Interpreting factor analysis means reading a series of statistical tables in a specific order: first confirming your data is suitable, then determining how many factors to keep, then examining which variables cluster together and why. Each table in your output tells you something different, and skipping ahead to factor loadings without checking the earlier steps can lead to meaningless results. Here’s how to work through the output from start to finish.
Check Whether Your Data Is Suitable
Before interpreting any factors, you need two pieces of evidence that your variables are correlated enough to produce meaningful groupings. Your software will generate both automatically if you request them.
The Kaiser-Meyer-Olkin (KMO) measure tells you whether the patterns of correlation among your variables are compact enough for factor analysis to produce distinct, reliable factors. KMO ranges from 0 to 1, and values above 0.6 are considered acceptable. Values above 0.8 are good, and above 0.9 are excellent. If your KMO falls below 0.6, your variables don’t share enough common variance to justify factor analysis.
Bartlett’s test of sphericity checks whether your correlation matrix is significantly different from an identity matrix, which would mean no variables are correlated at all. You want this test to be significant (p < 0.001). If it’s not significant, your variables are essentially unrelated and factoring them together won’t produce anything meaningful.
Decide How Many Factors to Keep
This is the most consequential judgment call in the entire analysis. Keeping too many factors splits real patterns into fragments; keeping too few forces unrelated variables together. Two tools help you decide, and they sometimes disagree.
The Kaiser criterion is the simpler rule: keep every factor with an eigenvalue greater than 1. An eigenvalue represents how much total variance a factor accounts for, expressed in units equivalent to one variable. A factor with an eigenvalue below 1 explains less variance than a single original variable would, so it’s not pulling its weight. You’ll find eigenvalues in the “Total Variance Explained” table, usually in the first numeric column. This rule tends to overestimate the number of factors, especially with large numbers of variables.
The scree plot is a line graph of those same eigenvalues plotted in descending order. You’re looking for the “elbow,” the point where the line drops steeply and then flattens out. The number of factors to retain is the number of points before the last large drop. Everything on the flat portion of the curve represents statistical noise. The tricky part is that some scree plots don’t have a clean elbow, which is where judgment and theory come in. If the Kaiser criterion says five factors and the scree plot suggests three, look at both solutions and see which one produces more interpretable groupings.
How Much Variance Should Factors Explain
Your “Total Variance Explained” table also shows the cumulative percentage of variance accounted for by your retained factors. In the natural sciences, a common benchmark is 70% to 80% of total variance. In social science research, that standard is usually unrealistic because human behavior is messy and measurement tools are imperfect. Extracted factors in social science typically explain 50% to 60% of total variance, and that’s considered adequate.
Don’t chase a high variance number by adding factors that don’t make conceptual sense. A three-factor solution explaining 52% of variance is more useful than a six-factor solution explaining 71% if those extra factors are uninterpretable.
Reading the Factor Loading Table
Factor loadings are the core of your interpretation. Each loading is a correlation coefficient between a variable and a factor, ranging from -1 to +1. A loading of 0.70 means that variable shares about half its variance with that factor (0.70 squared = 0.49). The higher the absolute value, the more strongly that variable belongs to that factor.
Threshold guidelines vary by source, but here’s a practical framework. Loadings below 0.32 are generally too weak to interpret. Loadings of 0.45 are fair, 0.55 are good, 0.63 are very good, and 0.71 or above are excellent. Sample size also matters: a loading of 0.30 is only statistically meaningful with a sample of roughly 350, while a loading of 0.50 becomes significant around 120 participants. With smaller samples, you need stronger loadings to trust the pattern.
A stricter rule of thumb suggests treating a factor as reliable only if it has four or more loadings of at least 0.6, regardless of sample size. For most applied work, using 0.40 as your interpretive cutoff and suppressing everything below it in your output makes the table much easier to read.
What Clean Structure Looks Like
The ideal result, sometimes called “simple structure,” is one where each variable loads strongly on exactly one factor and weakly on all others. In practice, you’ll see some variables that load moderately on two factors (called cross-loadings). If the difference between a variable’s highest and second-highest loading is less than about 0.15 to 0.20, that variable doesn’t clearly belong to either factor and may need to be removed.
Variables that don’t load above your threshold on any factor are also candidates for removal. After dropping problematic items, rerun the analysis. Factor analysis is iterative: it’s common to run it three or four times before arriving at a clean solution.
Choosing a Rotation Method
Unrotated factor solutions are mathematically correct but often hard to interpret because variance gets concentrated in the first factor. Rotation redistributes the variance to produce a cleaner pattern of loadings without changing how much total variance is explained.
Orthogonal rotation (varimax is the most common) forces factors to remain uncorrelated with each other. This produces a simpler, easier-to-interpret solution, but it assumes your underlying constructs are truly independent. Oblique rotation (promax or direct oblimin) allows factors to correlate, which is more realistic in most social and behavioral research where psychological constructs overlap. If your factors correlate at 0.32 or above after oblique rotation, that correlation is meaningful and oblique rotation was the right choice. If correlations between factors are all below 0.32, the orthogonal and oblique solutions will look nearly identical, and you can use either.
One practical difference: with oblique rotation, your output includes both a pattern matrix and a structure matrix. The pattern matrix shows the unique contribution of each factor to each variable (controlling for other factors), while the structure matrix shows the simple correlation between each variable and factor. Interpret the pattern matrix. That’s where you’ll identify which variables belong to which factors.
Checking Communalities
The communalities table shows how much of each variable’s variance is explained by the retained factors combined. Each communality value is the sum of that variable’s squared loadings across all factors. Think of it as an R-squared value for that variable: a communality of 0.65 means the factors collectively explain 65% of that variable’s variance.
Variables with low communalities (below 0.30 or 0.40, depending on your field) aren’t well-represented by the factor solution. They’re essentially sitting outside the groupings you’ve identified. Low-communality items may be poorly worded survey questions, conceptually distinct from the other items, or simply unreliable. Consider removing them and rerunning the analysis.
Naming Your Factors
Once you have a clean loading pattern, naming each factor is a qualitative step. Look at the variables that load most strongly on a given factor and ask what they have in common. If a factor’s highest-loading items are “I feel nervous in social situations,” “I avoid speaking in groups,” and “Meeting new people makes me uncomfortable,” a label like “Social Anxiety” captures the shared theme.
Good factor names are specific enough to distinguish one factor from another but broad enough to encompass all high-loading items. Avoid naming a factor after a single variable. If you can’t find a coherent theme, that’s a sign the factor may not be meaningful, or that the number of factors you retained isn’t quite right.
Validating Your Factors With Reliability
After identifying your factors, calculate Cronbach’s alpha for each group of items that loaded onto the same factor. Alpha measures internal consistency: whether the items in a group tend to move together. Values above 0.70 are generally considered acceptable for research purposes.
A low alpha for a factor could mean the items aren’t measuring the same underlying construct, which contradicts what the factor analysis suggested. This sometimes happens when a factor has only two or three items, since alpha is sensitive to the number of items in the scale. It can also happen when items within a factor have very different response distributions. If alpha is low, look at the “alpha if item deleted” column in your reliability output to identify which specific item is dragging consistency down.
One important nuance: alpha assumes all items in a scale measure a single dimension. If your factor analysis revealed that a set of items is multidimensional, calculating alpha across all of them will underestimate the true reliability. Run alpha separately for each factor’s items, not for the whole instrument at once.