I am sometimes confused about different interpretations of internal consistency indices, specificly (Cronbach's) Alpha. Some find .30 for each subscale adequate, yet some find .50 inadequate. I would be glad to encounter ideas and references on this issue.
Hi,
I am not aware of anyone who would consider .3 as adequate. Most people consider .7 as the minimum adequate level and simply reject anything below that.
See the link below for an interesting discussion on the topic and references
https://www.researchgate.net/post/Cronbachs_alpha_below_07?_sg=Y7PC%2B0zOPYYUeZhkO7m2fXKSojbe2Enyz%2F9SLXeehN6KzZzKznru9jYVzmgneidz_UUdn0i7RJfA%2BH%2BCsYCK7Y7a48HY96H7YT%2F77JzxVASMtXesAI1o7gNe5QhS1dSq9
One of the examples that baffled me was that passage:
"....Estimates of internal consistency (alpha) coefficients were low to modest, ranging from .33 (Purpose in Life) to .56 (Positive Relations With Others). The alpha coefficient is a conservative estimate of internal reliability for most (congeneric indicators) scales (Bollen, 1989). The modest alpha coefficients likely reflect the small number of indicators per scale and the fact that items were chosen to represent the conceptual breadth within each construct (see rationale for item selection in Measures and Procedure section) rather than to maximize internal consistency."
Ryff, C. D., Keyes, C. L. M. (1995). The structure of psychological well-being revisited. Journal of Personality and Social Psychology, 69, 4, 719-727.
The paragraph essentially justifies the low alphas using two arguments:
(a) to capture the breadth of the construct they selected diverse items which resulted in the lower alpha. Had they used more similar items they could artificially increase the value of alpha but they would not have captured the construct itself.
(b) they used a small number of items for each scale. A large number of items can inflate the value of alpha and therefore a .5 value for a 3 item scale suggests higher reliability than a .5 value for a 10 item scale.
But what I see in many occasions is that there seems to be a huge number of studies strictly use the .70 cut-off criteria. I wonder whether I can use the same argumentation and cite the related source as the given passage? This would be a good point to refer when the brilliant scales are under risk of becoming dumped as negative data.
Internal consistency reliability is the degree of correlation between the items of a scale. The internal consistency is calculated by Kuder-Richardson’s formula 20 for dichotomous; and Cronbach alpha, polytomous answers. The internal consistency is acceptable between 0.70 and 0.95. To compute the internal consistency a twenty item scale are need between five to twenty participants for each item.
@Cinar: Depends on what are you using your measure for. If your goal is only to compute a correlation or a t-test, 0.70 or 0.80 might be fine. For a high-stakes individual decisions, .90 may be too low. See for instance Nunnally & Bernstein, 1994, p. 264.
@George: "A large number of items can inflate the value of alpha and therefore a .5 value for a 3 item scale suggests higher reliability than a .5 value for a 10 item scale." This is not true. A .5 value suggests the same reliability of the sum score regardless of the test length. It is true, however, that in you second case the inter-item correlations are lower than in your first case. But as long as you are interested in the test score reliability, the interpretation of the value of a reliability coefficient does not depend on the number of items.
Internal consistency depends on the contruct been measured. Usually in medical, sport, physiotherapy scales an α considered as appropriate with values >.70. In other cases you have to look the results of other studies if you want to learn what is acceptable in the scale you intend to use or cross-culturally adapt.
- Badges
- Science topic
- Similar topics
- Psychology
- Psychometrics