DIY statistics

>>1081723
Try asking /sci/, but beware that they're often snobby shitheads.

What does your data look like? A paired t-test is almost certainly what you want.

My experience is with R, not SSPS, but I can for sure help you figure out the right test to do.

>>1081801
Well as far as I understand it, you use the Student-T-Test when you have something like gender and your test result, and that you check if there is a difference between men and women scoring.

I have two tests, before and after, with numbers ranging between 0 and 120. After some research last night i finnaly setteled on the wilcoxon test. Any thoughts about my choice?

>>1081905
Well, theres more than one test that you could call a 'Wilcoxon' test. If its the Wilcoxon-Mann-Whitney test, it is plainly wrong for you data because you violate independence (you're working with before-after pairs). If its the WIlcoxon signed-rank test, its a valid test for your data. However, the WIlcoxon signed rank test is a crutch for when you'd want to use a paired t-test but you violate normality (e.g. your data doesn't look like a bell curve and your sample size is only like 6 or something). And there might be other 'Wilcoxon' tests with other assumptions.

Anyway, the if you can get your software to spit out an answer for the signed-rank test, you can get it to spit out an answer for a paired-t. And you really should use a paired-t if you can. A paired t-test is going to give you more meaningful results to report, but its up to you the investigator to know if your data calls for the signed-rank.

Many medical studies have issues from small sample sizes. Is yours 12? or more like 120? How about the distribution? I'm guessing its one bell curve with no real difference before after? Or do you have distinct before and after curves? If your distribution is something just plain screwy, there are three options: use signed rank because its for this situation; if you have good enough demographic info maybe you can fix it by breaking out multiple populations that got into your data and run an ANOVA (e.g. young women with infection related hearing loss vs. old men with occupational hearing loss etc…); or if your sample size is large its valid to assume normality and run the paired t-test anyway.
(cont.)

>>1082006
The way you describe it makes me think that you're running into some trivial issue like your data isn't formatted properly or you're not correctly calling out a paired t-test in the software. A paired t-test really is preferred, please only use a WIlcoxon signed-rank test if you data actually has one of the validity contradictions I talked about, not because you never had the chance to learn about your software well.

The last thing I have to add is post more info and maybe you'll get more help.


tl;dr: you should work at this until you can run a PAIRED t-test on this data, but please at least do a little more reading

>>1081723
There was a paper somewhere about papers in the medical field. An alarmingly large percentage (40-ish) used fucked up statistics when analysing data. I don't remember what paper or where as this was a long time ago when I wrote my masters thesis. Shouldn't be too hard to find if you are interested in failures in both data analysis and peer review.

I had the opportunity to use a the house statistics guy which helped a lot.
Using statistics is hard if you aren't used to it. An option I have heard of is to use a shared statistican between groups, so that you have someone to consult when needed. But then again you need to have a few research groups to share the cost of those wages. Anyway something to consider for small groups.

>>1082008
>>1082006
Ok, I see I still have a lot of reading to do.

Thanks a lot, I'll do the reading, take a better look at my data, and probably some time tomorrow post more details.
Thanks once again!

File: 1440223327425.png (297KB, 376x488px) Image search: [Google]

297KB, 376x488px

>>1082062

>mfw I see ibm spss logo on /diy/

I agree with the other guy, it sounds like you need to use a paired two-tailed t-test.

Also, is your study controlled? I.e. control and treatment groups? If so, you can subtract the post treatment dB from the pre treatment dB and get a value for improvement, but in this case you would compare that total improvement value for the control and treatment groups with an independent two-tailed t-test. If not, just ignore that and go with the paired t-test.

>my Prof won't give me any cash for a professional to do the statistics

What? Money? Go to the stat department and make some friends.

>>1082751
Came here to post this. There must be some professor or other PhD student that would be glad to help you out, or at least push you in the right direction.
While I am still undergrad, I have never had a professor refuse to spend a bit of his spare time to answer a few questions related to his work or domain.