Factor analysis uncovers the smaller set of hidden latent factors that explain why your questionnaire items move together, while Cronbach's alpha measures the internal consistency of the items that load on a factor. Factor analysis tells you how many underlying dimensions your scale has and which items belong to each; alpha then tells you whether the items within a dimension are reliable enough to combine into one score. In a dissertation that uses a survey, you almost always run them as a pair.
Why scale validation comes before hypothesis testing
Before you can test relationships between constructs, you have to show that your measurement instrument actually captures those constructs. Running factor analysis first establishes the construct validity and dimensionality of the scale, so that the variables feeding your later models are meaningful rather than arbitrary sums of items. This is the same discipline that lets you treat a tidy total score as a clean predictor or outcome variable in the next stage. Skipping validation and jumping straight to correlations is a frequent cause of viva questions, because nobody can be sure the numbers mean what you claim. To check where a scale stands, the Cronbach's alpha calculator returns alpha plus the per-item statistics that show which items hold it together.
Running factor analysis in SPSS
In SPSS, factor analysis lives under Dimension Reduction. You begin by checking that the data are suitable: the Kaiser-Meyer-Olkin measure should sit above roughly 0.6, and Bartlett's test of sphericity should be significant. You then choose an extraction method, most often principal axis factoring for exploratory factor analysis, and decide how many factors to keep using eigenvalues over 1 and a scree plot. A rotation such as varimax or oblimin makes the factor loadings easier to read, and you interpret the rotated pattern matrix to see which items load cleanly on which factor. Reading these tables is the same skill set covered in how to interpret SPSS output.
Exploratory vs confirmatory factor analysis
The two approaches answer opposite questions. Exploratory factor analysis is what you run when you do not yet know how many dimensions a scale has or which items load where; it lets the data suggest the structure, which suits a new instrument, a translated questionnaire, or a scale you have adapted for a fresh population. Confirmatory factor analysis is what you run when theory or prior research already proposes a specific structure and you want to test whether your data fit it. Confirmatory work belongs to the structural equation modelling family and reports fit indices such as the comparative fit index, the Tucker-Lewis index, and the root mean square error of approximation rather than eigenvalues. A common dissertation pattern is to explore on one half of the sample and confirm on the other, which is why a grasp of structural equation modelling pays off once your measurement model is settled.
Factor loadings and how to read them
A factor loading is the correlation between an item and the underlying factor, and it is the number you scan to decide which items define each dimension. Loadings run from negative one to positive one, and a common working threshold is that an item should load at about 0.40 or higher on its intended factor to be retained, with some authors preferring 0.50 for cleaner structure. Two patterns spell trouble. A cross-loading item loads moderately on two or more factors at once, which muddies what the item measures and usually justifies dropping it. A low-loading item, sitting below the threshold everywhere, contributes little and weakens the scale. After a rotation, you read the pattern matrix row by row, keeping items that load strongly on one factor and weakly on the rest. The cleaner this simple structure, the more defensible the scores you carry into your later models become.
Reading Cronbach's alpha
Once factor analysis has grouped the items, you compute Cronbach's alpha for each resulting subscale. Alpha ranges up to 1, and a value of about 0.7 or higher is the usual threshold for acceptable reliability, with 0.8 and above considered good. Very low alpha suggests the items are not measuring a single underlying thing, while an alpha above roughly 0.95 can signal redundant items that say the same thing twice. Always inspect the item-total correlations and the alpha-if-item-deleted column, because a single poorly worded item can drag a whole subscale down. Alpha is also sensitive to the number of items, so a short scale can post a modest alpha and still be sound.
Cronbach's alpha and scale reliability beyond one number
Reliability is the degree to which a scale measures its construct consistently, and Cronbach's alpha is the most reported index of one kind of reliability, internal consistency. It is worth knowing that alpha rests on a strong assumption called tau-equivalence, meaning every item contributes equally to the true score, and when that assumption is violated alpha can understate true reliability. For that reason many methodologists now prefer McDonald's omega, which relaxes the assumption and draws directly on your factor loadings. Internal consistency is also not the whole of reliability: test-retest reliability captures stability over time, and composite reliability is the version typically quoted alongside a confirmatory model. A thorough measurement chapter reports more than a lone alpha, and pairs it with the validity evidence from your factor structure. The Cronbach's alpha calculator gives the per-item diagnostics that show whether a single weak item is masking an otherwise reliable scale.
Common pitfalls to avoid
A few mistakes recur in dissertation survey work. Treating principal component analysis as identical to factor analysis blurs an important distinction: components summarise total variance, whereas factors model the shared variance behind observed items. Computing one global alpha across a multidimensional scale is meaningless; you report alpha per factor. Reverse-coded items must be recoded before alpha is calculated, or reliability will look artificially poor. Finally, ordinal Likert items should be screened for distribution before extraction, which connects to the wider habit of testing your data for normality, and to the analysis choices set out in analysing Likert scale data. Keeping the measurement layer clean is part of the same descriptive groundwork that precedes inferential testing.