I am trying to build a log log model for which I transformed dataset taking logs for all the variables( independent and dependent continuous variables).
After taking log, zero values in some of columns are transformed with -Inf.
So , I replaced those -Inf with zero and then computed the correlation matrix. I am not convinced with doing this as I feel that there should be any other way for treating those - Inf values.
Looking for suggestions and inputs.
Thanks
You need to understand the nature of your '0's first.
They can be:
- missing values, i.e. measurement not done, '0' was used to record that
- values measured, and found to be below the detection limit
- true zeros
if missing, replace with NAs, drop the rows from analysis
if below detection limit, try to define that limit, then use a random value generator to make up one between 0 and the limit
if true zeros are abundant, consider box-cox instead of a log, a more generic transform that has a log as a particular case
In any case, replacing your -Inf with 0 is definitely wrong!
- Badges
- Science topic
- Similar topics
- Analytical Chemistry
- Column