Larger
samples are better than smaller samples (all other things being equal)
because larger samples tend to minimize the probability of errors,
maximize the accuracy of population estimates, and increase the generalizability
of the results. Unfortunately, there are few sample size guidelines
for researchers using EFA or PCA, and many of these have minimal empirical
evidence (e.g., Guadagnoli & Velicer, 1988).
This is problematic
because statistical procedures that create optimized linear combinations
of variables (e.g., multiple regression, canonical correlation, and
EFA) tend to "overfit" the data. This means that these procedures
optimize the fit of the model to the given data; yet no sample is
perfectly reflective of the population. Thus, this overfitting can
result in erroneous conclusions if models fit to one data set are
applied to others. In multiple regression this manifests itself as
inflated R2 (shrinkage) and misestimated variable regression
coefficients (Cohen, Cohen, West, & Aiken, 2003, pp. 83-84). In
EFA this “overfitting” can result in erroneous conclusions
in several ways, including the extraction of erroneous factors or
misassignment of items to factors (e.g., Tabachnick & Fidell,
2001, p. 588).