Dept of PHS Division of Biostats amp Bioinf Biostatistics Shares Resource Hollings Cancer Center Cancer Control Journal Club March 3 2016 Motivating Example Goals of paper 1 See if previously defined measurement model of hopelessness in advanced cancer fits this sample Confir ID: 681835
Download Presentation The PPT/PDF document "Factor Analysis Liz Garrett-Mayer, PhD" is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
Factor Analysis
Liz Garrett-Mayer, PhD
Dept
of PHS, Division of
Biostats
&
Bioinf
Biostatistics Shares Resource, Hollings Cancer Center
Cancer Control Journal Club
March 3, 2016Slide2
Motivating ExampleSlide3
Goals of paper
1. See if previously defined measurement model of hopelessness in advanced cancer fits this sample. Confirmatory Factor Analysis.
2.
Describe the factor structure in these two subpopulations (curative and palliative). Exploratory Factor Analysis.
3. 4. Evaluate ‘stability’ of factor structure: does the factor structure stay the same after 12 months? Confirmatory Factor Analysis.Slide4
(Exploratory) Factor
Analysis
Data reduction tool
Removes redundancy or duplication from a set of correlated variables
Represents correlated variables with a smaller set of “derived”
variables (aka “factors”).
Factors are formed that are relatively independent of one another.
Two types of “variables”:
latent variables: factors
observed
variables (aka manifest variables; items)Slide5
Examples
Diet
Air pollution
Personality
Customer satisfaction
Depression
Quality of
LifeSlide6
Some Applications of Factor Analysis
1.
Identification of Underlying Factors
:
clusters variables into homogeneous sets
creates new variables (i.e. factors)
allows us to gain insight to categories
2.
Screening of Variables
:
identifies groupings to allow us to select one variable to represent many
useful in regression (recall collinearity)
3.
Summary
:
Allows us to describe many variables using a few factors
4.
Clustering of objects
:
Helps us to put objects (people) into categories depending on their factor scoresSlide7
“Perhaps the most widely used (and misused) multivariate
[technique] is factor analysis. Few statisticians are neutral about
this technique. Proponents feel that factor analysis is the
greatest invention since the double bed, while its detractors feel
it is a useless procedure that can be used to support nearly any
desired interpretation of the data.
The truth, as is usually the case,
lies somewhere in between
.
Used properly, factor analysis can
yield much useful information; when applied blindly, without
regard for its limitations, it is about as useful and informative as
Tarot cards.
In particular, factor analysis can be used to explore
the data for patterns, confirm our hypotheses, or reduce the
Many variables to a more manageable number.
-- Norman
Streiner
,
PDQ Statistics
Slide8
Exploratory Factor Analysis
Takes a set of variables thought to measure an underlying latent variable:
Determines which ones “hang together”
Identifies how many ‘dimensions’ there are to the latent variable of interest
Determines which are the strongest variables:
which ones contain most of the information
Which ones might be able to be removed due to either redundancy or because they don’t “hang” with other variables.
One
of the primary goals of factor analysis is often to identify a
measurement model
for a latent variable
This includes
:
I
dentifying
the items to include in the model
Identifying
how many ‘factors’ there are in the latent
variable (i.e. the dimensionality of the latent variable).
I
dentifying
which items are “associated” with which
factorsSlide9
In our example: “Hopelessness” as measured by the Beck hopelessness scaleSlide10
Beck hopelessness scaleSlide11Slide12
Goals of this EFA
What ‘structure’ is there to hopelessness in these patient populations?
Can we come up with ‘components’ of hopelessness? If so, what do they look like?
How do we do this?
Factor analysis is PURELY based on the correlation matrix (or covariance matrix) of your items.
It searches for ‘commonality’ among the items based on the correlations
Think about it: if the items are all measuring the same ‘traits’. The correlation patterns can be used to see which of the items cluster together, which don’t and how much of each item is simply ‘noise.’
Note about Likert scale variables: Very noisy! Correlation matrices are more apt for truly continuous variables. Slide13
Graphically, a two factor EFA (with 7 items in the scale)
F1
F2
y2
y1
y3
y4
y5
y6
y7Slide14
Mathematically, the 2 factor EFA (with 20 items in the scale)
Interpretations:
F
1
and F
2
are the latent variables (e.g. F
1i
is the value of the
i
th
person’s F
1
)
The
λ
’s are called “loadings” and each one represents, in a loose sense, the correlation between each item from the scale and the factor.Slide15
Graphically, a two factor EFA (with 7 items in the scale)
F1
F2
y2
y1
y3
y4
y5
y6
y7
λ
11
λ
12
λ
2
1
λ
2
2
λ
31
λ
71
λ
32
λ
72
e
1
e
3
e
4
e
5
e
6
e
7
e
2Slide16
Some statistical stuff
Without additional assumptions, the model would be ‘unidentifiable’ and also hard to interpret.
Assumptions
:
F
1
and F
2
are statistically independent (uncorrelated
) in most implementations
F
1 and F2 are
each normally distributed with mean 0, variance 1
Conditional on the latent variables, the error terms are independent.Slide17
Spangenberg Paper
Recruited participants as part of a prospective observational study, investigating meaning-focused coping and mental health in cancer patients.
732 eligible adult cancer patients receiving treatment with curative or palliative intention in inpatient and outpatient cancer care facilities in Northern Germany were asked to participate.
At baseline, 315 patients participated.
At follow-up, 158 could be reassessed at 12 months.Slide18
Results from Beck, curative treatment group (n=145)Slide19
Factor 1 comprises items reflecting mainly pessimistic/resigned beliefs (e.g. Item 12), whereas Factor 2 especially contains items reflecting positive beliefs explicitly referring to the future (e.g. Item 5).
It is noteworthy that Factor 2 solely includes positively worded items, whereas Factor 1 includes negatively worded items only.
Factor 1 includes Items 2, 9, 12, 14, 16, 17, and 20 (curative sample, Cronbach alpha 0.88; palliative sample, alpha=0.85).
Factor 2 includes Items 5, 6, 8, 10, 15 and 19 (alpha=0.73 in both samples). Both factors are moderately correlated (r=0.45)Slide20
Let’s reverse: why two factors?
How do you figure out how many factors? That is, what is the “
dimensionality
” of the latent variable?
For a dataset with, let’s say 20 variables, fit a principal components analysis (aka PCA, which is, loosely, a kind of factor analysis).
The PCA creates a matrix of weights from the data from which you can generate composite variables.
The weights are chosen so that the 1
st
component (i.e. a weighted average of the variables) explains the maximum amount of variance in the variables.
The weights for the 2
nd
component are chosen to maximize the variance remaining in the data AFTER having already computed the 1
st
PC
And so on for the remaining 18 (20 – 2 = 18) components.
What?
This isn’t as abstract as it sounds
There are scales out there that use this kind of ‘component’ or ‘composite’ variable approach.
Example: 0.5x
1
+ 0.75x
2
+ x
3
= z
This approach just finds the optimal weights to maximize the variability explained. Slide21
Eigenvalues
Each of the principal components (aka eigenvectors) has a corresponding value called an ‘eigenvalue’ which represents
the amount of variability explained by the component
.
The sum of the eigenvalues is equal to the number of variables in your analysis (e.g. for
Spangenberg
paper, the sum of the eigenvalues is 20).
There are several rules of thumb for using the eigenvalues to help determine the number of ‘components’ or ‘factors’ to keep.
1. Keep as many components as have eigenvalues greater than 1.
2. use a
screeplot
to determine how many components to keep.
3. Preset a threshold for percent variance explained and keep enough to explain a sufficient amount of variance.
Think about what it means to have an eigenvalue of 1 or greater?Slide22
Screeplot examplesSlide23
From Spangenberg:
Initially, five eigenvalues were >1 in the curative sample (7.42, 2.04, 1.53, 1.25, and 1.01).
In the palliative sample, four eigenvalues were >1 (7.13, 1.96, 1.43 and 1.14).
The scree plot suggested a two factor structure in both patient groups.
Note: THESE SCREE PLOTS ARE INCOMPLETE!Slide24
In terms of percent variance explained
Divide the eigenvalue by the number of variables in the model and you get the incremental variance explained.
You can also calculate cumulative variance of the first, for example, 3 components.Slide25
Two factor solution
Explains about 50% of the variance in the data.
Thus, 50% of the information in the data is ‘discarded’ when only two components are retained.
However, there were TWENTY variables that you started with.
And, it’s quite possible that there is a lot of noise
Likert scale variables
How much is noise? How much is ‘signal’
Glass half-full: With only TWO components, you can explain as much as almost TEN variables (on average). Slide26
Pick number of factors
Based on the examination of eigenvalues, determine the number of ‘factors’ you want to retain in a factor analysis.
Fit a factor analysis where you pre-specify the number of factors. Slide27
Interpretable?
The problem with PCA and an
unrotated
factor analysis is that the factors are hard to interpret.
The first component or factor usually has about equal loadings (+/- depending on the direction of the item) for all items.
The second component may have some high, some low loading, but usually not very interpretable.
Solution? Factor rotation.Slide28
Rotation: More statistical stuff
In
principal components, the first factor describes most of variability.
After
choosing number of factors to retain, we want to
spread
variability
more
evenly among factors.
To
do this we “rotate” factors
:– redefine factors such that loadings on various factors tend to be very high (-1 or 1) or very low (0)
– intuitively
, it makes sharper distinctions in the meanings of the factors
We
use “factor analysis
” for
rotation NOT principal components!Slide29
How can we do this? Doesn’t it change our ‘answer’?
Statistically, it doesn’t. The percent variance is the same, etc.
For a factor analysis solution to be calculated, there have to be constraints, or assumptions.
Some of them are
Factors are normally distributed
Factors have mean 0 and variance 1
In the initial solution, another constraint is that the first factor explains the most variance, the 2
nd
factor explains the ‘next most’ conditional on the first factor, etc.
What if we change this last assumption and instead create a different constraint to make it identifiable? That is, focus on ‘shrinking’ loadings?Slide30Slide31
Rotation types
“orthogonal”: maintains independent factors (
i.e
uncorrelated factors)
“oblique”: allows some dependence. Usually not terribly different from orthogonal, but loadings are often ‘shrunk’ more towards 0 or 1 (or -1).
Spangenberg
assumed that factors would be likely to be correlated, so they used oblique rotation. Slide32
Aside
Principal factors vs. principal components.
The defining characteristic that distinguishes between the two factor analytic models is that
in principal components analysis we assume that
all
variability
in an item should be used in the analysis
, while in
principal factors analysis we only use the variability in an item that it has in common with the other items
. In most cases, these two methods usually yield very similar results. However, principal components analysis is often preferred as a method for data reduction, while
principal factors analysis is often preferred when the goal of the analysis is to detect structure
.
(http://www.statsoft.com/textbook/stfacan.html)Slide33
Factor 1 comprises items reflecting mainly pessimistic/resigned beliefs (e.g. Item 12), whereas Factor 2 especially contains items reflecting positive beliefs explicitly referring to the future (e.g. Item 5).
It is noteworthy that Factor 2 solely includes positively worded items, whereas Factor 1 includes negatively worded items only.
Factor 1 includes Items 2, 9, 12, 14, 16, 17, and 20 (curative sample, Cronbach alpha 0.88; palliative sample, alpha=0.85).
Factor 2 includes Items 5, 6, 8, 10, 15 and 19 (alpha=0.73 in both samples). Both factors are moderately correlated (r=0.45)Slide34
A few more details
Keeping vs. dropping items
Should look at the fitted model (before or after rotation) to determine the variable’s “uniqueness.”
Communality = 1 – Uniqueness
Communality is a measure of how much is ‘shared’ between the item and the latent variable structure
Uniqueness is what is left over (i.e. noise)
Uniqueness DOES depend on the number of factors retained.Slide35
Example of uniquenessesSlide36
A few more details
EstimationSlide37
Next steps
You can use your model to calculate the estimated factor scores for each subject in your dataset
Confirmatory factor analysis: Restrictive approach which forces some of the arrows (i.e. loadings) to be zero.
EFA: descriptive approach to determine structure
CFS: test a particular structure. Applications?Slide38
CFA: compare structure to a fixed model
F1
F2
y2
y1
y3
y4
y5
y6
y7
λ
11
λ
2
1
λ
31
λ
4
2
λ
72
e
1
e
3
e
4
e
5
e
6
e
7
e
2
λ
52
λ
62Slide39
Spangenberg: Used CFA
CFA was used to determine if the patients in this study demonstrated the same factor structure for BHS as advanced cancer patients.
They found that the structure in this sample of patients differed from previous studies. Slide40
CFA
CFA is often thought of as a ‘special case’ of structural equation models
You make assumptions about how the variables are associated using ‘arrows’ to join variables (both latent and observed). Slide41
“Stability”
Spangenberg
et al. also evaluated structure over time. Did the factors stay the same in their structure (i.e. loadings, dimensionality)?
Some measures said it was acceptable, other that is wasn’t.
Problem with this paper: they provide the test statistics, but do not show us the descriptive statistics (i.e. what were the loadings in the fit on the 12 month data?)Slide42
More details on the process
And, a lot more math, statistics and matrices!
http://people.musc.edu/~
elg26/teaching/psstats1.2006/factoranalysis.pdf