Raymond J. CarrollTexas A&M University
http://stat.tamu.edu/[email protected]
Postdoctoral Training Program:http://stat.tamu.edu/B3NC
Non/Semiparametric Regression and Clustered/Longitudinal Data
Where am I From?
College Station, home of Texas A&MI-35 I-45
Big Bend National Park
Wichita Falls, my hometown
Raymond Carroll Alan Welsh
Naisyin Wang Enno Mammen Xihong Lin
Oliver Linton
Acknowledgments
Series of papers are on my web site
Lin, Wang and Welsh: Longitudinal data (Mammen & Linton for pseudo-observation methods)
Linton and Mammen: time series data
Outline
• Longitudinal models: • panel data
• Background: • splines = kernels for independent data
• Correlated data: • do splines = kernels?
• Semiparametric case: • partially linear model: • does it matter what nonparametric method is
used?
Panel Data (for simplicity)
• i = 1,…,n clusters/individuals• j = 1,…,m observations per cluster
Subject Wave 1 Wave 2 … Wave m
1 X X X
2 X X X
… X
n X X X
Panel Data (for simplicity)
• i = 1,…,n clusters/individuals• j = 1,…,m observations per cluster• Important points:
• The cluster size m is meant to be fixed• This is not a multiple time series
problem where the cluster size increases to infinity
• Some comments on the single time series problem are given near the end of the talk
The Marginal Nonparametric Model
• Y = Response• X = time-varying covariate
• Question: can we improve efficiency by accounting for correlation?
ij ij ij
ij
Y =Θ(X )+εΘ • unknownfunctioncov(ε )=Σ
The Marginal Nonparametric Model
• Important assumption• Covariates at other waves are not
conditionally predictive, i.e., they are surrogates
• This assumption is required for any GLS fit, including parametric GLS
ij ij ik ijE(Y| X ,X for k j)=Θ(X )
Independent Data
• Splines (smoothing, P-splines, etc.) with penalty parameter =
• Ridge regression fit• Some bias, smaller variance• is over-parameterized least squares• is a polynomial regression
n T 2
i i i ii 1
Y (X ) Y (mini X )m tize ( ) dt
0
Independent Data
• Kernels (local averages, local linear, etc.), with kernel density function K and bandwidth h
• As the bandwidth h 0, only observations with X near t get any weight in the fit
n1 i
ii 1
n1 i
i 1
t
t
Xn Y K h(t) Xn K h
Independent Data
• Major methods• Splines • Kernels
• Smoothing parameters required for both
• Fits: similar in many (most?) datasets• Expectation: some combination of bandwidths
and kernel functions look like splines
12
Independent Data
• Splines and kernels are linear in the responses
• Silverman showed that there is a kernel function and a bandwidth so that the weight functions
are asymptotically equivalent • In this sense, splines = kernels • This talk is about the same result for correlated
data
n
-1n i i
i=1Θ( ) = n G ( ,t Xt )Y
nG (t,x)
The weight functions Gn(t=.25,x) in a specific case for independent data
Kernel Smoothing Spline
Note the similarity of shape and the locality: only X’s near t=0.25 get any weight
Working Independence
• Working independence: Ignore all correlations
• Fix up standard errors at the end• Advantage: the assumption
is not required• Disadvantage: possible severe loss of
efficiency if carried too far
ij ij ik ijE(Y| X ,X for k j)=Θ(X )
Working Independence
• Working independence: • Ignore all correlations• Should posit some reasonable marginal variances• Weighting important for efficiency
• Weighted versions: Splines and kernels have obvious analogues
• Standard method: Zeger & Diggle, Hoover, Rice, Wu & Yang, Lin & Ying, etc.
Working Independence
• Working independence: • Weighted splines and weighted kernels are
linear in the responses• The Silverman result still holds• In this sense, splines = kernels
Accounting for Correlation• Splines have an obvious analogue for non-
independent data• Let be a working covariance matrix
• Penalized Generalized least squares (GLS)
• GLS ridge regression• Because splines are based on likelihood ideas,
they generalize quickly to new problems
n T 2
i i i ii 1
-1wY (X ) Y (X ) (t) dtΣ
wΣ
Accounting for Correlation
• Splines have an obvious analogue for non-independent data
• Kernels are not so obvious• Local likelihood kernel ideas are standard
in independent data problems• Most attempts at kernels for correlated
data have tried to use local likelihood kernel methods
Kernels and Correlation
• Problem: how to define locality for kernels?• Goal: estimate the function at t• Let be a diagonal matrix of standard
kernel weights • Standard Kernel method: GLS pretending
inverse covariance matrix is
• The estimate is inherently local
iK(t,X )
-1/ 2 1/w
2i
1iK (t,X ) K (t,XΣ )
Kernels and CorrelationSpecific case: m=3, n=35
Exchangeable correlation structure
Red: = 0.0
Green: = 0.4
Blue: = 0.8
Note the locality of the kernel method
The weight functions Gn(t=.25,x) in a specific case
18
Splines and CorrelationSpecific case: m=3, n=35
Exchangeable correlation structure
Red: = 0.0
Green: = 0.4
Blue: = 0.8
Note the lack of locality of the spline method
The weight functions Gn(t=.25,x) in a specific case
Splines and CorrelationSpecific case: m=3, n=35
Complex correlation structure
Red: Nearly singular
Green: = 0.0
Blue: = AR(0.8)
Note the lack of locality of the spline method
The weight functions Gn(t=.25,x) in a specific case
Splines and Standard Kernels
• Accounting for correlation:• Standard kernels remain local• Splines are not local
• Numerical results can be confirmed theoretically
• Don’t we want our nonparametric regression estimates to be local?
Results on Kernels and Correlation
• GLS with weights
• Optimal working covariance matrix is working independence!
• Using the correct covariance matrix • Increases variance• Increases MSE• Splines Kernels (or at least these
kernels)
-1/ 2 1/w
2i
1iK (t,X ) K (t,XΣ )
24
Pseudo-Observation Kernel Methods
• Better kernel methods are possible• Pseudo-observation: original method• Construction: specific linear
transformation of Y• Mean = (X)• Covariance = diagonal matrix
• This adjusts the original responses without affecting the mean
* 1i i i i
-1/ 2w= diag( )Σ
Y Y ( ) Y (X )
Pseudo-Observation Kernel Methods
• Construction: specific linear transformation of Y• Mean = (X)• Covariance = diagonal
• Iterative:• Efficiency: More efficient than working
independence• Proof of Principle: kernel methods can be
constructed to take advantage of correlation
* 1i i i iY Y ( ) Y (X )
0
0.5
1
1.5
2
2.5
3
3.5
Excnhg AR Near Sing
SplineP-kernel
Efficiency of Splines and Pseudo-Observation Kernels
Exchng: Exchangeable with correlation 0.6
AR: autoregressive with correlation 0.6
Near Sing: A nearly singular matrix
Better Kernel Methods: SUR
• Simulations of the original pseudo-observation method: it is not as efficient as splines
• Suggests room for a better estimate• Naisyin Wang: her talk will describe an
even better kernel method• Basis: seemingly unrelated regression ideas• Generalizable: based on likelihood ideas
SUR Kernel Methods
• It is well known that the GLS spline has an exact, analytic expression
• We have shown that the Wang SUR kernel method has an exact, analytic expression
• Both methods are linear in the responses
SUR Kernel Methods
• The two methods differ only in one matrix term
• This turns out to be exactly the same matrix term considered by Silverman in his work
• Relatively nontrivial calculations show that Silverman’s result still holds
• Splines = SUR Kernels
29
Nonlocality
• The lack of locality of GLS splines and SUR kernels is surprising
• Suppose we want to estimate the function at t
• If any observation has an X near t, then all observations in the cluster contribute to the fit, not just those with covariates near t
• Splines, pseudo-kernels and SUR kernels all borrow strength
Nonlocality
• Wang’s SUR kernels = BLUP-like pseudo kernels with a clever linear transformation. Let
• SUR kernels are working independence kernels
-1 jkjk
Σ = σ
jjweight =σs
jk
*ij ij ik ikjjk j
σpseudo Y =Y + Y (- X )σobs
Locality of Kernels
• Original pseudo-observation method: pseudo observations uncorrelated
• SUR kernels: pseudo-observations are correlated
• SUR kernels are not local• SUR kernels are local in (the same!)
pseudo-observations
jk
*ij ij ik ikjjk j
σpseudo Y =Y + Y (- X )σobs
Locality of Splines
• Splines = SUR kernels (Silverman-type result)
• GLS spline: • Iterative• standard independent spline smoothing• SUR pseudo-observations at each iteration
• GLS splines are not local• GLS splines are local in (the same!) pseudo-
observations
Time Series Problems
• Time series problems: many of the same issues arise
• Original pseudo-observation method• Two stages• Linear transformation• Mean (X)• Independent errors• Single standard kernel applied
• Potential for great gains in efficiency (even infinite for AR problems with large correlation)
Time Series: AR(1) Illustration, First Pseudo Observation Method
• AR(1), correlation :
• Regress Yt0 on Xt
0t t t-1 t-1Y =Y - Y -Θ(Xρ )
t t-1 tε - ε =u (whitenρ oise)
Time Series Problems
• More efficient methods can be constructed• Series of regression problems: for all j,
• Pseudo observations • Mean• White noise errors• Regress for each j: fits are asymptotically
independent• Then weighted average
• Time series version of SUR-kernels for longitudinal data?
jiY
t j(X )
Time Series: AR(1) Illustration, New Pseudo Observation Method
• AR(1), correlation :
• Regress Yt0 on Xt and Yt
1 on Xt-1
• Weights: 1 and 2
0t t t-1 t-1Y =Y - Y -Θ(Xρ )
t t-1 tρε - ε =u
1 -1t t-1 t tY =Y - Y -Θ(Xρ )
Time Series Problems
• AR(1) errors with correlation • Efficiency of original pseudo-observation
method to working independence:
• Efficiency of new (SUR?) pseudo-observation method to original method:
21 : as 11
21 : 2 as 1
36
The Semiparametric Model
• Y = Response• X,Z = time-varying covariates
• Question: can we improve efficiency for by accounting for correlation?
ij ij ij ij
ij
Y =Z +Θ(X )+εcov(ε
β)=Σ
Profile Methods
• Given , solve for say• Basic idea: Regress
• Working independence• Standard kernels• Pseudo –observations kernels• SUR kernels
*ij ij ij ijY =Y -Z on Xβ
ijΘ(X ,β)
ij ij ij ij
ij
Y =Z +Θ(X )+εcov(ε
β)=Σ
Profile Methods
• Given , solve for say• Then fit GLS or W.I. to the model with mean
• Question: does it matter what kernel method is used?
• Question: How bad is using W.I. everywhere?• Question: are there efficient choices?
ij ijZ + (XΘβ ,β)
ijΘ(X ,β)
ij ij ij ij
ij
Y =Z +Θ(X )+εcov(ε
β)=Σ
The Semiparametric Model: Special Case
• If X does not vary with time, simple semiparametric efficient method available
• The basic point is that has common mean and covariance matrix
• If were a polynomial, GLS likelihood methods would be natural
ij ij i ijY =Z +Θ(X )+εβ
ij ijY -Z βiΘ(X ) Σ
Θ( )
The Semiparametric Model: Special Case
• Method: Replace polynomial GLS likelihood with GLS local likelihood with weights
• Then do GLS on the derived variable
• Semiparametric efficient
ij ij i ijY =Z +Θ(X )+εβ
iXK ht
ij ijZ + (XΘβ ,β)
Profile Method: General Case• Given , solve for say• Then fit GLS or W.I. to the model with mean
• In this general case, how you estimate matters• Working independence• Standard kernel• Pseudo-observation kernel• SUR kernel
ij ijZ + (XΘβ ,β)
ijΘ(X ,β)
Profile Methods• In this general case, how you estimate
matters• Working independence• Standard kernel• Pseudo-observation kernel• SUR kernel
• We have published the asymptotically efficient score, but not how to implement it
Profile Methods
• Naisyin Wang’s talk will describe • These phenomena• Search for an efficient estimator• Loss of efficiency for using working
independence to estimate • Examples where ignoring the correlation
can change conclusions
Conclusions (1/3): Nonparametric Regression
• In nonparametric regression• Kernels = splines for working
independence (W.I.)• Weighting is important for W.I.• Working independence is inefficient• Standard kernels splines for
correlated data
Conclusions (2/3): Nonparametric Regression
• In nonparametric regression• Pseudo-observation methods improve
upon working independence• SUR kernels = splines for correlated
data• Splines and SUR kernels are not local• Splines and SUR kernels are local in
pseudo-observations
Conclusions (3/3): Semiparametric Regression
• In semiparametric regression• Profile methods are a general class• Fully efficient parameter estimates are
easily constructed if X is not time-varying• When X is time-varying, method of
estimating affects properties of parameter estimates
• Ignoring correlations can change conclusions (see N. Wang talk)
Conclusions: Splines versus Kernels
• One has to be struck by the fact that all the grief in this problem has come from trying to define kernel methods
• At the end of the day, they are no more efficient than splines, and harder and more subtle to define
• Showing equivalence as we have done suggests the good properties of splines