I need help with statistical power.
I have 40 participants, 2 groups of 20. I want to compare them, so would want to do either a T-test or ANOVA.
I've used G*power, but can't work out what the hell its going on about. I just want to know whether I have enough people to do a t-test or whether I should do a mann-whitney instead???
ok I've worked out that if I use a t-test the power will be about .45 which is RUBBISH.
So what do I do now???
Do I just say in my article that "the power was rubbish" or do I do a different test?
From what I've read non-parametric tests have less power. So confused.
This is what I'd do:
Do the t-test. If the answer is significant you had sufficient participants.
If the answer in non significant look at the effect size, if it is a reasonable effect size then maybe it is a genuine effect but you had insufficient power to detect it (ie not enough participants). If it is very small then you can argue that it is a genuine finding that there is no difference. t-tests are often done on an n or 20 per group in psychology.
Check in Andy Field's book for how to calcuate effect size and what is a small, medium or large effect.
Excellent! thank you
I have some non-significant results and some significant. I have run 5 independent t tests, and then another 5 (1 set for positive and 1 set for negative) So - here's my next question.
Do I have to do boneferroni correction .05/5 OR .05/10 because overall there are 10 tests, but only 5 in each set?????
I seem to have some non sig results, some very sig and some approaching signficance - so hopefully i can argue that the non-sig ones would be significant! :-) :-):-)
Most papers I read don't seem to bother with bonferroni corrections (even if a statistician would argue they should) so I'd try and get away with doing as few as possible! If you can argue that the positive and negative are independent then you could do each divided by 5 for 5 t-tests in each set ie p of 0.01 (I think but could be wrong)
I'm just about to get a call from sup - I think she is going to advocate MANOVA :-s which I really don't want to do. I'd rather do t-tests or ANOVA followed by t-tests. I'll wait and see what she says!
In the end I went with repeated-measures ANOVA followed up by post-hoc t-tests (dependent or independent, depending on the hypotheses) and I used bootstrapping to help with one of the non-normal variable I had. Transforming didn't help, and I didn't want to lose power with non-parametric tests. I think I'm ok, going to pass it through a stats guru asap though!
PostgraduateForum Is a trading name of FindAUniversity Ltd
FindAUniversity Ltd, 77 Sidney St, Sheffield, S1 4RG, UK. Tel +44 (0) 114 268 4940 Fax: +44 (0) 114 268 5766
An active and supportive community.
Support and advice from your peers.
Your postgraduate questions answered.
Use your experience to help others.
Enter your email address below to get started with your forum account
Enter your username below to login to your account
An email has been sent to your email account along with instructions on how to reset your password. If you do not recieve your email, or have any futher problems accessing your account, then please contact our customer support.
or continue as guest
To ensure all features on our website work properly, your computer, tablet or mobile needs to accept cookies. Our cookies don’t store your personal information, but provide us with anonymous information about use of the website and help us recognise you so we can offer you services more relevant to you. For more information please read our privacy policy
Agree Agree