Be careful about getting stuck viewing the veins on the leaves of the trees and not the whole forest here.  Certainly, that may improve Cronbach's Alpha. But, does it improve the theoretical structure of the scale?  Second, be cautious about a statistical fishing expedition in search of better statistics.  Do you have a theoretical rationale or other psychometric rationale for omitting these items (s)?  Do they have low loadings, cross loadings, or other problematic indicators? These would be more potent reasons for omitting an item.

Hello Sam,

I think you are probably in the happy situation of worrying about small things when all is likely to be OK.  From a "naive" perspective, your Cronbach's alpha values are good, and I would be inclined to leave all of your items in because they might provide you with valid variability within your final data.

From a more sophisticated perspective, you might need to consider the number of items in your scale as a basis for removing some items.  Do you have a large number of items, so removing some of them might be good for subsequent rapport with respondents, for example?

There is some useful information in an article by Cortina titled What Is Coefficient Alpha? An Examination of Theory and Applications, published in 1993 in the Journal of Applied Psychology.  I have also published an article that deals with some ins and outs of Cronbach's alpha - at a less dauntingly sophisticated level.  That article is in the Journal of the American Podiatric Medical Association: https://doi.org/10.7547/0990460.

All the best with your work.  Feel free to get back if you have any queries.

Robt.

The values are both fine and above the given treshhold. I see no need to rework the scale and throw out items, especially not, in case you apply a well-established scale for measurement. 

Hi, Sam,

Cronbach's alpha is merely measuring the consistency of responses between items on your scale. Your question, to me, suggests that you are wondering if the factor structure of your instrument is strong. This can be assessed rather quickly and simply with exploratory factor analysis in SPSS. Items with low factor loadings (> .4) may improve the instrument. At the very least, EFA will tell you the proportion of variance in your construct that is explained by the items you keep.

The SPSS option for calculating alpha if an item is deleted is basically telling you that an item has a lower average correlation with the rest of the items. That is not at all the same as saying that that an item is "unreliable," especially when you are dealing with scores at the level you have.

So yes, retain such items.

Hi,

If your aim is to "purify" the scale, go with what Peter A. Kindle has suggested, otherwise, follow David's comment that tells what exactly alpha depicts.

The elephant in the room is the need to establish the validity of your scales, towards which establishing reliability is only the first (but essential) step. Now you need to either (a) run the scales together with other established scales that measure the same constructs, or (b) run them with a wider set of scales that, taken together, suggest that you are really measuring the constructs you are pursuing, or (c) run the scales with groups that are known to have high and low levels of the constructs. You can then check both the scale scores and the individual item scores against these criteria. You may find that the whole scales are measuring reasonably satisfactorily, but that some items are not contributing much, or even correlating negatively.

Another issue is whether the constructs are narrowly definable or 'pure,' or more general constructs encompassing a number of different ideas. In the latter case, you do not necessarily want a high alpha, which points to a narrow construct.

Both aspects need to be taken into account, and if you are trying to perfect scales of newish constructs, you will need to go round the cycle more than once; but if you just want rough-and-ready practical measures, less finesse is called for.

Hope that helps! [I am a Manchester Psychology graduate,  BA 1965 and MA 1967]. Most of this is 'old hat' and discussed in depth in books such as Nunnally & Berstein's 'Psychometric Theory' (1967!) or Gatewood & Feild's 'Human Resource Selection' (1987).

 Be careful about getting stuck viewing the veins on the leaves of the trees and not the whole forest here.  Certainly, that may improve Cronbach's Alpha. But, does it improve the theoretical structure of the scale?  Second, be cautious about a statistical fishing expedition in search of better statistics.  Do you have a theoretical rationale or other psychometric rationale for omitting these items (s)?  Do they have low loadings, cross loadings, or other problematic indicators? These would be more potent reasons for omitting an item.

Hi everybody! 

Thanks for your responses. Since I'm using PLS-SEM I have decided to keep my scale items, despite their potential to increase my CA value. 

The composite reliability check indicated that all questions are in a reasonable range (I had 2 in 0.6, but Hair [2016] indicates that you should really only delete scale items if they're below 0.4) 

Thank you for your help!

The .40 threshold is somewhat arbitrary.  Some even suggested retaining items as low as .30 (Child, 2006; Schmitt, 2011) but it depends upon what you are attempting to accomplish.  Keep in mind that the variance associated with an item with a loading of .40 is 16% versus 9% for a .30 loading.  

Hi all,

I'm a little late to this discussion but am writing up results on a study I did using the Witmer-Singer immersive tendencies Questionnaire. I did my own Cronbach's alpha and got an alpha of .678 (not bad considering I had a small sample). I also did a Cronbach's alpha for the subscales in the questionnaire and a few of the questions identified in the overall reliability are definitely messing up the subscale reliability.

My question is whether I remove the questions identified from the subscale analysis to better reflect the constructs or is this 'cherry picking' the data considering this is an established questionnaire?

Thanks,

Zara

I have a similar situation my alpha is equal to 0.8 for a scale with 29 items but I have items with negative Corrected Item-Total Correlation scores and most of those items are reversed items ... so what shall I do ?

Did you re-score the reversed items so that all 29 are in the same direction? if not you should, because that is an assumption for alpha.

Dear Sam Penny

Check the indicator weights (are they significant?) and Item-total correlation (every item above the recommended threshold 0,3?). Are they ok? Then you have empirical support to keep them. And also, even If the Cronbach alpha will have a little improvement you should be aware that may be losing a part of your measurement tool.

Good luck in your research.

Best

Yves