just an addon regarding alpha in general. To interprete alpha as a signal of reliability, the set of items has to fulfill two prerequisites:
a) they have to be measures of the same latent variable and this is only one variable (homogeneity)
b) the factor loadings have to be the same
Together, these assumptions are called "essential tau-equivalence".
Very often (if it is not the major reason at all), low alphas are a signal of the items tapping different things, hence, are measures of different latent variables.
Therefore, instead of killing items based on the information on "alpha if item deleted", the better option is to really test the factor structure and to react to misfit with meaningfull respecifications of the factor structure. Sometimes, single items are valid indicators of important latent variable. If you eliminate the item, you possibly eliminate a whole latent variable.
Dear Mohamoud, you can use alpha to calculate reliability for categorical data. If the alpha is to low (that will depend on the context and the criteria that you are using this scale) you can delete items that have bad loading on the factor or if the item loads in two different factor as well (you have quite few criteria to determine if you should delete an item). If you are using SPSS you can ask for improving alpha when you delete item and then, you can observe what happen if you delete certain item.
just an addon regarding alpha in general. To interprete alpha as a signal of reliability, the set of items has to fulfill two prerequisites:
a) they have to be measures of the same latent variable and this is only one variable (homogeneity)
b) the factor loadings have to be the same
Together, these assumptions are called "essential tau-equivalence".
Very often (if it is not the major reason at all), low alphas are a signal of the items tapping different things, hence, are measures of different latent variables.
Therefore, instead of killing items based on the information on "alpha if item deleted", the better option is to really test the factor structure and to react to misfit with meaningfull respecifications of the factor structure. Sometimes, single items are valid indicators of important latent variable. If you eliminate the item, you possibly eliminate a whole latent variable.
the best approach is to test the supposed factor structure within a structural equation modeling framework. If the model fits via the chi-square test, you can apply the composity reliability measure which can be simply calculated with e.g., excel.
The formula is contained here:
Fornell, Claes, & Larcker, David F. (1981). Evaluating structural equation models with unobservable variables and measurement error. Journal of Marketing Research, 18, 39-50.
I am not a big fan of this article as it weakens the goal of statistical testing in SEM.
The advantage of Composite reliability is that it does not require equal loadings (as alpha does).
As we know, Cronbach's alpha determines the internal consistency. It is usually used for categorical data, either binary or Likert Scale. For reliability, you may use Composite Reliability.