Scribd
Upload
26 views

Module 14

This document summarizes noise models for fMRI data. It discusses how serial correlation in fMRI data can be modeled using autoregressive (AR) processes like AR(1). The AR(1) model represents errors as depending on the previous error plus new disturbances. Estimating the noise covariance matrix V is important for GLM analysis but requires iterative procedures since V and model parameters depend on each other. Maximum likelihood and restricted maximum likelihood methods can be used to estimate variance components. Spatial maps of AR model parameters show the complex spatiotemporal behavior of fMRI noise.

Uploaded by

augustus1189
Copyright
© Attribution Non-Commercial (BY-NC)
Available Formats
Download as PDF, TXT or read online on Scribd
26 views

Module 14

This document summarizes noise models for fMRI data. It discusses how serial correlation in fMRI data can be modeled using autoregressive (AR) processes like AR(1). The AR(1) model represents errors as depending on the previous error plus new disturbances. Estimating the noise covariance matrix V is important for GLM analysis but requires iterative procedures since V and model parameters depend on each other. Maximum likelihood and restricted maximum likelihood methods can be used to estimate variance components. Spatial maps of AR model parameters show the complex spatiotemporal behavior of fMRI noise.

Uploaded by

augustus1189
Copyright
© Attribution Non-Commercial (BY-NC)
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 16

Module 14: Noise Models

GLM
A standard GLM can be written:

Y = X" + !
where

! ~ N (0, V )

&Y1 # &1 X 11 " X 1 p # & ) 0 # & (1 # ! $) ! $ ! $Y ! $1 X " X (2 ! 21 2p! $ 1 ! $ 2! = $ $ ' + ! ! ! $ ! ! $!! $ ! ! $! ! $ ! $ ! $ ! $1 X " X Y np np " % ) p " % n" % %( n "

V is the covariance matrix whose format depends on the noise model.

fMRI Data

Design matrix

Noise

Regression coefficients

The quality of the model depends on our choice of X and V.

Design Matrix
! We has previously discussed various signal and nuisance components that can be included in the design matrix to improve the model.
! Temporal Basis functions
!! Allows for flexible HRF

! Parametric modulation
!! Allows for trial-specific variation in amplitude

! Motion parameters
!! Corrects for spin history artifacts

fMRI Noise
! Functional MRI data typically exhibit significant autocorrelation.
! Caused by physiological noise and low frequency drift, that has not been appropriately modeled. ! Typically modeled using either an AR(p) or an ARMA(1,1) process. ! Single subject statistics are not valid without an accurate model of the noise.

AR(1) model
! Serial correlation can be modeled using a first-order autoregressive model, i.e.

!t = "!t!1 + ut

ut ~ N (0, ! 2 )

! The error term !t depends on the previous error term !t-1 and a new disturbance term ut.

AR(1) model
! The autocorrelation function (ACF) for an AR(1) process at lag h:
" $ 1, if h = 0, ! (h ) = # |h| if h ! 0 $ % "
"=0.7
0.5 -1.0 -0.5 0.0 q 1.0

5 1:16

10

15

Error Term
! The format of V will depend on what noise model is used.
IID Case
" $ $ V !$ $ $ $ # 1 0 0 " 0 0 1 0 " 0 0 0 1 " 0 ! 0 % ' ! 0 ' ! 0 ' " ' ' ! 1 ' &

AR(1) Case
# 2 1 ! ! % % ! 1 ! % 2 V !% ! ! 1 % ! ! ! % n"1 ! n"2 ! n"3 % ! $ & ! ! n"1 ( ! ! n"2 ( ( n"3 ! ! ( ! ( ( ! 1 ( '

GLM Summary
model estimate

Y = X" + !
fitted values

!1 !1 !1 ! = ( X' V X) X' V Y

residuals

= X! Y

r = Y!Y
= (I ! (X ' V!1X )!1 X ' V!1 )Y

= RY

Estimating V
! In general the form of the covariance matrix is unknown, which means it has to be estimated. ! Estimating V depends on #s, and estimating #s depends on V. Need iterative procedure. ! Methods for estimating variance components:
! Method of moments ! Maximum likelihood ! Restricted maximum likelihood

Iterative Procedure
1.! Assume that V=I and calculate the OLS solution. 2.! Estimate the parameters of V using the residuals. 3.! Re-estimate the # values using the from step 2. estimated covariance matrix V 4.! Iterate until convergence.

Yule-Walker Estimates
! Assume !t is an AR(1) process, i.e.

!t = "!t!1 + ut
where ut ~ WN(0,$2)

t = 0,1,

! The Yule-Walker estimates are:


(1) ! "= ! (0)

!(1) 2 = !(0) $ " #


Auto Covariance Function!

MLE
! Maximum likelihood estimators (MLEs) are obtained by maximizing the log-likelihood:
T 1 1 l (! ) = ! log ( V ) ! Y ! X! V!1 Y ! X! 2 2 *

where % are parameters associated with V.

ReML
! Restricted maximum likelihood (ReML) requires maximizing the restricted loglikelihood:
T 1 1 1 T l (" ) = ! log( V ) ! log X VX ! Y ! X! V !1 Y ! X! 2 2 2 *

) (

Extra ReML variance term

where % are parameters associated with V.

ML vs ReML
! Maximum Likelihood
! Maximize likelihood of data y ! Used to estimate mean parameters ! ! But can produce biased estimates of variance
2 ML # =

1 ( yi " y ) 2 ! n

! Restricted Maximum Likelihood


! Maximize likelihood of residuals e = y - Xb ! Used to estimate variance parameters ! Provides unbiased estimates
2 ReML # =

1 ( yi " y ) 2 ! n "1

Spatio-temporal Behavior
! The spatiotemporal behavior of these noise processes is complex.

Spatial maps of the model parameters from an AR(2) model estimated for each voxels noise data.

End of Module

@fMRIstats

You might also like