When determining the recommended value for a physical quantity, the evaluator is sometimes faced with the problem of how best to derive such a value. A literature search often yields discrepant values from different experiments or laboratories. In the present work a number of data evaluation procedures proposed in the last two decades have been tested on randomly generated data sets with different types of outliers. A modified Bayesian procedure is shown to be the most reliable, and this method has been applied to half-life data to give the following values: 137Cs = 10985±12 d, 90Sr = 10544±20 d, 154Eu = 3138.5±0.3 d and 252Cf = 968.0±0.8 d.

A comparison of maximum likelihood (or chi-square) and Bayesian peak fitting techniques shows that the latter can reduce peak intensity uncertainties by a factor of up to ten in the case of closely separated doublets, leading to greatly improved doublet resolution. The need for laboratories to demonstrate the reliability of their chosen peak fitting techniques and measures of peak intensity is also demonstrated.

A class of fitting functions which provides an adequate representation of a wide variety of Ge detector spectral peaks is proposed. Members of this class are composed of a summation of five functions, containing a total of ten parameters. A new fitting technique, based on Bayesian statistics, is also proposed. By adding a penalty function to the normal chi-square term, this technique improves the speed of convergence and the reliability of the minimization algorithms used during peak fitting. The Bayesian approach offers a satisfactory statistical treatment of peak-area uncertainties, avoiding the often unreliable approximations inherent in most other approaches to uncertainty estimation.

Objective data evaluation procedures involve the use of a formal set of rules rather than subjective judgements to detect, adjust and eliminate rogue measurements during the course of an evaluation. It is necessary to specify explicitly the error probability density function (a function specifying the extent to which experimentalists misestimate the uncertainties in their measurements) in order to establish the reliability of a data evaluation procedure, or to compare the effectiveness of different data evaluation procedures.

When a gamma-ray spectrum contains a number of photopeaks originating from the same nuclide, the calculation of nuclide activity depends not only on the values of, and uncertainties in, the photopeak areas, photopeak efficiencies, gamma-ray emission probabilities and peak area correction factors, but also on the correlations that exist between these quantities. When photopeaks corresponding to a group of nuclides overlap or are unresolved, simultaneous estimation of the activities of all the nuclides present in the group is necessary if the information contained in the spectrum is to be used efficiently — a procedure which can prove important when the only prominent photopeak associated with a nuclide of interest overlaps with the photopeak of another nuclide. Two procedures for activity estimation are described: both procedures take correlations into account; one procedure deals with the special case of activity estimation for a single nuclide, where unresolved photopeaks are absent or neglected; the other procedure deals with simultaneous activity estimation using all the photopeaks associated with an interrelated group of nuclides.

The statistical and computational advantages which arise from using a linear rather than a non-linear class of fitting functions are discussed. Two linear classes of Ge(Li) detector efficiency functions are constructed, and their range of applicability assessed using a large number of independent efficiency data sets. One of these classes, defined by ϵ(E, p) = [p1 + p2ln(E) + p3ln2(E) + p4ln3(E) + p5ln5(E) + p6ln7(E)]E, where E represents energy in MeV, is shown to provide a satisfactory representation of Ge(Li) detector efficiency over the energy range (80/120–1850) keV, for efficiency measurements with standard deviations greater than 1%. It is shown that when the full covariance matrix of the measured efficiency values is not used, the goodness-of-fit test based on the chi-square approximation is no longer valid, and the likelihood of inconsistencies between measured and fited efficiency values remaining undetected is increased.