book meme
I've done the book meme that's been going around before, but just so you can share my pain, here's the result for my current closest book:
1. Grab the nearest book.
2. Open the book to page 123.
3. Find the fifth sentence.
4. Post the text of the next 4 sentences on your blog along with these instructions.
5. Don't you dare dig for that "cool" or "intellectual" book in your closet! I know you were thinking about it! Just pick up whatever is closest.
This is an important principle to keep in mind as you conceptualize research problems. For example, if you initially select individuals who scored "severe" on the Beck Depression Inventory, it would make little sense to then correlate their depressiveness with some variable of interest. Similarly, if you are doing research on gifted students, you probably should hesitate to correlate achievement with some other variable. And if you are the admissions officer at a selective university, do not be surprised to find that your students' SAT scores bear little relation to their college GPAs.Okay, that doesn't really convey the full horror of the fact that I've been reading statistics books for the past hour and a half, and R help files* for four hours before that (I think that's a record for total hours of work! And with only 3 hours of wasting time in between!). It's so demoralising when I actually try to do some work for a change and still accomplish nothing...
Since all that reading hasn't helped me at all, does anyone out there know how I test whether there's a statistically significant difference in noise between multiple groups? I don't care what the actual values are in each group; I just want to figure out which one has the least variation.
*also, why does R have to have such an un-google-able name?
Labels: memes
posted by Lucy @ 7:29 PM
13
comments
13 Comments:
http://www.le.ac.uk/biology/gat/virtualfc/Stats/variance.html
http://www.itl.nist.gov/div898/handbook/eda/section3/eda35a.htm
there are also stats within SPSS output that tell you how much variance there is.
ps. What's a ceiling or floor effect?
(is that what you're reading about?)
Thanks, Shrinky; you rock! Although, I'm not sure if I can use those tests, since my null hypothesis is that the variance is the same and I want them to be different...
I don't know what a ceiling or floor effect is. Would it be where values are squished against an upper or lower boundary?
It strikes me that I should know that answer to your stat question but it would require a little thought. However, ceiling and floor effects are also called range restriction. Range restriction occurs if there's a reason that a bottom score (or top score) can't occur.
Here's an example. You select people into a job based on a cognitive ability test. You then train those who got the job on something. You test to see effectiveness of traning. Chances are there will be floor effects (i.e., no one will test particularly low) because everyone in the job is relatively smart and your sample of score are unlikey to spread down to the lowest possible scores on the training test. Also your average variance is lower when RR exists because in my example, the bottom half of the possible variance that would have existed in a non-restricted sample is cut off.
Does that make sense?
R is also called multiple regression coeficient I believe.
Sorry for the long comment. I blame it on Luckybuzz for sending me over here with a link.
Long comments aren't a problem :) I feel a bit better that there isn't something obvious I'm missing (although that is always likely since I know almost no stats). Thanks.
I just ran the levene test (my first successful R function all day!) and I have a nice p value of < 2.2e-16, but now I don't know which groups are different...
Plot your points! You should always be plotting things so you can look at them and see if you have any kind of ceiling or floor effect - and to see how they all fall (that is, do scatterplots of each).
I have plotted them. I can see differences and there are differences in the CVs. I just don't know if they're significant.
Wait, I'm confused. Your p value is sig, right? So there is a sig difference in your variances. Are you comparing more than two groups? If so, do post hocs?
I have 26 groups, so when I compare them all at once, there are significant differences between some of them, but I don't know if it's between groups I care about. I've been trying the same test on smaller subsets of my groups, but it's kind of a pain to compare every possible pair of groups.
I don't know what post hocs are, either...
Sorry, I didn't mean to get you involved in my stats mess. I was really just whingeing :)
I feel like we are chatting. :)
Well, to be above board, your post hoc ( as in, after the fact) should be planned. That is, which groups would you hypothesize would be different? You can then do n-1 post hoc tests (n=26 in this case0. So, I would just choose the pairs you think are likely to be different based on your hypotheses - otherwise you are "fishing" so to speak (just running a bunch of analyses hoping to find differences).
Where's statgirl???
It might have been easier if we were really chatting :) I don't have the data in the right form to compare just the groups I want, but I'll try that tomorrow. I do know which ones should/could be different for non-random reasons.
Thanks for your help!
So, just don't forget to list me as statistics consultant! :)
Post a Comment
<< Home