======= Date Modified 22 50 2010 19:50:13 =======
The final bit of data collection and analysis for me is a reliability study using dreaded intra class correleation co-efficients. I want to, hopefully, demonstrate that there is a strong level of agreement between my scores over the space of a week. Now, I don't have any fancy confidence interval statistics because I haven't started using it yet and it just seems like everything o do with sample size calculations is largely free fancy and supposition. Thinking that at least an attempt at a sample size calculation is better than nothing, I've done one that says I need 33 patients.

The null hypothesis is very conservative and says I will have an ICC (a 2 way random effects one) of 0.55. My alternative hypothesis says that it will be 0.7 or greater. 0.7 is acceptable for my questionnaire, but if I go for a null hypothesis like 0.7 and alternative hypothesis of 0.8 (which would admittedly be much better), I'd need over 100 subjects. I'm using sample size calculation software. This will be impossible based upon the time I have left and the fact that it would cost too much money. At least I'll have some evidence suggesting it's reliable. Does anyone else think that this is okay?

Apologies for the typo in the thread title.

For me, if I think there might be a chance of the stats techniques looking a little dodgy, I just whack in a ton of references that say its ok, so find a load of journals that say *something* statistical and back up your use of 33 participants.

Thank you, Sneaks. I think I've sorted it now. The absolute minimum I can go for is 46 patients. That's if I have 19 observations, a Ho of 0.7 and a H1 of 0.8. I couldn't have done the other, no because of my strong moral compass, but because my viva examiner could argue that it's unethical, as it would be me choosing numbers to fit my own circumstances. Meh.