The NIST Uncertainty Machine (http://uncertainty.nist.gov) is a web-based software application produced by the National Institute of Standards and Technology (NIST) to evaluate the measurement uncertainty associated with a scalar output quantity that is a known and explicit function of a set of input quantities for which estimates and evaluations of measurement uncertainty are available. Input and output quantities are modeled as random variables and their probability distributions are used to characterize measurement uncertainty. The NIST Uncertainty Machine implements the approximate method of uncertainty evaluation described in the “Guide to the expression of uncertainty in measurement” (GUM), and the Monte Carlo method of the GUM Supplement 1 (GUM-S1). The NIST Uncertainty Machine is freely accessible worldwide from any computer connected to the World Wide Web. Using a Web browser, the user specifies all the necessary information in an easy-to-use form. The calculations are performed on a NIST server, which returns the results to the user by posting them on a Web page that is displayed in the user's browser, and by writing them to files that the user may download. The NIST Uncertainty Machine also produces a conventional uncertainty budget that lists the contributions that the uncertainties associated with the input quantities make to the uncertainty associated with the output quantity. We show how these contributions can be derived from the Monte Carlo samples, and in the process illustrate an interesting and informative fact that conventional uncertainty budgets ignore.
1 January 2017 in Review of Scientific Instruments
Measurements involve comparisons of measured values with reference values traceable to measurement standards and are made to support decision-making. While the conventional definition of measurement focuses on quantitative properties (including ordinal properties), we adopt a broader view and entertain the possibility of regarding qualitative properties also as legitimate targets for measurement. A measurement result comprises the following: (i) a value that has been assigned to a property based on information derived from an experiment or computation, possibly also including information derived from other sources, and (ii) a characterization of the margin of doubt that remains about the true value of the property after taking that information into account. Measurement uncertainty is this margin of doubt, and it can be characterized by a probability distribution on the set of possible values of the property of interest. Mathematical or statistical models enable the quantification of measurement uncertainty and underlie the varied collection of methods available for uncertainty evaluation. Some of these methods have been in use for over a century (for example, as introduced by Gauss for the combination of mutually inconsistent observations or for the propagation of “errors”), while others are of fairly recent vintage (for example, Monte Carlo methods including those that involve Markov Chain Monte Carlo sampling). This contribution reviews the concepts, models, methods, and computations that are commonly used for the evaluation of measurement uncertainty, and illustrates their application in realistic examples drawn from multiple areas of science and technology, aiming to serve as a general, widely accessible reference.
In several disciplines, measurement results occasionally are expressed using coverage intervals that are asymmetric relative to the measured value. The conventional treatment of such results, when there is the need to propagate their uncertainties to derivative quantities, is to replace the asymmetric uncertainties by ‘symmetrized’ versions thereof. We show that such simplification is unnecessary, illustrate how asymmetry may be modeled and recognized explicitly, and propagated using standard Monte Carlo methods. We present three distributions (Fechner, skew-normal, and generalized extreme value), among many available alternatives, that can be used as models for asymmetric uncertainties associated with scalar input quantities, in the context of the measurement model considered in the GUM. We provide an example where such uncertainties are propagated to the uncertainty of a ratio of mass fractions. We also show how a similar, model-based approach can be used in the context of data reductions from interlaboratory studies and other consensus building exercises where the reported uncertainties are expressed asymmetrically, illustrating the approach to obtain consensus estimates of the absorption cross-section of ozone, and of the distance to galaxy M83 in the Virgo cluster.