Standard deviation
Standard deviation is the square root of variance, describing spread around a mean in the same units as the observations.
For n observations with mean xbar, the usual sample statistic is s = sqrt(sum((x_i - xbar)^2)/(n - 1)), requiring n > 1. Population SD instead uses the population's variance; specify whether a reported number is a population quantity or a sample estimate. Squared deviations make s sensitive to extreme observations. [1]
Latency SD in milliseconds describes dispersion, not the uncertainty of estimated mean latency. Under independent, identically distributed observations with finite variance, the mean's estimated standard error is s/sqrt(n). A confidence interval needs an appropriate procedure as well as a standard error. [2] [3]
Neither SD nor standard error specifies p99 without further distributional information. In Hesela's separate synthetic example, copying each observed record four times preserves its sample mean but does not add independent draws. Equal measured values alone do not identify copied records; legitimate ties must be retained. [4]
Relations
Citations
- NIST e-Handbook 1.3.5.6: Measures of Scale, sample variance and standard deviation
- JCGM 100:2008 (GUM), 4.2.1-4.2.3 and 4.2.7: repeated observations and correlation
- NIST e-Handbook 1.3.5.2: Confidence Limits for the Mean
- Hesela: spread, mean uncertainty and copied latency records; exact synthetic example