4
Multilevel regression
Marginal models

Regression equation

Saturated model?

Transform categorical predictors into dummy variables

include every interaction

Parts

Fixed

Random

Error

Same for all participants

Individual for participants

Described in covariance matrix

V-matrix

R matrix

G matrix

Group

Residual

Covariance structures

Unstructured

Most general

summarizes all (co)variance

Determines regression coefficient b

In population all (co)variances are different

depend on scale of dependent var.

Make correlation

devide by SDx*SDy

Correlation metric

Goodness of fit

can be modelled by -2 log likelihood

Also known as deviance

how strongly sample deviates from expectation( that all variances are equal?)

The smaller the better

Compound symmetry

Resembles sphericity

variance equal at each timepoint
correlation between all timepoints are same

R matrix

only 2 values

average variance

average covariance

R matrix

Are they sig. different?

Likelihood ratio test

Calculate chi-square

difference between deviances

Df

how many parameters do they estimate

Per parameter difference +1 Df

Look up p-value in book

P. 509

Compound symmetry heterogeneous

Softer version of compound symmetry

R matrix

covariance matrix

Different when SDs differ

Correlations are equal

Toeplitz

Equal variances, different correlations

Different (co)variance, equal correlations

adjacent time points have same correlation

Also exists with unequal variances (heterogeneous)

AR 1

First-order autoregressive

Correlations decrease as timepoints are further apart

Variances are equal

corr 1-3=corr1-2^2

Equally parsimoneous as CS

AR heterogeneous

first-order autoregressive heterogeneous

Multiple variances

equal correlation

Other structures

MA1

first-order moving average

no pattern in the structure

ARMA1,1

Combines autoregressive pattern with moving average

AR1+ME

fist order autoregressive plus measurement error

Score combines true score with random measurement error

Errors depend on reliability

ID

Scaled Identity

All timepoints have same var

No correlation

Same as linear regression

Useful in module 5

with random intercept/slope

Fixed effects

Spss

should be univariate

Model dimension

overview of regression model

Maximum likelihood

underestimates covariance

Look at Restricted maximul likelihood (REML)

combination of multiple regression and repeated measures ANOVA

nested data structure

Comparing Likelihood ratio

Only if model is special case of the other

Not in marginal models

Good when big sample few measures

Variances +covariances

Choosing a model

Identifiability

nr of unknown parameters </= nr of equations

Equations

fixed(beta) vs mean

covariance vs residual variances

For fixed and random part

Estimation & testing

ML

REML

restricted maximum likelihood

Maximum likelihood

Steps

  1. General fixed and random
  1. reduce random
  1. Reduce fixed