FACTOR ANALYSIS
NASC
Factor Analysis…
A data reduction technique designed to represent a wide range of attributes on a smaller number of dimensions.
Aim is to identify groups of variables which are Aim is to identify groups of variables which are relatively homogeneous.
Groups of related variables are called ‘factors’.
Purposes
The main applications of factor analytic techniques are:
(1) to reduce the number of variables and
NASC@Courtesy Dr. Thagunna
(1) to reduce the number of variables and (2) to detect structure in the relationships
between variables, that is to classify
variables.
Factor 1 Factor 2 Factor 3
Conceptual Model for a Factor Analysis with a Simple Model
e.g., 12 items testing might actually tap only 3
underlying factors
Conceptual Model for Factor Analysis (with cross-loadings)
NASC@Courtesy Dr. Thagunna
Common Factor Model
It is suggested that X
1, X
2, and X
3are functions of two underlying factors, F
1and F
2. It is assumed that each X variables are linearly related to the two factors as shown in the following model.
X
1= β
11F
1+ β
12F
2+ e
1β β
X
2= β
21F
1+ β
22F
2+ e
2X
3= β
31F
1+ β
32F
2+ e
3The error terms e
1, e
2, and e
3, serve to indicate that the hypothesized
relationships are not exact. In the vocabulary of factor analysis, the
Expected Structure of Loadings
Loading (F1) β
i1Loading (F2) β
i2X1 + 0
It is expected that the loadings have roughly the structure as shown in the table.
X1 + 0
X2 0 +
X3 0 +
Of course, the zeros in the preceding table are not expected to be exactly equal to zero. By `0' we mean approximately equal to zero and by `+' a positive number substantially different from zero.
NASC@Courtesy Dr. Thagunna
Model Assumptions
A1: The error terms e
iare independent of one another and
E(e
i) = 0 and Var(e
i) = σ
i2,
A2:The unobservable factors are independent of one another.
It is also assumed that the factors and error terms are It is also assumed that the factors and error terms are independent.
As for the factor means and variances, the assumption is that the
factors are standardized: E(F
j) = 0 and Var(F
j) = 1. It is an
assumption made for convenience; since the factors are
unobservable, we might as well think of them as measured in
Implications of Assumptions…
The variance of X
ifrom the model can be expressed as
Var(X
i) = Var(F
1) + Var(F
2)+ Var(e
i) = + +
We see that the variance of X
iconsists of two parts:
( + ) and .
• The first part is called communality of the variable. It is the
• The first part is called communality of the variable. It is the part of Var(X
i) explained by the common factors F
1and F
2.
• The second part is called specific variance of the variable. It is the part of Var(X
i) unable to explain by the common factors.
The covariance of any two observable variables, X
iand X
j, from the model can be expressed as
Cov(X
i, X
j) = β
i1β
j1Var(F
1)+ β
i2β
j2Var(F
2) = ββββ
i1ββββ
j1+ ββββ
i2ββββ
j2 NASC@Courtesy Dr. ThagunnaHistory of Factor Analysis?
Invented by Spearman (1904)
Usage hampered by onerousness of hand calculation
Since the advent of computers, usage has thrived, esp. to develop:
esp. to develop:
• Theory
– e.g., determining the structure of personality
• Practice
– e.g., development of 10,000s+ of psychological
screening and measurement tests
Assumption Testing – Factorability
It is important to check the factorability of the correlation matrix
(i.e., how suitable is the data for factor analysis?)
• Check correlation matrix for correlations
NASC@Courtesy Dr. Thagunna
• Check the anti-image matrix for diagonals
• Check measures of sampling adequacy (MSAs)
Bartlett’s
KMO
Rule of thumb: Measures of Sampling Adequacy
Are there several correlations over .3?
Are the diagonals of anti-image matrix > .5?
Is Bartlett’s test significant?
Is Bartlett’s test significant?
Is KMO > .5 ?
Assumption Testing – Factorability (Correlation and partial correlation)
Medium effort, reasonably accurate
Examine the diagonals on the anti-image
correlation matrix to assess the sampling adequacy of each variable
NASC@Courtesy Dr. Thagunna
of each variable
Variables with diagonal anti-image correlations of
less that .5 should be excluded from the analysis –
they lack sufficient correlation with other variables
Assumption Testing – Factorability (Bartlett’s and KMO measure)
Quickest method, but least reliable
Sampling Adequacy predicts whether the data you have collected are likely to "factor well" based on correlation and partial correlation and this is measured by the Kaiser- Meyer-Olkin (KMO) statistic
Global diagnostic indicators - correlation matrix is factorable if:
Bartlett’s test of sphericity is significant and/or
(Null: no correlation among the variables(unit R matrix)
Kaiser-Mayer Olkin (KMO) measure of sampling
adequacy > .5
NASC@Courtesy Dr. Thagunna
Communalities
The proportion of variance in each variable which can be explained by the factors
Also called the explained variation due to factor.
Communalities range between 0 and 1
High communalities (> .5) show that the factors High communalities (> .5) show that the factors
extracted explain most of the variance in the variables being analysed.
Low communalities (< .5) mean there is considerable
variance unexplained by the factors extracted
Eigen Values
EV = sum of squared correlations for each factor
EV = overall strength of relationship between a factor and the variables
Successive EVs have lower values
NASC@Courtesy Dr. Thagunna
Eigen values over 1 are ‘stable’
Explained Variance
A good factor solution is one that explains the most variance with the fewest factors
Realistically happy with 50-75% of the
variance explained
Example: interpreting the communality
Variable (1)
Variance (2)
Loadings of F
1(3)
Loadings of F
2(4)
Communality (5)
% explained (6) = 100×(5)/(2)
Finance 1,0000 .0299 .9995 0.9999 99.9910
Marketing 1.0000 .9941 -.0815 0.9949 99.4940
Policy 1.0000 .9961 .0514 0.9949 99.4920
Overall 3.0000 1.9815 1.0083 2.9898 99.6590
The loadings on F
1are relatively large for marketing and policy but close to zero for finance. On the contrary, the loadings on F
2are relatively large for finance but relatively low for marketing and policy. This solution supports the expectation.
F
1could be interpreted as verbal ability, and F
2as quantitative ability.
NASC@Courtesy Dr. Thagunna
The communalities show that the factor model explains nearly 100%, 99.5%, and 99.5% respectively of the observed variance of finance, marketing and policy grades. Overall, the two factors explain 99.65% of the sum of all observed variances.
The sum of squared loadings on F
1can be interpreted as the contribution of F
1, and that on F
2as the contribution of F
2in explaining the sum of the observed variances.
In our example F explains about 1.9815/3 or 66%, and F about 33.7% of the
Assessment of the First Solution based on R
In our example F
1explains about 1.9815/3 or 66%, and F
2about 33.7% of the sum of the observed variances.
Theoretically,
• the sum of squared loadings, 1.9815, is the largest eigenvalue of R and the loadings on F1 constitute the corresponding eigenvector.
• the sum of squared loadings, 1.0083, is the second largest eigenvalue of R and
NASC@Courtesy Dr. Thagunna
How Many Factors?
A subjective process.
Seek to explain maximum variance using fewest factors, considering:
1. Theory – what is predicted/expected?
Eigen Values > 1? (Kaiser’s criterion)
2. Eigen Values > 1? (Kaiser’s criterion)
3. Scree Plot – where does it drop off?
4. Factors must be able to be meaningfully interpreted &
make theoretical sense?
Cattell & Jaspers (1967) suggest that the number of factors be taken as the number of eigenvalues immediately before the straight line begins.
before the straight line begins.
NASC@Courtesy Dr. Thagunna
Scree Plot
A bar graph of Eigen Values
Depicts the amount of variance explained by each factor.
Look for point where additional factors fail to add appreciably to the cumulative explained variance.
variance.
1st factor explains the most variance
Last factor explains the least amount of variance
Factor Rotation
Factor loadings are not unique. There exist an infinite sets of factor loadings yielding the same theoretical dispersion matrix. The process of obtaining a new set of loadings with some specific objective is called
NASC@Courtesy Dr. Thagunna
factor rotation.
Orthogonal (Varimax)
Oblimin
Factor loading stages
In practice, FA can be carried out in two stages.
• In the first stage, one set of loadings is estimated. These loadings may not agree with the prior expectations, or may not lend themselves to a reasonable interpretation.
• In the second stage, the first set of factor loadings are
• In the second stage, the first set of factor loadings are
"rotated" in an effort to arrive at another set that are more consistent with prior expectations or more easily interpretable.
• variables with cross-loading shall be omitted from the
How do I eliminate items?
A subjective process, but consider:
Size of main loading (min=.5) Size of cross loadings (max=.3?)
Eliminate 1 variable at a time, then re-run, before deciding which/if any items to eliminate next
Number of items already in the factor
More items in a factor -> greater reliability
NASC@Courtesy Dr. Thagunna
More items in a factor -> greater reliability Minimum = 3
Maximum = unlimited
Factor Analysis: an example
suppose that an automobile company asked a large number of questions about different vehicles.
Consider how the different items (features) might be more parsimoniously represented by just a few
constructs (factors).
- Ideally, interval data (e.g., a rating on a k- point
scale), regarding the perceptions of consumers are
required regarding a number of features
NASC@Courtesy Dr. Thagunna
NASC@Courtesy Dr. Thagunna
We are looking for an eigenvalue above 1.0.
Cumulative percent of variance explained.
NASC@Courtesy Dr. Thagunna
Expensive Exciting Luxury
Appeals to Others Attractive Looking Trend Setting
Reliable
Latest Features Trust
Luxury
Distinctive
Not Conservative Not Family
Not Basic
Trend Setting Trust
NASC@Courtesy Dr. Thagunna
Expensive Exciting Luxury
Appeals to Others Attractive Looking Trend Setting
Reliable
Latest Features Trust
What shall these components be called?
Luxury
Distinctive
Not Conservative Not Family
Not Basic
Trend Setting Trust
Expensive Exciting Luxury
Appeals to Others Attractive Looking Trend Setting
Reliable
Latest Features Trust
EXCLUSIVE TRENDY RELIABLE
Luxury
Distinctive
Not Conservative Not Family
Not Basic
Trend Setting Trust
NASC@Courtesy Dr. Thagunna
= (Expensive + Exciting + Luxury + Distinctive – Conservative – Family – Basic)/7
EXCLUSIVE
TRENDY
Calculate Component Scores(summated score)
= (Appeals to Others + Attractive Looking + Trend Setting)/3
= (Reliable + Latest Features + Trust)/3
RELIABLE
NASC@Courtesy Dr. Thagunna
Exclusive Trendy Reliable
Beetle 1.4 6.7 6.9
Hummer 3.9 6.2 6.7
Lotus 4.1 7.3 6.7
Lotus 4.1 7.3 6.7
Minivan -1.67 4.83 6.5
Pick-Up -0.43 4.93 6.3
Not much differing on this dimension.
NASC@Courtesy Dr. Thagunna
Exclusive Trendy Reliable
Beetle 1.4 6.7 6.9
Hummer 3.9 6.2 6.7
Lotus 4.1 7.3 6.7
Lotus 4.1 7.3 6.7
Minivan -1.67 4.83 6.5
Pick-Up -0.43 4.93 6.3
Practical session : using SPSS
Step 1: Open the data file, for example, Example.SAV
Step 2: Click on sequentially: Analyze → Data Reduction → Factor….
Step 3: Move the three variables – X1, X2 & X3 - from Source to Variable box
NASC@Courtesy Dr. Thagunna
Step 4: Click on Descriptives. Activate
• Coefficients
• Significance levels
• KMO and Bartlett’s test of sphericity
• Anti-image Click on Continue.
This will produce correlation matrix and significance of correlations, sampling adequacy and test of sphericity.
Step 5: Click on Extraction. Activate Step 5: Click on Extraction. Activate
• Correlation Matrix
• Unrotated factor solution
• Eigenvalues greater than 1 Click on Continue.
This will produce loadings from correlation
matrix and the number of factors is same as
the number of eigenvalues greater than 1.
Step 7: Click on OK
Step 6: Click on Rotation.
Activate
• Varimax
• Rotated Solution Click on Continue
SPSS will produce 8 tables as outputs with table titles →
1. Correlation Matrix 2. KMO & Bartlett’s Test 3. Anti-image Matrices 4. Communalities
5. Total Variance Explained 6. Component Matrix
7. Rotated Component Matrix
8. Component Transformation Matrix
NASC@Courtesy Dr. Thagunna
Composite Factor Values
Frequently, FA is not an end in itself but an intermediate step on the way to further analysis of the data. In such case we may require the composite values of each factor based on original/standardized data. In recent years, the composite values are generated through three techniques.
• Surrogated variables (A surrogated variable of a factor is a single variable that has the highest factor loading)
• Summated scale (The values of several variables defining a factor are summed and their total or average scores are considered)
• Factor scores (computer generated scores available under
Advantages & Disadvantages of the Techniques
Advantages Disadvantages
Surrogate Variables
Simple to administer and interpret
Does not represent all facets of a factor
Prone to measurement error Factor Scores Represent all variables through
loadings
Best method for complete data reduction
Interpretation more difficult because all variables contribute through loadings
NASC@Courtesy Dr. Thagunna
reduction
By default orthogonal Summated
Scales
Compromise between the surrogate variable and factor score options
Reduce measurement error
Represent multiple facets of a concept
Include only the variables that load highly on the factor and exclude those having little or marginal impact
Not necessarily orthogonal
Require extensive analysis of reliability and validity
Source: Hair et al
Judging Practical Significance of FA
In interpreting factors, a decision must be made regarding the factor loadings. A factor loading is the correlation of the variable and the factor, the squared loading is the amount of the variable's total variation accounted for by the factor. Thus, a 0.3 loading translates to 9 per cent explanation; and a 0.5 loading denotes that 25% of the variation is accounted for by the factor. The loading must exceed 0.7 for the factor to account for 50% of the variation of the variable. Thus larger the absolute size of the factor loading, the more improvement the loading in interpreting the factor matrix using the practical significance as the criteria, we can the factor matrix using the practical significance as the criteria, we can assess the loadings as follows.
• Factor loadings in the range of ± 0.3 to ± 0.4 are considered to meet the minimal level for interpretation of structure
• Absolute value of loading 0.5 or greater are considered practically significant
•
Some Relations Among Output Values
A number of relations exist among outputs, which help us to understand and interpret outputs better. The major relations are the followings when input matrix is p × p correlation matrix.
1. Sum of all eigenvalues = p = total variance of p standardized variables.
2. Sum of squared factor loadings for the j
thfactor = λ
j= j
thlargest eigenvalue
NASC@Courtesy Dr. Thagunna