A statistic is sufficient for a parameter when it retains all information in the observed data relevant to that parameter under a specified probability model. After computing it, further analysis of the original observations adds no information for inference about that parameter. section 3.2.3
Sufficiency can make data reduction lossless for a particular inferential goal: a compact summary can replace the full sample without sacrificing information relevant to the parameter. Sufficiency depends on the statistical model, so a statistic may be sufficient under one model but not another. section 3.2.3
Estimating a normal mean
For observations 5, 7, 6, 8, and 4 from a normal distribution with known variance, the sample mean, 6, is sufficient for the population mean. The individual observations add no information about that mean beyond their average. section 3.2.3
Sufficiency is one of Fisher’s criteria for a good estimator, alongside consistency and efficiency. It concerns preserving information relevant to the parameter, not merely producing a compact summary. section 3.2 section 3.9
