We know that the estimation error (in percent) can be calculated with EEx = a*sqrt(var(x))*100/x, where EEx is the estimation error, a is a constant, var(x) is the variance (came from (co)kriging), and x is the estimated value by (co)kriging.
Considering an estimated value below 1%, the above equation may leads to large and meaningless estimation errors (thousands of percentages).
The question is that what should we do for estimated values less than 1%?
in the other side, we need to report the estimation error in percentages to classify mineral resources based on their estimation error.
p.s.: an example is Phosphorus grades at an iron ore mine. They vary between 0 and 1, but after performing (co)kriging, their estimation error ranges around 1500%.
Despite performing compositional data analysis (CoDa), this happens, more or less.
Truncated normal distribution should be the appropriate model for such situation. However, the mathematics behind is quite messy.
Perhaps, if the values are bounded from 0 and are far from their upper physical bound (1), log-normal distribution can be a suitable model (that is, one would assume that ln(x) instead of x is normally distributed).
One simple solution is Feature Scaling, for defining a bounded interpretable result.
If it is a measurement you can convert the unit. for example replace mg/kg for ng/ kg or km for m etc.
Your problem arises because you start with "We know that the estimation error (in percent) can be calculated with ..." It can be avoided by replacing your indication of estimation error with a direct representation of a confidence interval for your estimate. This might be obtained as an outcome from your co-kriging of the (presumably transformed) variate.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics