We have tried to make log10 of one variable normal, its skewness is reduced by 50%, but normality test of Shapiro Wilk in SPSS, with significance level is still .000 with statistics 0.821. It is still non normal. What can we do for converting non normal data into normal data?
Abhishek
Plot distribution curve over your vars.
Check the curve shape:
General guideline:
left skewed (left tail) = use square or cube.
Right skewed (right tail) = use log or reciprocal or square root.
For a more precise method: check the attached diagram. If everything still fails, opt for non-parametric tests.
Best wishes
Saiyidi
Article Introduction to SPSS
May be you can use dummy variables or non-parametric tests. Consider generalized additive models (GAM) and recursive partitioning analysis. G'd luk! :)
It depends of what kind of statistical analysis you are doing. You can use robust statistics or non-linear analysis. Please, see these two papers:
https://www.statmodel.com/download/GMM.pdf
https://www.statmodel.com/download/Skew.pdf
Hi.
It depends on the distribuition you have. However, you should have in mind that with large samples it is possible that the Shapiro test ALWAYS reject normality eventhough the sample is near normal. I would recommend you to look at the histogram (with Sturges rule) to look for normality. On the other hand, the law of large numbers states that for samples larger than 30 you can assume normality even if the Shapiro or Kolmogorov test says otherwise.
Hope this may help you.
Best
We can use non parametric test but I have to check moderation effect of variable on relationships, so for that I need to have normal data. So how we can transform, I have already used log,sqrt,box-cox,arscine etc, but still significance level of Shapiro Wilk is 0.000, How we can increase the significance level.
Maybe a stupid idea, but check whether the skew is to the left or to the right. Many random variables are skewed towards the left (many low values), and the transformations you mention can correct for that. But if your data are the rare case where the skew is to the right, the usual transformations only help if you mirror-flip the distribution. Skew to the right is strange, however, and may be the result of outliers. -- Also check whether your procedue is robust against non-normality. Don't just rely on the normality test, it is probably more sensitive than the procedure you are using. Plot the (transformed) data against the normal distribution in a probability plot, and if it looks okay, igore the normality test. -- If all that doesn't help, turn the variable into a binomial one (dummy coding).
Unfortunately, SPSS does not have an easy guide to finding the best transformation to make a variable approximately normal. If you were using STATA the ladder command would provide you with Tukey's Ladder for transformations to make variables approximate a normal distribution. There are ways of looking at the data and seeing how it deviates from a normal distribution, and than deciding which transformation to use, but that would take up a lot of space/time to explain. So wIthout getting into the details of each you might try different transformation (square root, squared, inverse, ect). Trial and error and see which approximate the most normal.
The larger the sample the more likely Shapiro Wilk will say the data is not normal. Have you tried QQ plots? I find those helpful in deciding on if a large sample is approximately normal.
Abhishek
Plot distribution curve over your vars.
Check the curve shape:
General guideline:
left skewed (left tail) = use square or cube.
Right skewed (right tail) = use log or reciprocal or square root.
For a more precise method: check the attached diagram. If everything still fails, opt for non-parametric tests.
Best wishes
Saiyidi
Article Introduction to SPSS
Why are you trying to transform to achieve normality? Please provide more information. For example:
1. Is the variable in question an outcome variable for some kind of multivariable model? (I'm guessing this is what you mean when you say "multivariate data".)
2. If so, what is the variable?
I ask question 2, because it may be that you should be considering a different kind of model rather than trying to transform to achieve normality. (E.g., if the DV is a count, you should be considering a regression model for count data, such as Poisson or Negative Binomial regression.)
HTH.
Have you examined the data using a Q-Q plot and/or box-plots to check whether the non-normality is a result of a few outliers? Often I've found data having severe skewness is plagued by one or two extreme outliers. There are a host of methods that are accepted for dealing with outliers that may make more sense than applying a data transformation.
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software
- SPSS