γ =δ/ n )

 Let D be the limiting distribution of 1/2 ˆ

Semiparametric Information Bound

 Model Averaging Result : Then the model average estimator has limiting distribution

(D) red

1 (D)

 full

 This is exactly the same result as in Hjort and

Claeskens, except their Fisher information matrix is replaced by the semiparametric information bound

Semiparametric Information Bound

 Inference : Hjort and Claeskens propose a method for inference that is asymptotically correct.

 In numerous simulations and examples for AIC, we have found that it is essentially the same as fullmodel inference

 BIC : In this local misspecification framework,

BIC is an inconsistent model selector: it selects the reduced model with probability  1.

Summary

 SUMEM : Model averaging can, in real life, result in estimators that decrease variance by

2/3 while retaining robustness

 Very important in measurement error problems, where information is sparse

 Inference : The Burnham and Anderson standard error formulae are reasonably precise

 Because of non-normal limit distributions, coverage probabilities are much less than nominal

Summary

 BIC and Oracles : Simply disastrous in our example

 Lack of uniform consistency means the: “ it sounds too good to be true ” is correct

 Bootstrap : Does not work

 Not new, but striking numerics

 Semiparametrics :

 In the local framework, same as parametric methods with the semiparametric information bound replacing the Fisher information (I contiguity)

Burnham and Anderson Formula

 You want to estimate a standard errror for a model-average estimate of ( )

 

R reduced mod el estimate

 

F full mod el estimate

ˆ 2

R estimated var iance of

ˆ 2

F estimated var iance of

 

R

 

F

ˆ

F

2

ˆ

MA

ˆ ˆ

R

(1

ˆ

)

2

(1

ˆ

)

ˆ

F

2

Bootstrap Problem

Estimator :

Bootstrap :

1/2

ˆ

ˆ ˆ

ˆ

R

)

)

1

1 1/2

ˆ

ˆ b

F

ˆ

) b )

Difference :

 1/2 ˆ b 1/ 2

( ˆ b ) (

 1/2 ˆ b

(n γ )

  1/ 2 ˆ

ˆ

R

)

 

1/ 2

1/2 ˆ b

(

ˆ

R

) (

 

F

)

( ˆ b

F

) ( ˆ

F

)

 First two terms have same limit as the limit in the data world.

 3 rd term is not  0, hence inconsistency

Download