Hi dear experts and colleagues,
Actually, I have a questionnaire and for testing its reliability, it has been applied to 20% of my sample. This questionnaire embraces 8 domains, the scores of Cronbach's alpha according to SPSS analysis were as follow: 0.747, 0.706, 0.346, 0.7, 0.477, 0.754, 0.638, and 0.79.
Please, could I know if it is reliable or not? If NOT, what should I do?
Best regards
Iyad,
First, a scale isn't "reliable or not" -- it is a mater of degree. Next, although there are Rules Of Thumb (I typically label these "ROT" as they are always wrong in some parameter space), it is you who needs to decide whether a scale is reliable enough for your purposes. [Some may, at this point wish to apply such a Procrustean .70 ROT and say that # 1, 2, 4, 6, and 8 are OK, the rest are not. I refuse to do so as such decisions are inherently contextual. Although to be fair, the politics are such that some reviewers follow that ROT.]
Cronbach's alpha is an estimate of proportion of true score variance that your instrument can theoretically predict. It depends on only two data characteristics: the average inter-item covariance (or correlation if the standardized version is used) and the number of items. As either increases, alpha increases.
So apply that notion to the alphas that you specified. You are looking at several scales that account for 70%+ of the true variance, and some that are performing markedly lower that that. Outside of individual assessment, why do we care? Because you would like to relate these scales to something -- perhaps each other, perhaps other scales, perhaps demographics, I don't know, but I am sure you want to relate these scales to something. Therein, as such correlations require covariance, and assuming the true scores are actually correlated to those other variables, unreliability (read error) will attenuate covariance and resulting correlation estimates.
Do note that depending on the "overhead" you have afforded yourself in the number of items in each scale, there may be ways to improve alpha by deleting items. This may seem counter-intuitive as I have already said that alpha goes up as the number of items increases, however, that would only be if the average inter-item covariance remains the same. You can sometimes find a subset of items that manifest a larger alpha even though the number of items is less because the increased average inter-item correlation more than makes up for the decrease in number of items. Diagnostics that can help are item-total correlations, and alpha if the item is omitted.
My best on your scale development,
John
Thanks so much John Morris for this valuable answer you have delivered.
Kind regards,
Iyad
To determine the internal consistency of the eight variables the conbach alpha value should be 0.7 and above and individual variable dimensions should also be 0.5 and above using factory analysis. In using factory analysis any variable less than 0.5 is considered not good enough to be used for generalisation.
Thanks so much Mr. Timothy Adu Gyamfi for your valuable and brief answer.
Best wishes,
Iyad
Hi Iyad Shaqura
Alpha > 0.7 is the generally accepted cut-off for that test of internal consistency. Accordingly, dimensions 3, 5, and 7 are not reliable.
Please note that internal consistency implies that items within a dimension should be correlated (I use between 0.5 to 0.7), but not too high (e.g., 0.9) or they might indicate redundancy.
(1) Check you SPSS output for reliability test and look for Item-to-total correlations which are low.
(2) Make sure these items are not reverse worded. If they are, this might be the reason why the don't correlate with others.
(3) You can also check the "Alpha if item deleted" option.
(4) If you have many items (i.e., questions) within a dimension, a low value of Alpha might be an indication of the presence of sub-scales within that dimension. As such, if it happens to be a formative construct, these sub-scales (i.e., sub-dimensions) should not be expected to be correlated. This means that item-to-total correlation (at a dimension level or second-order if you like) will be lowered and thus underestimating the value of Alpha. Try to run exploratory factor analyses each dimension separately just to verify whether you have sub-scales.
(5) If all the above fail, you can try to increase the sample size as a last resort.
I hope this info can be useful to you and future readers.
Good luck,
Kimo
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Probability
- Probability Theory
- Reliability