# help on cronbach's alpha statistics

#### chaimae

##### New Member
Hi everyone,
I want to test the reliability of my questionnaire and I have calculated cronbach's alpha, the maximum value I found is 0.46. then What should I do?
thank you

#### katxt

##### Well-Known Member
Are all the questions asking about aspects of the same sort of thing in different ways?

#### chaimae

##### New Member
yes all the questions asking about the same topic in different ways
thank you

#### katxt

##### Well-Known Member
Are all the questions pointing in the same direction?

#### chaimae

##### New Member
they are around the same topic, questions are differents but they are around same subject they concerns the behaviours towards a concept.
if the maximum value of cronbach's alpha is 0.46, is there any Is there any other way to test reliability or should I modify my questions?

#### Karabiner

##### TS Contributor
Could you tell us something more about your study? How large was your sample size, how many items
are there and how does the answering format look like, and what do you want the scale use for?

With kind regards

Karabiner

#### chaimae

##### New Member
the sample size is 100
there are 20 items, my first cronbach value is 0,21,
I deleted some items and the max value I obtained is 0,46
I checked the coding of my data, there is no problem.
thank you

#### katxt

##### Well-Known Member
You clearly recognize that 0.46 is too low to have much confidence in your questionnaire as an instrument to measure a particular single concept/construct. Your sample size is good and it probably won't help much to have more questions or subjects. You have done the obvious things - checked the direction of the responses and removed some questions that are clearly not pulling their weight. Now you may have to face the fact that this set of questions just isn't up to the task. You may be actually trying to work with two (or more) different constructs that are not as similar as they may first appear.
You could try finding the correlation matrix between the questions to see if they fall naturally into two groups, or do some other sort of multivariate discriminant analysis.

#### chaimae

##### New Member
You clearly recognize that 0.46 is too low to have much confidence in your questionnaire as an instrument to measure a particular single concept/construct. Your sample size is good and it probably won't help much to have more questions or subjects. You have done the obvious things - checked the direction of the responses and removed some questions that are clearly not pulling their weight. Now you may have to face the fact that this set of questions just isn't up to the task. You may be actually trying to work with two (or more) different constructs that are not as similar as they may first appear.
You could try finding the correlation matrix between the questions to see if they fall naturally into two groups, or do some other sort of multivariate discriminant analysis.
thank you for these guidelines
I'm in the learning phase and I'm applying at the same time, if in my case I find questions that fall naturally into two groups, do I apply Cronbach's alpha for each of these two groups; is that the idea?

and can you clarify more about doing other sort of multivariate discriminant analysis? is that in this case I won't need to specify cronbach's alpha?
thank you

#### katxt

##### Well-Known Member
First a caveat: I have to say that this is not my field, and I only know of these ideas because I have helped social scientists with their research statistics and learned a little along the way. As I understand things ...
Chrobach's alpha is usually used to show that the total score on a particular set of related questions is a good measure of a particular construct (eg self confidence) and so that the total score on that set can be used as a reliable instrument to measure that construct.
If your questions fall naturally into two groups measuring different things you are likely to get a low alpha overall. So a low alpha suggests to me that says that you are perhaps measuring two different concepts. If you can separate these groups and they both have high alphas then you have two reliable instruments that measure different things. What those different things are may not be obvious. You will still need alpha to justify the questions in each group if you want other people to trust or use your newly invented scale.
Determining the groups of related questions, if they exist, can be done in a variety of ways. Look up cluster analysis and discrimination analysis if you think there may be two different scales. If this is an important project, like a masters or PhD, or a journal paper, I suggest you talk to a social science statistician for an hour. It would be money well spent. kat