FY 1983

Data Intercomparison Theory—Vol. IV, Tercile tests for location, spread and pattern differences

Preisendorfer, R.W., and C.D. Mobley

NOAA Tech. Memo ERL PMEL-41, NTIS: PB83-182377, 45 pp (1982)

Two data sets can be examined for closeness by considering them as field maps and pretending that one is trying to forecast the other. This is done by applying some recently devised forecast verification techniques to the pair of data sets. The technique we apply for the purpose is the tercile (or Trinomial Stochaster) technique wherein, over a given set of points in space, the two fields have their 0-class, 1-class and 2-class errors tallied, and examined for statistical significance. These class errors can be used to gauge the closeness of the three main attributes of the data sets: their locations (akin to averages), spreads (akin to variances), and (spatial or temporal) patterns. In illustration, the tercile technique is applied three times: to show how to gauge the effects of different objective analysis methods on the same raw data set; to examine the self-predictability (and hence noise or information content) of a data set; and to devise a new principal-component selection rule using the concept of self-predictability.