Scribd
Upload
21 views

PSYC588Lecture 01

Pscyh 588 Lecture 1

Uploaded by

kaiwenkaer
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
21 views

PSYC588Lecture 01

Pscyh 588 Lecture 1

Uploaded by

kaiwenkaer
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 9

Covariance structure and factor models

PSYC/EPSY/SOC/STAT 588 Spring 2011

Instructor: Sungjin Hong


Office: Psych 429 (244-8296, hongsj AT illinois DOT edu)
Class meetings: TR, 3:00-4:50, room 29 (219A/289 to
access AMOS; 289 also reserved for Wed
3-5pm)
Office hours: W, 2:00-3:00 or by appointment
Course website: https://compass.illinois.edu
Historical backgrounds

1. Path analysis

2. Factor analysis (measurement model in SEM terms)

3. General estimation procedure


1. Path analysis

• To parsimoniously describe relationships between a set of


observed variables with directed paths by which we wish to
represent causal influences of one variable to another

a b
IQ GPA AI

EG EA

GPA  a IQ  E G , AI  b GPA  E A

cov  GPA,AI   a 2b var  IQ   b var  E G 


2. Factor analysis

• Scientifically useful variables are often not directly observable


(measurable) --- e.g., IQ

• The FA model defines the unobservable (latent) variables such


that they linearly relate to (influence or determine) a set of
correlated, yet distinctive, observable variables (indicators)

X 1  a1  b1 IQ  E1
  
X q  aq  bq IQ  Eq

cf . IQ  c1 X 1  c2 X 2    cq X q  EIQ
3. General estimation procedure

• Relationships among observed and unobserved variables are


quantitatively defined --- “parameterized”, “modeled”

• A number is derived to indicate discrepancy between the


modeled structure and the data with some side conditions
(assumptions) --- model fit

F = f (data, parameters, combining rule)


F  tr  S  Σ  θ    : ULS
 2
 
• Distributional characteristics of F (as a random variable) are
established according to a set of parametric assumptions (e.g.,
multivariate normal data), which allows for inference on
statistical reliability of some hypotheses, e.g., F = 0 or F1 –
F2 = 0
SEM related myths

• Confirmatory vs. exploratory:

− a priori model specified


− statistical test available (parametric or nonparametric)
− what software used (e.g., AMOS or SPSS)

• An SEM analysis proves something:

“I got a dataset from somewhere and model fit from an AMOS


output is acceptable. Now, I have proved my hypothesis (the
model) and so I am entitled a Ph.D.”

• Hopefully, we will unveil many more in this course


Notational conventions

• Data matrix tall or wide?

The principal component model written for tall data matrix ---
“mathematical” notation

J R J J
X  AB  E X = A B + E
R

I I I
The principal component model written for a set of random
variables (wide data matrix) --- “statistical” notation

x = Ba + e x = B a + e

q q q

• Variables in the data vector x are considered as random


quantities; not interested in their realized values themselves but
their distributions --- equivalent to the data matrix represented
horizontally
• The “LISREL” notation follows this convention as is most
common for SEM; and we’ll do so in this course
Sample covariance matrix

• Given data matrix X (N × q), deviation matrix (column mean


centered) is written as:

Z  LX, L  I  N1 11 : idempotent centering matrix

And unbiased sample covariance matrix:

S 1
N 1
ZZ or 1
N 1
zz with the horizontal notation

c11 , c12  1 z1 z1 , z2 , 

c21 , c22  N 1 z
2

   

You might also like