Practical issues and assumptions

Question 1

Assumptions checklist

Answer 1

SMMNOF

1) Sample size
2) Missing data
3) Multicolinearity
4) Normality
5) Outliers
6) Factorisability of the correlation matrix

Question 2

1) Sample size

Answer 2

Rule of thumb: 300+

OR 10:1 (participant: variable)

Requirements depend of the factorisability

in SPSS:

KMO (Keiser-Meyer-Olkin) measure of sampling adequacy: >.6 is fine, higher is better

Question 3

2) Missing data

Answer 3

:) If there is a large sample size, it is not too problematic to lose some data

Remove ALL of that participants’ data

Question 4

3) Multicolinearity

Answer 4

:) not a problem for PCA

:( may be a problem for FA

SPSS: look at the OBSERVED CORRELATION MATRIX: r>.9 is a problem

Look at INITIAL COMMONALITIES: this shows the overlap in variables using the squared multiple correlations (SMCs- each V correlated with all the others)
>.9 is a problem (too correlated)
close to 0 is a problem (it doesn’t add anything, it is explained by all the other variables)

Question 5

4) Normality

Answer 5

Multivariate normality required when testing for a number of factors .

Normality of single variables not required, but useful for finding an adequate solution.

Test skewness and kertosis of single variables

Question 6

Skewness:

Answer 6

Used for normality or single variables

0=normal

Skewness statistic/ its standard error= z= (+ or -) 3.29 and p

Question 7

Kertosis

Answer 7

Used for normality or single variables

(The concentration of scores)

Kertosis statistic/ its standard error= z= (+ or -) 3.29 and p

Question 8

5) Outliers

Answer 8

May distort correlational relationships between variables.

Address by: transforming the data (if you need to based on the skewness and kertosis) by winorizing (move the outliers tot he next highest or lowest score’s value) or triming (cut them off a %),

or remove

Question 9

6) Factorisability

Answer 9

SPSS:

1) KMO AND BARTTLET’S TABLE: Barttlet’s test of spericity: not significant=problem (this is sensitive so will generally no be significant even if it’s poor :( )
2) CORRELATION MATRIX: look for variables with >.30 relationships (also look at matrices of partial correlation to needed)
3) ANTI-IMAGE CORRELATION MATRIX: shows the negatives of partial correlations between pairs with the effects of other variables removed- look for small values of the off-diagonal elements.