Cronbach's Alpha

[WeeG]

Hello,

If I have to build some variable from a poll, and my Cronbach's Alpha is 0.66, and no removal of a variable increase that, what do I do ?
use this variable anyway and just make a remark about it, or give this var up ?

[lumhearts]

Is your sample size large? If so, a .66 with a large sample may get criticized. If your sample is small you can make the comment that alpha is sample size dependent.

[WeeG]

my sample size is N=59 cases.

if it's "criticized", what does it mean, I can't use it at all ?

[lumhearts]

I've seen alphas in the .6's before. I think they are workable. n = 59 is good news I think. Just another thought, you could always run a confirmatory factor anlaysis and see if the model fits well. Maybe include a model fit statistic instead of or along side of alpha as evidence of consistency.

[Lazar]

How many item do you have. Alpha is biased against scales with a few items. There is a paper by David Greyson in Understanding statistics Journal called something like "Some Myths and Legends in Quantitative Psychology". Read that as a basis.

[CowboyBear]

Hi WeeG, I think you might find the below article interesting:

http://mres.gmu.edu/pmwiki/uploads/M...rtina_1993.pdf

I'm pretty skeptical of simple cutoff rules for alpha - I'd tend to prefer a scale with good factorial fit but a less impressive alpha rather than the opposite. Have you used any exploratory or confirmatory factor analysis as yet?

[WeeG]

no, I didn't....the measures I was trying to build where based on questions from a poll, according to the demand I got. If I'll run Factor Analysis I will get different measures, won't I ? I also understand (not too well I am afraid), that Factor Analysis is problematic because of some severe assumptions.

[Lazar]

no, I didn't....the measures I was trying to build where based on questions from a poll, according to the demand I got. If I'll run Factor Analysis I will get different measures, won't I ? I also understand (not too well I am afraid), that Factor Analysis is problematic because of some severe assumptions.

I agree with CowboyBear comments.


A confirmatory factor analysis (or given that you have only one factor a one factor congeneric model) should be fine in your case.

EDIT: the article I was discussing http://www.informaworld.com/smpp/con...ent=a787471370