Hello all

I know that it is often useful to log-transform a variable if the data is positively skewed, so I am planning to do this for both my DVs. The main element of my dissertation is a 2x2 ANOVA. However, I have also gathered data on a lot of other variables which I consider to be my IVs in that I will see if there are correlations between them and my DVs. Some of these IVs have data which is positively skewed as well. Not all, but some. My question is, should you transform any positively skewed data you find, or is it only applicable to DVs?

Thanks!

I always think that you should transform everything, if you are transforming one variable - if you are using it in the same analysis.

I thought you only did that with different levels of the same variable e.g. Day 1, Day 2, Day 3. But then if more than one variable has a skew and one will end up examining different levels of it, it needs transforming.

If you were going to transform every IV and every DV in the same analysis regardless of skew, surely you'd risk transforming normal data into negatively skewed data?

ooh, I guess so. Tbh, I've been using bootstrapping recently, rather than transform, you can get it on SPSS 18.

Bootstrapping is beyond my ken - think I will just stick with the (limited) wisdom of Andy Field for now.

Now I'm discovering variables with negative skew. You can obtain reverse scores and log-transform them to make them more normal, but I don't know if you can then have them in the same analysis as variables you have log-transformed in the ordinary way.

I feel I am being lowered down a very deep well.

:-)

not sure about that. ANOVA is usually robust to non-normal data, as long as its not too non-normal and sample sizes are equal (ish)

My sample sizes are great, but there is such large skew on many of my variables I think that ANOVA would be out of the question in this instance. Mostly now I'm just weighing the benefits of a transformation against just going straight for non-parametric tests.

What kind of ANOVA is it? within/between groups?

I considered transformations at one point for my data but once you do it you can't really go back. At least in the sense of consistency, if one experiment has transformed data then they all have to! You can try and get away with not doing it for some experiments but if it's picked up on someone is likely to a tear a huge stats shaped hole in you :(

As for just DV's or IV's too I don't know for sure. I'm leaning more towards just DV's, unless the IV's are some sort of strange shared/linked measure.

The ANOVA is between-participants.

It seems both of my DVs are positively skewed most of the time (I've checked different groups and levels of the IVs). The DVs are k-value (a measure of impulsivity) and sociosexuality (a measure of how promiscuous someone is/wishes to be). Then the ANOVA IVs are sex (M,F) and exposure to certain stimuli (ladies, landscapes).

But I've collected other IVs like amount of money spent per week on booze, number of hours watching pornography each week, and these are positively skewed. There's one that is negatively skewed and that's religiosity.

I'm inclined to check what a transformation does to the data and then decide whether to go for parametric or non-parametric. Most of the IVs and DVs are crying out for a transformation as they are positively skewed, but as religiosity is negatively skewed it throws a spanner in the works as log-transformation is only for positive skew. You can reverse the scores with negative skew and then log it, but I don't know if I want to mix up this kind of operation in the same analysis with a conventionally log-transformed variable.

Therefore non-parametric tests are looking increasingly appealing...

After rending my garments somewhat yesterday, I have decided that ANY decision must be made today, no matter what.

I hate data when it does this (i.e most of the time :-s )

I'd say non-parametric. I assume you've checked for outliers? I had to remove about 3/300 people because they'd scored highly on one of my scales and it completely corrected all my histograms :-) Turned out they were idiots who'd answered the same for every question.

Does anyone know if a non-parametric version of a 2x2 between-participants ANOVA even exists? The things I've been looking up seem to either have only one IV or only one DV. And I have two of both. Do I have a design which can't actually be analysed?

isn't it a friedman's test (or something!?) or a krusal wallis?