The confusion between Confirmatory Factor Analysis (CFA) and Principal Component Analysis (PCA) is one of the most persistent misunderstandings in psychological methodology. Despite having coexisted in textbooks for decades and their differences being conceptually clear, they continue to be used interchangeably in a surprising percentage of published articles. Part of the problem is that SPSS groups both under the "Dimension Reduction" menu, which contributes to the idea that they are variants of the same thing. They are not.
The fundamental difference is conceptual, not technical. PCA is a data reduction method that seeks to find linear combinations of observed variables that capture the maximum possible variance. It does not assume an underlying model: it simply transforms your variables into a smaller set of components that retain most of the original information. CFA, on the other hand, starts from an explicit theoretical model that specifies which observed variables should cluster under which latent factors, and evaluates whether the data fit that model. While PCA tells you how your data are organized, CFA tells you whether your data are compatible with the theoretical structure you have proposed.
When to use each technique
PCA is appropriate in purely exploratory situations where you have no prior hypotheses about the structure of your data and your primary goal is to reduce a large number of variables to a manageable set of components. It is useful, for example, when you are developing a new questionnaire and want to see how the items cluster empirically, without imposing any predefined structure. However, even in this exploratory context, many methodologists prefer Exploratory Factor Analysis (EFA), which unlike PCA does assume a latent variable model and distinguishes between common variance and specific variance.
The difference between PCA and EFA may seem technical, but it has practical consequences. PCA includes all the variance of each variable (including error variance) in the solution, which tends to inflate factor loadings and produce solutions that appear cleaner than they actually are. EFA, by estimating communalities and separating common variance from specific and error variance, produces more honest estimates of the relationships between observed variables and latent factors. If your goal is to identify latent constructs (such as personality dimensions or questionnaire factors), EFA is conceptually more appropriate than PCA.
CFA, for its part, is used when you already have a theory about the factor structure and want to test it with new data. It is the natural step after an EFA: you explore the structure with a first sample and confirm it with a second. It is also the fundamental tool for evaluating the construct validity of a questionnaire adapted from another language or culture: if the original questionnaire has three factors, CFA allows you to verify whether that same structure holds in your sample.
Common errors and practical considerations
The most serious error committed with CFA is using it as an exploratory tool. Some researchers specify a CFA model, obtain poor fit, and start modifying the model based on modification indices until the fit is acceptable. This is problematic because modification indices are suggestions based on the data, not on theory, and following them systematically capitalizes on sampling error and produces models that will not replicate in new samples. If your CFA model does not fit the data, the answer is not to modify it until it fits, but to reconsider your theory or conduct an EFA to explore what structure the data actually suggest.
Regarding CFA fit, it is important to evaluate multiple indices and not rely on a single criterion. The chi-square is the formal test of fit, but it is very sensitive to sample size and rejects reasonable models with large samples. The most commonly used fit indices are the CFI (Comparative Fit Index, which should be above 0.95 for good fit), the RMSEA (Root Mean Square Error of Approximation, ideally below 0.06), and the SRMR (Standardized Root Mean Square Residual, ideally below 0.08). These cutoff points, proposed by Hu and Bentler (1999), are guidelines and should not be applied mechanically, but they are a useful reference.
Before you submit: running a PCA and calling it factor analysis is one of the things a methodological reviewer spots on the first read, along with fit indices reported halfway. If your manuscript is already written, run it through the AI Paper Reviewer: a free Reviewer 2 style pre-review that tells you what they will challenge about your factor structure while you can still change it.
Final recommendations
If you are developing a questionnaire or exploring the structure of a dataset, use EFA (preferably with principal axis or maximum likelihood extraction, not PCA) and apply an appropriate rotation (oblique if you expect the factors to be correlated, which is typical in psychology). If you are confirming a theoretical structure or validating an existing instrument, use CFA with lavaan in R, Mplus, or the AMOS module in SPSS. And in any case, report enough information for the reader to evaluate the quality of your solution: factor loadings, explained variance, communalities, fit indices (in CFA), and the theoretical justification for the proposed structure.