Cronbach's Alpha

Alpha asks a simple question: do the items of a scale move together enough to be treated as one measure? It rises with inter-item correlation and with the number of items — which also means a long scale can score well while containing weak items.

Convention treats 0.70 as acceptable and 0.80+ as good, but alpha above ~0.95 can signal redundant items. Report alpha per construct in your instrument section, and check 'alpha if item deleted' before dropping anything.

Alpha rests on an assumption that is rarely met exactly: that every item contributes equally to the construct (tau-equivalence). When loadings differ substantially, alpha tends to underestimate reliability, which is why SEM-based studies also report composite reliability and, increasingly, McDonald's omega. Alpha is also only meaningful for reflective scales, where items are interchangeable indicators of one underlying construct — it should not be computed for formative indices or for a mixed set of unrelated questions.

A low alpha usually has a diagnosable cause. Negatively worded items that were not reverse-coded before analysis are the most common culprit and can pull alpha down sharply. Very short scales of two or three items also tend to produce modest values, and a sample that is homogeneous on the construct restricts variance and depresses the coefficient. Fix coding errors first, then consider item deletion only where it is justified both statistically and conceptually — dropping an item changes what the scale measures.