6.2.3 Statistical Acceptance Sampling
統計驗收抽樣
Statistical acceptance sampling is another commonly used statistical tool for validation. The general principle is that the sampling used for validation should provide higher confidence than sampling used during routine production. In validation, larger sample sizes, more replicates, and other such factors are typically used. Commonly used acceptance sampling plans for validation to ensure that a high percentage of individual units (e.g., tablets, vials) are conforming are:
統計驗收抽樣是另一個(gè)驗證中常用的統計工具。其一般原則是驗證取樣應提供更高（相比于日常生產(chǎn)?。┑闹眯哦?。在驗證中，更大的樣本量、更多的重復次數或其他類(lèi)似的因素常被使用。在驗證中為確保單項（如：藥片、小瓶）具有較高符合要求的百分率常用的抽樣方法有：
· Single sampling for attributes data
屬性數據的一次抽樣法
· Double sampling for attributes data
屬性數據的二次抽樣法
· Variables sampling for quantitative data
數據的變量抽樣法
Samples should be representative of the entire population being sampled. Random, stratified, and periodic/systematic sampling are the most commonly used approaches. Targeted sampling to include suspected worst-case locations within the batch or process may be used when appropriate. For example, samples from the very beginning and end of the batch may be selected to assure that these potential trouble spots are included, while the rest of the required samples are randomly selected from throughout the batch.
樣品應具有代表性。隨機抽樣、分層法和定期/系統抽樣是常用的方法。在適當的時(shí)候可以采用針對性的抽樣方法（包含批內或工藝中疑似最壞情況的位置）。例如：對一批產(chǎn)品在最開(kāi)始和結束時(shí)進(jìn)行取樣以確保潛在的故障點(diǎn)都被包含其中，余下所需的樣品可以從整個(gè)批次中隨機選擇。
Reaching at least 90% confidence at the end of PPQ is desirable when using statistical acceptance sampling for validation with little prior confidence. This means that the combined information from the PPQ runs shows that there is at least 90% confidence that the validation performance level has been met; 90% confidence is recommended as the minimum because it is the traditional confidence associated with detecting unacceptable quality levels (called the Rejection Quality Level [RQL], Lot Tolerance Percent Defective [LTPD], or Limiting Quality [LQ]) (46). Note that this use of the term ”confidence” is different than the traditional 95% confidence of acceptance associated with the Acceptance Quality Limit (AQL) in routine lot acceptance sampling. The AQL relates to the Type I error of incorrectly rejecting an acceptable lot, while the 90% minimum confidence recommended here refers to the Type II error of incorrectly accepting an unacceptable process.
驗證時(shí)使用具有較小的置信度的用統計驗收抽樣后，在 PPQ 結束階段，其置信度至少達到 90%。這表明驗證性能水平符合規定的置信度至少有 90%。 90%置信度是建議的最小值，因為它是傳統檢測不可接受質(zhì)量水平的置信度（常稱(chēng)為“不良產(chǎn)品率” [RQL]，“批內允許次品率” [LTPD],或“極限品質(zhì)” [LQ]）。需要注意的是這里的術(shù)語(yǔ)“置信度”不同于傳統日常取樣中與接受質(zhì)量限度[AQL 相關(guān)的]95%“置信度”。 接受質(zhì)量限[AQL]中的“置信度”表示錯誤的拒絕了原本合格的批次（第 I 類(lèi)錯誤），而這里所謂的 90%最小“置信度”表示錯誤的接受了原本不合格的工藝（第 II 錯誤）。
Single sampling for attributes is the simplest type of sampling. For example, a sampling plan of n=388 units, accept on 1 nonconformance, reject on 2, would detect a 1% nonconformance rate with 90% confidence. The statistical operating characteristic curve for this sampling plan is shown in Figure 6.2.3-1.
Figure 6.2.3-1 Example of an Operating Characteristic Curve
Double sampling plans for attributes may take a second set of samples depending on the results of the first set. For example, the double sampling plan n1=250, a1=0, r1=2; n2=250, a2=1, r2=2 will also detect a 1% nonconformance rate with 90% confidence. The values n1 and n2 are the stage 1 and stage 2 sample sizes; a1 and a2 are the accept numbers; r1 and r2 are the reject numbers. If a1 =0 nonconformances are found in the first set of n1=250 samples, the sampling plan is passed. If exactly 1 nonconformance is found in the first sample of n1=250 units, an additional n2=250 units are sampled.If the total number of nonconformances found in the combined 500 samples is no more than a2=1, the sampling plan is passed. If the total number of nonconformances found in the combined 500 samples is r2=2 or
greater, the sampling plan is failed. One advantage of double sampling plans is that they often have lower false reject rates; i.e., good processes will not fail the sampling plan as often.
用于屬性測定的二次抽樣計劃可能需要取第二組樣品，這取決于第一組樣品的結果。例如：這樣一個(gè)檢測 1%的不合格率抽樣方案(90%的置信度)n1=250, a1=0, r1=2; n2=250, a2=1, r2=2。其中(n1、 n2)，（ a1、 a2），（ r1、 r2）分別是第（ 1、 2）次抽樣的樣本量，接受數，拒收數。如果第 1 次抽樣時(shí) a1=0 且不合格數被抽到了。如果第 1 次抽樣時(shí)不合格數恰好為 1，就需要進(jìn)行第2 次抽樣。如果 2 次抽樣的不合格總數不超過(guò) a2=1，則取樣計劃通過(guò)。如果 2 次抽樣的不合格總數≥ 2，則取樣計劃失敗。二次抽樣的優(yōu)點(diǎn)是具有較低的假的拒絕率；也就是說(shuō)好的工藝通常不會(huì )使抽樣計劃失效。
Several types of variables sampling plans may be used for validation, one of the most common being the normal tolerance interval. For example, one normal tolerance interval sampling plan for twosided specifications is n=30, k=3.17. If the average ± 3.17 standard deviation is contained within the specification limits, the sampling plan is passed. This plan also provides 90% confidence in detecting a 1% nonconformance rate. Variables sampling plans assume the data are normally distributed, and this assumption should be confirmed with a suitable normality test. An advantage of variables sampling plans is that they often are able to use much smaller sample sizes than attributes plans to provide the same confidence.
許多變量抽樣法也可用于驗證中，其中最常用的是正態(tài)容許區間。例如：對于一個(gè)具有兩側規格的正態(tài)容許區間（ n=30,k=3.17），如果平均值±3.17 倍標準差包含在控制限內，則取樣方案通過(guò)。該方案能檢出 1%的不合格率（具有 90%的置信水平）。由于假設抽樣方案中所有數據均符合正態(tài)分布，因此應對該假設進(jìn)行正規檢驗。變量抽樣法的優(yōu)點(diǎn)是在相同置信水平的條件下需要的樣本量更?。ㄏ啾扔趯傩猿闃臃ǎ?。
Example: The validation will show with 90% confidence that the process averages ≤0.1% leaking containers after simulated shipping. This requires an attributes sampling plan of n=2300, accept=0, reject=1. Three lots will be used for the Stage 2 PPQ, so n = 2300/3 = 767 containers per lot will be inspected for leakage after simulated shipping. If no leakers are found in the combined n=2300 samples, the sampling plan is passed.
例如：要驗證在模擬運輸后容器泄漏率的平均值≤0.1%（90%的置信度）。屬性取樣方法需要n=2300,接受數=0, 拒收數=1。而工藝驗證的第二階段(PPQ) 需要進(jìn)行 3 批 ，所以每一批都需要在模擬運輸后對n = 2300/3 = 767 個(gè)容器進(jìn)行泄漏檢測。如果在所有2300 個(gè)樣品中均未發(fā)現泄漏的，取樣方案通過(guò)。
ANSI/ASQ Z1.4 “Sampling Procedures and Tables for Inspection by Attributes” and ANSI/ASQ Z1.9 “Sampling Procedures and Tables for Inspection by Variables” are commonly used sampling plans for routine production (47,48). They should be used with care for validation, since they may not provide a high enough level of confidence. For example, one Z1.4 tightened sampling plan for AQL 0.4% is n=315, a=2, r=3. If a validation lot has 2 nonconforming units in a sample of n=315, the validation lot would pass the sampling plan. (However, note that 2/315 = 0.63% is substantially larger than the AQL of 0.4%.) Finding 0.63% nonconforming units in a sample does not provide high confidence that the process is ≤0.4% nonconforming, if that was the goal of the PPQ. If Z1.4 and Z1.9 are used for validation, the Operating Characteristic curves in the standards should be consulted to verify that the desired confidence is achieved. ANSI/ASQ Z1.4 “Sampling Procedures and Tables for Inspection by Attributes” 和 ANSI/ASQ Z1.9 “Sampling Procedures and Tables for Inspection by Variables”（ 47,48）。
在驗證使用這些方法時(shí)應謹慎，因為他們無(wú)法提供更高的置信水平。例如：依據 Z1.4 中的加嚴取樣法(AQL=0.4%,n=315,a=2,r=3).如果一個(gè)驗證批次的樣本中（樣本量 n=315）不合格數是 2，該批驗證將通過(guò)。（但是，注意到 2/315=0.63%大于 AQL=0.4%.）。在樣品中發(fā)現 0.63%的不合格率并不能給出較高的置信度說(shuō)明工藝符合要求（如果該工藝驗證的目標是不合格率≤0.4%）。如果將 Z1.4 和 Z1.9中的方法應用于驗證中，那么應證明其 OC 曲線(xiàn)可以達到預期置信度。
Not all sampling plans used to make accept/reject decisions are for percent nonconforming units. For example, the USP test for content uniformity (of dosage units) is specified in terms of a two-stage sampling plan given in USP. In this case, validation sampling should provide confidence that the USP test can be passed with high confidence (49).
不是所有用于做接收/拒絕決策的取樣方法都適合不合格百分數。例如：對于含量均勻度（單位劑量）USP 明確規定要采用二層抽樣法。在這種情況下，驗證取樣應提供置信度以確保該測試能以較高的置信度通過(guò)。
Example: The sampling plan will show with 95% confidence that the routine USP content uniformity (of dosage units) test requirements can be met.
例如：抽樣方法(具有95％的置信度)表明常規USP 含量均勻度（劑量單位）的測試可以滿(mǎn)足要求。
聯(lián)系客服
