Statistical Mixing

Statistical Methods for
Assessment of Blend
Homogeneity
Camilla Madsen
LYNGBY 2002
IMM-PHD-2002-99
ATV Erhvervsforskerprojekt EF 767
IMM
IMM-PHD-2002-99
ATV Erhvervsforskerprojekt EF 767
ISSN: 0909-3192
ISBN: 87-88306-15-1

c Copyright 2002 by Camilla Madsen.
This document was prepared with LATEX and printed by DTU-tryk.

Preface
This thesis has been prepared at the department Informatics and Mathematical
Modelling (IMM), Technical University of Denmark (DTU), and at the phar-
maceutical company Novo Nordisk A/S in partial fulfilment of the require-
ments for the industrial Ph.D. degree within the Mathematical Ph.D. Program
at DTU.
The thesis is concerned with statistical methods to assess blend uniformity in

tablet production.
Lyngby, May 2002.
Camilla Madsen
iii
iv
Acknowledgements
First of all, I wish to thank my academic supervisor, Prof. Poul Thyregod,

IMM, DTU for many hours of discussions and support. His guidance has been
crucial for the outcome of this project and for that I am very grateful. I also
want to thank my three industrial supervisors at Novo Nordisk A/S, Jørgen Iw-
ersen, Quality Support Statistics, Per Grønlund, SDF Pilot Plant and Charlotte
Tvermoes Rezai, SDF Production, for their help and for our enlightening dis-
cussions throughout the project. Further I would like to thank Jørgen Iwersen
for taking the initiative to this project and for making an effort on finding the
right collaborators, and Per Grønlund and Charlotte Tvermoes Rezai for their
help and willingness to always answer my questions regarding pharmaceutical
matters. They also deserve thanks for their effort in trying to find the time for
experiments in a very tight production schedule.
I also would like to thank colleagues in the Quality Support Statistics depart-
ment, Novo Nordisk A/S and at IMM, DTU for a good and pleasant scientific
and social environment. Águsta Haflidadottir as well as my two officemates
Dorte Rehm and Dorte Vistisen deserves special acknowledgements for many
pleasant and constructive hours.
Last but not least I am grateful to my fiancee, my family and my friends for
their support, patience and encouragement during the hard parts of this work.
v
vi
Summary
In this thesis the use of various statistical methods to address some of the prob-
lems related to assessment of the homogeneity of powder blends in tablet pro-
duction is discussed.
It is not straight forward to assess the homogeneity of a powder blend. The

reason is partly that in bulk materials as powder blends there is no natural unit
or amount to define a sample from the blend, and partly that current technol-
ogy does not provide a method of universally collecting small representative
samples from large static powder beds.
In the thesis a number of methods to assess (in)homogeneity are presented.

Some methods have a focus on exploratory analysis where the aim is to in-
vestigate the spatial distribution of drug content in the batch. Other methods
presented focus on describing the overall (total) (in)homogeneity of the blend.
The overall (in)homogeneity of the blend is relevant as it is closely related to
the (in)homogeneity of the tablets and therefore critical for the end users of the
product.
Methods to evaluate external factors, that may have an influence on the content
in blend samples, as e.g. sampling device, have been presented. However,
the content in samples is also affected of internal factors to the blend e.g. the
particle size distribution. The relation between particle size distribution and
the variation in drug content in blend and tablet samples is discussed.
A central problem is to develop acceptance criteria for blends and tablet batches
vii
viii
to decide whether the blend or batch is sufficiently homogeneous (uniform) to

meet the need of the end users. Such criteria are most often criteria regarding
sample values rather than criteria for the quality (homogeneity) of the blend
or tablet batch. This inherently leads to uncertainty regarding the true quality
of a specific blend or batch. In the thesis it is shown how to link sampling
result and acceptance criteria to the actual quality (homogeneity) of the blend
or tablet batch. Also it is discussed how the assurance related to a specific
acceptance criteria can be obtained from the corresponding OC-curve.
Further, it is shown how to set up parametric acceptance criteria for the batch
that gives a high confidence that future samples with a probability larger than
a specified value will pass the USP three-class criteria.
Properties and robustness of proposed changes to the USP test for content uni-
formity are investigated by the use of simulations, and single sampling accep-
tance plans for inspection by variables that aim at matching the USP proposal
have been suggested.
Resumé (in Danish)
Denne afhandling omhandler brugen af statistiske metoder til at belyse forskel-

lige problemstillinger i forbindelse med vurdering af homogeniteten af en pul-
verblanding i tabletfremstilling.
Det at bestemme homogeniteten af en pulverblanding er ikke simpelt. Det

skyldes dels at bulkmaterialer som pulverblandinger ikke indeholder en naturlig
enhed eller mængde, der kan afgrænse en prøve fra blandingen og dels at der
med den nuværende teknologi ikke findes en universel metode til at indsamle
små repræsentative prøver fra store statiske pulverblandinger.
I afhandlingen er forskellige metoder til at vurdere homogenitet beskrevet.

De første metoder kan anvendes i forbindelse med explorative undersøgelser,
hvor formålet er at undersøge fordelingen af aktivt stof i blandingen. De sid-
ste metoder har til formål at beskrive den overordnede (totale) homogenitet i
blandingen. Den overordnede homogenitet i blandingen er relevant, da den
har betydning for homogeniteten af tabletterne, og derfor er den kritisk for de
endelige forbrugere af tabletterne.
Metoder til at vurdere ydre faktorers betydning for indholdet af aktivt stof i
prøver fra blandingen er blevet diskuteret. En ydre faktor kan f.eks. være det
redskab, prøverne udtages med. Indholdet af aktivt stof i prøver fra blandingen
samt i tabletterne afhænger også af indre faktorer som f.eks. partikel størrelses-
fordelingen. Sammenhængen mellem partikel størrelsesfordeling og variation
i indholdet af aktivt stof i prøver fra blandingen og tabletbatchen er blevet
diskuteret.
ix
x
Et central problem er at opstille accept kriterier for blandinger og tablet batche,

der sikrer at homogeniteten er tilfredsstillende i forhold til de endelige for-
brugeres behov. Sådanne accept kriterier bliver ofte formuleret som krav til
resultatet af stikprøven i stedet for mere direkte som krav til kvaliteten (ho-
mogeniteten) af blandingen eller tabletbatchen. Sådanne krav fører nødvendigvis
til usikkerhed angående den sande kvalitet af den enkelte blanding eller tablet-
batch. I afhandlingen er sammenhængen mellem på den ene side stikprøvere-
sultat og accept kriterium og på den anden side kvaliteten (homogeniteten) af
en blanding eller tablet batch beskrevet. Derudover diskuteres det hvordan,
den sikkerhed, der opnås ved et specifikt accept kriterium, kan udledes fra den
tilsvarende OC-kurve.
Det vises, hvordan et parametrisk accept kriterium for blandingen eller tablet-
batchen, der giver en fastlagt (stor) sikkerhed for at fremtidige stikprøver vil
have mindst en fastlagt sandsynlighed for godkendelse under et USP three-
class krav.
Egenskaber og robusthed ved ændringsforslag til USPs test for Content Unifor-
mity er belyst ved simuleringer, og enkeltprøvningsplaner for inspektion ved
kontinuert variation, der tilstræber at matche USP-forslaget er blevet forslået.
Contents
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . vi
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix
Resumé (in Danish) . . . . . . . . . . . . . . . . . . . . . . . . . . xi
I 1
1 Background 3
1.1 Outline of the Thesis . . . . . . . . . . . . . . . . . . . . . . 5
2 Introduction 7
2.1 Principles of tablet production . . . . . . . . . . . . . . . . . 7
2.2 Assessment of the uniformity of the blend . . . . . . . . . . . 10
2.3 Regulatory Affairs . . . . . . . . . . . . . . . . . . . . . . . 11
xi
xii CONTENTS
2.3.1 Organizations . . . . . . . . . . . . . . . . . . . . . . 13
2.3.2 Requirements and Recommendations . . . . . . . . . 15
3 Results and discussion 17
3.1 Variances as a measure of homogeneity . . . . . . . . . . . . 18
3.2 Methods to assess homogeneity and factors that may influence

homogeneity . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.2.1 The effect of particle size distribution . . . . . . . . . 22
3.2.2 Assessment of homogeneity in specific batches . . . . 23
3.2.3 Example . . . . . . . . . . . . . . . . . . . . . . . . 25
3.3 Analysis of acceptance criteria . . . . . . . . . . . . . . . . . 26
4 Conclusion 31
II Included papers 35
A Robustness and power of statistical methods to assess blend

homogeneity 37
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2 Models of batch homogeneity . . . . . . . . . . . . . . . . . 40
2.1 The aggregated model . . . . . . . . . . . . . . . . . 40
2.2 The hierarchical model . . . . . . . . . . . . . . . . . 41

CONTENTS xiii
2.3 Simulated samples . . . . . . . . . . . . . . . . . . . 44
3 Assessing factors with influence on the mean content of the

active component in a sample . . . . . . . . . . . . . . . . . . 45
3.1 Large and medium scale homogeneity assessed accord-

ing to the aggregated model (1) . . . . . . . . . . . . 46

ing to the hierarchical model (2) . . . . . . . . . . . . 48
3.3 The effect of including sampling thieves in the model . 50
3.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . 54
4 Assessing factors with influence on the variation between repli-

cates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.1 Conclusion . . . . . . . . . . . . . . . . . . . . . . . 59
5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
B Comprehensive measures of blend uniformity 63
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
2 Batches with medium scale variation (variation between areas) 68
2.1 The total variation from the ANOVA table . . . . . . . 68
2.2 The variance on a randomly sampled unit from the batch 69
3 Batches with large scale variation (variation between layers) . 71
3.2 The variance on a sample from the batch . . . . . . . 75

xiv CONTENTS
4 Batches with both large and medium scale variation . . . . . . 75
4.2 The variance on a sample from the batch . . . . . . . 82
4.3 The direct relation between large/medium scale varia-

tion and the acceptance criteria . . . . . . . . . . . . . 82
5 Discussion and conclusion . . . . . . . . . . . . . . . . . . . 82
C On a test for content uniformity in pharmaceutical products

Presented at the First Annual ENBIS Conference, Oslo 2001 87
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
2 The proposed test . . . . . . . . . . . . . . . . . . . . . . . . 90
2.1 Historical notes . . . . . . . . . . . . . . . . . . . . . 90
2.2 Description of the proposed test . . . . . . . . . . . . 91
3 Properties of the proposed test . . . . . . . . . . . . . . . . . 93
3.1 Description of the OC-surface of the test . . . . . . . 94
3.2 “Specification limits” for individual tablets . . . . . . 95
3.3 Details on the effect of individual elements of the test . 97
4 Comparison to other test procedures . . . . . . . . . . . . . . 101
5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
CONTENTS xv
D Statistical tests for uniformity of blend/dose samples 105
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
2 Acceptance criteria and statistical hypothesis testing . . . . . . 111
2.1 Choice of null hypothesis and alternative hypothesis . 111
2.2 Confidence intervals and statistical tests . . . . . . . . 113
3 Notation and distributional assumptions . . . . . . . . . . . . 114
4 Acceptance criteria for the dispersion of doses . . . . . . . . . 116
4.1 Criterion based upon a specified limiting value of sam-

ple standard deviation . . . . . . . . . . . . . . . . . 116

ple coefficient of variation . . . . . . . . . . . . . . . 118
4.3 Criterion based upon prediction of standard deviation

of future samples . . . . . . . . . . . . . . . . . . . . 123
4.4 A direct approach in terms of population values . . . . 129
5 Acceptance criteria with limits on individual measurements . . 133
5.1 The USP 21 criteria . . . . . . . . . . . . . . . . . . . 133
5.2 Three-class attributes and a parametric approach . . . 136
5.3 Confidence region approach . . . . . . . . . . . . . . 144
5.4 Relation to theories of acceptance sampling by variables 145
5.5 Design of a test with a trapezoidal acceptance region . 148
5.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . 159

xvi CONTENTS
6 Assessment of the properties of the USP preview dosage uni-

formity test . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
6.1 The acceptance value . . . . . . . . . . . . . . . . . . 162
6.2 The simulation study . . . . . . . . . . . . . . . . . . 163
6.3 Robustness against deviation from distributional assump-

tions . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
6.4 Equivalent single sampling plan . . . . . . . . . . . . 175
7 Further issues . . . . . . . . . . . . . . . . . . . . . . . . . . 177
8 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
9 List of symbols . . . . . . . . . . . . . . . . . . . . . . . . . 181
E On particle size distributions and dosage uniformity for low-

dose tablets 183
1 Introduction and summary . . . . . . . . . . . . . . . . . . . 185
2 Lognormal distribution of particle radii . . . . . . . . . . . . 186
2.1 Distribution of particle radii . . . . . . . . . . . . . . 187
2.2 Distribution of particle mass for spherical particles . . 188
3 Distribution of dose content under random mixing of particles 193
3.1 Modelling random mixing . . . . . . . . . . . . . . . 193
3.2 Constant number of particles in tablets . . . . . . . . . 193
3.3 Random variation of number of particles in tablets . . 195

CONTENTS xvii
3.4 Minimum number of particles necessary to secure a

specified dose coefficient of variation . . . . . . . . . 200
3.5 Coefficient of skewness and excess for distribution of

dose content . . . . . . . . . . . . . . . . . . . . . . 202
4 Distribution of dose content under non-random mixing . . . . 205
4.1 Spherical particles . . . . . . . . . . . . . . . . . . . 209

specified dose coefficient of variation . . . . . . . . . 210
5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
6 List of symbols . . . . . . . . . . . . . . . . . . . . . . . . . 214
F Case: Analysis of homogeneity in production scale batches 217
1 Purpose . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
2 Experimental . . . . . . . . . . . . . . . . . . . . . . . . . . 219
3 Statistical Analysis . . . . . . . . . . . . . . . . . . . . . . . 221
3.1 Assessment of tablet samples . . . . . . . . . . . . . . 222
3.2 Repeatability . . . . . . . . . . . . . . . . . . . . . . 224
3.3 Mean content of the active component . . . . . . . . . 228

xviii CONTENTS
I
Part I
1
Chapter 1
Background
The background for the Ph.D. project is the general requirements to the phar-
maceutical industry of scientific based documentation of the methods used for
validation of processes.
To meet the requirements extensive validation testing should be performed at

various stages of the manufacturing process to show that various unit opera-
tions accomplish what they are supposed to do. The validation testing in the
pharmaceutical industry is especially strict compared to requirements in most
other industries because failure of meeting a high standard for pharmaceutical
products could lead to quite grave consequences.
The pharmaceutical process under consideration in this thesis is the production

of tablets which is not an inconsiderable part of pharmaceutical production as
an estimated 80% of pharmaceutical products are tablets [1].
Tablets are compacts of powders. Essentially tablets are produced by blending

the powdered ingredients until satisfactory uniformity is obtained. Then the
tablets are compressed from the powder blend. Hence a critical unit operation
in the manufacturing of tablets is the mixing of the final blend. Poor blending
or the inability to maintain a blend, i.e. segregation, will inherently lead to
problems with the drug content of the tablets compressed from the blend. This
is costly in terms of rejected material, extra blending time, and defective end
3
4 C HAPTER 1. BACKGROUND
products. See [2] for a detailed overview over the compliance and science of
blend uniformity analysis.
Blend uniformity can be validated by sampling a number of samples from the

blend. If the content of the active ingredient in these samples conform to the
relevant acceptance criteria the blend is accepted.
Throughout the pharmaceutical industry process validation programs for the

manufacturing of tablets have been influenced by the Wolin decision in the
U.S. vs. Barr Laboratories [3]. Judge Alfred Wolin defined some of the CGMP
(Current Good Manufacturing Practice) requirements for process validation of
oral solid dosage forms in greater detail than specified in 21 CFR Part 211 [4].
Particularly it was ruled that the appropriate sample size for content uniformity
testing of the final blend in validation and ordinary production batches is up to
three times the run weight of the finished product. Larger sample sizes increase
the risk of masking insufficient homogeneity on a tablet scale.
The decision caused FDA (The Food and Drug Administration) to reexamine
and modify its policies on blend uniformity and sampling techniques. The
resulting policies are based on the assumption that current technology provides
a means to consistently collect minute representative samples from much larger
static powder beds.
However, limitations in the sampling technology makes it difficult to apply

scientifically valid methods to blend uniformity validation, because current
technology does not provide a method of universally collecting small repre-
sentative samples from large static powder beds. The problem is the potential
for sampling bias. As a result the mean and/or variation between the content in
the samples may be significantly different from the mean content / the variation
in the blend.
At the moment the tablet press could be viewed as the ultimate sampling de-
vice, because the whole batch is being sampled at this stage of the process.
However it is not allowed solely to rely on this when demonstrating blend uni-
formity.
1.1. O UTLINE OF THE T HESIS 5
1.1 Outline of the Thesis
This thesis is organized into two parts. The first part is four chapters and con-
tains a description of tablet production and the relevant regulatory affairs as
well as a discussion of the results presented in part two.
Part two is six appendices with five articles and manuscripts to articles as well
as a case studies. The appendices deals with different aspects of assessment of
blend homogeneity and as such they represent the main part of the thesis.
In Appendix A statistical methods to assess blend homogeneity and factors

that may have an influence on homogeneity are presented. The focus is on
exploratory analysis of blend homogeneity. Appendix F contains a case study
using these methods.
In Appendix B two methods to assess the overall (total) variation in the blend
are discussed.
The effect of the particle size distribution on the distributed content in blend
and tablet samples are discussed in Appendix E.
Finally acceptance criteria for blend and tablet samples are discussed in Ap-
pendix C and Appendix D.
The appendices should be read together with the discussion in Chapter 3 as the
results are put in a larger perspective in this chapter.
6 C HAPTER 1. BACKGROUND
Chapter 2
Introduction
The technical area of this thesis is applied statistics and the field of application
is the pharmaceutical industry, specifically tablet production. Thus, the content
of the thesis is in the borderland between the areas of applied statistics and
pharmaceutical science.
For the reader with a non-pharmaceutical background the principles of tablet

production and some relevant concepts are briefly introduced in the following.
2.1 Principles of tablet production
An example of a tablet production is shown schematically in Figure 2.1.
A tablet consists of one or more active ingredients; the drug substance, and
some filling materials which have the main purpose to give the tablet suitable
physical, biological and chemical properties. For example assure that the drug
is released after a certain amount of time in the body or to assure the breaking
strength so the tablet does not break into pieces before it is consumed by the
patient.
All raw materials are in a powdered form. As the particle size distribution is
7
8 C HAPTER 2. I NTRODUCTION
Raw materials
Sieving
Blending
Granulation
Sieving
Compression
Figure 2.1: Flowchart of tablet production phases.
very important for a number of tablet properties as well as for the variation in
content in the tablets the raw materials are initially sieved to eliminate lumbs.
Then the raw materials are mixed in a blender until the blend is considered
homogeneous. Many different types of blenders exists. The differences may
e.g. be due to the physical presentation or due to the mechanical principles
used. Further, some blenders may also be used for an eventual granulation of
the blend. The purpose of granulation is to obtain particles of more uniform
size. This can be done either by breaking larger particles but most often by
combining smaller particles to larger particles.
Granulation of the blend is not always necessary, however depending on the

properties of the blend granulation may result in improved floating properties
which is important at the tablet press. Further, by granulation the state of blend
is partly fixed thus reducing the risk of deblending, etc.
The granulate may be sieved again to eliminate lumbs. Then the blend is com-
pressed at the tablet press. Some products are produced by direct compression
of the blend without granulation.
2.1. P RINCIPLES OF TABLET PRODUCTION 9
The process in the tablet press is in principle a successive series of dosings

of the blend or granule to the die in the tablet press, whereupon each dose is
compressed to a tablet by compaction between two punches.
Figure 2.2: Tablet press
Figure 2.2 shows a row of punch stations on the tablet press. A and B are upper
and lower stamp respectively. C is the place where the blend is led to the press.
Beneath this place the matrix (D) passes at the same time as the dosing takes
place. Then the filled matrices and the corresponding punches passes two rolls
(E and F) that presses the punches together. Immediately after this the tablet is
ejected and pushed away.
After the compression at the tablet press the tablets may be coated for example
to protect the active ingredient from decomposition due to light or moisture or
to facilitate packaging, mask unpleasant taste or smell etc.
2.2 Assessment of the uniformity of the blend
The content uniformity of the tablets is an important quality measure of the

final product. As the content uniformity is closely related to the uniformity
of the blend, it is important to monitor blend uniformity. Inhomogeneities of
a blend can either be due to insufficient mixing or to deblending under trans-
portation or storage of the blend.
In practice blend uniformity is assessed by collecting a number of samples

from the blend, each sample being of the size of 1-3 times the corresponding
tablets. The sampling locations must be carefully chosen to provide a rep-
resentative cross-section of the granulation. The resulting samples are then
assayed using the same methods used to analyze the finished product. Con-
tent uniformity is established if the drug content of the samples conform to a
predetermined criterion.
The current state of the art regarding sampling technology is a device referred
to as a sampling thief. Many different types of sampling thieves have been
developed. However, in general a sampling thief consists of two concentric
tubes. The inner tube is solid except for one or more chambers that allow for
sample collection. The outer tube is hollow and contains openings that can
align with the chambers in the inner tube. A handle, located at the top of the
thief is used to rotate the inner tube within the outer tube in order to open or
close the thief. A sample is collected by inserting the closed thief into the
blend. Then the handle is rotated in order to open the thief allowing the sample
to flow into the sampling chamber in the inner tube. Then the thief is closed
and pulled out of the blend.
Figure 2.3 shows examples of two different types of sampling thieves. In these
thieves the sampling chamber is located at the tip of the thief rather than on the
side as described above.
However, the two “Golden Rules” of sampling [5] states that a powder should
be sampled when in motion and that the whole of the stream of powder should
be taken for many short increments of time. Any sampling methods which does
not adhere to these rules should be regarded as a second-best method liable to
lead to errors.
Collecting blend samples by the use of sampling thieves violates these Golden
2.3. R EGULATORY A FFAIRS 11
Figure 2.3: Two examples of sampling thieves. In these thieves the sampling
chamber is located in the tip of the thief.
Rules, and the result is the risk of sampling error or bias. The presence and
the size of such sampling errors and bias depends on factors such as sampling
device, sampling technique, blend formulation, blender size, sample location
and size of the collected sample. For a more detailed discussion of these factors
see e.g. [2].
Figure 2.4 shows a boxplot for the results of samples from blend and the cor-
responding tablets for seven batches produced at Pilot Plant, Novo Nordisk
A/S. Each batch corresponds to a different value of label claim (LC). It is seen
that for all batches the mean content in blend samples is larger than the mean
content in the tablet samples. This could be a typical result of sampling bias.
2.3 Regulatory Affairs
Pharmaceutical companies often sell their products in several parts of the world.
In order to do that the companies have to comply with all the requirements
covering the countries or areas in which they sell their products. In order to
reduce the need to duplicate the testing carried out during the research and de-
velopment of new medicines efforts are done at making greater harmonisation
in the interpretation and application of technical guidelines and requirements
for product registration. This work is particularly organized via The Interna-
tiona Conference on Harmonisation, ICH, which is a joint initiative between
Figure 2.4: Samples from blend and tablets.

the regulatory authorities of Europe, Japan and the United States and experts
of the pharmaceutical industry in the three regions. Until this hamonization is
completed the pharmaceutical companies have to comply with various sets of
requirements and in this situation the American legislation is very important
because for many pharmaceutical companies the American market is one of
the most important markets. Therefore the focus in this thesis is on rules and
requirements in the American legislation.
2.3.1 Organizations
The main actors in the U.S. are the FDA, USP, PDA and PQRI. In the following
these organizations are introduced.
The Food and Drug Administration, FDA
Most countries have a governmental drug administration which approves drug

products. In the U.S. the drug administrative organ is the FDA. FDA is an
agency, charged with protecting American consumers by enforcing the Federal
Food, Drug, and Cosmetic Act and several related public health laws. Among
other things it monitors the manufacture, import, transport, storage and sale of
medicines and medical devices.
In deciding whether to approve new drugs, FDA does not itself do research,
but rather examines the results of studies done by the manufacturer. A part of
this investigation is to assess whether the new drug produces the benefits it is
supposed to without causing side effects that would outweigh those benefits
[6].
The United States Pharmacopeia, USP
USP is the American pharmacopoeia responsible for developing public stan-

dards and information concerning public health.
In pursuit of its mission to promote public health, USP establishes standards to

ensure the quality of medicines for human and veterinary use. Manufacturers
must meet these standards to sell their products in the U.S. The standards are
officially recognized standards of quality and authoritative information for the
manufacturing and use of medicines and other health care technologies [7].
Parental Drug Association, PDA
PDA is a non-profit international association involved in the development,

manufacture, quality control and regulation of pharmaceuticals and related
products. PDA is a leading technical organization in the fields of parental sci-
ence and technology that tries to influence the future course of pharmaceutical
products technology.
The mission is to support the advancement of pharmaceutical technology by

promoting scientifically sound and practical technical information and educa-
tion for industry and regulatory agencies [8].
The Product Quality Research Institute, PQRI
PQRI is designed to provide a neutral environment where FDA, academia and

industry can collaborate on pharmaceutical product quality research and de-
velop information in support of policy relating to regulation of drug products.
PQRI supports the priorities of FDA to improve and enhance its science base
and provides scientific evidence for policy enactment or changes. PQRI also
serves the pharmaceutical industry by promoting efficiency and consistency in
the regulatory processes.
A number of working groups are established. The ultimate goal of these work-
ing groups is to develop scientific knowledge that will result in appropriate
changes to regulatory policies to make them less burdensome [9].
2.3.2 Requirements and Recommendations
A very important document is 21 Code of Federal Regulations. 21 CFR is

a very general law describing current good manufacturing practice (CGMP)1 .
Of special interest is 21 CFR Part 210 and Part 211 [4] describing respectively
processing, packing, or holding of drugs (part 210) and for finished pharma-
ceuticals (part 211).
21 CFR is published by FDA. Pharmaceutical companies on the American

market have to comply with this law. As the law is very general it does not
give many specific technical details on how to comply with the law. Some of
these details are found in the current American pharmacopoeia, USP 24. As an
example the pharmacopoeia specify how to perform content uniformity testing,
i.e. how to test the uniformity of tablets. Also a lot of guidance documents
and guidelines on various topics are published by FDA. The content of these
documents are not directly ’law’ but they contain detailed information on how
FDA interpret the law and more detailed suggestions and recommendations
on what the manufactures can do if they want a drug to be approved. As an
example [11] gives guidelines on blend uniformity testing. For a more detailed
description of these documents see e.g. [2].
In 1996 FDA proposed some changes to 21 CFR Part 210 and Part 211. Re-
garding blend uniformity testing the most important change is a new paragraph
211.110(d) that specifically require blend samples to approximate the dosage
size of the product for blend uniformity analysis. Thus, this proposed amend-
ment would for the first time legally oblige the pharmaceutical industry to con-
duct blend uniformity analysis using unit dose testing.
1
CGMP regulations are based on fundamental concepts of quality assurance: (1) Quality,
safety, and effectiveness must be designed and built into a product; (2) quality cannot be in-
spected or tested into a finished product; and (3) each step of the manufacturing process must
be controlled to maximize the likelihood that the finished product will be acceptable. [10]
Chapter 3
Results and discussion
With a background in the legal requirements for the pharmaceutical industry

to validate critical unit operations as for example the mixing of the final blend
in the tablet production, this thesis addresses some of the problems related to
assessing the homogeneity in powder blends.
Before starting the production of a new product or changing an existing blend-

ing or blend sampling process it is important to investigate factors that may
have an influence on the processes. For this kind of exploratory investigations
it is meaningful not just to evaluate the overall homogeneity but to consider
homogeneity on different scales in the blend. More specific in this thesis the
homogeneity is evaluated on a large, a medium and a small scale. Such an eval-
uation on more than one scale will enhance the understanding of the processes.
Statistical methods to assess blend homogeneity on different scales and to eval-
uate factors that have a possible influence on the homogeneity are presented in
Appendix A. An example of an explorative analysis is given in Appendix F.
Even though the number of actually conducted experiments in this example
was smaller than originally planned and therefore the resulting design is not
’optimal’ for the statistical methods used this experiment has been chosen as
an example, as it includes both blend and tablet samples. Comparing blend
and tablet samples is a more holistic approach than analysing blend and tablet
results separately. The example should be seen as a guidance on considerations
and conclusions with relevance for this type of analysis.
17
18 C HAPTER 3. R ESULTS AND DISCUSSION
Regarding the patients using the final tablets it is of less relevance if the varia-
tion between the doses is due to large, medium or small scale variation in the
blend. In this relation the magnitude of the total variation in the batch of tablets
is relevant. The total variation between the content of the tablets is closely re-
lated to the total variation in the blend. Therefore for practical purposes it
is relevant to control the total variation in the blend. In Appendix B the three
scales of homogeneity discussed in Appendix A are related to overall measures
of blend homogeneity. The measures of overall homogeneity are compared by
relating them to an acceptance criterion for blend uniformity.
Acceptance criteria for both blend and tablets are usually assessed assuming
a normal distribution of content in the samples. However, actual distributions
of particle sizes are often seen to be skewed. This might have an effect on the
shape of the distribution of content in blend and tablet samples. Therefore, in
Appendix E the effect of a skewed particle size distribution on the distribution
of content in the samples is discussed.
Keeping in mind, that for example a skewed particle size distribution can in-
fluence the distribution of the content in the blend and tablet samples, the
statistical properties of acceptance criteria for blend and tablet samples are
discussed under the normal assumption in Appendix C and Appendix D. Ap-
pendix C gives background and preliminary considerations to the analysis in
Appendix D. Further, the acceptance criteria analysed in Appendix C is an ear-
lier version of the corresponding acceptance criteria analysed in Appendix D.
In the following the results and discussions of these are given in more detail.
3.1 Variances as a measure of homogeneity
It comes natural to think of homogeneity as some kind of variance being small.

However, even though variation is an often used parameter in various relations
it is not straight forward in case of bulk materials to define homogeneity as
a variance. The reason is that bulk materials essentially are continuous and
do not consist of discrete, identifiable, unique units or items, i.e. there is no
natural unit or amount of material that may be drawn into the sample [12].
A single particle is not a suitable unit as it is to small for practical purposes.

3.1. VARIANCES AS A MEASURE OF HOMOGENEITY 19
Rather, the ultimate sampling unit must be created, at the time of sampling, by
means of some sampling device. The size and form of the units depend on the
particular device employed, how it is used, the nature, condition, and structure
of the material, and other factors.
However, this definition of a unit is convenient and conceptual and further for
practical purposes the size of a sample do not differ much from the size of a
tablet produced from the blend. Thus, a unit defined in this way is in agreement
with the tabletting process and therefore makes it less complicated to compare
homogeneity in the blend to homogeneity in the tablets.
By adapting a sample as a definition of a unit the variance between the drug

content in a number of units can be calculated and used as a measure of homo-
geneity.
When a unit has been defined the next problem is to decide where to sample
and how many samples to collect to be able to estimate a variance that is rep-
resentative for the blend homogeneity. In this relation it should be mentioned
that as an example the total amount of drug in a 360 kg batch (drug and filling
material) could be as little as 0.5 kg, and the weight of a sample less than f.ex.
80 mg. With these orders of magnitude and in case of batch inhomogeneity a
variance estimated between samples sampled close to each other differs from
a variance estimated from samples collected far apart. Hence, for exploratory
purposes it is relevant to assess different types of variances, i.e. variances esti-
mated from samples sampled closely and variances based on samples sampled
far apart.
In Appendix A a model that describes blend inhomogeneity (variation between

sample ’units’) on three scales is introduced. The three scales are referred to
as small, medium and large scale variation and they correspond to respectively
variation between the content in neighbouring samples/replicates, variation be-
tween the mean content in areas within a layer in the blend and variation be-
tween the mean content in different layers in the blend. In statistical terms this
is a hierarchical or a nested model. In Appendix A large scale variation refers
to inhomogeneities between layers in the blend as vertical inhomogeneity is a
very likely result of deblending. However, in case of suspicion of inhomogene-
ity in the horizontal direction the model could easily be changed to model this
kind of inhomogeneity. Further, the hierarchical model can also be changed
into modelling inhomogeneity on e.g. four or two scales of inhomogeneity if
this seems to be more relevant.
In case of blend homogeneity the large and the medium scale variation (mea-
suring differences between the mean content in respectively layers and areas
within a layer) are zero. The small scale variation is an inherent variation in
the blend and therefore it is not zero in case of homogeneity. However, in case
of homogeneity the small scale variation is independent of in which layer of
the batch it is estimated.
It should be noted that in the literature several examples exist of models taking
into account correlation between the samples measured as a function of the
distance between the spots in the blend from which the samples are collected.
(See e.g. [13]). However, these models are generally not used in practice. With
future techniques as e.g. NIR (near infra red) techniques correlation as a func-
tion of distance may be used in relation to image analysis methods. However,
NIR technology is not commonly introduced in production yet, and the focus
of this thesis is to develop and improve methods to assess uniformity within
the scope of current sampling technology, the sampling thieves.
For explorative purposes assessing inhomogeneity on different scales is rele-

vant. However, when it comes to the patients using the tablets a single measure
of the overall homogeneity in the blend is relevant as the overall blend homo-
geneity corresponds to the overall homogeneity of the content in the tablets.
In Appendix B two methods of measuring the overall variation in the blend
is discussed under the assumption that homogeneity can be modelled by the
hierarchical model presented in Appendix A. Both methods relate the overall
variation to the variation measured on the three scales of homogeneity defined
in Appendix A.
The first method is to use the total variation from the analysis of variance
(ANOVA) table corresponding to the hierarchical model for the variation in
the batch as an estimate for the overall variation. The other method is to use
the total variation on a randomly collected sample from the blend as an esti-
mate of the overall variation in the blend.
The difference between the estimates of the overall variation obtained with
each of these two methods depends on the sampling plan used to collect the
samples on which the estimates are based.
3.1. VARIANCES AS A MEASURE OF HOMOGENEITY 21
If a patient only uses one randomly sampled tablet for example when taking
an aspirin to relieve the pain of a headache, he/she will experience a deviation
from LC corresponding to the variance on a randomly chosen tablet. However,
if the patient uses more than one tablet as part of an ongoing treatment, the total
variation in drug content experienced may depend on the way the tablets are
collected. Are they randomly chosen from the batch or do they all come from
the same part of the batch etc. The tablets in a single package will in general not
be sampled from a balanced, hierarchical sampling plan as in Appendix A and
Appendix B, and even if the tablets by accident were sampled in accordance
with a hierarchical sampling plan, the "sampling plan" would be unknown.
Hence, regarding the total variation experienced by a patient using more than
one tablet neither method of estimating the overall variation is ideal.
Another criteria for deciding which estimate to use as a measure for the overall
variation in the blend is to consider the properties of the acceptance criteria for
the blend. Acceptance criteria are discussed in more detail in Section 3.3. In
case of uncorrelated samples, which corresponds to the model in Appendix A
with no variation between layers and a hierarchical sampling plan with only
one replicate per area, the two measures for the overall variation in the blend
are identical. Otherwise the measure of the total variation corresponding to the
ANOVA table in general leads to more efficient and less ambiguous properties
of the acceptance criteria for blend homogeneity.
In conclusion variance can be used as a measure of (in)homogeneity. For ex-

plorative purposes it is relevant to look at variances at different scales. In other
situations an overall measure of the batch homogeneity may be more conve-
nient and relevant. Two methods to estimate the overall variance are presented.
None of these truly describes the total variation experienced by a patient using
more than one tablet - but it is very complicated if possible at all, to estimate
this total variation. However, regarding acceptance criteria for blend unifor-
mity the total variation from the ANOVA table is relevant.
3.2 Methods to assess homogeneity and factors

that may influence homogeneity
In the thesis two different approaches to assess homogeneity and factors that
may influence homogeneity have been introduced.
The first approach described in Appendix E leads to a model of the best ob-
tainable blend and content uniformity derived from the distribution of particle
radii. However, the best obtainable homogeneity is a ’theoretical’ limit that
holds for all batches with the same distribution of particle radii, and therefore
this approach does not lead to information on the actual homogeneity of a given
batch. The second approach described in Appendix A introduces two methods
to assess the homogeneity of a specific batch.
3.2.1 The effect of particle size distribution
Particle size distributions are often seen to be skewed and it has been shown
in Appendix E that this feature affects the distribution of content in blend and
tablet samples.
For a log-normal distribution of particle diameters, the resulting distribution

of particle mass (volume) is also a log-normal distribution. It is found that
skewness and excess (heavy-tailedness) of the distribution of particle radii is
amplified when transformed to particle mass. The larger the coefficient of
variation in the distribution of particle radii the more pronounced the amplifi-
cation. The relation between the coefficient of variation for particle mass and
the coefficient of variation for particle radii is given in a table.
Beside the variation in particle mass the variation in dose content is affected
by variation in the number of particle in a sample. For a homogeneous blend
with a random scattering of particles over the blend it is demonstrated that for
a given distribution of particle sizes the variation in the distribution of absolute
doses is proportional to the average number of particles in the samples (tablets).
Further, it is shown that the larger the average number of particles in the sample
the closer the distribution of content in the samples is to a normal-distribution.
3.2. M ETHODS TO ASSESS HOMOGENEITY AND FACTORS THAT MAY
INFLUENCE HOMOGENEITY 23
For spherical particles an explicit relation between the variation in the relative
doses and the mean and the coefficient of variation in the distribution of particle
radii is given.
3.2.2 Assessment of homogeneity in specific batches
Two methods are introduced in Appendix A to assess homogeneity and factors

that may influence homogeneity in a specific blend. The two methods are based
on respectively Generalized Linear Models (GENMOD) and General Linear
Models (GLM). General Linear Models can be used to assess differences in
mean content in respectively layers and areas within a layer. Generalized Lin-
ear Models are here used to assess differences in variance. More specific to
assess whether the size of the small scale variation is constant throughout the
batch.
In practical applications a Generalized Linear Model should be applied first

to assess if the small scale variation/variation between replicates is constant
throughout the blend. If this is not so, it should be accounted for in the General
Linear Model.
GENMOD
In Appendix A a Generalized Linear Model is used to assess the influence of

layers on the small scale variation. For samples simulated from a hierarchi-
cal model with three layers, four areas within each layer, and three replicates
within each area the following was found: For the 5% level test the standard
deviation between replicates within an area has to be 4.5 times larger in one
layer than the standard deviation corresponding to replicates within an area in
another layer for the effect of layers to be declared significant with a probabil-
ity of at least 0.95.
However, depending on the experimental design and the assumptions made, the
method presented in this section can also be used to assess factors influencing
sample error (variation). For example the method can be used to test if one
sampling thief leads to larger variation between replicate samples than another
thief.
Using the thief leading to the smallest variation between replicates will reduce
the risk of incorrectly rejecting a batch because of suspicion of inhomogeneity.
Other external sources of variation (as e.g. the sampling procedure) may have
a similar effect on the small scale variation. Examples of such analysis are
given in Appendix F.
GLM
When a Generalized Linear Model has been applied a General Linear Model
can be applied specially to assess the medium and the large scale variation as
well as factors that may influence these types of variation. In case the variation
between replicates is found not to be constant throughout the blend, this should
be corrected for in the analysis by introducing appropriate weights
Under the assumption that the variation between replicate samples is indepen-
dent of the layer and that there is no interaction between the factors in the
model, two statistical methods to describe blend homogeneity have been in-
vestigated.
The first statistical method (using the aggregated model) corresponds to an

’aggregated’ definition of homogeneity in the sense that large and medium
scale variation in the batch is assessed as a whole.
The other statistical method (using the hierarchical model) corresponds to a

homogeneity definition with two different criteria; one explicitly regarding the
large scale variation and the other explicitly regarding the medium scale varia-
tion.
The analysis showed that the two methods are approximately equally good
at detecting inhomogeneity. That is, an analysis according to the aggregated
model can be used to detect inhomogeneity even in situations with large scale
variation (variation between layers).
The most important difference between the two types of analysis is that when
inhomogeneity is declared according to the aggregated analysis the result does
not reveal whether this inhomogeneity is due to large or medium scale variation
in the batch. However, the hierarchical model explicitly assesses respectively
large and medium scale variation.
3.2. M ETHODS TO ASSESS HOMOGENEITY AND FACTORS THAT MAY
INFLUENCE HOMOGENEITY 25
The power of the respective tests as a function of the standard deviation corre-
sponding to respectively variation between layers and variation between areas
within a layer are shown in Figure 3 to Figure 5 in Appendix A. The standard
deviation corresponding to respectively variation between layers and variation
between areas within a layer are measured relatively to the standard deviation
corresponding to replicates.
Finally, for the given design (and for σrep = 1) the power of the test of a thief
effect was shown in Figure 6 in Appendix A. The test of the thief effect is
independent of σlayer and σarea,hi .
With the given design (and for σrep = 1) a difference greater than 1.5 in mean
content in samples from two different thieves will be detected with a probabil-
ity of at least 0.95.
3.2.3 Example
As an example of how the robustness and power of a test can be used to evaluate
the test result the test of the small scale variation in Appendix F is discussed in
relation to the results from Appendix A.
As the resulting design in Appendix F is neither balanced nor identical to the

design from which the robustness and power are assessed, the example should
be seen as a guidance on the type of considerations to make when evaluating
test results.
Variation between replicate samples
2 , tends to be larger the

In Appendix F the variation between replicates, σrep
lower the layer the samples are sampled from. However, the tendency is not
significant.
The estimated variance components, σrep2 , are given in the Table 3.1. The esti-
mates in the table are multiplied with 1000 compared to the results in Case F.
The reason is that in Appendix A the unit is % LC rather than fraction LC.
Batch 1 Batch 1
Thief 1 2 3 1 2 3
2
σrep,top 0 0 1 0 3 8
2
σrep,bottom 1 7 17 4 35 98
Table 3.1: Estimates of variance components from Appendix F.
2
It is seen from the three cases in the table where σrep,top 6= 0 that the estimate
2 2
of σrep,bottom is between 10 and 20 times as large as σrep,top . At first a dif-
ference that large may be expected to be found significant. However, for the
design in Appendix A it was showed that σrep in one layer should be more than
4.5 times as large as σrep in an other layer for the test to show significance with
a high probability. This corresponds to one variance estimate being 4.52 ≈ 20
times larger than the smallest. This result is valid for the balanced design with
12 pairs of replicates in each batch and analysed with a model with one factor
(layer).
The design in Case F is not balanced, it has only 6 pairs of replicates in each
batch and two factors included in the model (layer and thief). Hence, with
2
an estimate of σrep,bottom 2
being 10 to 20 times as large as σrep,top it is not
surprising that the test shows no significance - however, the estimated differ-
2
ence between σrep,top 2
and σrep,bottom may still be real and give practical and
valuable insight to the sampling process.
3.3 Analysis of acceptance criteria
When a parameter expressing the overall (total) (in)homogeneity of the blend

has been estimated the batch quality can be evaluated by comparing the param-
eter estimate to some acceptance criteria or critical value. The critical value
could be determined from some theoretical model of the ultimate limit of ho-
mogeneity. These theoretic limits could for example be the variance in a com-
pletely ordered or in a completely random blend. The most common definition
of a perfectly random blend is one in which the probability of finding a particle
of a constituent of the blend is the same for all points in the blend. More than
30 different criteria relating the sample variance to theoretical limits have been
proposed by various investigators [14]. These criteria are referred to as mixing
indices in the literature. The analysis of variance method e.g. presented in Ap-
3.3. A NALYSIS OF ACCEPTANCE CRITERIA 27
pendix A could also be taken as basis for a mixing index (see also [15]). In this
case the theoretical limit of homogeneity is the various variance components
being zero. It should be noted that a variance component being zero serves
more as a theoretical value for homogeneity. It is not useful as an acceptance
criteria. Further, it should be noted that in a homogeneous blend the large and
medium scale variation are negligible. However, the small scale variation is an
inherent variation that is non-zero even in a random blend.
Alternatively to the models for ultimate limits of homogeneity the quality of

the blend can be evaluated in accordance to some practical criteria assessing if
the homogeneity is satisfactory for the blend to serve its purpose.
The properties of such acceptance criteria can be investigated as a function of

e.g. the true mean and total variation in the batch similarly to the analysis of
the properties of the acceptance criteria in Appendix C and Appendix D.
The discussion of the acceptance criteria and the derivation of expressions for
acceptance probabilities in the appendices is performed under the assumption
that individual sample values may be represented by independent, identically
distributed variables and that the distribution of sample results may be de-
scribed by a normal distribution.
Thus, when the samples are tablets selected from a batch, the assumption cor-
responds to assuming that the overall distribution of dose content in the batch
may be represented by a normal distribution and that individual dosage units
are selected at random from the dosage units in the batch. When the samples
are blend samples, the assumption analogously corresponds to assuming that
the overall distribution of such potential samples from the blend may be repre-
sented by a normal distribution and that samples are taken at randomly selected
positions in the blend. Thus, the model will not be adequate when the overall
distribution in the blend (or batch) is bimodal or multimodal corresponding e.g.
to stratification, when the distribution is skewed, e.g. as a result of deblending,
or when the distribution has heavier tails than the normal distribution, e.g. as
a result of imperfect mixing (clustering) or of using drug particles that are too
large for the intended dosage (see Appendix E).
When sampling is performed under a hierarchical (or nested) scheme as sug-

gested e.g. by PQRI, the model will only be adequate in such (rather unlikely)
situations where there is no correlation between subsamples from the same
location in the blend (see [16]).
However, even despite these restrictions the mathematical analytical discussion

serves a purpose of clarifying and illustrating the statistical issues involved,
thereby providing further insight in the properties of various tests that have
been proposed in the pharmaceutical literature.
Under the assumption that the distribution of sample results may be described
by a normal distribution, the following results regarding acceptance criteria
have been derived in Appendix C and Appendix D.
In essence the purpose of using acceptance criteria is to secure a certain quality

of the product under concern. Thus in industrial or commercial practice, prod-
uct requirements are often formulated as specifications for individual units of
product, but may also include specifications for such batch or process charac-
teristics as batch fraction nonconforming or standard deviation between units
in the batch.
However, regulatory practice for pharmaceutical products has most often speci-
fied criteria for sample values rather than providing specifications for the entity
under test. As therefore control and monitoring procedures in tablet production
are based upon samples from the blend, or from the batch of tablets, there is
an inherent uncertainty concerning the actual dispersion in the blend or batch
being sampled. This uncertainty is partly due to sampling and partly due to the
(in)homogeneity of the blend/batch.
The statistical tool used to link sample result and acceptance criteria to the
actual dispersion in the blend or batch is an OC-curve (or surface) that shows
the probability of passing the acceptance criteria as a function of e.g. fraction
nonconforming or true mean and standard deviation in the blend or batch, as
an OC-curve (or surface) reflects the effect of such sampling uncertainty.
When properties of an acceptance criteria have been described through the

corresponding OC-curve the next issue is to determine the assurance related
to the acceptance criteria. This assurance can also be determined from the
OC-curve.
In Appendix D statistical tools and methods that can be used to determine how
assurance depends on sample size (i.e. how to set up a criterion that gives a
3.3. A NALYSIS OF ACCEPTANCE CRITERIA 29
certain assurance with an appropriate sample size) are described and discussed
for simple acceptance criteria (sample standard deviation and coefficient of
variation as well as an USP criterion that includes a test by attributes).
Also in the appendix it is shown that a three-class attribute criteria as e.g. in

USP 24 for content uniformity in essence controls the proportion of tablet
samples outside the inner set of limits for individual observations. For nor-
mally distributed observations this is identical to control the combination of
batch mean and standard deviation, i.e. a parametric acceptance criterion. It
is shown how to set up parametric acceptance criteria for the batch that gives
a high confidence that future samples with a probability larger than a specified
value passes the USP three-class criteria.
In the literature changes to the procedure in USP have been proposed. In gen-
eral the proposed test procedure is similar to the parametric criteria mentioned
above. In the thesis simulations have been performed both for normally and
log-normally distributed content in the tablets. The simulations revealed that
the test is relatively robust to deviations from the normal distribution. This is
relevant as such deviations for example is seen in case of low-dose tablets with
large particle radii as also discussed in the thesis.
Finally single sampling acceptance plans for inspection by variables that aim
at matching the USP proposal have been suggested.
Chapter 4
Conclusion
In this thesis the use of statistical methods to address some of the problems
related to assessment of the homogeneity of powder blends in tablet production
is discussed.
When assessing homogeneity the first problem is how to define homogeneity

of the blend. This is not straight forward as bulk materials have no natural
unit or amount of material that may be drawn into a sample. However, a blend
sample of the size of one to three times the size of a tablet is a convenient unit.
With this definition of a unit, variances between blend samples can be used as
a measure of the (in)homogeneity in a blend.
In the thesis a hierarchical (or nested) as well as an aggregated model has been
introduced to describe (in)homogeneity. The hierarchical model specifically
takes into account deblending in a specified direction. Both the hierarchical
and the aggregated model can be used to detect inhomogeneity. However, in
case of inhomogeneity the hierarchical model provides the most detailed infor-
mation on the type of inhomogeneity.
Regarding the end users of the tablets the total variation between the tablet
content is relevant. This variation is closely related to the overall variation in
the blend. Two methods to determine the overall variation in the blend have
31
32 C HAPTER 4. C ONCLUSION
been suggested. One of the methods (estimating the overall variation from the
ANOVA table) leads to less ambiguous properties of acceptance criteria for
blend uniformity. However, none of the methods truly describes the variation
experienced by a patient, as this variation depends on how the tablets in the
package are selected from the batch of tablets.
It has been shown that particle size distribution may have an influence on the
distribution of content in blend and tablet samples. Specially for low-dose
tablets it is important to keep the particle radii small (and the number of parti-
cles large) to minimize the variation in content in the blend and tablet samples.
Assuming perfect mixing, and a log-normal distribution of particle sizes, the

requirement on the coefficient of variation in the distribution of dosage units is
essentially a general requirement on the minimum average number of particles
in a dosage unit. This minimum average number of particles does not depend
on label claim. However, as the average number of particles in tablets depend
on label claim, a blend that might produce a satisfactory distribution of doses
(in terms of the coefficient of variation in the distribution of relative doses) for
large dose tablets need not be satisfactory for smaller dose tablets.
Interpretating the results in terms of blend samples rather than samples of

tablets from a batch, it is of interest to note that the practical necessity of using
blend samples that are larger than the dosage units imply that the coefficient
of variation in such blend samples is smaller than the coefficient of variation
in the resulting dosage units. For blend samples that are four times the size of
the final dosage units, the coefficient of variation in the blend samples is only
half the size of the coefficient of variation in the final dosage units. Moreover,
a larger blend sample might mask departure from normality in the distribution
of dose content in low-dose tablets.
Generalized Linear Models (GENMOD) can be used to assess factors that may
have an influence on a variance (e.g. the effect of layers on the replicate vari-
ance). General Linear Models (GLM) can be used to assess factors that may
have an influence on the mean content in blend samples, e.g. a sampling de-
vice leading to sampling bias. For a specific sampling design the power and
robustness of the statistical tests related to the GENMOD and GLM models
have been assessed.
A central problem is to develop acceptance criteria for blends and tablet batches
33
to decide whether the blend or batch is sufficiently homogeneous to meet the

need of the end users. Under the assumption that the content in blend and tablet
samples is normally distributed properties of a number of acceptance criteria
have been discussed.
Regulatory practice related to tablet production are most often criteria specify-
ing limits for sample values rather than for the actual homogeneity in the blend
or batch of tablets. This leads to an inherent uncertainty concerning the homo-
geneity in the blend or tablet batch. This uncertainty is partly due to sampling
and partly due to (in)homogeneity of the blend or batch.
In the thesis it is shown how to link sampling result and acceptance criteria
to the actual quality (homogeneity) of the blend or tablet batch. Further it is
discussed how the assurance related to a specific acceptance criteria can be
obtained from the corresponding OC-curve.
Also in the thesis it is shown that a three-class attribute criteria as e.g. in USP
24 for content uniformity in essence controls the proportion of tablet samples
outside the inner set of limits for individual observations. For normally dis-
tributed observations this is identical to control the combination of batch mean
and standard deviation, i.e. a parametric acceptance criterion. It is shown how
to set up parametric acceptance criteria for the batch that gives a high confi-
dence that future samples with a probability larger than a specified value passes
the USP three-class criteria.
In the literature changes to the procedure in USP have been proposed. In gen-
eral the proposed test procedure is similar to the parametric criteria mentioned
above. In the thesis simulations have been performed both for normally and
log-normally distributed content in the tablets. The simulations revealed that
the test is relatively robust to deviations from the normal distribution. This is
relevant as such deviations for example is seen in case of low-dose tablets with
large particle radii as also discussed in the thesis.
34 C HAPTER 4. C ONCLUSION
II
Part II
Included papers
35
A
Paper A
Robustness and power of

statistical methods to
assess blend homogeneity
37
39
1 Introduction
In this appendix the robustness and power of statistical methods to assess blend
homogeneity is discussed. The statistical methods assess various factors that
may alter batch homogeneity and therefore one could also say that the statisti-
cal methods assess blend inhomogeneity. For convenience the term homogene-
ity as well as the term inhomogeneity will be used in relation to the statistical
models. The discussion is based on simulations of blend samples in SAS [17].
However, before simulations of the blend samples are made it is necessary to
have a definition and a model for blend homogeneity.
Two different models for blend homogeneity are presented: An overall or

aggregated model that does not distinguish between horizontal and vertical
(in)homogeneity and a hierarchical model that specifically takes into account
the situation where deblending or insufficient mixing causes inhomogeneity
in the vertical direction. This model also accounts for (in)homogeneity in the
horizontal direction. The aggregated and the hierarchical models assess large
scale and medium scale variation1 . Small scale variation is assessed in a sepa-
rate analysis.
All samples are simulated in accordance with the hierarchical model. How-
ever, both the aggregated and the hierarchical model are used in the statistical
analysis of the simulated samples.
Finally, a number of batches are simulated in accordance with a hierarchical

model including small scale variation, and a statistical method specifically use-
ful to assess small scale variation is presented.
1
The terms large, medium and small scale variation should be taken loosely. They are meant
as a convenient way to distinguish between different types of (in)homogeneities in a batch. By
large scale variation is meant variation between samples collected so far apart that the distance
between them should be measured on a ’large scale’, e.g. top, middle and bottom layer of the
batch. Similarly small scale variation means variation between samples collected so close to
each other (e.g. replicate samples) that the distance between them should be measured on a
’small scale’. However, it is very important to understand that large and small scale does not
refer to the size of the the respective variations. Thus, the large scale variation could actually be
smaller than the variation measured on the small scale.
40
2 Models of batch homogeneity
In the following two different statistical models of batch homogeneity are in-
troduced: the aggregated and the hierarchical models.
2.1 The aggregated model
No definitive definition of homogeneity of a powder blend exists. One defini-

tion relates to the acceptance criteria for blend uniformity analysis as suggested
by the FDA in a Draft Guidance [11] as well as the the first stage of the blend
sample criteria suggested by PQRI [18]. In essence these acceptance crite-
ria controls the variation between a number of locations (areas) in the batch.
PQRI [18] explicitly recommends to assess this variation on the basis of sam-
ples from at least ten different areas with at least three replicates from each
area.
Fig. 1: A batch in which twelve areas are identified for sampling.
Figure 1 shows an example of a batch with twelve areas representatively (and

randomly) distributed in the batch. Sampling three replicates from each of the
twelve areas would be in compliance with the recommendations from PQRI.
The variation between samples collected in accordance with this sampling plan
41
can be modelled by the following statistical model:
(
i = 1, .., 12
Yij = µ + Ai + Eij , . (1)
j = 1, 2, 3
Yij is the content of a sample, µ the mean content in the batch (including bias
due to sampling and chemical analysis), Ai the effect corresponding to the
area from which the sample is taken, and Eij is the error term corresponding
2
to the j’th replicate from the i’th area. Further Ai ∈ N (0, σarea,agg ) and
2 2
Eij ∈ N (0, σrep ). The variation, σarea,agg , can be thought of as an aggregate
of the large and the medium scale variation in the batch.
If all samples are collected in random order the design is a simple single-factor
design. The experimental design, the analysis, assumptions and the statistical
terminology related to such a design is given by Montgomery [19].
2.2 The hierarchical model
When basing the assessment of batch homogeneity on a set of blend samples

it is important that the samples are representative, i.e. randomly selected loca-
tions. However, according to e.g. FDA [20]it is also important that the specific
areas of the blender which have the greatest potential to be non-uniform is rep-
resented. One common type of inhomogeneity is that the mean content in the
top, middle and bottom layer of a batch differ due to deblending or insufficient
mixing. Hence, in these cases it is important to collect samples from all layers.
Several examples of sampling plans that divide the batch into layers are seen in
the literature (see e.g. Kræmer [21] and Berman [5]). Further FDA [20] states
that for drum sampling samples should be collected from three layers of the
drum: the top, middle and bottom layer.
Figure 2 shows an example of a batch divided into three layers. Within each
layer samples are taken from the points of a triangle, as well as from the centre,
thus four areas within each layer are chosen. By changing to the next level, the
triangle is rotated by 180o . In the left corner of the figure a top view of the
sampling scheme is shown.
42
Fig. 2: A batch divided in three layers: top, middle and bottom. Variation be-
tween the mean content in the layers represents large scale variation. Variation
between the mean content in the four areas within a layer represents medium
scale variation. To the left a top view of the sampling scheme is seen.
Variation between the mean content in the three layers can also be thought of
as large scale variation in the vertical direction. The aggregated model (1) does
not explicitly account for this type of variation or inhomogeneity. However, it
is accounted for in the following model:


i = 1, 2, 3
Yijk = µ + Li + A(L)j(i) + Eijk , j = 1, 2, 3, 4 . (2)


k = 1, 2, 3
Li is the effect from layer i, A(L)j(i) is the effect from the j’th area within
layer i and Eijk is the effect from the k’th replicate from area j in layer i.
2
Further it is assumed that Li ∈ N (0, σlayer 2
), A(L)j(i) ∈ N (0, σarea,hi ) and
2 2 2
Eijk ∈ N (0, σrep ). In contrast to σarea,agg , σarea,hi only accounts for the
medium scale variation in the blend, i.e. the variation between the areas within
a layer.
The model is a hierarchical (also called nested or stratified) model. A detailed

description of nested models is also given in [19].
43
2
When there is no difference between layers (i.e. σlayer = 0) the aggre-
2
gated model (1) and the hierarchical model (2) are identical, i.e. σarea,agg =
2 2
σarea,hi = σarea . The hierarchical model corresponds to the definition of ho-
mogeneity used in the cases in Appendix F.
The hierarchical model including differing small scale variation
In practice when sampling replicates from the same area the replicates are sam-
pled as close as possible without overlapping. The reason is to avoid spots with
a high degree of deblending caused by the insertion of the thief when collecting
previous samples. Thus the variation between replicates includes variation be-
tween content in neighbouring spots within an area, i.e. variation in the batch
measured on a very small scale.
In a situation where deblending has occurred in one layer the variation be-
tween the content in neighbouring spots within an area may differ from the
corresponding variation in other layers, hence the small scale variation may
depend on the layer.
Further, the static pressure in the blend is lower near the surface than near
the bottom. This difference in static pressure may affect the sampling results
(see e.g. Berman [2]). One example would be the difference in static pressure
causing the variation between the content in replicate samples being larger
when sampled from the bottom layer than when samples from the top layer
(see e.g. Appendix F).
The (simulated) variation between replicate samples is seen to be the result of

several causes: small scale variation in the blend and variation due to sampling
and chemical analysis. At least two physical causes (small scale variation and
variation due to sampling) may lead to a variation between replicates, σrep 2 ,
that depends on the layer.
Neither the aggregated model (1) nor the hierarchical model (2) accounts for
this type of variation. However, with the following extension to the hierarchical
model (2), small scale variation is accounted for:
44


i = 1, 2, 3
2
Eijk ∈ N (0, σrep,i ), j = 1, 2, 3, 4 . (3)


k = 1, 2, 3
2.3 Simulated samples
The full hierarchical model (the hierarchical model (2) with the extension (3))
is able to account for three different types of variation in the batch: variation
measured on a large scale (the variation between layers), a medium scale (vari-
ation between areas within a layer) and the small scale variation (variation be-
tween replicates within an area). This ability makes it a very versatile model.
The drawback is that the model includes six parameters (µ, σlayer 2 2
, σarea,hi ,
2 2 2
σrep,top , σrep,middle and σrep,bottom ), hence it is very difficult in a clear way
to analyse and present results with all parameters varying at a time. There-
fore in each series of simulated batches some of the parameters are kept fixed
depending on the aim (hypothesis) of the model under analysis.
Thus, the full hierarchical model (model (2) with the extension (3)) was used to
generate a number of datasets all with a total of 36 samples equally distributed
over 3 layers, 4 areas within each layer and 3 replicates within each area. The
sampling plan is shown in the first three columns of Table 1, page 52. On
the whole the number of samples corresponds to the number of samples in Ap-
pendix F as well as the sampling plan suggested by PQRI [18]. In each analysis
500 batches are simulated for each combination of the parameter values.
In the analysis the sample content is measured in percent of label claim, LC.
Thus µ = 100 corresponds to LC. However, as the overall mean content in the
batch is balanced out in the calculations the results are independent of the value
of the overall mean content. Further the results do only depend on the ratios
σlayer /σrep and σarea,hi /σrep (or σarea,agg /σrep ), and therefore the value of
the variance components is measured relatively to σrep 2 (or σ
rep,top in a model
with extension (3)), i.e. either σrep = 1 or σrep,top = 1.
Two types of analyses are made in the following sections. The first type of
analysis is described in Section 3. It focuses on factors influencing the mean
content of a sample corresponding to the terms Ai , A(L)j(i) and Li in the
45
aggregated model (1) and the hierarchical model (2). This corresponds to an
analysis of the large and the medium scale variation. The other type of anal-
ysis, described in Section 4 focuses on factors influencing the term Eijk in
extension (3) of the full hierarchical model. Thus this is an analysis of the
small scale variation.
3 Assessing factors with influence on the mean

content of the active component in a sample
In Section 2 it is described how a number of samples were simulated from a

hierarchical model. These simulated samples represent real samples and real
batches.
In practice a statistical method corresponding to the aggregated model (1) is

(most often) used to analyse blend sample results (see eg. [22]). In essence the
acceptance criteria suggested by PQRI [18] is also based on this model, i.e. the
acceptance criteria does not explicitly take into account a possible large scale
variation between layers.
In this section the simulated batches (some of which contain variation between
layers) are analysed in accordance with both the aggregated model (1) and
the hierarchical model (2). The robustness and the power of these methods is
assessed as a function of the variation between layers, σlayer , and the variation
between areas within a layer, σarea,hi .
Finally the effect of including an external factor (sampling thieves) in the anal-
yses of the mean content of the collected samples is discussed.
The tests are conducted using the GLM procedure in SAS. In simulations of
2 = 1.
samples for the analysis µ = 100 and σrep
46

ing to the aggregated model (1)
A definition of blend homogeneity related to the aggregated model (1) is the

2
situation where there is no variation between areas, i.e. σarea,agg = 0. A statis-
2
tical test of the hypothesis σarea,agg = 0 is described e.g. by Montgomery [19].
The test statistic is
2
Sarea,agg
Zarea,agg = 2
, (4)
Srep
where hP i
12
3× i=1 (ȳi. − ȳ.. )2
2
Sarea,agg = (5)
12 − 1
and P12 P3
2 i=1 j=1 (yij − ȳi. )2
Srep = . (6)
12 × (3 − 1)
ȳi. is the mean of the three replicates in area i and ȳ.. is the overall mean.
2
Under the aggregated model with σarea,agg = 0 the test statistic follows a
2
F (12 − 1, 12 × (3 − 1)) distribution, and therefore σarea,agg is declared signifi-
cantly different from 0 at the 5 % level when Zarea,agg > F (11, 24).95 = 2.18.
Under the aggregated model (1)
2
Sarea,agg 2
/(3 × σarea,agg 2 )
+ σrep
2 /(σ 2 )
∈ F (12 − 1, 12 × (3 − 1)). (7)
Srep rep
As the samples were simulated from a model including the variation between
layers, the samples from the same layer are centered around the mean content
in that layer. Hence the averages ȳi. s are correlated, and therefore the under-
lying assumption from the aggregated model (1), that the mean content in the
2
areas are uncorrelated, is violated. Strictly spoken this means that σarea,agg
2 2
from model (1) does not exist in a context with σlayer and σarea,hi .
Under the hierarchical model (2) - which corresponds to the model from which
the samples are simulated - the overall variation between areas is separated into
47
the variation between layers, σlayer2 , and the variation between areas within a
2
layer, σarea,hi . In relation to the specific sampling plan from which the samples
2
are sampled (i.e. four areas within each layer) the mean value of Sarea,agg is
2 24 2 33 2 2
E[Sarea,agg ] = 11 σlayer + 11 σarea,hi + σrep .
0.953
0.860
0.767
s 0.674
t
d
l 0.581
a
y 2
e
r 0.487
0.394
0.301
0.208
0
0 1 2 0.115
std area
0.05
Fig. 3: The probability of declaring the variation between areas under the ag-
2
gregated model, σarea,agg , significant at the 5% level as a function of σarea,hi
and σlayer . This is also called the power of the test. The probability is
low when both σarea,hi and σlayer are small. When σarea,hi is larger than
1.25×σrep the probability is more than 0.95 no matter the value of σlayer . Sim-
ilarly when σlayer is larger than 3.5 × σrep the probability of testing σarea,agg
significant is more than 0.95, irrespective of the value of σarea,hi
This is also seen from Figure 3. The figure shows level curves for the power of
the test in (7) as a function of σarea,hi and σlayer . The power of the test is the
2
probability of declaring σarea,agg significant (here at the 5% level). The prob-
ability is found for each set of parameter values by testing simulated samples
from 500 batches.
2 the level curves correspond

From Figure 3 it is seen that for fixed values of σrep
q
to ellipses 24 2 33 2
11 σlayer + 11 σarea,hi = constant.
48
More explicit it is seen that, when σarea,hi is larger than 1.25 × σrep the prob-
2
ability of finding σarea,agg > 0 is at least 0.95. When σarea,hi is smaller the
2
probability of declaring σarea,agg significant depends on the value of σlayer .
The plot shows that because the variation between layers is not specifically ac-
counted for in the aggregated model (1), variation between layers in the batch
is identified as variation between areas. When the variation between layers,
2
σlayer , is more than 3.5×σrep the probability of testing σarea,agg > 0 is at least
24 2 33 2 2
0.95. In conclusion the test in 7 really measures 11 σlayer + 11 σarea,hi + σrep
2 .
relatively to σrep

ing to the hierarchical model (2)
A definition of blend homogeneity related to the hierarchical model (2) is

2
σlayer 2
= 0 and σarea,hi 2
= 0. The test statistic for the two hypotheses σlayer =
2
0 and σarea,hi = 0 are respectively (see e.g. [19])
2
Slayer
Zlayer = 2 (8)
Sarea,hi
and
2
Sarea,hi
Zarea,hi = 2
, (9)
Srep
where hP i
3
4×3× i=1 (ȳi.. − ȳ... )2
2
Slayer = (10)
3−1
and hP i
3 P4
3× i=1 j=1 (ȳij. − ȳi.. )2
2
Sarea,hi = (11)
3 × (4 − 1)
and P3 P4 P3
2 i=1 j=1 k=1 (ȳijk − ȳij. )2
Srep = (12)
3 × 4 × (3 − 1)
ȳ... is the overall mean, ȳi.. is the mean of the i’th layer and ȳij. is the mean of
the j’th area in the i’th layer.
49
The critical intervals for the test statistics are Zlayer > F (2, 9).95 = 4.26 and
Zarea,hi > F (9, 24).95 = 2.30
2
Under the hierarchical model the variation between areas, σarea,hi , is corrected
2
for the contribution from layers, σlayer . Hence, in this case the test of the
2
variation between areas, σarea,hi 2
= 0, is independent of the value of σlayer .
2
However, the test statistic, Zlayer , depends on Sarea,hi and therefore the test of
2 2
σlayer is expected to depend on the value of σarea,hi .
0.952
0.856
0.760
s 0.664
t
d
l 0.568
a
y 2
e
r 0.472
0.376
0.280
0.184
0
0 1 2 0.088
std area
0.05
2
Fig. 4: The power of the 5% level test of σarea,hi in the hierarchical model
(2). When σarea,hi is approximately 1.5 × σrep the probability of declaring
significance is more than 0.95. The probability is independent of σlayer .
2
For the test at the 5 % level Figure 4 shows the power of the test of σarea,hi as
a function of σarea,hi and σlayer .
As expected it is seen that now that the variation between layers is accounted
for in the model the test of the variation between areas does no longer depend
of the value of σlayer . In agreement with Figure 3, Figure 4 shows that when
2
σarea,hi is approximately 1.5 × σrep , the probability of declaring σarea,hi sig-
nificant is at least 0.95.
50
0.952
0.856
0.760
s 0.663
t
d
l 0.567
a
y 4
e
r 0.471
0.375
2
0.279
0.182
0
0 2 4 6 8 0.086
std area
0.05
2
Fig. 5: The probability of the 5% level test of σlayer in the hierarchical model
(2). The power is dependent of σlayer and σarea,hi .
2
Figure 5 shows the power of the test of σlayer as a function of σarea,hi and
2
σlayer . As expected the power of the test of σlayer depends on the value of
2
σlayer
σarea,hi , namely in essence of 2
1+3×σarea,hi
.
If the factors describing the variation between layers had been simulated with
three fixed (predefined) levels the test statistic would follow a non-central F-
distribution, but the overall picture would not be changed.
3.3 The effect of including sampling thieves in the model
So far the homogeneity of the blend has been modelled by models taking into
account the large and the medium scale variation. Blend samples results are
however not entirely a consequence of the batch homogeneity. Sometimes it is
relevant to investigate whether other factors such as sampling thief or sampling
technique have an effect on the content in a sample. This is the situation in
Appendix F. Also Berman [23] has assessed the significance of factors as
51
sampling thief and technique using an analysis of variance.
A number of batches in which two different thieves were used for sampling
were simulated. The experimental design is shown in Table 1. The model is


 i = 1, 2, 3


j = 1, 2, 3, 4
Yijkl = µ + Li + A(L)j(i) + tk + Eijkl , , (13)

 k = 1, 2


l = 1, 2, 3
P
where tk is the effect of thief k, and tk = 0. Apart from that the assumptions
are identical to those from the hierarchical model (2).
For practical reasons the following parameters are fixed: σlayer = 2, σrep = 1
and µ = 100.
2
The test statistic for the hypothesis tk = 0, σlayer 2
= 0 and σarea,hi = 0 are
now
2
Sthief
Zthief = 2
, (14)
Srep
2
Slayer
Zlayer = 2 , (15)
Sarea,hi
and
2
Sarea,hi
Zarea,hi = 2
. (16)
Srep
2 denotes the estimate of the variance between replicates, σ 2 , obtained

Srep rep
2
from the residual sum of squares corresponding to model (13), Sarea,hi denotes
the mean sum of squares corresponding to variation between areas, corrected
2
for layers and thieves in analogy with (11), Slayer in analogy with (10) denotes
the mean sum of squares corresponding to the variation between layers. Finally
52
No. Layer Area Thief

1 top A X
2 top A X
3 top A Y
4 top B X
5 top B Y
6 top B Y
7 top C X
8 top C X
9 top C Y
10 top D X
11 top D Y
12 top D Y
13 middle A X
14 middle A X
15 middle A Y
16 middle B X
17 middle B Y
18 middle B Y
19 middle C X
20 middle C X
21 middle C Y
22 middle D X
23 middle D Y
24 middle D Y
25 bottom A X
26 bottom A X
27 bottom A Y
28 bottom B X
29 bottom B Y
30 bottom B Y
31 bottom C X
32 bottom C X
33 bottom C Y
34 bottom D X
35 bottom D Y
36 bottom D Y
Table 1: All simulated batches include 36 samples: 12 samples from each of

three layers and 3 samples from each of four areas in each layer. Some batches
were simulated with two different thieves included in the experimental design
as shown in the last column. In the simulation of other batches the variation
2 , depends on the layer from which the samples
between replicate samples, σrep
are sampled.
53
2
Sthief denotes the mean sum of squares corresponding to variation between
thieves.
The critical value for the test statistic corresponding to thieves is Zthief >
F (1, 23).95 = 4.3. The test statistic corresponding to layers and areas within a
layer are unchanged from the hierarchical model (1).
As there were only three replicate samples in each area, the use of two different
thieves could not be completely balanced within an area, and therefore the
partitioning of the variation is not straightforward. However, as the design
allows for an estimation of the variation between replicates and as the use of
the two thieves has been balanced over areas and layers, it is seen from (14)
that the test of the effect of sampling thief, tk , is independent of σlayer and
σarea,hi .
0.961
0.883
0.805
t 0.726
h
i 3
e
f 0.648
d
i
f
. 0.570
2
0.492
0.414
1
0.335
0 2 4 6 8 0.257
std area
0.05
Fig. 6: The power of the 5% level test of the thief effect, tk , i.e. the probability
of declaring the thief effect significant.
Figure 6 illustrates the power of the 5 % level test of tk . The power is plotted as
2
a function of σarea,hi , and the effect of the thieves. The effect is the difference
between the mean of samples collected with one thief and the mean of samples
collected with the other thief.
54
The power of the test of the thief effect is independent of σarea,hi and a dif-
ference between samples collected with each of the two thieves of larger than
1.5% of LC has a probability of at least 0.95 of being detected with the given
design.
2
From expression (16) the test of σarea,hi is expected to be independent of both
2
σlayer and tk . Expression (15) shows that the power of the test of σlayer is
2
independent of tk but not of σarea . Hence, plot of the power of these two tests
are similar to Figure 4 and Figure 5.
3.4 Conclusion
Samples from a number of batches have been simulated according to a hier-

archical model that explicitly takes into account respectively the large and the
medium scale variation in the batch. Further, it is assumed that the variation
between replicate samples is independent of the layer and that there is no in-
teraction between the factors in the model.
Under the assumption that the homogeneity of real batches in this way may
consist of both a large and a medium scale variation the power of two statistical
methods to assess batch homogeneity has been found.
The first statistical method (the aggregated model) corresponds to an ’aggre-

gated’ definition of homogeneity in the sense that large and medium scale vari-
ation in the batch is assessed as a whole.
The other statistical method (the hierarchical model) corresponds to a homo-

geneity definition with two different criteria; one explicitly regarding the large
scale variation and the other explicitly regarding the medium scale variation.
The probability of declaring inhomogeneity is basically the same for analy-

sis corresponding to each of the two model. The most important difference
between the two types of analysis is that when inhomogeneity is declared ac-
cording to the aggregated analysis the result does not reveal whether this inho-
mogeneity is due to large, medium or small scale variation in the batch. How-
ever, the hierarchical model explicitly assesses respectively large and medium
scale variation.
55
The power of the respective tests are shown in Figure 3 to 5. It should be noted
that σlayer and σarea,hi are measured relatively to σrep .
Finally, for the given design the power of the test of a thief effect was assessed
in Figure 6. In these simulations the variation between layers were held fixed
(σlayer = 2). However, the test of the thief effect is independent of σlayer and
σarea,hi .
With the given design a difference greater than 1.5 in mean content in samples
from the two thieves has a probability of at least 0.95 of being detected.
4 Assessing factors with influence on the varia-

tion between replicates
In Section 3 the robustness and power of statistical methods to evaluate the

large scale and medium scale variation in the blend were assessed by means of
a General Linear Model. In this section the robustness and power of a statistical
method to evaluate the small scale variation is assessed. Small scale variation
is variation in the blend measured on a small scale, i.e. variation between
neighbouring samples.
However in practice, variation between neighbouring samples includes both

small scale variation in the blend as well as sampling error and variation due
to the chemical analysis. In practice replicate samples from exactly the same
spot is avoided because of the risk of subsequent samples being biased due to
deblending from taking the first sample. Therefore neighbouring samples are
also referred to as replicate samples.
The method presented in this section can be used to assess factors with influ-
ence on either the small scale variation or eg. the size of the variation due to the
sampling, depending on the experimental design and the assumptions made.
In this analysis the factor under consideration is the effect of layers: Layers
having an effect on the small scale variation means the special kind of inhomo-
geneity where for example the variation between neighbouring spots within an
2
area is smaller in the top layer, σrep,top 2
, than in the bottom layer, σrep,bottom .
56
A tendency like this was found in Appendix F.
However, differences between the variation between replicates in the top layer,
2
σrep,top 2
, and the variation between replicates in the bottom layer, σrep,bottom ,
may also be due to the fact that it is more difficult to handle a long thief in the
bottom of the blend than handling a short thief in the top layer. Also the static
pressure in the bottom of a batch is greater than near surface, which may result
in different types of sampling bias [2] for example differences in the variation
between replicates. These are examples of layers having an effect on the size
of the sampling error (variation).
In this analysis it is assumed that the variation between replicates represents the
small scale variation. The variation due to the chemical analysis is assumed to
be neglectable.
For the analysis of the small scale variation samples from a number of batches
were simulated according to the hierarchical model (2) with extension (3).
In the simulations σrep,top = 1. As one set of replicates is sampled from each

of four areas within each of three layers, the design is balanced. See also
Table 1.
For each area,

P3 
i = 1, 2, 3
2 k=1 (yijk
− ȳij. )2
Srep,ij = , j = 1, 2, 3, 4 (17)
3−1 

k = 1, 2, 3
provides an estimate of the variance between replicates in that area.
Under the assumption of a normal distribution of replicates, (3), it follows that
2 2
Srep,ij ∈ σrep,i χ2 (2)/2, (18)
see e.g. Montgomery [19], and therefore standard methods for analysis of
linear models for normally distributed observations are not applicable to assess
2
hypotheses concerning σrep,i .
57
Bartlett and Kendall [24] have suggested approximate tests for the analysis
of variance heterogeneity. However, utilizing the relation between the χ2 -
distribution and the gamma-distribution, models for variance heterogeneity
may be investigated using the theory for the so-called generalized linear mod-
els [25].
The hypothesis
2 2 2
σrep,top = σrep,middle = σrep,bottom (19)
may be assessed e.g. by means of the GENMOD procedure of SAS specifying

the model
2
ln(σrep,i ) = µ∗ + αi , (20)
with αtop = 0 and testing the hypothesis
α∗middle = α∗bottom = 0. (21)
In the simulation a test at the 5 % level was performed.
Figure 7 shows the level curves for the power of the test as function of ln(σrep,middle )
and ln(σrep,bottom ). It is seen that the level curves have ellipsoid shape. The
reason is that the power depends on
" #2 " #2
2
ln(σrep,middle 2
) + ln(σrep,bottom ) 2
ln(σrep,middle 2
) − ln(σrep,bottom )
+3
2 2
= (ln(σrep,middle ) + ln(σrep,bottom ))2 + 3 [ln(σrep,middle ) − ln(σrep,bottom )]2
= C.
(22)
58
4
0.954
0.862
0.770
l
o
g
( 0.677
v
a
r
r 2
0.585
e
p
m 0.493
i
d
)
0.401
0.309
0.216
0
0 2 4 0.124
log(var rep bot)

0.05
2
Fig. 7: Level curves for the power of the 5% level test for a dependency of σrep
on layers. The power is plotted against σrep,middle and σrep,bottom .
0.954
0.862
6 0.770
s
t 0.677
d
r
e 0.585
p
b 4
o
t
. 0.493
0.401
0.309
2
0.216
2 4 6 8 0.124
std rep mid.

0.05
2
Fig. 8: Level curves for the power of the 5% level test for a dependency of σrep
on layers plotted against σrep,middle and σrep,bottom .
59
Furthermore Figure 8 shows the probability of declaring a significant differ-

2
ence between σrep,top 2
, σrep,middle 2
and σrep,bottom as a function of σrep,middle
and σrep,bottom . When at least one of σrep,middle and σrep,bottom is more than
4.5 times as large as σrep,top the probability of declaring significance is more
than 0.95.
The shape of the level curves is similar for other values of σrep,top .
4.1 Conclusion
In this section the theory of generalized linear models is presented as a tool to

analyse e.g. the influence of layers on the small scale variation.
For samples simulated from model (2) with the extension (3) it has been shown
that for the 5 % level test the variation between replicates σrep,i has to be 4.5
times larger in one layer than the variation between replicates in another layer
for the effect of layers to be declared significant with a probability of at least
0.95.
Depending on the experimental design and the assumptions made the method
presented in this section can also be used to assess factors influencing sample
error (variation). For example the method can be used to test if one sampling
thief leads to larger variation between replicate samples than another thief.
5 Conclusion
In this appendix a statistical model describing blend homogeneity is presented.

The model is a hierarchical model explicitly taking into account both large,
medium and small scale variation in the blend. Large scale variation is varia-
tion measured on a large scale, e.g., top, middle and bottom layer of the blend.
Medium scale variation is in this model variation between areas within a layer.
Small scale variation is variation between the mean content in the neighbour-
hood samples within an area.
As both large, medium and small scale variation may occur in real blend
60
batches all batches and samples were simulated from this hierarchical model.
However, the analysis of the simulated samples were conducted both in accor-
dance with an aggregated model that as a measure of inhomogeneity uses an
aggregate of the medium and the large scale variation, and in accordance with
a hierarchical model explicitly taking into account the large and the medium
scale variation. In essence the aggregated model relates to the acceptance cri-
teria suggested by FDA [11] and PQRI [18].
The power of the tests of the homogeneity according to respectively the aggre-
gated and the hierarchical model is found. Basically the probability of declar-
ing inhomogeneity is similar for analysis corresponding to each of the two
models. The power of the respective tests are given in more details in Figure 3
to Figure 5.
As the probability of declaring inhomogeneity is basically the same for anal-

ysis corresponding to each of the two models, the aggregated model could be
useful if the purpose of the analysis is to detect inhomogeneity and there is
no special reason to suspect large scale inhomogeneity (i.e. variation between
layers).
However, by using the hierarchical analysis explicit knowledge of respectively

the large and the medium scale variation in the batch is provided. Hence, in
a situation with suspicion of large scale inhomogeneity or if the purpose is to
get more specific knowledge about the homogeneity of the batch, an analysis
in accordance with the hierarchical model should be chosen.
The aggregated as well as the hierarchical model describe inherent variations

in the batch. By an example it has been shown how external factors such as
thieves can be included in the hierarchical model. Other external factors as
sampling technique or sampling personnel could be included in the model as
well (see Appendix F).
In the specific design used in this appendix (with σrep2 = 1) the 5% level test
for the thief effect has a probability of at least 0.95 of declaring significance
when the difference between the mean content in samples collected with each
of the two thieves is larger than 1.5.
The method of Generalized Linear Models was introduced to conduct a sep-

61
arate analysis of the variation between replicate samples. It was shown that
when the variation between replicates is more than 4.5 times as large as the
variation between replicates in another layer, probability of detecting an effect
of layers is at least 0.95.
For all tests in this appendix it should be noted that the power of the tests
corresponds to the specific experimental plan used. For other experimental
plans the power of the corresponding test can be found after same principles as
in this appendix.
Further, when using the methods to assess factors with influence on the mean
content (see Section 3) it is assumed that the variation between replicates is
the same. In the analysis in this appendix only batches with constant variation
between replicates were simulated for this analysis. However, when analysing
the mean content in samples from real batches there is a risk that the variation
between replicates is not constant. The method presented in Section 4 can be
used to verify this assumption. Therefore in practice a suitable procedure is
to use the methods presented in this appendix in reversed order, i.e. first to
analyse the variation between replicates and if this variation is constant, then
continue with the analysis of the mean content. If the variation between repli-
cates is not the same, a so-called weighted squares of means method should be
used (see e.g. [19]).
62
B
Paper B
Comprehensive measures of
blend uniformity
63
65
1 Introduction
An important quality characteristic in tablet production is the homogeneity of

the tablets, i.e. the variation in content of the active ingredient. As this homo-
geneity is closely related to the homogeneity of the blend, the homogeneity is
also an important quality measure of the blend.
One measure of homogeneity is defined in Appendix A. Here homogeneity

is defined as a relevant component of variance being equal to zero. However,
ultimative homogeneity like this is neither necessary nor practical obtainable -
neither for the tablets nor for the blend. Therefore the definition of homogene-
ity in Appendix A is not suitable as acceptance criteria. Instead the quality of
the blend (or the product) is evaluated relatively to practical criteria assessing
if the homogeneity of the blend (product) is satisfactory for the blend (product)
to serve its purpose. Satisfactory homogeneity is here referred to as uniformity.
The issue is then to find an acceptance criteria corresponding to satisfactory

homogeneity/uniformity.
In Appendix A it is shown how to model homogeneity in the blend using three

different scales of homogeneity: large, medium and small scale variation. For
exploratory investigations of the blending process, an analysis based on these
three scales of homogeneity is relevant. However, both existing and suggested
acceptance criteria for blend uniformity are based on the total variation in the
blend, for example through the sample coefficient of variation (see e.g. [2]
and [22]). The aim of the discussion in this appendix is to investigate how
the total variation in a blend can be interpreted in case of different types of
inhomogeneity.
The discussion is exemplified by the first stage of a procedure for blend vali-
dation suggested by PQRI [18]. However, the principles in the discussion of
interpreting the total variation in the blend would be the same had a differ-
ent acceptance criteria been chosen, e.g. the acceptance criteria recommended
by FDA [11] or the standard prediction interval (SDPI) method suggested by
PDA [22].
In the first stage of the procedure suggested by PQRI it is recommended to

assay one sample per area (location) in the batch. Samples should be collected
from at least ten different areas. The relative standard deviation (sample stan-
66 Paper B
dard deviation divided by the average) for the samples should be less than or
equal to 5.0% and all individual results should be within 90.0% - 110.0 % of
the mean result (of all samples). If these two criteria are met the first stage of
the suggested procedure is passed.
Fig. 1: The probability of passing the first stage of the PQRI acceptance criteria
for blend uniformity validation [18]. The probability is plotted as a function of
the total standard deviation in the blend.
The properties (OC-surface) of this acceptance criteria can be investigated for

example as a function of the mean content and the standard deviation in the
batch (see e.g. Appendix D and Appendix C). However, in Figure 1 the prob-
ability of acceptance is plotted as a function of the standard deviation in the
batch. The batch mean content is fixed at label claim, LC 1 . The OC-surface
correspond to a sampling plan where one sample is collected from each of 12
areas in the blend. The plot is based on simulations under the assumption that
the 12 samples are uncorrelated and identically (normally) distributed. The
question regarding the samples being uncorrelated will be discussed later in
1
In this appendix all measurents are given in percent of LC. Hence, for a batch with mean
content equal to label claim, µ = 100
67
this appendix.
Besides a slight variation due to the simulation procedure the plot shows a
slim, smooth curve indicating that in the situation of uncorrelated, identically
(normally) distributed samples, there exists an unambiguous relation between
the uniformity (total variation) of the blend and the probability for the batch
to pass the acceptance criteria. The question is how this relation is affected by
the interpretation of the total variation in the blend.
In the analysis in this appendix it should be noted that the values of the variance
components are no longer measured relatively to the small scale variation as in
Appendix A. All variance components are measured absolutely. For practical
purposes the small scale variation (variation between replicates from the same
area) has been fixed, σrep2 = 1, in these calculations and simulations. The
results in this appendix are only valid for the absolute value of the variance
components on which the calculations are based.
The sampling plan is seen in Table 1.
No. Layer Area

1 top A
2 top B
3 top C
4 top D
5 middle A
6 middle B
7 middle C
8 middle D
9 bottom A
10 bottom B
11 bottom C
12 bottom D
Table 1: Most of the simulations in this appendix are based on 12 samples from
each batch. The distribution of the samples in the batch appears from the table.
68 Paper B
2 Batches with medium scale variation (variation

between areas)
As mentioned above the first stage of the procedure suggested by PQRI [18]
recommends to assay one sample per area in the batch. Samples should be
collected from at least ten different areas. In the simulations discussed in this
appendix samples from twelve different areas were collected. This sampling
plan does not specifically refer to layers in the batch. Therefore, as a first ex-
ample batches with no variation between layers and a fixed standard deviation
corresponding to replicates, σrep = 1, are simulated. With no variation be-
tween layers the hierarchical model (2) in Appendix A from which the batches
are simulated, is identical to the aggregated model (1) in Appendix A, hence,
2
σarea,hi 2
= σarea,agg 2
= σarea .
Two different principles of determining the total variation in the batch is pre-
sented in the following.
2.1 The total variation from the ANOVA table
One way of defining the total variation in the batch is to use the expression for
the total variance obtained from the analysis of variance table corresponding to
an analysis of sample results obtained from the specific sampling plan, in this
case the plan sketched in Table 1. The model for this analysis is
Yij = µ + Ai + Ei , i = 1, .., 12. (1)
Yij is the content of a sample, µ the mean content in the batch, Ai the effect
corresponding to the area from which the sample is taken, and Ei is the error
term also including the variation due to sampling. As only one sample from
each area is collected the terms Ai and Ei have identical indices. Further
2
Ai ∈ N (0, σarea,agg 2 ). The analysis of variance table
) and Ei ∈ N (0, σrep
corresponding to this model is given in Montgomery [19].
In this case where one sample is collected from each of twelve areas the anal-
ysis of variance (ANOVA) table is
69
ANOVA Table
Source of Sums of Degrees of Expected mean square
variation squares freedom (df) (EMS)
2 2
Between areas SSarea,agg 12 - 1 σarea,agg + σrep
2
Error SSrep 0 σrep
Total SStotal 12 - 1
Table 2: The analysis of variance table corresponding to model (1).
As the terms in the ’sums of squares’ column are additive the total variation in
2
the blend, σtotal , assessed under this scheme is estimated from
2 SSarea,agg + SSrep
σtotal =
dftotal
dfarea,agg × M SEarea,agg + dfrep × M SErep
=
dftotal
2 2 )
11 × (σarea,agg + σrep
=
11
2 2
= σarea,agg + σrep . (2)
2.2 The variance on a randomly sampled unit from the batch
Another definition of the total variation in the blend is to consider the variance
2
on a randomly sampled unit. In relation to model (1) this variance is σsample =
2 2
σarea,agg + σrep . In this situation with one replicate per area and no variation
between layers, the total variation in the blend determined from the ANOVA
table is identical to the variance on a randomly sampled unit from the blend.
Figure 2 shows the level curves for the OC-surface of the PQRI acceptance
criteria as a function of the mean content and the total standard deviation in the
blend, σsample = σtotal , determined from the expression (2).
The figure shows that when the total standard deviation is less than approxi-
mately 3%, the probability of passing the first stage of the PQRI criteria is at
least 0.95. Further, it should be noted that even batches with a mean content
relatively far from LC (µ = 100) have a relative high probability of accep-
70 Paper B
10
0.950
0.850
8
0.750
s 0.650
t
d 6
b 0.550
a
t
c
h 0.450
4
0.350
0.250
2
0.150
90 100 110 0.050
mu
pass
Fig. 2: Level curves for properties (OC-surface) of the first stage of the pro-
cedure for blend validation suggested by PQRI [18]. The OC-surface is plot-
ted against the mean content in the blend, µ, and the total standard devia-
q in the blend in a situation with no variation between layers, σtotal =
tion
2
σarea,agg 2 . Under model (1) the expression for the total standard de-
+ σrep
viation determined from the ANOVA table is identical to the expression for
the total standard deviation on a randomly sampled unit from the blend, i.e.
σtotal = σsample . The standard deviation corresponding to replicates is fixed,
σrep = 1.
71
tance. The reason is that the PQRI criteria allows for sampling bias of content.
Allowing the mean content of samples to deviate from LC, reduces the risk of
rejecting homogeneous batches with a mean content close to LC in cases of
sampling bias. This is in contrast to the USP criteria for tablets which besides
uniformity also controls deviation from LC (see Appendix D and Appendix C).
Further it is seen that the level curves are rather smooth and almost linear and
almost independent of the mean content. In these simulations the level curves
2 +1
depend weekly on the mean content through CV = σtotal µ = σareaµ . Thus the
test is essentially a test on CV even though it also includes limits for individual
observations.
In this situation with fixed variation between replicates and no variation be-
tween layers, there exists an unambiguous relation between the probability of
passing the PQRI test and the total standard deviation in the batch. This is also
seen from Figure 3 which is a sectional plane of the OC-surface in Figure 2
for µ = 100. From superimposing Figure 3 on Figure 1 it is also seen that
the total standard deviation in the batch obtained from combining the variation
among areas and the replicate variation in accordance with expression (2) is
identical to the the total standard deviation among 12 uncorrelated, identically
(normally) distributed samples in Figure 1. This is not surprising as the areas
from which each of the 12 samples in model 1 is sampled are uncorrelated.
As the OC-surface is only moderately dependent of the mean content in the

batch as seen in Figure 2, the rest of the analysis in this appendix will for
simplicity focus on blends with a mean content equal to LC.
3 Batches with large scale variation (variation be-

tween layers)
2
In this section batches with large scale variation, σlayer , (and fixed small scale
2
variation, σrep = 1), but no medium scale variation, σarea,hi , are considered.
72 Paper B
Fig. 3: The probability of passing stage 1 of the PQRI criteria in a situation

with no variation between layers and fixed variation between replicates. The
probability is plotted against the total standard deviation in the batch deter-
mined from expression (2) and correspond to a sampling plan with one sample
from each of 12 different areas.
73
With three layers and four samples from each layer a model for the content in
a sample taking into account only large scale variation is
(
i = 1, 2, 3
Yij = µ + Li + Eij , . (3)
j = 1, 2, 3, 4
Yij is the content of a sample, µ the mean content in the batch, Li the effect
corresponding to the layer from which the sample is taken, and Eij is the error
2
term also including the variation due to sampling. Further Li ∈ N (0, σlayer )
2
and Eij ∈ N (0, σrep ).
The corresponding ANOVA table is
ANOVA Table
2 2
Between layers SSlayer 3-1 4σlayer + σrep
2
Error SSrep 3(4-1) σrep
Table 3: The analysis of variance table corresponding to the model (3).
From Table 3 the estimate of the total variation in the batch is
2 SSlayer + SSrep
σtotal =
dftotal
dflayer × M SElayer + dfrep × M SErep
=
dftotal
2 2 + 9σ 2
8σlayer + 2σrep rep
=
11
8 2 2
= σlayer + σrep . (4)
11
The OC-surface is plotted as a function of the standard deviation corresponding

to this expression with dots in Figure 4. µ is fixed at LC.
74 Paper B
Fig. 4: The OC-surface plotted as a function of various measures of the to-

tal standard deviation in the batch for µ = LC and σrep = 1. The dotted
line corresponds
q to the total standard deviation as derived from expression (4),
8 2 2 , and the circles correspond to the total standard de-
σtotal = 11 σlayer + σrep
viation defined as theqtotal standard deviation on a random unit sampled from
the blend, σsample = σlayer2 2 .
+ σrep
75
3.2 The variance on a sample from the batch
2
According to model (3) the variation on a sample from the blend is σsample =
2 2
σlayer + σrep . Thus, with the given sampling plan and under model (3) the
2
expression for the total variation in the batch according to Table 3, σtotal , and
the expression for the variation on a randomly sampled unit from the blend,
2
σsample , are not identical. The variation on a sample from the blend is al-
ways the sum of the variance components corresponding to the given model,
whereas the expression for the total variation derived from an ANOVA table is
only a sum with all coefficients equal to one in a few cases, depending on the
corresponding sampling plan.
As the dotted line in Figure 4 is more steep than the circled line it appears
that in a batch with large scale but no medium scale variation the acceptance
criteria is more efficient at separating batches with good quality from batches
with a low quality when the measure of total variation is as derived from the
ANOVA table rather than from the expression for the variance on a randomly
sampled unit from the batch.
Further as both lines in Figure 4 is less steep than the line in Figure 3 it is seen
that under a sampling plan with 12 observations from three layers and four
areas within each layer the acceptance criteria is more sensitive in the case of
medium scale variation but no large scale variation than in the case of large
scale variation and no medium scale variation in the blend.
4 Batches with both large and medium scale vari-

ation
In real batches both large and medium scale variation may occur at the same
time. Therefore, in this section expressions for the total blend homogeneity
is derived for batches with both large and medium scale variation (and fixed
small scale variation).
76 Paper B
Under a sampling plan with three layers, four areas within each layer and (in
accordance with the first stage of the PQRI procedure) only one sample from
each area, the hierarchical model (2) from Appendix A reduces to
(
i = 1, 2, 3
Yijk = µ + Li + A(L)j(i) + Eij , . (5)
j = 1, 2, 3, 4
Li is the effect from layer i, A(L)j(i) is the effect from the j’th area within
layer i and Eij is the error term also including the variation due to sampling.
2
Further it is assumed that Li ∈ N (0, σlayer 2
), A(L)j(i) ∈ N (0, σarea,hi ) and
2
Eij ∈ N (0, σrep ).
The corresponding ANOVA table is
ANOVA Table
2 2 2
Between layers SSlayer 3−1 4σlayer + σarea,hi + σrep
2 2
Between areas, SSarea,hi 3(4 − 1) σarea,hi + σrep
within a layer
2
Error SSrep 0 σrep
Table 4: The analysis of variance table corresponding to the hierarchical

model (5).
2
In analogy with the derivations of σtotal in the previous sections, the expression
2 2 8 2 2 2 . The OC-
for σtotal derived from Table 4 is σtotal = 11 σlayer + σarea,hi + σrep
surface is plotted as a function of the corresponding standard deviation, σtotal ,
with dots in Figure 5 to Figure 7. The difference between the three figures is
that the mean content is fixed at respectively 90%, 100% and 110 % of LC. It
is seen that the plots in the three figures are very alike in agreement with the
fact that the acceptance criteria is only moderately dependent on the value of
the mean content in the batch, µ.
It is also seen that for this OC-surface there is not an unambiguous relation
77
between the total variation in the blend and the probability of passing the test 2 .
Figure 5 to Figure 7 is the resulting plot from superimposing all plots for spec-
ified values of respectively variation between areas and variation between lay-
ers. If σarea is fixed at same value the plot is a parallel displaced version of
Figure 4. If σlayer is fixed at same value the plot is a parallel displaced version
of Figure 3.
Fig. 5: The dots correspond to the total standard deviation as derived from
2 8 2 2 2 , and the circles correspond to the total
σtotal = 11 σlayer + σarea,hi + σrep
2
standard deviation on a randomly sampled unit from the blend, σsample =
2
σlayer 2 2
+ σarea,hi + σrep . The mean content in the batch is 90% of LC.
It should be remarked that if there had been three replicates per area as in
the batches simulated in Section 2 in Appendix A, the expression for the total
variation in the batch derived from the corresponding ANOVA table would be
8 2
2
σtotal = 35 σlayer + 33 2 2
35 σarea,hi + σrep .
2
In the simulations σarea,hi and σlayer vary between 0 and 10. Had larger values for the
two parameters been included in the simulations the OC-band would have been even broader in
these figures
78 Paper B
2
2 2 2
σlayer + σarea,hi + σrep . The mean content in the batch is 100% of LC.
79
2
2 2 2
σlayer + σarea,hi + σrep . The mean content in the batch is 110% of LC.
80 Paper B
The OC-surface is plotted as a function of the standard deviation corresponding

to this expression with dots in Figure 8. Comparing the dotted OC-surfaces in
Figure 6 and Figure 8 shows that under this hierarchical sampling plan the
relation between the total variation in the blend and the probability for a batch
to pass the acceptance criteria is even less clear with three replicates per area
than under the hierarchical sampling plan with only one observation per area.
Fig. 8: The OC-surface for a hierarchical sampling plan with a total of 36

samples. qThe dots correspond to the total standard deviation derived from
8 2
σtotal = 35 σlayer + 33 2 2
35 σarea,hi + σrep , and the circles correspond to the to-
q standard deviation on a randomly sampled unit from the blend, σsample =

tal
2
σlayer 2
+ σarea,hi 2 . The mean content in the batch is 100% of LC.
+ σrep
Figure 9 is similar to Figure 8 except that the criteria for the individual tablets
are disregarded. It is seen that as long as the criteria for the sample coefficient
of variation is included, the criteria on the individual tablets does not have a
large effect on which batches are accepted by the acceptance criteria.
81
Fig. 9: The figure is similar to Figure 8 except that the criteria for the individual
tablets are left out. The dots correspond to σtotal and the circles correspond to
σsample .
82 Paper B
4.2 The variance on a sample from the batch
The total variance of a randomly sampled unit from the hierarchical model (5)
2
is σsample 2
= σlayer 2
+ σarea,hi 2 .
+ σrep
The circles in Figure 5 to Figure 9 corresponds to the total variation in the

batch expressed as σsample . As the OC-bands corresponding to dots in gen-
eral is more narrow than the OC-bands corresponding to circles it is seen that
both under the hierarchical model with 12 samples and the hierarchical model
with 36 samples, the acceptance criterion is more unambigious related to the
measure of total variation, σtotal , than to σsample .
4.3 The direct relation between large/medium scale varia-

tion and the acceptance criteria
In the previous sections various expressions for the total variation in the blend
has been derived as a function of the variation between respectively layers, ar-
eas and between replicates. These expressions are not always straight forward
and they are dependent on the specific sampling plane. Thus in case of both
medium and large scale variation (and fixed variation between replicates) in
the blend it may be more straight forward to investigate the OC-surface for the
acceptance criteria directly as a function of σlayer and σarea,hi respectively. In
Figure 10 through Figure 12 this relation is plotted for the hierarchical model
with 12 samples. In the three figures µ is fixed at respectively 85 % LC, 100
% LC and 115 % LC. The variation between replicates is fixed, σrep = 1.
5 Discussion and conclusion
The aim of this appendix is to investigate how a concept of the total variation
in the blend, can be interpreted in case of inhomogeneities in the blend.
The interpretation of the total variation is relevant as acceptance criteria for

blend uniformity is based on considerations on the total variation.
83
10
0.950
0.850
0.750
s 0.650
t
d
l 5 0.550
a
y
e
r 0.450
0.350
0.250
0.150
0
0 5 10 0.050
std area
pass
Fig. 10: Level curves for the OC-function of the first stage of the acceptance
criteria for blend uniformity suggested by PQRI. The mean content in the
batch, µ, is 85 % LC and the variation between replicates is fixed, σrep = 1.
84 Paper B
10
0.950
0.850
0.751
s 0.651
t
d
l 5 0.551
a
y
e
r 0.451
0.351
0.252
0.152
0
0 5 10 0.052
std area
pass
85
10
0.950
0.850
0.750
s 0.650
t
d
l 5 0.550
a
y
e
r 0.450
0.350
0.250
0.150
0
0 5 10 0.050
std area
pass
86 Paper B
In case of no variation between layers and under a hierarchical sampling plan

with only one replicate per area, the measure for the total variation in the blend
derived from the analysis of variance table and the total variation defined as
the variance on a random sample from the blend are identical. The reason is
that in this case the samples are uncorrelated. However, in presence of large
scale variation or when collecting more than one replicate from each area in
presence of medium scale variation the two expressions are not identical.
In general the total variation is defined from the analysis of variance table leads
to more effective (steeper OC-surface) and less ambiguous (more narrow OC-
band) acceptance criteria than the total variation defined as the total variation
on a random sample from the batch. However, the expression for total variation
defined as the total variance from the analysis of variance table, depends on the
sampling plan used.
Paper C
On a test for content

uniformity in pharmaceutical
products
Presented at the First Annual ENBIS Conference, Oslo 2001
C
87
Presented at ENBIS 2001 89
Abstract
A proposed amendment of the current procedure in the US Pharmacopeia for

testing content uniformity in tablets is a two-stage sampling plan to be applied
to each batch during validation.
In statistical terms, the procedure at each stage is a combination of a test by at-

tributes with acceptance number zero with “specification limits” ±25 % around
the target value, and a test by variables with “specification limits” ±16.5 %
around the target value. In the paper we discuss the statistical properties of this
procedure.
Keywords
Specification limits, inspection by attributes, inspection by variables, OC-curve,

acceptance sampling.
1 Introduction
To assure the therapeutic utility of dosage units such as tablets, the drug content
of each unit in a batch should not deviate too much from a chosen target value,
e.g. label claim (LC).
A procedure to control the variability (uniformity) of the content of compressed

tablets in a batch is the test for content uniformity specified by the US author-
ities [26].
The current test for content uniformity is a two-stage test including tests by
attributes for the content of single tablets in the sample expressed relatively to
the label claim (LC) as well as a limit for the sample coefficient of variation.
Pharmaceutical companies often sell their products in several regions of the

world and therefore have to comply with the requirements in each of the coun-
tries and areas in which they sell their products. In order to reduce the need to
90 Paper C
duplicate the testing carried out during the research and development of new
drugs, efforts are made to achieve greater harmonisation in the interpretation
and application of technical guidelines and requirements for product registra-
tion.
Regarding testing for content uniformity this has so far resulted in a proposal
from the US authorities for an alternative to the current test procedure.
The proposed test procedure, which is the subject of this article, is a two-stage
test using in each stage a test by attributes on the content of the individual
tablets combined with a test by variables.
2 The proposed test
2.1 Historical notes
The currently valid requirement from the US authorities concerning content

uniformity is termed USP 24 [26]. This test is a two stage sampling plan
with ten tablets tested on stage 1 and further twenty tablets tested on stage 2.
On each stage the test has two requirements to the drug content of individual
tablets in the sample: At most one tablet is allowed to be outside the limits
1.0 ± 0.15 LC and no tablet outside the limits 1.0 ± 0.25 LC. Actually, the
limits for the individual tablets depend on the sample mean. Further the sample
coefficient of variation shall be less than 6 % and 7.8 % respectively for stage
1 and stage 2.
The first proposed amendment to the requirements in the form described above
was moved in 1997 [27]. The amendment has been commented and discussed
over the last years for the present resulting in a proposed test described in Phar-
macopeial Forum [28]. The version of the proposed test under concern in this
article is described in Pharmacopeial Forum [29]. However, the content of the
various versions of the proposed test are essentially the same, and somewhat
different from the the approach in the current requirement.
The current USP specification controls the content variability by establishing

a limit on the sample coefficient of variation. Further the test includes limits
for the content in individual tablets in the sample. In this procedure the mean
content is not directly controlled. This is in line with the test being a test for
uniformity, i.e. variability, and not explicitly for mean content. The proposed
test, however, controls content variability by means of a linear combination of
the sample standard deviation and the absolute mean deviation from the target.
Also in this procedure a test for the content in individual tablets is included.
With the linear combination including sample standard deviation and mean
in the new procedure not only the variability but also the mean is controlled.
Thus, what historically started as a test for uniformity, still is termed a test for
content uniformity, but now also controls the mean content.
The details of the proposed test, see [29], are described in the following.
2.2 Description of the proposed test
The proposed test is a two stage test which combines a criterion on the sample
mean and standard deviation, (x̄, s), with a criterion on individual tablets. The
details of the testing procedure is outlined schematically in Figure 11 .
Before the testing procedure is started 30 tablets are selected from the batch.
On stage 1 of the testing procedure a sample of 10 tablets is investigated. Test

stage 1 may either lead to acceptance, rejection, or to further testing on stage 2.
The batch of tablets is accepted at this stage if the sample mean and standard
deviation, (x̄, s), are within the acceptance limits, and all individual tablets are
within limits for individual tablets specified by the procedure. These limits
depend on the sample mean. The batch is rejected if the content of one or more
individual tablet in the sample is outside the limits for individual tablets. If
no individual tablets is outside the limits but the combination of sample mean
and sample standard deviation is outside the acceptance limits, further testing
is performed in accordance with the criteria for test stage 2.
The sample used for test stage 2 consists of the 10 tablets from stage 1 com-
bined with the remaining 20 tablets. Stage 2 results in either acceptance or
1
For simplicity only the part of the proposed test that concerns compressed tablets with a
rubric mean less than 101.5 percent of label claim is considered. The rubric mean is the average
of the shelf limits specified in the potency definition of the drug.
92 Paper C
Stage 1
All of 10 tablets within the acceptance

limits for individual tablets
NO YES
REJECT
The combination of
sample mean and sample
standard deviation falls YES
within the acceptance area ACCEPT
for stage one
NO
Stage 2
Add 20 tablets to the first sample

to a total sample of 30 tablets. All 30
tablets within the acceptance limits
for individual tablets
NO YES
REJECT
The combination of
within the acceptance area
ACCEPT
for stage two
NO
REJECT
Fig. 1: Schematic representation of the proposed test for content uniformity.

Acceptance limits Acceptance limits

x̄ for s for individual
tablets in the sample
Stage 1
x̄ ≤ 0.835 - 0.74 − 1.23
0.835 ≤ x̄ ≤ 0.985 s ≤ (x̄ − 0.835)/2.3 0.74 − 1.23
0.985 ≤ x̄ ≤ 1.015 s ≤ 0.065 0.75x̄ − 1.25x̄
1.015 ≤ x̄ ≤ 1.165 s ≤ (1.165 − x̄)/2.3 0.76 − 1.27
1.165 ≤ x̄ - 0.76 − 1.27
Stage 2
x̄ ≤ 0.835 - 0.74 − 1.23
0.835 ≤ x̄ ≤ 0.985 s ≤ (x̄ − 0.835)/2.0 0.74 − 1.23
0.985 ≤ x̄ ≤ 1.015 s ≤ 0.075 0.75x̄ − 1.25x̄
1.015 ≤ x̄ ≤ 1.165 s ≤ (1.165 − x̄)/2.0 0.76 − 1.27
1.165 ≤ x̄ - 0.76 − 1.27
Table 1: Acceptance limits, measured relatively to label claim (LC), for the sample standard
deviation and for individual tablets. In general the limits are functions of the sample mean. If
x̄ ≤ 0.835 LC or x̄ ≥ 1.165 LC the batch is rejected. x̄ and s are based on a sample of 10
tablets on stage 1 and the total sample of 30 tablets on stage 2.
rejection of the batch. The batch is accepted if criteria for individual tablets as
well as criteria for the sample mean and sample standard deviation are satisfied.
Otherwise the batch is rejected.
Acceptance limits for individual tablets and the sample standard deviation as
a function of the sample mean are shown in Table 1. The corresponding ac-
ceptance area in the (x̄, s)-plane for the test on sample mean and standard
deviation is shown in Figure 2.
3 Properties of the proposed test
The proposed test evaluates a batch on the basis of the content and the varia-
tion of the active component (drug) in the batch. Therefore, assuming a nor-
mal distribution of tablet content, the properties of the test is described by the
94 Paper C
Acceptance area
0.14
0.12
Stage 1
Stage 2
sample std. dev. as fraction of LC
0.10
0.08
0.06
0.04
0.02
Acceptance area
0.0
0.7 0.8 0.9 1.0 1.1 1.2 1.3

sample mean as fraction of LC
Fig. 2: Acceptance area for sample mean content and standard deviation.
probability for a batch to be accepted by the test as a function of the mean

and standard deviation, (µ, σ), of the content in the batch. Figure 3 shows the
level-curves for the OC-surface for the test. The combination of requirements
on individual tablets with requirements on x̄ and s complicates the derivation
of an exact analytical expression for the OC-surface. Instead the OC-surface is
based on Monte Carlo simulations in S-Plus version 6 [30]. 3000 simulations
were performed for each of various combination of µ and σ and level curves
have been determined by interpolation.
In this article the influence of measurement errors on the OC-surface is not

taken into account. For a discussion of the effect of measurement errors in
acceptance sampling see [31] and [32].
3.1 Description of the OC-surface of the test
From Figure 3 it is seen that when the standard deviation, σ, exceeds 0.095 LC
the probability of accepting the batch is less than 0.1 no matter what the mean
content, µ, is. Similarly, when the mean content is less than 0.835 LC or larger
than 1.165 LC the probability of accepting the batch is less than 0.1, no matter
Total probability of acceptance,

levels: 0.1, 0.25, 0.5, 0.75, 0.90, 0.99
0.14
0.12
0.10
0.10
0.08
sigma
0.1
0.06
0.50
0.04
0.99
0.02
0.25
0.10
0.25
0.50
0.75
0.0 0.90
0.99
0.7 0.8 0.9 1.0 1.1 1.2 1.3
mu
Fig. 3: Level-curves for the OC-surface for the test. The curves are based on 3000 simulations
for each lattice point (µ, σ).
how small the standard deviation in the batch is.
Within the innermost triangular-like shape in Figure 3 the probability of ac-

cepting the batch is 0.99 or larger.
3.2 “Specification limits” for individual tablets
As a single quality measure for a batch, one might consider the proportion
of tablets in the batch with a content outside some specified range of values.
The U.S authorities have not explicitly specified limiting values for the content
of tablets in the batch, nor is an unacceptable proportion specified. However,
the acceptance limits for individual tablets in Table 1 suggests that it is not
considered satisfactory for the content in a tablet to deviate more than 25%
from label claim. Figur 4 shows the proportion of tablets with a content outside
1 ± 0.25 LC in a batch as a function of µ and σ.
Comparing the proportion of tablets in a batch outside these limits with the
probability for acceptance of the batch is a way to evaluate the properties of
96 Paper C
Proportion outside the limits 0.75-1.25,

levels 0.0001, 0.0005, 0.001, 0.005, 0.01, 0.05, 0.10, 0.15
0.14
0.12
0.10
0.08
sigma
0.06
0.15
0.04 0.0001
0.02
0.0005
0.1500
0.1000
0.0500
0.0100
0.0050
0.0005
0.0010
0.0001
0.0
0.7 0.8 0.9 1.0 1.1 1.2 1.3
mu
Fig. 4: Level-curves for the proportion of tablets in a batch outside the limits 1 ± 0.25 LC.
the test: Batches with a large proportion of tablets outside these limits should
have a high probability of being rejected by the test, and batches with a very
small proportion of tablets outside the limits should have a high probability of
acceptance.
Superimposing Figure 3, the probability for a batch to be accepted by the test,

upon Figure 4, the quality of the batch, reveals that batches with a probability
of 0.9 of being accepted have a proportion of tablets outside the specification
of less than 0.0001.
As the curves in the two figures do not have the same shape it is not possible
to determine the probability of rejecting a batch as a single-valued function
of the proportion of tablets outside the specification, i.e. it is not possible
unambiguously to determine which proportion of unacceptable tablets in the
batch will lead to rejection with a given probability. As an example in some
cases batches with a proportion of 0.0001 outside the specification is rejected
with a probability of more than 0.9. However, in other cases, depending on the
value of mu, such batches have a probability of 0.99 of acceptance.
The acceptance limits for the sample standard deviation as a function of the
sample mean (Figure 2 and Table 1) suggest also to consider the specification,
1 ± 0.165 LC, for the content in a tablet.
The proportion of tablets in the batch outside these limits as a function of the
batch mean and standard deviation (µ, σ), is shown in Figure 5.
Superimposing Figure 3 upon Figure 5 shows that the shape of the level-curves
in the two plots are similar. This means that it is (almost) possible to determine
a unique relation between the proportion of tablets outside these limits and the
probability of acceptance by the USP test.
Proportion of tablets outside the limits 0.835-1.165;

levels 0.005, 0.01, 0.02, 0.03, 0.04, 0.05, 0.09
0.14
0.12
0.10
0.08
sigma
0.06
0.09
0.04 0.005 0.090
0.02
0.090
0.050
0.040
0.030
0.020
0.010
0.005
0.0
0.7 0.8 0.9 1.0 1.1 1.2 1.3
mu
Fig. 5: Level-curves for the proportion of tablets in the batch outside the limits 1 ± 0.165 LC.
Batches with a proportion of less than 0.005 outside this specification have a
probability of more than 0.99 of acceptance by the test. Batches with a pro-
portion of 0.09 outside this specification corresponds to a probability of 0.1 of
acceptance.
3.3 Details on the effect of individual elements of the test
The purpose of the test is to assess whether a batch may be considered satisfac-
tory (accept) or not (reject). This discriminatory ability was assessed above.
98 Paper C
Probability of acceptance - disregarding the attribute test,

levels: 0.1, 0.25, 0.5, 0.75, 0.99
0.14
0.12
0.10
0.10
0.08
sigma
0.1
0.06
0.50
0.04 0.99
0.02
0.25
0.10
0.25
0.50
0.75
0.0 0.99
0.7 0.8 0.9 1.0 1.1 1.2 1.3
mu
Fig. 6: Level-curves for the OC-surface for the test - disregarding the attribute test. The
level-curves are based on 3000 simulations for each lattice point (µ, σ).
However, the discriminatory ability is the result of the combined properties

of the individual elements of the test. These elements are a test by attributes
and a test by variables in each of two stages of the test. The test by attributes
is (essentially) a test for the proportion of tablets outside 1 ± 0.25 LC. The
shape of the acceptance region for (x̄, s) in Figur 2, resembles the acceptance
region for the test by variables controlling the proportion of tablets outside a
given specification as derived by Lieberman and Resnikoff [33] and described
by Schilling [12]. This suggests that the proposed test by variables for prac-
tical purposes controls the proportion of tablets outside 1 ± 0.165 LC. As the
specification 1 ± 0.25 LC is less restrictive than the specification 1 ± 0.165 LC
it is indicated that the test by variables is the effective part of the test.
To further investigate the effect of respectively the test by variables and the test
by attributes the level-curves for the OC-surface is plotted for a test procedure
disregarding the attribute test. These level-curves are shown in Figure 6. Su-
perimposing this figure upon Figure 3, the level-curves for the total test, reveals
that the probability for a batch to be accepted by the total test virtually does not
depend on whether the attribute test is included in the test or not. This is in line
with the fact that under the assumption of normality (x̄, s) are jointly sufficient
for (µ, σ), and hence the conditional distribution of the content of individual
tablets in the sample for a given combination (x̄, σ) does not depend on the
value of (µ, σ).
Now, consider the two stages of the test. Figure 7 shows the probability of
invoking test stage 2. Batches corresponding to a (µ, σ)-combination in the
inner triangular area have a high probability of acceptance on stage 1. Batches
corresponding to a (µ, σ)-combination outside the plotted level-curves have a
high probability of being rejected on stage 1, whereas batches corresponding
to a (µ, σ)-combination in the ’sausage’-shape in the middle have a high prob-
ability of invoking test stage 2. For these latter batches 30 tablets instead of
10 shall be analysed, i.e. such batches are expensive regarding time and other
ressources to the testing and chemical analysis.
Probability of invoking stage 2,

levels: 0.05, 0.1, 0.25, 0.5, 0.75, 0.99
0.14
0.05
0.12
0.10
0.08
sigma
0.06
0.04
0.99
0.99 0.05 0.99
0.02
0.05
0.10
0.25
0.50 0.75
0.50 0.99 0.05
0.25
0.50
0.10
0.0 0.75
0.99 0.25
0.10
0.05
0.7 0.8 0.9 1.0 1.1 1.2 1.3
mu
Fig. 7: The probability of invoking stage 2 when both the variable and the attribute tests are
included in the test procedure. The level-curves are based on 3000 simulations for each lattice
point (µ, σ).
Figure 8 shows the probability of invoking test stage 2 when the attribute test
on individual tablets is disregarded in the procedure. Batches corresponding
to a (µ, σ)-combination in the inner triangular area have a high probability of
acceptance on test stage 1. However, disregarding the test for individual tablets
rejection on stage 1 is not possible, implying that all batches corresponding to
a (µ, σ)-combination outside the inner triangular area would have a high prob-
ability of invoking test stage 2, hence requiring the efforts of further testing.
100 Paper C
Probability of invoking stage 2 - disregarding the attribute test,

levels: 0.05, 0.1, 0.25, 0.5, 0.75, 0.95
0.14
0.12
0.10
0.08
sigma
0.06
0.95
0.04
0.05
0.02
0.95
0.75
0.50
0.0 0.25
0.10
0.05
0.7 0.8 0.9 1.0 1.1 1.2 1.3
mu
Fig. 8: The probability of invoking stage 2 - disregarding the attribute test. The level-curves
are based on 3000 simulations for each lattice point (µ, σ).
Thus, a very important effect of the test by attributes is to reduce ressources

to the testing and the chemical analysis by rejecting most of the unacceptable
batches already on stage 1. Actually, it is possible to design criteria for re-
jection on stage 1 that are even more effective than the above attribute test.
Schilling [12] describes a double sampling plan by variables that allows for
rejection on stage 1.
Other benefits of including the test by attributes in the test procedure is that the
test provides robustness of the procedure in case of a non-normal distribution
of the tablet content. The discriminating properties of the test in the case of
non-normal distribution of the tablet content have not been investigated. The
limits for individual tablets may in these situations serve as a safety precaution.
Finally the test by attributes may serve a psychological purpose. The reason
is that the use of a test by variables - without a test by attributes - might (in
extreme cases) lead to acceptance of a batch even though tablets of a very ex-
treme quality is found in the sample. The attribute specification in the proposed
test prevents such a situation.
4 Comparison to other test procedures
It is well understood in most commercial sectors that complete testing of prod-

uct is ressource-demanding, and - in case of destructive testing - even impossi-
ble, and therefore there is a long industrial tradition for acceptance of product
based upon inspection of a sample. Statistical theories for acceptance sampling
by attributes dates back to the pioneering work by Molina, Dodge and Romig
at Bell Telephone Laboratories in the 1920’s and the theoretical basis for ac-
ceptance sampling by variables was given by Lieberman and Resnikoff [33]
in the 1950’s. The current international standards for acceptance sampling by
attributes, ISO 2859 [34], has been widely accepted by industry as pragmatic
rules to be used in agreements by two parties for releasing product after inspec-
tion of only a limited sample of the product. Because of the greater efficiency
of sampling by variables, the complementary standard, ISO 3951 [35], for ac-
ceptance sampling by variables is also used in many industrial relations.
Although the current trend in quality management is to shift the focus from the
final product inspection to the monitoring of the process, acceptance sampling
may still serve a purpose as part of quality control procedures as described e.g.
in the quality management standard QS 9000 [36] developed by the automo-
tive industry and the US military standard, MIL-STD-1916 [37] for acceptance
of product. However, in line with the quest for “zero defects”, the acceptance
sampling plans suggested by these standards are plans with acceptance num-
ber zero (“accept zero plans”) for sampling by attributes, and equivalent plans
for sampling by variables. Although the assumption of normally distributed
item characteristics is seldomly questioned in the calculation of process ca-
pability and performance measures, there is however, some reluctance towards
using variables sampling plans, partly because of the underlying assumption of
a normal distribution, and partly because the use of a variables sampling plan
might (in extreme cases) lead to acceptance of product even when a noncon-
forming item is found in the sample. Such situations are in apparent conflict
with the “ accept zero” philosophy.
Although acceptance sampling of product in commercial transactions between

two parties has a more pragmatic purpose than sampling inspection for regula-
tory purposes they share the common goal of establishing evidence that product
is of a satisfactory quality.
Thus, the EC legislation on weights and measures of prepackaged goods [38]

102 Paper C
and [39] lays down a sampling procedure to be used by the authorities for ver-
ifying compliance to the labelled content. The procedure is a combination of a
test by attributes for the proportion of packages with content less than specified,
and a separate test by variables leading to rejectance when the sample average
is significantly smaller than the specified content in the packages. Although
the test by variables is only concerned with the mean content, the combined
result of the two tests is an economic incentive to the producer to maintain a
small variance in order to avoid overfill.
However, in case of drugs, “overfill” is of just as much concern as “under-

fill”, and therefore positive as well as negative deviations from label claim are
undesirable.
Standard acceptance sampling plans for industrial use are based upon the usual
industrial practice of setting up specifications for the final product (and all pre-
vious stages). Such specification limits, or tolerances, facilitate the communi-
cation between the parties and provide a basis for assessing quality simply by
verifying compliance to specifications.
In the USP-proposal no explicit specification limits for the content of individ-

ual tablets has been laid down. Thus, in contrast to standard acceptance sam-
pling procedures for product, the quality requirement to the product is specified
only through the acceptance sampling procedure.
However, our analysis of the operations characteristics of the procedure pro-

vides some guidance to the producer and shows some analogies to standard
procedures. Thus, the proposed procedure in essence is a procedure for con-
trolling the proportion of tablets with content outside the 1 ± 0.165 LC, com-
bined with an “accept zero” plan corresponding to the limits 1 ± 0.25 LC.
5 Conclusion
As a result of the efforts of achieving international harmonisation in the inter-

pretation and application of technical guidelines and requirements for product
registration the procedure for testing of content uniformity in tablets is under
revision.
The proposed procedures all include both a test by attributes and a test by
variables whereas the current procedure includes test by attributes and a test of
the relative coefficient of variation.
In this article the discriminating properties of the proposed test procedure have
been assessed. Under the assumption of normally distributed content in the
tablets the qualities, (µ, σ), leading to acceptance of a batch have been de-
lineated. The analysis in this article reveals that the proposed procedure in
essence is a procedure for controlling the proportion of tablets with content
outside the 1 ± 0.165 LC, combined with an “accept zero” plan corresponding
to the limits 1 ± 0.25 LC.
Further the effect of individual elements of the test procedure have been as-
sessed.
The acceptance of a batch is determined by the test by variables. The function

of the test by attributes is to reduce the ressources to the testing and chemi-
cal analysis, to reject unsatisfactory batches in situations of non-normally dis-
tributed content in the tablets and finally it has a psychological effect as the
use of a variables sampling plan might (in extreme cases) lead to acceptance
of product even when a tablet of very extreme content is found in the sample.
104 Paper C
Paper D
Statistical tests for

uniformity of blend/dose
samples
D
105
107
1 Introduction
Pharmaceutical companies are legally required to manufacture their products

using current Good Manufacturing Practices (cGMPs) as defined, e.g. in doc-
uments from regulatory authorities. In pharmaceutical production as well as in
other industrial sectors an important goal of good manufacturing practice is to
control variability in the characteristics of the end product, and therefore good
manufacturing practice implies monitoring of processes that may be responsi-
ble for causing variability in the characteristics of the end product, see e.g. the
guidelines provided in the International Standard ISO 11462-1, [40].
In pharmaceutical production of tablets, a key process is the blending process

producing the powder mix, and therefore most legal requirements prescribe
control of blend processes with the purpose of demonstrating that a satisfactory
degree of mixing has been achieved. Also, requirements on the monitoring of
the final product (tablets) are provided with the purpose of demonstrating that
the drug content of each unit in a lot is distributed within a narrow range around
the label claim.
In the last decade there has been an increasing interest as well among phar-
maceutical manufacturers as in regulatory agencies to clarify and standardize
cGMP procedures for demonstrating blend uniformity, see the discussion in
PDA Technical Report No. 25 [22] and the recent review by Berman [2], and
to harmonize requirements on final product testing, see the series of proposed
amendments to the United States Pharmacopeia tests for uniformity of dosage
units, [27], [28].
As such control and monitoring procedures are based upon samples from the
blend, or from the batch of tablets, there is an inherent uncertainty concerning
the actual dispersion in the blend or batch being sampled. Therefore, the as-
surance provided by such procedures is of a statistical nature (i.e. depending
on the pattern of variation in the entity under test), and in order to assess the
influence of the uncertainty due to sampling the properties of the procedures
may therefore be assessed using statistical concepts and techniques under due
consideration to the potential sources of variation in the processes being mon-
108 Paper D
itored.
In industrial and commercial practice, product requirements are often formu-

lated as requirements on quantifiable characteristics of the product. Such re-
quirements are most appropriately formulated as specifications for individual
units of product, but may also include specifications for such batch or process
characteristics as batch fraction nonconforming or standard deviation between
units in the batch. The International Standard ISO 10576 [41] provides general
guidelines on drafting specifications as well for commercial as for governmen-
tal regulatory purposes.
In particular, in ISO 10576 it is advised to separate the issue of designating

specifications (i.e. range of permissible values) for a product, or a process
from the issue of designating acceptance criteria to be used for assessment of
conformity to the specifications. This facilitates the discussion of conformity
assessment procedures in situations involving measurement- or sampling un-
certainty, and allows for declarations of conformity (or nonconformity) that do
not depend on the particular choice of measurement- or sampling method.
Thus, to be in line with this recommendation, requirements to e.g. uniformity

of the blend and of doses should be formulated in terms of blend and batch
characteristics, like mean dose content and standard deviation between dosage
units in the batch, proportion of dosage units in the batch exceeding specified
limits, etc. Although such a distinction between product requirements and ac-
ceptance criteria would be helpful e.g. in clarifying to what extent - if any -
manufacturers that assay a large number of samples are penalized or rewarded,
in regulatory practice it is often seen that requirements are expressed in terms
of acceptance criteria for samples from the process or product, rather than in
terms of product or process characteristics. Thus, the 1984 USP requirements
to content uniformity of tablets [42] was formulated as the following accep-
tance criteria
Stage 1: Assay 10 tablets. Pass if both of the following criteria are met:
1) sample coefficient of variation is less than or equal to 6.0%

2) no value is outside claim ±15%
109
Fail, if one or more values are outside claim ±25%. Otherwise go to

stage 2
Stage 2: Assay 20 further tablets. Pass if, for all 30 tablets, the following
criteria are met:
1) sample coefficient of variation is less than or equal to 7.8 %

2) no more than one value is outside claim ±15%, and no value is
outside claim ±25%.
Otherwise fail
Thus, in essence, the only requirement to product quality is that a (randomly

selected) sample from the batch shall pass this test.
The requirements above have subsequently been subject to various amend-

ments. The currently valid requirement, termed USP 24 [26], has been under
revision since 1997. Although the various proposals tend towards a more para-
metric approach, the requirements are still formulated in terms of criteria to
be applied to a randomly selected sample from the batch, and not as explicit
requirements to content uniformity in the batch.
As considerations regarding sampling uncertainty are not explicit in these re-

quirements, the implicit borderline between production that is considered sat-
isfactory according to these requirements, and production that is not, is deter-
mined by the operating characteristics of these criteria. Therefore, in order
for the manufacturer to design test and validation procedures and to make an
informed choice of the number of samples to be used in such procedures, it is
imperative to have an understanding of the operation of the criteria.
In the paper we shall therefore address some basic statistical issues related to
a crude demonstration of uniformity, i.e. provision of evidence that there is a
satisfactorily narrow dispersion of values in the entity under investigation.
The main issue of the paper is to discuss the influence of sampling uncertainty
in such demonstrations of uniformity, and in particular to discuss the assurance
provided against “unsatisfactorily large dispersions of values” under various
110 Paper D
testprocedures. Although acceptance criteria used in the pharmaceutical indus-

try mostly are formulated in terms of requirements to sample results, we shall
consider the operating characteristics of such procedures in terms of popula-
tion values rather than properties of future samples from the population under
investigation. Formulating the problem in terms of population values allows
for interpretation in terms of the formal statistical framework of hypothesis
testing and confidence intervals, and use of the concepts and techiques from
the statistical theory of hypothesis testing to specify the assurance provided by
the various procedures. In particular, concepts and techniques developed in
the field of acceptance sampling may be used to provide insight in the mecha-
nisms involved when providing assurance under due consideration to sampling
uncertainty.
In the paper we shall discuss some of the acceptance criteria for blend or
dose uniformity that have been suggested in the pharmaceutical literature. In
Section 4 criteria based solely upon a measure of dispersion in the sample
(standard deviation or coefficient of variation) are discussed. Such criteria are
mainly used for blend uniformity analysis where use of statistical measures of
location are not necessarily relevant (e.g. because of bias due to the sampling
device).
The main body of the paper, Sections 5 to 6 discuss acceptance procedures

that include a criterion on a measure of dispersion in the sample as well as
criteria on individual measurement values like the two stage procedure in USP
21 [42]. It is shown that - in terms of population requirements - this pro-
cedure essentially monitors the proportion of population values outside some
limiting values. The statistical problem of monitoring such a population value
has been discussed in the literature on acceptance sampling by variables. In
Section 5 this literature is reviewed and the assurance provided by various ac-
ceptance procedures based solely upon combinations of sample average and
sample standard deviation that have been suggested in the pharmaceutical lit-
erature are discussed. Finally, in Section 6 the operating properties of the USP
draft proposal [28] are discussed in light of these general results.
111
2 Acceptance criteria and statistical hypothesis

testing
2.1 Choice of null hypothesis and alternative hypothesis
Whereas acceptance criteria are formulated as criteria to be applied to the assay

results for a random sample from the blend or batch without reference to any
distributional assumptions, the assessment of the operating properties of such
criteria is usually best performed under specific assumptions on the distribution
of sample values.
As the statistical theories of hypothesis testing provides a convenient formal

statistical framework for design of acceptance criteria and assessment of dis-
criminatory properties, the discussion of the various acceptance criteria will be
related to concepts and results from statistical theories of hypothesis testing.
Thus, when a parametrized model for the sample results is assumed, we shall
formulate a hypothesis (null hypothesis) H0 regarding values of the parameters
of the model (i.e. quality of the blend or batch). Those parameter values that
do not belong to H0 constitute the alternative hypothesis, H1 .
It is an inherent feature of the Neyman-Pearson theory of hypothesis testing

that a test can only offer evidence against the null hypothesis. A small ob-
served significance, or p-value indicates that the alternative has significantly
larger explanatory power. However, a large p-value does not suggest that the
null hypothesis is true, but only that we lack evidence that it is not. Unfor-
tunately, this difficulty is often swept under the carpet, the technically correct
phrase “fail to reject the null hypothesis” being replaced by the term “accept”.
In general, when the goal of an experiment or sampling procedure is to estab-

lish an assertion, it is good statistical practice to take the negation of this asser-
tion as the null hypothesis. The assertion becomes the alternative hypothesis.
Therefore, when methods of statistical hypothesis testing are used for the pur-
poses of conformity testing, i.e. to provide assurance that an entity conforms to
a specified criterion, it is generally recommended to formulate the hypothesis
112 Paper D
H0 as the “nonconforming” parameter values, see also the discussion by Holst

et al. [43]. Following this recommendation, the null hypothesis should taken
to mean “the batch is not satisfactory”, with the alternative being “the batch is
satisfactory”. “Rejection” of the null hypothesis occurs when the sample has
provided strong evidence favouring the assertion that the batch is satisfactory,
and therefore, rejection of the null hypothesis should imply acceptance of the
batch, whereas failure to reject the hypothesis is taken to mean that the sample
has not provided sufficient evidence to warrant acceptance of the batch.
As this equivalence between “rejection of a hypothesis” and “acceptance of a

batch” invariably leads to confusion, we have chosen in the paper when refer-
ring to statistical hypothesis testing to let the null hypothesis correspond to a
region of “satisfactory population values”, and the alternative hypothesis corre-
spond to unsatisfactory values. In this way the terms “accept” and “reject” may
be used with the same meaning when referring to batch (or blend) acceptance
criteria, and when referring to some underlying statistical hypothesis.
In the terminology of hypothesis testing, an acceptance criterion defines a re-

gion of potential sample results that do not satisfy this criterion, the critical
region, such that whenever a sample result is within this region, the hypothe-
sis is said to be rejected. The significance level of the test is the (maximum)
probability of obtaining a sample result in the critical region when the hypoth-
esis H0 is true (i.e. the parameters of the distribution are as specified by the
hypothesis).
As a consequence of this choice of null hypothesis to denote the values of

population parameters that are “satisfactory” in some sense, it will sometimes
be practical to use a critical region corresponding to a formal significance level
of 95%, or 90%, say, instead of the traditional 5% or 10 % significance level.
Because of this duality in the choice of hypothesis and significance level it is

advisable that the assessment of an acceptance criterion is performed by means
of the operating characteristic of the criterion, i.e. the function showing the
probability of passing the test as function of the parameters of the distribution
of sample values (blend or batch quality) as this function shows the discrimi-
natory power of the actual acceptance criterion, and does not otherwise depend
113
on phrasing of statistical hypothesis etc.
2.2 Confidence intervals and statistical tests
In practical assessment of sample results it is good statistical practice to con-

sider not only the numerical estimate obtained from the sample, but also the
associated uncertainty, e.g. in terms of a confidence interval giving a set of
plausible values for the population parameter(s), i.e. a set of values that are in
agreement (do not contradict) the sample result.
Sometimes acceptance criteria are formulated in terms of requirements that a

1 − α confidence interval for the relevant parameter shall be fully included in
some specified region of “satisfactory” qualities. Such an approach might be
more easily understood than a criterion derived using formal statistical tests.
There is, however, no conflict between these two approaches.
Thus, consider the formal statistical hypothesis concerning a population pa-

rameter, θ
H0 : θ > θ0 vs Ha : θ ≤ θ0 (1)
with θ0 denoting some specified limiting value for the population parameter θ,
and assume that some statistic T provides information about θ.
Then the rule: “Reject H0 (and claim θ < θ0 ) whenever the upper (1 − α)
confidence limit for θ is less than the limiting value θ0 ” corresponds to a level
α test of the hypothesis H0 , (or, alternatively to a level (1 − α) test of the
hypothesis H0∗ : θ ≤ θ0 ). For a formal proof of the duality between statistical
confidence regions and tests seee e.g. [44] chapter 4.5.
The interpretation of any given acceptance criterion in terms of the critical

region for some test is, however, more directly suitable for the derivation of the
operating characteristic for that acceptance criterion. In particular it is useful
when assessing the effect on the discriminatory power of using e.g. larger
number of samples.
114 Paper D
3 Notation and distributional assumptions
Following the usual practice in the pharmaceutical literature we shall not con-
sider absolute dose values but assume that measurements are recorded relative
to some (known) target dose weight or label claim.
As the focus of the paper is a discussion of the statistical properties of various

acceptance criteria, we shall follow the practical convention in mathematical
statistical literature and distinguish between random variables (i.e. the un-
known sample result before sampling) denoted by capital letters, D1 , D2 , . . . , Dn ,
and actual values in a sample, d1 , d2 , . . . , dn . Thus, D1 , . . . , Dn refers to a
probability distribution of potential sample results, whereas d1 , . . . , dn shall
be understood as a generic representation of an actual set of sample results.
We shall let D and Sd denote the average relative dose and standard deviation
of relative doses in the sample,
X
n .
D = Di n (2)
i=1
X
n .
Sd2 = (Di − D)2 (n − 1) (3)
i=1
with the corresponding sample results d¯ and sd .
Moreover, as the term relative sample standard deviation may have different
interpretations (as standard deviation of measurements relative to target value,
or as standard deviation of measurements relative to the average sample value),
we shall throughout the paper use the unambiguous term sample coefficient of
variation to mean .
Z = Sd D. (4)
The discussion of the acceptance criteria and the derivation of expressions

for acceptance probabilities is performed under the assumption that individual
sample values, Di may be represented by independent, identically distributed
variables and that the distribution of sample results Di may be described by a
115
normal distribution,
D1 , . . . , Dn mutually independent with Di ∼ N(µd , σd2 ). (5)
Thus, when Di represents content of dosage units (tablets) selected from a

batch, the assumption corresponds to assuming that the overall distribution of
dose content in the batch may be represented by a normal distribution, and that
individual dosage units are selected at random from the dosage units in the
batch. When Di represents content of blend samples, the assumption analo-
gously corresponds to assuming that the overall distribution of such potential
samples from the blend may be represented by a normal distribution, and that
samples are taken at randomly selected positions in the blend.
To preserve the generality we shall use the term “population” when referring
to this distribution of potential values representing dose content measurements
from all doses in the tablet batch, or to the distribution of potential values
representing conceivable blend sample measurements.
It should be noted that the model (5) will not be adequate when the distribution
in the population is bimodal or multimodal corresponding e.g. to stratifica-
tion between locations (e.g. top, middle and bottom), when the distribution is
skewed, e.g. as a result of deblending, or when the distribution has heavier tails
than the normal distribution, e.g. as a result of imperfect mixing (clustering)
or of using dose particles that are too large for the intended dosage.
When sampling is performed from the “worst case positions in the blender”,
as suggested e.g. in the FDA guidelines for blend analysis [20], the population
refers only to sample results that might have been obtained under hypothetical
repeated sampling from these positions, and the model does not necessarily
reflect the population of values representing the totality of the blend.
When sampling is performed under a hierarchical (or nested) scheme as sug-

gested among other by the Product Quality Research Institute, PQRI [18]
where e.g. n = 36 sample values are obtained by repeatedly sampling from
only 12 locations, the model will only be adequate in such (rather unlikely)
situations where there is no correlation between subsamples from the same
location in the blend or batch.
116 Paper D
However, even despite these restrictions, the mathematical analytical discus-

sion serves a purpose of clarifying and illustrating the statistical issues in-
volved, thereby providing further insight in the properties of various tests that
have been proposed in the pharmaceutical literature.
4 Acceptance criteria for the dispersion of doses
In pharmaceutical production, the blending, or mixing process is generally

considered a key process aiming at producing a uniform, or homogenous blend.
The purpose of blend sample analysis is to ensure that the blending is adequate
for the end use.
As blend uniformity analysis might be subject to bias (i.e. systematic deviation

of the mean value in the distribution of sample values) caused by the sampling,
or by the analytical procedure, blend uniformity analysis puts main emphasis
on assessment of the dispersion of population values.

ple standard deviation
A direct estimate of the dispersion of dose-content is the standard deviation of

the dose-content for the units in the sample. This estimate is an obvious quan-
tity to use in making a judgment about the uniformity of a product. Further,
the standard deviation is not affected by a constant bias of the mean due to
sampling, handling (storage) or analytical disturbancies.
Therefore, it might be considered to specify a limiting value, σlim for the sam-
ple standard deviation, and use the following acceptance criterion
Pass when the sample standard deviation, sd ≤ σlim (6)

otherwise fail.
In the framework of statistical hypothesis testing, this acceptance criterion may

117
be interpreted as defining a critical region for a test of the hypothesis
H0 : σd ≤ σlim against the alternative H1 : σd > σlim . (7)
The significance level of this test (6) of the hypothesis (7) is given as the prob-
ability of rejection in the borderline case when σd = σlim . In this borderline
case there is approximately a 50% risk of obtaining a sample standard devia-
tion exceeding σlim , and hence (referring to the hypothesis (7)) the significance
level is approximately 50% regardless of the sample size.
A more detailed assessment of the operating characteristic for the acceptance

criterion may be performed under specific assumptions for the distribution of
sample values, Di . Assuming a normal distribution of dosage units, (5), it
follows that the sample variance Sd2 is distributed like a σd2 χ2n−1 /(n − 1) dis-
tributed variable, and therefore the probability of acceptance may be deter-
mined from the χ2 -distribution as
2
2
Pacc (σd ) = P[Sd ≤ σlim ] = P χn−1 ≤ (n − 1) σlim /σd (8)
i.e. from the cumulative distribution function of the χ2n−1 distribution. In the
SAS-system software the acceptance probability may be found by calling the
function PROBCHI(x,df) with the arguments x = (n − 1)(σlim /σd )2 and
df = n − 1 with σd denoting the (true) standard deviation between doses. In
Microsoft Excel, the probability may be found from the function CHIDIST(x,df)
providing a value which is one minus the probability of acceptance.
Figure 1 shows graphs of the probability of acceptance as function of the batch

standard deviation of relative doses, σd , for σlim = 0.06 and various values of
the sample size, n, between n = 10 and n = 40. It is seen that the set of curves
tend to intersect at the point σd = 0.06, Pacc = 0.5 (the significance level),
and that the discriminatory power (the steepness of the OC-curve) increases
with increasing sample size.
When a sample size n = 10 is used, it is necessary that the batch standard

deviation is less than 0.035 (producer’s risk quality) in order to assure a high
probability of acceptance, and the sampling plan does only provide a consumer
protection (10% consumer risk) against a batch standard deviation of 0.085
118 Paper D
(consumer’s risk quality). As the limiting value of the sample standard devi-
ation does not depend on the sample size, the consumer’s and producer’s risk
qualities move towards this limiting value with increasing sample sizes, i.e.
the (fixed) limiting value of the sample standarddeviation, σd = 0.06 is also
the borderline between batch standard deviations that are accepted under this
criterion and batch standard deviations that are not accepted.
Moreover, it is a consequence of this choice of a limiting value for the sam-

ple standard deviation, independent of the size of the sample, that acceptance
under this scheme provides a confidence of approximately 50% that the stan-
dard deviation in the blend (or batch) does not exceed the limiting value,
σlim = 0.06, independent of the sample size (increasing from 45% for n = 10
to 48% for n = 100). However, when confidence statements corresponding to
greater confidence are formulated, the assurance provided depends on the sam-
ple size. Thus, in the example considered, acceptance by a sampling plan with
n = 10 provides a 90% confidence that the standard deviation in the blend
does not exceed 0.09, whereas acceptance by a sampling plan with n = 40
provides a 90% confidence that the standard deviation in the blend does not
exceed 0.07.

ple coefficient of variation
In pharmaceutical applications, it is customary practice to use the sample co-

efficient of variation, (4) as a measure of dispersion.
This measure is used, eg. as part of the acceptance criterion in the test for
uniformity of dosage units in the currently valid version of the United States
Pharmacopeia, USP 24, [26].
It is interesting to note that although dosage values are already measured rela-
tive to the target, or label claim, and hence, the standard deviation of relative
dosage values already measures dispersion in relative units, it has been con-
sidered appropriate, to make a further adjustment (for possible sampling or
analysis bias ) by considering the standard deviation relative to the estimate, D
119
Fig. 1: The probability of accepting a batch as function of the relative standard

deviation, σd for various sample sizes. Accept when sample standard deviation
of relative doses is ≤ 0.06.
120 Paper D
of actual content.
Consider an acceptance criterion of the form
Pass when the sample coefficient of variation, sd /d¯ ≤ Clim (9)

otherwise fail.
The acceptance criterion may be interpreted as defining a critical region for a

test of the hypothesis
H0 : σd /µd ≤ Clim against the alternative H1 : σd /µd > Clim . (10)
The operating characteristic for this acceptance criterion may be assessed un-
der the assumption of a normal distribution of dosage units, (5) utilizing that
the distribution of √
D n
T = √ = (11)
Sd / n Z
is a non-central t-distribution with non-centrality parameter
√
nµd
δ=
σd
and f = n − 1 degrees of freedom (see e.g. Johnson and Kotz [45]) and hence
for normally distributed measurements the distribution of the sample coeffi-
cient of variation depends only on the sample size and the batch coefficient of
variation, Cd = σd /µd .
Disregarding negative values of Z, it is seen that

√
n
P[Z ≤ Clim ] = P T ≥ (12)
Clim
and therefore the probability of acceptance may be found from the cumulative
probability distribution of the noncentral t-distribution as
√ √
Pacc (Cd ) = P[Z ≤ Clim ] = 1 − P tn−1 ( n Cd ) ≤ n Clim (13)
√
with tn−1 ( n/Cd ) denoting a random variable distributed according to a non-
central t-distribution with n−1 degrees of freedom and noncentrality parameter
√
n/Cd .
121
In the SAS-system software the acceptance probability may be found by call-

√
ing the function PROBT(x,df,nc) with the arguments x = n/Clim , df =
√
n − 1 and nc = n/Cd . This gives the probability of non-compliance; the
probability of acceptance is then found as one minus the result provided by the
function call.
Figure 2 shows graphs of the probability of acceptance as function of the batch

coefficient of variation of relative doses, Cd = σd /µd for Clim = 0.06 and
various values of the sample size, n, between n = 10 and n = 40.
In analogy with the test of the sample standard deviation, Figure 1, the set
of curves tend to intersect at the point Cd = 0.06, Pacc = 0.5 indicating that
acceptance under this scheme provides a confidence of approximately 50% that
the coefficient of variation in the blend (or batch) does not exceed the limiting
value, Clim = 0.06, independent of the sample size.
The difference between the acceptance probability corresponding to the cri-

terion (7) on the sample standard deviation (relative to label claim) shown in
Figure 1, and the acceptance probability corresponding to the criterion (10) on
the sample coefficient of variation shown in Figure 2 is only marginal as long
as the batch mean, µd does not deviate too much from the target value.
However, because of the sampling variation of the average dose, D, that is in-
troduced in the criterion (10) on the sample coefficient of variation, the curves
corresponding to this criterion (Figure 2) are not quite as steep as their counter-
parts for the criterion on the sample standard deviation (Figure 1). Thus, when
the coefficient of variation of dosage units in the batch is less than the limiting
value, Clim = 0.06, the acceptance probability under the sample coefficient of
variation criterion is slightly less than under the sample relative standard devi-
ation criterion, whereas it is slightly larger when the coefficient of variation in
the batch is larger than 0.06.
122 Paper D
Fig. 2: The probability of accepting a batch as function of the coefficient of

variation, Cd = σd /µd of doses in the batch for various sample sizes. Accept
when sample coefficient of variation is ≤ 0.06.
123
4.3 Criterion based upon prediction of standard deviation

of future samples
When the acceptance criterion specifies a limiting value for some sample statis-
tic, independent of the size of the sample, this limiting value implicitly defines
a border value between population values that will pass under applicaton of
this criterion, and those that will not. For large samples, blends or batches
with a population value that does not exceed this limiting value will pass under
this criterion. For smaller samples the sampling uncertainty implies a lesser
discriminatory power, and the amount of sample evidence favouring the hy-
pothesis is given the same weight as the evidence against the hypothesis (the
statistical test is performed at a 50% significance level).
This feature is not very satisfactory when the limiting value used in the ac-
ceptance criterion has been chosen to represent some undesirable quality, and
when the purpose of the test is to provide some assurance against accepting
batches of such qualities, and therefore the sample standard deviation should
in some way be reduced by its uncertainty before comparison with the limiting
value.
Such an acceptance procedure that puts more focus upon the assurance pro-
vided by the procedure has been suggested in PDA Technical Report 25 [22].
The fundamental idea in this procedure is that the sampling uncertainty asso-
ciated with the sample under investigation shall be taken into account when
deciding whether or not to accept the blend or batch. The blend is passed only
if the sample provides strong evidence (even when accounting for the sampling
uncertainty) that the quality is better than some prescribed undesirable quality.
The statistical technical tool invoked in the design of this procedure is the so-
called Standard Deviation Prediction Interval to be explained below.
The starting point in the prediction interval approach as developed in [22] is a

requirement (by authorities) that the standard deviation in a sample from the
blend or batch shall not exceed a specified value, sspec. Taking this to be the
“specification” of satisfactory batch quality, the acceptance rule is then devised
124 Paper D
to provide a high assurance that batches that are accepted by the sampling plan
really are of a satisfactory quality. In [22] “satisfactory quality” is taken to
mean that the standard deviation in a sample (of size 10) from the batch does
not exceed sspec = 0.06, and the “assurance” is chosen to be 90%.
It is worth noting that the “specification” of batch quality is not explicitly stated
in terms of parameters like σd characterising the batch, but rather in terms of
properties of a future sample from the batch.
Standard Deviation Prediction Interval
The standard deviation prediction interval, as described e.g. by Hahn and

Meeker [46] is derived from the joint distributional properties of two sets of
samples from the same population of normally distributed values, like two sets
of samples, D1 , D2 , . . . , Dn (the current, actual sample) and D10 , D20 , . . . , Dm
0
(a future sample).
Let
X
n .
Sd2 = (Di − D)2 (n − 1) (14)
i=1
denote the sample variance found in the current sample (of size n), and let
moreover
X
m .
0
Sp2 = (Di0 − D )2 (m − 1) (15)
i=1
denote the sample variance in a future sample of size m. Then, under the dis-
tributional model (5) assuming independence between the two sets of samples
and assuming the same underlying population variance, σd2 , for both samples,
it follows that the distribution of
Sp2
F = (16)
Sd2
125
is a F-distribution with (m − 1, n − 1) degrees of freedom, and hence

" #
Sp2
P ≤ F(m − 1, n − 1)1−α = 1 − α (17)
Sd2
or, equivalently

P Sp2 ≤ Sd2 F(m − 1, n − 1)1−α = 1 − α (18)
with F(m − 1, n − 1)1−α denoting the 1 − α quantile in the F(m − 1, n − 1)

distribution. The distribution does not depend on the actual population value
σd2 neither does it depend on the mean content in the population, µd (or a
potentially different mean µd0 for the future sampling process).
The reader is cautioned that in some (mainly US) textbooks and software
quantiles and percentiles of non-normal statistical distributions are sometimes
given as the so-called critical values indexed by the probability associated with
more extreme values in the distribution. To avoid the ambiguity when using
this practice in situations where one-sided as well as two-sided tails may be
appropriate, we have chosen the unambiguous approach to let symbols like
F(m − 1, n − 1)1−α , χ2f,1−α , tf,1−α , etc. refer to quantiles or percentiles of the
distributions under consideration.
In the SAS-system software, the quantile may be found by calling the function
FINV(p,ndf,ddf) with the arguments p = 1 − α, ndf = m − 1, ddf = n −
1, and in Microsoft Excel, by calling the function FINV(p,df1,df2) with the
arguments p = α, df 1 = m − 1, df 2 = n − 1 (as the Microsoft Excel function
returns the critical value).
q
Thus, when a value, sd = s2d , of the sample standard deviation is found in
the sample, the upper 1 − α prediction bound for the sample standard deviation
in a future sample from that batch may be found as
p
spred (sd ) = sd F(m − 1, n − 1)1−α . (19)
It follows from the formal calculations that as well the sampling uncertainty
associated with the standard deviation in the actual sample, Sd , as the uncer-
tainty associated with the standard deviation in the future sample, Sp , has been
taken into account in the derivation of this upper bound.
126 Paper D
Acceptance criterion
Now, assume that a limiting value, sspec for the sample standard deviation in a
future sample of size m has been specified.
Following the approach in [22] the following acceptance criterion may be con-
sidered.
Pass when the standard deviation prediction bound, spred(sd ) ≤ sspec

(20)
otherwise fail.
Utilizing the relation (19), and observing that F(m − 1, n − 1)1−α = 1/F(n −
1, m−1)α the acceptance criterion may be phrased in terms of a limiting value,
slim of the sample standard deviation, sd in the current sample,
p
slim (n) = sspec F(n − 1, m − 1)α . (21)
In terms of this limiting value, the acceptance criterion is
Pass when the sample standard deviation, sd ≤ slim (n)

(22)
otherwise fail.
Clearly, the limiting value of the sample standard deviation depends on the
sample size, n.
It may be shown that when the sample size, n, increases, the limiting value of
the sample standard deviation, slim (n) → σlim with
q
σlim = sspec (m − 1) χ2m−1,1−α (23)
with χ2m−1,1−α denoting the 1 − α quantile in the χ2 distribution with m − 1

degrees of freedom. Thus, for α < 0.50 the limiting value, slim (n) of the
sample standard deviation will be less than the specified value, sspec , for future
samples.
127
n
10 15 20 25 30 35 40 ∞
3.841 4.119 4.260 4.346 4.440 4.444 4.476 4.697
Table 1: Limiting value, slim (n) of the sample standard deviation (in percent
of target value) ensuring that the 90% prediction bound for a future sample of
size 10 does not exceed 0.06
Table 1 shows values of the limiting value slim (n) for various sample sizes,
n = 10, 15, . . . , 40 for sspec = 0.06 and α = 0.10.
Under the distributional assumptions (5), the distribution of Sd , and, hence also
of spred (Sd ) depends on the population standard deviation, σd , and therefore
the acceptance criterion may be interpreted as defining a critical region for a
test of the hypothesis
H0 : σd ≤ σlim against the alternative H1 : σd > σlim (24)
with σlim given by (23). This limiting value of the population standard devia-
tion depends on the number, m, of prospective sample items, and the desired
confidence, 1 − α, that the standard deviation in this prospective sample shall
be less than the design value, sspec.
Operating characteristic
The operating characteristic for this acceptance criterion is given as

h p i
Pacc (σd ) = P[Sd ≤ slim (n)] = P Sd ≤ sspec F(n − 1, m − 1)α . (25)
Under the assumption of a normal distribution of measurement values, (5),

values of the operating characteristic may be calculated utilizing that Sd2 is
distributed like a σd2 χ2n−1 /(n − 1) distributed variable, and therefore the prob-
ability of acceptance may be determined from the χ2 -distribution as
2
2
Pacc (σd ) = P χn−1 ≤ (n − 1) slim /σd (26)
128 Paper D
with slim given by (21), i.e.

2
2
Pacc (σd ) = P χn−1 ≤ (n − 1) sspec/σd F(n − 1, m − 1)α . (27)
In the SAS-system software the acceptance probability may be found by call-

ing the function PROBCHI(x,df) with the arguments x = (n−1)(sspec /σd )2 F(n−
1, m − 1)α and df = n − 1, and in Microsoft Excel, invoking the function
CHIDIST(x,df) with the arguments x = (n − 1)(sspec /σd )2 F(n − 1, m − 1)α
and df = n − 1 returns a value which equals one minus the probability of
acceptance.
Figure 3 shows graphs of the probability of acceptance as function of the pop-

ulation standard deviation of relative doses, σd , for various sizes, n, of the
sample used for testing for the test with sspec = 0.06, m = 10 and α = 0.10,
i.e. an assurance of 90% that the standard deviation in a future sample of size
10 does not exceed 0.06.
Comparing these graphs with the acceptance probability corresponding to a

crude comparison of the actual sample standard deviation with the limiting
value σlim = 0.06 shown in Figure 1, it is seen that the steepness of the curves
remains unchanged, but the set of curves has been shifted to the left. Thus,
the borderline between those batch standard deviations that will be accepted
according to the SDPI-criterion, and those that are not accepted.
Under the standard deviation prediction interval acceptance criterion the prob-
ability of acceptance when σd = 0.06 is between 7% (for n = 10) and 1% (for
n = 40).
When σd = σlim = 0.047 the probability of acceptance is around 20%. Thus,

for the sample sizes under consideration, the significance level of the statistical
test of the hypothesis H0 formulated in (24) is 80%. In other words, the ac-
ceptance rule provides an assurance of 80% that the standard deviation in the
population does not exceed σlim = 0.047.
In general, the significance level of the test of the hypothesis (24) performed
using an acceptance criterion of the form (20) is found as 1 − Pacc (σlim ) with
129
Pacc (·) given by (27), i.e.

∗ n−1 2
α =1−P χ2n−1 ≤ χ F(n − 1, m − 1)α (28)
m − 1 m−1,1−α
The significance level depends on the sample size. For increasing sample size,
n, the significance level α∗ → 0.5.
The borderline between those population standard deviations that will lead
to acceptance according to this criterion, and those that are not accepted is
σd = 0.047. Thus, for a sample size n = 10, it is necessary that the batch
standard deviation is less than 0.03 (producer’s risk quality) in order to as-
sure a high probability of acceptance, but the sampling plan provides a con-
sumer protection (10% consumer risk) against a batch standard deviation of
0.06 (consumer’s risk quality). Using a larger sample, e.g. n = 35 allows the
manufacturer to operate with a population standard deviation, σd up to 0.04
and still retain a high probability of acceptance, and, moreover, for this sample
size, a batch with a standard deviation exceeding σd = 0.05, say, will only be
accepted with a probability less than 15%.
4.4 A direct approach in terms of population values
Realizing that the approach based upon the standard deviation prediction inter-
val is basically a statistical test of a hypothesis (24) concerning the population
standard deviation σd , a more direct approach using concepts from statistical
theories of hypothesis testing, one might choose a more direct approach and
use the standard test for the hypothesis
H0∗ : σd ≤ σlim vs Ha∗ : σd > σlim (29)
with some specified limiting value, σlim , e.g. given by (23). The test may be
found in introductory statistical textbooks (e.g. [47], sec. 10.13), and is also
provided in the ISO Standard ISO 2854 [48] with the corresponding power
curves given in ISO 3494 [49].
The acceptance criterion corresponding to a test of the null hypothesis (29) at

130 Paper D

deviation, σd for various sample sizes. Accept when standard deviation 90%
prediction bound for a sample of size m = 10 is ≤ 0.06.
131
significance level α∗ is
Accept H0∗ when sd ≤ slim (n)

otherwise reject
with the limiting value, slim (n), for the sample standard deviation given by
q
slim (n) = σlim χ2n−1,1−α∗ /(n − 1) (30)
and with χ2f,p denoting the p-quantile in the χ2f -distribution.
As the purpose of the procedure is to provide assurance (even under con-

sideration of the sampling uncertainty) that the value of σd is adequate, i.e.
σd ≤ σlim , the significance level of this test should be chosen rather high, e.g.
a significance level, α∗ = 0.8
It follows from the discussion in Section 2.2 that the acceptance rule of this
test is equivalent with an acceptance rule of the form “accept when the upper
α∗ confidence limit for σd (as calculated from sd ) does not exceed σlim . Thus,
whenever a sample passes this test, there is a confidence of α∗ (e.g. α∗ = 0.8
corresponding to 80% confidence) that the population standard deviation does
not exceed σlim .
As the acceptance rule is intended to provide evidence that σd ≤ σlim , one

might have chosen (in line with good statistical practice, and the provisions in
ISO 1576-1, [41]) instead to formulate a null hypothesis of the form
H0 : σd > σlim vs Ha : σd ≤ σlim (31)
with the interpretation that the content uniformity test is passed whenever the
sample leads to rejection of this hypothesis (31) at a significance level, α =
0.20, say.
The formulation in terms of a statistical test of this null hypothesis (31) might
be more easily understood, as rejection of the hypothesis (31) takes place
whenever the sample result provides strong evidence (at the magnitude 1 − α)
against the hypothesis σd > σlim in favour of the alternative σd ≤ σlim , and,
hence failure to reject the hypothesis σd > σlim means that there is not suffi-
cient evidence (considering the sample uncertainty) to claim that σd ≤ σlim .
132 Paper D

deviation, σd for various sample sizes. Test the hypothesis σd ≤ 0.0470 at
significance level 0.80.
133
However, in order to maintain equivalence between the phrasing “passing the

test” and “accepting the hypothesis” the statistical hypothesis to be tested has
been formulated as in (29) in accordance with the discussion in Section 2.1.
5 Acceptance criteria with limits on individual mea-

surements
5.1 The USP 21 criteria
Traditionally, tests for content uniformity of final product (dosage units) have
not only been concerned with the dispersion of dosage values, but require-
ments to dispersion (e.g. to the sample coefficient of variation) have been
supplemented by requirements on individual measurements.
Thus, in USP 21 [42] on each stage the criterion on the sample coefficient of
variation is combined with criteria on individual measurements.
The requirements for passing this test for content uniformity are
Stage 1: Assay 10 tablets. Pass if both of the following criteria are met:
1) sample coefficient of variation is less than or equal to 6.0%

2) no value is outside claim ±15%.
Fail, if one or more values are outside claim ±25%. Otherwise go to

stage 2
Stage 2: Assay 20 further tablets. Pass if, for all 30 tablets, the following
criteria are met:
1) sample coefficient of variation is less than or equal to 7.8 %

2) no more than one value is outside claim ±15%, and no value is
outside claim ±25%.
Otherwise fail.
134 Paper D
From a statistical point of view two different issues are involved in these re-
quirements
a) In each stage, the criteria include as well a limit on the sample coeffi-
cient of variation (inspection by variables) as limits on individual mea-
surements (inspection by attributes).
b) The plan is a two-stage plan with a decision on whether to invoke the

second stage depends on the result in the first stage, and the criteria to
be applied in the second stage depends on the results in the combined
sample (30 tablets).
The formulation of an acceptance sampling plan as a two-stage plan is a well-

known principle in statistical theories of acceptance sampling. Most often a
suitably designed two-stage plan allowing for acceptance as well as rejection
in the first stage is developed as an alternative to an existing one-stage plan
in order to provide the possibility of reaching a decision with less inspection
effort, but without loosing discriminatory power when compared with the one-
stage plan. Use of a two-stage plan also has the psychological advantage that
batches with borderline sample results in the first sample are given a second
chance. For a discussion of two-stage sampling plans see Schilling, [12]. How-
ever, as the criteria invoked at the second stage involve the sample results from
the first stage, the assessment of the properties of two-stage plans by analytical
methods is not as straightforward as the assessment of one-stage plans. There-
fore, in order to obtain insight in the statistical properties of the procedures, the
assurance provided, and the interplay between the limits on the sample coeffi-
cient of variation (inspection by variables) and the limits on individual values
(inspection by attributes), in the following we shall initially consider only one-
stage procedures. In Section 6.2 we shall return to a discussion of the interplay
between the two stages.
As the literature on assessment of the operating characterics of acceptance pro-

cedures for dosage/blend uniformity mainly assumes a normal distribution of
dosage units, the theoretical discussion will assume that the distribution of
dosage values may be described by a normal distribution, and that sample
measurements may be considered to be independent, identically distributed
135
according to this normal distribution (5).
Operating characteristic of the USP-21 test
Assuming a normal distribution of relative dosage units, di ∼ N(µ, σd2 ), and

independent samples the probability of passing the USP-21 test is a function,
Pusp (µd , σd ), of the mean and standard deviation (µd , σd ) of the dosage units
in the population, and in principle, this function may be expressed by an ana-
lytical expression involving integration over a region in a 30-dimensional space
of sample results.
It is well-known from statistical theory (see e.g. [44]) that under the assump-
tion of independent, normally distributed sample values the sample average
D and sample standard variance Sd2 are jointly sufficient for the population
mean and variance, (µd , σd2 ), and therefore, when sample average and sample
standard deviation (D, Sd ) are known, the distribution of individual measure-
ment values (D1 , . . . , Dn ) does only depend on these sample statistics, and
not upon population mean and variance (µd , σd2 ). Thus, it is formally pos-
sible to device a test based solely upon sample average and sample standard
deviation in the two stages (otherwise disregarding individual measurements)
having the same probability of passing as the USP-21 test for all values of
population mean and standard deviation. However, as the level curves of the
probability, Pusp (µd , σd ) of passing the USP-test do not allow for a simple an-
alytical representation, the actual mathematical derivation of such a test is not
straightforward, and therefore more insight is gained by considering various
approximations to Pusp (µd , σd ).
By disregarding the correlation between the sample coefficient of variation and

the individual measurements, Bergum [50] developed a simple expression giv-
ing a lower bound on the probability, Pusp (µd , σd ), of passing the USP-test:
For this lower bound level curves in the (µd , σd )-plane giving combinations,
(µd , σd ), that have the same lower bound for this probability were provided.
Bergum, op.cit. moreover compared the calculated lower bounds with simu-
lated values of Pusp (µd , σd ) and found a reasonable agreement.
136 Paper D
5.2 Three-class attributes and a parametric approach
Operating characteristic of a three-class attribute plan
In the following we shall use a more direct analytical approach by (initially)

disregarding the requirement on the sample coefficient of variation in the USP-
21 criteria, and consider the following modification of the requirements for
passing the USP-test obtained by disregarding the requirements on the sample
coefficient of variation:
Stage 1: Assay 10 tablets. Pass if no value is outside claim ±15% (LC ± ∆)1 .
continue to stage 2 if no more than one value is outside claim ±15%
(LC ± ∆), and no value is outside claim ±25% (LC ± ∆1 ).
Stage 2: Assay 20 further tablets. Pass if, for all these 20 tablets, no value is
outside claim ±15% (LC ± ∆).
Otherwise fail.
The two sets of limiting values, LC ± ∆ and LC ± ∆1 serve to specify a

classification of each individual measurement values into one of three mutually
exclusive classes defined by the following three zones around the target value,
LC, viz.
Green zone: dose values in the interval LC ± ∆.
Amber zone: dose values in either of the intervals LC − ∆1 < D < LC − ∆,

or LC + ∆ < D < LC + ∆1 .
Red zone: dose values in either of the intervals D < LC − ∆1 , or LC + ∆1 <

D.
1
In this paper calculations and discussions are based on measurements relatively to label
claim, LC. Therefore in the formulae in the rest of the article, LC is equal to 1 (100%)
137
To aid the intuition the zones have been labelled with the colours on traffic
light signals.
In terms of the three zones introduced above the requirements are:
Stage 1: Assay 10 tablets. Pass if all values are in the “green zone”,
continue to stage 2 if no more than one value is in the “amber zone” and
no values in the “red zone”.
Stage 2: Assay 20 further tablets. Pass if, for all these 20 tablets, all values
are in the “green zone”.
Otherwise fail.
For single sampling plans (i.e. acceptance sampling plans with only one stage
of sampling), sampling plans involving such attribute criteria with three classes
are often referred to as a “three-class” attribute sampling plans. Such sampling
plans were introduced by Bray et al [51] inspired by applications of lot accep-
tance sampling in food microbiology, and subsequently recommended by the
International Commison on Microbiological Specifications for Foods (ICMSF)
of the International Union of Microbiological Societies [52]. In the terminol-
ogy of three-class sampling plans, the “green” zone is often labelled “good”,
or “in compliance with upper limit of GMP”, the “amber” labelled “marginal”
and the “red” labelled “bad”. Three class attribute sampling plans (to be used
for one-stage sampling) usually specify sample size, maximum number, cm ,
of non-good (i.e. non-green) items in the sample and maximum number, cM ,
of bad (i.e. red) items in the sample. In microbiological applications cM is
often taken to be zero.
For normally distributed measurement values, Di , the probability of an indi-

vidual measurement value falling in the “red”, and the probability of a value in
either the “amber” or the “red” zone, respectively, are
pr (µd , σd ) = pnonc (µd , σd ; ∆1 ) (32)
pa (µd , σd ) = pnonc (µd , σd ; ∆) (33)
138 Paper D
with the generic function pnonc (µ, σ; ∆) given by

LC − ∆ − µ LC + ∆ − µ
pnonc (µ, σ; ∆) = Φ +1−Φ (34)
σ σ
Hence, for normally distributed measurement values, the probability of a batch

passing under this three-class criterion is
P3−c (µd , σd ) = (1 − pa )10 + 10(pa − pr ) × (1 − pa )29

= (1 − pa )10 × 1 + 10(pa − pr )(1 − pa )19 (35)
with pa denoting the probability of a measurement value outside the interval

LC ± ∆ (e.g. 85 to 115% of label claim) given by (33), and pr given by (32)
denoting the probability of a measurement value outside the interval LC ± ∆1
(e.g. 75 to 125% of label claim) such that pa − pr gives the probability of a
measurement value outside LC ± ∆ ( 85 and 115%), but between LC ± ∆1 (75
to 125%). Clearly, this latter probability does not exceed pa .
Disregarding the contribution from pr in (35), one obtains an approximation to

the probability of passing the test as

P3−c,appr (µd , σd ) = (1 − pa )10 × 1 + 10pa (1 − pa )19 . (36)
This probability is slightly larger than the value given by (35). For any value
of pa (µd , σd ), the largest value of pr (µd , σd ) occurs when µd = LC and
σd = ∆/z1−pa /2 . For values of ∆ and ∆1 used in pharmaceutical practice,
the approximation error incurred when substituting (35) by (36), i.e. by disre-
garding pr , is only marginal.
Figure 5 shows the maximal value of the difference, P3−c,appr (µd , σd )−P3−c (µd , σd )
as function of the the proportion, pa = pnonc (µd , σd ; ∆) for ∆ = 15% and
∆1 = 25%. The largest potential difference occurs for pa = 0.09 (i.e. 9% of
the population values falling in the amber or the red zone) with a maximum ap-
proximation error, P3−c,appr (µd , σd ) − P3−c (µd , σd ) = 0.003, i.e. a difference
in acceptance probability of 0.3 %.
139
maximal approximation error when disregarding ’red class’ criterion

0.0030
0.0025
0.0020
max approx error
0.0015
0.0010
0.0005
0.0
0.0 0.1 0.2 0.3 0.4

pamber
Fig. 5: Maximal absolute error in acceptance probability committed when dis-

regarding the “red” class in calculation of acceptance probability versus total
proportion of “amber” and “red” class units in the population. ∆ = 0.15 and
∆1 = 0.25.
140 Paper D
P p0 (P )
0.99 0.0070
0.95 0.0165
0.90 0.0244
0.50 0.0800
Table 2: Relation between acceptance probability, P , under the three class

sampling plan, and limiting value, p0 (P ) of the population proportion of values
outside LC ± ∆.
(Dahms and Hildebrandt [53] have discussed the choice of limits ∆ and ∆1 for
applications in microbiological quality control under single sampling).
Population requirements derived from requirements to three-class

acceptance probability
The approximative probability of passing the three-class attribute test, P3−c,appr

is a decreasing function of pa , and therefore a requirement
P3−c,appr (µd , σd ) ≥ Pspec (37)
is equivalent with a requirement
pa (µd , σd ) ≤ p0 (Pspec ) (38)
with the limiting value, p0 (P ), of the population proportion nonconforming

defined as the solution, p0 , to

(1 − p0 )10 1 + 10p0 (1 − p0 )19 = P. (39)
Table 2 shows the limiting value, p0 (P ), of the population proportion non-

conforming corresponding to different values of the acceptance probability, P
under the three class attribute criterion.
Hence, under the assumption of a normal distribution, the requirement of a

specified probability of passing under this three-class test is equivalent to a
141
requirement that the proportion, pa (µd , σd ), of non-green values in the popu-

lation does not exceed some limiting value that depends on the specified prob-
ability of passing. Thus, the apparent three class criterion (green, amber, red)
is essentially just a two class criterion, (green, non-green).
In particular, a requirement of a 95% probability of passing under the three-

class modification of the USP test is equivalent to requiring that no more than
1.65% of the units in the batch are outside the limits LC ± ∆ ( 85 and 115%).
Interpretation of a two-class requirement in terms of combinations

of population mean and variance
It is well-known, see e.g. the textbook by Schilling [12] that for a specified
fraction, p0 , of nonconforming product the set of solutions (µd , σd ) to
pnonc (µd , σd ; ∆) = p0 (40)
are found on a curve in the (µd , σd )-plane.

Thus, under the assumption of a normal distribution, control of the proportion
of dosage units with content exceeding specified limiting values ( LC ± ∆) is
equivalent to controlling the combination of batch mean and variance.
Wallis [54] has provided a parametric description of the curve using an auxil-
iary variable, π, (0 < π < p0 ) as

µd = LC + ∆ × (zp0 −π − zπ ) (zp0 −π + zπ ) (41)

σd = (µd − LC − ∆) zπ (42)
with zp denoting the p-quantile in the standardized normal distribution. The

idea is attributed by Wallis to Kenneth J. Arnold.
For π = 0 the point on the curve is µd = LC − ∆, σd = 0; for π = p0 the

point is µd = LC + ∆, σd = 0. When π = p0 /2, the parametric form gives
µd = LC, σd = ∆/z1−p0 /2 indicating that the maximum standard deviation in
the population allowing for at most the fraction p0 of product outside the limits
LC ± ∆ is σd = ∆/z1−p0 /2 which is only admissible when the population
mean is on target (LC).
142 Paper D
Figure 6 shows examples of the curve (41), (42) for ∆ = 16.5% and for p0 =
0.0165 and p0 = 0.0278, respectively.
The curve is bounded by a trapezoid with corners in the points
A: µd = LC − ∆ ; σd =0
B: µd = LC − ∆(1 − z1−p0 /z1−p0 /2 ) ; σd = ∆/z1−p0 /2
(43)
C: µd = LC + ∆(1 − z1−p0 /z1−p0 /2 ) ; σd = ∆/z1−p0 /2
D: µd = LC + ∆ ; σd = 0.
The length of the baseline is 2∆, and the height of the trapezoid is ∆/z1−p0 /2 .
The slope of the left-hand side of the trapezoid is 1/z1−p0 . For p0 > 0.5
the oblique sides of the trapezoid are sloping outwards, and the upper side of
the trapezoid will be longer than the baseline. For p0 = 0.5 the trapezoid
degenerates to a rectangle, and for p0 < 0.5 the oblique sides of the trapezoid
are sloping inwards. The bounding trapezoids have also been shown in Figure
6.
Summary of properties of the three class attribute plan
Disregarding the requirements to the sample coefficient of variation it was

found that under a normal distribution of population values does the operat-
ing characteristic of the three class attribute criteria invoked at each stage in
the USP 21 acceptance procedure only depend on the population mean and
standard deviation through the proportion, pa (µd , σd ) = pnonc (µd , σd ; ∆), of
units in the population with values outside the “green” zone (LC ± ∆).
Hence, assurance of a specified probability, Pspec , of acceptance under this

three class attribute criterion is equivalent to assuring that the proportion, pnonc (µd , σd ; ∆)
of units in the population with values outside the interval LC ± ∆ does not ex-
ceed some value, p0 (Pspec ), depending only upon the specified value, Pspec , of
the acceptance probability.
In turn, assurance that the proportion, pnonc (µd , σd ; ∆), of units in the popu-
lation with values outside the interval LC ± ∆ does not exceed some specified
value, p0 , is equivalent to assuring that the combination (µd , σd ) is within an
143
level-curve for p(mu,sig) = 0.0165, delta= 0.165

0.08
0.06
0.04
sig
0.02
0.0
0.9 1.0 1.1

mu
p0 = 0.0165
level-curve for p(mu,sig) = 0.0278, delta= 0.165

0.08
0.06
0.04
sig
0.02
0.0
0.9 1.0 1.1

mu
p0 = 0.0278
Fig. 6: Combinations of batch mean, µd , and standard deviation, σd , corre-

sponding to a specified fraction, p0 , of units outside the limits LC ± 16.5%.
144 Paper D
(almost trapezodial shaped) region in the (µd , σd ) plane.
Hence, under the assumption of a normal distribution of population values, use

of the three class attribute criterion is equivalent to a parametric requirement
to the combination of population values (µd , σd ).
5.3 Confidence region approach
Consider a level curve for the probability, Pusp (µd , σd ), of passing the USP-
test under a normal distribution of dosage content, e.g. the curve corresponding
to a probability γ of passing the test, i.e. values (µd , σd ) satisfying
Pusp (µd , σd ) = γ (44)
for that value of γ. The curve defines a region in the (µd , σd )-plane of combi-
nations of batch mean and standard deviation with the properties that at least
the proportion γ of all samples tested for content uniformity under the USP-test
plan will pass the test.
Bergum [50] suggested a confidence region approach to construct acceptance

limits for d¯ and sd such that there is a specified assurance, 1 − α, that future
samples from that batch will have a specified probability, γ, of passing the
USP-test. The idea being that for each potential sample result, d¯ and sd , a con-
fidence region is constructed in the (µd , σd )-plane such that, for that sample
result, there is a confidence of (1 − α) that the combination of batch mean and
standard deviation is within this confidence region. Then, for each value, d, ¯ of
the sample mean in the interval LC ± ∆, a value of the sample standard de-
viation, sd,cr , may be determined such that the point in the confidence region
with the lowest probability of passing the USP-test is on the level curve corre-
sponding to γ. The sample standard deviation determined in this way defines
the acceptance limit for sd corresponding to that value of the sample average,
¯
d.
Thus, Bergums approach shares the spirit of the prediction interval approach
discussed in Section 4.3 in the sense that acceptance of the batch is based
upon a strong confidence that a future sample from the batch will conform to
145
some specified requirements on that sample. In this way, when assessing the
current sample, the uncertainty associated with the current sample is taken into
consideration in order to provide the assurance desired.
PDA Technical Report 25 [22] provides tables of acceptance limits (expressed

as the limiting value of the sample coefficient of variation, sd,cr /d¯ for each
value of d¯ ) for n = 10, γ = 0.95, and 1 − α = 0.10.
Analogously, the relations (37) and (39) might be used to formulate an accep-
tance criterion based upon a (1 − α) confidence interval for the proportion
pa (µd , σd ) of doses in the batch with values in the amber and red zones, and
accept the batch when the upper bound of this interval does not exceed p0 (γ).
Alternatively one could use a formulation of this criterion in terms of a sta-
tistical tolerance interval that contains at least the proportion 1 − p0 (γ) with
confidence (1 − α), and accept the batch when the tolerance limits are within
the green zone, see e.g. Hahn and Meeker [46].
5.4 Relation to theories of acceptance sampling by vari-

ables
The problem of designing sample-based acceptance criteria for monitoring the

proportion, pnonc (µ, σ; ∆), of nonconforming product items (items outside the
limits LC ± ∆) in a batch has been investigated in the literature on acceptance
sampling. When the acceptance rule is based upon sample average and sam-
ple standard deviation, the term acceptance sampling by variables for fraction
nonconforming is often used. In the following relevant results from that lit-
erature will therefore be summarized. When referring to the proportion of
nonconforming units, we shall suppress the explicit reference to the limit ∆,
and use the symbol pa (µd , σd ) to denote the proportion of units that do not
conform to the limits LC ± ∆.
The decision whether to accept or reject a batch may be formulated in terms of

146 Paper D
a test of the statistical hypothesis,

H0 : pa (µd , σd ) ≤ p0 with the alternative Ha : pa (µd , σd ) > p0 (45)
for a specified value of p0 .
In order to maintain equivalence between the phrasing “passing the test” and
“accepting the hypothesis” the statistical hypothesis to be tested in (45) has
been formulated in accordance with the discussion in Section 2.1.
The statistical problem of testing a hypothesis of the form (45) was adressed by
Wallis, [54], who suggested to use a test based upon the estimator of pa (µd , σd )
obtained by substituting µd and σd in the generic function (34) by the usual
estimates, d¯ and sd
¯ sd ) ≤ plim , otherwise reject
Accept H0 if pa (d, (46)
i.e. use the sample average and sample standard deviation to compute an es-
timate of the fraction nonconforming items using (33), and compare this esti-
mate to some limiting value, plim .
Subsequently, Lieberman and Resnikoff [33] suggested an approach based

upon an optimal (UMVU) estimator of pa (µd , σd ). This approach was adapted
in the US MIL STD 414, [55], and as this standard was solving a need in indus-
try, this standard was later adopted as an ISO Standard, ISO 3951, [35], and as
a US national standard ANSI Z1.9 [56], see also Schilling [57] and Boulanger
et al. [58].
In light of more recent developments of theories of hypotheses testing, various

authors have investigated other approaches than the MVU estimator-based ap-
proach suggested by Lieberman and Resnikoff, see e.g. Bruhn-Suhr et al. [59],
Lei and Vardeman [60].
In the following we shall follow the direct approach suggested by Wallis [54].
The difference between this direct approach by Wallis and use of the likelihood-
ratio principle is that under the likelihood-ratio approach, sd of
p the estimate,
σd is replaced by the maximum-likelihood estimate, (n − 1)/n sd . For
a comparison of the direct approach and the likelihood-ratio approach see Lei
and Vardeman [60].
147
Although the problem is a rather common problem in industrial practice, tests

of the hypothesis (45) are seldomly described in statistical textbooks.
Once, the limiting value, plim , of the estimated fraction nonconforming has
been established, the acceptance region (in the (d, ¯ sd )-plane) for the test may
¯
determined as the set of values, (d, sd ), such that
¯ sd ) ≤ plim
pa (d, (47)
with the function pa (·, ·) given by (33) and (34). The acceptance region is of
a similar shape as shown in Figure 6. The boundary of the region may be de-
termined from (41) and (42) substituting plim for p0 . When the significance
level of the test is less than 50%, plim will be larger than p0 (the sampling un-
certainty is “subtracted” from the estimated proportion nonconforming before
comparing with p0 ), and the acceptance region will include the region corre-
sponding to pa (d, ¯ sd ) ≤ p0 . When the significance level of the test is greater
than 50%, plim will be less than p0 (the sampling uncertainty is “added” to
the estimated proportion nonconforming before comparing with p0 ), and the
acceptance region will be inside the region corresponding to pa (d, ¯ s d ) ≤ p0 .
¯
The acceptance region may be expressed as a list of limiting values, slim (d),
for the sample standard deviation, sd , with the limiting values depending on
the sample average, d.¯ The maximal acceptable value of the sample standard
deviation is found from (42) for π = plim /2, d¯ = LC as
smax = ∆/z1−plim /2 . (48)
Alternatively, the acceptance region may be expressed as a list of limiting val-

¯ = slim (d)/
ues CVlim (d) ¯ d¯ for the sample coefficient of variation, with the
limiting values depending on the sample average, d¯ and with the maximal ac-
ceptable sample coefficient of variation given by (48).
For any combination, (µd , σd ), of batch mean and standard deviation, the prob-
ability of passing this dosage uniformity test with a specified limiting estimated
proportion nonconforming, plim , may be determined by numerical integration
¯ sd ) over the acceptance region
of the joint probability density function for (d,
¯ sd )-plane). It follows from the mutual statistical independence of d¯
(in the (d,
148 Paper D
and sd (see [44]) that the joint probability function may be determined as the
product of density functions corresponding to the normal distribution of d¯ and
the σd2 χ2 (n − 1)/(n − 1) distribution of s2d .
Hence, for any specified limiting value, plim , the probability of passing this
test may be determined as function of population parameters (µd , σd ). As the
test is based upon a comparison of the estimated fraction nonconforming items
with some limiting value, it is to be expected that the probability of passing the
test depends only on the proportion, pa (µd , σd ) of nonconforming items in the
population. Although this is not exactly true, the assertion generally holds to
a degree of approximation satisfactory for all practical purposes, see Lei and
Vardeman [60].
The limiting value, plim , is determined from the requirement to the assurance
desired from the test procedure. In statistical terms, the assurance is expressed
in terms of the significance level, α, of the statistical test. Thus, using sig-
nificance level, α = 0.90 say, the borderline case with population proportion
nonconforming, py (µd , σd ) = p0 , the probability of rejecting H0 is 90% (the
probability of rejecting H0 is α).
When the design value p0 in (45) has been selected as the value p0 (γ) derived
from (37) and (39), it follows from the duality between statistical confidence
regions and tests that acceptance by the test at a significance level α provides
assurance with confidence 1−α that there is at least a probability γ that a future
sample from the batch will pass the three-class modification of the content
uniformity test.
5.5 Design of a test with a trapezoidal acceptance region
The shape of the acceptance region in (d, ¯ sd )-space corresponding to a spec-

ified significance level, α, is more easily understood by first considering the
derivation of a statistical test referring to a one-sided specification limit, e.g. a
specified upper limit, U = LC + ∆.
149
Test of a one-sided upper specification
Consider a test concerning a one-sided specification limit, e.g. an upper limit,

U . The quantity of interest is

+ U − µd
p (µd , σd ) = P[U < D] = 1 − Φ . (49)
σd
Thus, we shall consider the hypothesis
H0+ : p+ (µd , σd ) ≤ p0 vs H1+ : p+ (µd , σd ) > p0 (50)
for some specified value p0 and with p+ (·, ·) given by (49).
Now introduce the transformed random variable, Z = U − D. It follows that

Z ∼ N(U − µd , σd2 ) , and hence,

+ 0 − (U − µd )
p (µd , σd ) = P[Z < 0] = Φ = Φ(−θ)
σd
with θ = (U − µd )/σd .
In terms of θ we have Z ∼ N(θσd , σd2 ).
Thus, there is a one-to-one correspondance between p+ (µd , σd ) and θ, viz.

θ = z1−p , corresponding to µd = U − σz1−p .
Accordingly, the hypothesis (50) is equivalent with a hypothesis

+ +
H0,θ : θ ≥ z1−p0 vs H1,θ : θ < z1−p0 . (51)
Let Z and Sz2 denote average and empirical variance of Z1 , Z2 , . . . , Zn . It fol-

lows from a result by Lehmann [61] that the uniformly most powerful invariant
test for H0+θ is a test with critical region of the form:
+
Reject H0,θ for z/sz < k (52)
The significance level of this test is determined from the distribution of W + =

Z/Sz = (U − d)/S¯ d . A large value of W + indicates a small fraction in the
population exceeding the upper limiting value U .
150 Paper D
The distribution of
√ U −D √
T = n = nW
Sd
is a non-central t-distribution with n − 1 degrees of freedom and with non-
√
centrality parameter δ = θ n.
√
As W + = T / n it follows that
√ √
P[W + ≤ k] = P[T / n ≤ k] = P[T ≤ k n]
and, hence
√
P[(U − D)/sd ≤ k] = P[W + ≤ k] = P[t(n − 1, δ(p)) ≤ k n] (53)
√ √
with δ(p) = θ n = z1−p n.
Thus, the critical value, k (for w+ ) at a level α test for the hypothesis (51) is
determined from
√
k = k(p0 ) = t(n − 1, δ(p0 ))α / n (54)
viz. the α percentile in the t(n − 1, δ(p0 ))-distribution. The hypothesis (51) is
accepted when w+ ≥ k(p0 ), and therefore, whenever w+ ≥ k(p0 ) it has been
demonstrated with a confidence α that p+ (µd , σd ) ≤ p0 .
For p0 = 0.0165, n = 10 and α = 0.90 one finds
k(0.0165) = 3.255.
¯ sd )-plane
The requirement determines a region in the (d,
z ≥ sd k
or,
U − d¯ ≥ sd k.
Inserting U = LC + ∆, the requirement w+ ≥ k may be expressed as
LC + ∆ − d¯ ≥ sd k(p0 ). (55)
151
Thus, the acceptance criterion is of the form
Accept H0+ when d¯ + ksd ≤ LC + ∆ (56)

otherwise reject
with k = k(p0 ) given by (54). In this form the criterion simply compares
a suitably chosen linear combination of sample average and sample standard
deviation with the upper specification for individual measurement values.
Graphically, the criterion, (55) determines a line

¯
sd = (LC + ∆ − d)/k (57)
¯ sd )-plane, passing through the point (d¯ = LC + ∆, sd = 0) with
in the (d,
slope −1/k for k = k(p0 ) given by (54).
For sample values, (d, ¯ sd ), to the left of this line the hypothesis (50) is ac-
cepted. It follows from (53) that the probability of obtaining a sample value,
¯ sd ), to the left of this line depends on (µd , σd ) only through p+ (µd , σd ). In
(d,
other words, all combinations of batch mean and standard deviation (µd , σd )
with the same proportion, p = p+ (µd , σd ) of nonconforming units (with re-
spect to the upper limit LC + ∆) will have the same probability of rejection
under this test. In statistical terms the critical region is said to be “similar”.
Introducing the estimated fraction nonconforming units (with respect to the

upper limit U ),
¯ sd )
pb+ = p+ (d, (58)
it may be verified that the acceptance criterion (56) is equivalent with
¯ sd ) ≤ plim
Accept H0+ when p+ (d, (59)
otherwise reject
with the limiting value, plim for the estimated fraction nonconforming deter-
mined from
plim = Φ(−k) (60)
with k = k(p0 ) given by (54). Thus, the limiting value for the estimated
fraction nonconforming depends as well upon the sample size as upon the sig-
nificance level, α, of the test (the desired level of assurance of p ≤ p0 ).
152 Paper D
For a level α-test, the value of k is given by (54) and the acceptance probability
for this test is determined from (53) with that value of k, i.e.
P[(U − D)/sd ≥ k] = P[t(n − 1, δ(p)) > t(n − 1, δ(p0 ))α ] (61)

√ √
with δ(p) = θ n = z1−p n.
In the SAS-system, the acceptance probability may be determined by means of

the function PROBT(T,DF,NCNTL) as
PACC = 1 - PROBT( K, N-1, NCNTL)
with K = TINV(ALFA, N-1, PROBIT(1-P0)*SQRT(N), and

NCNTL = SQRT(N)*PROBIT(1-P).
Test of a one-sided lower specification
Consider a test concerning the lower limit, L = LC − ∆, with the quantity of

interest
L − µd
p− (µd , σd ) = P[D < L] = Φ (62)
σd
and the hypothesis
H0− : p− (µd , σd ) ≤ p0 vs H1− : p− (µd , σd ) > p0 . (63)
By symmetry, it is found that a level α-test of the hypothesis (63) rejects the
hypothesis when
d¯ − LC − ∆ < sd k (64)
with k = k(p0 ) given by (54).
Accordingly, the acceptance criterion corresponding to a test at significance

level α of the hypothesis (63) regarding the lower limit is
¯ + ksd ≤ ∆
Accept H0− when (LC − d) (65)
otherwise reject.
153
The criterion (65) determines a line
sd = (d¯ − LC − ∆)/k (66)

¯ sd )-plane, passing through the point (d¯ = LC − ∆, sd = 0), and with
in the (d,
slope 1/k where k = k(p0 ) is given by (54).
Introducing the estimated fraction nonconforming units (with respect to the

lower limit L)
¯ sd )
pb− = p− (d, (67)
it is seen that the acceptance criterion (65) is equivalent with the criterion:
¯ sd ) ≤ plim
Accept H0− when p− (d, (68)
otherwise reject
with plim determined from (60) for k = k(p0 ) given by (54).
Combination of the two tests
So far, the upper and lower limits, U and L, have been considered separately.
However, the quantity of interest is the proportion,
pa (µd , σd ) = p− (µd , σd ) + p+ (µd , σd ) (69)
of units violating either of the limits.
Accordingly, the estimates pb+ from (58) and pb− from (67) may be combined
to form an estimate,
¯ sd ) = p− (d,
pba (d, ¯ sd ) + p+ (d,
¯ sd ) (70)
of the proportion of units that do not conform to the combined specifica-

tion. It seems natural to base acceptance of the hypothesis (45) concerning
pa (µd , σd ) upon a comparison of this estimated proportion of nonconforming
¯ sd ) with some limiting value plim .
units, pba (d,
¯ sd ) ≤ plim
Accept H0 when pba (d, (71)
otherwise reject
154 Paper D
where plim shall be determined to secure a significance level, α, of the test.
¯ sd )-plane satisfying
A parametric form of the boundary of the region in the (d,
¯ sd ) ≤ plim is given by (41) and (42).
pba (d,
For any combination, (µd , σd ) of the population values, the acceptance prob-
ability under this criterion may be determined by numerical integration of the
joint probability density function for (D, Sd ) over this region.
However, as noted by Lei and Vardeman [60] for any specified limiting value,
plim , there is a whole band of operating characteristic curves, and there the
determination of a limiting value plim corresponding to a specified significance
level, α, of the test of the hypothesis (45) requires extensive use of numerical
search methods.
Instead, following the suggestion by Wallis, one might construct an acceptance

region corresponding to a level α test of the two-sided specification by com-
bining the acceptance regions corresponding to level α-test of each of the two
one-sided specifications, and supplement with a requirement to the maximum
sample standard deviation derived from a level α-test of the two-sided specifi-
cation. Following this suggestion, the acceptance region will be a trapezoidal
¯ sd )-space with oblique sides given by the lines (66), (57) corre-
region in (d,
sponding to that value of α, and with upper side given by
s = smax = ∆/z1−p∗ /2 (72)
with p∗ = Φ(−k).
Thus, the resulting criterion is of the form
¯ sd ) is within the polygon specified by k

Accept H0 when (d, (73)
otherwise reject.
¯ sd )-space
It follows from the considerations in Section 5.4 that the region in (d,
¯ ¯
of values (d, sd ) such that pa (d, sd ) ≤ plim is bounded by this acceptance
polygon. Thus, the acceptance criterion (73) corresponding to the polygon
will be slightly more conservative (i.e. accept in more cases) than the criterion
155
(71) based upon a direct comparison of pa (d, ¯ sd ) with plim . However, as indi-
cated in Figure 6, the difference between these two acceptance regions is only
marginal. The difference is most apparent for larger values of sd close to smax .
Hence, for any given value of the proportion nonconforming, pa (µd , σd ), the
largest difference in acceptance probability occurs when µd = LC. This max-
imal difference in probability of acceptance under the two criteria has been
illustrated as function of pa (µd , σd ) in Figure 7. The largest difference occurs
for acceptance probabilities around 0.50.
Alternative representations of the acceptance rule
The acceptance criterion (73) specifying a trapezoidal acceptance region in

¯ sd )-space may be represented in various equivalent ways.
(d,
Rearranging (55), the requirement for acceptance at the upper limit may be
expressed as
¯
sd ≤ (LC + ∆ − d)/k
or,
sd /d¯ ≤ (LC + ∆ − d)/(k
¯ ¯
d) (74)
which, corresponding to each value of the sample average, d, ¯ specifies an upper
limit to the sample coefficient of variation to be applied for LC < d.¯
Similarly, from rearranging (64), the requirement for acceptance at the lower
limit may be expressed as
sd ≤ (d¯ − LC − ∆)/k
or,
sd /d¯ ≤ (d¯ − LC − ∆)/(kd)
¯ (75)
providing an upper limit to the sample coefficient of variation to be applied for
d¯ < LC.
For values of d¯ in the interval,
LC − ∆(1 − z1−plim /z1−plim /2 ) ≤ d¯ ≤ LC + ∆(1 − z1−plim /z1−plim /2 ) (76)

156 Paper D
acceptance probability for plim (lower curve), and trap. region(upper curve)
1.0
0.8
0.6
Pacc
0.4
0.2
0.0
0.0 0.05 0.10 0.15 0.20
acceptance probability under plim vs acc.prob. for trap.reg.

1.0
0.8
0.6
ocplim
0.4
0.2
0.0
0.0 0.2 0.4 0.6 0.8 1.0

octrap
∆ = 0.165; n = 21, k = 1.91, smax = 0.075; plim = 0.0278; µd = LC.
Fig. 7: Top: Acceptance probability under the criterion (71) based upon the
estimated proportion nonconforming, and under the criterion (73) for the cor-
responding trapezoidal rule. Bottom: P − P -plot of corresponding values of
the acceptance probability under the two criteria. The identity line has been
superimposed on the graph.
157
corresponding to the center part of the trapezoid (see (43)), the criterion to the
sample standard deviation is
sd ≤ smax
with the maximum sample standard deviation, smax given by (48). Hence, for
these values, the criterion to the sample coefficient of variation is
sd /d¯ ≤ smax /d.

¯ (77)
The requirements (74), (75) and (77) may be presented in a tabular form as in
PDA Technical Report No. 25 [22].
Combining (56) and (65) it is seen that the region bounded by the oblique sides
of the trapezoid may be delimited by the requirement
¯ + ksd < ∆
|LC − d| (78)
i.e. the absolute value of the deviation between sample average and LC plus
(or minus, depending on the sign of k) some adjustment for sampling uncer-
tainty shall not exceed the specification ∆ (e.g. 16.5%). It follows from (54)
that when the level of significance, α has been chosen to be larger than 0.50,
then k > 0 and, hence the uncertainty adjustment is added as an extra penalty,
whereas when α < 0.50, the uncertainty adjustment is subtracted before com-
parison with ∆.
The adjustment for sampling uncertainty
Irrespective of the representation of the acceptance rule, considerations on the

approach to the influence of the sampling uncertainty are reflected through the
choice of significance level, α, that in turn is determinant for the value k given
by (54) that is used in the acceptance rule.
In this section we shall briefly describe the effect of the choice of significance
level on the designation of the various limiting values. For ease of trans-
parency, we shall first consider a test of a one-sided upper specification as
discussed in the subsection page 149.
158 Paper D
From (55) it is seen that in essence the criterion determines an estimate of

z1−p0 as (LC + ∆ − d)/s¯ d , and compares this estimate with a limiting value k.
If there was no sampling uncertainty, and d¯ and sd represented the population
values (µd , σd ) then k should be chosen to be z1−p0 . However, accounting for
the sampling uncertainty (in accordance with the specified level of significance,
α,) the limiting value, k will also depend upon the level of significance, α, and
the sample size, n, as seen from (54).
The approach to the adjustment for sampling uncertainty depends on the value
of the significance level α. When α > 0.50 (and p0 < 0.5) then k will be
larger than z1−p0 , implying that the sample average shall be even less than
d¯lim (sd ) = (LC + ∆) − z1−p0 sd (79)
in order to provide satisfactorily strong evidence of pa < p0 .
For significance levels α < 0.50 the value of k will be smaller than z1−p0 , im-
plying that even sample averages exceeding d¯lim (sd ) may lead to acceptance,
if only the exceedance is not larger than what may be attributed to the sampling
uncertainty. Thus, for α < 0.50 the procedure accepts unless there is strong
sample evidence contradicting pa < p0 .
The larger the sample, the smaller is the adjustment for sampling uncertainty,
and the closer will k be to z1−p0 .
The interpretations above are valid also for the combined test in the sub sec-
tions page 153 and page 155.
Analogously, considering the formulation (71) of the acceptance criterion in

terms of the limiting value, plim , for the estimated fraction nonconforming, it
is seen that in order for the test to provide sufficient assurance that the pop-
ulation proportion nonconforming, pa (µd , σd ), really is smaller than p0 , the
significance level, α, shall be chosen to be larger than 50%, and consequently
the limiting sample fraction nonconforming, plim , is smaller than p0 to allow
for sampling uncertainty. The larger the sample, the closer plim will be to p0 .
This is in line with an interpretation of the statistical test in terms of a con-

fidence interval for pa (µd , σd ). Acceptance by the statistical test of H0 :
159
pa (µd , σd ) ≤ p0 at significance level α is equivalent with accepting when-

ever the upper α confidence limit for pa (µd , σd ) does not exceed p0 . Hence, in
order for the test to provide assurance that pa (µd , σd ) ≤ p0 , the significance
level, α, of the test shall be larger than 50%. In the tables provided in PDA
Technical Report No. 25 [22], an assurance, α = 0.90 has been suggested.
In terms of acceptance sampling, a requirement of a 90% assurance corre-

sponds to specifying the quality, p0 as the quality that should have a very low
probability of acceptance, Pacc = 0.10, under the acceptance sampling proce-
dure. Such a quality is often termed Limiting quality, (LQ10 ), indexed by the
specified low acceptance probability.
5.6 Discussion
Disregarding the requirement to the sample coefficient of variation in the USP 21

content uniformity test, it was found in Section 5.2 that acceptance probabil-
ity under the three-class attribute requirements on the individual measurement
values in the two-stage procedure only depends upon the proportion of values
in the population outside the innermost set of limiting values LC ± ∆ (e.g.
∆ = 0.165). In Section 5.2 a table of the relation between this proportion of
population values outside LC ± ∆ and the acceptance probability under the
USP 21 test was provided showing that a requirement of a 95% probability of
passing the USP test is equivalent with a requirement that no more than 1.65%
of the population values are outside the limits LC ± ∆. Thus, a requirement
on the probability of passing the USP-test may be translated to a population
requirement to the proportion of values in the population outside the innermost
limits for individual values in the USP-test.
The statistical procedures for acceptance sampling by variables have been tay-
lored for assessing such population values, and the assurance provided by these
procedures under due consideration to the sampling uncertainty has been dis-
cussed in the statistical literature. An operational acceptance criterion may be
formulated in terms of a trapezoidal acceptance polygon in (d, ¯ sd )-space that
approximates the criterion (71) on the estimated proportion of values outside
the limits.
160 Paper D
The acceptance polygon criterion may be formulated in various equivalent

ways. In particular the criterion may be phrased in terms of a requirement
to an acceptance value (78), i.e. linear combination of sample average and
standard deviation combined with a requirement to the maximal value of the
sample standard deviation.
6 Assessment of the properties of the USP pre-

view dosage uniformity test
As part of the effort for global harmonization of Uniformity of Dosage Units

tests, various changes to the USP general test chapter Uniformity of Dosage
Units < 905 > have been suggested. The first USP suggestion of changes
[27] were modelled after the test provided in the Japanese Pharmacopeia that
was based upon such procedures for acceptance sampling by variables that are
used in the International Standard ISO 3951 [35] for acceptance sampling by
variables for proportion nonconforming items (equivalent to [56]).
Responding to this suggestion, the Statistics Working Group of the Pharma-

ceutical Research and Manufacturers of America (PhRMA) proposed an alter-
native that provided the Japanese, the European and the US Pharmacopeia with
a harmonized Content Uniformity Test [29], [62]. During the review process
various amendments have been introduced resulting in the USP draft proposal
[28] published in 2001 as a stage 4 draft.
In the following the statistical properties of the acceptance criteria for unifor-
mity of dosage units by the Content Uniformity method will be investigated.
We shall consider the acceptance criteria only for such situations where the
target test sample amount at the time of manufacture equals the label claim
(LC = 100%, or, in fraction LC = 1).
The proposal has been formulated in terms of an “acceptance value”, and limits
for individual units in the sample as follows:
161
“Calculate the acceptance value as
¯ sd ) = |M − d| + ksd
A = A(d, (80)
with (
2.4 when n = 10
k = k(n) = (81)
2.0 when n = 30
and (
¯ 0.985}
max{d, when d¯ ≤ 1.0
¯ =
M = M (d) (82)
¯
min{d, 1.015} ¯
when 1.0 < d.
The requirements are met if the acceptance value of the first 10 dosage units is
less than or equal to 0.15 (denoted L1 ). If the acceptance value is greater than
0.15, test the next 20 units and calculate the acceptance value.
The requirements are met if the final acceptance value of the 30 dosage units
is less than or equal to 0.15 (L1 ), and no unit is over the deviation of 0.25
¯
(denoted L2 ) from the calculated value of M (d).”
Two different statistical issues are involved in these requirements
a) In each stage, the criteria include as well a parametric test based upon
the acceptance value, (inspection by variables), as limits on individual
measurements (inspection by attributes)2 .
As the requirements in the attribute test involves an “accept zero” crite-
rion, we shall use the term “non-satisfactory units” for units with mea-
surements outside the limits for individual units.
b) The plan is a two-stage plan with a decision on whether to invoke the

second stage depends on the result in the first stage, and the criteria to
be applied in the second stage depends on the results in the combined
sample (30 units)
2
In [28] the attribute criterion is applied only in the second stage. However, as the attribute
criterion also was applied in the first stage in the previous draft [29] and [62] we have chosen
this criterion also in the first stage.
162 Paper D
As the criteria involve the individual values of the 30 sample units, a direct as-
sessment of the operating properties of the test by analytical methods requires
numerical integration over a region in a 30-dimensional space, and therefore
we have chosen to assess the properties using numerical simulation.
However, before reporting the results of the simulations, the criterion on the
acceptance value will be further discussed.
6.1 The acceptance value
Rearranging (82) one finds



 0.985 when d¯ < 0.985
¯
M (d) = d¯ when 0.985 ≤ d¯ ≤ 1.015 (83)

 1.015 when 1.015 < d¯
and, hence the criterion

¯ s d ) ≤ L1
A(d, (84)
on the calculated acceptance value is equivalent with the set of criteria
when d¯ < 0.985 ¯ + ksd ≤ L1

: (0.985 − d)
when 0.985 ≤ d¯ ≤ 1.015 : ksd ≤ L1 (85)
when 1.015 < d¯ ¯ + ksd ≤ L1
: (1.015 − d)
where the inequality to be satisfied in order to pass the criterion (84) depends
¯
on the value of d.
As the sample standard deviation, sd ≥ 0, no value of d¯ less than 0.985 − L1 ,

or greater than 1.015 + L1 will satisfy the criterion, and hence the criterion
(84) on the acceptance value may be reformulated as
when d¯ < 0.985 − L1 : not applicable

when 0.985 − L1 ≤ d¯ < 0.985 : ¯ + ksd ≤ L1
(0.985 − d)
when 0.985 ≤ d¯ ≤ 1.015 : ksd ≤ L1 (86)
when 1.015 < d¯ ≤ 1.015 + L1 : ¯ + ksd ≤ L1
(1.015 − d)
when 1.015 + L1 < d¯ : not applicable.
163
Thus, the criterion on the acceptance value determines a trapezoidal region in

¯ sd )-space with corners in the points
(d,
A: d¯ = 0.985 − L1 ; sd = 0
B: d¯ = 0.985 ; sd = L1 /k
¯ (87)
C: d = 1.015 ; sd = L1 /k
D: d¯ = 1.015 + L1 ; sd = 0.
Comparing the region defined by (87) with the trapezoidal acceptance region
derived in the subections page 153 and page 155, it is seen that letting ∆ =
L1 + 0.015 and k(p0 ) be given by (81), the two approaches lead to trapezoids
with coinciding oblique sides. They only differ in the position of the upper
horizontal line specifying the maximum sample standard deviation. Hence,
on each stage, the criterion on the acceptance value essentially controls the
proportion of population units outside the interval LC ± (0.015 + L1 )
Table 3 summarizes the criteria in terms of sample average and standard devi-
¯ sd ), and limits for individual measurements3 .
ation (d,
Thus, the procedures in the proposal represent a shift of paradigm compared to

the procedures in the 1984 USP 21 [42]. Most notably, the role of the sample
average has changed from serving only a normalizing quantity for the sample
standard deviation (to form the sample coefficient of variation) to serve now
a descriptor of the location of the distribution used in the acceptance value to
monitor the proportion of units outside the interval LC±0.165 more effectively
than the previous attribute criterion on values in the “amber” zone.
6.2 The simulation study
Overall probability of acceptance
Under the distributional assumption (5), the distribution of individual sam-

ple values depends only on population mean and standard deviation, (µd , σd ).
3
In the simulations, we have used the sample average in the first sample to determine the
limits to be applied to measurements in the first sample, and (when applicable) the total sample
average of all 30 measurements to be applied to the subsequent 20 measurements.
164 Paper D
Limits
Acceptance limits
for individual
Value of d¯ for sd
units in the sample
Stage 1, 10 units
d¯ < 0.835 - -
0.835 ≤ d¯ ≤ 0.985 ¯
sd ≤ (d − 0.835)/2.4 0.74 - 1.23
0.985 ≤ d¯ ≤ 1.015 sd ≤ 0.0625 0.75d¯ − 1.25d¯†
1.015 ≤ d¯ ≤ 1.165 ¯
sd ≤ (1.165 − d)/2.4 0.76 − 1.27
1.165 < d¯ - -
Stage 2, all 30 units
d¯ < 0.835 - -
0.835 ≤ d¯ ≤ 0.985 ¯
sd ≤ (d − 0.835)/2.0 0.74 - 1.23
0.985 ≤ d¯ ≤ 1.015 sd ≤ 0.075 0.75d¯ − 1.25d¯
1.015 ≤ d¯ ≤ 1.165 ¯
sd ≤ (1.165 − d)/2.0 0.76 − 1.27
1.165 < d¯ - -
Table 3: Acceptance limits for sample standard deviation sd and limits for
individual sample units at each stage in the dosage uniformity test
165
Therefore the operating properties of the criteria for a distribution of popula-

tion values characterized by its mean and standard deviation, (µd , σd ), may be
assessed by simulating random samples from a normal distribution with mean
and standard deviation (µd , σd ).
The simulations were performed in a lattice in (µd , σd )-space, with 3000 sam-
ples (each consisting of 30 units) simulated in each lattice point. Hence, the
uncertainty due to simulation is at most ±0.02 (for values of the acceptance
probability in the neighbourhood of Pacc = 0.5) and less than ±0.0001 when
Pacc ≥ 0.75, or Pacc ≤ 0.25.
Figure 8 shows the level-curves (in (µd , σd )-space) for the overall probabil-
ity of acceptance under the procedure. The shape of the curves resembles the
curves of the solution to pnonc (µd , σd ; ∆) = p (see (40) and Figure 6) for ∆ =
0.165. For comparison, Figure 9 shows the solution to pnonc (µd , σd ; 0.165) =
p for various values of p. Superimposing the graphs in the two figures it is seen
that there is a good agreement between the shapes of the two set of curves.
Thus, the criterion on the acceptance value that at each stage controls the pro-
portion of population units outside the interval LC ± (0.015 + L1 ), apparently
also has a dominant effect on the properties of the overall two-stage procedure
with supplementary attribute criteria on individual measurement values, di .
Comparing the graphs in the two figures it is seen that populations with a pro-
portion pnonc (µd , σd ; 0.165) = 0.005 of values outside the interval 1.0±0.165
will have a probability of more than 99% of being accepted by the procedure;
when the population proportion of values outside the interval 1.0 ± 0.165 is
0.09, the acceptance probability is 10%.
For each lattice point (µd , σd ), the value of the acceptance probability Pacc (µd , σd )
has been plotted against the corresponding value of pnonc (µd , σd ; 0.165) in
Figure 10. The rather narrow sigmoid-shaped scatter of points confirms the
impression from Figure 8 that the criteria essentially controls the population
proportion of values outside the interval 1.0 ± 0.165.
The graph allows for an assessment of the approximate probability of accep-

tance corresponding to any given value of the population proportion of values
166 Paper D
Total probability of acceptance,

levels: 0.1, 0.25, 0.5, 0.75, 0.99
0.14
0.12
0.10
0.08
sigma
0.1
0.06 0.50
0.04
0.99 0.75
0.99
0.02
0.10
0.25
0.50
0.75
0.0 0.99
0.7 0.8 0.9 1.0 1.1 1.2 1.3
mu
Fig. 8: Level-curves for the overall probability of acceptance for the procedure.

levels 0.005, 0.01, 0.015, 0.02, 0.025, 0.03, 0.035, 0.04
0.14
0.12
0.10
0.08
sigma
0.06
0.04
0.010
0.04 0.005
0.02
0.0 0.040
0.035
0.030
0.025
0.020
0.015
0.010
0.005
0.7 0.8 0.9 1.0 1.1 1.2 1.3
mu
Fig. 9: Combinations, (µd , σd ) of population mean and standard deviation cor-

responding to specified values, pnonc (µd , σd ; 0.165) of the proportion of pop-
ulation values outside the interval 1.0 ± 0.165.
167
outside the interval 1.0 ± 0.165 as indicated in the table below.
pnonc 0.006 0.010 0.075 0.10

Pacc (p) 0.95 0.90 0.10 0.05
Thus, whenever a batch is accepted by the procedure, there is a confidence

of 90% that no more 7.5% of the units in the batch have measurement values
outside the interval LC ± 0.165.
However, as the scatter in the vertical direction is larger than what can be ex-
plained by simulation uncertainty, there is not a unique value of the acceptance
probability corresponding to a given value of pnonc (µd , σd ), but rather an in-
terval of values with the specific value of the acceptance probability depending
on the particular combination (µd , σd ) giving rise to this proportion noncon-
forming. Thus, the acceptance probabilities are represented by an OC “band”
rather than a single OC-curve. It is well-known that even in the case of sin-
gle sampling plans such an OC band is an inherent feature of the statistical
test, but the thickness of the band depends on the test method selected (see
Lei and Vardeman [60]). The thickness of the band clearly has an effect on the
steepness of the OC-curve. The thicker the band the less steep is the OC-curve.
For the two-stage plan under study it is believed that the thickness is further
enhanced by the choice of the position of the upper horizontal line of the
trapeziodal acceptance region in the two stages. Moreover, as discussed in
the subsection below (page 167), the additional attribute-criterion on individ-
ual measurement values further contribute to the thickness of the band.
The discriminatory effect of the limits for individual measurements
Grossly speaking, the limits to be applied for the individual measurement val-
ues correspond to an interval ±25% (termed L2 ) around the label claim.
In order to obtain an overall impression of the protection against population

proportion of nonsatisfactory units (i.e. with values outside the interval LC ±
168 Paper D
Probability of acceptance versus fraction outside 0.835 - 1.165

1.0
•••••••••••
••••
•••••
••••
••••
•••••
••
•••
0.8
•••••
•
••• •
•
••••••
•••••
Probability of acceptance
• •
•••
0.6
••••••
••• •
•••••••••
•••••
•• ••
•••••••
0.4
•• •
•••
•• •
•••••
••• ••••
• ••
••••
••••••• •
•••••
0.2
••••••••••
• •••••••
•• ••••• •
••••••••••••••••
••• ••••••••• •••
• • ••••••••••
• ••••• ••••••••••••••••••••••••••
• • • •• • •••••••• •••••• ••••••••• •••••••••••••••••••••••• •• • • • •• • •
0.0
• • • ••• • • •
0.0 0.05 0.10 0.15 0.20

Proportion of tablets outside 0.835 - 1.165 LC
Fig. 10: Overall probability of acceptance vs. the proportion,

pnonc (µd , σd ; 0.165), of population values outside the interval 1.0 ± 0.165.
0.25), Figure 11 shows the solution to pnonc (µd , σd ; 0.25) = p for various
values of p. Superimposing the graphs in Figure 11 on the graph of the level-
curves of the OC-surface of the procedure (Figure 8) it is seen that the dis-
crimatory power against proportion of nonsatisfactory units depends strongly
upon the particular combination of population mean and standard deviation,
(µd , σd ). In some cases, populations with 0.01 % nonsatisfactory units will
be rejected with a high probability (larger than 0.9) whereas in other cases
(depending on the value of µd ) such populations will be accepted with a prob-
ability higher than 0.99.
However, in greater detail as seen from Table 3 (page 164), the limits depend
on the value of the sample average, thereby extending the range of values to the
slightly asymmetric interval 0.74 − 1.27. Moreover, as the limit to be applied
depend on the sample average d¯ that is subject to random error, the criteria on
the individual limits might contribute some random “noise” to the acceptance
value criterion and result in a less steep OC-surface.
In order to assess the effect of the attribute criteria on the discriminatory power
169

levels 0.0001,0.0005,0.001,0.005,0.01,0.05,0.10,0.15
0.14
0.12
0.10
0.0500
0.08
sigma
0.06
0.15
0.04 0.0001
0.02
0.0005
0.0001
0.1500
0.1000
0.0500
0.0100 0.1000
0.1500
0.0 0.0050
0.0010
0.0001
0.0005
0.7 0.8 0.9 1.0 1.1 1.2 1.3
mu
Fig. 11: Combinations, (µd , σd ) of population mean and standard deviation

corresponding to specified values, pnonc (µd , σd ; 0.25) of the proportion of
nonsatisfactory units in the population (i.e. with values outside the interval
1.0 ± 0.25).
170 Paper D
of the procedure, the simulation resuls were recalculated disregarding the cri-
teria on individual measurement values and using only the criterion based upon
the acceptance value for determining acceptability. Figure 12 shows the level-
curves (in (µd , σd )-space) for the probability of acceptance when disregarding
the criteria for the individual measurement values. Superimposing these graphs
upon the level-curves in Figure 8, it is seen that there is virtually no difference
between the two sets of level-curves. Thus, under the assumption of normally
distributed measurement values, the criteria for the individual measurement
values does not affect the ability to discriminate between different combina-
tions (µd , σd ) of population values. This is in line with the fact that (d, ¯ sd )
are jointly sufficient for (µd , σd ) and therefore knowledge about the individual
measurement values does not add to the information provided by (d, ¯ sd ).
Probability of acceptance - disregarding the attribute test,

levels: 0.1, 0.25, 0.5, 0.75, 0.99
0.14
0.12
0.10
0.08
sigma
0.1
0.06 0.50
0.04
0.99 0.75
0.99
0.02
0.10
0.25
0.50
0.75
0.0 0.99
0.7 0.8 0.9 1.0 1.1 1.2 1.3
mu
Fig. 12: Level-curves for the overall probability of acceptance for the proce-
dure, disregarding the criteria on individual measurement values.
The two stages
An overview of the procedure and the possible conclusions at each stage has
been provided in Figure 13.
171
Stage 1
All of 10 tablets within the acceptance

limits for individual tablets
NO YES
REJECT
The combination of
within the acceptance area ACCEPT
for stage one
NO
Stage 2
Add 20 tablets to the first sample

to a total sample of 30 tablets. All 30
tablets within the acceptance limits
for individual tablets
NO YES
REJECT
The combination of
within the acceptance area
ACCEPT
for stage two
NO
REJECT
Fig. 13: Schematic representation of the procedure for content uniformity test-
ing.
After assaying the first sample, three different actions are possible
1. Accept without further testing when
a) all 10 units are within the attribute limits, and

¯ sd )
b) the calculated acceptance value does not exceed 0.15 (i.e. (d,
is within the trapezoidal region)
2. Test the next 20 units when
a) all 10 units are within the attribute limits, but

¯ sd ) is outside
b) the calculated acceptance value exceeds 0.15 (i.e. (d,
the trapezoidal region)
3. Reject (i.e. the test is not passed) when one or more units is beyond the
attribute limits, irrespective of the acceptance value4 .
4
The wording in [28] does not explicitly specify this option; however in previous versions
explicit provisions were given for this option. Moreover, the option makes sense in practice, as
172 Paper D
Thus, it is possible to conclude the test after testing only 10 units, by outright
acceptance (when the acceptance value is satisfactorily small), or by outright
rejection (when one or more nonsatisfactory units are found in the first sample).
Figure 14 shows level curves in (µd , σd )-space of the probability of invoking

the second stage of the procedure. Corresponding to a given level of the prob-
ability of invoking the second stage there are two curves, viz. one (innermost)
curve corresponding to a constant value of pnonc (µd , σd ; 0.165) and another
(outermost) corresponding to a constant value of the proportion of nonsatisfac-
tory units, pnonc (µd , σd ; 0.25). This reflects the trade-off between the effect of
the criterion on the acceptance value monitoring pnonc (µd , σd ; 0.165), and the
attribute criterion pnonc (µd , σd ; 0.165) monitoring pnonc (µd , σd ; 0.25).
Probability of invoking stage 2,

levels: 0.05, 0.1, 0.25, 0.5, 0.75, 0.99
0.14
0.05
0.12
0.05
0.10
0.08
sigma
0.06
0.04
0.05
0.02
0.99
0.05
0.10
0.25
0.50 0.75
0.50
0.25 0.99 0.25
0.10
0.05
0.0 0.75
0.99 0.10
0.05
0.7 0.8 0.9 1.0 1.1 1.2 1.3
mu
Fig. 14: Level-curves of the probability of invoking the second stage of the
procedure.
Populations corresponding to (µd , σd )-combinations in the inner triangular

area with a low probability of invoking the second stage have a high probability
a situation with at least one nonsatisfactory unit found in the first sample would lead to rejection
after the second sample, anyhow. Here we have disregarded the extra complication arising from
the fact that the attribute limits depend on the sample average, and therefore may change from
the first sample to the combined sample.
173
of being accepted in stage 1, whereas populations corresponding to (µd , σd )-

combinations outside the plotted level-curves with a low probability of invok-
ing the second stage have a high probability of being rejected on stage 1. Pop-
ulations corresponding to (µd , σd )-combinations in the “sausage-shaped” area
in the middle have a high probability of invoking stage 2, requiring assay of
further 20 units.
Comparing with the level-curves of the overall acceptance probability in Figure

8, the effect of giving “a second chance” for such “mediocre populations”, (i.e.
populations corresponding to (µd , σd )-combinations in the middle “sausage-
shaped” area that are not accepted in stage 1) may be assessed. Consider e.g.
populations corresponding to the innermost 75%-level curve on Figure 14. The
curve is seen to correspond to the 50%-level curve of the overall acceptance
probability in Figure 8. Thus, such populations are accepted in stage 1 with a
probability of 25%, and with 75% probability they are given a second chance,
but only one third of these second chances lead to final acceptance.
Thus, from a purely statistical point of view, the test by attributes in stage 1
serves the purpose of saving testing ressources for batches that would have
been rejected anyhow (after testing all 30 units). This might, however, have
been achieved by using also a criterion based upon sample average and stan-
dard deviation for rejection in the first stage. Schilling [12] describes the design
of two-stage sampling plans by variables that allow for rejection also at stage
1.
As already noted, under the assumption of a normal distribution and indepen-

dent samples, requirements on individual measurement results are redundant.
However, in contemporary industrial applications of acceptance sampling pro-
cedures a so-called “accept zero” principle is sometimes considered, viz. a
lot can only pass if no nonconforming items are found in the sample, see e.g.
[36]. Although the discriminatory power achieved using only such a criterion
by attributes is inferior to the power when using sample average and standard
deviation, the psychological advantage of invoking this principle is that it con-
veys a signal that items outside specification are of great concern, and such
items should not be found, neither in a sample, nor in the batch. Moreover,
the use of an accept zero criterion on individual measurements to supplement
174 Paper D
the criterion on sample average and standard deviation has the statistical ad-
vantage to make the test procedure more robust towards deviations from the
assumption of a normal distribution of dosage content in the tablets.
6.3 Robustness against deviation from distributional assump-

tions
In order to assess the robustness of the procedure to the assumption of a nor-

mal distribution of sample values the simulations were performed also for a
lognormal distribution of individual measurements.
The simulations were performed in the same lattice in (µd , σd )-space as in the
previous section. In each lattice point, (µd , σd ), 3000 samples (each consisting
of 30 units) were simulated as random values from a lognormal distribution
with that mean and standard deviation.
Figure 15 shows the level-curves (in (µd , σd )-space) for the overall probability
of acceptance under this distributional assumption. Comparing with the level-
curves in Figure 8 it is seen that the part of the curves corresponding to values
1.0 < µd and larger values of σd have been shifted slightly to the right un-
der the lognormal distribution, indicating a slighty larger probability of accep-
tance under the lognormal distribution than under a normal distribution with
the same mean and variance. This may be interpretated by the fact that under
the lognormal distribution there is a heavier concentration of probability mass
to the left of the mean than under the corresponding normal distribution, and
this outweighs the heavier right-hand tail of the lognormal distribution.
However, the heavier tail of the lognormal distribution might imply that the
attribute criterion on individual measurement values would have a greater in-
fluence than under the normal distribution. In analogy with Figure 12 for nor-
mally distributed measurement values, Figure 16 shows the level-curves for
the probability of acceptance when disregarding the criteria for the individual
measurement values. Superimposing these graphs upon the graph in Figure
15 it is seen that there is virtually no difference between the two sets of level
curves. Thus, also under a lognormal distribution of individual measurement
175
Total probability of acceptance, LN,

levels: 0.1, 0.25, 0.5, 0.75, 0.99
0.14
0.12
0.10
0.08 0.25
sigma
0.1
0.06
0.04
0.99
0.02
0.10
0.25
0.50
0.75
0.0 0.99
0.7 0.8 0.9 1.0 1.1 1.2 1.3
mu
Fig. 15: Level-curves for the overall probability of acceptance for the proce-
dure under a lognormal distribution of measurement values.
values, the criteria on the accectance values are the dominant criteria of the
procedure.
6.4 Equivalent single sampling plan
As an illustration of the direct approach in Section 5.5, consider the determi-

nation of a single sampling plan with the same discriminatory properties as the
USP preview dosage uniformity test.
The following approximate values for the probability of acceptance as function

of the proportion, p, of units outside the interval LC ± 0.165 are read off from
Figure 10:
p 0.006 0.010 0.075 0.10

Pacc (p) 0.95 0.90 0.10 0.05
176 Paper D
Probability of acceptance - disregarding the attribute test, LN,

levels: 0.1, 0.25, 0.5, 0.75, 0.99
0.14
0.12
0.10
0.08 0.25
sigma
0.1
0.06
0.04
0.99
0.02
0.10
0.25
0.50
0.75
0.0 0.99
0.7 0.8 0.9 1.0 1.1 1.2 1.3
mu
Fig. 16: Level-curves for the overall probability of acceptance for the pro-
cedure under a lognormal distribution, disregarding the criteria on individual
measurement values
When a single point, (p, Pacc (p)) on the OC-curve has been specified, it fol-
lows from the considerations in Section 5.5, that such a specification defines a
relation between sample size n and limiting value, plim = plim (n), of the es-
timated proportion nonconforming units such that the OC-curves of all such
sampling plans with (n, plim ) satisfying this relation will pass through the
specified point (p, Pacc (p)). The larger the sample size, the steeper the OC-
curve.
Correspondingely, when two points (pa , 1 − α) and (pr , β)) on the OC-curve
have been specified, it is possible to determine a unique combination, (n, plim )
such that the OC-curve for that sampling plan satisfies
Pacc (pa ) ≥ 1 − α , and Pacc (pr ) ≤ β
i.e. that the plan provides at least the protection specified.
Often pa is termed the “producer’s risk quality” with α denoting the corre-
sponding producer’s risk, and pr is termed the “consumer’s risk quality” with
177
β denoting the corresponding consumer’s risk.
In terms of assurance, 1 − β denotes the assurance that an accepted batch will

have a proportion nonconforming that does not exceed pr .
The table below shows the single sampling plans “matching” the acceptance
procedures in the USP preview dosage uniformity test for various choices of
matching points, (pa , 1 − α) and (pr , β)). It is seen that the discriminatory
power of the 10 − 20 two-stage plan corresponds to a single sampling with
sample size slightly larger than 20 units. Thus, for good quality productions
the savings when using the two-stage plan is 50% (corresponding to acceptance
in stage 1), and the extra effort when analysis of the second sample is called
for, also amounts to 50% of the sample size for the single sampling plan.
pa α pr β n k smax plim
0.010 0.10 0.10 0.05 22 1.89 0.076 0.0299
0.006 0.05 0.10 0.05 21 1.91 0.075 0.0278
0.010 0.10 0.075 0.10 24 1.90 0.075 0.0278
The OC-curves have been depicted in Figures 17 to 19.
7 Further issues
We have only discussed situations with independent, identically distributed

measurement values where (µd , σd ) may be considered to reflect the popula-
tion mean dose and the dispersion of doses in the population. This assumption
represents an “ideal” situation that does not necessarily reflect all situations
occurring in practice. In practice, some other factors might influence the assay
mean and the dispersion of the population of potential assay values.
One such factor could be the variation introduced by the analytical procedure.
The effect of measurement error on the operating properties of procedures for
acceptance sampling by variables has been discussed in [31] and [32].
178 Paper D
acceptance probability for plim (lower curve), and trap. region(upper curve)
1.0
0.8
0.6
Pacc
0.4
0.2
0.0
0.0 0.05 0.10 0.15 0.20
Fig. 17: OC-curve for single sampling plan with n = 21, k = 1.91, smax =
0.075; plim = 0.0278.
acceptance probability for plim (lower curve) and trap. region (upper curve)
1.0
0.8
0.6
Pacc
0.4
0.2
0.0
0.0 0.05 0.10 0.15 0.20
0.076; plim = 0.0299.
179
acceptance probability for plim (lower curve) and trap. region (upper curve)
1.0
0.8
0.6
Pacc
0.4
0.2
0.0
0.0 0.05 0.10 0.15 0.20
0.075; plim = 0.0278.
Another important factor could be the sampling design used. In practice, in

particular in blend sampling, a so-called nested design is sometimes used,
where e.g. n = 36 sample values are obtained by repeatedly sampling sub-
dividing from only 12 locations. [22] and [62] presents examples showing
the operating characteristics of the procedures in the draft USP-proposal un-
der various assumptions on the magnitude of the within location and between
location variation. A general discussion of the operating characteristics of a
single sampling plan under a nested sampling scheme has been given in [63].
8 Discussion
In the paper we have discussed the statistical properties of various testproce-

dures used for content uniformity and blend uniformity analysis.
Traditionally, in pharmaceutical regulatory practice such procedures have spec-

ified limiting values of sample statistics rather than requirements to population
180 Paper D
values. Thus, considerations on the sampling uncertainty have mainly been im-
plicitly considered in the specification of the limiting values, and the assurance
provided by the acceptance criteria has not been very transparent.
To overcome this deficiency, various approaches using acceptance criteria based

upon prediction intervals or statistical tolerance intervals for future samples
have been suggested in the pharmaceutical literature.
In the paper we have discussed the acceptance criteria in terms of the statistical
hypothesis concerning population values that is implicitly underlying the cri-
teria. Using the concepts from statistical theory of hypothesis testing, a more
transparent description of the assurance provided by the criteria, and the con-
siderations on the sampling uncertainty is obtained.
In particular, we have studied the statistical properties of the acceptance criteria

for uniformity of dosage units in the USP draft proposal [28]. The criteria have
been related to procedures for acceptance sampling by variables as described
in the statistical literature and the individual components of the procedures
have been interpretated in terms of concepts from theories of acceptance sam-
pling. The overall properties of the USP draft proposal have been assessed by
means of simulation, and as an example of the use of theories and procedures
from theories of acceptance sampling a single sampling plan with operating
characteristics matching this proposal has been determined.
181
9 List of symbols
Symbol
E[X] Mean of the population distribution of X
V[X] Variance of the poulation distribution of X
D Relative dose i.e. mass of drug (in blend sample, or in
tablet) as fraction, or percentage of target value (ran-
dom variable)
d Relative dose in sample unit (actual value)
µd Mean relative dose (population value)
σd Standard deviation in distribution of relative doses
(population value)
Cd Coefficient of variation in distribution of relative doses
(population value)
Cd = σd /µd
n Number of units in sample
D sample average relative dose per sample unit (random
variable)
d¯ sample average relative dose per sample unit (actual
value)
Sd sample standard deviation of relataive doses in sample
(random variable)
sd sample standard deviation of doses in sample (actual
value)
LC Required dose, Label Claim. In Section 5 and 6
measurements are relatively to LC, i.e. LC=1 (100%).
χ2f Random variable distributed according to a χ2 -
distribution with f degrees of freedom
χf , p 2 p’th quantile in χ2 -distribution with f degrees of free-
dom, (i.e with probability mass p to the left of this
value)
Some (mainly US) textbooks use the notation χ2α,ν to
denote the so-called χ-squared critical value, denot-
ing the number on the measurement axis such that the
probability mass for the χ2ν -distribution to the right of
this value is α.
F(f1 , f2 ) Random variable distributed according to a F-
distribution with (f1 , f2 ) degrees of freedom
F(f1 , f2 )p p’th quantile in F-distribution with (f1 , f2 ) degrees of
freedom, (i.e with probability mass p to the left of this
value)
182 Paper D
Symbol
tf (δ) Random variable distributed according to a noncen-
tral t-distribution with f degrees of freedom and non-
centrality parameter δ
t(f, δ)p p’th quantile in noncentral t-distribution with f degrees
of freedom and non-centrality parameter δ, (i.e with
probability mass p to the left of this value)
zp p’th quantile in standard normal distribution, (i.e with
probability mass p to the left of this value)
∆, ∆1 quantity serving to specify the limits for individual val-
ues (usually in the form of LC ± ∆ )
P3−c (µd , σd ) Acceptance probability under a 3-class attribute one-
stage sampling plan when mean and standard deviation
in the population are (µd , σd )
pnonc (µ, σ; ∆) Generic function (34) expressing the probability mass
outside the limits LC ± ∆ in a normal distribution with
mean µ and standard deviation σ.
Pusp (µd , σd ) probability of passing the USP-21 test when mean and
standard deviation in the batch is (µd , σd )
pa (µd , σd ) fraction of units outside limits LC ± ∆ when mean and
standard deviation in the batch is (µd , σd )
p+ (µd , σd ) fraction of units violating upper limit, LC + ∆ when
mean and standard deviation in the batch are (µd , σd )
p(µd , σd ) fraction of units violating lower limit, LC − ∆ when
mean and standard deviation in the batch are (µd , σd )
¯ sd )
A(d, Acceptance value used for determining acceptability,
see (80)
Paper E
On particle size distributions

and dosage uniformity for
low-dose tablets
183
185
1 Introduction and summary
In pharmaceutical production of tablets it is of natural concern that the active

ingredient is distributed uniformly among the individual units of the batch.
Therefore, during production various tests are performed to ensure uniformity
of the blend and uniformity of content in the final product.
The statistical properties of such tests of content uniformity are usually as-
sessed assuming a normal distribution of content in the tablets. However, ac-
tual distributions of particle sizes are often seen to be skewed, and it is there-
fore conceivable that this feature will also affect the shape of the distribution of
content in the dosage units. In this paper the particles referred to are the drug
particles in the blend.
In the paper we investigate the effect of relative variation, skewness and ex-
cess (heavy-tailedness) of the distribution of particle diameters on the result-
ing distribution of particle volume and particle mass under the assumption that
particles are spherical. For a log-normal distribution of particle diameters, the
resulting distribution of particle volume and mass is also a log-normal distribu-
tion with a mean that is larger than the mass corresponding to the mean radius,
and a coefficient of variation that is larger than the coefficient of variation for
the distribution of the radii. It is shown that this implies that the skewness
and excess of the distribution of particle radii is amplified when transformed
to the distribution of particle mass. The larger the coefficient of variation in
the distribution of particle radii, the more these departures from normality are
amplified.
Along with the variation in particle mass, an important source of dose varia-
tion for low-dose tablets is the variation in the number of drug particles in the
tablets. We investigate the transfer of the variation of particle sizes and of num-
ber of particles in the tablets to dose variation in tablets under the assumption
of a homogeneous blend with a random scattering of particles over the blend,
and derive expressions for skewness and excess of the distribution of tablet
doses. It is demonstrated (as previously shown by Yalkowsky and Bolton [64])
that for a given distribution of particle sizes the variance in the distribution of
186
absolute doses is proportional to the average number of drug particles in the

tablets, and moreover we show that the larger the average number of particles
in the tablets, the closer will the distribution of tablet doses be to a normal
distribution.
It is a common practice in the pharmaceutical industry to perform tests for con-

tent uniformity on measures of dispersion of relative doses (relative to label
claim) rather than on absolute doses. Translating the above results to relative
content it follows that the standard deviation of the relative content in tablets
decreases when the average number of particles in the tablets is increased.
Thus, for a given distribution of particle sizes in the blend, the standard devi-
ation of the relative content depends on the label claim (or target value). The
smaller the label claim, the larger is the relative standard deviation, and the
more pronounced is the departure from a normal distribution of relative doses.
In practical production, the assumptions of an ideal blend (random scattering

of drug particles) is not always satisfied, but some clustering of particles is con-
ceivable. We propose a simple and transparent hierarchical model for particle
distribution that reflects a varying intensity of particles over the blend. The
properties of the distribution of relative doses under this model are derived.
Although the results are presented in terms of content uniformity of tablets, the
statistical results are also valid for blend samples, under the assumption that
no systematic sampling error is introduced by the sampling process. The only
difference being that the size of individual blend samples usually is larger than
the size of the dosage units, and therefore in the interpretation of the results,
label claim shall be substituted by the target value of the blend sample.
2 Lognormal distribution of particle radii
Assume that the distribution of particle sizes may be adequately described by a

log-normal distribution. In pharmaceutical practice this distributional assump-
tion is generally considered to provide a good description of particle size data.
187
Furthermore, the use of the log-normal distribution to characterize particle size

data has the advantage that this assumption allows for explicit mathematical
expressions for the distribution of spherical particle volumes.
An implicit assumption underlying the use of the log-normal distribution model

is that the sieving/milling process has been satisfactory in the sense that no sin-
gle, large particles are present.
2.1 Distribution of particle radii
Let R denote the radius of a particle. The assumption of a log-normal dis-

tribution of particle radii is equivalent to assuming that the distribution of the
logarithm of the radii is a normal distribution.
The probability density function for a log-normal distribution is

!
1 1 ln(r) − αR 2
fR (r) = √ exp − (1)
ln(r) ln(βR ) 2π 2 β
where the parameters αR and βR are related to the so-called geometric mean,
µgR , and the geometric standard deviation by
µgR = exp(αR ) (2)

g
σR = exp(βR ). (3)
For the purpose of interpretation it is usually more convenient to work in terms

of the moments of the distribution,
2
µR = E[R] = exp(αR + (1/2)βR ) (4)
p q
σR = V[R] = µR × exp(βR 2 ) − 1. (5)
The coefficient of variation is

q
CR = σR /µR = 2 ) − 1.
exp(βR (6)
188
g
Hence, the parameter βR = ln(σR ) is in a one-to-one correspondence with the
coefficient of variation, CR
q
βR = log(1 + CR 2) (7)
1 2
αR = log µR − log(1 + CR ) (8)
2
and
2 2
exp(βR ) = 1 + CR . (9)
Expressed in terms of the mean, µR and the coefficient of variation, the mo-
ments about the mean are
µ2 (R) = V[R] = µ2R CR

2
(10)
3
µ3 (R) = E[(R − µR ) ] = µ3R 4
× CR × 2
(CR + 3) (11)
µ4 (R) = E[(R − µR )4 ]
h i
= µ4R × CR
4 2 4
× (1 + CR 2 3
) + 2(1 + CR 2 2
) + 3(1 + CR ) − 3(12)
.
Hence, the coefficients of skewness, γ1 , and excess, γ2 are (see e.g. [45])
3 2
γ1 (R) = µ3 (R)/σR = CR × (CR + 3) > 0 (13)
4
γ2 (R) = µ4 (R)/σR −3
2 4 2 3 2 2
= (1 + CR ) + 2(1 + CR ) + 3(1 + CR ) −6
2 2 4 6
= CR (16 + 15CR + 6CR + CR ) > 0. (14)
Thus, the distribution is positively skewed, and has heavier tails than the nor-
mal distribution (is leptokurtic).
2.2 Distribution of particle mass for spherical particles
Assuming that particles are spherical, the particle volume, V , is V = (4π/3)R3

and, hence the moments in the distribution of particle volume may be expressed
in terms of the moments in the distribution of particle radii, R
189
CR γ1 γ2
0.1 0.30 0.16
0.2 0.61 0.66
0.3 0.93 1.57
0.4 1.26 2.97
0.5 1.63 5.04
0.6 2.02 8.00

0.7 2.44 12.21
0.8 2.91 18.12
0.9 3.43 26.42
1.0 4.00 38.00
1.2 5.33 76.36

1.4 6.94 148.92
1.6 8.90 282.88
1.8 11.23 523.58
2.0 14.00 944.00
Table 1: Skewness, γ1 , and excess, γ2 , for log-normal distributions of particle

radii for different values of CR .
190

E[R3 ] = µ3R CR
4 2
(CR 2
+ 3) + 3CR +1 (15)
V[R3 ] = E[R6 ] − (E[R3 ])2 . (16)
For a log-normal distribution of particle radii, the distribution for the particle
volume, V = (4π/3)R3 , is also a log-normal distribution with parameters
αV = ln(4π/3) + 3αR (17)

βV = 3βR . (18)
Assuming that all drug particles have the same mass density, ρ, the distribution
of particle mass M = ρV is also log-normal with parameters
αM = ln(4π/3) + ln(ρ) + 3αR (19)

βM = 3βR . (20)
Hence the mean particle mass is

2
µM = E[M ] = exp(αM + (1/2)βM ) = (4π/3)ρµ3R (1 + CR
2 3
) (21)
and the coefficient of variation is

q q
2 2 )9 − 1
CM = exp(βM ) − 1 = (1 + CR (22)
with the relation

2 2 9
1 + CM = (1 + CR ) (23)
q
σM = µM × (1 + CR 2 )9 − 1. (24)
Utilizing the fact that the distribution of particle mass is also a log-normal
distribution, the third and fourth moment about the mean are found in analogy
with (11) and (12)
µ3 (M ) = E[(M − µM )3 ] = µ3M CM
4 2
(CM + 3) (25)
4
µ4 (M ) = E[(M − µM ) ]
h i
= µ4M CM
4 2 4
(1 + CM 2 3
) + 2(1 + CM 2 2
) + 3(1 + CM ) − 3 (. 26)
191
In analogy with (13) and (14) one obtains the coefficients of skewness and
excess for the distribution of particle mass
3 2
γ1 (M ) = µ3 (M )/σM = CM × (CM + 3) > 0 (27)
4
γ2 (M ) = µ4 (M )/σM −3
2 4 2 3 2 2
= (1 + CM ) + 2(1 + CM ) + 3(1 + CM ) −6
2 2 4 6
= CM (16 + 15CM + 6CM + CM )> 0. (28)
Inserting µM and CM from (21) and (22) the moments of the distribution of
particle mass may be expressed in terms of the parameters in the distribution
of particle radii as
µ3 (M ) = E[(M − µM )3 ]
= [(4π/3)ρµ3R ]3 (1 + CR
2 9 2 9
) [(1 + CR ) − 1]2 [(1 + CR
2 9
) + 2] (29)
µ4 (M ) = E[(M − µM )4 ]
= [(4π/3)ρµ3R ]4 [(1 + CR
2 9
) − 1]2 (1 + CR
2 12 2 9
) [(1 + CR ) − 1]2 ×
2 36 2 27 2 18
[(1 + CR ) + 2(1 + CR ) + 3(1 + CR ) − 3]. (30)
Further, the coefficients of skewness and excess for particle mass may be ex-
pressed in terms of the coefficient of variation, CR in the distribution of the
radii as
q
2 )9 2 9
γ1 (M ) = (1 + CR −1 (1 + CR ) +2 (31)
2 36 2 27 2 18
γ2 (M ) = (1 + CR ) + 2(1 + CR ) + 3(1 + CR ) − 6. (32)
Table 2.2 shows the coefficient of variation in the distribution of particle mass
and the corresponding coefficients of skewness and excess for different values
of the coefficient of variation in the distribution of particle radii. It is seen that
the relative variation in the distribution of the radii is amplified when consider-
ing the distribution of particle mass, resulting in a similar marked amplification
of the coefficients of skewness and excess.
192
CR CM γ1 γ2
0.1 0.31 0.95 1.64
0.2 0.65 2.23 9.95
0.3 1.08 4.52 50.89
0.4 1.67 9.72 356.55
0.5 2.54 24.00 4.07×103
0.6 3.86 69.20 7.30 ×104

0.7 5.93 226.6 1.82 ×106
0.8 9.21 808.8 5.55×107
0.9 14.40 3.03×103 1.91×109
1.0 22.61 1.16×104 6.90 ×1010
1.2 55.36 1.70×105 8.84×1013

1.4 132.07 2.30×106 9.26×1016
1.6 303.06 2.78 ×107 7.12×1019
1.8 665.50 2.95×108 3.85 ×1022
2.0 1 397.54 2.73 ×109 1.46 ×1025
Table 2: Coefficient of variation, CM , skewness, γ1 , and excess, γ2 , in log-

normal distributions of particle mass for different values of coefficient of vari-
ation, CR , in log normal distribution of particle radii.
193
3 Distribution of dose content under random mix-

ing of particles
3.1 Modelling random mixing
Most of the probabilistic models for mixtures in the literature discuss statis-
tical properties of samples from mixtures of two (or more) components when
sampling a fixed number of particles, or a fixed volume from the mixture, see
e.g. the survey by Gjelstrup Kristensen [13] and concern the relative distribu-
tion (by mass) of the two components in the sample. In these models the two
components enter symmetrically and the main issue is to describe the overdis-
persion compared to a binomial distribution of particles of the key component.
For low-dose tablets focus is upon the distribution of the small fraction of key
component particles in the blend, and this symmetry between key component
particles and other particles is not necessarily of concern. Therefore, in the
following we shall investigate the distribution of dose content in tablets un-
der various assumptions on the distribution of key component particles in the
blend. As the resulting variation in dose content in tablets is caused by varia-
tion of the number of drug particles in the sample as well as variation in particle
sizes, the discussion aims at separating the contribution from these two sources
of variation, thereby allowing for more transparent analytical formulae for the
variation.
3.2 Constant number of particles in tablets
The dose content of a tablet is the sum of the masses of the N drug particles in
that tablet
XN
D= Mi (33)
i=1
with Mi denoting the mass of the i’th particle.
If all tablets (or samples) contained the same number of drug particles, N , dose
194
variation would only be caused by the variation in the mass of the individual
drug particles selected by the tabletting process. Assume that particles are
sampled at random from the distribution (by number) of particle radii.
For later reference the moments in the distribution of doses under this simpli-
fying assumption are given below.
P
µD|N= E[ Ni=1 Mi ] = N × µM (34)
2 P 2
µ2 (D|N ) = σD|N = V[ N i=1 Mi ] = N × σM (35)
3 3
µ3 (D|N ) = E[(D − µD|N ) ] = E[(D − N µM ) ] = N µ3 (M ) (36)
µ4 (D|N ) = E[(D − µD|N )4 ] = N µ4 (M ) + 3N (N − 1)σM
4
(37)
where (37) is found from
κ4 (D|N ) = µ4 (D|N ) − 3(µ2 (D|N ))2
such that
µ4 (D|N ) = κ4 (D|N ) + 3(µ2 (D|N ))2
but
κ4 (D|N ) = N κ4 (M ) = N (µ4 (M ) − 3µ2 (M )2 )
and
µ2 (D|N ) = N µ2 (M )
such that
µ4 (D|N ) = N (µ4 (M ) − 3µ2 (M )2 ) + 3N 2 µ2 (M )2 (38)

2
= N µ4 (M ) + 3N (N − 1)µ2 (M ) . (39)
The coefficient of variation in the distribution of doses is found from (34) and
(35) as
σD|N 1
CD|N = = √ × CM (40)
µD|N N
195
and the coefficients of skewness and excess are

1
γ1 (D|N ) = √ × γ1 (M ) (41)
N
1
γ2 (D|N ) = × γ2 (M ). (42)
N
For a log-normal distribution of particle masses, γ1 (M ) and γ2 (M ) may be
expressed in terms of CM by (27) and (28), respectively.
Spherical particles
For spherical particles with a log-normal distribution of particle radii, the co-
efficient of variation in the distribution of doses may be expressed in terms of
the coefficient of variation, CR , in the distribution of radii using (22) such that
q
1 2 )9 − 1
CD|N = √ × (1 + CR (43)
N
and γ1 (M ) and γ2 (M ) are given by (31) and (32), respectively.
The expressions (41) and (42) show that the effect of skewness and excess
of the distribution of particle masses is diminished when particle masses are
added to form tablets. The larger the number of particles, the smaller the skew-
ness and excess, and the better will the distribution of doses be approximated
by a normal distribution. This is also a consequence of the central limit theo-
rem in probability theory.
3.3 Random variation of number of particles in tablets
In practice, the number of particles in a tablet exhibits variation. For low-

dose tablets the drug particles occupy only a small fraction of the blend vol-
ume, and therefore under perfect mixing the distribution of dose particles in the
three-dimensional space should share the properties of a random distribution
of points, i.e. a spatial Bernoulli process, which for practical purposes may be
approximated by a spatial Poisson process, see Stoyan et al [65].
196
When the intensity of drug particles is small, perfect mixing should imply
that the distribution of particle volume in a sample unit is independent of the
number of particles in the unit. For large particles, however, this assumption
is not necessarily satisfied, and more structured random measures need to be
invoked, see Stoyan et al [65].
In the present discussion we shall assume that the variation in the number of
particles in tablets may be described by independent samples from a Poisson
distribution with mean N , and that the distribution of particle sizes in a sam-
pled unit correspond to a random sample from the (number) distribution of
particle sizes.
Thus, the probability mass function for the number, N , of particles in a tablet
is n
N
fN (n) = exp − N . (44)
n!
It is a well-known property of the Poisson distribution (see e.g. Johnson and

Kotz, [66]) that the variance, V[N ], of the distribution is equal to the mean,
E[N ] = N .
For a specified absolute tablet dose, LC, and a drug with mean particle mass,
µM , the mean number of particles will be (assuming that the mean dose equals
label claim, LC)
N = LC/µM . (45)
Taking the variation of the number N of particles into account, and utilizing
that for the Poisson distribution V[N ] = N , the mean and variance of the dose
under an arbitrary distribution of particle masses are found as
µD = E[D] = EN [µD|N ] = EN [N µM ] = N µM (46)

2 2 2
σD = V[D] = EN [σD|N ] + VN [µD|N ] = EN [N σM ]+ VN [N µM ](47)
2
= N σM + µ2M V[N ] = N µ2M × (CM 2
+ 1) (48)
and hence the coefficient of variation in the distribution of doses is

q
1 2 .
CD = √ 1 + CM (49)
N
197
We note that the coefficient of variation, CD , of dose content under this as-
sumption of a random spatial distribution of particles mixing satisfies
q
CD = CD|N × 1 + 1/CM 2 (50)
with CD|N given by (40).
Hence, as in the case of a constant number of particles in all dosage units,

the√coefficient of variation decreases with the average number of particles as
1/ N . However, the variation of the number of particles in the individual
dosage units implies that the coefficient of variation of dose content is larger
than the coefficient of variation corresponding to a fixed number of particles in
each tablets (for the same average number of particles in tablets under the two
schemes). The smaller the coefficient of variation in the distribution of particle
mass, the more pronounced is the difference between the two schemes. For
large variation of particle masses, the variation of particle masses outweighs
the effect of the varying number of particles.
Expressed in terms of label claim (LC) one finds using (45)

2 2
σD = V[D] = LC × µM (CM + 1) (51)
s
µM (CM2 + 1)
CD = (52)
LC
which shows that under the assumption of random distribution of particles and
a given distribution of particle sizes does the absolute standard deviation in-
crease proportional to the square root of the magnitude of the dose (LC), and
hence the coefficient of variation decreases proportional to the square root of
the dose. The larger the dose, the smaller a coefficient of variation in the dis-
tribution of doses. This is in agreement with practical experience where it is
found that problems mainly occur for small dose tablets.
When relative doses, d = D/LC, are considered, the variance, σd2 , of the
distribution of relative doses is
µM
σd2 = V[D/LC] = CD 2
= (CM2
+ 1) (53)
LC
with the coefficient of variation, Cd , in the distribution of relative doses given
by (52) as Cd = CD .
198
This is in line with the general observation from theories on blend uniformity
that the standard deviation of relative content decreases as the square root of
sample weight (or volume).
The relation (53) has previously been established by Yalkowsky and Bolton
[64]. The relation is in agreement with the well known results on variance-
sample size relations for random binary mixtures. Noting that µM /LC = 1/N
the relation (53) simply states that the variance in the distribution of relative
doses is proportional to 1/N .
Spherical particles
Under the assumption of spherical particles and a log-normal distribution of

particle radii, (21) gives µM as a function of mean particle radius, µR , and
coefficient of variation, CR , in the distribution of particle radii
µM = (4π/3)ρµ3R × (1 + CR
2 3
) (54)
which inserted into (45) gives the average number of particles as function of
label claim and of the parameters in the distribution of particle radii
3 × LC
N= 2 )3 . (55)
4π × ρ × µ3R × (1 + CR
For spherical particles, µM and CM are given by (21) and (22),
and one obtains

µD = E[D] = N µM = N × (4π/3)ρµ3R × (1 + CR
2 3
) (56)
h i2
2
σD = V[D] = N × (4π/3)ρ × µ6R (1 + CR2 15
) (57)
σD 1 2 9/2
CD = = √ × (1 + CR ) (58)
µD N
or, in terms of label claim, using (45)
2 4πLC × ρ
σD = V[D] = × µ3R (1 + CR
2 12
) (59)
r 3
4πρ 3/2 2 6
CD = × µR (1 + CR ) . (60)
3LC
199
Coefficient of variation in distribution of doses, LC = 10, rho = 1.3 g/cc

0.0250.030.035
0 0.0050.010.0150.02
C_D
0.5
0.4
5
0.3
4
C_ 0 3
R .2
2 R
0.1 mu_
1
0
Fig. 1: Coefficient of variation, CD , in the distribution of doses as function of

mean particle radius, µR , and coefficient of variation, CR , in the distribution
of particle radii. LC is given in µg and particle radius is given in µm.
When relative doses, d = D/LC are considered, the variance, σd2 (53) under
the assumption of spherical particles may be expressed as
4π × ρ
σd2 = × µ3R (1 + CR
2 12
) . (61)
3LC
As an illustration, Figure 1 shows the coefficient of variation in the distribution

of doses as function of mean particle radius and coefficient of variation in
distribution of particle radii when LC = 10 mg and ρ = 1.3 g/cm3 . It is seen
that the coefficient of variation in the distribution of doses increases rapidly
with mean particle radius and coefficient of variation in the distribution of radii,
reflecting the decrease in average number of drug particles.
200

specified dose coefficient of variation
In practice, tests for demonstrating blend uniformity or content uniformity con-

tain requirements on the standard deviation of the content of the samples (blend
samples or dosage units) expressed relative to the target value (LC).
It is therefore of interest to assess the implications for the distribution of par-

ticle sizes of such requirements to blend or content uniformity. For simplicity
only a requirement on the coefficient of variation, CD will be considered.
For a specified maximum value, Cmax of CD , it follows from (52) that under
the assumption of perfect mixing, the mean particle mass, µM , shall satisfy
2
Cmax
µM ≤ LC × 2 (62)
CM + 1
with CM denoting the coefficient of variation in the distribution of particle

mass.
Thus, the limit for acceptable particle masses increases proportional to LC, and
decreases proportional to (1 + CM 2 )−1 as C
M increases.
Under the assumption of spherical particles, the limit may be expressed in

terms of the coefficient of variation, CR , for the distribution of particle radii
using (60)
[C 2 × 3LC/(4πρ)]1/3
µR ≤ max 2 )4 . (63)
(1 + CR
Thus, the limit for acceptable particle radii increases proportional to the third
root of LC, and decreases with increasing coefficient of variation, CR , for the
2 )−4 .
distribution of particle radii, the increase being proportional to (1 + CR
Reducing the label claim by 50% implies that the maximum allowable average
particle radius is reduced by 20% (if the coefficient of variation in the distribu-
tion of particle radii remains unchanged). Reducing the label claim to a tenth
implies that the maximum allowable average particle radius is reduced by 50%.
201
Maximum particle radius to secure dose coeff.of var. less than

0.05, rho = 1.3 g/cc
20
15
max r
105
0
25
20
0.8
15 0.7
LC 0.6
10 0.5 C_r
0.4
5
Fig. 2: Maximum mean particle radius necessary to secure a coefficient of

variation in the distribution of dosage values less than Cmax = 0.05 shown as
function of label claim and coefficient of variation in distribution of particle
radii. LC is given in µg and particle radius is given in µm.
Thus, a blend that might produce a satisfactory distribution of doses (in terms
of the coefficient of variation in the distribution of relative doses) for large dose
tablets need not be satisfactory for smaller dose tablets.
Figure 2 shows the limit (63) for particle radii necessary to assure that the
coefficient of variation in the distribution of doses does not exceed Cmax =
0.05. The limit is shown as function of label claim, LC, and coefficient of
variation, CR , in the distribution of particle radii for ρ = 1.3 × 10−6 µg/µm3 .
It is seen that the maximal radius increases rather slowly with the label claim
(proportional to the third root of label claim).
The requirement (62) to the mean drug particle size may be expressed as a
202
requirement to the average number of particles in a dosage unit:

2 )
(1 + CM 2 )9
(1 + CR
N≥ 2
= 2
. (64)
Cmax Cmax
Thus, for low-dose tablets, assuming perfect mixing, and a log-normal distri-
bution of particle sizes, the requirement on the coefficient of variation in the
distribution of dosage units is essentially a general requirement on the mini-
mum average number of particles in a dosage unit. This minimum average
number of particles does not depend on label claim.
3.5 Coefficient of skewness and excess for distribution of

dose content
In order to assess the validity of the approximation of the distribution of dose

content by a normal distribution, the higher moments, i.e. the coefficients of
skewness and excess of the distribution of dose content under perfect mixing
are of interest.
2 , in the distribution of dose content was found as

In (48), the variance, σD
2
σD = N µ2M × (CM
2
+ 1).
In order to determine the coefficients of skewness and excess consider
κ3 (D) = E[(D − µD )3 ] = EN [E[(D − N µM )3 |N ]] (65)

4 4
κ4 (D) = E[(D − µD ) ] − 3σD
= EN [E[(D − N µM )4 ]|N ] 4
− 3σD . (66)
But
(D − N µM )3 = (D − N µM )3 + 3(D − N µM )2 (N − N )µM
+3(D − N µM )(N − N )2 µ2M + (N − N )3 µ3M (67)
203
and therefore
E[(D − N µM )3 |N ] = µ3 (D|N ) + 3µ2 (D|N )(N − N )µM + 0

+(N − N)3 µ3M
2
= N µ3 (M ) + 3(N − N )µM N σM
+(N − N)3 µ3M (68)
EN [E[(D − N µM )3 ]|N ] = N µ3 (M ) + V[N ]µM σM
2
+ µ3 (N )µ3M .(69)
For the Poisson-distribution V[N ] = µ3 (N ) = N , and hence

κ3 (D) = EN [E[(D − N µM )3 ]|N ] = N µ3 (M ) + µM σM
2
+ µ3M . (70)
Thus, from (70)

2 + µ3
N µ3 (M ) + µM σM
κ3 (D) M
γ1 (D) = 3 = 3/2 3
σD N µM (CM2 + 1)3/2
µ3 (M ) + µM σM2 + µ3
1
= √ × M
.
N µ3M (CM2 + 1)3/2
Using (25) we then find
1 C 4 × (CM
2 + 3)
γ1 (D) = √ × M 2 (71)
N (CM + 1)3/2
which shows that also in this case, the skewness tends to zero when the number
of particles increase.
In order to assess µ4 (D) consider
(D − N µM )4 = [(D − N µM ) + (N − N )µM ]4
= (D − N µM )4 + 4(D − N µM )3 (N − N )µM
+6(D − N µM )2 (N − N )2 µ2M
+4(D − N µM )(N − N )3 µ3M + (N − N )4 µ4M
204
such that
E[(D − N µM )4 ]|N = µ4 (D|N ) + 4µ3 (D|N )(N − N)µM

+6µ2 (D|N )(N − N )2 µ2M + 0
+µ4 (N )µ4M
= N µ4 (M ) + 3N (N − 1)µ2 (M )2 + 4N (N − N )µM µ3 (M )
2
+6N × σM (N − N )2 µ2M + (N − N )4 µ4M
and hence

µ4 (D) = EN E[(D − N µM )4 ]|N

= Nµ4 (M ) + 3 V[N ] + N (N − 1) µ2 (M )2 + 4V[N ]µM µ3 (M )

2 2
+6 µ3 (N ) + N V[N ] σM µM + µ4 (N )µ4M .
2
For the Poisson distribution we have µ4 (N ) = 3N + N , and therefore
2
µ4 (D) = N µ4 (M ) + 3N µ2 (M )2 + 4N µM µ3 (M )

2 2 2
+6N (N + 1)σM µM + 3N + N µ4M .
Using (25) and (26) we obtain

h
µ4 (D) = µ4M N CM 4 2 4
(1 + CM 2 3
) + 2(1 + CM 2 2
) + 3(1 + CM ) −3
4 4 2
+3N CM + 4CM (CM + 3)
i
2
+6(N + 1)CM + 3N + 1
2
= 3µ4M N (1 + CM 2 2
)
h
+µ4M N CM 4 2 4
(1 + CM 2 3
) + 2(1 + CM 2 2
) + 3(1 + CM ) −3
i
4 2 2
+4CM (CM + 3) + 6CM +1
such that we obtain

C 4 (1 + C 2 )4 + 2(1 + C 2 )3 + 3(1 + C 2 )2 − 3 + 4C 4 (C 2 + 3) + 6C 2 + 1
1 M M M M M M M
γ2 (D) = × 2 2
N (CM + 1)
(72)
205
which shows that the excess of the distribution of the doses tends to zero when
the number of particles increase.
Thus, the effect of the deviation from the normal distribution of particle sizes
becomes less pronounced the larger the average number of particles.
Invoking (45) to express the average number of particles in terms of the label
claim, (71) may be expressed as
r
µM C 4 × (CM2 + 3)
γ1 (D) = × M 2 (73)
LC (CM + 1)3/2
and similarly, (72) yields

C 4 (1 + C 2 )4 + 2(1 + C 2 )3 + 3(1 + C 2 )2 − 3 + 4C 4 (C 2 + 3) + 6C 2 + 1
µM M M M M M M M
γ2 (D) = × 2 2
LC (CM + 1)
(74)
4 Distribution of dose content under non-random

mixing
We still consider the model (33) for the dose, D.
However, instead of the random dispersion of dose particles over the blend
modelled by a distribution of the number of particles in the blend as a homoge-
neous Poisson point process in the three-dimensional space, we shall consider
a model for a nonhomogeneous blend.
The homogeneous Poisson point process model is naturally extended to a Pois-

son point process where the intensity is varying over the blend. One such
extension is the doubly stochastic Poisson Process, or Cox Process, see e.g.
Stoyan et al [65]. Such a process can be thought of as arising from a two-step
random mechanism. The first step generates a random intensity measure over
the blend, and the second step generates a Poisson process corresponding to
that intensity measure.
206
Let Λ(·) denote the random intensity measure defined over the blend, and let
Z
Λv = Λ(u)du
u∈v
denote the random volume-intensity corresponding to a volume v. Analo-

gously let λ(·) denote a realization of this measure with
Z
λv = λ(u)du
v
denoting the realized volume-intensity in a volume v.
When a sample of a fixed volume, v, is selected from the blend, the sampling
process may similarly be thought of as a two-step random mechanism. First,
a random intensity Λv corresponding to that volume is selected from a distri-
bution of volume-intensities, and secondly a number of particles is generated
as a Poisson distributed random variable with mean λv corresponding to that
intensity.
In order to obtain a mathematically tractable form of the resulting distributions,

it will be assumed that the doubly stochastic Poisson process is such that the
distribution of Λv under repeated sampling from the blend may be described
by a gamma distribution with mean
E[Λv ] = v × ν (75)
and with variance
V[Λv ] = E[Λv ] × γv2 = v × ν × γv2 (76)
where ν denotes the overall average number of particles in the blend per unit
volume ([µm3 ]) and where the measure, γv , of the clustering tendency is re-
lated to the coefficient of variation in the distribution of λv ,
p
V[Λv ] γv
CΛ = =√ . (77)
E[Λv ] v×ν
The suffix v on γv2 serves to emphasize that the variance in the distribution of
Λv may depend on the sample amount v. Moreover, γv2 depends on the second
207
order (spatial correlation) properties of the random intensity measure Λ(·):

one extreme being complete independence between intensities in infinitesimal
neighbouring volumes in which case γv2 does not depend on the sample amount
v, i.e. γv2 = c. Under a perfect random mixing, i.e. a homogeneous Poisson
process with (constant) intensity ν one has γv2 = 0.
It will moreover be assumed that the spatial correlation corresponding to the

intensity measure Λ(·) fades away so rapidly that the volume-intensities, Λv ,
corresponding to independently sampled volumes, v, may be approximated by
independent samples from the above distribution.
Formally, it is assumed that the conditional distribution of the number, N , of

particles in a random volume, v, given the volume intensity Λv = λv corre-
sponding to that volume may be described by a Poisson distribution with mean
λv , and that the volume-intensities corresponding to the volumes sampled may
be described by independent identically gamma-distributed variables Λv with
mean (75) and variance (76).
Under this assumption it may be shown (see e.g. Johnson and Kotz [66]) that
the marginal distribution of the number of particles in a sample of volume v is
a negative binomial distribution with mean and variance
N = E[N ] = v × ν (78)
V[N ] = EΛ [V[N |Λv ]] + VΛ [E[N |Λv ]] (79)
= E[Λv ] + V[Λv ] = vν(1 + γv2 ) (80)
= E[N ] × (1 + γv2 ) (81)
i.e. γv2 expresses the overdispersion of counts as compared with the Poisson
distribution. For γv = 0 the negative binomial distribution of particle counts
degenerates to a Poisson distribution with mean N = vν.
The intraclustercorrelation in volumes of size v is

γv2 1
ρv = 2
= . (82)
1 + γv 1 + 1/γv2
Assuming further that particle sizes vary independent of the volume-intensity,

i.e. particles in clusters are selected independently from the distribution of
208
particle sizes, then the moments in the conditional distribution of doses given
the number of particles in the tablet are given by (34) to (37).
It then follows from (46) and (47), that the (marginal) mean and variance in
the distribution of doses are
µD = E[D] = N µM = v × ν × µM (83)
2 2
σD = V[D] = N σM + µ2M V[N ] (84)
2
= N × (σM + µ2M (1 + γv2 ))
= N µ2M × (CM
2
+ 1 + γv2 ) (85)
and
q
1 2 + γ 2.
CD = √ 1 + CM v (86)
N
Under the homogeneity (Poisson) assumption we have V[N ] = E[N ] = N ,

and in that case the expression for the variance simplifies to (48).
In the general case the coefficient of variation in the distribution of doses is

q
σD 1 2 + γ2
CD = = √ × 1 + CM v (87)
µD N
(with (49) as a special case for γv = 0).
Comparing (87) and (49) for a given number distribution of particle masses,
it is seen that the overdispersion of the distribution of the number of particles
under imperfect mixing implies an overdispersion of doses as compared to the
distribution of doses under perfect mixing (Poisson distribution of number of
particles).

2 2
σD = V[D] = LC × µM (CM + 1 + γv2 ) (88)
s
µM (CM2 + 1 + γ2)
v
CD = . (89)
LC
209
Consequently, when relative doses, d = D/LC are considered, the variance,

σd2 of the distribution of relative doses is
µM
σd2 = V[D/LC] = CD
2
= 2
(CM + 1 + γv2 ) (90)
LC
4.1 Spherical particles
For spherical particles, µM and CM are given by (21) and (22) and one obtains

2
σD = = N µ2M × (CM 2
+ 1 + γv2 ) = N µ2M × (1 + CR 2 9
) + γv2 (91)
h i2
= N × (4π/3)ρµ3R × (1 + CR 2 3
) × (1 + CR2 9
) + γv2 . (92)
The coefficient of variation in the distribution of doses is

q
σD 1 2 )9 + γ 2
CD = = √ × (1 + CR v (93)
µD N
(with (58) as a special case for γ = 0).
Comparing (93), (58) and (40) for a given number distribution of particle radii,
it is seen that the overdispersion of the distribution of the number of particles
(as compared to the Poisson distribution) under imperfect mixing implies an
overdispersion of doses as compared to the distribution of doses under perfect
mixing.
2 4πLC × ρ
σD = V[D] = × µ3R (1 + CR
2 3
) (1 + CR 2 9
) + γv2 (94)
s 3
4πρµ3R (1 + CR
2 )3
2 )9 + γ 2

CD = × (1 + CR (95)
3LC v
which shows that under the assumption of perfect mixing and a given distribu-
tion of particle sizes does the absolute standard deviation increase proportional
210
to the square root of the magnitude of the dose (LC), and hence the coefficient
of variation decreases proportional to the square root of the dose. The larger
the dose, the smaller a coefficient of variation in the distribution of doses.
Consequently, when relative doses, d = D/LC are considered, the variance,

σd2 of the distribution of relative doses is
4π × ρ
σd2 = V[D/LC] = × µ3R (1 + CR2 3
) (1 + CR 2 9
) + γv2 (96)
3LC

specified dose coefficient of variation
In analogy with Section 3.4 consider a requirement CD ≤ Cmax . Under the

model for nonrandom mixing, the resulting requirement on the average number
of particles is found from (93)
2 )9 + γ 2
(1 + CR
N≥ 2
v
. (97)
Cmax
In analogy with (63) the the requirement CD ≤ Cmax may be expressed as a

requirement to the mean particle radius (for given LC, CR and γv2 ) as
2
[Cmax × 3LC/(4πρ)]1/3
µR ≤ 2 ) × [(1 + C 2 )9 + γ 2 ]1/3
. (98)
(1 + CR R v
Thus, also in this case the limit for acceptable particle radii increases pro-
portional to the third root of LC, and decreases with increasing coefficient of
variation, CR , for the distribution of particle radii, although in a less simple
manner as in (63).
5 Discussion
In production of tablets, variation is introduced at various stages in the process:

211
sieving/milling process: the purpose of the sieving/milling of the drug parti-

cles is to reduce the particles to an adequate size. Natural variation in
process output is described by a unimodal distribution, often a normal,
or a log normal distribution.
blending/mixing process: a purpose of the blending/mixing is to obtain a uni-

form mixture of drug particles with other components. Natural variation
in a satisfactory blend might be modelled by a random spatial distribu-
tion of drug particles.
transfer to tablet press and tabletting process: ideally, the transfer aims at
transferring the spatial distribution in the blend at tablet size level.
In the paper we have investigated the transfer of variation resulting from the
sieving/milling process through the mixing process to samples from the blend,
or to tablets in ideal situations where only natural variation is present, i.e.
no sampling bias, no deblending etc., (only some clustering tendencies in the
blend have been considered).
In Section 2 properties derived from a lognormal distribution of particle radii

were discussed with the aim of assessing the departure from normality of the
resulting distribution of particle mass. It was found that the relative variation
(the coefficient of variation) in the distribution of radii is amplified when con-
sidering the distribution of particle mass. As the coefficient of variation in a
lognormal distribution is determinant for the departure from normality (skew-
ness and excess), the departure from normality of the distribution of particle
mass is heavily dependent on the magnitude of the coefficient of variation in
the distribution of radii.
The variation in dose content of tablets is caused by variation in the number

of dose particles in the tablets as well as variation in the mass of individual
particles. In Section 3.2 the contribution from the variation in particle mass has
been investigated keeping the number of particles fixed. Expressions for the
moments in the distribution of dose content have been derived and expressed
in terms of label claim, mean particle size and coefficient of variation in the
distribution of particle radii. Alternative expressions in terms of number of
particles and coefficient of variation in the distribution of particle radii have
212
been provided. It was found that the effect of skewness and excess in the
distribution of particle masses is diminished when particle masses are added to
form tablets. The larger the number of particles, the smaller the skewness and
excess in the distribution of doses, and the better will the distribution of doses
be approximated by a normal distribution.
Moreover, it was found that the coefficient of variation in the distribution of

dose content (and of relative content) is inversely proportional to the square
root of the number of particles. Thus, the smaller the number of particles,
the larger a variance in the distribution of relative doses. Therefore, in order
for low dose tablets to comply with usual dosage uniformity test, it is impera-
tive that particle diameter is small, and it is furthermore advantageous that the
coefficient of variation in the distribution of particle sizes is suitably small.
In Section 3.3 the further variation introduced by the variation of the number
of particles in the tablets has been investigated under the assumption of a ho-
mogeneous blend, i.e. a random spatial distribution of dose particles in the
blend. Expressions for the moments in the distribution of dose content have
been provided. As could be expected, the variation of the number of particles
amplifies the variation in dose due to varying particle sizes. However, unless
the distribution of particle masses is very narrow, the major contribution to the
variation in dose content originates from the variation in particle masses, and
hence, also in this case the coefficient of variation in the distribution of dose
content is inversely proportional to the square root of the number of particles.
Thus, assuming perfect mixing, and a log-normal distribution of particle sizes,

the requirement on the coefficient of variation in the distribution of dosage
units is essentially a general requirement on the minimum average number of
particles in a dosage unit. This minimum average number of particles does
not depend on label claim. However, as the average number of particles in
tablets depend on label claim (for a given particle size distribution), a blend that
might produce a satisfactory distribution of doses (in terms of the coefficient
of variation in the distribution of relative doses) for large dose tablets need not
be satisfactory for smaller dose tablets.
213
In Section 3.4, explicit expressions for the minimum average number of par-
ticles per tablets necessary to secure a specified coefficient of variation in the
distribution of doses have been provided under the assumption of a homoge-
neous blend, and in Section 3.5 expressions for coefficients of skewness and
excess for the distribution have been provided.
Finally, in Section 4, the results are extended to cover also such nonrandom
mixtures that may be modelled by a hierarchical model representing a spatial
variation of the intensity of particles.
Results have been derived under the ideal assumption of spherical particles.
Therefore, for practical use, a “nonsphericity factor” might be introduced in
the transformation from distribution of particle radii to particle mass and vice
versa. In the paper we have allowed for this possibility by providing separate
expressions for moments in the distribution of doses in terms of properties of
the distribution of particle masses, valid regardless of shapes, supplemented
by expressions in terms of properties of the distribution of particle radii, valid
only under the further assumption of spherical particles.
Interpretating the results in terms of blend samples rather than samples of

tablets from a batch, it is of interest to note that the practical necessity of using
blend samples that are larger than the dosage units imply that the coefficient
of variation in such blend samples is smaller than the coefficient of variation
in the resulting dosage units. For blend samples that are four times the size of
the final dosage units, the coefficient of variation in the blend samples is only
half the size of the coefficient of variation in the final dosage units. Moreover,
a larger blend sample might mask departure from normality in the distribution
of dose content in low-dose tablets.
214
6 List of symbols
Symbol Units
E[X] Mean of the distribution of X
V[X] Variance of the distribution of X
EN [X] Mean (under variation of N ) of the distribution of X
VN [X] Variance (under variation of N ) of the distribution of X
µ3 (X) Third moment about the mean in the distribution of X
µ3 (X) = E[(X − E[X])3 ]
µ4 (X) Fourth moment about the mean in the distribution of X
µ4 (X) = E[(X − E[X])4 ]
γ1 (X) - Coefficient of skewness of the distribution of X (fraction)
- γ1 (X) = µ3 (X)/(V[X])3/2
γ2 (X) - Coefficient of excess of the distribution of X (fraction)
- γ2 (X) = µ4 (X)/(V[X])2 − 3
R µm Radius of a particle
µR µm Mean particle radius
σR µm Standard deviation in particle radius distribution
CR - Coefficient of variation in particle radius distribution (fraction)
CR = σR /µR
βR Standard deviation in distribution of log particle radius
αR log(Median) in particle radius distribution
V µm3 Particle volume
ρ µg/µm3 Mass density of particles
M µg Particle mass
µM µg Mean particle mass
σM µg Standard deviation in particle mass distribution
CM - Coefficient of variation in particle mass distribution (fraction)
CM = σM /µM
N Number of particles in tablet
N Average number of particles in tablet
D µg Dose (Total mass of drug in tablet)
P
D= N i=1 Mi
µD µg Mean dose
σD µg Standard deviation in dose distribution
CD - Coefficient of variation in dose distribution (fraction)
CD = σD /µD
LC µg Required dose, Label Claim
d - Relative dose (fraction)
d = D/LC
µd - Mean relative dose
σd - Standard deviation in distribution of relative doses
215
216
Paper F
Case: Analysis of
homogeneity in production
scale batches
217 F
219
1 Purpose
This case relates to the sampling from a low dose (25µg) product. An experi-
mental design was set up with the purpose to investigate:
1. whether the type of thief shown to the left in Figure 2.3 in Chapter 2 is
qualified as sampling device for the product. Tablet samples from the
same batch are used to evaluate this problem.
2. bias (difference in mean content) and repeatability of samples sampled

by respectively different thieves, persons and thief tips.
(a) whether samples taken with respectively a long and a short thief
lead to the same result (it might be more difficult to handle a long
thief, or the samples taken with the long thief may be biased be-
cause they are primarily sampled from the bottom of the batch).
(b) whether the size (1X or 3X unit dose, 80 mg) influences the result.
(c) whether the sampling procedure is sensitive of the person collect-
ing the sample.
3. the homogeneity of each batch.
(a) whether the three layers in a batch lead to equally representative

samples.
2 Experimental
The original experimental design included twenty-six blend samples, fifty tablet
samples and a larger blend sample for sieving analysis from each of three
batches of the product.
However, blend samples were taken from the first two batches, but the third
batch was transferred to the tablet press before the blend samples were taken.
220 Paper F
From each of the first two batches ten tablet samples were collected after one
hour of tabletting.
In total 26 samples were collected from each of two batches as shown in Ta-
ble 1.
Nr. Batch Layer Area Theif Pers. Nr. Batch Layer Area Theif Pers.
1 1 Top C 1 X 1 2 Top A 1 Y
2 1 Top C 1 X 2 2 Top A 1 Y
3 1 Top B 1 X 3 2 Top C 1 Y
4 1 Top B 1 Y 4 2 Top B 1 X
5 1 Top A 1 X 5 2 Top D 1 X
6 1 Top D 1 Y 6 2 Top D 1 Y
7 1 Middle B 2 X 7 2 Middle D 3 Y
10 1 Middle C 3 Y 10 2 Middle A 3 X
13 1 Middle D 3 Y 13 2 Middle B 2 Y
16 1 Middle A 1 X 16 2 Middle C 1 X
19 1 Bottom D 3 Y 19 2 Bottom B 2 Y
20 1 Bottom C 3 Y 20 2 Bottom B 2 Y
21 1 Bottom C 2 Y 21 2 Bottom B 3 Y
22 1 Bottom A 2 X 22 2 Bottom B 3 Y
23 1 Bottom A 2 X 23 2 Bottom A 3 X
24 1 Bottom A 3 X 24 2 Bottom D 2 Y
25 1 Bottom A 3 X 25 2 Bottom C 2 X
26 1 Bottom B 2 X 26 2 Bottom C 3 X
Table 1: The table shows the samples collected from two blend batches.
Samples were collected from three layers, and four areas within each layer by
two persons using three different sampling thieves.
Regarding areas, four areas are identified in each layer as shown in Figure 1. In
the middle layer the orientation of the areas are rotated 60 degrees compared to
the top and the bottom layer. If more than one sample is taken in the same area
221
and layer, the samples are sampled as close as possible - but not from exactly
the same position. Estimates of the small scale variation in the blend are based
on these samples.
B D
Fig. 1: A sectional plane of a layer in the blend. Each layer contains four
areas: A, B, C and D.
The three different thieves used for sampling are:
• Thief 1 is a short thief with a 1X dose tip.
• Thief 2 is a long thief with a 1X dose tip.
• Thief 3 is a long thief with a 3X dose tip.
The same two persons are sampling in the two batches. Person X is a person
with above average skills in sampling techniques.
Because of several physical restrictions the blend samples are not collected in
a completely randomized order.
3 Statistical Analysis
The analyses are performed in SAS using procedures for Generalized Linear
Models (GENMOD) and General Linear Mixed Models (MIXED).
222 Paper F
The actual collected blend samples correspond to a design with confounding

between factors. Further only tablets corresponding to the bottom of the blend
have been collected. The results of the statistical analysis is therefore only a
guide of the conditions in these two batches.
The statistical analysis is performed on the relative dose, i.e. content of the
active component divided by label claim.
The mean relative content is 0.912 in blend samples from batch 1 and 1.016
in blend samples from batch 2. The relative dose in the individual samples are
shown in Figure 2.
Batch 1 Batch 2
1.15
1.15
1.10
1.10
1.05
1.05
Relative Dose
Relative Dose
1.00
1.00
0.95
0.95
0.90
0.90
0.85
0.85
Bottom Middle Top Tablets Bottom Middle Top Tablets

Layer Layer
Fig. 2: Results from all blend- and tablet samples.
3.1 Assessment of tablet samples
Summary results for the tablets are given in Table 2.
Figure 2 indicates that the relative content in samples from batch 1 is less
than the relative content in samples from batch 2. A Welch modified two-
223
Dose (µg) Relative Dose

X̄ S2 X̄ S2 Conf. Int. S 2
Batch 1 23.18 0.0773 0.9272 0.0001 [0.0001;0.0004]
Batch 2 26.02 0.4684 1.0408 0.0007 [0.0004;0.0025]
Table 2: Mean, variance and 95% confidence interval for the dose and the
relative dose for tablet samples.
sample t-test on the tablet samples shows a significant difference between the
two samples. The plot also reveals that in general the variation in batch 2 is
larger than the variation in batch 1.
Figure 3 shows QQ-plot for the tablets from each batch. The content of active
component in the tablets may be considered to be normally distributed, with
mean and variance depending on the batch from which the tablets are sampled.
Tablets from Batch 1 Tablets from Batch 2

1.10
1.10
1.05
1.05
Relative Dose
Relative Dose
1.00
1.00
0.95
0.95
0.90
0.90
-1 0 1 -1 0 1
Quantiles of Standard Normal Quantiles of Standard Normal
Fig. 3: QQ-plot for ten tablets sampled from each of the two batches. The
tablets were sampled when the tablet press had been running for a period of 1
hour. The Normal distribution gives a satisfactory fit of the data.
An F-test shows that the variation between tablets from batch 2 is significantly
larger than the variation between the tablets from batch 1. This indicates that
224 Paper F
batch 1 is more homogeneous i.e. that the small scale variation in batch 1 is
less than the small scale variation in batch 2.
3.2 Repeatability
The repeatability variance is the variance between replicate samples from the
same batch, layer and area by the same person and sampling thief. The re-
peatability variance incorporates contributions from small scale variation in
the blend, variation due to sampling and due to the chemical analysis. Further,
as blend samples in contrast to tablet samples are adjusted for the weight of
the sample, uncertainty on the weight measurement also has an influence on
the repeatability variance between blend samples.
With 6 pairs of replicates in each batch it has been possible to assess the effect
of different sampling thieves and layers on the repeatability variance. The
number of actual collected samples is too small to include the effect of persons
in this statistical analysis. Figure 4 indicates that the overall variation between
samples collected by each person do not differ, indicating that the persons are
equally consistent in the way they take a sample.
The statistical model for the repeatability variance is a Generalized Linear

Model:
2
ln σlayer,thief = µ∗ + α∗layer + βthief
∗ ,
2
where Slayer,thief 2
∈ σlayer,thief χ2 (1).
The results are discussed below.
Thieves
The statistical analysis shows that use of different sampling thieves have no
significant influence on the repeatability variance. This means that the repeata-
bility variance introduced by the three thieves do not differ significantly. How-
225
Batch 1 Batch 2
1.15
1.15
Top Top
Midt Midt
Bund Bund
1.10
1.10
1.05
1.05
Relative Dose
Relative Dose
1.00
1.00
0.95
0.95
0.90
0.90
0.85
0.85
X Y Tablets X Y Tablets
Person Person
Fig. 4: Variation between replicate samples collected by each of the two per-
sons. This variation includes variation due to variations in the blend, varia-
tion among sampling thieves, analytical error and variation from adjusting the
blend samples for the weight. Results from the tablet samples are included as
reference in the plot.
226 Paper F
ever, the tendency for both batch 1 and batch 2 is that using sampling thief
1 introduces less sampling error than using thief 3. As seen in Table 3 the
estimated standard deviation between samples collected by thief 3 is approx-
imative four to five times as big as the estimated standard deviation between
samples collected by thief 1. It is conceivable that an analysis of the results
from an experimental design with more experiments will find the tendency
significant.
Thief Batch 1 Batch 2

Middle layer Middle layer
1 (0.0028)2 (0.0170)2
2 (0.0064)2 (0.0495)2
3 (0.0103)2 (0.0831)2
Table 3: Variation between samples collected with various thieves in the mid-
dle layer in batch 1 and batch 2.
Layers
The repeatability variance is not found to differ significantly in the three layers
in a batch. However for both batches the tendency is that the variation between
replicates is less in the top, than in the bottom layer. This can be interpreted
as the batches not being homogeneous on the small scale level. It is conceiv-
able that the tendency will be significant in an experimental design with more
observations.
Comparing variation in the blend with variation between tablets
Ten tablets from each batch were sampled as closely as possible in time after
one hour of tabletting, i.e. at the beginning of the tabletting process. The
variation between the tablets from a batch includes the variation in the blend,
the variation introduced by the tablet press (deblending and weight variation)
and the chemical analysis. However, it is assumed, that the weight variation
and the variation due to deblending introduced at the tablet press are negligible
and therefore the variation between tablets can be regarded as an estimate of
227
the small scale variation in the blend, i.e. a measure of the minimum obtainable
variation between samples from the blend. It is therefore relevant to compare
the variation among tablets to variations in the blends.
The problem is to decide which variations in the blends should be compared to

the variation between tablets.
Even though not statistically significant the tendency is that the variation be-
tween replicates in the top layer is smaller than the variation between replicates
in the bottom layer. Therefore, comparing variation between tablets to varia-
tion between replicates in respectively the top and the bottom layer may not
lead to the same result regarding an estimate of the small scale variation in the
blend.
To assess whether an estimate of the small scale variation in the blend should
be based on the top or the bottom layer, assumptions have to be made about a
physically explanation for the tendency for the variation between replicates to
be smaller in the top layer than in the bottom.
One explanation is that the real small scale variation in the blend differs among
layers. Another explanation is that it is more difficult to hit the same spot in a
lower layer of a blend when sampling replicates. It is not possible due to the
experimental design to decide which explanation is correct.
The two physically explanations lead to the following estimates of the small
scale variation in the blend:
1. Difficulties with hitting the same spot is the main reason for the ten-
dency of the variation between replicates being smaller in the top
than in the bottom of a batch. The best estimate of the small scale
variation in the blend is an estimate based on replicates sampled with
the thief that introduces the smallest sampling variance, i.e thief 1. The
estimate is based on the top layer because it is easier to hit the same
spot in this layer than in the middle and the bottom layer. A 95% con-
2
fidence interval for the small scale variation, σsmall scale , in batch 1 is
[0.0000;0.0001] and [0.0000;0.0005] for batch 2.
228 Paper F
2. The real small scale variation is not constant among the three layers
in the batch. As the tablets are sampled at the beginning of the tabletting
process, they correspond to the bottom layer of the blend. Therefore the
best estimate of the small scale variation in the blend corresponding to
the variation among the collected tablets is based on the replicates sam-
pled with thief 2 in the bottom of the blend. Thief 2 is chosen because
it is the thief used in the bottom layer that introduces the smallest sam-
pling variance. A 95% confidence interval for the small scale variation
in the bottom layer is [0.0001;0.0063] for batch 1 and [0.0004;0.0272]
for batch 2.
95% confidence interval for the small scale variation based on the tablets are
[0.0001;0.0004] for batch 1 and [0.0004;0.0025] for batch 2. Confidence in-
tervals for the small scale variation based on the tablets are more overlapping
to the confidence interval corresponding to situation 2 than the confidence in-
terval corresponding to situation 1.
If the small scale variation estimated from the tablets represents the true small
scale variation in the blend, the small scale variation estimated from blend
samples must be either of the same size or (when variation due to sampling
is not negligible) larger. However, in situation 1 the estimate based on blend
samples is smaller than the estimate based on tablet samples, thus implying
that the true small scale variation in the bottom layer actually is larger than in
the top layer.
3.3 Mean content of the active component
The statistical model for the relative content in a sample from a given batch is
ylayer,area,person,thief = µ+αlayer +βthief +γperson +Darea +Elayer,area,person,thief ,
2
where Darea ∈ N (0, σarea 2 ).
) and Elayer,area,person,thief ∈ N (0, σE
The model is a General Linear Mixed Model. Under this model it has been
tested whether the mean content of the active component differs in the three
229
layers in a blend and whether the mean content in a sample depends on the
thief and the person collecting the sample.
The residual variance from this analysis is an estimate of the variance between
replicates in the batch. The estimates are 0.000227 for batch 1 and 0.005141
for batch 2. These estimates are average of the variance between replicates
regardless of the actual person, thief, layer and area from which the sample
is taken. Estimates of the small scale variation (variation between replicates)
from the analysis of repetability (Section 3.2) take into account the the thief
used to collect the sample and the layer from which the sample is collected.
The results are discussed in more detail in the following.
Layers
For both batches the tendency is that the mean content in the top layer is less
than the mean content in the bottom layer.
The analysis also gives an estimate for the variance between the mean content
of areas within a layer. For batch 1 the estimate is 0.00036 = (0.01897)2 . For
batch 2 it has not been possible to distinguish the variation between areas from
the variation between replicates within an area.
Thieves
Thief 1 is a short thief used in the top and the middle layer, whereas both thief
2 and 3 are long thieves used in the middle and the bottom layer. Thief 1
and thief 2 collect samples of the size 1X dose and thief 3 collect samples of
the size 3X dose. Thus, the tip used to thief 1 and thief 2 are similar but not
identical. Because both thief 2 and 3 are long thieves they are interchangeable
and for future sampling it is relevant to know which thief to choose.
In both batches the tendency is that the mean content of the active component
is less in samples collected with thief 2 than in samples collected with thief 3.
230 Paper F
The difference is significant on the 10% level in batch 1 and 16% level in batch
2.
Batch 1 Batch 2
1.15
1.15
Top Top
Midt Midt
Bund Bund
1.10
1.10
1.05
1.05
Relative Dose
Relative Dose
1.00
1.00
0.95
0.95
0.90
0.90
0.85
0.85
1 2 3 Tablets 1 2 3 Tablets
Thief Thief
Fig. 5: In general the relative content in samples collected with thief 2 is

smaller than the relative dose in the tablets.
Regarding content the tablet samples are in general more similar to samples
collected by thief 3 than to samples collected by thief 2. This indicates that
thief 2 introduces bias. The fact that thieves may introduce bias causing the
content in blend samples to differ from the content in tablet samples is in agree-
ment with the literature [5] as well as with Figure 2.4 on page 12. The literature
(e.g. [21]) also gives several examples that the risk of bias is larger when the
blend sample is small, i.e. close to unit dose. This is in agreement with thief
3 taking samples of 3X dose, whereas thief 2, which is under suspicion of
sampling bias, takes samples of 1X dose.
In batch 1 the estimated bias on the mean content in samples collected with
thief 2 is −0.015 ± 0.0184 (95% confidence interval) relatively to the mean
content in samples collected with thief 3. That is, if the relative content in
a sample collected with thief 3 is 0.98, the relative content would have been
0.965 ± 0.0184 had thief 2 been used to collect the sample.
231
In batch 2 the estimated bias is −0.057 ± 0.0814 (95% confidence interval),

i.e. a relative content of 0.98 in a sample collected with thief 3 corresponds to
a relative content of 0.923 ± 0.0814 had thief 2 been used.
Conclusion
Homogeneity of the batches
From blend- and tablet samples it appears that the mean content in batch 1 is
less than the mean content in batch 2, and that in general the variation in batch
2 is larger than the variation in batch 1.
The mean content in the three layers in a batch does not differ significantly.
However, there is a tendency that the small scale variation in a top layer is
smaller than the small scale variation in the bottom.
In batch 1 the variation between replicates (small scale variation) are smaller
than the variation between areas within a layer. In batch 2 the variation be-
tween replicates is so large that it has not been possible to separate the variation
between replicates from the variation between areas within a layer.
It is concluded, that batch 1 is more well mixed than batch 2.
Persons
The mean content in samples collected by person 1 does not differ significantly
from the mean content in samples collected by person 2. It has not been possi-
ble to test whether sampling error (variation) introduced by the two persons is
of the same order of magnitude, but a plot indicates that neither person intro-
duces significantly larger sampling error than the other.
It can not be rejected that the persons have similar skills in sampling.
232
Thieves
The tendency is that thief 3 introduces the largest sampling error (variation)
and that thief 1 introduces the smallest sampling error. A possible explanation
is that it is easier to handle a short thief than a long thief.
Even though thief 2 generally introduces less sampling error than thief 3, thief
3 is recommended for sampling in the bottom layer, as thief 2 is under suspi-
cion of sampling bias.
It is noteworthy that the long thief with the small chamber in the tip (thief 2,
1X dose) is under suspicion of sampling bias, whereas there is no indication
of sampling bias for the long thief with the large chamber in the tip (thief
3, 3X dose). This conclusion is based on a comparison of variation between
replicates from the blends with variation between tablets.
Bibliography
[1] Muzzio, F. J., Roddy, M., Brone, D., Alexander, A. W., and Sudah, O.
An improved Powder-Sampling Tool. Pharmaceutical Technology, pages
92–110, April 1999.
[2] Berman, J. The Compliance and Science of Blend Uniformity Analysis.

PDA Journal of Pharmaceutical Science & technology, 55(4), 2001.
[3] United States of America vs. Barr Laboratories Inc. Civil action for the
District of New Jersey, February 1993.
[4] Center for Drug Evaluation FDA and Research. 21 Code of Federal Reg-
ulations Parts 210 and 211.
[5] Berman, J. and Planchard, J. A. Blend Uniformity and Unit Dose Sam-
pling. Drug Development and Industrial Pharmacy, 21(11), 1995.
[6] FDA. The Food and Drug Administration: An Overview. Publication

No. BG99-2, www.fda.gov, January 1999.
[7] USP. About USP. www.usp.org, 2000.
[8] PDA. About PDA. www.pda.org, 2000.
[9] PQRI. About PQRI. www.pqri.org, 2000.
[10] Department of Health and FDA Human Services. 21 CFR Parts 210
and 211 Current Good Manufacturing Practice: Amendment of Certain
Requirements for Finished Pharmaceuticals; Proposed Rule. In Federal
Register, Proposed Rules, volume 61, pages 20103 – 20115. FDA.
233
234
[11] Draft Guidance for Industry ANDAs: Blend Uniformity analysis, August
1999.
[12] Schilling, E. G. Acceptance Sampling in Quality Control. Marcel Dekker,

Inc., 1982.
[13] Kristensen, H. G. The Characterization of Non-Random Mixtures. A

survey. Powder Technology, 13:103–113, 1976.
[14] Fan, L. T., Chen, S. J., and Watson, C. A. Solids Mixing. Industrial and
Engineering Chemistry, 62(7):53–69, 1970.
[15] Rollins, D. K., Faust, D. L., and Jabas, D. L. A superior approach to

indices in determining mixture segregation. Powder Technology, 84:277–
282, 1995.
[16] Wilrich, P.-Th. In Proceedings of VIIth International Workshop on Intel-

ligent Statistical Quality Control September 5-7, 2001. Institute for Im-
provement in Quality and Productivity, University of Waterloo, Waterloo,
Canada., 2001.
[17] SAS version 8.02. SAS Institute, Inc., Cary, NC., 1999.
[18] PQRI. The use of stratified sampling of blend and dosage units to demon-
strate adequacy of mix for powder blends. www.pqri.org, September 3.
2001. DRAFT.
[19] Montgomery, D. C. Design and Analysis of Experiments. John Wiley &

Sons, Inc., fifth edition edition, 2001.
[20] Guide to Inspections of oral Solid Dosage forms Pre/Post Approval Is-
suses for Development and Validation. FDA Division of Field Investiga-
tions, January 1994.
[21] Kræmer, J., Svensson, J. R., and Melgaard, H. Sampling Bias in blending
Validation and a Different Approach to Homogeneity Assessment. Drug
Development and Industrial Pharmacy, 25(2), 1999.
[22] Berman, J., Elinski, D. E., Gonzales, C. R., Hofer, J. D., Jimenez, P. J.,
Planchard, J. A., Tlachac, R. J., and Vogel, P. F. Technical Report No. 25
235
Blend uniformity Analysis: Validation and in-Process Testing. Technical

Report 25, PDA Solid Dosage Process Validation Committee, 1997.
[23] Berman, J., Schoeneman, A. , and Shelton, J. T. Unit Dose Sampling:

A Tale of Two thieves. Drug Development an Industrial Pharmacy,
22(11):1121–1132, 1996.
[24] Bartlett, M. S. and Kendall, D. G. The Statistical Analysis of Variance-

heterogeneity and the Logarithmic Transformation. Journ. Roy. Statist.
Soc. Suppl. VIII, 1946.
[25] McCullagh, P. and Nelder, J. A. Generalized Linear Models. Chapman

and Hall, London, second edition edition, 1983.
[26] U.S. Pharmacopeia, USP 24-NF19, Second Supplement, July, 2000.
[27] Pharmacopeial Previews; <905> Uniformity of Dosage Units, 1997.
[28] Harmonization, <905> Uniformity of Dosage Units. Pharmacopeial

Forum, 27(3):2615 – 2619, 2001.
[29] Members of the Statistics Working Group Pharmaceutical Research and

Manufacturers of America (PhRMA). Uniformity - Alternative to the
USP Pharmacopeial. Pharmacopeial Forum, 25(2):7939 – 7948, 1999.
[30] S-PLUS. Insightful Corp., 1700 West Lake Avenue North, Suite 500,
Seattle, WA 98109, 2000.
[31] Melgaard, H. and Thyregod, P. Acceptance sampling by variables under

measurement uncertainty. In Wilrich P.-Th. Lenz H.-J., editor, Frontiers
in Statistical Quality Control 5, Heidelberg, 1997. Physica-Verlag.
[32] Wilrich, P.-Th. Part 1: Sampling inspection, Single sampling plans for in-
spection by variables in the presence of measurement error. Allgemeines
Statistisches Archiv, 84:239–250, 2000.
[33] Lieberman and Resnikoff. Sampling plans for Inspection by Variables.

Journ. Amer. Statist. Assoc, (50):457–516, 1955.
[34] ISO 2859: Sampling procedures and charts for inspection by attributes,
Part 1 to 4. International Organization for Standardization, Geneva, 1999.
236
[35] ISO 3951: Sampling procedures and charts for inspection by variables for
percent nonconforming. International Organization for Standardization,
Geneva, 1989.
[36] Quality System Requirements, QS - 9000, 3rd edition. Automotive In-

dustry Action Group, Soutfield, Michigan.
[37] MIL-STD-1916 Department of Defense Test Method Standard - Depart-

ment of Defense Preferred Methods for Acceptance of Product. US De-
partment of Defense, 1996.
[38] EEC Council Directive No. 75/106/EEC of 19. December 1974, pp. 1-13
. Technical report, Official Journal of the European Communities L42,
15. february 1975.
[39] EEC Council Directive of 20. January 1976, pp. 1-11. Technical report,
Official Journal of the European Communities L46, 21. February 1999.
[40] ISO 11462-1: Guidelines for implementation of statistical process control

- Part 1 Elements of SPC. International Organization for Standardization.
Geneva, 2001.
[41] ISO 10576-1: Statistical methods - Guidelines for the evaluation of con-
formity with specified requirements - Part 1: General principles. Inter-
national Organization for Standardization. Geneva, Final Draft submitted
for voting 2002.
[42] < 905 > USP Uniformity of Dosage Units. “USP XXI”, Mack Printing
Company, 1984.
[43] Holst, E., Thyregod, P., and Wilrich, P-Th. On Conformity Testing
and the use of Two-stage Procedures. International Statistical Review,
69:419–432, 2001.
[44] Mathematical Statistics, Basic Ideas and Selected Topics, volume 1.

Prentice-Hall, Upper Saddle River, 2 edition, 2001.
[45] Johnson, N. L., Kotz, S., and Balakrishnan, N. Continuoous Univariate

Distributions, volume 1. Wiley, New York, second edition, 1994.
237
[46] Hahn, G. J. and Meeker, W. Q. Statistical Intervals, A Guide for Practi-

tioners. John Wiley & Sons, New York, 1991.
[47] Probability and Statistics for Engineers and Scientists. Prentice-Hall,

Upper Saddle River, 6 edition, 1998.
[48] ISO 2854: Statistical interpretation of data - Techniques of estimation

relating to means and variances. International Organization for Standard-
ization. Geneva., 1976.
[49] ISO 3494: Statistical interpretation of data - Power of tests relating to

means and variances. International Organization for Standardization.
Geneva., 1976.
[50] Bergum, J. S. Constructing Acceptance Limits for Multiple Stage Tests.

Drug Development and Industrial Pharmacy, 16:2153–2166, 1990.
[51] Bray, D. F., Lyon, D. A., and Burr, I. W. Three Class Attributes Plans in
Acceptance Sampling. Technometrics, 15, 1973.
[52] International Commison on Microbiological Specifications for Foods

(ICMSF) of the International Union of Microbiological Societies, Uni-
versity of Toronto Press, Toronto. Sampling for microbiological analysis:
principles and specific applications, 2 edition, 1986.
[53] Dahms, S. and Hildebrandt. Some remarks on the Design of Three-Class

Sampling Plans. Journal of Food Protection, 61, 1998.
[54] Wallis, W. A. Lot quality measured by proportion defective. Chapter 1 in

Selected Techniques of Statistical Analysis (Eisenhart, Hastay and Wallis,
eds). McGraw-Hill Book Co. Inc., New York, 1947.
[55] MIL-STD-414 Sampling procedures and Tables for Inspection by Vari-

ables for Percent Defective. Superintendant of Documents, United States
Government Printing Office, Washington, D.C., 1957.
[56] American National Standard: Sampling Procedures and Tables for In-
spection by Variables for Percent Nonconformance, ANSI/ASQC Stan-
dard Z1.9. American National Standards Institute, 1980.
238
[57] Schilling, E. G. Revised variables acceptance sampling standards -

ANSI Z1.9 (1980) and ISO 3951 (1980). Journal of Quality Technol-
ogy, 13:131–138.
[58] Boulanger, M., Johnson, M., Perrouchet, C., and Thyregod, P. Interna-
tional statistical standards. Their place in the life-cycle of products and
services. International Statistical Review, (67):151–173, 1999.
[59] Bruhn-Suhr, M., Krumbholz, W., and Lenz, H.-J. On the Design of Exact
Single Sampling Plans by Variables. In Wilrich P.-Th. Lenz H.-J., editor,
Frontiers in Statistical Quality Control, volume 4, pages 83–90. Physica
Verlag, Heidelberg, 1992.
[60] Lei, D.-H. and Vardeman, S. B. The LRT method of constructing a

two-sided “variables” acceptance region and its comparison with other
methods. Communications in Statistics, Part A – Theory and Methods,
(27):329–351, 1998.
[61] Lehmann, E. L. Testing Statistical Hypotheses. Wiley, second edition

edition, 1986.
[62] PhRMA. Recommendations for a Globally Harmonized Uniformity of

Dosage Units Test. Pharmacopeial Forum, 25(4):8609–8624, 1999.
[63] Wilrich, P.-Th. Single sampling plans for inspection by variables under
a variance component situation. In Proceedings of VIIth International
Workshop on Intelligent Statistical Quality Control September 5-7, 2001.,
Waterloo, Canada. Institute for Improvement in Quality and Productivity,
University of Waterloo.
[64] Yalkowsky, S.H. and Bolton, S. Particle Size and Content Uniformity.
Pharmaceutical Research, 7:962–966, 1990.
[65] Stoyan, D., Kendall, W.S, and Mecke, J. Stochastic Geometry and its
Applications. Wiley, second edition, 1995.
[66] Johnson, N.L., Kemp, A.W., and Kotz, S. Univariate Discrete Distribu-
tions. Wiley, New York, 2 edition, 1993.
239
Ph. D. theses from IMM
1. Larsen, Rasmus. (1994). Estimation of visual motion in image se-

quences. xiv + 143 pp.
2. Rygard, Jens Moberg. (1994). Design and optimization of flexible

manufacturing systems. xiii + 232 pp.
3. Lassen, Niels Christian Krieger. (1994). Automated determination of

crystal orientations from electron backscattering patterns. xv + 136 pp.
4. Melgaard, Henrik. (1994). Identification of physical models. xvii +

246 pp.
5. Wang, Chunyan. (1994). Stochastic differential equations and a bio-

logical system. xxii + 153 pp.
6. Nielsen, Allan Aasbjerg. (1994). Analysis of regularly and irregularly

sampled spatial, multivariate, and multi-temporal data. xxiv + 213 pp.
7. Ersbøll, Annette Kjær. (1994). On the spatial and temporal correla-

tions in experimentation with agricultural applications. xviii + 345 pp.
8. Møller, Dorte. (1994). Methods for analysis and design of heteroge-

neous telecommunication networks. Volume 1-2, xxxviii + 282 pp., 283-
569 pp.
9. Jensen, Jens Christian. (1995). Teoretiske og eksperimentelle dy-

namiske undersøgelser af jernbanekøretøjer. viii + 174 pp.
10. Kuhlmann, Lionel. (1995). On automatic visual inspection of reflective

surfaces. Volume 1, xviii + 220 pp., (Volume 2, vi + 54 pp., fortrolig).
11. Lazarides, Nikolaos. (1995). Nonlinearity in superconductivity and

Josephson Junctions. iv + 154 pp.
12. Rostgaard, Morten. (1995). Modelling, estimation and control of fast

sampled dynamical systems. xiv + 348 pp.
13. Schultz, Nette. (1995). Segmentation and classification of biological

objects. xiv + 194 pp.
240
14. Jørgensen, Michael Finn. (1995). Nonlinear Hamiltonian systems. xiv

+ 120 pp.
15. Balle, Susanne M. (1995). Distributed-memory matrix computations.

iii + 101 pp.
16. Kohl, Niklas. (1995). Exact methods for time constrained routing and
related scheduling problems. xviii + 234 pp.
17. Rogon, Thomas. (1995). Porous media: Analysis, reconstruction and

percolation. xiv + 165 pp.
18. Andersen, Allan Theodor. (1995). Modelling of packet traffic with

matrix analytic methods. xvi + 242 pp.
19. Hesthaven, Jan. (1995). Numerical studies of unsteady coherent struc-

tures and transport in two-dimensional flows. Risø-R-835(EN) 203 pp.
20. Slivsgaard, Eva Charlotte. (l995). On the interaction between wheels

and rails in railway dynamics. viii + 196 pp.
21. Hartelius, Karsten. (1996). Analysis of irregularly distributed points.

xvi + 260 pp.
22. Hansen, Anca Daniela. (1996). Predictive control and identification -

Applications to steering dynamics. xviii + 307 pp.
23. Sadegh, Payman. (1996). Experiment design and optimization in com-

plex systems. xiv + 162 pp.
24. Skands, Ulrik. (1996). Quantitative methods for the analysis of elec-
tron microscope images. xvi + 198 pp.
25. Bro-Nielsen, Morten. (1996). Medical image registration and surgery

simulation. xxvii + 274 pp.
26. Bendtsen, Claus. (1996). Parallel numerical algorithms for the solution
of systems of ordinary differential equations. viii + 79 pp.
27. Lauritsen, Morten Bach. (1997). Delta-domain predictive control and

identification for control. xxii + 292 pp.
241
28. Bischoff, Svend. (1997). Modelling colliding-pulse mode-locked semi-

conductor lasers. xxii + 217 pp.
29. Arnbjerg-Nielsen, Karsten. (1997). Statistical analysis of urban hy-

drology with special emphasis on rainfall modelling. Institut for Miljøteknik,
DTU. xiv + 161 pp.
30. Jacobsen, Judith L. (1997). Dynamic modelling of processes in rivers

affected by precipitation runoff. xix + 213 pp.
31. Sommer, Helle Mølgaard. (1997). Variability in microbiological degra-

dation experiments - Analysis and case study. xiv + 211 pp.
32. Ma, Xin. (1997). Adaptive extremum control and wind turbine control.
xix + 293 pp.
33. Rasmussen, Kim Ørskov. (1997). Nonlinear and stochastic dynamics

of coherent structures. x + 215 pp.
34. Hansen, Lars Henrik. (1997). Stochastic modelling of central heating

systems. xxii + 301 pp.
35. Jørgensen, Claus. (1997). Driftsoptimering på kraftvarmesystemer.

290 pp.
36. Stauning, Ole. (1997). Automatic validation of numerical solutions.

viii + 116 pp.
37. Pedersen, Morten With. (1997). Optimization of recurrent neural net-

works for time series modeling. x + 322 pp.
38. Thorsen, Rune. (1997). Restoration of hand function in tetraplegics

using myoelectrically controlled functional electrical stimulation of the
controlling muscle. x + 154 pp. + Appendix.
39. Rosholm, Anders. (1997). Statistical methods for segmentation and

classification of images. xvi + 183 pp.
40. Petersen, Kim Tilgaard. (1997). Estimation of speech quality in telecom-

munication systems. x + 259 pp.
242
41. Jensen, Carsten Nordstrøm. (1997). Nonlinear systems with discrete

and continuous elements. 195 pp.
42. Hansen, Peter S.K. (1997). Signal subspace methods for speech en-
hancement. x + 226 pp.
43. Nielsen, Ole Møller. (1998). Wavelets in scientific computing. xiv +

232 pp.
44. Kjems, Ulrik. (1998). Bayesian signal processing and interpretation of

brain scans. iv + 129 pp.
45. Hansen, Michael Pilegaard. (1998). Metaheuristics for multiple ob-

jective combinatorial optimization. x + 163 pp.
46. Riis, Søren Kamaric. (1998). Hidden markov models and neural net-
works for speech recognition. x + 223 pp.
47. Mørch, Niels Jacob Sand. (1998). A multivariate approach to func-

tional neuro modeling. xvi + 147 pp.
48. Frydendal, Ib. (1998.) Quality inspection of sugar beets using vision.
iv + 97 pp. + app.
49. Lundin, Lars Kristian. (1998). Parallel computation of rotating flows.

viii + 106 pp.
50. Borges, Pedro. (1998). Multicriteria planning and optimization. -

Heuristic approaches. xiv + 219 pp.
51. Nielsen, Jakob Birkedal. (1998). New developments in the theory of

wheel/rail contact mechanics. xviii + 223 pp.
52. Fog, Torben. (1998). Condition monitoring and fault diagnosis in ma-
rine diesel engines. xii + 178 pp.
53. Knudsen, Ole. (1998). Industrial vision. xii + 129 pp.
54. Andersen, Jens Strodl. (1998). Statistical analysis of biotests. - Ap-

plied to complex polluted samples. xx + 207 pp.
243
55. Philipsen, Peter Alshede. (1998). Reconstruction and restoration of

PET images. vi + 132 pp.
56. Thygesen, Uffe Høgsbro. (1998). Robust performance and dissipation

of stochastic control systems. 185 pp.
57. Hintz-Madsen, Mads. (1998). A probabilistic framework for classifi-

cation of dermatoscopic images. xi + 153 pp.
58. Schramm-Nielsen, Karina. (1998). Environmental reference materials

methods and case studies. xxvi + 261 pp.
59. Skyggebjerg, Ole. (1999). Acquisition and analysis of complex dy-

namic intra- and intercellular signaling events. 83 pp.
60. Jensen, Kåre Jean. (1999). Signal processing for distribution network
monitoring. xv + 199 pp.
61. Folm-Hansen, Jørgen. (1999). On chromatic and geometrical calibra-

tion. xiv + 238 pp.
62. Larsen, Jesper. (1999). Parallelization of the vehicle routing problem

with time windows.xx + 266 pp.
63. Clausen, Carl Balslev. (1999). Spatial solitons in quasi-phase matched

struktures. vi + (flere pag.)
64. Kvist, Trine. (1999). Statistical modelling of fish stocks. xiv + 173 pp.
65. Andresen, Per Rønsholt. (1999). Surface-bounded growth modeling

applied to human mandibles. xxii + 125 pp.
66. Sørensen, Per Settergren. (1999). Spatial distribution maps for benthic
communities.
67. Andersen, Helle. (1999). Statistical models for standardized toxicity

studies. viii + (flere pag.)
68. Andersen, Lars Nonboe. (1999). Signal processing in the dolphin

sonar system. xii + 214 pp.
244
69. Bechmann, Henrik. (1999). Modelling of wastewater systems. xviii +

161 pp.
70. Nielsen, Henrik Aalborg. (1999). Parametric and non-parametric sys-

tem modelling. xviii + 209 pp.
71. Gramkow, Claus. (1999). 2D and 3D object measurement for control

and quality assurance in the industry. xxvi + 236 pp.
72. Nielsen, Jan Nygaard. (1999). Stochastic modelling of dynamic sys-

tems. xvi + 225 pp.
73. Larsen, Allan. (2000). The dynamic vehicle routing problem. xvi + 185
pp.
74. Halkjær, Søren. (2000). Elastic wave propagation in anisotropic inho-

mogeneous materials. xiv + 133 pp.
75. Larsen, Theis Leth. (2000). Phosphorus diffusion in float zone silicon
crystal growth. viii + 119 pp.
76. Dirscherl, Kai. (2000). Online correction of scanning probe micro-

scopes with pixel accuracy. 146 pp.
77. Fisker, Rune. (2000). Making deformable template models operational.

xx + 217 pp.
78. Hultberg, Tim Helge. (2000).Topics in computational linear optimiza-

tion. xiv + 194 pp.
79. Andersen, Klaus Kaae. (2001).Stochastic modelling of energy systems.

xiv + 191 pp.
80. Thyregod, Peter. (2001). Modelling and monitoring in injection mold-

ing. xvi + 132 pp.
81. Schjødt-Eriksen, Jens. (2001). Arresting of collapse in inhomogeneous

and ultrafast Kerr media.
82. Bennetsen, Jens Christian. (2001). Numerical simulation of turbulent

airflow in livestock buildings. xi + 205 pp + Appendix.
245
83. Højen-Sørensen, Pedro A.d.F.R. (2001). Approximating methods for

intractable probabilistic models: - Applications in neuroscience. xi +
104 pp + Appendix.
84. Nielsen, Torben Skov. (2001). On-line prediction and control in non-
linear stochastic systems. xviii + 242 pp.
85. Öjelund, Henrik. (2001). Multivariate calibration of chemical sensors.

xviii + 182 pp.
86. Adeler, Pernille Thorup. (2001). Hemodynamic simulation of the heart

using a 2D model and MR data. xv + 179 pp.
87. Nielsen, Finn Årup. (2001). Neuroinformatics in functional neuroimag-

ing. 330 pp.
88. Kidmose, Preben. (2001). Blind separation of heavy tail signals. viii +
136 pp.
89. Hilger, Klaus Baggesen. (2001). Exploratory analysis of multivariate

data. xxiv + 180 pp.
90. Antonov, Anton. (2001). Object-oriented framework for large scale air
pollution models. 156 pp. + (flere pag).
91. Poulsen, Mikael Zebbelin. (2001). Structural analysis of DAEs. 192

pp.
92. Keijzer, Maarten. (2001). Scientific discovery using genetic program-

ming.
93. Sidaros, Karam. (2002). Slice profile effects in MR perfusion imaging

using pulsed arterial spin labelling. xi + 202 pp.
94. Kolenda, Thomas. (2002). Adaptive tools in virtual environments. -

Independent component analysis for multimedia. xiii + 117 pp.
95. Berglund, Ann-Charlotte. (2002). Nonlinear regularization with ap-

plications in geophysics. xiii + 116 pp.
96. Pedersen, Lars. (2002). Analysis of two-dimensional electrophoresis

Gel images. xxii + 161 pp.
246
97. Rasmussen, Thomas Marthedal. (2002). Interval logic. Proof theory

and theorem proving. xii + 150 pp.
98. Vistisen, Dorte. (2002). Models and methods for hot spot safety work.
ix + 168 pp.
99. Madsen, Camilla. (2002). Statistical methods for assessment of blend

homogeneity. xviii + 246 pp.
100. Holten-Lund, Hans. (2002). Design for scalability in 3D computer

graphics architectures. x + 169 pp.
101. Bache, Morten. (2002). Quantum noise and spatio-temporal pattern-

formation in nonlinear optics. ix + 123 pp.
102. Stidsen,Thomas Jacob Kjær. (2002). Optimisation problems in optical

network design. xii + 55 + Appendix.
103. Andersen, Irene Klærke. (2002). Dynamic contrast enhanced MRI for
perfusion quantification. iv + 181 pp.
104. Xia, Fujie. (2002). The dynamics of the three-piece-freight truck. x +

205 pp.
105. Johansen, Steffen Kjær. (2002). All-optical signal processing in quadratic

nonlinear materials. x + 90 pp.
106. Nyeland, Martin Erik. (2002). Sampling strategy and statistical mod-
elling of exposure. viii + 39 + appendix.

Statistical Mixing

Uploaded by

Copyright:

Available Formats

Statistical Mixing

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Statistical Mixing

Uploaded by

Copyright:

Available Formats

Statistical Methods for

This document was prepared with LATEX and printed by DTU-tryk.

The thesis is concerned with statistical methods to assess blend uniformity in

Lyngby, May 2002.

First of all, I wish to thank my academic supervisor, Prof. Poul Thyregod,

It is not straight forward to assess the homogeneity of a powder blend. The

In the thesis a number of methods to assess (in)homogeneity are presented.

to decide whether the blend or batch is sufficiently homogeneous (uniform) to

Denne afhandling omhandler brugen af statistiske metoder til at belyse forskel-

Det at bestemme homogeniteten af en pulverblanding er ikke simpelt. Det

I afhandlingen er forskellige metoder til at vurdere homogenitet beskrevet.

Et central problem er at opstille accept kriterier for blandinger og tablet batche,

Resumé (in Danish) . . . . . . . . . . . . . . . . . . . . . . . . . . xi

1.1 Outline of the Thesis . . . . . . . . . . . . . . . . . . . . . . 5

2.1 Principles of tablet production . . . . . . . . . . . . . . . . . 7

2.2 Assessment of the uniformity of the blend . . . . . . . . . . . 10

2.3 Regulatory Affairs . . . . . . . . . . . . . . . . . . . . . . . 11

2.3.2 Requirements and Recommendations . . . . . . . . . 15

3 Results and discussion 17

3.1 Variances as a measure of homogeneity . . . . . . . . . . . . 18

3.2 Methods to assess homogeneity and factors that may influence

3.2.1 The effect of particle size distribution . . . . . . . . . 22

3.2.2 Assessment of homogeneity in specific batches . . . . 23

3.3 Analysis of acceptance criteria . . . . . . . . . . . . . . . . . 26

A Robustness and power of statistical methods to assess blend

2 Models of batch homogeneity . . . . . . . . . . . . . . . . . 40

2.1 The aggregated model . . . . . . . . . . . . . . . . . 40

2.2 The hierarchical model . . . . . . . . . . . . . . . . . 41

2.3 Simulated samples . . . . . . . . . . . . . . . . . . . 44

3 Assessing factors with influence on the mean content of the

3.1 Large and medium scale homogeneity assessed accord-

3.2 Large and medium scale homogeneity assessed accord-

3.3 The effect of including sampling thieves in the model . 50

4 Assessing factors with influence on the variation between repli-

B Comprehensive measures of blend uniformity 63

2 Batches with medium scale variation (variation between areas) 68

2.1 The total variation from the ANOVA table . . . . . . . 68

2.2 The variance on a randomly sampled unit from the batch 69

3 Batches with large scale variation (variation between layers) . 71

3.1 The total variation from the ANOVA table . . . . . . . 73

3.2 The variance on a sample from the batch . . . . . . . 75

4 Batches with both large and medium scale variation . . . . . . 75

4.1 The total variation from the ANOVA table . . . . . . . 76

4.2 The variance on a sample from the batch . . . . . . . 82

4.3 The direct relation between large/medium scale varia-

5 Discussion and conclusion . . . . . . . . . . . . . . . . . . . 82

C On a test for content uniformity in pharmaceutical products

2 The proposed test . . . . . . . . . . . . . . . . . . . . . . . . 90

2.1 Historical notes . . . . . . . . . . . . . . . . . . . . . 90

2.2 Description of the proposed test . . . . . . . . . . . . 91

3 Properties of the proposed test . . . . . . . . . . . . . . . . . 93

3.1 Description of the OC-surface of the test . . . . . . . 94

3.2 “Specification limits” for individual tablets . . . . . . 95

3.3 Details on the effect of individual elements of the test . 97

4 Comparison to other test procedures . . . . . . . . . . . . . . 101

D Statistical tests for uniformity of blend/dose samples 105

2 Acceptance criteria and statistical hypothesis testing . . . . . . 111

2.1 Choice of null hypothesis and alternative hypothesis . 111

2.2 Confidence intervals and statistical tests . . . . . . . . 113

3 Notation and distributional assumptions . . . . . . . . . . . . 114