Low cronbach's alpha - any remedy?

#1
Hi there,

I want to construct a scale of 'contacting government' based on the question 'Please indicate if in the past year you contacted govt about any of the following services.' --There is a list of 14 different topics. The factor analysis (result attached) shows what to me what are logical groupings of services (for instance, police and fire together, education and youth center, etc), but the Cronbach's alpha is only .592. This of course is lower than the conventionally accepted .7. I looked at the 'show alpha if item deleted' scale and nothing will increase the alpha even over .6.

Is it necessary to abandon this scale? Is there anything else I could try to make an acceptable measure? Other suggestions?

My apologies for the open ended questions. I appreciate any advice.
 

CowboyBear

Super Moderator
#2
Is it necessary to abandon this scale?
Flippant answer: Just abandon Cronbach's alpha. See On the use, the misuse, and the very limited usefulness of Cronbach’s alpha (Sijtsma, 2009).

Slightly less flippant answer: Alpha tends to be biased downward as a measure of reliability when responses to a scale are determined by more than one underlying latent variable. No doubt this is the case here: There are probably lots of causative factors driving how often someone contacts government, with different factors underlying contacts about different issues. (Rather than one underlying factor of "propensity to contact government" that explains most the variation and covariation in the frequency of different types of contact). You can of course investigate the factor structure of responses to the instrument, though PCA is not necessarily ideal for this purpose (it isn't a true factor analysis).

Maybe it would be worthwhile reading a little about cronbach's alpha and deciding whether it really measures something you care about and that is relevant for your scale, rather than sticking rigidly to the orthodoxy of always reporting alpha, and always sticking to the .7 rule of thumb?
 

Jake

Cookie Scientist
#3
You could try removing items iteratively until the scale reaches what you deem to be a satisfactory alpha.

Edit: plus what CB said about abandoning alpha :p
 

trinker

ggplot2orBust
#4
I think SPSS gives you the alpha if deleted item scores (referencing Jake's answer). Note that cronbach's uses N in the calculation, with such a low number of items alpha tends to be lower. You can add more items if this is a test run.
 
#5
Thank you very much for the answers.

@CowboyBear: Both the flippant and the much less flippant answers are helpful. There is a somewhat strong theoretical motivation for using this measure, so I will take your advice and look into how to pre-empt journal reviewer criticism.

@Jake and trinker: Also, thanks. Unfortunately, the removal of individual or groups of items does not produce an alpha even over .6.
 

Jake

Cookie Scientist
#6
You're sure? So you're saying you tried removing all possible subgroups of items and checking the alpha? Removing all possible subgroups of only sizes 1 through 5 would require 14 + 91 + 364 + 1001 + 2002 = 3472 analyses, and with 14 total items one could certainly keep going. This must have taken a while.
 

Dragan

Super Moderator
#7
You might want to use the L-comoment (robust) based estimator of coefficient alpha, which is based on the L-correlation.
 
#8
Thanks for the continued interest.

@Jake: Thank you for clarifying. No, I did not perform the tests as you suggest. Rather, I checked if removing any individual item or subset of related items would increase the alpha. It did not. I don't think I can do what you are suggesting manually. However, if there is a way to automate this process in the SPSS syntax, that might be an option.

@Dragon: I did a quick search of journal articles in my academic field for L-correlation-based coefficient alpha estimator. If I can find a way to produce this, I will try it.
 
#9
hey Jesse,

whats your conclusion now about the low cronbach alph value. I am in the same boat. I have a pretest and post test model. pretest show .1 and post test shows .4. after removing one particular item both score increases but are at .5. I am totally confused at what to do next/

twlight