Hugo Kornelis (1/28/2013)
I though I understood the difference between STDEV and STDEVP, but apparently - I don't. And the explanations in Books Online don't help me either.
STDEVP: "the statistical standard deviation for the population for all values"
STDEV: "the statistical standard deviation of all values"
The only difference is the three words "for the population". Probably makes a lot of sense to someone who knows statistics inside out, but I am not that person. *lol*
The problem is that STDEVP is not actually the standard deviation of anything. It is actually an estimate of the standard deviation of a population for which only the values of a sample are known. BoL expresses it very badly. A better way of describing it would be:
STDEV: the standard deviation of an available sample, which is the standard deviation of a whole population calculated from all its members only if the sample consists of all the members.
STDEVP: an estimate of the standard deviation of a whole population based on data for a sample which is significantly smaller than the whole population.
The relation between the two is easy to express: STDEVP = N*STDEV/(N-1) where N is the number of members in the (available) sample. STDEVP is statistically a significantly better estimate of the standard deviation that STDEV is (in the sense that the statistically expected error in the estimate of the whole population is smaller for STDEVP than for STDEV, if one considers all possible populations and all possible sensibly-sized samples of them), provided the sample is enough smaller than the whole population. Of course in the extreme cases (e.g. sample contains only one member - STDEVP is indeterminate; and sample is whole population - STDEV is accurate and STDEVP is not) STDEV is more useful than STDEVP, but many practical cases are not extreme.
STDEVP is relevant only to finite populations or to finite samples of infinite populations for which the distribution is not known analytically, not to known continuous distributions or to random variables with known distribution. Of course some random distributions/random variables don't have a standard deviation anyway, and there can be no estimate for what doesn't exist.