Factor Analysis Flashcards
Exploratory Factor Analysis
Used to explore the underlying structure within a set of variables. No correct solution because variables can be extracted and rotated differently.
Steps in EFA
- Check assumptions
- Check factorizability
- Select a method of identification
- Determine how many factors to extract
- Rotate factors
- Interpret factors
- Check fit
Assumptions of EFA
theoretical and statistical relationship between variables (correlations above .32), large sample size, normality and no outliers, continuous data, equal intervals, independence and no multicollinearity (correlations below .85)
Testing Factorizability
Assess whether the data is suitable for factor analysis.
Correlations between .32 and .85
KMO - assesses common variance (above .9 is best, .5 is bare minimum)
Bartlett’s - assesses whether matrices are different (should be significant)
Communality - proportions of variance in each variable that is explained by other variables
Identifying Factors
SPSS provides multiple ways to identify factors.
PCA - assumes no error, default option, DON’T USE
Principal Axis Factoring - extracts for maximum variance
ULS - minimise differences between observed and expected data
GLS - same as ULS but communality given more weight
Maximum Likelihood - most generalisable
Determining Extraction
How many factors should be included in the model? Typically use a combination of methods and test one above and one below.
Kaiser’s - retain all with eigenvalue above 1
Joliffe’s - retain all with eigenvalue about .7
Cattell’s - Use a scree plot and extract at point of inflection
Parallel Analysis - tell SPSS to generate a random data set and compare raw data to random data. When Random > Raw, extract
Rotating Factors
Aids interpretation. SPSS can rotate each factor so that each item loads highly on one factor and low on all the others. Method of rotation depends on aim and theory.
Orthogonal Rotation
Preserves the independence of all factors.
Quartimax (maximises spread)
Varimax (maximises loadings)
Equamax (does both)
Oblique Rotation
Allows factors to correlate (more reflective of real world)
Oblimin (maximises loadings while allowing correlation between factors)
Promax (like oblimin but for large samples)
Interpreting and Naming Factors
Examine pattern matrix. Meaning of factors can be inferred from examining high loading items (above .32). Naming is arbitrary, but be sensible.
Checking EFA Fit
Check residuals (should be small) Chi Square is usually significant which indicates poor fit between data and model, but it is highly influenced by sample size (which should be large)
Common Issues in EFA
Factors are unclear - check extraction and rotation
Items load on multiple or no factors - poor items
Low factor loadings - to many factors, not enough data
Poor model fit - check assumptions and factorizability, over/under extraction.