factor analysis caroline van baal march 3 rd 2004, boulder

Factor analysis

Caroline van Baal

March 3rd 2004, Boulder

Phenotypic Factor Analysis

• (Approximate) description of the relations between different variables– Compare to Cholesky decomposition

• Testing of hypotheses on relations between different variables by comparing different (nested) models– How many underlying factors?

Factor analysis and related methods

• Data reduction– Consider 6 variables:– Height, weight, arm length, leg length,

verbal IQ, performal IQ– You expect the first 4 to be correlated, and

the last 2 to be correlated, but do you expect high correlations between the first 4 and the last 2?

Data analysis in non-experimental designs using latent

constructs

• Principal Components Analysis

• Triangular Decomposition (Cholesky)

• Exploratory Factor Analysis

• Confirmatory Factor Analysis

• Structural Equation Models

Exploratory Factor Analysis

• Account for covariances among observed variables in terms of a smaller number of latent, common factors

• Includes error components for each variable• x = P * f + u• x = observed variables• f = latent factors• u = unique factors• P = matrix of factor loadings

SIMINF VOC CODCOM ARI DIG BLC MAZ PIC PIA OBA

Factor 1IQ, “g”

1

SIMINF VOC CODCOM ARI DIG BLC MAZ PIC PIA OBA

Factor 1verbal

Factor 2performal

1 1

EFA equations

• C = P * D * P’ + U * U’• C = observed covariance matrix

• Nvar by nvar, symmetric

• P = factor loadings• Nvar by nfac, full

• D = correlations between factors• Nfac by nfac, standardized

• U = specific influences, errors• Nvar by nvar, diagonal

Exploratory factor analysis

• No prior assumption on number of factors

• All variables load on all latent factors

• Factors are either all correlated or all uncorrelated

• Unique factors are uncorrelated

• Underidentification

SIMINF VOC CODCOM ARI DIG BLC MAZ PIC PIA OBA

Factor 1verbal

Factor 2performal

Fix to 0

1 1

Confirmatory factor analysis• An initial model is constructed, because:

– its elements are described by a theoretical process

– its elements have been obtained from a previous analysis in another sample

• The model has a specific number of factors• Variables do not have to load on all factors• Measurement errors may correlate• Some latent factors may be correlated,

while others are not

SIMINF VOC CODCOM ARI DIG BLC MAZ PIC PIA OBA

Factor 1verbal

Factor 2performal

1 1

SIMINF VOC CODCOM ARI DIG BLC MAZ PIC PIA OBA

Factor 1verbal

Factor 2performal

1 1

SIMINF VOC CODCOM ARI DIG BLC MAZ PIC PIA OBA

VC FD PO

SIMINF VOC CODCOM ARI DIG BLC MAZ PIC PIA OBA

VC FD PO

CFA equations

• x = P * f + u• x = observed variables, f = latent factors• u = unique factors, P = factor loadings• C = P * D * P’ + U * U’• C = observed covariance matrix• P = factor loadings• D = correlations between factors• U = diagonal matrix of errors

Structural equations models

• The factor model x = P * f + u is sometimes referred to as the measurement model

• The relations between latent factors can also be modeled

• This is done in the covariance structure model, or the structural equations model

• Higher order factor models

SIMINF VOC CODCOM ARI DIG BLC MAZ PIC PIA OBA

VC FD PO

2nd order Factor“g”

F3F2F1

• Second order factor model: C = P*(A*I*A’+B*B')*P' + U*U’

Five steps characterize structural equation models

• Model specification• Identification

– E.g., if a factor loads on 2 variables only, multiple solutions are possible, and the factor loadings have to be equated

• Estimation of parameters• Testing of goodness of fit• Respecification

• K.A. Bollen & J. Scott Long: Testing Structural Equation Models, 1993, Sage Publications

Practice!• IQ and brain volumes (MRI)

• 3 brain volumes– Total cerebellum, Grey matter, White matter

• 2 IQ subtests– Calculation, Letters / numbers

• Brain and IQ factors are correlated

• Datafile: mri-IQ-all-twinA-5.dat

Script: phenofact.mx

• BEGIN MATRICES ;• P FULL NVAR NFACT free ; ! factor loadings• D STAND NFACT NFACT !free ;! correlations between factors• U DIAG NVAR NVAR free ; ! subtest specific

influences• M Full 1 NVAR free ; ! means • END MATRICES ;

• BEGIN ALGEBRA;• C= P*D*P' +U*U' ; ! variance covariance matrix• END ALGEBRA;

• Means M /• Covariances C /

• in exploratory factor analysis, if nfact = 2, one of the factor loadings has to be fixed to 0 to make it an identified model

• fix P 1 2

• In confirmatory factor analysis, specify a brain and an IQ factor• SPECIFY P• 101 0• 102 0• 103 0• 0 204• 0 205• 0 206

• (if a factor loads on 2 variables only, it is not possible to estimate both factor loadings. Equate them, or fix one of them to 1)

Phenotypic Correlations: MRI-IQ, Dutch twins (A), n=111/296 pairs

brain

cereb

brain

grey

brain

white

IQ

calc

IQ

L/n

Cerebellum 1

Grey .63 1

White .61 .55 1

calculation .23 .25 .26 1

Letter/numb. .30 .19 .19 .46 1

• What is the fit of a 1 factor model?– C = P * P’ + U*U’, P = 5x1 full, U = 5x5 diagonal

• What is the fit of a 2 factor model?– Same, P = 5x2 full with 1 factor loading fixed to 0– (Reducion: fix first 3 factor loadings of factor 2 to 0)

• Data suggest 2 latent factors: a brain (first 3) and an IQ factor (last 2): what is the evidence for this model?– Same, P = 5x2 full with 5 factor loadings fixed to 0

• Can the 2 factor model be improved by allowing a correlation between these 2 factors?– C = P * D * P’ + U*U’, P = 5x2 full matrix (5 fixed),

D = stand 2x2 matrix, U = 5x5 diagonal matrix

Principal Components Analysis

• SPSS, SAS, Mx (functions \eval, \evec)

• Transformation of the data, not a model

• Is used to reduce a large set of correlated observed variables (xi) to (a smaller number of) uncorrelated (orthogonal) components (ci)

• xi is a linear function of ci

PCA path diagram

• D

• P

• S = observed covariances = P * D * P’

x1 x2 x3 x4 x5

c1 c2 c3 c4 c5

PCA equations

• Covariance matrix qSq = qPq * qDq * qPq’

• P = full q by q matrix of eigenvectors• D = diagonal matrix of eigenvalues• P is orthogonal: P * P’ = I (identity)

Criteria for number of factors• Kaiser criterion, scree plot, %var• Important: models not identified!

x1 x2 x3 x4 x5

c1 c2 c3 c4 c5

Correlations: satisfaction, n=100

Var 1

work

Var 2

work

Var 3

work

Var 4

home

Var 5

home

Var 6

home

Var 1 1

Var 2 .65 1

Var 3 .65 .73 1

Var 4 .14 .14 .16 1

Var 5 .15 .18 .24 .66 1

Var 6 .14 .24 .25 .59 .73 1

++++ ++

00

0

0

0

0++

++++

work home

Var 1 Var 2 Var 3 Var 4 Var 5 Var 6

PCA: Factor loadings(eigenvalues 2.89 & 1.79)

Factor 1 Factor 2

Var 1 (work) .65 .56

Var 2 (work) .72 .54

Var 3 (work) .74 .51

Var 4 (home) .63 -.56

Var 5 (home) .71 -.57

Var 6 (home) .71 -.53

Triangular decomposition (Cholesky)

x1 x2 x3 x4 x5

y1 y2 y3 y4 y5

1 operationalization of all PCA outcomes

Model is just identified! Model is saturated (df=0)

1 1 1 1 1

Triangular decomposition

• S = Q * Q’ ( = P# * P# ‘, where P# is P*D)•

5Q5 = f11 0 0 0 0f21 f22 0 0 0f31 f32 f33 0 0f41 f42 f43 f44 0f51 f52 f53 f54 f55

• Q is a lower matrix• This is not a model! This is a transformation of the

observed matrix S. Fully determinate!

Saturated model, # latent factorsscript: phenochol.mx

• BEGIN MATRICES ;• P LOWER NVAR NVAR free ; ! factor loadings• M FULL 1 NVAR free ; ! means • END MATRICES ;

• BEGIN ALGEBRA;• C= Q*Q' ; ! variance covariance matrix• K=\stnd(C) ; ! correlation matrix• X=\eval(K) ; ! eigen values (i.e., variance of latent factors)• Y=\evec(K) ; ! eigenvectors (i.e., regression coefficients)• END ALGEBRA;

• Means M /• Covariances C /