It's a simple question! Please, anyone may explain me if we need transform data in all regression analyzes? I have 2 variables (counts) to do a regression (or correlation) analysis. The first one is touch of a species in a 1 m2 quadrat (pointing with a grid frame) and second one is species richness in this quadrat. I ran the analysis without data transformation, modeled by Poisson family on R. But I am asking myself if it's correct! Should I transform these data to run my analysis? (not for normalize!). Thanks in advanced!
One of the assumptions of a correlation is that the data for each variable is normally distributed (i.e., “bell curve”). I recommend making a histogram of each of your variables. If one or both are not normal, then use transformations to get them closer to normal. But if your data already looks fairly normal, there is no need for a transformation. Look for outliers in the histograms and the scatterplot. Outliers could dramatically impact the results, especially if you have small numbers. There is no need to normalize the variables, as normalization is part of the formulas that your computer is using to calculate the correlation. Best wishes with your study Jhonny! ~ Kevin
One of the assumptions of a correlation is that the data for each variable is normally distributed (i.e., “bell curve”). I recommend making a histogram of each of your variables. If one or both are not normal, then use transformations to get them closer to normal. But if your data already looks fairly normal, there is no need for a transformation. Look for outliers in the histograms and the scatterplot. Outliers could dramatically impact the results, especially if you have small numbers. There is no need to normalize the variables, as normalization is part of the formulas that your computer is using to calculate the correlation. Best wishes with your study Jhonny! ~ Kevin
Thanks Kevin. But look: after transformation, my data are not closer to normal. do you have any more suggestion?
Were your data normally distributed beforehand? A log transformation usually works well for counts if most of your counts are at zero and it declines precipitously from there. Another popular transformation for counts is a square root, so you could try that too. But if no transformation makes your data more normal, I wouldn't do it.
If doubts about distribution still remains, try Spearman's correlation, it does not assume any distribution (it is a distribution -free or non-parametric method)
you never need to transform. check out our new paper
Article TO DETERMINE SKEWNESS, MEAN AND DEVIATION WITH A NEW APPROAC...
- Badges
- Science topic