China
Africa
Explore chapters and articles related to this topic
Manufacturing and Process Control Issues
Published in Laszlo Endrenyi, Paul Jules Declerck, Shein-Chung Chow, Biosimilar Drug Product Development, 2017
Alan Fauconnier, Lyudmil Antonov
This is the two-sided equivalence hypothesis. There are two one-sided variants of this hypothesis. In the noninferiority testing, one attempts to establish that mean difference and CI do not fall below the lower EL, while in the nonsuperiority testing, those should not be above the higher EL. It makes sense, for example, to test impurities with a one-sided nonsuperiority test rather than with a two-sided equivalence test. Like other tests, the equivalence testing starts with a null hypothesis (which for the two-sided test is that reference and test samples are inequivalent) and then asks if one has enough evidence (data) to reject that null hypothesis, and, in effect, to prove the alternative hypothesis that the samples are equal for the particular quality attribute (QA) tested. It is said that one makes a Type I error when one incorrectly rejects a valid null hypothesis/accepts a false alternative hypothesis. The probability for a Type I error is the significance level of the test (α-level, p-value). The significance level of the equivalence test is usually .10 (90% CI), while for the one-sided tests (noninferiority and nonsuperiority) the significance level is .05 (95% CI). Type II error is an incorrect acceptance of a false null hypothesis/rejection of a valid alternative hypothesis. With regard to equivalence testing, α is the probability for concluding equivalence when reference and test are inequivalent and β is the probability for concluding inequivalence when reference and test are equivalent. From the aspect of risk, α is the consumer’s risk and β is the producer’s risk. Power is the reverse of β (power = 1 – β) and corresponds to the probability for correctly concluding equivalence when reference and test are equivalent (Table 5.3).
Bayesian evaluation of system structure for reliability assessment
Published in Quality Engineering, 2019
An alternative to the standard hypothesis test proposed is the equivalence test (Wellek, 2010). Instead of assuming the system reliability and the component aggregated reliability are the same and seeking statistically significant evidence for contradicting the null hypothesis, the equivalence test assumes the two ways of calculating the reliabilities are different. The test then focuses on if the difference is large enough to be practically meaningful and worthwhile to treat as different. In our case, if the discrepancy between the system and component data are not sufficiently large to be of practical importance in terms of assessing the system reliability, then we can comfortably use the cheaper and easier method to predict the system reliability based on component test data. Stevens et al. (2017) propose the probability of agreement (PoA) as an easier alternative to the equivalence test for comparing measurement systems. Stevens and Anderson-Cook (2017a, 2017b) extend the PoA to assess the similarity between reliability curves and surfaces. Stevens et al. (2018) develop the Bayesian PoA for measuring the predictive probability of two surfaces being practically equivalent with the differences within a specified threshold value.
Discussion of “Scaling-up process characterization”
Published in Quality Engineering, 2018
I welcome the introduction of practical difference and equivalence tests to determine which changes are practically relevant. Equivalence tests are common in the biotech industry to demonstrate, for example, that two products or processes are equivalent. John argues that “it is reasonable to choose a practical difference that is 10% of 6.” Although this is a good starting point, we should remember that the practical threshold should be based on scientific and engineering knowledge of what constitutes a difference that it is practically irrelevant. For the equivalence tests three outcomes are listed “practically different”, “equivalent”, or “inconclusive” but it is not clear how an “inconclusive” is defined. An equivalence test is still a significance test, two one-sided t-tests, and therefore we either do not reject the null hypothesis, or we favor the alternative hypothesis of equivalence. In other words, the test has two outcomes, not equivalent, what John calls “practically different”, or equivalent.