Hi Johanna,

In my opinion, there are two reasons why you might want to take the logarithm of raw RTs. First, it is your intention to analyze the raw data (instead of aggregated means, see below) and you want to take care of the skewness of the individual RT distributions. Second, you want to handle outliers. However, doing both makes little sense if you intend to analyze mean RTs. Those are (usually) distributed normally anyway.

In sum, if you plan to analyze mean RTs, don’t worry about the skewness of the individual RT distributions. You do need to take care of outlying values though. I know that taking the logarithm of raw RTs is sometimes recommended as a means to achieve that, but it is my experience that a real outlier-exclusion algorithm works best. I prefer to exclude values that exceed the mean of an experimental condition with more than 2.5 standard deviations. The nice thing about this approach is that it is adaptive. You can always use it, irrespective of the mean RT observed in a particular study.

There is just one downside to this particular algorithm: You need a substantial amount of observations in each cell of the design. Otherwise, the outlying values (especially when they are really extreme) inflate the SDs so much that it becomes impossible to identify the outliers. In that case you might want to get rid of the most extreme values by using a fixed cutoff (e.g., everything above 3000 ms). In a second operation, you can use the traditional 2.5 SD cutoff.

More information on this matter can be found in the paper by Ratcliff (1993).

Good luck,

Adriaan Spruyt

Thanks Adriaan! That is very helpful information.

Just out of curiosity, you talk about reaction times, would you treat reaction times the same way as decision times?

Of course. Any timed response yields RTs, be it a binary decision, a go/nogo response, whatever ...

Thanks again Adriaan und thanks a lot Marcellin!

I appreciate your comments and I will look into the MAD method as well.

Johanna,

When I was a PhD student I used the log-transformation on the RTs then run regression analyses on the log-transformed RTs.

Now, that was a lot of time ago (sigh).

Since then I have learned that this may work sometimes but that it is a very unnatural way to force a variable that is not normally distributed to become a "well-behaved" normally distributed one. It's like wearing 37-size shoes (European sizes, sorry, no idea how this translates into American or UK sizes) when you'd need 39-size ones: it will do sometimes but you don't want to make a habit of doing this.

More technically the problem is that the normal distribution is unbounded in both directions, while RTs clearly have a bound at 0.

So the elegant way of dealing with this is to run regressions that suppose a distribution law that can be left-bounded (as are RTs). Gamma law is such a law.

You can easily run gamma law regression analyses under R (http://cran.r-project.org/), which is a free and extremely powerful statistics program. I do recommend you an excellent GUI (graphical interface) named R2STATS, that can be installed into R once you have installed R on your computer. Or you can do this through code lines in R using the lmer function from the lme4 package (that you'll have to install into R), but if I were you I wouldn't (R2STATS is based on lme4 but you issue commands by making mouse clicks, it's much more user-friendly).

If you are afraid of doing analyses the bases of which you do not master, well, don't! Because in a nutshell this is how it works:

When you do a regression "that assumes a normal distribution of the dependent variable", in fact it assumes that the distribution of the dependent variable is normal for each value (or level, for factors) of the independent variable (i.e., normally distributed conditional to each value or level of the independent variable).

In a gamma law regression analysis, the assumption is instead that the distribution of the dependent variable follows a gamma law for each value (or level, for factors) of the independent variable.

I think this way of dealing with RTs is more elegant AND more correct (you won't have to trim out very large RTs because gamma law accommodates them -- it is not a symmetrical law as the normal law, so data very far in the upper values are expected).

HTH.

Cheers,

SCM

Thanks! This is actually the way I am going right now after getting some similar tipps outside RG. Maybe I'll try that R2STATS, sounds promising.

You are welcome!

R2STATS is the easy way to go, and it also allows you to test mixed-effect models.

I am glad to hear you've got similar advise elsewhere and that it concurs with mine. Always good to hear :)

Cheers,

SCM

do not log transform, do not exclude outliers. here is a new approach

Article TO DETERMINE SKEWNESS, MEAN AND DEVIATION WITH A NEW APPROAC...