Tutorial on evaluation of type I and type II errors in chemical analyses: From the analytical detection to authentication of products and process control

Abstract

Uncertainty is inherent in all experimental determinations. Nevertheless, these measurements are used to make decisions including the performance of the own measurement systems. The link between the decision and the true implicit system that generates the data (measurement system, production process, category of samples, etc.) is a representation of this uncertainty as a probability distribution. This representation leads to the probabilistic formalization of the possibility of making errors. In the context of regulations established by official agencies, it is important to use these statistical decision methods in some cases because the own norm makes them mandatory and, in general, because this is the way of reasonably evaluating whether a working hypothesis is rejected on the basis of the experimental data. The aim of the present tutorial is to introduce some ideas and basic methods for the critical analysis of experimental data. With this goal, the basic elements of the Neyman–Pearson theory of hypothesis testing are formally introduced in connection with the common problems in chemical analysis and, if this is the case, their relation to the norms of regulatory agencies. The notion of decision with ‘enough quality’ is modelled when explicitly considering: (1) the null, H 0, and alternative, H 1, hypotheses. (2) The significance level of the test, which is the probability, α, of rejecting H 0 when it is true, and the power of the test, 1 − β, β being the probability of accepting H 0 when it is false. (3) The difference between H 0 and H 1 that has to be detected with experimental data. (4) The needed sample size. These four concepts should be explicitly defined for each problem and, under the usual assumption of normal distribution of the data, the mathematical relations among these concepts are shown, which allow the analyst to design a decision rule with pre-set values of α and β. To illustrate the unifying character of this inferential methodology, several situations are exposed along the tutorial: the design of a hypothesis test to decide on the performance characteristics of analytical methods, the capability of detection of both quantitative and qualitative analytical methods (including its generalization to the case of multivariate and/or multiway signals), the analytical sensitivity with multivariate signals, the class-modelling and the process control.