There exist multiple regression applications in engineering, industry and medicine where the outcomes are not conditionally independent given the covariates, but where instead the covariates follow an adaptive experimental design in which the next measurement depends on the previous observations, introducing dependence. Such designs are commonly employed for example for choosing test values when estimating the sensitivity of a material under physical stimulus. In addition to estimating the regression parameters, we are also interested in tasks such as hypothesis testing and constructing confidence intervals, both of which rely on asymptotic normality of the maximum likelihood estimator. When adaptive designs are employed, however, the large-sample theory of the maximum likelihood estimator is more involved than in the standard regression setting, where the outcomes are assumed conditionally independent given the covariates. Hence, asymptotic normality must be verified explicitly case by case. For some classic adaptive designs like the Bruceton up-and-down designs, this issue has been resolved. However, general results for adaptive designs relying on minimal assumptions are to a large extent lacking. In this paper we establish a general large-sample theory for a wide selection of adaptive designs. Motivated by the theory, we propose a new Markovian version of the Langlie design and verify asymptotic normality for this proposal. Our simulations indicate that the Markovian design is more stable and yields better confidence intervals than the original Langlie design.
翻译:暂无翻译