초록
선형 캘리브레이션 실험계획 문제에 대하여, 베이지안 의사결정론을 이용하여 평균제곱오차손실을 최소화한 Kim(1988, 1993)의 실험계획과 관련 문헌의 결과인 몇 가지 최적계획을 비교한다. 비교대상 실험계획으로서 고전적 추정량의 점근분산을 최소화하는 Buonaccorsi(1986)의 최적계획, 회귀분석 모형에서 $ M(x) = \sum x_i x_i '$의 함수를 최대화 또는 최소화하는 D-optimal 또는 A-optimal 계획, Hunter and Lamboy(1981)가 베이지안 추정량의 특성을 설명하기 위하여 그 논문에서 예로 들었던 실험계획을 고려한다. 서로 다른 기준에 의한 최적계획을 비교하기 위해서 우선 기대사후분산을 계산하여 비교하고 몇가지 사전분포에 대하여 몬테칼로 시뮬레이션을 통한 평균분산과 HPD 구간의 크기를 비교한다.
We consider a linear calibration problem, $y_i = $$\alpha + \beta (x_i - x_0) + \epsilon_i$, $i=1, 2, {\cdot}{\cdot},n$ $y_f = \alpha + \beta (x_f - x_0) + \epsilon, $ where we observe $(x_i, y_i)$'s for the controlled calibration experiments and later we make inference about $x_f$ from a new observation $y_f$. The objective of the calibration design problem is to find the optimal design $x = (x_i, \cdots, x_n$ that gives the best estimates for $x_f$. We compare Kim(1989)'s Bayesian design which minimizes the expected value of the posterior variance of $x_f$ and some optimal designs from literature. Kim suggested the Bayesian optimal design based on the analysis of the characteristics of the expected loss function and numerical must be equal to the prior mean and that the sum of squares be as large as possible. The designs to be compared are (1) Buonaccorsi(1986)'s AV optimal design that minimizes the average asymptotic variance of the classical estimators, (2) D-optimal and A-optimal design for the linear regression model that optimize some functions of $M(x) = \sum x_i x_i'$, and (3) Hunter & Lamboy (1981)'s reference design from their paper. In order to compare the designs which are optimal in some sense, we consider two criteria. First, we compare them by the expected posterior variance criterion and secondly, we perform the Monte Carlo simulation to obtain the HPD intervals and compare the lengths of them. If the prior mean of $x_f$ is at the center of the finite design interval, then the Bayesian, AV optimal, D-optimal and A-optimal designs are indentical and they are equally weighted end-point design. However if the prior mean is not at the center, then they are not expected to be identical.In this case, we demonstrate that the almost Bayesian-optimal design was slightly better than the approximate AV optimal design. We also investigate the effects of the prior variance of the parameters and solution for the case when the number of experiments is odd.