Problems with correlated data

A misunderstanding exists in the PRA (probabilistic risk assessment) field over what constitutes correlated data. This report clarifies the applications that fit the initial intent of the definition. In addition, even when used as intended, current theory appears to give an overly conservative answer. A more realistic answer, which is still conservative (e.g., overestimates the failure frequency of the group being estimated), is obtained by assuming the data are uncorrelated. Definition: “correlated” data (as used in this paper) are data linked by a common data-distribution; i.e., if two separate components derive their failure rate from this same distribution, they are “correlated”. This is not the same as statistically correlated data wherein the data can always be statistically correlated (e.g., race, sex, age, and education of poor people)