醫學儀器測量方法比較一致性試驗之樣本數模擬研究

Sample Size Simulation for Measurement Method Comparison Studies

黃信明(Hung Sing-Ming)

指導教授 ： 林建甫

國立臺北大學/商學院/統計學系/碩士(2015年)

摘要

統計評估一致性的方法廣泛地使用在檢測新發展的醫學儀器的測量比較上,
這些研究常稱為測量方法比較 (measurement method comparison).
例如, 發展出無輻射線的遠紅外線儀器測量骨質疏鬆取代現有具輻射線的 DXA 儀器,
需進行臨床試驗以確定這二種醫學測量儀器的一致性 (agreement).
檢定一致性的方法常用的是Bland(1986)所提出的預測區間進行判定.
但 Bland 與 Altman 並未提出樣本數的考量,
僅建議樣本數為 100-200. Lin(2002)提出一些樣本數近似計算方法,
但計算過程複雜, 且其結果與 Bland 與 Altman 相差許多.

DXA為放射性測量,
在醫學倫理的考量上, 發展臨床試驗過程中有其額外的限制.
研究前尋找出最佳的本樣本數是重要的考量,
然而, 測量方法比較一致性會受到分配的位置參數與尺度參數,
難以得到精確的樣本數,
因此本研究目的使用統計模擬,
分析最佳化的樣本數,
並與 Lin(2002)提出的樣本數計算方法進行比較.

研究結果顯示,
在合理的型一與型二誤差下, 參數設定與一致性臨界值之下,
使用統計模擬出最佳化樣本數,
在最佳化的樣本數下,
雖然雙邊型一誤差遠比事先給定的小,
但在檢定力都是接近事先給定型二誤差,
最佳化的樣本數表現, 比 Lin 樣本數計算較佳,
但研究結果同時發現不同醫學測量儀器的一致性對參數與一致性臨界值的變動非常敏感.

模擬與分析最佳化的樣本數可提供在不同試驗的情境下比較樣本數大小, 探討試驗的可行性,
建議在進行醫學測量方法比較一致性的臨床試驗之前,
同時使用多種樣本數計算分析, 尋找出合理的樣本數.

關鍵字: 一致性, 測量方法比較.

關鍵字

測量方法比較 ； 一致性

並列摘要

Measurement method comparison is commonly used to assess agreement between two measurements
both in clinical medicine and statistics.
For example,
if we want to develop a new non-radiation far-infrared (FIR) instrument to replace
the standard radiation dual-energy x-ray absorptiometry (DXA) for osteoporosis,
we need to conduct clinical trials
to assess agreement between these two methods of clinical measurements.
The most common method to assess agreement between methods in clinical practice
is using ``prediction interval'' proposed by Bland(1986).
However, Bland and Altman did not provide the sample size calculation formula.
Bland only recommended that sample size should be around 100-200 in clinical practices.
citet{lin2002statistical} proposed several formal tests for measurement method comparison
and also provided sample size calculation formulas for assessing agreements.
However, the calculations are complicated and often require repeated measurements.

Subjects would be exposed to radiation when using DXA to measure osteoporosis.
To conduct a clinical trial to assess agreement between two methods, DXA and new FIR,
we need to find the optimal sample size to decrease the chance
for subjects being exposed radiation.
The agreement between two methods actually depend both on
the location and scale parameters for measurements.
It is unethical to allow subjects being exposed to radiation frequently.
And it would be also unethical to take repeated measurements
to calculate the sample size in advance to conduct a formal test
based on Lin's methods.Lin(2002)
Therefore, the purpose of this study is to use statistical simulation
to analyze the optimal sample size for assessing agreement based on prediction interval in clinical trials.
We also compare our simulated sample size to those methods proposed by Lin.Lin(2002)
We found that using statistical simulation allow us easily
to find an optimize sample size given pre-specified type one, type two error rates, location and scale parameters.
Based on the simulated optimize sample size,
the observed type one error rate is smaller than the pre-specified type one error.
And the observed power is close to the pre-specified power.
The simulated optimize sample size usually had better performances than those of Lin's.
We also found that the simulated optimize sample size is very sensitive to
the change of parameters and marginal vales for agreements.

Simulating the optimize sample size
provided essential information to compare different scenarios.
It is critical to conduct a variety of sample size calculations before
the initiating a clinical trials for assessing agreement
between two methods of clinical measurements.

Key words: Agreement, measurement method comparison

並列關鍵字

measurement method comparison ； Agreement

參考文獻

analysis of method comparison studies. The statistician, pages 307–317.

lation coefficient for evaluating agreement among multiple observers. Biometrics

agreement with continuous measurements. Journal of biopharmaceutical

ter and total agreement with replicated readings. Statistics in Medicine,

agreement between two methods of clinical measurement. The lancet,