You are on page 1of 20

Received: 5 January 2018 Revised: 31 March 2018 Accepted: 27 May 2018

DOI: 10.1002/qre.2353

TUTORIAL

Optimized sampling design and rationale for verification


and validation

Abstract appropriate sampling plans for V&V are less well known.
An optimized sampling design that meets customer, design, The development of an adequate sampling plan for V&V
or process requirements, while balancing technology limi- has been challenging and has been frequently cited by
tations, is still a common challenge to engineering commu- Food and Drug Administration (FDA) warning letters or
Form 483 Observations.
nities. This is especially true in the medical device industry.
In V&V, a common parameter to estimate is the reli-
Acceptance sampling plans for manufacturing are widely ability (R), defined as the proportion of a population that
available, but the appropriate sampling plans for verifica- conforms to the specification, eg, the proportion of
tion and validation (V&V) are less well known. This paper in vitro diagnostic tests that report a valid test result.
applies established statistical theory to derive sampling Medical device V&V sets the requirement of protecting
plans appropriate for estimating product reliability during the consumer by demonstrating that R ≥ RL at the appro-
priate level of statistical confidence, where RL is the
V&V, where reliability must exceed an established thresh-
predetermined minimum acceptable quality level. Verifi-
old with an appropriate margin of statistical confidence. cation and validation therefore assumes that the require-
The paper provides insight on how to estimate parameters ments have not been met unless the testing demonstrates
of interest and interpret acceptance criteria. Operating that they are met. A statistical sample of size N is drawn
characteristic curves are used to examine if a design or pro- from the relevant population of interest; eg, N in vitro
cess is capable of producing future product that meets diagnostic test strips are drawn from the product line
and run on a control specimen producing N test results.
design specifications and/or customer requirements in
R is the proportion of valid test results for the product line
terms of confidence and reliability. The methodology is
(the population proportion in statistical terminology). RS
applied to both attribute and variable sampling plans, is the estimate of this proportion based on the results
including examples showing how to achieve a high proba- observed in the statistical sample. It is important to
bility of passing the acceptance criteria. Formulas, sample understand that R has some true but unknown value in
size tables, and operating characteristic curves are provided the population and that RS is an estimate of this parame-
for engineering practitioners to use. The paper aims at pro- ter. The logic of the sampling plan concerns the accuracy
of the estimate RS based on the size of the statistical sam-
viding a practical quantitative approach and a valid statisti-
ple. In statistics, the accuracy of an estimate is often
cal rationale to assess overall product quality during V&V. assessed by a confidence interval. The statistical confi-
dence is given by a confidence level defined here as
KEYWORDS (1 − β), where typical values of β are 0.025, 0.05, and
confidence, reliability, sampling, validation, verification 0.10, corresponding to 97.5%, 95%, and 90% confidence.
The lower 1‐sided (1 − β) confidence limit of the estimate
of R is determined; the V&V passes its acceptance criteria
1 | INTRODUCTION when the lower confidence limit of the estimate of R is
equal to or exceeds RL. Another way of stating this
Development of an adequate sampling plan with a valid criteria is that the hypothesis R < RL must be rejected
statistical rationale that balances the customer or design with probability (1 − β), so that we may conclude R ≥ RL
requirements and technology limitations is a pivotal ele- with the stated level of statistical confidence. In this
ment in verification and validation (V&V), especially for paper, we speak of 1‐sided lower confidence limits
the medical device industry. Acceptance sampling plans because the upper confidence limit of the estimate of R
for manufacturing are widely available, but the is not relevant.

Qual Reliab Engng Int. 2019;35:483–502. wileyonlinelibrary.com/journal/qre © 2018 John Wiley & Sons, Ltd. 483
484 TUTORIAL

Acceptance sampling is often applied to manufactur- Often, the reliability is defined as the proportion of
ing, wherein the Operating Characteristic (OC) curve is the population of a continuous variable that must be
used to accept or reject lots with the trade‐off between between a lower and upper specification to be considered
Producer's Risk (α) and Consumer's Risk (β). The OC curve acceptable. In such cases, the reliability is related to the
plots the probability of acceptance versus the population mean and the standard deviation of the distribution of
proportion defective, which is 1 − R in our notation. The the continuous variable. Typically, when variable sam-
probability of acceptance associated with the Producer's pling plans are considered, the estimate of the reliability
Risk is (1 − α), and the corresponding population propor- (in the sample) is not explicitly calculated, as it is more
tion defective is the Acceptable Quality Limit (AQL). The common to report the sample mean and the sample stan-
probability of acceptance associated with the Consumer's dard deviation, together with the tolerance interval and
Risk is β, and the corresponding population proportion the pass/fail disposition. This is because the statistical
defective is the Rejectable Quality Level (RQL), also theory of tolerance intervals is well established
known as the Lot Tolerance Percent Defective (LTPD). (Krishnamoorthy and Mathew,4 Wilks,5,6 ISO 16269‐6,7
The OC curve is readily applied to the problem of V&V Owen,8 and NIST/SEMATECH9); but the statistical the-
by taking the confidence level as (1 − β), the population ory of estimating the reliability (given the sample mean
proportion defective as (1 − R), and the Rejectable Quality and the sample standard deviation) is less familiar to
Level (RQL) as (1 − RL). In the same fashion, we may label engineers, although it is on equally strong footing.2,3,10
the Acceptable Quality Limit (AQL) as (1 − RA), where RA The technical problem arises when applying the
is the value of the reliability in the population such that the familiar 2‐sided tolerance interval to a problem with both
probability of acceptance is (1 − α). In practice, α should be a lower and upper specification limit (LSL and USL). If
suitably small for the success of the V&V, but the sample one requires, for example, as in Ferryanto,1 that both
size may be constrained given the practical limitations of sides of the tolerance interval must be within the interval
conducting the V&V. Generally speaking, when RA is large (LSL and USL), then a precise statement regarding the
compared with RL, the sample size can be relatively small, confidence level can no longer be achieved. In this case,
but when RA approaches RL, the sample size must be large. the V&V may fail against a confidence level that exceeds
Minimal sampling plans are often inappropriate because the required confidence level by an unspecified amount.
of the low probability of acceptance when the reliability The rigorous method to overcome this deficiency is to
in the population is close to the reliability requirement. estimate the sample mean and the sample standard devi-
Some discussion of notation is unfortunately required. ation and to use these to estimate the reliability in the
Here the confidence levels are labeled by the Greek letter sample (RS). The acceptance criterion is then given in
β, because this is the conventional notation for terms of a threshold that RS must meet (or exceed), the
Consumer's Risk. In conventional statistical notation, con- threshold depending on both RL (the predefined quality
fidence levels are labeled by α instead of β, which is just an level that meets the requirement), and the confidence
exchange of Greek letters. The probability of passing the level (1 − β). This makes the variable sampling plan sim-
acceptance criteria is traditionally labeled as (1 − α) on ilar to the attribute sampling plan, in that both plans esti-
the OC curve, where α is the Producer's Risk. In conven- mate the same underlying parameter (the reliability). It
tional statistical notation, the probability of passing would also allows the OC curve of each plan to be compared
typically be called “statistical power” and it would be directly in the same type of plot. In the attribute sampling
labeled as (1 − β), again just an exchange of Greek letters. plan, RS is calculated by counting the data points that fall
A statistician would consider R < RL to be the “null within the interval (LSL and USL), whereas in the vari-
hypothesis” and would calculate the power of the sample able sampling plan, RS is calculated from the sample
size under the “alternative hypothesis” that R = RA. mean and the sample standard deviation. The variable
The statistical theories to develop a sampling plan to sampling plan is more efficient than the attribute plan
estimate reliability with confidence are well established in terms of the sample size, requiring smaller N to
and have recently been reviewed in the context of V&V achieve the same level of confidence.
for medical devices.1 However, this reference did not In addition to using appropriate statistical techniques,
cover the probability of passing the acceptance criteria the statistical rationale for the sampling plan requirement
of the V&V protocol. As compared with Ferryanto,1 this should be based on voice of customer (VOC) and risk
paper covers the probability of acceptance as a function management analyses in accordance with ISO 1497111
of the true (but unknown) product reliability and also together with failure modes and effects analysis (IEC
includes an improved method for variable sampling as 6081212) and fault tree analysis (IEC 6102513), where the
indicated in Lieberman and Resnikoff2 and Bruhn‐Suhr appropriate confidence and reliability levels are deter-
and Krumbholz.3 mined on the basis of the risks. Some discussion of the
TUTORIAL 485

appropriate levels for confidence and reliability is pro- and solved with double‐precision floating point using
vided in Durivage.14 MATLAB version R2016a, unless otherwise noted.
This paper has 5 key elements. The first is to delineate The sample sizes at designated confidence and reli-
the statistical principles of estimating reliability with con- ability are listed in Table 1. For example, the required
fidence and its application to V&V. The second is to pres- minimum sample size is n = 59 when allowing zero fail-
ent how to design optimized sampling plans using such ures to demonstrate 95% reliability (RL = 0.95) with 95%
statistical techniques. The third is to consider the proba- confidence (β = 0.05).
bility of passing the acceptance criteria given some When C = 0 (zero failures allowed), the above
assumption about the true (but unknown) level of R in equation is reduced to
the population. The fourth is to estimate the reliability
with confidence for both the attribute and variable sam- lnβ
n≥ : (3)
pling plans, in a comparable methodology allowing direct lnRL
comparison of OC curves. Finally, examples of typical
Equation 3 provides a convenient calculation of the
applications of the methodology in the medical device
minimum sample size in circumstances where zero fail-
industry are presented in the context of V&V.
ures are allowed in the acceptance criteria of the V&V.
For example, to demonstrate 90% reliability with 95%
2 | CONF IDENCE AND confidence, the minimum sample size is calculated as
RELIABILITY METHOD BASED
SAMPLING PLANS lnð1−0:95Þ= ln0:9≥28:4≈29

ðrounding up to the nearest integerÞ:


2.1 | Attribute sampling plan
Each measurement in an attribute sampling plan has a
dichotomous result, eg, PASS or FAIL. An example of 2.2 | Variable sampling plan
the application is for conformance or nonconformance
of product characteristics. For an independent random The variable sampling plan follows the standard statistical
sample selected from an infinite population, the sample methodology as indicated in Lieberman and Resnikoff,2
size n can be determined from the following equation Bruhn‐Suhr and Krumbholz,3 and ANSI/ASQ Z1.9‐2003.10
based on the binomial distribution, Montgomery15: The technique is often perceived as complex because of
the special mathematical functions used in the calculation.
C n! Nonetheless, it is relatively simple to explain. The problem
β≥ ∑ ð1−RL Þi RLn−i (1) statement relies on 2 additional input parameters, the LSL
i¼0 i! ð n−i Þ!
and the USL. Performance is considered acceptable when
the continuous variable of interest has a high probability
C n! of falling within the interval (LSL and USL). The problem
Probability of Acceptance ¼ ∑ ð1−RÞi Rn−i (2)
i¼0 i!ðn−iÞ!
statement can be adapted to the case where either the LSL
goes to negative infinity or the USL goes to positive infinity,
where (1 − β) is the required confidence, RL is the speci- thereby turning the 2‐sided problem statement into a 1‐
fied or required minimum reliability, R is the true (but sided problem statement. The goal is to estimate the propor-
unknown) reliability in the population, C is the number tion of the population outside of the specification limits, ie,
of allowable failures, and n is the sample size. The case the proportion that does not conform to the specification.
of sampling from a finite population, as may be appropri- The maximum allowable estimate of the proportion defec-
ate for manufacturing lot release when the lot size is rel- tive p* is used in the acceptance criteria shown below:
atively small, is based on the more complicated
hypergeometric distribution, which is not relevant here pL þ pU ≤p* (4)
because the design or process (capable of making all the
where pL and pU are the estimated probabilities of noncon-
future lots in manufacturing) is the subject of the V&V.
formity below the LSL or above the USL, respectively. The
It is worth noting that Equation 1 follows from Equation 2
probabilities are calculated as follows:
by setting R = RL and setting the probability of accep-
tance ≤ β. Equation 1 is solved iteratively to find the min- pL ¼ ∅ð−Z L Þ; pU ¼ ∅ð−Z U Þ; and p* ¼ ∅ð−k 1 Þ
imum sample size n for the maximum allowable failures,
or the inverse binomial distribution is used for the same where ϕ is the cumulative density function of the standard
purpose. In this paper, all the equations are programmed normal distribution.
486 TUTORIAL

TABLE 1 Attribute sampling plans for required minimum reliability RL with confidence (1 − β)

Confidence Level 97.5% Confidence Level 95% Confidence Level 90%


Reliability RL Reliability RL Reliability RL
C 0.99 0.95 0.9 0.85 0.8 0.99 0.95 0.9 0.85 0.8 0.99 0.95 0.9 0.85 0.8

0 368 72 36 23 17 299 59 29 19 14 230 45 22 15 11


1 555 110 54 35 26 473 93 46 30 22 388 77 38 25 18
2 720 142 70 46 34 628 124 61 40 30 531 105 52 34 25
3 874 173 85 56 41 773 153 76 50 37 667 132 65 43 32
4 1022 202 100 66 48 913 181 89 59 44 798 158 78 52 38
5 1164 230 114 75 55 1049 208 103 68 50 926 184 91 60 45
6 1303 258 127 84 62 1182 234 116 76 57 1051 209 104 68 51
7 1439 285 141 93 69 1312 260 129 85 63 1175 234 116 77 57
8 1573 312 154 102 75 1441 286 142 93 69 1297 258 128 85 63
9 1705 338 167 110 82 1568 311 154 102 76 1418 282 140 93 69
10 1835 364 180 119 88 1693 336 167 110 82 1538 306 152 100 75
11 1964 390 193 127 94 1818 361 179 118 88 1658 330 164 108 81
12 2092 415 206 136 101 1941 386 191 126 94 1776 353 175 116 86
13 2219 440 218 144 107 2064 410 203 134 100 1893 377 187 124 92
14 2345 465 231 152 113 2185 434 215 142 106 2010 400 199 132 98
15 2470 490 243 160 119 2306 458 227 150 112 2127 423 210 139 104

The quantities ZL and ZU are determined by the sam- statistical confidence level (1 − β), and the sample size
ple mean X and the sample standard deviation (S) of the (n) are included in the expression for k1.
continuous variable of interest, as shown in the formulas To complete the variable sampling plan, Bruhn‐Suhr
below. (The “sample mean” and the “sample standard and Krumbholz3 provide a formula for the probability of
deviation” refer to estimates of the mean and standard acceptance as shown below.
deviation calculated from the data collected in a statistical  pffiffiffi 
sample selected from the population of interest.) Probability of Acceptance ¼ 1− F k1 n; n−1; δ2 (6)

Here F is the noncentral t cumulative distribution



X−LSL
ZL ¼ ; with n − 1 degrees of freedom and noncentrality param-
S pffiffiffi
eter δ2 ¼ zp2 n, where zp2 is the critical value of the nor-
mal distribution with probability p2 = R (the population

USL−X true but unknown reliability) and where k1 is given by
ZU ¼ :
S Equation 5. The factor k1 is equal to the exact 1‐sided sta-
tistical tolerance limit factor k1.9
The quantity k1 is computed using the inverse cumu- Tables 2–4 list commonly used k1 in medical device
lative distribution function for the noncentral t distribu- V&V. Simplifying the problem statement by setting the
tion, as shown below. USL to infinity allows one to see the exact connection
to the 1‐sided tolerance interval. In this case, pU is zero
t 1−β ðn−1; δ1 Þ and Equation 4 can be solved explicitly, as follows:
k1 ¼ pffiffiffi (5)
n

X−LSL
≥ k1 (7)
pffiffiffi S
Here δ1 ¼ zp1 n is the noncentrality parameter,
which depends on zp1 , the critical value of the normal dis- However, the 2‐sided problem statement (where both
tribution with probability p1 = RL, and the sample size n. the LSL and the USL are finite) is not exactly the same as
Therefore, all the parameters determining the sampling a tolerance interval and Equation 4 must be solved in units
plan, the minimum acceptable reliability (RL), the of probability. Solving the problem in terms of
TUTORIAL 487

TABLE 2 Exact 1‐sided statistical tolerance limit factor and corresponding minimum acceptable estimate of the reliability (1 − p*) for
variable sampling plan at (1 − β) confidence level of 97.5%a

Confidence Level 97.5%


Reliability RL=0.99 Reliability RL=0.95 Reliability RL=0.9 Reliability RL=0.85 Reliability RL=0.8
N k1 1 − p* k1 1 − p* k1 1 − p* k1 1 − p* k1 1 − p*

20 3.5291 0.9998 2.5760 0.9951 2.0792 0.9812 1.7507 0.9600 1.4948 0.9326
30 3.2331 0.9994 2.3508 0.9907 1.8894 0.9706 1.5835 0.9434 1.3445 0.9106
40 3.0780 0.9990 2.2320 0.9872 1.7886 0.9632 1.4941 0.9325 1.2635 0.8968
50 2.9798 0.9986 2.1564 0.9845 1.7243 0.9577 1.4368 0.9247 1.2114 0.8872
60 2.9109 0.9982 2.1032 0.9823 1.6788 0.9535 1.3962 0.9187 1.1744 0.8799
70 2.8593 0.9979 2.0633 0.9805 1.6446 0.9500 1.3656 0.9140 1.1464 0.8742
75 2.8379 0.9978 2.0467 0.9797 1.6304 0.9485 1.3529 0.9120 1.1348 0.8718
80 2.8189 0.9976 2.0319 0.9790 1.6177 0.9472 1.3415 0.9102 1.1243 0.8696
90 2.7861 0.9974 2.0065 0.9776 1.5959 0.9448 1.3219 0.9069 1.1064 0.8658
100 2.7590 0.9972 1.9854 0.9765 1.5777 0.9427 1.3055 0.9042 1.0913 0.8625
125 2.7072 0.9967 1.9450 0.9742 1.5430 0.9386 1.2742 0.8987 1.0626 0.8561
150 2.6701 0.9963 1.9160 0.9724 1.5179 0.9355 1.2516 0.8947 1.0418 0.8513
175 2.6418 0.9959 1.8939 0.9709 1.4988 0.9331 1.2344 0.8915 1.0258 0.8476
200 2.6194 0.9956 1.8763 0.9697 1.4836 0.9311 1.2206 0.8889 1.0131 0.8445
250 2.5857 0.9952 1.8499 0.9679 1.4607 0.9280 1.1999 0.8849 0.9939 0.8399
300 2.5613 0.9948 1.8308 0.9665 1.4440 0.9257 1.1848 0.8820 0.9800 0.8365
350 2.5426 0.9945 1.8161 0.9654 1.4313 0.9239 1.1732 0.8797 0.9692 0.8338
400 2.5277 0.9943 1.8043 0.9645 1.4211 0.9224 1.1639 0.8778 0.9606 0.8317
450 2.5154 0.9941 1.7947 0.9637 1.4127 0.9212 1.1563 0.8763 0.9535 0.8299
500 2.5051 0.9939 1.7866 0.9630 1.4056 0.9201 1.1498 0.8749 0.9475 0.8284
600 2.4887 0.9936 1.7736 0.9620 1.3943 0.9184 1.1396 0.8728 0.9380 0.8259
700 2.4761 0.9934 1.7636 0.9611 1.3856 0.9171 1.1316 0.8711 0.9306 0.8240
800 2.4660 0.9932 1.7556 0.9605 1.3786 0.9160 1.1253 0.8698 0.9247 0.8225
a
All the data in Tables 2–4 are calculated using 16 digits of precision. The k1 values shown in the table are rounded up at the fourth decimal place to align with
ISO 16269‐67. The 1 − p* values shown in the table are also rounded up at the fourth decimal place to be conservative.

probabilities makes the connection between the variable It is necessary to add some statistical footnotes to the
sampling plan and the attribute sampling plan. The left above methodology for the variable sampling plan. First,
side of Equation 4 is the estimate of the proportion defec- the original methodology2 that is used in ANSI/ASQ
tive (1 − RS) based on the sample mean and the sample Z1.9‐2003 (R2013)10 is not an exact method, as discussed
standard deviation. The value of p* on the right side of in Bruhn‐Suhr and Krumbholz,3 which compares the
Equation 4 is always less than (1 − RL) to allow for the original method and the exact method. Second, we have
required level of confidence given a finite sample size. modified the original methodology following Bruhn‐Suhr
Tables 2–4 list 1 − p* as a function of the sample size for and Krumbholz3 to simplify the interpretation of the
combinations of commonly used confidence levels equations. The original methodology uses a minimum
(1 − β) and minimum required reliability (RL). It is impor- variance unbiased (MVU) estimate of the 3 probabilities
tant to note that p* plays the same role in the variable sam- pL, pU, and p*. The modified methodology used in this
pling plan as C plays in the attribute sampling plan; ie, the paper uses the maximum likelihood (ML) estimate of
maximum allowable estimate of the proportion defective the same 3 probabilities. This replacement of the MVU
in the attribute plan is simply C/N. (For a specific example estimate by the ML estimate is profound in terms of sim-
comparing the attribute sampling plan and the variable plifying the equations, because the probabilities pL and
sampling plan, see the description for Figure 4.) pU are independent of sample size and because the ML
488 TUTORIAL

TABLE 3 Exact 1‐sided statistical tolerance limit factor and corresponding minimum acceptable estimate of the reliability (1 − p*) for
variable sampling plan at (1 − β) confidence level of 95%

Confidence Level 95%


Reliability RL=0.99 Reliability RL=0.95 Reliability RL=0.9 Reliability RL=0.85 Reliability RL=0.8
N k1 1 − p* k1 1 − p* k1 1 − p* k1 1 − p* k1 1 − p*

20 3.2952 0.9996 2.3961 0.9918 1.9260 0.9730 1.6146 0.9468 1.3714 0.9149
30 3.0640 0.9990 2.2199 0.9868 1.7774 0.9623 1.4833 0.9311 1.2531 0.8950
40 2.9410 0.9984 2.1255 0.9833 1.6972 0.9552 1.4121 0.9211 1.1885 0.8827
50 2.8625 0.9979 2.0650 0.9806 1.6456 0.9501 1.3661 0.9141 1.1465 0.8742
60 2.8071 0.9976 2.0222 0.9785 1.6090 0.9462 1.3333 0.9088 1.1165 0.8679
70 2.7654 0.9972 1.9899 0.9767 1.5813 0.9431 1.3084 0.9047 1.0938 0.8630
75 2.7482 0.9971 1.9765 0.9760 1.5697 0.9418 1.2981 0.9029 1.0843 0.8609
80 2.7327 0.9969 1.9645 0.9753 1.5594 0.9406 1.2888 0.9013 1.0758 0.8590
90 2.7061 0.9966 1.9438 0.9741 1.5416 0.9385 1.2728 0.8985 1.0611 0.8557
100 2.6840 0.9964 1.9266 0.9730 1.5268 0.9366 1.2595 0.8961 1.0488 0.8529
125 2.6418 0.9959 1.8937 0.9709 1.4984 0.9330 1.2338 0.8914 1.0252 0.8474
150 2.6114 0.9955 1.8699 0.9693 1.4778 0.9303 1.2153 0.8879 1.0081 0.8433
175 2.5882 0.9952 1.8517 0.9680 1.4621 0.9282 1.2011 0.8852 0.9950 0.8402
200 2.5698 0.9950 1.8373 0.9670 1.4496 0.9265 1.1897 0.8830 0.9845 0.8376
250 2.5421 0.9945 1.8155 0.9653 1.4307 0.9238 1.1726 0.8796 0.9686 0.8337
300 2.5219 0.9942 1.7997 0.9641 1.4170 0.9218 1.1601 0.8770 0.9570 0.8308
350 2.5065 0.9940 1.7875 0.9631 1.4064 0.9202 1.1505 0.8751 0.9481 0.8285
400 2.4941 0.9937 1.7778 0.9623 1.3979 0.9190 1.1428 0.8735 0.9410 0.8267
450 2.4840 0.9936 1.7698 0.9617 1.3910 0.9179 1.1365 0.8722 0.9351 0.8252
500 2.4755 0.9934 1.7631 0.9611 1.3851 0.9170 1.1311 0.8710 0.9301 0.8239
600 2.4619 0.9931 1.7523 0.9602 1.3757 0.9156 1.1226 0.8692 0.9222 0.8218
700 2.4514 0.9929 1.7441 0.9595 1.3685 0.9145 1.1160 0.8678 0.9160 0.8202
800 2.4430 0.9928 1.7374 0.9589 1.3627 0.9136 1.1107 0.8667 0.9111 0.8189

estimate depends on the normal distribution, which is Therefore, the approximate plan using the MVU estimate
more familiar than the beta distribution. Monte Carlo slightly overestimates the probability of acceptance,
simulations conducted as background material for this which is why p* is adjusted downward in the exact plan
manuscript (unpublished) were not able to distinguish to achieve the required level of confidence. This can also
substantially between the probabilities computed via sim- be seen by calculating the probability of acceptance for
ulation of the exact plan and the probabilities computed R = RA = 0.99. The exact plan has a probability of accep-
via simulations of the approximate plan, either in its orig- tance of 0.900 (corresponding to α = 0.10) whereas the
inal form (using the MVU estimates) or in the modified approximate plan has a probability of acceptance of
form, as presented here using the ML estimates of the 0.908. Again, the approximate plan using the MVU esti-
probabilities. For example, the exact plan for RL = 0.94, mate slightly overestimates the probability of acceptance.
β = 0.10, N = 34 has p* = 0.02262, as given by example The same calculations via Monte Carlo simulation were
1 of Bruhn‐Suhr and Krumbholz3 using the MVU esti- also made for the approximate plan using the ML esti-
mate, whereas the approximate plan (also using the mate following the formulas presented in this paper. In
MVU estimate) has p* = 0.02337. The probability of this case, for R = RL, the approximate plan has a probabil-
acceptance calculated by Monte Carlo simulation of the ity of acceptance of 0.091, as compared with the probabil-
approximate plan is 0.106 for R = RL, as compared with ity of acceptance of 0.100 for the exact plan, whereas for
the probability of acceptance of 0.100 for the exact plan. R = RA, the approximate plan has a probability of
TUTORIAL 489

TABLE 4 Exact 1‐sided statistical tolerance limit factor and corresponding minimum acceptable estimate of the reliability (1 − p*) for
variable sampling plan at (1 − β) confidence level of 90%

Confidence Level 90%


Reliability RL=0.99 Reliability RL=0.95 Reliability RL=0.9 Reliability RL=0.85 Reliability RL=0.8
N k1 1 − p* k1 1 − p* k1 1 − p* k1 1 − p* k1 1 − p*

20 3.0516 0.9989 2.2078 0.9864 1.7653 0.9613 1.4711 0.9294 1.2407 0.8927
30 2.8838 0.9981 2.0799 0.9813 1.6571 0.9513 1.3754 0.9155 1.1542 0.8758
40 2.7932 0.9974 2.0103 0.9778 1.5979 0.9450 1.3228 0.9071 1.1063 0.8658
50 2.7349 0.9969 1.9653 0.9754 1.5595 0.9406 1.2884 0.9012 1.0750 0.8588
60 2.6936 0.9965 1.9333 0.9734 1.5321 0.9373 1.2638 0.8969 1.0525 0.8537
70 2.6623 0.9962 1.9091 0.9719 1.5113 0.9347 1.2451 0.8935 1.0353 0.8498
75 2.6493 0.9960 1.8990 0.9713 1.5026 0.9336 1.2373 0.8921 1.0282 0.8481
80 2.6377 0.9959 1.8899 0.9707 1.4948 0.9326 1.2303 0.8907 1.0217 0.8466
90 2.6177 0.9956 1.8743 0.9696 1.4813 0.9308 1.2182 0.8885 1.0106 0.8439
100 2.6010 0.9954 1.8613 0.9687 1.4701 0.9293 1.2081 0.8865 1.0012 0.8417
125 2.5690 0.9949 1.8363 0.9669 1.4485 0.9263 1.1886 0.8827 0.9833 0.8373
150 2.5459 0.9946 1.8182 0.9655 1.4329 0.9241 1.1744 0.8799 0.9702 0.8341
175 2.5282 0.9943 1.8044 0.9645 1.4209 0.9224 1.1636 0.8777 0.9601 0.8315
200 2.5141 0.9941 1.7934 0.9636 1.4113 0.9210 1.1549 0.8760 0.9521 0.8295
250 2.4930 0.9937 1.7767 0.9622 1.3969 0.9188 1.1418 0.8733 0.9399 0.8264
300 2.4775 0.9934 1.7646 0.9612 1.3863 0.9172 1.1322 0.8713 0.9310 0.8241
350 2.4657 0.9932 1.7553 0.9604 1.3782 0.9160 1.1248 0.8697 0.9241 0.8223
400 2.4562 0.9930 1.7478 0.9598 1.3717 0.9150 1.1189 0.8684 0.9186 0.8209
450 2.4484 0.9929 1.7416 0.9593 1.3663 0.9141 1.1140 0.8674 0.9141 0.8197
500 2.4418 0.9927 1.7365 0.9588 1.3618 0.9134 1.1099 0.8665 0.9103 0.8187
600 2.4314 0.9925 1.7282 0.9581 1.3546 0.9123 1.1033 0.8651 0.9041 0.8171
700 2.4233 0.9924 1.7218 0.9575 1.3490 0.9114 1.0982 0.8640 0.8994 0.8158
800 2.4168 0.9922 1.7167 0.9570 1.3445 0.9107 1.0941 0.8631 0.8956 0.8148

acceptance of 0.887, as compared with the exact plan that (RL), and a sample size (N). However, the selection of an
has a probability of acceptance of 0.900. Therefore, the appropriate sample size must consider the true (but
approximate plan using the ML estimate in this paper unknown) product reliability.
slightly underestimates the probability of acceptance The probability of acceptance is given by Equation 2 for
and, therefore, is conservative. the attribute sampling plan, or by Equation 6, for the var-
iable sampling plan. The OC curve can be plotted from
Equation 2 or 6. The OC curve is a plot of the probability
3 | OPTIMIZED S AMPLI N G DE SI GN of acceptance on the y‐axis versus the population propor-
BALANCING CONSUMER tion defective (1 − R) on the x‐axis for a given sample size
PROTECTION AN D P R OD U CE R 'S N at the selected levels for the minimum acceptable reli-
T E C H N O L O G Y LI M I TA T I O N S ability RL and statistical confidence (1 − β). The value of
the reliability in the population is a parameter that is con-
An optimized sampling plan captures both the VOCs and sidered unknown during the V&V; hence, it is sometimes
technology limitations commensurate of risks. Therefore, called the “true but unknown” parameter, which we have
an optimized sampling plan needs to be (a) adequate and denoted as R. Given data from previous studies (design cal-
(b) practical. In the attribute and variable sampling plans culation, dry runs, or historical data), it should be possible
shown in Tables 1–4, each sampling plan is specified by a to make an initial estimate for the assumed value of this
confidence level (1 − β), a minimum acceptable reliability parameter when choosing the sample size.
490 TUTORIAL

The OC curves for the attribute sampling plan of 95%


reliability and 90% confidence requirement are shown in
Figure 1. Each of the OC curves is labeled by the sample
size, N, and the corresponding maximum allowable fail-
ures, C.
In Figure 1, the OC curves are shown for 95% reliabil-
ity (RL = 0.95) and 90% confidence (β = 0.10) for sample
sizes N = 45, 77, 105, 158, 209, 306, 400, 492, and 807 with
corresponding allowable failures C = 0, 1, 2, 4, 6, 10, 14,
18, and 32. For the attribute plan, all of the OC curves
have a probability of acceptance that is less than or equal
to β when R = RL. In the case of Figure 1, when the prod-
uct true but unknown reliability is equal to the required
reliability of 95%, or the proportion defective is 5%, the
probability of acceptance is 10% for all the OC curves cor-
responding to 95% reliability at 90% confidence.
In the case of Figure 1, the minimum sample size N FIGURE 2 Operating characteristic (OC) curves for attribute
was selected for each level of C, following the convention sampling for 95% reliability (RL) at 90% confidence (1 − β) when
sample sizes are not selected from Table 2 [Colour figure can be
in Table 1. In this case, because the OC curves are labeled
viewed at wileyonlinelibrary.com]
by increasing levels of C, the probability of acceptance
increases with the increase of sample size N. As a more
general example, consider the OC curves for the sample (release 15, 2017, NCSS, LLC). Figure 3 shows that when
size selections shown in Figure 2, where different values the sample size is increased from 76 to 77, the probability
of N may share the same value of C. This is the case for of acceptance jumps from 68.3% to 94.3%. The power
the first 2 curves in Figure 2, where the probability of decreases as N increases (above the minimum required
acceptance decreases when going from N = 50 to N) until C is allowed to increase by one, thereby making
N = 75, because both curves allow the same maximum a saw tooth–shaped power curve. To summarize, Figure 1
number of failures (C = 0). shows the minimum sample size for each level of C (the
This behavior of the power versus sample size curve maximum allowable number of failures), where the prob-
for estimating a binomial proportion is well known16 ability of acceptance is at a local maximum of the saw
and is demonstrated in Figure 3, assuming a true (popu- tooth power curve. For this reason, it is conventional to
lation) reliability of 99.5%, as calculated in PASS software select the minimum sample size for each level of C. (Note,
however, that the number of units selected for evaluation
in the V&V protocol typically exceeds the minimum
required sample size to allow for the possibility that some

FIGURE 1 Operating characteristic (OC) curves for attribute


sampling for 95% reliability (RL) at 90% confidence (1 − β) when FIGURE 3 Probability of acceptance for 95% reliability at 90%
sample sizes are selected from Table 1 [Colour figure can be viewed confidence when the true but unknown (population) reliability is
at wileyonlinelibrary.com] 99.5% [Colour figure can be viewed at wileyonlinelibrary.com]
TUTORIAL 491

units may give an indeterminate or unevaluable response. between the variable sampling plan and the attribute sam-
For example, to get a minimum of N = 45 evaluable pling plan. Each of the OC curves of the variable sampling
in vitro diagnostic test results, the number of human plan corresponds to a unique value of p*, as provided in
specimens selected for testing may exceed N = 45 to allow Table 4. As previously noted, the value of p* plays the same
that some specimens may repeatedly give an error code or role as C/N in the attribute plan. For example, the OC
no reportable result.) curve in Figure 2 for the attribute sampling plan of
Additional OC curves for the attribute sampling plans (N = 100, C = 1) corresponds to a minimum acceptable
corresponding to the reliability and confidence combina- reliability estimate of (1 − C/N) = 0.990, as compared with
tions listed in Table 1 are provided in Appendix A for (1 − p*) = 0.9687 for the N = 100 variable sampling plan to
V&V practitioners to use. achieve the same required level of reliability (RL = 0.95) at
In Figure 4, the OC curves for the variable sampling the same level of confidence (β = 0.10). It is seen that,
plan are shown for 95% reliability (RL = 0.95) and 90% con- when the sample size is the same, the population reliabil-
fidence (β = 0.10) for sample sizes N = 50, 75, 100, 150, ity needed from the variable sampling plan is much lower
200, 300, 400, 500, and 800. For the variable sampling plan, than that from the attribute sampling plan to pass at the
all of the OC curves have a probability of acceptance of β same confidence (1 − β) and required reliability (RL). Note
when R = RL. In the case of Figure 4, when the product that while the attribute sampling plan for C = 0 imposes a
true but unknown reliability is equal to the required reli- minimum sample size, for example, N ≥ 45 for 95%
ability of 95%, or the proportion defective is 5%, the prob- reliability at 90% confidence, there is no minimum
ability of acceptance is 10% for all the OC curves. The sample size for the variable sampling plan, because the
OC curves in Figure 4 are labeled by the same sample sizes standard deviation of the variable of interest can be
as the OC curves in Figure 2 to allow a direct comparison arbitrarily small.
A variable sampling plan requires a smaller sample
size than an attribute one to demonstrate the fulfillment
of the same confidence and reliability requirement under
V&V. Table 5 compares some of the sampling plan OC
curves based on Figures 1 and 4. It is shown that an attri-
bute sampling plan needs more than double the sample
size of a variable sampling plan to achieve the same prob-
ability of acceptance.
To summarize, when developing an optimized sam-
pling plan for V&V, one must balance consumer protec-
tion and producer's technology limitations. On the basis
of design calculations or historical data, or other initial
engineering estimates, some level of product reliability
must be assumed to represent the population. Then, the
V&V requirement is analyzed to determine if an attribute
or variable sampling plan is appropriate.
FIGURE 4 Operating characteristic (OC) curves for variable An attribute sampling plan has its advantage when
sampling for 95% reliability (RL) at 90% confidence (1 − β) the V&V characteristics are evaluated from binary gauges
[Colour figure can be viewed at wileyonlinelibrary.com] because an exact measurement is excessively time‐

TABLE 5 Comparisons of sample sizes in variable vs attribute sampling plans for 95% reliability (RL = 0.95) and 90% confidence (β = 0.10)

Sample Size Additional Samples


Minimum Probability Product True Needed for Attribute
of Acceptance, % Reliability, % Attribute Variable Sampling

90 99 N = 105, C = 2 N = 50 55
90 97.7 N = 306, C = 10 N = 150 156
90 96.8 N = 807, C = 32 N = 400 407
80 98.1 N = 158, C = 4 N = 75 83
80 97.8 N = 209, C = 6 N = 100 109
80 97.1 N = 400, C = 14 N = 200 200
492 TUTORIAL

consuming or expensive or when multiple binary gauges Solution: Select Figure 1 that shows OC curves for
can be combined into a single binary outcome at the unit 95% reliability at confidence level of 90% requirement.
level. To select the appropriate sample size for an attri- Assume that the true but unknown reliability is 98%. It
bute sampling plan, the OC curves shown in Figure 1 or follows from Figure 1 that the minimum sample size is
Appendix A are used. For these attribute plans, the accep- falling between the OC curve of (N = 158, C = 4) with
tance criterion is based on comparing the number of fail- the probability of acceptance of 78.9% and the OC curve
ures observed with the number of failures allowed. of (N = 209, C = 6) with the probability of acceptance of
Variable sampling plans are applicable to a single 87.1% (rounded down). From Table 1, the sampling plan
continuous characteristic and use the sample mean and (N = 184, C = 5) is likely optimal. From Equation 2, the
sample standard deviation to calculate the reliability in probability of acceptance calculated for the (N = 184,
the sample. When applicable, the variable sampling plans C = 5) plan is 83.4% at the assumed true but unknown
are encouraged because a much smaller sample size is reliability of 98%. Provided that 83.4% probability of
required, as compared with an attribute sampling plan, acceptance is a reasonable target (as may be the case
as shown in Table 5. Operating characteristic curves for when the assumed true but unknown reliability is based
variable sampling, as shown in Figure 4 or Appendix B, on a pessimistic scenario), the minimum sample size is
are used to select the appropriate sample size. Equation 4 N = 184 with the maximum allowable failures of C = 5
is used to calculate the reliability from the data by referring to Table 1.
collected during the V&V. The calculated reliability is
compared with the allowable minimum reliability listed The exact probability of acceptance (Equation 2) can
in Tables 2–4 to evaluate the acceptance criteria. be then calculated using Excel using the function
Furthermore, the attribute sampling plan and the var- BINOM.DIST (c, n, 1‐true reliability, TRUE). In this case,
iable sampling plan are on equal footing in that both BINOM.DIST (c, n, 1‐true reliability, TRUE) = BINOM.
plans estimate the reliability in a statistical sample DIST (5, 184, 0.02, TRUE). The V&V is passed when less
(drawn from a population) and compare this to a mini- than or equal to 5 failures are seen.
mum acceptable estimate. In the attribute sampling plan,
the estimate of the reliability in the statistical sample is Example 2: Engineering dry runs will be performed
the proportion (number of passing results)/(total number prior to a V&V to evaluate if the assumed reliability of
of results), which is compared with the minimum accept- a binary attribute is correct. The quality requirement is
able estimate (1 − C)/N. In the variable sampling plan, to demonstrate the design meet specification of 95% reli-
the estimate of the reliability in the statistical sample is ability at confidence level of 90%. The target probability
the probability (1 − pL − pU), which is compared with of acceptance is 90%. The assumed device reliability is
the minimum acceptable estimate (1 − p*). 99.4%.
Solution: From Figure 1, it is shown that the mini-
mum sample size is between N = 45 (C = 0) and
4 | I M P L E M E N T A T I O N O F TH E N = 77 (C = 1) based on the assumed device reliability
OPTIMIZED S AMPLI N G DE S I G N of 99.4%. The larger sample size of N = 77 is selected
for the dry run. After the execution of the dry run, the
Development of an adequate sampling plan with a valid result was 2 out of 77 failures. Therefore, the device reli-
statistical rationale for V&V has been challenging for ability was estimated to be 75/77 or 97.4% in the dry run.
industry although the statistical theories behind are On the basis of Figure 1 or Table 1, a sampling plan of
widely available. Some examples are provided below to (105, 2) was chosen to redo the dry run and the second
show how to design an optimized attribute or variable dry run passed with 2 out of 105 failures. The device reli-
sampling plan in the context of V&V. ability was reestimated to be 178/182 (97.8%) by pooling
the 2 dry runs. On the basis of the best estimate avail-
Example 1: Design verification for a medical device is able, the true but unknown reliability was assumed to
conducted to demonstrate that a characteristic meets the be 97.8% and the sampling plan (N = 282, C = 9) was
specification of 95% reliability at confidence level of 90%. selected from Table 1 for the V&V to target a high prob-
Four different failure modes are expected from failure ability of acceptance of approximately 90%.
modes and effects analysis, which can be combined into Example 3: Medical device design verification is con-
a single binary attribute. The target probability of accep- ducted to determine if a quantitative in vitro diagnostic
tance is 80%. The initial estimate of the device reliability result meets the requirement of 95% reliability with
calculated from design parameters based on a pessimis- 95% confidence for a 1‐sided specification of the USL of
tic scenario is 98%. 2.2. The target probability of acceptance is 80%. The
TUTORIAL 493

initial estimate of the device reliability is 98.1%, and the the predefined specification, the required sample size gets
data are assumed to be normally distributed. large. Often the reliability is defined as the proportion of
Solution: From Figure B7 in Appendix B, assuming a the population of a continuous variable that must be
true but unknown reliability of 98.1% and targeting the between the LSL and the USL to be considered accept-
probability of acceptance of 80%, the sample size chosen able. In such cases, the reliability is related to the mean
is 100. After the data are collected, the mean value (X ) is and the standard deviation of the distribution of the con-
−0.02654, and the standard deviation (S) is 1.059. From tinuous variable and a variable sampling plan is more
Table 3, the 1‐sided tolerance limit k factor for a sample efficient than an attribute sampling plan.
of 100 at 95% confidence level and 95% reliability is This paper aims at providing optimized sampling
k1 = 1.9266, and the minimum acceptable estimate of the design with a valid statistical rationale for V&V, espe-
reliability is 0.9730 (=1 − p*) or 97.3%. ZU = (2.2 + 0.02654)/ cially for the medical device industry, to demonstrate
1.059 = 2.1025, so pU = Φ(−2.1025) = 0.0178. There is no future product quality or processes ability in addition to
LSL (so pL = 0), and therefore, the estimated reliability is comply with regulations. To start, one needs to determine
1 − pU = 0.9822 (or 98.2%), which is greater than the min- the appropriate levels of statistical confidence and mini-
imum acceptable estimate of 97.3%. Therefore, the design mum acceptable reliability that the V&V is going to dem-
verification passes. onstrate. This determination is based on the VOC and
Example 4: Medical device design verification is con- risk analysis. The manufacturing and technology limita-
ducted to determine if a quantitative in vitro diagnostic tions are then considered by plotting the probability of
meets the requirement of 95% reliability with 95% confi- acceptance versus the population reliability (OC curve)
dence for a 2‐sided specification of the USL of 2.5 and and selecting the smallest sample size N that gives a prac-
the LSL of −2.5. The target probability of acceptance is tical probability of acceptance. The attribute sampling
80%. Following an engineering study, the true product plan estimates the reliability (a proportion) by counting
reliability is assumed to be 98.1% and the data are the number of instances of a binary variable, whereas the
assumed to be normally distributed. variable sampling plan estimates the same proportion as
Solution: From Appendix B, Figure B7, with true reli- an integral over a probability density that depends on the
ability of 98.1% (=1 − x‐axis) and the probability of mean and standard deviation of a continuous variable.
acceptance of 80% (=y‐axis), the sample size chosen is The variable plan requires that the distribution of the
100. After test, it is calculated from the data collected continuous variable be normal or approximately normal.
that the mean value (X Þ is −0.02654, and the standard When the continuous variable of interest is not normally
deviation S is 1.059. distributed, a transformation may be used to make the
transformed variable approximately normal. For example,
From Table 3, k1 = 1.9266 for sample size n = 100 at the signal in an immunoassay is often log normally distrib-
95% reliability with 95% confidence, ZL = (−0.02654 + 2.5)/ uted because of the multiplicative effect of the various
1.059 = 2.3357, ZU = (2.5 + 0.02654)/1.059 = 2.3858, components of the system noise. Finally, if the variable is
pL = Φ(−2.3357) = 0.00975, pU = Φ(−2.3858) = 0.00852. not normally distributed and there is no suitable continu-
pL + pU = 0.0183, which is less than p* = 0.0270 ous transformation, then an attribute sampling plan may
(=1 − 0.9730). Therefore, the verification is passed with be used by transforming the continuous variable into a
the confirmation that the estimated reliability of 98.3% binary variable. In this case, the attribute plan is a non-
(=1 − 0.0183) is greater than the minimum acceptable parametric statistical technique, free of assumptions about
estimate of 97.3%. the distribution of the underlying continuous variable. If
the nonparametric method is used, then it is encouraged
to report descriptive statistics related to the variable data
5 | SUMMARY AND CONCLUSIONS (eg, mean, standard deviation, median, and interquartile
range) so that these values are readily available in the ver-
In a V&V of safety‐critical products, it is necessary to pro- ification/validation report. This will help understand the
tect the consumer by demonstrating that the reliability process, uncover trends, predict future performance, and
meets a predetermined minimum quality level at the support continuous improvement.
appropriate level of statistical confidence. Attribute sam- In this paper, the statistical theories for estimating reli-
pling plans are popular, but minimal plans are often ability by sampling a population are presented. Equations
inappropriate and the engineers responsible for V&V pro- and tables of commonly used attribute sampling plans
tocols need to understand the relationship between the based on the binominal distribution are provided; equa-
probability of passing the acceptance criteria and the tions and tables for the variable sampling plan based on
sample size. When the product true reliability is close to noncentral t distribution are provided; the probabilities
494 TUTORIAL

of acceptance associated with the sampling plans are pre- 4. Krishnamoorthy K, Mathew T. Statistical Tolerance Regions:
sented and OC curves are calculated and provided. In the Theory, Applications, and Computation. Hoboken, New Jersey:
statistical theory, one must be careful to distinguish John Wiley & Sons, Inc.; 2009. ISBN:978‐0‐470‐38026‐0.

between the reliability of the population (R) sometimes 5. Wilks SS. Determination of sample sizes for setting tolerance
called the “true but unknown” parameter, the minimum limits. Ann. Math. Stat. 1941;12(1):91‐96.

acceptable reliability based on the product requirements 6. Wilks SS. Statistical prediction with special reference to the
(RL), and the estimate of the reliability that is calculated problem of tolerance limits. Ann. Math. Stat. 1942;13(4):400‐409.

from the statistical sample of the population (RS). The sam- 7. ISO 16269‐6:2014. Statistical interpretation of data—part 6:
pling plans guarantee that a prespecified level of statistical determination of statistical tolerance intervals. 2014.

confidence (1 − β) can be achieved to limit the Consumer's 8. Owen DB. Factors for one‐sided tolerance limits and for variables
Risk and assure product quality, as appropriate during the sampling plans, Monograph No. SCR‐607. Sandia Corp., 1963.

V&V activities of product development or process valida- 9. NIST/SEMATECH. e‐Handbook of statistical methods. http://
tion. This is equivalent to saying that the V&V passes its www.itl.nist.gov/div898/handbook/ [30 Oct. 2013]

acceptance criteria when the lower 1‐sided (1 − β) confi- 10. ANSI/ASQ Z1.9‐2003 (R2013): Sampling procedures and tables
dence limit of the estimate of R is equal to or exceeds RL. for inspection variables for percent nonconforming. 2013.

In this paper, the method to develop appropriate attri- 11. BS EN ISO 14971:2012. Medical devices. Application of risk
bute and variable sampling plans to meet the reliability management to medical devices. 2012.

and confidence requirement for V&V is demonstrated, 12. IEC 60812:2006. Analysis techniques for system reliability—pro-
including examples showing how to achieve a high proba- cedure for failure mode and effects analysis (FMEA). 2006.

bility of passing the acceptance criteria. Formulas, 13. IEC 61025:2006. Fault tree analysis (FTA). 2006.
sample size tables, and OC curves for V&V as well as step‐ 14. Durivage M. Risk‐based approaches to establishing sample sizes for
by‐step examples are provided for engineering practitioners process validation. http://www.asqrd.org/risk‐based‐approaches‐to‐
to use. This paper also provides a statistical rationale for establishing‐sample‐sizes‐for‐process‐validation/ [16 June 2016]

selecting a sampling plan appropriate for conducting a 15. Montgomery DC. Introduction to Statistical Quality Control. Sixth
V&V. The V&V sampling method presented in this paper ed. Hoboken, New Jersey: John Wiley & Sons, Inc; 2009:638.
ISBN:978–0–470‐16992‐6.
captures both the VOCs and technology limitations. This
is significant because the appropriate sampling plans for 16. Chernick MR, Liu CY. The saw‐toothed behavior of power ver-
sus sample size and software solutions: single binomial
V&V are less well known than for manufacturing, as
proportion using exact methods. Am. Stat. 2002;56(2):149‐155.
evidenced by the fact that many companies have been cited
by Food and Drug Administration warning letters or
Form 483 Observations for lacking a valid statistical
Dr S. Cheng is a Certified Manager of Quality/Orga-
rationale when choosing a sampling plan for V&V.
nizational Excellence and a Certified Reliability Engi-
neer with ASQ and also a Certified Six Sigma Black
S. Cheng
Belt with both Motorola and ASQ. She has practiced
K. Kupfer
quality, Six Sigma, reliability, and structural analysis
M. Dixon
using statistical methods at many major companies
S. Shammas
in the medical device industry including Alere, Baxter
Alere San Diego, Inc, San Diego, CA, USA
Healthcare, Life Technologies, St. Jude Medical, and
Correspondence Johnson and Johnson. She has also worked in the
S. Cheng, Alere San Diego, Inc, San Diego, CA, USA. automotive, telecommunications, and aerospace
Email: shumin.cheng@quidel.com industries. She holds a PhD degree in Engineering
and Applied Science from the Memorial University
of Newfoundland in Canada.
R EF E RE N C E S
1. Ferryanto L. Statistical sampling plan for design verification and Dr K. Kupfer is a Fellow Biostatistician at Alere, Inc,
validation of medical devices. http://www.ivtnetwork.com/ and holds a PhD degree in Applied Plasma Physics
print/article/statistical‐sampling‐plan‐design‐verification‐and‐ and Nuclear Engineering from the Massachusetts
validation‐medical‐devices [6 May 2015] Institute of Technology. He was previously an Assis-
2. Lieberman GJ, Resnikoff GJ. Sampling plans for inspection by tant Professor of Molecular Biology and Oncology at
variables. J Am Stat Assoc. 1955;50:457‐516. Southwestern Medical Center in Dallas, and he has
3. Bruhn‐Suhr M, Krumbholz W. Exact two‐sided Lieberman‐ worked in the pharmaceutical and medical device
Resnikoff sampling plans. Stat. Pap. 1991;32(1):233‐241. industry for approximately 20 years.
TUTORIAL 495

Mr M. Dixon is the Director of Quality Engineering


at Alere, San Diego, and holds a BS degree in Chemi-
cal Engineering from the University of California,
Santa Barbara. He is a certified Six Sigma Black Belt
and has worked in the medical device industry for
approximately 15 years with an emphasis on continu-
ous improvement and new product development in
the rapid diagnostics space.

Mr S. Shammas is a Quality Engineer at Alere San


Diego. He has practiced quality and reliability analysis
at major companies in medical device and diagnostics
industry including Alere, Beckman, and KONE. He
holds a bachelor's degree in Chemistry from the Uni-
versity of Massachusetts.
FIGURE A2 Operating characteristic (OC) curves for attribute
sampling for 95% reliability at 97.5% confidence [Colour figure can
A P P EN D I X A: A TT RI B U T E SA M P LI N G O C be viewed at wileyonlinelibrary.com]
CU RVES WH E N S A M PLIN G SI ZES A RE
S E L E C T E D F R O M T A B L E 2 ( E F F E C TI V E
SAMPLING PLAN)

FIGURE A3 Operating characteristic (OC) curves for attribute


sampling for 90% reliability at 97.5% confidence [Colour figure can
be viewed at wileyonlinelibrary.com]

FIGURE A1 Operating characteristic (OC) curves for attribute


sampling for 99% reliability at 97.5% confidence [Colour figure can
be viewed at wileyonlinelibrary.com]
496 TUTORIAL

FIGURE A4 Operating characteristic (OC) curves for attribute FIGURE A6 Operating characteristic (OC) curves for attribute
sampling for 85% reliability at 97.5% confidence [Colour figure can sampling for 99% reliability at 95% confidence [Colour figure can
be viewed at wileyonlinelibrary.com] be viewed at wileyonlinelibrary.com]

FIGURE A5 Operating characteristic (OC) curves for attribute FIGURE A7 Operating characteristic (OC) curves for attribute
sampling for 80% reliability at 97.5% confidence [Colour figure can sampling for 95% reliability at 95% confidence [Colour figure can
be viewed at wileyonlinelibrary.com] be viewed at wileyonlinelibrary.com]
TUTORIAL 497

FIGURE A8 Operating characteristic (OC) curves for attribute FIGURE A10 Operating characteristic (OC) curves for attribute
sampling for 90% reliability at 95% confidence [Colour figure can sampling for 80% reliability at 95% confidence [Colour figure can
be viewed at wileyonlinelibrary.com] be viewed at wileyonlinelibrary.com]

FIGURE A9 Operating characteristic (OC) curves for attribute FIGURE A11 Operating characteristic (OC) curves for attribute
sampling for 85% reliability at 95% confidence [Colour figure can sampling for 99% reliability at 90% confidence [Colour figure can
be viewed at wileyonlinelibrary.com] be viewed at wileyonlinelibrary.com]
498 TUTORIAL

FIGURE A12 Operating characteristic (OC) curves for attribute FIGURE A14 Operating characteristic (OC) curves for attribute
sampling for 95% reliability at 90% confidence [Colour figure can sampling for 85% reliability at 90% confidence [Colour figure can
be viewed at wileyonlinelibrary.com] be viewed at wileyonlinelibrary.com]

FIGURE A13 Operating characteristic (OC) curves for attribute FIGURE A15 Operating characteristic (OC) curves for attribute
sampling for 90% reliability at 90% confidence [Colour figure can sampling for 80% reliability at 90% confidence [Colour figure can
be viewed at wileyonlinelibrary.com] be viewed at wileyonlinelibrary.com]
TUTORIAL 499

A P P EN D I X B: VA R I A B L E SA M PL I N G O C
CURVES

FIGURE B3 Operating characteristic (OC) curves for variable


sampling for 90% reliability at 97.5% confidence [Colour figure can
be viewed at wileyonlinelibrary.com]

FIGURE B1 Operating characteristic (OC) curves for variable


sampling for 99% reliability at 97.5% confidence [Colour figure can
be viewed at wileyonlinelibrary.com]

FIGURE B4 Operating characteristic (OC) curves for variable


sampling for 85% reliability at 97.5% confidence [Colour figure can
be viewed at wileyonlinelibrary.com]

FIGURE B2 Operating characteristic (OC) curves for variable


sampling for 95% reliability at 97.5% confidence [Colour figure can
be viewed at wileyonlinelibrary.com]
500 TUTORIAL

FIGURE B5 Operating characteristic (OC) curves for variable FIGURE B7 Operating characteristic (OC) curves for variable
sampling for 80% reliability at 97.5% confidence [Colour figure can sampling for 95% reliability at 95% confidence [Colour figure can
be viewed at wileyonlinelibrary.com] be viewed at wileyonlinelibrary.com]

FIGURE B6 Operating characteristic (OC) curves for variable FIGURE B8 Operating characteristic (OC) curves for variable
sampling for 99% reliability at 95% confidence [Colour figure can sampling for 90% reliability at 95% confidence [Colour figure can
be viewed at wileyonlinelibrary.com] be viewed at wileyonlinelibrary.com]
TUTORIAL 501

FIGURE B9 Operating characteristic (OC) curves for variable FIGURE B11 Operating characteristic (OC) curves for variable
sampling for 85% reliability at 95% confidence [Colour figure can sampling for 99% reliability at 90% confidence [Colour figure can
be viewed at wileyonlinelibrary.com] be viewed at wileyonlinelibrary.com]

FIGURE B10 Operating characteristic (OC) curves for variable FIGURE B12 Operating characteristic (OC) curves for variable
sampling for 80% reliability at 95% confidence [Colour figure can sampling for 95% reliability at 90% confidence [Colour figure can
be viewed at wileyonlinelibrary.com] be viewed at wileyonlinelibrary.com]
502 TUTORIAL

FIGURE B13 Operating characteristic (OC) curves for variable FIGURE B15 Operating characteristic (OC) curves for variable
sampling for 90% reliability at 90% confidence [Colour figure can sampling for 80% reliability at 90% confidence [Colour figure can
be viewed at wileyonlinelibrary.com] be viewed at wileyonlinelibrary.com]

FIGURE B14 Operating characteristic (OC) curves for variable


sampling for 85% reliability at 90% confidence [Colour figure can
be viewed at wileyonlinelibrary.com]

You might also like