Statistical Considerations for Establishing

Acceptance Criteria for Content Uniformity

and Stratified Sampling

Alex Viehmann
PQRI
06 October 2015
Disclaimer
This presentation is not official Agency policy. My
statements and advice do not bind or otherwise
obligate or commit the agency to the views expressed.

2
Agenda
• Sampling Methods
• Distribution considerations
• Quality levels – producer / consumer risk
• USP <905> - what does it provide?
• Acceptance criteria design

3
Sampling Methods
• Simple Random Sampling
– Randomly pick units from lot
• Stratified Sampling
– Divide Lot into strata where the combined strata
cover the entire lot
– Use random sampling within each strata.
• Systematic Sampling
– Sample at specified points usually equally spaced
throughout the lot.
• Random vs. Stratified/Systematic
– Random = no estimation of between/within location
variability
 Comparison of Sampling Methods
•CU(%LC)
Distribution considerations

• What is the underlying distribution?

– Probability plot
– Normality test

Note: Statistical tests are

sensitive to high sample
sizes.

6
Parametric vs Distribution-free
• Parametric: Assume the data follow a specific distribution
– Example: Not enough evidence to conclude the distribution is non-
normal
• Distribution-free: Do not assume the data follow a specific
distribution
• Parametric tests can perform well with skewed and non-
normal distributions and when the spread of each group is
different
– Follow sample size guidelines
• Limited knowledge/data to confirm distribution or severely
skewed distributions (best represented by median) =
distribution-free techniques

7
Quality Levels
Acceptable: The quality level that would be
accepted with a high probability (e.g. 95%)
Producers risk

Unacceptable: The quality level that would be

rejected with a high probability (e.g. 90%)
Consumers risk

21 CFR 210.3(20): …. Acceptance criteria means

product specifications and acceptance criteria, such
as acceptable quality level and unacceptable
quality level, with an associated sampling plan….
USP <905> GC for demonstrating Uniformity of
Dosage Units
Table 1. Uniformity of Dosage Units Test Procedure
All measurements of dosage units and criteria values are in percentage label claim
(%LC).
At each stage calculate the sample average, X , and the sample standard deviation s.
Stage Number tested Pass stage if:
S1 10 |M -X| + 2.4s ≤ 15.0, where M is defined below.

i) |M -X| + 2.0s ≤ 15.0 using all 30 results (S1 + S2)

S2 20
ii) No dosage unit is outside the maximum allowed range of 0.75*M
to 1.25*M.

M is defined as follows:
If T is less than or equal to 101.5%LC, and
(i) If X is less than 98.5%LC, then M = 98.5%LC.
(ii) If X is between 98.5 and 101.5%LC, then M = X .
(iii) If X is greater than 101.5%LC, then M = 101.5%LC.
If T is greater than 101.5%LC, and
(i) If X is less than 98.5%LC, then M = 98.5%LC.
(ii) If X is between 98.5 and T, then M = X .
(iii) If X is greater than T, then M = T.
T is the Target content per dosage unit at the time of manufacture, expressed as percentage label
claim. Unless otherwise specified in the individual monograph, T is 100.0%LC.

9
USP <905> background
• The procedure is based on a two-sided tolerance interval
approach.
– An interval that contains p percent of the population
measurements
• N=10 / k2=2.4
• N=30 / k2=2.0
– k2 is a tolerance interval factor that is affected by
sample size, desired confidence, and coverage (k2 is
specific to a two sided tolerance interval)
– K2 is determined so that the interval covers at least a
proportion p of the population with a confidence c
• Based upon the criteria of the test (k2 value), the metrics
provided are – 84% confident that 91% of the population lies
between +/- 15%
• The second aspect of the procedure is no tablet will be outside ~
73.9-126.9%

10
USP <905> - Acceptance Region / OC curve

Acceptance Limits

8
7
6
5
Max SD

n=30
4
3 n=10
2
1
0
117

115
113

111
109

107
105
103
102

99.5

98.5
97

95
93

91
89

87
85
83.5
Mean

11
Blend Uniformity
• Draft guidance acceptance criteria was inconsistent
with current Agency thinking
– Old recommendation: sample multiple units (replicates) from
multiple locations (e.g. 3 samples from 10 locations) and only
assay 1 per location = total of 10 measurements
• Mean (90-110%) / SD < 5%
– Level II Q/A recommendation: assay all replicates
• Allow for estimation of between/within location variance
components

http://www.fda.gov/Drugs/GuidanceComplianceRegulatoryInforma
tion/Guidances/ucm124782.htm

12
Variance Components
• Partitions total variability
into its individual sources
of variation
• Examples:
– Between/within locations
– Between/within batches

13
Designing appropriate acceptance criteria
• What is the sampling method?
– Random, stratified or systematic

• What is the underlying distribution?

• What are the quality levels that are required for this
product/process?

14
Example I
• Sampling Method = Systematic
• Data = Continuous / Normal
– Sample size = 60 (15x4)
• Assurance requirements = 90% confident that at least
95% of samples will meet USP standard
– Coverage = true proportion of samples from the batch that
will meet USP <905>
• Method = ASTM E2709 / 2810; sampling plan 2

15
Example: Sampling Plan 2 (15 X 4) - Data
Result
Location 1 2 3 4
1 97.08 99.72 98.37 97.50
2 99.72 100.32 101.01 100.29
3 99.90 98.27 98.88 97.96
4 98.78 98.17 98.94 97.78
5 96.32 96.61 99.66 97.20
6 100.97 102.17 99.06 98.80
7 97.02 97.35 98.65 99.98
8 99.39 98.81 98.63 98.06
9 99.59 97.80 97.67 98.95
10 97.97 98.54 100.26 98.74
11 96.09 98.61 97.49 97.50
12 98.87 97.81 97.28 98.80
13 101.10 102.60 100.48 98.62
14 100.80 100.34 98.49 100.93
15 99.70 100.09 100.14 99.20
Example: Sampling Plan 2
Example: Sampling Plan 2
Descriptive Statistics
Overall Mean 98.93
SE (within-location Std Dev) 1.07
Standard deviation of Location Means 1.06

90%CI/95%Cov Standard Deviation of Location Means

0.9 1.0 1.1 1.2

SE LL UL LL UL LL UL LL UL

0.9 88.1 111.9 88.5 111.5 88.9 111.1 89.3 110.7

1.0 88.2 111.8 88.6 111.4 89.0 111.0 89.4 110.6

1.1 88.4 111.6 88.7 111.3 89.1 110.9 89.5 110.5

1.2 88.5 111.5 88.9 111.1 89.2 110.8 89.6 110.4

1.3 88.7 111.3 89.0 111.0 89.4 110.6 89.7 110.3

Example II
• Sampling Method = Random
• Data = Distribution unknown; attribute data
– Attribute: 85-115% LC
• Assurance requirements = 90% confident that at least
98% batch will be in spec.
– Coverage = true proportion of batch within desired spec
• Method = Attribute sampling plan
Example II: OC curve

AQL = 0.25% (95% probability the lot will be accepted if % defective ≤ 0.25%)
UQL = 2% (90% probability the lot will be rejected if % defective ≥ 2%)

20
Example III
• Sampling Method = Random
• Data = Continuous / Normal
– Spec: 90-110%
• Sample size = 60
• Assurance requirements = 95% confident that at
least 95% batch will be in spec.
– Coverage = true proportion of batch within desired spec
• NMT 2.5% below 90%; NMT 2.5% above 110
• Method = Two one-sided Tolerance Interval test

21
Example III acceptance criteria
• Xbar + (2.38*s) < 110
• Xbar – (2.38*s) > 90

2.38 = one-sided tolerance interval

factor representing 95% confidence,
97.5% coverage and sample size=60

22
Summary
• Sampling method needs to be defined
– Dictates the ability to quantitate between/within variance
components
• Distribution needs to be evaluated
– Dictates the method to apply
• Attribute (spec range) needs to be relevant for the
product
– Is 85-115% appropriate for all drug products?
• Quality levels = Risk Based
– Should not be static across all products
• Firm needs to clearly identify the assurance the plan
provides
– OC curve
23