The general idea behind transformation is to make a variable linear. Therefore, you can try various transformations and test it for linearity using tests for normality, as well as visual displays, Q-plots, etc.

The general idea behind transformation is to make a variable linear. Therefore, you can try various transformations and test it for linearity using tests for normality, as well as visual displays, Q-plots, etc.

One more thing: Most software packages have a command that will perform this for you. For example, in Stata you can run the command ladder which will generate several transformations and give you the statistic to show which fits the linear assumption the best

Dear Ariel Linden,

Thank you for your valuable suggestions.

Does a transformation method depend on the skewness of the data distribution. In other words, can you suggest any transformation method, which is most suitable/not suitable for negatively skewed data?

You can use power transformation techniques that will indicate the best transformation to normalize your data based on maximum likelihood principles. Like Ariel Linden indicated, there are many software packages that will perform these tests for you. For example, the TRANSREG procedure in SAS will determine the most appropriate transformation for your data via a Box-Cox transformation, and will then conduct an ANOVA on the transformed data. 

http://support.sas.com/documentation/cdl/en/statug/63962/HTML/default/viewer.htm#statug_transreg_sect059.htm

You'll need to transform the data and review it. I can't tell you which is the best transformation for your particular data set without testing it.

As far as possible, data should preferably be analyzed on its original scale as this helps better and straightforward interpretation of results. For statistical tests and modelling, what is important is the normality of the RESIDUALS not of the data. I would suggest that you analyze your data as it is and see if the residuals exhibit reasonable normality (and constancy of variance). If they do, that is it - no need to transform your data. For non-normal data (eg binomial, Poisson, ...), you can fit a model that directly takes the the distribution into account - avoiding the need for data transformation.  

Thank you David Glover and Subhash Chandra for your suggestions.

Dear Thomas,

You should choose several transformations for a set of data in order to normalize your data through Kolmogorov-Smirnov goodness-of-fit test.

 I hope you solve your request.

Angel

Hi Jobim,

I think that Subhash Chandra answer is quite appropriate, the statistical analysis must to tend to see what model fits better to data and not try to fit data to a given model. And Ariel Linden gives (for me) the best definition and maybe the only reasonable application of data transformation, to test it for linearity.

Best regards