I'm working on data that I want to transform in order to get a normal distribution. However, some of these values are negative. I can of course add a random constant, but to use it on multiple variables, I would like to add the lowest number in the list so everything will turn positive.
So far I tried this in the COMPUTE VARIABLE menu:
"Transformed_variabel"=LG10("Variable"+ABS(MIN("Variable")))
With LG10 being the log transformation, ABS to return the absolute value of the lowest number and MIN to return the lowest number.
What am I doing wrong? Can I use "MIN" in this way?
UPDATE:
To answer your comments so far: I have of course looked at the data first. A histogram of the data showed a left-skewed distribution. The data that I'm talking about is a difference in weight% before and after an intervention. Nevertheless, when applying a logarithmic formula to any type of data, the answer should be the same.
Hi,
You use the Box-Cox transformation without any data preprocessing.
I would not recommend doing this without knowing what is going on. At the least you should have provided some type of graph of the function that you want to transform. what type of variable are we talking about? etc, etc, etc. I think it is more likely that there is an error of some type here than that you need to shift the axes. Best, David Booth
1) if you have negative data in your dataset, logarithmic transformation is a strange idea. You speak of differences: there is no reason for a difference to be always positive, so log-transform them is suspicious. You may consider ratios instead of differences, and then log-transform them.
2) your formula makes the smallest value of the variable be 0, and log1(0) is -Inf, so may cause problems. If you really want to do a log transformation, you should add a 0.1 or something like that to ensure strictly positive values.
But once again, with the example you give, log-transformation seems a very bad idea with no support. log-transformation is situed for situations when you know your data must be positive.
Thank you for your answers. I will try to work with ratios as Emmanuel Curis suggested
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software
- SPSS