Factor analysis is robust for non normal data. in factor analysis we statndardise the variable and then we use FA. but before go to the FA, it is very important to check sampling adequecy of the data by using KMO method and ur correlation matrix it should be diffrent from identity matrix. if KMO values getting more than 0.5 then u can move for FA and u can check whether correlation matrix deffrent from identity matrix or not by Bartlet test of sphericity. then extract factors by using eign value criterion or scree plot. then rotate factors by varimax rotation( orthogonal rotation) , factors should be independent.

Especially data generated using
Likert-type items - a common technique in IS research - frequently do not meet
the requirements of the method applied, such as normal distribution of the individual
variables or multivariate normal distribution

In many cases, latent variables
represent users' attitudes, norms, values and intentions

Although the underlying latent construct may be seen as a
continuum, the items are measured on an ordinal scale

researchers have a multitude of options such as data
transformation, factor extraction method, rotation method, and choosing
the optimum number of factors when conducting EFA

Although most of the attention is given to confirmatory procedures,
EFA is frequently used when new scales are developed or the validity
of a measurement model is assessed

Its major objective is to reduce a number of observed variables to fewer
unobserved factors in order to enhance general interpretability and to detect
hidden structures in the data. Frequently, these structures are used as constructs
in sophisticated models displaying aspects of human behavior

Exploratory factor analysis relies on the estimation of the correlation matrix.
Once this correlation matrix is available, the loadings and uniquenesses, and
subsequently the factor scores are estimated. The correlation matrix is usually
estimated with the sample correlation matrix, which is the empirical sample covariance
matrix standardized by the empirical variances. This classical approach is
most adequate if the data are multivariate normally distributed. However, if the
data distribution is deviating from this ideal distribution, the estimated correlation
matrix can be severely biased. Figure 1 shows this effect. The estimated correlations
are visualized by the ellipses which, in case of bivariate normal distribution

A prominent way is to use the Minimum Covariance Determinant (MCD)
estimator, which looks for a subset of observations (e.g. at least 50% of the observations)
with minimal determinant of the empirical covariance matrix

A multivariate normal distribution is required when using maximum
likelihood as the factor extraction method, whereas a principal component analysis
and a principal factor analysis require elliptical symmetry. In these cases,
normal distribution is not a prerequisite, but the results may still be strongly influenced
by the occurrence of non-normally distributed data and outliers because of
their dependence on the correlation and the covariance matrix [40]. In such cases
a robust method is preferable.

Ordinal items of the Likert type are often seen as amenable to factor analysis, particularly when it is thought that ‘assignment of ordinal categories to the data does not seriously distort the underlying metric scaling’ (Mueller and Kim).
Likert-type items are more often than not regarded as conforming to this requirement. They are then referred to as quasi-interval, and seen as appropriate for being factor analysed

A confirmatory factor analysis (CFA), in which the model is tested that all k items are indicators of a single underlying dimension

Faktör Analizi

İSTATİSTİK