I am currently doing my thesis in college. It is about the customer satisfaction on the 7ps of marketing mix on a restaurant and I have a 5-point Likert's Scale type of Questionnaire (Very Satisfied, Satisfied, Neutral, Dissatisfied, Very Dissatisfied). The questionnaire consists of 35 questions. They are categorized by every P's (Product, Price, Place, Promotion, People, Process and Physical Environment) 5 questions per category.. I did a cronbach-alpha test on this survey and came up with a result of .95 . My panels said that there might be redundant questions in the survey and I need to find them to have an efficient survey. What should I do?
Hi, With alpha value of .95, I wonder how can there be redundant questions. Getting 0.95 cronbach alpha with seven dimensions, you may have issues with discriminant validity. Anyway, analysis of Item wise cronbach alpha values may help you examine which questions are problematic but alpha value of .95 exhibit very good internal consistency so I would recommend you try factor analysis/structural equation modeling to examine your measurement model.
Regards
Such a large value for Cronbach's alpha means that the items are highly intercorrelated. If that value was for all 35 items, it would indicate that you might have trouble discriminating between the 7ps as separate sub-scales.
I suggest that you use alpha to evaluate each of the seven sub-scales. If they are reliable, then add the items in each sib-scale together to form 7 scales, and then inter-correlate that set of scales.
Look into the inter items correlation if they are very high ( Greater then .8) consider eliminating one if them as they probably are measuring the same thing.
Usually what I do, because I tend use large samples, is to run a CFA.
Thank you everyone for your comments, these answer really helps a lot.
I'm not sure why your committee is concerned, because having "redundant" items does not harm the creation of a scale. (Highly correlated items would pose a problem of multicoliearity if they were both entered into the same regression, but that is another matter.)
But, redundant items may not help your scale very much, so if you want to avoid having a fight with your committee, you might be able to remove everything with a correlation above .9 and still get an adequate alpha.
I support the suggestions given by José and Smitha - looking at intercorelations and running a factor analysis.
I will suggest that you determine the Cronbach alpha for each scale rather than for the entire scale. This way you may be able to detect the subscale(s) with poor internal consistency and then deal with that. However, if all the scale has Cronbach alpha >0.7, then you will need to look at the inter-items correlation if high. removing an item and re-running the analysis may help remove the redundant item. lastly, factor analysis might just be your best way out for each of the subscales. it may be cumbersome, but ist sure will sort the items out nicely.
Start off by having a good pilot study where you would ask the questions. Analyse them and you would be able to weed out the questions that appear to be redundant. On the other hand it is better to have some extra questions rather then be short of questions. The extra questions can be weeded out ( even in the actual study) but if you are short of the questions it would be difficult for your work.
I agree with the Segun, calculate Cronbach alpha for each scale (or "P"). Factor analysis is also useful.
obtain (cronbach alpha) reliability index or coefficient of the questionnaire you developed for the study and use the score obtained to judge whether a question is redundant or not. for you to do this any question that scores less than five reliability is actually redundant
I am not clear about something that seems important to me: Do you are testing your scale in a pilot or you actually applied it to your survey sample?
Passing thru this concern I agree with David about why your alpha level is stepping concerns. Regarding statistical analysis, if you actually collected your data I agree with Principal Component Analysis or Exploratory Factor Analysis specially because I strongly believe that is something very easy to do in R or in any commercial software. In R there are a very good package 'psych' of William Revelle that helps a lot analyzing survey data.
If the questionnaire are not yet applied to the sample I will suggest to read Rossiter C-OAR-SE that is a very sound approach besides the author's reluctant claims to classical psycometrics, we actually get results that shows both positions are congruent besides their assumptions are not. I like a lot Rossiter's assumptions and I strongly believe that many people agree but clasical psychometrics is, as Kuhn teach us, the mainstream paradigmatic view and figthing with advantage in order to repeal alternative approaches. The main question in any approach when dealing with your question could be the same: Why so many items if less are better?
Best of luck in your quest!
One thing that others have suggested where I disagree is using Principal Components Analysis. I do think that Exploratory Factor Analysis (EFA) would be useful, but I would use a method that is better suited to correlated factors (i.e., oblique rotation). In particular, I recommend Maximum Likelihood Factor Analysis, because that would be the closest to Confirmatory Factor Analysis (CFA), in case you go that way.
David is right regarding use of EFA and not PCA. Most of time both these techniques are considered same, which is not the case. In EFA, number of factors/components to be retained is a crucial decision. I would suggest to apply parallel analysis to identify number of factors to be retained rather than using eygen value rule or looking at scree plot only.
- Badges
- Science topic
- Similar topics
- Linguistics
- Composition Studies
- Publication