I have a query regarding data transformation if anyone can provide any guidance please?
I was wondering if, generally, it is possible to transform a variable's raw data twice, using 2 different methods, for the purpose of 2 different tests? I will provide you with a little background to my study first. I have a variable for 'Adverse Childhood Experiences' containing 1 score per participant. N = 113; however, 65 of these are 0 values and 3 are missing data - which I believe is disrupting my data considerably. I understand that it is not advised to simply remove the cases that read 0 just because there are many (however, if you recommend otherwise please let me know if so and why).
Useful to note here is that this variable has a skewness of 1.943, and because of this, I have made the decision to transform it.
I am carrying out a path analysis with 1 IV, one DV and 2 mediators. In the first instance I am carrying out a t-test (IV - gender, DV - ACE score) and then in the second instance I am carrying out a linear regression (IV age, DV - ACE score), to understand whether age and gender need to be included in my path analysis as covariates. In order to meet the assumptions of the t-test (namely, normal distribution across both levels of the IV: male and female) I have transformed the raw ACE data this using Tukey's formula, which brought the skewness to < 1 for each IV level - great. But then when I go to carry out the linear regression, and aim to meet the assumption of approx. normal distribution of residuals, the assumption is not met on the Tukey transformed ACE data. I have carried out a number of other transformations on the raw ACE data and the only one where the residuals are normally distributed for the regression is through a Log10 transformation.
My question is this: am I able to carry out the t-test with the Tukey transformed variable data, and then the linear regression with the Log10 transformed data? Or is it the case that I need to use the same transformed data for each stage of the analysis (ie. both Tukey or both Log10 for t-test and linear regression and then the same onward path analyses?)
If it is the case that I will need to use the Log10 ACE data to go back and carry out the gender t-test, it is useful to note here that I have done this already and when inspecting the Log10 transformed ACE data across the gender variable descriptives table the results come out very strange - for example, N for males goes down from 15 to 6, and N for females goes down from 115 to 59, and there are outliers, where there were none in the Tukey transformed data descriptives, so it is confusing me a little.
Any guidance welcome!
Thank you
I assume your data are count values (i.e., the number of adverse childhood experiences). If so, then use a count model (negative binomial is usually a good choice). Transfroming your data is not required at all. If excess zeros are a problem, consider a zero-inflated negative binomial model.
Hello Nicola,
I don't see any problem in transforming the data with different methods for each test. At the end you will have to interpret the results after back transforming to original scale.
However, I think you don't need to perform a t-test. You can fit the regression model and add gender and age as covariates. A model selection can be used to exclude covariates from the model, but in your case there is no need for it maybe, you have two covariates only. You can add an interaction, age*gender, in the model too.
Good luck with your work!
I assume your data are count values (i.e., the number of adverse childhood experiences). If so, then use a count model (negative binomial is usually a good choice). Transfroming your data is not required at all. If excess zeros are a problem, consider a zero-inflated negative binomial model.
Log-transforming count data is discouraged. So you had better go with a count model (like Poisson etc).
With model fitting just like shoes you can force a foot in a shoe by transforming or get a shoe that fits the foot. As an exploratory analysis transformation is ok for me, and for a second pass I would try a different model without transformation. My other colleagues above have already answered this.
Have a look at Andrew Gelman's book GELMAN, A., & HILL, J. (2007). Data analysis using regression and multilevel/hierarchical models. Cambridge, Cambridge University Press. It has a chapter on count data GLMs and some examples of negative binomial and zero inflated Poisson models with social inequality data.
count data can be ordinal also why not try non parametric tests
Generally, as was suggested above, it seems that a zero-inflated negative binomial model is the best approach here.
But if you want to use OLS, perhaps you can try the SQRT transformation, which ensures good properties if you data follows the Poisson distribution. The SQRT transformation can be applied to 0's as well.
Another approach is just to replace ACE values with ranks..
- Badges
- Science topic