normalization for standard deviation in PropertyStats about matminer HOT 3 CLOSED

(note that currently, we are normalizing by n always, not n-1)

from matminer.

@computron, perhaps a first issue to address is "what is our intended use for the PropertyStats.std_dev function?"

There seem to be two possible cases for employing property stats. A first one is measuring the variance in some property across a dataset. In that case, I agree that using the Bessel correction is appropriate because we are estimating the variance with only a sample of all possible entries.

The other case is using the function is more like measuring the variance as it relates to a certain entry in the dataset. For example, this could be the variance a property across the elements present in the material or across all sites in a crystal. In those cases, we are measuring the variance across the entire population (e.g., all sites in a crystal) and the uncorrected measurement seems more appropriate.

Given that the only current use of PropertyStats in matminer is for the the latter case, I vote we keep using the uncorrected standard deviation by default. I think we should support both as an option, as that we help ensure users of the PropertyStats functions consider which is most appropriate for their application.

Sound reasonable?

from matminer.

@WardLT Ok agreed. I was getting std_dev statistics across rows confused with std_dev stats within a row. The option for n-1 probably isn't needed at the moment.

from matminer.