Stats question

[Ollie123]

---

Members do not see this ad.

So...past experience has told me this is usually not the best place for more complex stats questions, but I'm going to give it a try anyways🙂

1) Is anyone aware of a non-parametric analog to a mixed between-within repeated measures ANOVA? The short version is we have data that significantly violates normality (extremely leptokurtic). Traditional transformations do not appear to be helping (they seem more effective for dealing with skew and/or platykurtosis rather than the current situation). Part of me thinks the reason is that leptokurtosis should (I think) inflate power, so people may not be as motivated to look for transformations for things that are more likely to be significant😉The only thing I have found suggest converting the data to an ordinal format and running the usual RM analysis on the ordinal data may have some merit but it is from a dubious source and I cannot find any other papers confirming it. I think GEE might be a solution, and I can use a Poisson link function, but there are some problems with that (see below).

2) I'm not sure if anyone has used GEE before, but my understanding is that main effects are rendered uninterpretable when interactions are included in the model and have to be tested separately. Can anyone confirm this? It poses a problem for the analysis I hope to conduct that is certainly not insurmountable but seems unnecessarily convoluted to me.

Lastly, has anyone ever encountered a situation like this before (extremely non-normal data that they could not correct with transformation)? Reading the literature would imply that this is a rare situation, but based on my experiences it seems to be incredibly common (at least for psychophysiological data). I am increasingly convinced that people just run the RM ANOVA anyways and cross their fingers...something I am sorely tempted to do myself🙂

 

[Therapist4Chnge]

I got nothing.



I would like to one day be able to answer your questions, so I am going to sign up for a few stat seminars that are being offered this Spring and Fall....ugh. I'm not sure if division 5 has a listserv, but it may be a good place to start.

ps. You may want to PM Jon Snow and ask, he is pretty knowledge about stats stuff. He pops in from time to time.

 

[Psy1]

I took all of Dr. Wuensch's stats courses. He is an incredible teacher and an EXCELLENT online resource. This is just a link to his stats lessons page, but there are links to SAS and SPSS programs as well, including syntax. He is amazing. I would check it out.


http://core.ecu.edu/psyc/wuenschk/StatsLessons.htm

 

[Ollie123]

I appreciate the link, but am I missing something? Everything on there seems relatively basic, and I didn't see anything that addressed GEE or non-parametric tests for complex designs. He has a section on boostrapping which I actually think would work for this as well, albeit to my knowledge no one has even tried to apply bootstrapping procedures to a design of this type before. I'm a bit of a stats geek for a clinical psych student, but am not sure I'm up to the task of developing new techniques😉

 

[JockNerd]

What's "extreme" kurtosis? You can have values up to a raw kurtosis value of 10 before it breaks down GLM analyses.

 

[Ollie123]

Oh, we are wayyyyy past that. I am close to 10 on the "good" ones. Many are 20+. Some are 50+. I can pull them in some by transforming, but not enough for my liking.

I ran it on the trial-by-trial data before I aggregated it because I was contemplating a different type of analytic approach that hasn't been used before. Now I know why...I almost passed out when i saw the values. I did not realize skew and kurtosis could reach into the 2000's, but it can😉 Aggregation pulls that down a great deal, but not enough I'm comfortable calling it normal.

I'll look into the haseman-elston regression. If its what I'm thinking of I think I've mostly seen it used with twin data and behavioral genetics type work, but I'm always up for trying something novel🙂 Maybe it would get other people looking at alternative approaches because as I said...I guarantee I'm not the first person to come across this problem given there are thousands of papers using similar methodology, but a seeming lack of "statistical caution" so to speak.

 

[Existenz]

why don't you tell us more about your study. I'm not even sure what it is about so I cannot comment on your attempts at analysis.

 

[Ollie123]

Well, the overall design is more complicated but this analysis is relatively straightforward.

Two groups: Drug and Placebo (between subjects)
Picture Type: Three experimental conditions and a true neutral (within subjects). They were presented with 10 of each type in a random order, but these are averaged together.

We have ratings and psychophys data for their responses to the pictures. The ratings came out fine with simple analytic approaches, but the psychophys is giving us problems since the distributions prevent us from using the same approach.

 

[DynamicDidactic]

I am far from a "stats person" but here it goes.

It looks like you have a very tricky situation that requires some uncommon statistical procedures. I am sure there is a nonparametric equivalent to the mixed between-within repeated measures ANOVA. However, you will need to look through the literature to find it and I don't know if any stats programs have an option for it.

Due to your extreme violations of normality it seems you may want to consider bootstrapping your data. Fortunately, SPSS now allows for bootstrapping if you found a way to run this nonpara test in SPSS.