I have a lot of data with 0 as a value, its non normal and skewed heavily to the the right which is fine and expected with the 'positive' values, however if i normalise using log etc it removed these 0 values on me as it cannot exist. Are there any other normalising calculations that can keep the 0 values in so as not to falsely inflate the predictive power of my biomarkers. Have tried square root transformations with and without subsequent log transformations but they are still quite non-normal
Thanks
Hello,
this approach may help you;
Article TO DETERMINE SKEWNESS, MEAN AND DEVIATION WITH A NEW APPROAC...
Thanks Mehmet Guven Gunver
Quick query, my data is right handed in its weighting as I have numerous 0 values and none in the negative realm, would this help with that at all?
Thanks
You could try the cube-root.
If you are interested in predictive power then if your event is binary the possible equations to use in a logistic regression model may be (bm=biomarker)
event ~ bm^(1/3) [cube root]
or possibly better create a second variable bmo which is 1 of bm==0, and 0 other wise
event ~ bmo + bm^(1/3)
or perhaps use a restricted cubic spline with about 4 knots.
event ~ bmo + rcs(bm, 4)
A similar issue has been recently discussed here: https://discourse.datamethods.org/t/troponin-its-use-and-misuse-to-rule-out-and-rule-in-heart-attacks/509/57
- Badges
- Science topic
- Similar topics
- Analytical Chemistry
- Column