Chapter 15: Factor Analysis and Reliability Analysis
Chapter 15: Factor Analysis and Reliability Analysis
I. Overview
b. Researchers often measure or observe several variables and then want to group highly
correlated variables together into a smaller number of factors, or constructs.
i. This is what data reduction means: To reduce many measured variables into a
smaller number of constructs.
c. The observed, or measured, variables that are grouped together are called indicators of
the underlying, or latent, construct.
i. For example, the latent construct of self-esteem may be indicated by several
observed variables, including how one feels about one’s appearance, one’s
intelligence, one’s ability to get along with others, and one’s athletic ability.
d. Factor analysis reveals which observed variables are strongly correlated with each other
and therefore should perhaps be grouped together.
b. Principal Components Analysis (PCA) arranges the variables into separate factors based
on how strongly correlated variables are with each other.
c. These factors are then rotated to create maximum similarity among the strongly
correlated variables within each factor and maximum distance between each of the
factors.
i. Orthogonal factor rotation is designed to create maximum distinctness among the
factors and works well when the researcher does not expect the factors to
correlate strongly with each other.
ii. Oblique factor rotation assumes that factors will be correlated with each other
and produces factors that allow overlap, or similarity, among them.
d. Eigenvalues and percentage of variance explained are used by the researcher to determine
how many factors from the PCA should be retained.
i. This decision is also influenced by the researcher’s interpretations of the factors.
e. The items with the strongest factor loadings for a given factor tend to define that factor
and reveal what the underlying construct is for that factor.
i. Factor loadings are an indication of how strongly individual items are associated
with each factor.
f. Items that do not load strongly on any factor indicate that these items are not good
indicators of any of the underlying constructs represented by the factors created in the
PCA.
g. Items that have strong factor loadings on more than one factor are said to cross load with
multiple factors.
i. These items are also considered to be weak indicators of the underlying
constructs because they do not distinguish well between constructs.
a. In this method of factor analysis, a number of measured variables are put into the analysis
and are separated into separate factors based on statistical measures.
i. The researcher does not define which items belong on which factors before
performing the factor analysis.
ii. It is a method of exploring the data to find out which measured items correlate, or
load, with each other.
b. This is a data- and statistics-driven model of data reduction more than a theory-based
method of grouping measured variables together.
i. It is important for the researcher to closely examine the factors that emerge from
an EFA to see if they make sense conceptually and theoretically.
a. The researcher decides ahead of time how to organize the observed variables (i.e.,
measured items) into factors.
c. A variety of fit statistics are used to determine how well a hypothesized factor structure
fits the data.
d. Confirmatory factor analysis is often used as part of a larger set of statistical techniques
known as structural equation modeling.
V. Reliability Analysis
a. The purpose of a reliability analysis is to determine how well a set of items, i.e., observed
variables, go together into a single scale.
i. This analysis also reveals how strongly each item in the scale is associated with
the overall scale. This is called item-total correlations.
ii. In survey research, building scales out of several individual measured items is
important for at least two reasons:
1. Using multiple items to measure a single latent construct allows
researchers to build scales that are broader than any single items could
measure
a. E.g., Satisfaction with work involves several aspects (pay,
respect, interesting work, likeability of co-workers, etc.) that
cannot be assessed with a single item.
2. Using multiple items to measure a single latent construct helps
researchers have some confidence that participants are understanding the
items as the researcher intended.
b. The statistic that results from a commonly used reliability analysis is the Cronbach’s
alpha coefficient.
i. This coefficient has a maximum value of 1.0.
ii. Generally speaking, when a collection of items (i.e., a scale) has a Cronbach’s
alpha of .70 or larger, the scale is considered to be reliable.
c. This type of reliability analysis refers to the internal consistency of a set of variables.
i. This differs from test-retest reliability.
VI. Summary
a. Both factor analysis and reliability analysis are statistical techniques used to reduce a
larger set of measured items (i.e., observed variables) into a smaller set of latent
constructs.
b. Researchers typically use factor analysis first to organize the items into constructs and
then use reliability analysis to determine how well each construct holds together.