The data from my last work, most of them are obtain normal distribution. Only one of them are lognormality and I log transfered the lognormality data. But a reviewer ask me to log transfer all my data, even they are normal distributed. I think it's wrong.
So, do I need to do log to all my data?
Hi,
See this link, I think you can get answer to your question
https://stats.stackexchange.com/questions/18844/when-and-why-should-you-take-the-log-of-a-distribution-of-numbers
Best Regards
Xinghua -
I would not consider transforming data unless I had a really good reason. If you give your reason you may get more helpful responses.
If it is for purposes of controlling heteroscedasticity, that can reduce it, but i would not generally recommend using it. Below is a link to a discussion of heteroscedasticity, with several examples. Then there is a link to a tool you could use, and finally the project that covers this:
https://www.researchgate.net/publication/333642828_Estimating_the_Coefficient_of_Heteroscedasticity
with
https://www.researchgate.net/publication/333659087_Tool_for_estimating_coefficient_of_heteroscedasticityxlsx
Heteroscedasticity for estimated residuals in regression is not a bug, it's a feature.
The following is on the nature and magnitude of heteroscedasticity, for regressions of form y = y* + e:
https://www.researchgate.net/project/OLS-Regression-Should-Not-Be-a-Default-for-WLS-Regression
Best wishes - Jim
As you are working in a hospital, I assume that the variables you are talking about are biological/physiological variables, usually representing some kind of concentration. Such variables typically have a positive domain (the values can only be positive, never negative). Further, it makes much sense for such variables to analyze relative (not absolute) effects [distinguishing this is not relevant when the effects are tiny, but when effects can be larger, it makes an important difference]. Biology is relative, so to say. Cells, tissues, organs, organisms, populations and species actually do not mind absolute changes, but relative changes (the impact of a change always depends on what is already there).
The standard statistical procedures we use analyze variables without restricted domains (that can have positive and negative values) and analyze additive (not relative) effects. If you analyze the logs of such a biological variable, the domain boundary is resolved (what usually also resolves issues with symmetry and variance homogeneity; things I did not mention above) and absolute differencesin logs represent relative changes on the original scale (so the standard procedures on the logs actually analyze relative changes on the original scale).
So, actually, all you have to do is to think if a variable is allowed to take negative values. If not, use the log transformation. It is not really relevant if the data of the untransformed variable is insufficient to reject the null hypothesis that the data is from a normal distribution.
So it' likely that the reviewer is right.
The reviewer should have given you reasons for the suggestion. If a measurement variable does not fit a normal distribution or has greatly different standard deviations in different groups, you should try a data transformation. If you have a large number of observations, compare the effects of different transformations on the normality and the homoscedasticity of the variable. If you have a small number of observations, you may not be able to see much effect of the transformations on the normality and homoscedasticity; in that case, you should use whatever transformation people in your field routinely use for your variable. Many variables in biology have log-normal distributions, meaning that after log-transformation, the values are normally distributed. This is because if you take a bunch of independent factors and multiply them together, the resulting product is log-normal. Remember to back transform your results after the test because you cannot report test results of mean, standard deviation or error etc of transformed data, otherwise the significance of your result will be masked. The reviewer may just be right.
If you use these variables in modelling the only way to prove you're right is to use residual tests.
Andrey Davydenko Maurice Ekpenyong Jochen Wilhelm James R Knaub Hussein A Mustafa thank you all.
One measure of my data are log distributed, I transferred them. So it has a different y-axis from the others in the same figure. So the reviewer ask me to make them all the same by transfering them all whether they are log of nomarl distributed. That's why I don't think it right.
As James R Knaub and others state, you need to give more details. It depends what you are doing with the variables. If I want to see which of 7 variables (say the count for something on 7 days in the week) has the highest mean or median, transforming one (and thereby changing the scale) means that the comparison would not make much sense unless you transformed others. In some examples it would make sense only to do one. You need to say what you are doing in order to get useful responses. Also, you haven't said why you wanted the variables to be more similar to the Normal distribution. I assume this was for more than aesthetics.
- Badges
- Science topic
- Similar topics
- Correction