Statistics in Biopharmaceutical Research

ISSN: (Print) 1946-6315 (Online) Journal homepage: http://www.tandfonline.com/loi/usbr20

Improved Acceptance Limits for ASTM Standard

E2810

Richard A. Lewis & Ailin Fan

To cite this article: Richard A. Lewis & Ailin Fan (2016) Improved Acceptance Limits
for ASTM Standard E2810, Statistics in Biopharmaceutical Research, 8:1, 40-48, DOI:
10.1080/19466315.2015.1093959

To link to this article: http://dx.doi.org/10.1080/19466315.2015.1093959

Published online: 22 Mar 2016.

Submit your article to this journal

Article views: 21

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

http://www.tandfonline.com/action/journalInformation?journalCode=usbr20

Download by: [University of Nebraska, Lincoln] Date: 31 May 2016, At: 23:30
 Improved Acceptance Limits for ASTM
Standard E2810
Downloaded by [University of Nebraska, Lincoln] at 23:30 31 May 2016

Richard A. LEWIS and Ailin FAN

intended to serve as a lot-release procedure (USP Gen-

ASTM Standard E2810 provides a methodology for es- eral Notices and Requirements 2010). ASTM Standard
tablishing confidence in passing the USP <905> Unifor- E2709 (ASTM International 2012) provides an approach
mity of Dosage Units (UDU) test, and provides accep- for establishing confidence in the probability of passing
tance limits for sample means and standard deviations a test, and ASTM Standard E2810 (ASTM International
that can be used as elements of lot release. These accep- 2011) applies this approach to the USP <905> UDU
tance limits are quite conservative, however, due to both test. E2810 provides acceptance limits for sample means
the nature and shape of an inverted triangular joint con- and standard deviations for a variety of confidence levels
fidence region for the lot mean and standard deviation. and lower bounds on the probability of passing the UDU
We obtain improved (wider) acceptance limits by using a test, and states that these acceptance limits can be used as
Bayesian approach that focuses on the posterior distribu- elements of lot release. It turns out that these acceptance
tion of the probability of passing the USP <905> UDU limits are quite conservative, however, which can lead to
test. The Bayesian approach has good sampling proper- unnecessary rejection of batches.
ties, and the improvement in acceptance limits can be We obtain improved (wider) acceptance limits by us-
considerable. For example, for sample size of 10 units ing a Bayesian approach that focuses on the posterior
and sample means between 97 and 103, the Bayesian ap- distribution of the probability of passing the UDU test.
proach results in acceptance limits for sample standard The Bayesian approach has good sampling properties,
deviations that are at least 25% greater than those in and the improvement in acceptance limits can be con-
E2810. The impact of the improved acceptance limits is siderable. The Bayesian approach is straightforward, but
illustrated with operating characteristic curves. computationally intensive. We determine acceptance lim-
its for a sufficient number of cases to contrast the Bayesian
Key Words: ASTM Standard E2709; ASTM Standard E2810;
Lot-release, USP <905>; Uniformity of dosage units (UDU).
and E2810 approaches, but do not attempt to provide a
comprehensive tabulation.
Our article is organized as follows. Section 2 de-
1. Introduction scribes the UDU test and the approach E2810 uses to
establish confidence in passing it. Section 3 describes the
Content uniformity is a key quality characteristic for Bayesian approach and explains why it results in wider
drug products. United States Pharmacopeia (USP) gen- acceptance limits. Section 4 explores sampling properties
eral chapter <905>, Uniformity of Dosage Units, pro- of the Bayesian approach and provides operating charac-
vides a two-stage test for content uniformity, but is not teristic curves for a few cases. Section 5 summarizes our

© 2016 GlaxoSmithKline
Statistics in Biopharmaceutical Research
February 2016, Vol. 8, No. 1
DOI: 10.1080/19466315.2015.1093959

40
 Improved Acceptance Limits for ASTM Standard E2810

results and discusses potential extensions of the Bayesian First, specify a desired level of confidence on a lower
approach to additional situations, such as blend unifor- bound for P, such as 90% confidence that P is at least
mity and dissolution. 95%. We designate the lower bound for P by LB and the
confidence level by 100(1 − α)%, 0 < α < 1. Second, find
the contour line for LB in two-dimensional (μ, σ ) space.
Third, obtain a 100(1 − α)% confidence region for μ and
2. Establishing Confidence in Passing the σ , based on ȳ and s; E2810 uses an inverted triangular
USP <905> UDU Test confidence region (Lindgren 1968). Then for any specific
value of ȳ, the acceptance limit for s is the largest value
We begin by establishing some notation and describ- of s for which the confidence region abuts the contour
ing the USP <905> UDU test. Suppose dosage units line for LB. This is illustrated in Figure 1 for α = 0.10
within a lot are normally distributed with mean μ and (90% confidence), LB = 95%, ȳ = 97, and n = 10; the
standard deviation σ . The units for μ and σ are percent E2810 acceptance limit is s = 2.63. Table 1 gives selected
of label claim (%LC). As in E2810, we assume the target acceptance limits from ASTM Standard E2810.
The solid black curve in Figure 1 is the contour line
Downloaded by [University of Nebraska, Lincoln] at 23:30 31 May 2016

content per dosage unit is 100%LC. A random sample

from the lot can be described by for LB = 95%. The inverted triangle is a 90% confi-
dence region (CR) for μ and σ , based on ȳ = 97, s =
Yi = μ + ei , for i = 1, 2, . . . , n, (1) 2.63, and n = 10. The CR abuts the contour line for
LB = 95%, so 2.63 is the E2810 acceptance limit for s.
where the ei , which include genuine differences be-
The blue, roughly ellipsoidal region is the 90% highest
tween dosage units as well as variation in the mea-
posterior density (HPD) region for μ and σ . One would
surement system, are iid Normal (0, σ 2 ). The Yi are
obtain an improved acceptance limit by using the HPD
random variables with corresponding observed values
region instead of the CR, by increasing s until the HPD
y1 , y2 , . . . , yn . We shall make frequent use
n of the sam- abuts the contour line for LB = 95%. The dashed green
ple mean
 and standard deviation: ȳ = i=1 yi /n and curve represents combinations of μ and σ associated with
n
s= i=1 (yi − ȳ) / (n − 1). Key to tabulating accep-
2
Q 0.10 (P), the 10th percentile of the posterior distribution
tance limits is the fact that ȳ and s are sufficient statistics for P. The Bayesian acceptance limits for s are obtained
for μ and σ , so neither E2810 nor the Bayesian approach by increasing s until Q 0.10 (P) = 95%, at which point the
requires the values of individual observations. solid black and dashed green curves coincide with one
The UDU test is a two-stage test. Stage 1 is based on another. For this example, the Bayesian acceptance limit
a random sample of size n1 = 10 and Stage 2, if needed, for s is 3.57.
involves an additional random sample of size 20 for a total
sample size of n2 = 30. Let ȳ1 and s1 represent the sample
mean and standard deviation of the 10 observations in 3. Bayesian Approach
Stage 1, and let ȳ2 and s2 represent the sample mean
and standard deviation of the 30 observations in Stage Improved (wider) acceptance limits for the sample
2. In addition, let Mi = 98.5 if ȳi < 98.5, let Mi = ȳi if mean and standard deviation can be obtained by us-
98.5 ≤ ȳi ≤ 101.5, and let Mi = 101.5 if ȳi > 101.5, for ing a Bayesian approach that focuses on the posterior
stages i = 1, 2. When the target content per dosage unit distribution of P, the probability of passing the USP
is 100%LC, the two stages of the UDU test are: <905> UDU test. More specifically, we focus on a
Stage S1. Pass the UDU test if |M1 − ȳ1 | + 2.4s1 ≤ 100(1 − α)% one-sided credible interval—similar to a
15.0; otherwise proceed to Stage 2. confidence interval—of the form (Q α (P) , 1], where
Stage S2. Pass the UDU test if |M2 − ȳ2 | + 2.0s2 ≤ Pr (Q α (P) <P ≤ 1| y1 , y2 , . . . , yn ) = 1 − α
15.0 and all individual results are between 0.75M 2 and
1.25M 2 ; otherwise fail the UDU test. for 0 < α < 1. It follows that Q α (P) is the α quantile
Let P represent the probability of passing the UDU (or 100 α th percentile) of the posterior distribution for P,
test. Although P is a complex function of μ and σ with which depends only on ȳ, s, n and the prior distribution
no closed-form solution, it can be estimated via Monte for μ and σ . Given α, a lower bound LB for P, and any
Carlo simulation. specific value of ȳ, the Bayesian acceptance limit for s is
ASTM Standard E2709 provides a general approach the largest value of s for which Q α (P) is greater than or
for establishing confidence in the probability of passing a equal to LB. This may appear to be completely different
test, based on an approach developed by Bergum (1990). than the approach in E2810, which places a confidence
ASTM Standard E2810 applies E2709 to the probability region on μ and σ , but, since P is a function of μ and σ ,
of passing the UDU test, following Bergum and Li (2007). a credible interval for P also places limits on μ and σ .

41
 Statistics in Biopharmaceutical Research: February 2016, Vol. 8, No. 1
Downloaded by [University of Nebraska, Lincoln] at 23:30 31 May 2016

Figure 1. A comparison of the E2810 and Bayesian approaches for α = 0.10, LB = 95%, ȳ = 97, s = 2.63, and n = 10.

 
Following the approach in Box and Tiao (1973, sec. bution for μ given τ is Normal ȳ, 1/ (nτ ) = σ 2 /n . One
2.4) and Glickman and van Dyk (2007, chap. 16), we use can generate samples from the joint posterior distribu-
the improper, noninformative prior p (μ, σ ) = 1/σ . This tion for μ and τ by alternating between samples from
prior is widely used and yields inference for μ and σ the marginal posterior distribution for τ and samples
that mirrors sampling theory results. The marginal pos- from the conditional posterior distribution for μ given
terior distribution for τ = 1/σ 2 is Gamma (a, b), with τ , because the joint posterior distribution for μ and τ
probability density function is p (μ, τ | ȳ, s) = p(τ | ȳ, s) · p (μ|τ, ȳ, s). The posterior
distribution for P can be obtained from the posterior dis-
p(τ | ȳ, s) = ba τ a−1 exp (−bτ ) /  (a) , σ > 0, tribution for (μ, τ ) via Monte Carlo simulation; this is ex-
plained in greater detail below. We conducted all Monte
where a = (n − 1) /2, b = (n − 1) s 2 /2, and E(τ | ȳ, s) = Carlo simulation using SAS R
software, Version 9.3 of
a/b = 1/s 2 . In addition, the conditional posterior distri-

Table 1. Acceptance limits for the sample standard deviation, s target = 100% LC

ASTM Standard E2810 Bayesian approach

Confidence/ Sample mean Sample size, n Sample size, n
Probability LB ȳ 10 30 50 10 30 50

95% 90% 100 2.91 4.36 4.84 3.71 4.93 5.31

95% 90% 99 or 101 2.74 4.10 4.57 3.67 4.85 5.19
95% 90% 98 or 102 2.56 3.84 4.28 3.53 4.61 4.97
95% 90% 97 or 103 2.39 3.58 3.99 3.33 4.34 4.66
95% 95% 100 2.81 4.18 4.64 3.55 4.73 5.08
95% 95% 99 or 101 2.64 3.94 4.39 3.51 4.65 4.98
95% 95% 98 or 102 2.47 3.68 4.11 3.39 4.42 4.75
95% 95% 97 or 103 2.30 3.43 3.82 3.17 4.15 4.43
95% 99% 100 2.62 3.88 4.31 3.29 4.37 4.70
95% 99% 99 or 101 2.46 3.65 4.06 3.27 4.29 4.59
95% 99% 98 or 102 2.31 3.42 3.80 3.10 4.07 4.36
95% 99% 97 or 103 2.15 3.19 3.54 2.93 3.80 4.06
90% 95% 100 3.21 4.47 4.88 4.02 5.01 5.29
90% 95% 99 or 101 3.02 4.21 4.61 3.95 4.91 5.19
90% 95% 98 or 102 2.82 3.95 4.32 3.78 4.71 4.96
90% 95% 97 or 103 2.63 3.68 4.03 3.57 4.40 4.63

42
 Improved Acceptance Limits for ASTM Standard E2810

the SAS System for Windows (© 2002–2010, SAS In- Table 2. Ratios of acceptance limits for the sample standard
deviation, s target = 100% LC
stitute Inc., SAS and all other SAS Institute Inc. product
or service names are registered trademarks or trademarks
of SAS Institute Inc., Cary, NC, USA). Ratio (Bayesian/E2810)
Determining the Bayesian acceptance limits is Confidence/ Sample mean Sample size, n
straightforward, but computationally intensive. Given α, a Probability LB ȳ 10 30 50
lower bound LB for P, ȳ, and n, we proceeded as follows.
95% 90% 100 1.27 1.13 1.10
1. Specify a value for s, and generate N 1 = 5000 95% 90% 99 or 101 1.34 1.18 1.14
samples from the posterior distribution for μ and 95% 90% 98 or 102 1.38 1.20 1.16
95% 90% 97 or 103 1.39 1.21 1.17
τ . Our goal is to estimate Q α (P) for this and 95% 95% 100 1.26 1.13 1.09
several other values of s. 95% 95% 99 or 101 1.33 1.18 1.13
95% 95% 98 or 102 1.37 1.20 1.16
2. For each sample from the posterior distribution for 95% 95% 97 or 103 1.38 1.21 1.16
μ and τ , generate N 2 = 5000 datasets to estimate 95% 99% 100 1.26 1.13 1.09
Downloaded by [University of Nebraska, Lincoln] at 23:30 31 May 2016

the true value of P for that particular combination 95% 99% 99 or 101 1.33 1.18 1.13
95% 99% 98 or 102 1.34 1.19 1.15
of μ and τ . This results in N 1 = 5000 samples
95% 99% 97 or 103 1.36 1.19 1.15
from the posterior distribution for P, from which 90% 95% 100 1.25 1.12 1.08
we estimate Q α (P) using SAS/STAT © PROC 90% 95% 99 or 101 1.31 1.17 1.13
UNIVARIATE. 90% 95% 98 or 102 1.34 1.19 1.15
90% 95% 97 or 103 1.36 1.20 1.15
3. Repeat Steps 1 and 2 as many times as needed to
obtain seven values of Q α (P) that bracket LB, us-
the Bayesian approach provides a better improvement.
ing a step size of 0.05 for s. At least three estimates
Second, the inverted triangular confidence region—and
of Q α (P) should be above LB, and at least three
the HPD region, as well—excludes values of μ and σ
should be below LB. Fit a quadratic equation to
that have high probabilities of passing the UDU test. In
the seven pairs (s, Q α (P)) and, using a step size
particular, the confidence region excludes the combina-
of 0.01, find the largest value of s for which the
tion μ = 100 and sigma σ = 0, unless ȳ = 100. One
fitted value for Q α (P) exceeds LB.
would do better by partitioning the (μ, σ ) space into two
As an example, for α = 0.05 (95% probability), regions, one containing values of μ and σ with higher
LB = 90%, ȳ = 100, and n = 30, we used s ranging from probabilities of passing the UDU test and one containing
4.75 to 5.05 by 0.05, which gave Q 0.05 (P) estimates of values of μ and σ with lower probabilities of passing the
94.85%, 93.56%, 91.42%, 92.06%, 89.54%, 87.65%, and UDU test. Such a partition is provided by contour lines
86.59%, respectively. The estimates of Q 0.05 (P) are not for P. These two reasons are illustrated in Figure 1, for α
monotonic due to simulation error. The resulting accep- = 0.10, LB = 95%, ȳ = 97, s = 2.63, and n = 10.
tance limit for s is 4.93. The results for 48 combinations It turns out that while the Bayesian approach is based
of probability 100(1 − α)%, LB, ȳ, and n are shown in on the posterior distribution for P, it can be viewed as
Table 1, alongside the corresponding acceptance limits using contour lines for P to partition the (μ, σ ) space
in E2810. The ratio of the Bayesian acceptance limits to into two regions. The dashed line in Figure 1 represents
the limits in E2810 is shown in Table 2. For sample size combinations of μ and σ that correspond to Q 0.10 (P),
10 the Bayesian acceptance limits are 25%–39% greater which is approximately 99.999% when ȳ = 97, s = 2.63,
than the corresponding limits in E2810; for sample size and n = 10. Combinations of μ and σ below the dashed
30 the Bayesian acceptance limits are 12%–21% greater; line have higher probabilities of passing the UDU test,
and for sample size 50 the Bayesian acceptance limits are and combinations of μ and σ above the dashed line have
8%–16% greater. This is a substantial improvement to the lower probabilities of passing the UDU test. Using the
acceptance limits in E2810. Bayesian approach, s can be increased until the contour
There are two reasons for the wider acceptance lim- line for Q α (P) exactly matches the contour line for LB
its. First, E2810 uses an inverted triangular confidence = 95%; in this example this occurs when s = 3.57.
region that forces acceptable values of ȳ and s relatively
far from the contour line for LB. One would do better
by using the roughly ellipsoidal highest posterior density 4. Sampling Properties and Operating
(HPD) region for μ and σ shown in Figure 1 or a similarly Characteristic Curves
shaped confidence region, such as one of the approximate
confidence regions given by Arnold and Shavelle (1998). As previously mentioned, Bayesian inference for
We did not pursue this improvement to E2810 because μ and σ mirrors sampling theory results when one

43
 Statistics in Biopharmaceutical Research: February 2016, Vol. 8, No. 1

Table 3. Estimated coverage of the one-sided credible interval for P for sample size n = 30

Nominal and estimated Coverage, %

μ σ P n 90 95 99

100 5.726 99% 30 93.6 97.6 99.2

100 6.185 95% 30 93.7 97.6 99.6
100 6.461 90% 30 92.9 97.4 99.4
100 7.638 50% 30 95.1 97.2 99.6
100 9.392 10% 30 92.5 96.2 98.8
100 10.116 5% 30 92.7 97.1 99.5
100 12.098 1% 30 94.5 97.9 99.5
Average: 93.6 97.3 99.4
90 or 110 2.395 99% 30 90.7 95.0 99.3
90 or 110 2.623 95% 30 91.5 95.9 98.3
90 or 110 2.760 90% 30 90.3 94.6 98.8
90 or 110 3.396 50% 30 89.0 94.6 98.3
Downloaded by [University of Nebraska, Lincoln] at 23:30 31 May 2016

90 or 110 4.546 10% 30 88.6 94.2 99.0

90 or 110 5.125 5% 30 89.3 95.3 99.1
90 or 110 7.037 1% 30 90.6 94.9 98.9
Average: 90.0 94.9 98.8

uses the noninformative prior p (μ, σ ) = 1/σ (Box and from the posterior distribution for P; use these
Tiao 1973). Consequently, it seems reasonable to expect to calculate Q α (P) using SAS/STAT © PROC
Q α (P), the α quantile of the posterior distribution of P, UNIVARIATE.
to have good sampling properties. We briefly explored
3. Calculate the percentage of the N 1 = 1000 samples
the sampling properties of Q α (P) for 14 combinations
of size n = 30 from Step 1 for which the true value
of μ and σ with associated values of P ranging from 1%
of P is within the interval (Q α (P), 1]. This is the
to 99%. The 14 combinations of μ and σ are shown in
estimated coverage of the 100(1 − α)% one-sided
Table 3, where we estimated the true value of P using
credible interval (Q α (P), 1].
Monte Carlo simulation with N = 1,000,000 simulated
datasets for each combination of μ and σ . The results are summarized in Table 3. Estimated cov-
For each of the 14 combinations of μ and σ in Table 3, erage is in line with nominal coverage for lot means of 90
we are interested in how often the 100(1 − α)% one-sided or 110, and somewhat conservative for a lot mean of 100.
credible interval (Q α (P), 1] contains the (estimated) true This indicates that the Bayesian approach’s wider accep-
value of P. We refer to this as the coverage of the credible tance limits are due to the nature and shape of the inverted
interval. We fixed sample size at n = 30 and estimated triangular confidence region used by E2810, rather than
coverage for α = 0.01, 0.05, and 0.10. We proceeded as a liberal credible interval for Q α (P).
follows. Now suppose we release a lot if we have 90% con-
fidence or probability that P is a least 95%, based on a
1. For each combination of μ and σ in Table 3, gen-
sample size of n = 30. The resulting operating character-
erate N 1 = 1000 samples of size n = 30 from a
istic (OC) curves for the E2810 and Bayesian approaches
normal distribution with mean μ and standard de-
are shown in Figures 2 and 3, along with a curve showing
viation σ . We only need ȳ and s for each sample,
the probability of passing the UDU test. OC curves show
and generating these sufficient statistics is straight-
the probability of releasing a lot versus true properties
forward: ȳ and s are independent, ȳ has a nor-
of the lot. Figure 2 shows the probability of lot release
mal distribution with mean μ and variance σ 2 /n,
versus σ for μ = 100, and Figure 3 shows the probability
and (n − 1) s 2 /σ 2 has a chi-squared distribution
of lot release versus σ for μ = 97. The OC curves for the
with n − 1 degrees of freedom. Our goal is to find
Bayesian approach are about 0.6 to 0.7 units to the right
Q α (P) for each sample, and compare it to the true
of E2810’s OC curve, while providing 90% probability
value of P.
that P is a least 95%. These 0.6 to 0.7 units are a win-
2. For each sample from Step 1, generate N 2 = 1000 dow in which the acceptance limits in E2810 can lead to
samples from the posterior distribution for μ and unnecessary rejection of batches.
σ . For each of these samples from the posterior dis- To produce the OC curves, we first estimated Bayesian
tribution, generate N 3 = 1000 datasets to estimate acceptance limits for sample means ȳ = 90, 91,. . ., 96
the true value of P for that particular combina- using a lookup table to estimate P, rather than the Monte
tion of μ and σ . This gives N 2 = 1000 samples Carlo simulation used in Step 2 of Section 3. These

44
 Improved Acceptance Limits for ASTM Standard E2810
Downloaded by [University of Nebraska, Lincoln] at 23:30 31 May 2016

Figure 2. OC curves for α = 0.10, LB = 95%, μ = 100, and n = 30.

acceptance limits are shown in Table 4 along with the Bayesian acceptance limits and used the fitted curve to
corresponding E2810 acceptance limits. The ratio of determine lot release status. Finally, we varied σ from 0 to
Bayesian to E2810 limits is 1.19–1.20 for all values of ȳ 10 by 0.1 and, for each value of σ , used 10,000 simulated
from 90 to 96. LeBlond and Mockus (2014) used a lookup datasets to estimate the probability of lot release.
table for a more complex one-way random effects model. We note that one can obtain similar OC curves—for
To create our lookup table, we varied μ from 100 to 125 both the E2810 and Bayesian approaches—by using a
by 0.5 units and σ from 0 to 14 by 0.1 units, using σ = parametric tolerance interval test (PTIT) for lot release.
0.0001 in place of σ = 0. We estimated P from 40,000 A tolerance interval allows one to make a statement such
simulated datasets of size n = 30 for each combination as “We are 100(1 − α)% confident that 100 γ % of dosage
of μ and σ , where each dataset either passed or failed the units are between L and U.” Lostritto (2012) proposed us-
UDU test. We then fit a fourth-order polynomial to the ing a PTIT as an alternative to the USP <905> UDU test,

Figure 3. OC curves for α = 0.10, LB = 95%, μ = 97, and n = 30.

45
 Statistics in Biopharmaceutical Research: February 2016, Vol. 8, No. 1

Table 4. Estimated acceptance limits for the sample standard deviation, s target = 100% LC, Confidence/Probability = 90%, LB = 95%, n = 30

Sample mean ȳ
90 91 92 93 94 95 96

ASTM Standard E2810 1.77 2.05 2.32 2.59 2.86 3.13 3.41
Bayesian approach 2.12 2.44 2.77 3.10 3.43 3.72 4.07
Ratio (Bayesian/E2810) 1.20 1.19 1.19 1.20 1.20 1.19 1.19

and Hudson-Curtis and Novick (2016) proposed using a mately 95% probability of passing the UDU test, and
PTIT for lot release of solid oral dose drug products. 99.4% coverage of (83.5, 116.5) provides approximately
Novick et al. (2009) described a two-stage PTIT for de- 99% probability of passing the UDU test. The PTIT pro-
livered dose uniformity of inhaled products. In this arti- cedure individually controls the left and right tails of
cle, we use a single-stage PTIT, as in Lewis and Novick the distribution. Consequently, 98.1%, 98.7%, and 99.4%
(2011). coverage are only achieved when μ = 100, because in-
Downloaded by [University of Nebraska, Lincoln] at 23:30 31 May 2016

Hudson-Curtis and Novick present the PTIT approach dividual tail areas are restricted to (1 − 0.0190)/2 =
as an alternative to the USP <905> UDU test. Here we 0.0095, (1 − 0.0130)/2 = 0.0065, and (1 − 0.0060)/2 =
show how the PTIT approach indirectly provides con- 0.0030, respectively. For lot means sufficiently far from
fidence in passing the USP <905> UDU test by con- 100, the actual levels of coverage are 99.05%, 99.35%,
trolling tail areas in the distribution of dosage units, and 99.70%, respectively. Table 5 shows acceptance lim-
thereby providing a rational approach for selecting spe- its for PTIT-based lot-release procedures based on estab-
cific PTIT parameters. Bergum (2015) also described how lishing 100(1 − α)% confidence of 98.1%, 98.7%, and
tolerance intervals can provide confidence in passing the 99.4% coverage of (83.5, 116.5). The PTIT acceptance
USP <905> UDU test, employing a different approach. limits are similar to the corresponding Bayesian accep-
Figure 4 shows three possible selections for (γ , L, U) tance limits for sample means outside the range 98–102,
overlaid on contours for P = 90%, P = 95%, and P but somewhat wider for sample means close to 100.
= 99%. For sample means outside the range 98–102 it Figure 5 compares the OC curves for a PTIT-based lot-
can be seen that 98.1% coverage of (83.5, 116.5) pro- release procedure with α = 0.10 and 98.7% coverage of
vides approximately 90% probability of passing the UDU (83.5, 116.5) to OC curves for the Bayesian procedure
test, 98.7% coverage of (83.5, 116.5) provides approxi- with α = 0.10 and P = 95%. The PTIT and Bayesian OC

Figure 4. 98.1%, 98.7%, and 99.4% coverage of (83.5, 116.5) and contours for P = 90%, P = 95%, for P = 99%.

46
 Improved Acceptance Limits for ASTM Standard E2810

Table 5. Acceptance limits for the sample standard deviation, s, 5. Summary

based on a PTIT approach confidence and coverage apply to the
interval (83.5, 116.5)
ASTM Standard E2810 provides an approach for es-
tablishing confidence in passing the USP <905> Unifor-
PTIT approach
mity of Dosage Units (UDU) test, but gives conservative
Sample mean Sample size, n acceptance limits due to both the nature and shape of its
Confidence Coverage LB ȳ 10 30 50 joint confidence region for the lot mean and standard de-
viation. We have shown that a Bayesian approach based
95% 98.1% 90% 100 4.11 5.34 5.72
95% 98.1% 90% 99 or 101 3.86 5.02 5.37
on a one-sided credible interval for the probability of
95% 98.1% 90% 98 or 102 3.61 4.70 5.03 passing the USP <905> UDU test leads to improved
95% 98.1% 90% 97 or 103 3.37 4.37 4.68 (wider) acceptance limits. The credible interval has good
95% 98.7% 95% 100 3.90 5.06 5.41 sampling properties, and can be viewed as a more sen-
95% 98.7% 95% 99 or 101 3.66 4.75 5.09
95% 98.7% 95% 98 or 102 3.43 4.45 4.76 sible “confidence region” for the lot mean and standard
95% 98.7% 95% 97 or 103 3.19 4.14 4.43 deviation. The improvement in acceptance limits can be
Downloaded by [University of Nebraska, Lincoln] at 23:30 31 May 2016

95% 99.4% 99% 100 3.55 4.59 4.91 considerable, as shown in Table 2. For sample means be-
95% 99.4% 99% 99 or 101 3.33 4.32 4.61 tween 97 and 103, the Bayesian approach’s acceptance
95% 99.4% 99% 98 or 102 3.12 4.04 4.32
95% 99.4% 99% 97 or 103 2.90 3.76 4.02 limits are 25%–39% greater than those in E2810 for a
90% 98.7% 95% 100 4.39 5.37 5.66 sample of size 10, 12%–21% greater for a sample of size
90% 98.7% 95% 99 or 101 4.13 5.05 5.32 30, and 8%–17% greater for a sample of size 50.
90% 98.7% 95% 98 or 102 3.86 4.72 4.98
Although we focused on the USP <905> UDU test,
90% 98.7% 95% 97 or 103 3.59 4.40 4.63
the Bayesian approach can be used for any test where the
NOTE: LB is not a parameter for the PTIT approach, but is provided as
a link to the acceptance limits for corresponding lot-release procedures simple mean model given in (1) is appropriate, as well as
in Table 1. for more complex models. A useful generalization of (1)
is the balanced one-way random effects model
curves are very similar for μ = 97 and, more generally, for Yi j = μ + Bi + ei j , i = 1, 2, . . . , k,j = 1, 2, . . . , n,
lot means outside the range 98–110. They are less similar (2)
near μ = 100 due to a straight-line approximation to the where the Bi are iid Normal (0, σ B2 ) and the ei j are
curved contours for P (see Figure 4). Acceptance limits iid Normal (0, σe2 ). Section 5.2 in Box and Tiao (1973)
for the PTIT procedure are easy to calculate, making it gives the likelihood function
k for (2) in terms ofthe suffi-
computationally attractive. cient statistics ȳ = i=1 ( nj=1 yi j /n)/k = i=1k
ȳi /k,

Figure 5. A PTIT approximation to the Bayesian procedure’s OC curve forμ = 100 and n = 30.

47
 Statistics in Biopharmaceutical Research: February 2016, Vol. 8, No. 1

 
SS B = n nj=1 ( ȳi − ȳ)2 and SSe = nj=1 (yi j − ȳi )2 . and Content Uniformity. Technical Discussion of Sampling Plans
Hence, the values of individual observations are not and Application of ASTM E2709/E2810,” Journal of Pharmaceuti-
cal Innovation, 10, 84–97. [48]
needed to carry out inference. LeBlond and Mockus
Bergum, J. S. (1990), “Constructing Acceptance Limits for Multiple
(2014) used (2) to model between- and within-lot vari- Stage Tests,” Drug Development and Industrial Pharmacy, 16, 2153–
ability in content uniformity, to assess sample size re- 2166. [41]
quirements for process performance qualification, which ——— (2015), “Tolerance Interval Alternative to ASTM E2709/E2810
is often referred to as PQ or PPQ. Garcia et al. (2014) Methodology to Provide Assurance of Passing the USP Uniformity
recommended using variance components analysis to as- of Dosage Unit (UDU) Test <905>,” Pharmaceutical Engineering,
sess blend and content uniformity; in this regard the ran- 35, 68–79. [46]
dom components in model (2) can be used to represent Bergum, J. S., and Li, H. (2007), “Acceptance Limits for the New ICH
USP 29 Content Uniformity Test,” Pharmaceutical Technology, 31,
between-location and within-location variation. Bergum 90–100. [41]
et al. (2014) applied the approach in ASTM Standards Box, G. E. P., and Tiao, G. C. (1973), Bayesian Inference in Statistical
E2709 and E2810 to the recommendations in Garcia et al. Analysis, Reading, MA: Addison-Wesley. [42,44,47]
to derive acceptance limits for ȳ, SS B , and SSe . We briefly Garcia, T., Bergum, J., Prescott, J., Tejwani, R., Parks, T., and Clark, J.
Downloaded by [University of Nebraska, Lincoln] at 23:30 31 May 2016

examined these acceptance limits and, not surprisingly, (2014), “Recommendations for the Assessment of Blend and Content
found they are conservative. As with model (1), improved Uniformity: Modifications to Withdrawn FDA Draft Stratified Sam-
acceptance limits can be obtained using a Bayesian ap- pling Guidance,” Journal of Pharmaceutical Innovation, 10, 76–83.
[48]
proach that focuses on the probability of passing the USP
Glickman, M. E., and van Dyk, D. A. (2007), “Basic Bayesian Methods,”
<905> UDU test. in Topics in Biostatistics, ed. W. T. Ambrosius, Totowa, NJ: Humana
In addition to generalizing the random-effects portion Press, pp. 324–325. [42]
of the model, one could also generalize the fixed-effects. Hudson-Curtis, B., and Novick, S. (2016), “Content Uniformity,” in
For example, Nonclinical Statistics for Pharmaceutical/Biotechnology Industries,
ed. L. Zhang, Berlin: Springer, pp. 631–651. [46]
Yi (t) = μ (t) + Bi + ei (t) (3) LeBlond, D., and Mockus, L. (2014), “The Posterior Probabil-
ity of Passing a Compendial Standard, Part 1: Uniformity of
could be used to model tablet dissolution, letting the fixed- Dosage Units,” Statistics in Biopharmaceutical Research, 6,
effect μ (t) model dissolution over time with a polyno- 270–286. [45,48]
mial, say, and letting the random effect Bi model tablet- Lewis, R. A., and Novick, S. J. (2011), “A Generalized Pivotal Quan-
to-tablet variation. Note that Yi (t) in (3) might represent a tity Approach to the Parametric Tolerance Interval Test for Dose
transformed response, such as a logistic transformation of Content Uniformity Batch Testing,” Statistics in Biopharmaceutical
percent dissolved, to stabilize the variances of the random Research, 4, 28–36. [46]
effects Bi and ei (t). Lindgren, B. W. (1968), Statistical Theory, New York: Macmillan.
[41]
Lostritto R. (2012), “Content Uniformity (CU) testing for the 21st
Century: CDER Perspective,” in AAPS Annual Meeting, available at
Acknowledgments http://www.fda.gov/downloads/AboutFDA/CentersOffices/Officeof
MedicalProductsandTobacco/CDER/UCM341168.pdf. [45]
The authors thank Buffy Hudson-Curtis and Steven Novick for alerting
Novick, S., Christopher, D., Dey, M., Lyapustina, S., Golden, M., Leiner,
us to the use of parametric tolerance interval tests for lot-release of solid
S., Wyka, B., Delzeit, H. J., Novak, C., and Larner, G. (2009), “A Two
oral dose drug products.
One-Sided Parametric Tolerance Interval Test for Control of Deliv-
ered Dose Uniformity. Part 1—Characterization of FDA Proposed
[Received December 2014. Revised August 2015.] Test,” American Association of Pharmaceutical Scientists Pharm-
SciTech, 10, 820–828. [46]
United States Pharmacopeia General Chapter <905> Uniformity of
References Dosage Units. [40]
United States Pharmacopeia General Notices and Requirements (2010),
Arnold, B. C., and Shavelle, R. M. (1998), “Joint Confidence Sets for United States Pharmacopeial Convention, USP 34, General Notices
the Mean and Variance of a Normal Distribution,” The American 3, Section 3.10. [40]
Statistician, 52, 133–140. [43]
ASTM E2709-12 (2012), Standard Practice for Demonstrating Capa-
bility to Comply with an Acceptance Procedure, West Conshohocken,
PA: ASTM International. Available at www.astm.org [40] About the Authors
ASTM E2810-11e2 (2011), Standard Practice for Demonstrating Ca- Richard A. Lewis is Director of Statistics, Global Manufacturing
pability to Comply with the Test for Uniformity of Dosage Units, West and Supply, GlaxoSmithKline, Five Moore Drive, Research Tri-
Conshohocken, PA: ASTM International. Available at www.astm.org angle Park, NC 27709-3398 (E-mail: [email protected]).
[40] Ailin Fan is PhD Candidate, Department of Statistics, North
Bergum, J., Parks, T., Prescott, J., Tejwani, R., Clark, J., Brown, W., Carolina State University, 2311 Stinson Drive, P.O. Box 8203,
Muzzio, F., Patel, S., and Hoiberg, C. (2014), “Assessment of Blend Raleigh, NC 27695-8203 (E-mail: [email protected]).

48