06/10/2023

RESEARCH GROUP: MARKETING, INNOVATION AND ORGANIZATION OBJECTIVE

FACTOR ANALYSIS
• General term for a class of multivariate techniques

• Interdependence technique
- No distinction DV/IV: all variables considered simultaneously
- Search for structure/patterns in data

• Data reduction while preserving as much info as possible

FACTOR ANALYSIS • Discover latent dimensionality

AY 2023-2024 • Highly correlated variables (= variables that share variance) are grouped into distinct sets (i.e. “factors”)
Laura De Kerpel

1 2

RUNNING A PCA IN SPSS RUNNING A PCA IN SPSS

DATASET: pleasure_and_planning.sav STEP BY STEP GUIDE

A supermarket chain asked 500 of its customers to fill in a questionnaire which 1. Data suitability
contained 12 questions about shopping behaviour, all on a 7 point Likert scale 2. Sample requirements
3. Usefulness of Factor Analysis
2 underlying dimensions found: 4. Number of factors
ü “Pleasure”
5. Interpretation
ü “Planning”
6. Data reduction

3 4

3 4

1
 06/10/2023

STEP 1 STEP 2
ü Suitability of data ü Sample requirements
• Correlation matrix: Interval or ratio scaled variables
• Min 50 (better: min 100)
(Avoid non-metric variables)
• 5:1 ratio (better 10:1 or even 20:1)
• Item scales: Likert scales, semantic differential…
• High case-to-variable ratio à to avoid overfitting!
à Same level of measurement (standardize)
• Underlying structure exists? Factor analysis will always produce factors!
• Homogeneous sample?
“garbage in, garbage out”
E.g. Men vs. women and shoe characteristics
Men: 3 dimensions à appearance-situation-comfort
Women: 5 dimensions à style-material-colour-situation-
comfort

5 6

5 6

STEP 3 STATISTICAL ASSUMPTIONS IN FA

ü Factor analysis = useful?
① Correlation matrix
Check if statistical assumptions are met? A FA is only meaningful if the variables involved are
= Table showing the intercorrelations among all variables
sufficiently correlated to one another, so some degree of multicollinearity is desirable

Visual inspection: high (>.30) and not equal (because some structure must exist)

Measures to check statistical assumptions:

①Correlation matrix
②Partial correlation/ Anti-image correlation
matrix
③Bartlett’s test of sphericity
④KMO
7 8

7 8

2
 06/10/2023

STATISTICAL ASSUMPTIONS IN FA STATISTICAL ASSUMPTIONS IN FA

Analyze > Correlate > Bivariate

9 10

9 10

STATISTICAL ASSUMPTIONS IN FA STATISTICAL ASSUMPTIONS IN FA

2) STATISTICAL ASSUMPTIONS
② Partial correlation/Anti-image correlation matrix
A FA is only meaningful if the variables involved are sufficiently correlated to one another,
• Calculate partial correlation
so some degree of multicollinearity is desirable
= the correlation between 2 variables that remains when the
effects of other variables are taken into account

Partial correlation should be low (high = .70)

Measures to check statistical assumptions:

①Correlation matrix • Anti-image correlation matrix: negative values of partial correlation should all be
low
②Partial correlation/ Anti-image correlation
matrix
③Bartlett’s test of sphericity
④KMO
11 12

11 12

3
 06/10/2023

STATISTICAL ASSUMPTIONS IN FA
STATISTICAL ASSUMPTIONS IN FA
Analyze > Dimension Reduction > Factor > Descriptives

13 14

13 14

STATISTICAL ASSUMPTIONS IN FA STATISTICAL ASSUMPTIONS IN FA

2) STATISTICAL ASSUMPTIONS ③ Bartlett’s test of sphericity
Are the variables significantly correlated?
A FA is only meaningful if the variables involved are sufficiently correlated to one another,
H 0: correlation matrix = identity matrix (i.e. the variables are uncorrelated)
Identity matrix
so some degree of multicollinearity is desirable
④ Measure of sampling adequacy (KMO*)
= Measure calculated for both the entire correlation matrix and each individual
variable (see diagonal on anti-image for individual scores); ranges from 0 to 1
Measures to check statistical assumptions:
.80: meritorious
.70: average
①Correlation matrix .60: mediocre
.50: absolute minimum
②Partial correlation/ Anti-image correlation <.50: unacceptable

matrix
③Bartlett’s test of sphericity
④KMO
15 * Kaiser-Meyer-Olkin Measure of Sampling Adequacy 16

15 16

4
 06/10/2023

STATISTICAL ASSUMPTIONS IN FA STATISTICAL ASSUMPTIONS IN FA

Analyze > Dimension Reduction > Factor >
Descriptives

17 18

17 18

STEP 4 & 5 TYPES OF FACTOR EXTRACTION METHODS

ü Number of factors and interpretation CFA (Common or True Factor Analysis) ≠ PCA (Principal Component Analysis)
Related but not identical à different calculation of factors
Analyze > Dimension Reduction > Factor >
Why PCA?
Extraction

1. Default in SPSS
2. Data reduction: account for max total variance in min factors
3. No factor indeterminacy problem (>< Common FA)
4. Invalid calculation of communalities

However, similar results if:

• #variables > 30
• communalities >.60 for most variables

19 20

19 20

5
 06/10/2023

STEP 4 & 5 STEP 4 & 5

Analyze > Dimension Reduction > Factor > Rotation & ü Number of factors & interpretation
Options
COMMUNALITIES
• Communalities: indicator of explained variance in every
variable by the factor solution

• Extraction: % variance explained by the factors, so low

values are bad

e.g. 70,3% of the total variance of “Organize my

shopping well” is explained by the 3 underlying
dimensions
The eigenvalue cutoff point
ROTATION METHOD = VARIMAX

21 22

21 22

STEP 4 & 5
ü Number of factors & interpretation COMPONENT MATRIX

TOTAL VARIANCE EXPLAINED Only contains values for the 3 relevant factors
• Initial Eigenvalues = explained variance by each factor
• Extraction Sums of Squared Loadings : the factors that passed the Kaiser Criterion (= Latent root criterion) Factor loadings: correlation between a set of factor
(i.e. Eigenvalue > 1) scores and a set of scores for an original variable (only
• Total variance explained: social sciences > 60% (rule of thumb) if orthogonal!)
Total variance = 12

Stop extracting

23 24

23 24

6
 06/10/2023

ROTATED COMPONENT MATRIX STEP 4 & 5

• More exclusive relationship between factor and WHAT ARE “GOOD” FACTOR LOADINGS?
variables
There is no univocal way of deciding on which variables to retain per factor, this is where
• Ideally a variable loads high on one single factor and statistics stop and common sense takes over.
low on all the others
(= unidimensionality) Some rules of thumb:
• Variables loading high on several factors are not • Absolute minimum: .50
desirable • A variable should load high on 1 factor and low on the others (unidimensionality, no
(= crossloading) remove cross-loadings)
unidiemnsionality
• The higher the loading, the better
• Items 10, 11 and 12 will be removed remove
• Factor composition must make logical sense
because their composition doesn’t make sense
crossloading

• Item 9 will be removed because it doesn’t match remove

the researchers expectation -> Throw the variables out

25

26

25 26

STEP 4 & 5 STEP 4 & 5

SCREE TEST CRITERION 1 4

• All factors contain some unique variance, the proportion of unique variance increases in
later factors
• Extract factors before unique variance > common variance 2
Iterative
Scree Plot
• Where curve straightens (“elbow”) 4. 5

4 process
3. 5

3
Eigenvalue

2. 5

1. 5

0. 5

0
1 2 3 4 5 6 7 8 9 10 11 12
C om ponent N um ber

Extract 3 factors 27 28

27 28

7
 06/10/2023

STEP 4 & 5 STEP 6

After rerunning the analysis, only 2 factors appear: üData reduction
Reversing scores:

Transform > Recode into different variables

29 30

29 30

STEP 6 STEP 6
üData reduction üData reduction
Reversing scores:
CREATING COMPOSITE MEASURES
= combining several variables that measure the same concept into a single variable

1) Summated scale: average score of selected high loading variables

Transform > Compute variable > Mean(var1,var2,var3)

2) Factor scores: factor loading and individual score on all variables

Analyze > Data Reduction > Factor > Scores > Save as variables > Regression

31 32

31 32

8
 06/10/2023

COMPUTE SUMMATED SCALE RELIABILITY

Transform > Compute Variable Use mean() function (separate with commas)
Analyze > Scale > Reliability Analysis

Name new variable

! Don’t forget to reverse negative variables

à A new variable named ‘ss_planning’ will be added to your dataset

Transform > Recode into different variables
33 34

33 34

RELIABILITY RELIABILITY
Always calculate the Cronbach’s Alpha

Rule of thumb
If the increase is very
small, just keep as the α ≥ 0.9: excellent
The higher the better analysis 0.8 ≤ α < 0.9: very good
If it’s the big increase, 0.7 ≤ α < 0.8: good
it’s worth 0.6 ≤ α < 0.7: acceptable
α < 0.6: unacceptable

Warning!
• Items should logically match (garbage in, garbage out)
• Min. 3 items (preferably) à If only 2 items, calculate Pearson’s r (Analyze > Correlate > Bivariate)

Alpha could not be increased by

deleting one of the items
35 36

35 36

9
 06/10/2023

PEARSON’S R EXERCISES - DIY

• r = value between -1 and 1 How many and which factors can you extract?

• SAQ_2019.sav (Statistics Anxiety Questionnaire)

High degree: If the coefficient value lies between ± 0.50 and ± 1, then it is said to be a strong
correlation. • Personality_2019.sav (Big Five)
Moderate degree: If the value lies between ± 0.30 and ± 0.49, then it is said to be a medium
correlation.
Low degree: When the value lies below + .29, then it is said to be a small correlation.

When the value is zero, there is no correlation. When the value is (near) +1 or -1, there is a
perfect correlation.

37 38

37 38

EXERCISES FIVE FACTOR MODEL

Dataset: SAQ_2019.sav (Andy Field) Hofstee & De Raad (1999)

Statistics Anxiety Questionnaire “Big Five”

Items Q001 to Q023, n = 2571 • 5 general personality traits > all possible adjectives (lexical approach)
• Widely used: solution stable across cultures, ages, gender…
• Used for psychological diagnostics, assessment and recruiting
1. Run a FA on the data to see how many and which factors emerge

2. Check the assumptions

3. Create summated scales and test your dimensions for reliability

39 40

39 40

10
 06/10/2023

BIG FIVE TRAITS FIVE FACTOR MODEL

1) OPENNESS TO EXPERIENCE Dataset: personality_2019.sav
Novelty, adventure, creation vs. Conservatism, routine
44 personality traits, n = 459 (5 point scale)
2) CONSCIENTIOUSNESS
Organised, disciplined vs. Easy-going, careless

3) EXTRAVERSION/INTROVERSION 1. Run a FA on the data to see whether the Big Five structure is supported.
Social, gregarious vs. Solitary, reserved Remove variables that do not load satisfactory

2. Check the assumptions

4) AGREEABLESNESS
Friendly, altruistic vs. Cold, unkind 3. Create summated scales and test your dimensions for reliability

5) NEUROTICISM
Nervous, mentally unstable vs. Stable, resilient

41 42

41 42

11