Alpha is not a great way of measuring reliability for a number of reasons. I would recommend using item response methods like Mokken scaling to make sure that the scale or its subscales meet minimum criteria for a unidimensional scale (Loevinger's H of at least 0.3, but better to have it higher).

That said, sample size for alpha depends on the number of items, but also on 1) the desired alpha and 2) the 'null hypothesis alpha'. If you expect an alpha of 0.8, and you want your confidence interval to exclude a value of 0.7, for example, you need a sample size of 31.

Here are the sample sizes for a 12-item test for alpha = 0.8,0.85 and 0.9

. ssalpha 12 .8

For a 12 items-scale, and an expected alpha of 0.8

sample sizes for desired 95% lower bound CI of alpha (LB-alpha):

----------------------------

LB-alpha Sample size

----------------------------

0000.790 2,423

0000.780 622

0000.770 284

0000.760 164

0000.750 108

0000.740 77

0000.730 58

0000.720 46

0000.710 37

0000.700 31

. ssalpha 12 .85

For a 12 items-scale, and an expected alpha of 0.85

sample sizes for desired 95% lower bound CI of alpha (LB-alpha):

----------------------------

LB-alpha Sample size

----------------------------

0000.840 1,375

0000.830 356

0000.820 164

0000.810 96

0000.800 63

0000.790 46

0000.780 35

0000.770 28

0000.760 23

0000.750 19

. ssalpha 12 .9

For a 12 items-scale, and an expected alpha of 0.9

sample sizes for desired 95% lower bound CI of alpha (LB-alpha):

----------------------------

LB-alpha Sample size

----------------------------

0000.890 622

0000.880 164

0000.870 77

0000.860 46

0000.850 31

0000.840 23

0000.830 18

0000.820 14

0000.810 12

0000.800 10

Sample sizes go down a bit with more items, but for 24 items the sample size is only about 5% to 10% smaller.

Calculations in Stata using the cialpha command written by Lin Naing and Than Winn of University Sains Malaysia

Alpha is not a great way of measuring reliability for a number of reasons. I would recommend using item response methods like Mokken scaling to make sure that the scale or its subscales meet minimum criteria for a unidimensional scale (Loevinger's H of at least 0.3, but better to have it higher).

That said, sample size for alpha depends on the number of items, but also on 1) the desired alpha and 2) the 'null hypothesis alpha'. If you expect an alpha of 0.8, and you want your confidence interval to exclude a value of 0.7, for example, you need a sample size of 31.

Here are the sample sizes for a 12-item test for alpha = 0.8,0.85 and 0.9

. ssalpha 12 .8

For a 12 items-scale, and an expected alpha of 0.8

sample sizes for desired 95% lower bound CI of alpha (LB-alpha):

----------------------------

LB-alpha Sample size

----------------------------

0000.790 2,423

0000.780 622

0000.770 284

0000.760 164

0000.750 108

0000.740 77

0000.730 58

0000.720 46

0000.710 37

0000.700 31

. ssalpha 12 .85

For a 12 items-scale, and an expected alpha of 0.85

sample sizes for desired 95% lower bound CI of alpha (LB-alpha):

----------------------------

LB-alpha Sample size

----------------------------

0000.840 1,375

0000.830 356

0000.820 164

0000.810 96

0000.800 63

0000.790 46

0000.780 35

0000.770 28

0000.760 23

0000.750 19

. ssalpha 12 .9

For a 12 items-scale, and an expected alpha of 0.9

sample sizes for desired 95% lower bound CI of alpha (LB-alpha):

----------------------------

LB-alpha Sample size

----------------------------

0000.890 622

0000.880 164

0000.870 77

0000.860 46

0000.850 31

0000.840 23

0000.830 18

0000.820 14

0000.810 12

0000.800 10

Sample sizes go down a bit with more items, but for 24 items the sample size is only about 5% to 10% smaller.

Calculations in Stata using the cialpha command written by Lin Naing and Than Winn of University Sains Malaysia

The old standard for alpha that I've used is at least 10 per item. So 12 items = 120 subjects. That's really a lower bound and the best scale creation calls for more in depth psychometrics. Examine item correlations, item-total correlations, squared multiple correlations (SMC), alpha, and PCA at the least. If you can get a larger sample - 500-1000, I'd suggest using IRT (item response theory) to get a really clear view of whether your items fit and define a consistent construct. The more items the better for IRT also.

For a smaller sample or less items, you won't be able to use IRT. See my 2012 paper for a psychometric analysis of 3-4 item scales.

Dear Ronán Michael Conroy und Karen H Schmeelk-Cone,

i have the same question. Your answers helped me alot. Would you please let me know your sources? It will be very helpful.

Thanks a lot.

Dear Karen and Ronan, you noted some useful answers.

I was wondering if you could mentioned minimum sample size for validity too.

I used some scales in my equasi-experimental research, by 60 subjects.

Some of my scales that I used to research, they have never uses in Persian and they used just in English.

I want to publish the results of my research in English. but: How I can report my research, without any similar study aligned with my research in Persian?

Any suggestion will be appreciate.

Dear Ronan - thank you very much. May I ask what is the syntax for the ssalpha command? And secondly, I have not come across sample size calculations comonly for questionnaire validation studies. Is it critical that it is done?

The article is about minimum sample size for Cronbach's coefficient...

https://www.researchgate.net/publication/268809872_Minimum_Sample_Size_for_Cronbach%27s_Coefficient_Alpha_A_Monte_Carlo_Stdy

Article Minimum Sample Size for Cronbach's Coefficient Alpha: A Mont...

Dear All

Can anybody answer my questions please?

1.How can I correlate the domains to each other using Cronbach's?

2. Can I use Cronbach's alpha for a domain of 3 items?

For correlating items, use correlation matrix among items.

Your second question is not clear, I think, you mean, three levels of rating. Whether three point or five point or more, no matter you can use cronbach's alpha for internal consistency test ...

Thanks a lot for your answer

I am developing a questionnair on eating related quality of life in orthodontic patients.

It has 6 domains : domains 1 and  2 consist of 4 questions, domains 3 and 5 consist of 3 questions, domain 4 has 7 questions and domain 6 with 5 questions.

Do I need to have cronbach's alpha for all of that domains? What about having it for all the questions in one goes?

Does the number of questions affect the outcome of cronbach's correlation?

Many thanks

Adham 

i used the objective meaurement tool for measuring joint range of motion in my study. would any one tell me how to calculate its sample size. what i got from your oveall discussion that this sample size calculation is for subjective tools