Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Technical Report No. 59

Utilization of Statistical Methods for
Production Monitoring

Paradigm Change in
Manufacturing OperationsSM

Bethesda Towers
4350 East West Highway
Suite 200
Bethesda, MD 20814 USA
Tel: 1 (301) 656-5900
Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

PDA Utilization of Statistical Methods for Production Monitoring Task Force Members
Authors

Greg Flexman, Grifols (Chair) Stephan Rönninger, F. Hoffmann-La Roche Ltd.

Michael Husovich, Amgen Mark A. Skoog, Genentech

Jason J. Orloff, PharmStat Lynn D. Torbeck, PharmStat

Jayesh Patel, F. Hoffmann-La Roche Inc. Dan Weese, Amgen

This technical report was developed as a part of PDA’s Paradigm Change in Manufacturing Operation (PCMO) proj-
ect. The content and views expressed in this Technical Report are the result of a consensus achieved by the authoriz-
ing Task Force and are not necessarily views of the organizations they represent.
Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Utilization of Statistical
Methods for Production
Monitoring
Technical Report No. 59

ISBN: 978-0-939459-44-5
© 2012 Parenteral Drug Association, Inc.
All rights reserved.

Bethesda Towers
4350 East West Highway
Suite 200
Bethesda, MD 20814 USA
Tel: 1 (301) 656-5900
Fax: 1 (301) 986-0296
E-mail: [email protected]
Web site: www.pda.org
Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Paradigm Change in Manufacturing Operations (PCMOSM)

PDA launched the project activities related to the PCMO program in December 2008 to help imple-
ment the scientific application of the ICH Q8, Q9 and Q10 series. The PDA Board of Directors ap-
proved this program in cooperation with the Regulatory Affairs and Quality Advisory Board, and the
Biotechnology Advisory Board and Science Advisory Board of PDA.

Although there are a number of acceptable pathways to address this concept, the PCMO program fol-
lows and covers the drug product lifecycle, employing the strategic theme of process robustness with-
in the framework of the manufacturing operations. This project focuses on Pharmaceutical Quality
Systems as an enabler of Quality Risk Management and Knowledge Management.

Using the Parenteral Drug Association’s (PDA) membership expertise, the goal of the Paradigm
Change in Manufacturing Operations Project is to drive the establishment of ‘best practice’ docu-
ments and /or training events in order to assist pharmaceutical manufacturers of Investigational
Medicinal Products (IMPs) and commercial products in implementing the ICH guidelines on Phar-
maceutical Development (ICH Q8, Q11), Quality Risk Management (ICH Q9) and Pharmaceutical
Quality Systems (ICH Q10).

The PCMO program facilitates communication among the experts from industry, university and regula-
tors as well as experts from the respective ICH Expert Working Groups and Implementation Working
Group. PCMO task force members also contribute to PDA conferences and workshops on the subject.

PCMO follows the product lifecycle concept and has the following strategic intent:
• Enable an innovative environment for continual improvement of products and systems
• Integrate science and technology into manufacturing practice
• Enhance manufacturing process robustness, risk based decision making and knowledge management
• Foster communication among industry and regulatory authorities

The Product Life Cycle

Pharmaceutical Technology Commercial Product

Development Transfer Manufacturing Discontinuation

For more information, including the PCMO Dossi.e., and to get involved, go to
www.pda.org/pcmo
Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Table of Contents

1.0  Introduction..........................................................1 4.0 Acceptance Sampling........................................24

1.1 Purpose and Scope........................................... 1 4.1 Typical Applications....................................... 24
1.2 Implementation to Support Decision Making.... 1 4.2 Key Terms....................................................... 24
4.2.1 Acceptable Quality Limit (AQL)................ 24
2.0  Glossary of Terms................................................2 4.2.2 Rejectable Quality Level (RQL)................. 24
4.3 Types of Sampling.......................................... 25
3.0  Statistical Process Control Tools.........................8
4.3.1 Attributes Sampling................................. 25
3.0.1 Prerequisites for Data Analysis.................. 8 4.3.2 Variable Sampling.................................... 25
3.1 Run Charts...................................................... 11 4.4 Types of Acceptance Plans............................ 25
3.1.1 Typical Applications................................. 12 4.4.1 Single Sampling Plans............................. 25
3.1.2 Pros......................................................... 12 4.4.2 Double Sampling Plans............................ 25
3.1.3 Cons........................................................ 12 4.4.3 Individual Sampling Plan.......................... 25
3.2 Control Charts: Individuals.............................. 12 4.4.4 Sampling Scheme.................................... 25
3.2.1 Attribute Control Charts........................... 12 4.5 Pros and Cons................................................ 25
3.2.2 Typical Applications................................. 13
3.2.3 Pros......................................................... 13 5.0 Appendices: Technical Details and Examples...27
3.2.4 Cons........................................................ 13 5.1 Run Charts...................................................... 27
3.3 Moving Range Control Charts......................... 13 5.1.1 Technical Details...................................... 27
3.3.1 Typical Applications................................. 13 5.1.2 Example................................................... 27
3.3.2 Pros......................................................... 14 5.2 Control Charts: Individuals.............................. 29
3.3.3 Cons........................................................ 14
5.2.1 Technical Details...................................... 29
3.4 Average and Variability Charts....................... 14
5.2.2 Example................................................... 30
3.4.1 Typical Applications................................. 14
5.3 Moving Range Control Charts......................... 33
3.4.2 Pros......................................................... 14
5.3.1 Technical Details...................................... 33
3.4.3 Cons........................................................ 15
5.3.2 Example................................................... 34
3.5 Histograms..................................................... 15
5.4 Average and Variability Charts....................... 37
3.5.1 Typical Application.................................. 16
5.4.1 Technical Details...................................... 37
3.5.2 Pros......................................................... 16
5.4.2 Example................................................... 39
3.5.3 Cons........................................................ 16
3.5.4 Distributions – 5.5 Histograms..................................................... 47
Interpretation of Histograms.................... 16 5.5.1 Example Case Worked in
3.5.5 Hints for Use............................................ 18 Parallel with Theory/Principles................. 47
3.6 Process Capability (Cpk, Ppk)............................ 18 5.6 Cpk, Ppk for Process Capability.......................... 48
3.6.1 Assumptions........................................... 18 5.6.1 General Procedure for
3.6.2 Typical Applications................................. 19 Finding Cpk and Ppk................................... 48
3.6.3 Pros......................................................... 20 5.6.2 Example Cpk for a Single Group................ 49
3.6.4 Cons........................................................ 20 5.6.3 Cpk for Several Groups.............................. 50
3.7 Exponentially Weighted 5.6.4 Example 3, Ppk for Several Groups........... 52
Moving Average Charts .......................... 20 5.7 Exponentially Weighted Moving
3.7.1 Typical Applications................................. 21 Area Charts (EWMA)..................................... 52
3.7.2 Pros......................................................... 21 5.7.1 Technical Details...................................... 52
3.7.3 Cons........................................................ 21 5.7.2 Calculation Example................................ 53
3.8 CuSum Charts................................................ 22 5.7.3 Control Limits.......................................... 53
3.8.1 Process Features Suitable for 5.7.4 Example................................................... 54
This Type Method.................................... 22 5.7.4.1 Interpretation of Example................... 54
3.8.2 Typical Applications................................. 22 5.8 CuSum Charts................................................ 55
3.8.3 Pros......................................................... 22 5.8.1 Example Calculation................................ 56
3.8.4 Cons........................................................ 23 5.8.2 Details of Principles................................. 57
3.9 Examples of Efficient Mixture of the 5.8.3 Interpretations Worked in Parallel with
Statistical Toolbox.......................................... 23 Theory/Principles..................................... 57
Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

5.8.4 Implementation of CuSum Charts............ 59 5.9 Sampling........................................................ 61

5.8.5 Example: Gradual Trend .......................... 59 5.9.1 Example................................................... 61
5.8.6 Preventative Action Before 6.0 References..........................................................63
Failure Occurrence................................... 60
5.8.7 Example: For Process Optimization.......... 60 7.0 Additional Reading..............................................64

FIGURES AND TABLES INDEX

Figure 2.0- 1 Example of a High Capability Figure 3.5.4-3 Bimodal Distribution........................... 17

(Low Variability) Process Capability Figure 3.5.4-4 Left Skewed Distribution.................... 18
Histogram............................................. 2
Table 3.6.1-1 Significance of Cp............................... 19
Figure 2.0-2 Example of a Low Capability
(High Variability) Process Capability Table 3.6.2-1 Interpretation of Cp
Histogram............................................. 3 Regarding Limits Taken....................... 20
Figure 2.0-3 Examples of a Stable Process Figure 3.7-1 Data Plot Compared to EWMA Chart.....21
(Statistically In Control)........................ 4 Figure 3.8-1 Run Plot Compared to CuSum
Figure 2.0-4 Examples of an Unstable Process Chart of the Same Data...................... 22
(Statistically Out Of Control)................. 4 Figure 3.9-1 Example of a Tool Bar for Statistical
Figure 2.0-5 Examples of Trends (Points Beyond Control of a Process........................... 23
Control Limits)...................................... 6 Table 5.1.2-1 Lot Data by Manufacture date............ 27
Figure 2.0-6 Examples of Trends (Points Figure 5.1.2-1 Run Chart (Plot of Lot Data)................ 28
within the control limits)....................... 6
Figure 5.1.2-2 Run Chart .......................................... 28
Table 3.0.1-1 Suggestions and Proposals on
When to Use Statistical Tools .............. 9 Table 5.2.2-1 First 15 data points/ lots..................... 30

Table 3.0.1-2 Areas of Potential Implementation...... 10 Figure 5.2.2-1 Run Chart (Plot of Lot Data)................ 30

Figure 3.1-1 Run Chart........................................... 11 Figure 5.2.2-2 Run Chart .......................................... 31

Figure 3.2.1-1 Individual Control Chart...................... 12 Table 5.2.2-2 Moving Range.................................... 32
Figure 3.3-1 Moving Range Control Chart............... 13 Figure 5.2.2-3 Run Chart (UCL and LCL shown)........ 33
Figure 3.4-1 Average and Range Chart................... 14 Table 5.3.2-1 Lot Data in Time Order........................ 34
Figure 3.5-1 Example of a typical Histogram Table 5.3.2-2 Lot Data with Moving Range.............. 35
Showing Data Location In Relation Figure 5.3.2-1 MR Chart............................................ 35
To Specification Limits........................ 15 Figure 5.3.2-2 Moving Range Chart (Upper &
Figure 3.5.4-1 Normal Distribution............................ 16 Lower Limits Shown)......................... 36
Figure 3.5.4-2 Histogram of a Normal Distribution.... 17 Table 5.4.2-1 Sample Data Collection Table............. 40
Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Table 5.4.2-2 Sample Average Data Table 5.6.2-1 Example of Cpk for a Single Group....... 49
Collection Table.................................. 40
Table 5.6.2-2 Example of 95% Confidence
Figure 5.4.2-1 Sample X Chart................................... 41 Interval for Cpk..................................... 50
Table 5.4.2-3 Sample Data Range Collection Table... 41 Table 5.6.3-1 Cpk for Several Groups........................ 51
Figure 5.4.2-2 Sample Range Chart........................... 42 Table 5.6.3-2 For d2................................................. 51
Table 5.4.2-4 Sample Data Average and
Table 5.7.2-1 Step-Change Example
Range Collection Table........................ 42
(No process noise)............................. 53
Figure 5.4.2-3 Sample Average X Chart.................... 43
Figure 5.7.4-1 Sample Random Distribution Plots..... 54
Figure 5.4.2-4 Sample Range X Chart........................ 43
Figure 5.7.4.1-1 EWMA Plot (UCL & LCL).................... 55
Figure 5.4.2-5 Sample Grand Range X Chart.............. 44
Figure 5.4.2-6 Sample X Chart of Average Figure 5.8.1-1 Example Calculation........................... 56
Range on the Range........................... 44 Figure 5.8.1-2 X-bar Chart vs. CuSum Chart.............. 56
Figure 5.4.2-7 Sample UCL and the LCL on the Figure 5.8.1-3 CuSum Chart with V-mask.................. 57
Range Chart........................................ 45
Figure 5.8.3-1 CuSum Chart with Values Around 100.... 58
Figure 5.4.2-8 Sample UCL and the LCL on the
Average Chart.................................... 46 Figure. 5.8.3-2 CuSum Chart with a
Slide Move at Event 50...................... 58
Table 5.4.2-5 Sample Average and
Range Data Table................................. 46 Figure. 5.8.3-3 CuSum Chart with an Upward Shift
Figure 5.4.2-9 Sample Average and Range X Chart ....46 in Mean Between Points 23 to 50...... 58
Figure 5.4.2-10 Sample Average and Range Figure 5.8.3-4 Example of a CuSum Chart................. 59
X Chart (UCL & LCL)........................... 47 Figure 5.8.5-1 Gradual Trend..................................... 60
Figure 5.5.1-1 Sample Average and Range X Chart Figure 5.8.7-1 For process optimization.................... 60
(UCL & LCL)........................................ 47
Table 5.9.1-1 Characteristics of Example Test Plan....62
Figure 5.5.1-2 Sample Statistical Distribution
of Results........................................... 48 Figure 5.9.1-1 OC Chart Curve................................... 62
Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

1.0  Introduction

As manufacturers seek to improve the quality of their goods, statistical methods have been rediscov-
ered as vital tools for successful development and manufacturing. Industries like automotive, elec-
tronics, and consumer products grow and change partly as a result of adopting statistical methods.

The pharmaceutical and biopharmaceutical industry increasingly recognizes the importance of statis-
tical methods to consistently create products that conform to predetermined quality characteristics.
Statistical methods provide objective evidence in meeting this goal and are fundamental for under-
standing the process, which enables further improvement and development.

Industry and regulatory bodies are working together to provide guidance and frameworks on the
use of statistical methods. The International Conference on Harmonisation, International Standards
Organization and European Union have provided guidance on the use of statistical methods.

In light of the increased focus on this topic, this PDA Task Force recognized the need to provide guid-
ance to help companies identify and use statistical methods. The primary objective of this Task Force
was to convey the appropriate use of statistical methods at a level most can understand.

1.1 Purpose and Scope

The purpose of this document is to present relevant and easy-to-use statistical process control (SPC)
methods that are applicable to the pharmaceutical/biopharmaceutical industry. Advanced statistical
methods, such as multivariate models and Design of Experiment (DoE) will not be considered. An
overview of acceptance sampling is also included in Section 4.0.

1.2 Implementation to Support Decision Making

Statistical methods are intended to improve the quality of decision-making. They are simply a means
to a result. If the manufacturer does not first understand why it is utilizing a statistical method, prob-
lems such as failing to detect important signals or over-detecting unimportant normal variation can
occur. Caution should be exercised to first establish the question to be answered and then the statisti-
cal method to aid in answering the question.

The statistical methods may be used in an ongoing program to analyze collected data. Timely evalu-
ation of data allows the prompt detection of undesired process variation, which facilitates process
understanding and may support responses to control variability.
To best aid the end-user, each statistical method is described in the following format:
• Description • Pros and Cons
• Typical Applications • Technical Details and Examples (see appendices)

The guidance contained in this document is not intended to establish mandatory standards for using
statistical methods across a product’s lifecycle.

Technical Report No. 59 © 2012 Parenteral Drug Association, Inc. 1

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

2.0  Glossary of Terms

Some of the key concepts are illustrated below; additional explanation will be in subsequent sections
Limits
There are different types of limits to be considered. Some are binding (specification limits); others are
for orientation (process control limits).
Specification Limits
Specification limits are set by the manufacturer based on therapeutic, product and regulatory require-
ments. ICH Q6A defines specification as: A list of tests, references to analytical procedures, and appropriate
acceptance criteria which are numerical limits, ranges, or other criteria for the tests described. It establishes the
set of criteria to which a drug substance or drug product should conform to be considered acceptable for its in-
tended use. “Conformance to specifications” means that the drug substance and/or drug product, when tested
according to the listed analytical procedures, will meet the listed acceptance criteria. Specifications are critical
quality standards that are proposed and justified by the manufacturer and approved by regulatory authorities
(1,2).
Specification limits:
• Denote the boundary between acceptable and unacceptable, the quality threshold
• Describe what the process must achieve. When specifications are exceeded, there is a loss of value, time or
cost
• Should be driven by therapeutic effect and toxicological impact as relevant to the patient. Specifications
control risk to the patient

Process Capability
Process capability describes how the process performs in relation to the specifications.
High Capability Processes
High capability processes have low inherent variation relative to the specification or goal. When
a number of results are plotted on a histogram, it is unlikely that there will be many result occur-
rences near the specification limits.
Figure 2.0- 1 Example of a High Capability (Low Variability) Process Capability Histogram

2 © 2012 Parenteral Drug Association, Inc. Technical Report No. 59

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Process Capability (cont.)

Low Capability Processes
Low capability processes have greater inherent variation relative to the specification or goal. When
a number of results are plotted on a histogram, it is likely that results will occasionally occur be-
yond the specification limits.
Figure 2.0-2 Example of a Low Capability (High Variability) Process Capability Histogram

Statistical Process Control Limits

Statistical process control limits are statistically derived measures that are used to define the typical
operating range for the process. Process control is the focus of this document. Unlike specification
boundaries which are related to product impact, control limits are boundaries that annunciate when
process performance may have shifted.
Process control limits:
• Denote the boundary between typical and unusual operational performance ranges for the process.
• Are calculated from prior performance data and used to detect when a process is unstable or “out of statistical
control”.

The prompt recognition of deviations from typical performance enables review. The review can lead to
an understanding of how a process may be improved. It is essential to understand that process control
boundaries only relate to the ability to discern numerical differences. They are not equivalent to speci-
fication limits that describe the conformance/non-conformance boundaries. Statistical significance is
also not the same as practical significance, which is a difference that has a meaningful impact on the
process. Depending on the circumstances of the process, when there are sufficient samples present, it
is possible to detect statistical differences that have no practical importance. It is also possible in pro-
cesses with tight specification limits and noisy measurement systems for it to be challenging to detect
important performance changes. The users of process control tools must apply their understanding of
the process when evaluating the relevance of process control detections.

Technical Report No. 59 © 2012 Parenteral Drug Association, Inc. 3

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Statistical Process Control Limits (cont.)

Stable Process (Statistically In Control)
Stable Process (statistically in control) is a process that is consistent and predictable. It does not ex-
hibit special (or assignable) cause variation as identified through the use of statistical techniques.
The variation present is due to only common cause variation.
Figure 2.0-3 below shows a stable process (statistically in control). The distribution of the process
is consistent over time (as shown on the left). Each successive outcome is random, but aligns with
the expectation for the process.
Figure 2.0-3 Examples of a Stable Process (Statistically In Control)

Stable Proces (In-Control) 58.8

UCL=58.02
Prediction
57.5
56.5
55.5
Mean=54.94
54.5
53.5
52.5
LCL=51.87
51.5
0 5 10 15 20 25 30 35 40 45 50
Week

Unstable Process (Statistically Out of Control)

Unstable process (statistically out of control) is a process that is not consistent or predictable. It ex-
hibits special cause variation as identified through the use of statistical techniques. This is also some-
times called “Out of Trend”, where the current performance is detected to be not part of the prior (level,
uniform) trend. Figure 2.0-4 below illustrates an unstable process. The distribution of the process
is not consistent over time (as shown on the left). Each successive outcome is random, but does
not align with the expectation for the process.
Figure 2.0-4 Examples of an Unstable Process (Statistically Out Of Control)

UnStable Proces (Out-of-Control) ?

65

Prediction

60 UCL=59.78

Mean=55.96
55

LCL=52.15

50
0 5 10 15 20 25 30 35 40 45 50
Week

4 © 2012 Parenteral Drug Association, Inc. Technical Report No. 59

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Statistical Process Control Limits (cont.)

Warning or Alert Limits
These types of limits are determined and used typically during development, environmental or
manufacturing process monitoring for the purpose of detecting trends or to get a better under-
standing of the manufacturing process and its consistency. These limits are usually not intended
to require a formal investigation, a report or an involvement of Quality Assurance in case of
detection. In some instances (e.g., microbiological monitoring) a Standard Operating Procedure
(SOP) can define follow-up steps that may include actions, such as increased frequency of testing
or possibly proactively increased actions to mitigate risk of future impact.

Variability
Variability is categorized here by two types; i.e., common cause and special cause. It is important to recog-
nize the distinctions when understanding a process and taking the appropriate action to improve the process.
Common Cause Variability
Common cause variation is the result of the combination of all of the typical variability in the
materials, process and measurement system. These many small components of variation are
expected to be present in the manufacturing process.
Special Cause Variability
Special cause variation is a change caused by special circumstances not expected from the process.
These occurrences are not predictable, and may come and go sporadically. Special cause events
are detectable when compared to statistical control limit techniques, hence the term used to de-
note this condition is described as “out of statistical control” (limits). The prompt detection of
special causes allows the proper investigation and evaluation of impact to subsequent production
or downstream processes.
Trends
Trends are changes in the average or variability of the result. An analysis of data often exhibits
an ongoing upward or downward pattern that is not due to random noise. Analyzing trends is
useful in detecting patterns that could lead to future quality problems, and in anticipating upcom-
ing performance. Statistical process monitoring tools as defined in this document can be used to
detect trends using objective numeric tools.
In the interpretation of GMP requirements, regulatory authorities increasingly request an assess-
ment of trends in inspections. The prompt detection and evaluation of trends (3) supports the
implementation of corrective and preventive actions (CAPA) as suggested by ICH Q10 (4). There
is always an uncertainty if a trend is relevant to product quality. When establishing monitoring
practices, the potential impact of process shifts should be considered when balancing risks of
failure to detect potential hazards versus the risk of annunciating unimportant changes. “The
level of effort, formality and documentation of the quality risk management process should be
commensurate with the level of risk” described as the second principle of ICH Q9 (5).
Some examples of trends are shown1 in Figures 2.0-5 and 2.0-6.
Points beyond control limits (not ‘specification limits’) which are isolated high or low points.

1 There are collections of rule sets such as “Western Electric Rules” designed to distinguish certain features in
data

Technical Report No. 59 © 2012 Parenteral Drug Association, Inc. 5

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Variability: Trends (cont.)

Figure 2.0-5 Examples of Trends (Points Beyond Control Limits)

Points Beyond Control Limits

Upper Control
Limit (UCL)

Average

Lower Control
Limit (LCL)

If all points lie within the control limits, there are still features that may be of process interest. Erratic ups and downs in
groups of points with sparse values in middle.

Figure 2.0-6 Examples of Trends (Points within the control limits)

All points lie within the control limits

Upper Control
Limit (UCL)

Average

Lower Control
Limit (LCL)

Variation small compared to control limits – where almost all of the points are within ne-third (of the distance between the
control limits) of the centerline.

Small Variation

Upper Control
Limit (UCL)

Average

Lower Control
Limit (LCL)

6 © 2012 Parenteral Drug Association, Inc. Technical Report No. 59

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Variability: Trends (cont.)

Sudden shift in level where the points seem to move to a new average over a short period of time.
Sudden Shift in Level

Upper Control
Limit (UCL)

Average

Lower Control
Limit (LCL)

A cycle produces a pattern of up and down points, as though the values of the points were time dependent.

A Cycle Produced

Upper Control
Limit (UCL)

Average

Lower Control
Limit (LCL)
There are several systems of “rules”, such as in the examples listed above, that can be applied to de-
tect non-random conditions. These may present opportunities to increase process understanding by
allowing appropriate reaction. Two of the more common sets include the Western Electric rules, and
the Nelson Rules. For every additional criteria applied to the evaluation of data, there will be more
“false positive” detections. It is, therefore, appropriate to evaluate the potential benefit, and apply only
the rules that are relevant for that process situation. The use of rules helps ensure a consistent thresh-
old of detection, and can be a useful mechanism to annunciate potential events for review. These
rules are typically evaluated as each new data point is added; they typically reference regions which
are defined as standard deviations around the process mean.

Technical Report No. 59 © 2012 Parenteral Drug Association, Inc. 7

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

3.0  Statistical Process Control Tools

The following sections introduce some of the more common statistical methods for monitoring man-
ufacturing processes. Most of these methods involve plotting the process data or a statistic calculated
from those data on a chart. The vertical axis of the chart represents the dependent range of values in
the process data or the statistic calculated from the process data. For many of these tools, the hori-
zontal axis represents the independent timing of the data being plotted. It is important to arrange and
then plot the data in “time-order,” as close as possible to the date and time the data were actually gen-
erated. Control charts are often plotted by date manufactured to better illustrate potential patterns
in production, review by date tested may be appropriate for monitoring measurement systems. Since
the primary purpose of process monitoring is to ensure that the process remains in a state of statisti-
cal control, most of the methods include the addition of limits or numeric boundaries that provide
one way of assessing control. All of the charts can be examined for data patterns which may indicate
when a process is not in a state of statistical control. There are several statistical software packages
that create the charts described, and even when these are used, an understanding of the underlying
principles will aid in the appropriate selection and application of statistical tools (see Table 3.0-1). For
each method, typical applications, technical details and the pros/cons are presented (see Table 3.0-2).

3.0.1 Prerequisites for Data Analysis

Any meaningful statistical evaluation has several prerequisites:
• The data integrity is assured and the measurement system is acceptable. The analytical method creating the results
assures a valuable set of data (e.g., accuracy, precision, repeatability, specificity, detection limit, and quantitation
limit are understood and maintained).
• The attributes where statistics should be applied are meaningful for the expected information. The data describe
critical process parameters (CPP) or critical quality attributes (CQA) (see also ICH Q8/Q11 and Q10) (4,6,7).
• The level of effort, formality and documentation in the use of statistical tools should be commensurate with the
level of risk (based on ICH Q9) (5).
• Results should maintain the order of production, so that trends which evolve over time are apparent.
• The means by which a sample is taken and measured are defined so that the sample accurately highlights important
features of the process. The design of the sampling approach should ensure that relevant sources of variability
can be detected. If there are potentially important differences between multiple samples, consider strategies for
grouping. This is described as the “rational subgroup.” To effectively select groups, one must also have reasonable
knowledge of the process and an understanding of the sources of variability. The selection of a rational subgroup
will maximize the chance of detecting differences between groups and minimize the differences within a group.
Groupings may support comparisons such as:
• Lot to Lot – common in parenteral manufacturing where a “batch” may be a certain (homogeneous) volume of
liquid progressing through manufacturing steps.
• Time to Time – this compares processing at one time to another, possibly within a batch. This may be samples
within a lot (e.g., verifying homogeneity of liquid filled vials during a fill) or across broader divisions (e.g., between
manufacturing campaigns).
Comparisons could also evaluate variability by comparing measurements within a item, across items, between
production lines, plants or products.

8 © 2012 Parenteral Drug Association, Inc. Technical Report No. 59

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Table 3.0.1-1 Suggestions and Proposals on When to Use Statistical Tools

The chart uses the following notation: + + for “preferred tool”; + for “useful”; o for “could be used”;
− for “not to recommended”; − − for “ not useful”; () for “depending on the individual process”.
These considerations do not create new requirements.

Control Charts
Moving Range
Control Charts
Distribution
Run Charts

Histogram
Areas of potential Characteristics for

Capability
Individual

Charts**
EWMA*
Process
implementation statistical purposes

CuSum
Data
Pharmaceutical Few lots/batches
process development under the same + + – – – –
process conditions

PAT Real time data

+ ++ + +
implementation

During Commercial Historic data available

++ ++ ++ ++ ++
manufacturing

Stability studies Few data over the

–– ––
years only

Process Validation Large data pool from a

+ o +
few lots/batches only

Continuous process Continuous big data

++ (+) (+) + ++
verification pool

In process control (IPC) Back log analysis (1d) ++ o + ++

Yield Trends Can be noisy, shifts in

average are relevant to + ++ +
business

Microbiological testing Time lag to any analysis

(1 week); no NON- Requires statistics suited for non normal distributions.
NORMAL distribution in Non-parametric tools are outside scope of this document.
the material

Conventional release Back log analysis (1

+ o (+)
testing week)

Real time release Lot of data in a short

+ +
testing time

Annual Product review Historic evaluation ++ ++ ++ ––

* Exponentially Weighted Moving Average (EWMA)

** Cumulative Sum (CuSum)

Technical Report No. 59 © 2012 Parenteral Drug Association, Inc. 9

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Table 3.0.1-2 Areas of Potential Implementation

Awareness of statistical
Statistical tool Strength Purpose
challenges

Run Charts Mapping of data • Look for evidence of Interpretation of trends

patterns in process data

Data Distribution Assess and compare Assess and compare properties Distribution alone assumes
properties of distributions of distributions, such as: the process is not shifting and
1) Where sample values are gives no indication of time
trends.
centered
2) Whether a sample distribu-
tion is symmetrical or skewed
3) Whether sample data follow
a specific distribution
4) How many peaks exist in the
sample distribution (more than
one peak can indicate that
data are from multiple popula-
tions)
5) What the most commonly
observed values in the sam-
ple are

Individual Control Detect the presence of special • Track the process level and Only detect subtle shift with
Charts causes detect the presence of spe- rule sets that evaluate a series
cial causes of points together.

Moving Range Track process variations • Track the process variation Range between successive
Control Charts and detect the presence of points is a coarse estimate of
special causes process variability.
Easy to do by hand or on the
manufacturing line.

S Charts Track process variations • Track the process variation Standard deviation is more ro-
and detect the presence of bust than Range against extreme
special causes values and outliers in the sample.
• Becomes preferred over the R Best done with a calculator or
chart when subgroup sample computer.
sizes reach 8-10. S chart is
preferred for sample sizes
>10. For smaller sample
sizes the R and S chart give
similar results.
• Should be used for variable
sample sizes.

Histogram Visualization of statistical • Show distribution Size of categories needs to be

anomalies • Examine the shape and spread appropriate for the data.
of sample data

10 © 2012 Parenteral Drug Association, Inc. Technical Report No. 59

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Awareness of statistical
Statistical tool Strength Purpose
challenges

Process Capability Performance evaluation • Continual improvement Risk to over interpret statistical
out of control as related to unac-
Determine if it is capable, and • Determine capability by com-
ceptable product quality
that is meeting specification lim- paring the width of the pro-
its and producing “good” parts. cess variation with the width
(Process needs to be in control of the specification limits.
before assessing its capability;
if it is not, then result will be
incorrect estimates of process
capability.)

EWMA A type of time-weighted control • Monitor in-control processes Dampens variability of single
chart that plots the exponentially for detecting small shifts events. Has a lag when dis-
weighted moving averages away from the target playing step changes.

CuSum Charts Real time trend analysis • Early warning on selected Risk to over interpret statistical
CQA/CPP out of control or a discernible
A type of time-weighted control
mean shift as related to product
chart that displays the cumulative • For detecting small shifts
quality
sums of the deviations of each away from the target
sample value from the target • Pinpointing the particular time
value when a small shift occurred

3.1 Run Charts

Run Charts are a simple visual monitoring method. Data or statistics calculated from measurements
are plotted on a chart in “time-order,” as in Figure 3.1-1 Technical details can be found in Section 5.1.
Figure 3.1-1 Run Chart

6.8
6.7
6.6
Quality Attribute

6.5
6.4
6.3
6.2
6.1
6
5.9
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Lot (by mfg date)

Technical Report No. 59 © 2012 Parenteral Drug Association, Inc. 11

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

3.1.1 Typical Applications

Run charts should be the first tool applied for monitoring all data in sequence. Once a reasonable
number of data points are collected, and the distribution of the data is reviewed (see Section 3.5 on
histogram), the run chart may be converted into one of the other control charts below.

3.1.2 Pros
• Simple to create
• Easy to visually identify general patterns in the data
• Do not assume a distribution

3.1.3 Cons
• Variability is not considered in assessing process control

3.2 Control Charts: Individuals

Individual Charts (I Charts) are used when only single, or individual, numeric measurements are used
to evaluate a Quality Attribute (see Figure 3.2.1-1). The I Chart should be used in tandem with the
Moving Range Chart described in Section 3.3. If the variability of a Quality Attribute is not in a state of
statistical control, the value of an I Chart is questionable. Technical details can be found in Section 5.2.

3.2.1 Attribute Control Charts

Attribute charts can be used for data that represent counts or proportions of a classification as op-
posed to measurements. A common application is evaluating multiple instances of conformance tests
applied to elements within a group (e.g., 12 of 30 units that are nonconforming or a nonconforming
proportion of 40%). Because the counts or proportions of these types of measures generally exhibit
a non-normal distribution, different methods need to be used to establish limits. The generation of
control charts based on non-normal distributions such as the “c chart” (for counts) or the “p chart”
(for proportions) is outside of the scope of this document.
Figure 3.2.1-1 Individual Control Chart

7.0
6.9
6.8
Quality Attribute

6.7
6.6
6.5
6.4
6.3
6.2
6.1
6.0
5.9

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Lot (by mfg date)

12 © 2012 Parenteral Drug Association, Inc. Technical Report No. 59

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

3.2.2 Typical Applications

The Run Chart above is an individuals chart before adding control limits. An I Chart is used to moni-
tor attributes where there is a single numeric measurement for each lot.

3.2.3 Pros
• Simple to create
• Easy to visually identify general patterns in the data
• Variability is considered in assessing process control

3.2.4 Cons
• Variability is based only on the change between individual data values potentially captured from each lot. That
variability may include multiple sources.

3.3 Moving Range Control Charts

Moving Range (MR) Charts as shown below in Figure 3.3-1 (also called ‘moving average charts’),
are used in combination with an I Chart to evaluate the variability of a Quality Attribute. Technical
details can be found in Section 5.3.
Figure 3.3-1 Moving Range Control Chart

0.60

0.50
MR

0.40

0.30

0.20

0.10

0.00
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Lot (by mfg date)

3.3.1 Typical Applications

The MR Chart is essentially an I Chart for monitoring the variability of a Quality Attribute. Although
this chart is sometimes omitted, it should be used in tandem with the I Chart. If the variability of a
Quality Attribute is not in a state of statistical control, the value of an I Chart is questionable.

Technical Report No. 59 © 2012 Parenteral Drug Association, Inc. 13

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

3.3.2 Pros
• Simple to create
• Easy to visually identify general patterns in the data
• Variability is considered in assessing process control

3.3.3 Cons
• Variability is based only on the change between individual data values potentially captured from each lot. That
variability may include multiple sources.

3.4 Average and Variability Charts

These charts are used to monitor the averages of small sets of data. Average and range charts take
advantage of the Central Limit Theorem, which states that average results will tend to be normally
distributed, regardless of the parent distribution. Of course, the larger the sample set, the greater the
state of normality. The range of results within each data set is used to estimate overall variability. The
difference between the sample average and the overall average is compared to the average range of
results within each sample to assess statistical control. Variability is plotted on a separate chart.

Process variability can be monitored by either a Range chart (R chart) or the standard deviation chart
(S chart). If the Variability Chart (Figure 3.4-1) indicates a statistically out of control situation, the
Average Chart may not be meaningful. Technical details can be found in Section 5.4.
Figure 3.4-1 Average and Range Chart
XBar & R chart of Variable
Average (Variable)

7.0
6.5
6.0
5.5
5.0
4
Range (Variable)

3
2
1
0
2 4 6 8 10 12 14 16 18 20 22 24 26 28 30
Lot (by mfg date)

3.4.1 Typical Applications

Average and Range charts are used in a wide range of applications.

3.4.2 Pros
• Simple to create
• Lessens concerns about non-normal data
• Easy to identify general patterns in the data
• Variability is considered in assessing process control

14 © 2012 Parenteral Drug Association, Inc. Technical Report No. 59

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

3.4.3 Cons
Variability is based only on the change between individual data values potentially captured from each
lot. That variability may include multiple sources.

If there are enough “within” batch samples to estimate standard deviation, the trend of these devia-
tions on an S chart can be a better measure of change in variability than moving range charts.

3.5 Histograms
A histogram is a graphical representation showing how often each value in a set of data occurs. A
histogram is used to understand the distribution of a data set by graphing rectangles to represent the
count of observations in a data set over the class intervals. The height of the rectangles is equal to the
number of observations in that class. The histogram, or distribution graph, is often an early technique
to be applied in analyzing a data set and may aid in the selection of statistical techniques for further
analysis.

This method is useful in analyzing processes to understand the distribution of process outputs. This
can be used to monitor changes to processes as well as variations from one time period to another by
comparing multiple histograms. The method is also valuable as an initial exploration of a data set in
order to better understand data and infer the distribution of the entire population.

The histogram may also be plotted using a relative frequency distribution in which the number of
observations in each class is divided by the total number of observations (Figure 3.5-1). In this case,
the vertical axis will be the percentage and the rectangles will represent a percentage of the total data
set for each class interval. Technical details can be found in Section 5.5.
Figure 3.5-1 Example of a Typical Histogram Showing Data Location in Relation To Specification Limits
LSL=20.0, Nominal=40.0, USL=60.0

18

15

12

0
10 20 30 40 50 60 70

Several different characteristics can be observed from a histogram.

Technical Report No. 59 © 2012 Parenteral Drug Association, Inc. 15

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

3.5.1 Typical Application

This method is useful in analyzing discrete and continuous data sets. It is an important early step in
analyzing a data set or distribution.

3.5.2 Pros
This method allows visualization of where values fall on a measurement scale in relation to the fre-
quency with which they occur. The histogram can summarize large data sets graphically.

3.5.3 Cons
Observation of a single histogram does not give any indication of any changes over time in the data
points. For example, a process that is drifting may appear as a wider distribution, without a clear indi-
cation that a drift is underway.

3.5.4 Distributions – Interpretation of Histograms

The standard normal distribution is 1.0 where the distribution of values has been standardized so that
the mean is 0.0 and the standard deviation is 1.0 (Figure 3.5.4-1). The following graph provides the
theoretical percentage of results that fall between + 1, + 2 and + 3 standard deviations around the
mean of 0:0.

Perfectly normal distributions are not found in actual practice. The percentages shown are only ap-
proximate for an actual data set or distribution.
Figure 3.5.4-1 Normal Distribution

68.26%

95.44%
99.73%

-3 -2 -1 0 1 2 3
Number of Standard Deviations

In the case of a normal distribution, the counts of data points are equally distributed around the mean
or average value and have a specific contour. The assumption of approximate normality is common
for data monitored with many of the charts presented here. In theory, when the data are normally
distributed, 99.73% of results are expected to fall with + 3 standard deviations of the mean. The 3
standard deviations in one direction represent a probability limit of 0.00135, or in both directions
0.0027, that a value will fall outside these limits by chance alone. For example, control limits set at
these ranges would rarely (in this case = 3 out of 1000) falsely annunciate a point as out-of-statistical-
control limits, when the process performance remained typical. This reasoning of a probability limit

16 © 2012 Parenteral Drug Association, Inc. Technical Report No. 59

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

can be expanded to non-normal distributions, such as the Poisson or the binomial distribution of at-
tribute data (P, nP charts) with the appropriate mean and variance, and to more powerful statistical
tools such as Prediction and Tolerance Intervals. The choices used for control limits should be based
on a balance of the risk implications of failing to detect a real difference (limits are too wide), and
falsely detecting a difference that isn’t there (limits are too tight).

The interpretation of the below graph (Figure 3.5.4-2) is that the data set is approximately nor-
mally distributed. It can infer that the population is approximately normally distributed and has a
bell-shaped curve.
Figure 3.5.4-2 Histogram of a Normal Distribution

20

15
Count

10

6.9 6.95 7 7.05 7.1

In case of a bimodal distribution, the data suggest that there may actually be two distinct populations
in the data set, as in Figure 3.5.4-3. This could be as simple as the data were a measurement of the
output from two systems. Further exploration of the data set is required in order to understand the
nature of the variation in the sample set. Depending on the size and nature of the data set, there may
be multiple modes or means within the data set.
Figure 3.5.4-3 Bimodal Distribution

25
Count

15

7.2 7.3 7.4 7.5 7.6 7.7 7.8 7.9 8

In case of a skewed distribution, the mean is skewed to one side of the distribution, as in Figure 3.5.4-4.

This type of distribution is sometimes encountered when a system is unable to vary in one area or
side of the range (constrained), but remains able to vary in another area or side of the range.

Technical Report No. 59 © 2012 Parenteral Drug Association, Inc. 17

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Figure 3.5.4-4 Left Skewed Distribution

40
Count

20

3.2 3.3 3.4 3.5 3.6 3.7

3.5.5 Hints for Use

The histogram provides an overview of the distribution of a data set. It can be performed manually
with graph paper or generated quickly using basic spreadsheet or statistical programs. After initial
generation of the histogram, it is useful to plot the specifications of process output on the graph to
get a rough visual estimate of the capability of the process.2

After generating the histogram, it is important to explore any variation from a visually normal distri-
bution. Changes to the shape of the distribution can be caused by natural limits or by the grouping of
several different populations. Potential differences from normality should be understood prior to use
of a “standard” deviation to calculate ranges. After a review and understanding of the histogram, the
data can be explored by more sophisticated techniques.

3.6 Process Capability (Cpk, Ppk)

This section illustrates the calculation of the process capability indices Cpk and Ppk. These indices are
defined as the ratios of the specification range to the natural variability of the data. They answer the
question: “Is this process capable of meeting its limits now or in the future?” The variability in the data,
which is compared to the limits, can be based on an estimate of the process variability. The variabil-
ity is expressed as the average within the group variability (Cpk) or as the variability of the observed
points, often expressed as the standard deviation of all data (Ppk). It is worth noting that for a process
which is statistically in control, Ppk is essentially equal to Cpk. Graphical examples of the relationship
between variability and limits are illustrated in Table 3.6.1-1. Technical details can be found in Sec-
tion 5.6.

3.6.1 Assumptions
The current specifications are realistic, relevant to the potential for product impact, and were estab-
lished appropriately. The data are approximately normally distributed. The data collected are truly
representative of the process and were obtained from an independent sample. Finally, measurement
variability is small or a small percentage of the process variability. Use of the Cpk index requires that
the process is in a state of statistical control, with consistent variability between groups.

2 Excel histograms are not calculated the same way as is done in statistical packages where some basis for the bin
width is applied. Proper techniques for histogram construction are outside the scope of this document.

18 © 2012 Parenteral Drug Association, Inc. Technical Report No. 59

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Table 3.6.1-1 Significance of Cp

Value of Cp Cp = 0.5 Cp = 1 Cp = 3

Graphical view
of a process at
different values
of Cp

Sigma value in 3s 6s 18 s
the given limits (±1.5 s) (±3 s) (±9 s)

Statistical
number of
13.58 % 0.27 % Really 0
values outside
the limits

Statistical
number of
86.42 % 99.73 % > 99.999999 %
values inside
the limits

Process statistically Process statistically unlikely to make non conforming

Result expected to routinely make product
nonconforming product

3.6.2 Typical Applications

These indices are one tool for studying processes and comparing relative risk between various mea-
sures. The interpretation of capability index values is shown in Table 3.6.2-1. Process capability indi-
ces should be used with other techniques, such as histograms, run charts and control charts, to assess
the process. They should not be used to accept or reject lots of a product.

Technical Report No. 59 © 2012 Parenteral Drug Association, Inc. 19

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Table 3.6.2-1 Interpretation of Cp Regarding Limits Taken

Value of Cp Result Comment

Remedy process to decrease failure rate:

Process is statistically poorly capable for
Cp < 0.6 Assess process design, measurement sys-
these limits
tems, and limits as appropriate

The variation is wide in relation to the limits.

Process statistically modestly capable for
Cp < 1.0 Occasional deviations can be expected from
these limits
the process as is. Remedy as appropriate.

Cp ≥ 1.0 Warn limit of 1.0 Tolerance of 6 Sigma (99.7%) is achieved.

Process statistically capable, likely to pass

1.0 ≤ Cp ≤ 1.33
specifications routinely.

Cp ≥ 1.33 Process statistically very capable

Consider if the limits are relevant to the pro-

cess (product impact basis). It may be that
Process is much better than needed to meet
Cp ≥ 3 a risk based review would indicate that the
the limits
process is adequately controlled without on-
going measurement

3.6.3 Pros
Because process capability indices are not based on any specific units, they can be used as a common
basis to compare between processes or measurement types, and among manufacturing steps and
products.

3.6.4 Cons
Observation of a single calculated capability index does not give any indication of change over time
of the included data points. A process that is drifting may show a diminished capability, without clear
indication that a drift is underway. For a precise estimate of capability, as with any summary statistic,
considerations should be made for sample size using a lower confidence bound (as in ASTM E2281).
With sample sizes less than 200 data points, consider comparing a tolerance interval of the process
data to the specifications.

3.7 Exponentially Weighted Moving Average Charts

Exponentially Weighted Moving Average (EWMA) charts are used to display the central tendency
of a series of data points (Figure 3.7-1). The variability of individual values is dampened, allowing
any underlying drift to be more easily observed. The weighted averaging favors the most recent data
point, with prior points having a decreasing influence as distance increases. Technical details can be
found in Section 5.7.

20 © 2012 Parenteral Drug Association, Inc. Technical Report No. 59

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Figure 3.7-1 Data Plot Compared to EWMA Chart

Sample Data Plot EWMA Plot

14 11.5
13
11.0
12
10.5
11
10 10.0
9
9.5
8
9.0
7
6 8.5

0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Sample The same data, plotted as Sample
EWMA, with l=0.3

3.7.1 Typical Applications

The EWMA chart may be used for variable or attribute data. This chart is able to highlight small
changes in mean, but may not react as quickly to large changes or brief transient events. The net
effect of the EWMA algorithm is the same as a first-order filter or dampening of individual values.

3.7.2 Pros
The EWMA chart is useful for summarizing trends. It is also helpful for monitoring if a relatively
noisy measurement system obscures the underlying process. This chart is also robust to non-normal
data, due to the averaging across multiple points.

As successive points are designed to influence each other, this chart will make common-cause varia-
tion more visually apparent.

This moving average may also be relevant for tracking key performance indicators, where the overall
business outcome is dependent on the average result (e.g., production rates or yield).

The EWMA chart is sensitive to small shifts in mean, and can display them with higher resolution
than plotted raw data. The prompt detection of process mean changes can allow response to correct
prior to a subsequent excursion. The CuSum chart (described in the following section) may be slightly
more powerful for annunciating mean shifts, but it is less intuitive to interpret. The EWMA chart has
the comparative advantage of being in the same units as the measurement.

3.7.3 Cons
The EWMA chart is not helpful for detecting changes in variability. This chart is designed such that
the variability of individual points is reduced, so as to not distract from observation of the central
trend of the data.

The dampening of the EWMA causes a lag in the response. The magnitude of a step change will only
become apparent after a number of samples appear at the new average.

Technical Report No. 59 © 2012 Parenteral Drug Association, Inc. 21

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Individual special cause events are not highlighted by the EWMA chart. Other means of detection of
outlying individual results should be employed (e.g., alert/action limits, individuals chart, etc.).

3.8 CuSum Charts

Chart of the cumulative sum (CuSum) represents a quality control chart with a memory. As samples
accumulate, these charts total the sum of the deviations from a given specification value.

This approach is different than most other quality control charts (e.g., x-bar charts) that map indi-
vidual results on a defined ‘window’ of ranges. On these charts, previous values are not taken into
account when trying to detect unusual events.

The CuSum chart is the most sensitive for detecting changes in mean. It highlights changes in mean
as a shift in slope (Figure 3.8-1). Technical details can be found in Section 5.8.
Figure 3.8-1 Run Plot Compared to CuSum Chart of the Same Data
4
3 U 40
Cumulative Sum of Data

2
1 30
Data

0 A 20
-1
-2 10
-3 L 0
-4
10 20 30 40 50 60 70 80 90 100 10 20 30 40 50 60 70 80 90 100
Sample Sample

3.8.1 Process Features Suitable for This Type Method

CuSum charts should be implemented as a statistical tool at the shop floor, where processes must be
monitored and guided.

This method can be considered to use for all types of data recording, where an understanding of
changes in the average is important. Such data could be taken from any monitored value, such as
process parameters from in-line, at-line or off-line recording.

3.8.2 Typical Applications

Generally CuSum charts are used for the detection of anomalous behavior so they are implemented
and typically used for:
• Monitoring and detection of a change
• Identify changes in the probability distribution of a stochastic process
• Taking the advantage to stop processes as soon as significant trends are observed

3.8.3 Pros
The CuSum charts represent a sequential analysis technique. The sample size is not fixed in advance
and the data are evaluated as they are collected. The early detection of a change in mean sometimes
allows for the initiation of corrective actions at a much earlier stage at consequently lower costs.

22 © 2012 Parenteral Drug Association, Inc. Technical Report No. 59

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

The location of the inflection point provides a valuable clue as to which time frame to review for
potential changes.

3.8.4 Cons
There is a potential for over-interpretation of the prompt detection of statistically significant differ-
ences as relevant to product impact. The presence of a discernible difference in mean does not imply
that specifications have been exceeded. Consider the risk that CuSum charts provides an alert system
with potential to overreact.

3.9 Examples of Efficient Mixture of the Statistical Toolbox

As an example, CuSum Charts have been used in a multi-product manufacturing facility of Active
Pharmaceutical Ingredients (APIs). Many different products and conditions are manufactured e.g.,
larger campaigns with good knowledge of the manufacturing process and small campaigns with less
historical data. The grouping of an x-bar chart, histogram and CuSum charts combined into the same
display have been employed to efficiently convey process performance (Figure 3.9-1).
Figure 3.9-1 Example of a Tool Bar for Statistical Control of a Process

I Chart with specifications Historgram with process

and 3 sigma levels capabilities (Cpk)

CuSum Chart

Technical Report No. 59 © 2012 Parenteral Drug Association, Inc. 23

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

4.0 Acceptance Sampling

Acceptance sampling is the process of taking a representative sample from a lot, which may be com-
posed of distinct individual units and then making a decision about the disposition of that lot. The
sample is inspected (i.e., measured, examined or tested) and compared to the requirements. The deci-
sion is to reject or accept the lot, based on the conformance to specified quality levels of the number
of defects detected in the sampled units. Defects are defined as a departure of quality characteristics
from intended level. Technical details can be found in Section 5.9.

4.1 Typical Applications

Acceptance sampling is used to decide whether or not the lots, on average, are likely to be acceptable.
This will ensure against the release of highly defective lots as well as assure that the average quality
going to customer is at or better than some specified level. Sampling will protect customers from
exposure to high defect levels and reject obviously bad lots. Acceptance sampling is also often applied
on incoming materials.

A sampling approach may be appropriate for incoming material inspections, in-process components,
and possibly finished goods inspections where 100% inspection is not required. Sampling can be ap-
plied across a supplier/customer junction or between steps or business sections within a company.
Sampling may also be applied at line or in-line as a part of ongoing monitoring during production.
Sampling approaches are clearly required for unit attributes that can only be tested destructively.
When sampling is considered, the costs of sampling should be balanced against the hazard presented
by defective components. The selection of samples and subgroups that are required for consideration
are as previously described in the prerequisites for data analysis in Section 3.0.

4.2 Key Terms

Sampling is a complex topic in it of itself and formal training is highly recommended. This document
is only intended to provide an overview of general concepts to put sampling into context with other
statistical tools. Further resources (e.g., ANSI/ASQ Z1.4-2008: Sampling Procedures and Tables for
Inspection by Attributes and/or ANSI/ASQ Z1.9-2008: Sampling Procedures and Tables for Inspec-
tion by Variables for Percent Nonconforming) should be studied for further details or guidance in
developing any specific plan.3

4.2.1 Acceptable Quality Limit (AQL)

The AQL represents the worst tolerable quality level in a continuous series of lots that can be consid-
ered to be acceptable as a process average by the customer. AQL-based sampling plans are designed to
have a high probability of accepting lots at a given AQL.

4.2.2 Rejectable Quality Level (RQL)

Sometimes called Lot Tolerance Percent Defective (LTPD), Limiting Quality (LQ) or Unacceptable
Quality Limit (UQL); it is the highest percentage of defective units in any individual lot before it is
considered unacceptable by the customer. RQL-based sampling plans are designed to have a high
probability of rejecting lots at a given RQL.

3 Much has also been written outside of the pharmaceutical industry on selecting sample sizes, for further
discussion also see ASTM E2587 “Use of Control Charts in Statistical Process Control”, ASTM E2281-08a
“Standard Practice for Process and Measurement Capability Indices”, ASTM E2709-09 “Standard Practice for
Demonstrating Capability to Comply with a Lot Acceptance Procedure”, or other statistics texts as appropriate.

24 © 2012 Parenteral Drug Association, Inc. Technical Report No. 59

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

4.3 Types of Sampling

There are different sampling approaches can be applied to both attributes and variables.

4.3.1 Attributes Sampling

Attribute Sampling is used on discrete outcome type of data (go /no go gauge, missing parts in an assem-
bly, individual units under/oversized). Count type of data (e.g., number of under- or over-filled tablets in
a bottle) is also an attribute sampling. The attribute case is the most common for acceptance sampling.

4.3.2 Variable Sampling

When each sample is measured on a continuous scale, the sampling type is denoted as variable sam-
pling. Examples might be weight, torque, potency, or moisture. The variable measurement provides
much more information with a smaller sample size than attribute sampling. Variable sampling plans
assume that the data is normally distributed.

4.4 Types of Acceptance Plans

Different sampling plans can be used to balance test costs against complexity and risk of error.

4.4.1 Single Sampling Plans

One sample of items is selected at random from a lot and the disposition of the lot is determined from
the resulting information. These plans are usually denoted as (n,c) plans for a sample size n, where
the lot is rejected if there are more than c defectives. These are the most common, and easiest, plans
to use, although not the most efficient in terms of average number of samples needed. There are two
widely used ways of picking (n,c):
• Use tables (Standards) that focus on either the desired AQL or the RQL (LTPD)
• Specify two desired points on the Operating Characteristics (OC) curve and solve for the (n,c) that uniquely
determines an OC curve going through these points.

4.4.2 Double Sampling Plans

After the first sample is tested/inspected, there are three possibilities:
1. Accept the lot
2. Reject the lot
3. No decision (extend samples)

If the outcome is (3), a second sample is taken. The procedure is to combine the results from both
samples and make a final decision based on the information.

4.4.3 Individual Sampling Plan

A specific plan is the sample size and the accept/reject numbers.

4.4.4 Sampling Scheme

A sampling scheme consists of a normal sampling plan, a tightened sampling plan, a reduced sam-
pling plan, discontinuation and rules for switching from one to the other.

4.5 Pros and Cons

Sampling is a means of adjudicating quality at a certain step. When properly applied, it allows a

Technical Report No. 59 © 2012 Parenteral Drug Association, Inc. 25

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

likelihood-based discussion of measurement costs and risk trade-offs to be made at a specific process
juncture. This also allows a rational basis when deciding on appropriate levels of destructive testing.

Acceptance sampling is not a means of assuring defect-free outcomes. It does not assure an exact
quality level is present, as all detections of intermittent events are based on likelihoods, and there is
inherent randomness to any particular detection outcome. The value of acceptance sampling is as
an independent guard against the catastrophic failure of the testing of in-process materials and drug
products. It monitors the output and validates the performance of those manufacturing processes
that may be responsible for causing variability in the characteristics of in-process materials and the
drug product. Acceptance sampling may be costly both in terms of materials and time. It should not
be used as an alternative to process monitoring and improvement.

26 © 2012 Parenteral Drug Association, Inc. Technical Report No. 59

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

5.0 Appendices: Technical Details and Examples

These appendices relate to the previous sections on statistical tools. Each section in the body of the
text above has a corresponding section in these appendices, which provides technical details along
with examples.

5.1 Run Charts

The following detailed instructions include a step-by-step method for calculating initial control limits
as a process and then refining them as more data become available and the process becomes routine.
This illustrates one particular approach using the moving range; other approaches such as multiple
of standard deviations (sigma) and choices for numbers of lots to incorporate in averages are valid as
well and may be employed at the user’s discretion.

5.1.1 Technical Details

Step 1 — Place data or statistics in time-order.
Step 2 — Create a plot of the data with the range of data values on the dependent vertical axis and
the timing of the data on the independent horizontal axis. The plot should be created as soon as the
first datum value is available.

Step 3 — Connect plotted points with lines.

Step 4 — Add subsequent data values and connect to the previously plotted point as the data become
available.

5.1.2 Example
Step 1 — Place data or statistics in time-order. The following data in Table 5.1.2-1 were collected
from individual measurements of 20 independent lots. The lots are ordered by manufacturing date:
Table 5.1.2-1 Lot Data by Manufacture Date

Quality Quality
Lot Lot
Attribute Attribute
1 6.23 11 6.29
2 6.50 12 6.61
3 6.59 13 6.38
4 6.26 14 6.44
5 6.47 15 6.31
6 6.23 16 6.25
7 6.41 17 6.54
8 6.32 18 6.37
9 6.44 19 6.46
10 6.52 20 6.32

Technical Report No. 59 © 2012 Parenteral Drug Association, Inc. 27

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Step 2 — Create a plot of the data with the range of data values on the vertical axis and the timing
of the data on the horizontal axis (Figure 5.1.2-1). The plot should be created as soon as the first
value is available.
Figure 5.1.2-1 Run Chart (Plot of Lot Data)
Run Chart

7.0
6.9
6.8
6.7
Quality Attribute

6.6
6.5
6.4
6.3
6.2
6.1
6.0

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Lot (by mfg date)
Step 3 — Connect plotted points with lines as shown in Figure 5.1.2-2.
Figure 5.1.2-2 Run Chart
Run Chart

7.0
6.9
6.8
6.7
Quality Attribute

6.6
6.5
6.4
6.3
6.2
6.1
6.0

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Lot (by mfg date)

28 © 2012 Parenteral Drug Association, Inc. Technical Report No. 59

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Step 4 — Add subsequent data values and connect to the previously plotted point as the data become
available.

5.2 Control Charts: Individuals

5.2.1 Technical Details
Step 1 — Place data or statistics in time-order
Step 2 — Create a plot of the data with the range of data values on the vertical axis and the timing of
the data on the horizontal axis. The plot should be created as soon as the first datum value is available.
Step 3 — Connect plotted points with lines.
Step 4 — Add subsequent data values and connect to the previously plotted point as the data become
available.
Step 5 — After 15 data points4 are collected, compute the overall average of the data values (X).
Step 6 — Add a line to the plot at the X value on the vertical axis.
Step 7 — Compute the moving range (MR) for each consecutive pair5 of data values by subtracting
the lower value from the higher value. MR will always be a positive number.
Step 8 — Compute the average of all MR values, (MR).
Step 9 — Compute the Upper Control Limit (UCL) using the following formula:6
UCL = X + 2.66 × MR
Step 10 — Compute the Lower Control Limit (LCL) using the following formula:6
LCL = X – 2.66 × MR
Step 11 — Add UCL and LCL to the chart for monitoring data points 16-30. These limits should not
be used retrospectively on prior data points
Step 12 — After 30 data points are collected, re-compute Upper and Lower Control Limits
Step 13 — Adjust UCL and LCL to the chart for monitoring points 31+
Step 14 — If any data point falls outside a control limit, investigate the root cause. If the deviation
can be assigned to a “special” cause, correct the cause. If the deviation cannot be assigned to a special
cause, evaluate the impact prior to continuing. When a special cause can be assigned to a deviation,
that data point should continue to be included on the chart. But it should be excluded from calcula-
tions of the average, UCL and LCL.

4 It is typical to review an initial set of data against expectations from validation or development, and then to
establish the initial working control limits on first series of commercial production. This isthen revisited as
a broader set of data become available

5 MR may be computed using a “run” of data longer than two consecutive points, but that is a more advanced
discussion
6 The multiplier of 2.66 is directly related to the use of two consecutive data points, or a run length of two and
the assumption that 3SD limits are employed

Technical Report No. 59 © 2012 Parenteral Drug Association, Inc. 29

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

5.2.2 Example
The first 15 data points represent the first fifteen lots in time-order:
Step 1 — Place data or statistics in time-order.

Table 5.2.2-1 First 15 Data Points/ Lots

Quality Quality
Lot Lot
Attribute Attribute
1 6.23 9 6.44
2 6.50 10 6.52
3 6.59 11 6.29
4 6.26 12 6.61
5 6.47 13 6.38
6 6.23 14 6.44
7 6.41 15 6.31
8 6.32

Step 2 — Create a plot of the data with the range of data values on the vertical axis and the timing
of the data on the horizontal axis. The plot should be created as soon as the first data value is available

Step 3 — Connect plotted points with lines

Figure 5.2.2-1 Run Chart (Plot of Lot Data)
Run Chart

7.0
6.9
6.8
6.7
Quality Attribute

6.6
6.5
6.4
6.3
6.2
6.1
6.0

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Lot (by mfg date)

30 © 2012 Parenteral Drug Association, Inc. Technical Report No. 59

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Step 4 — Add subsequent data values and connect to the previously plotted point as the data become
available
Step 5 — After 15 data points are collected, compute the overall average of the data values (X)
(X) = 6.40

Step 6 — Add a line to the plot at the X value on the vertical axis
Figure 5.2.2-2 Run Chart
Run Chart
6.70

6.60
Quality Attribute

6.50

6.40 Avg=6.4000

6.30

6.20
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Lot (by mfg date)
Step 7 — Compute the Moving Range (MR) for each consecutive pair of data values by subtracting
the lower value from the higher value. MR will always be a positive number.

Technical Report No. 59 © 2012 Parenteral Drug Association, Inc. 31

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Table 5.2.2-2 Moving Range

Lot Quality Attribute Moving Range

1 6.23
2 6.50 0.27
3 6.59 0.02
4 6.26 0.26
5 6.47 0.21
6 6.23 0.24
7 6.41 0.18
8 6.32 0.09
9 6.44 0.12
10 6.52 0.08
11 6.29 0.23
12 6.61 0.32
13 6.38 0.23
14 6.44 0.06
15 6.31 0.13

Step 8 — Compute the average of all MR values.

MR = 0.184
Step 9 — Compute the UCL using the following formula:
UCL = 6.40 + 2.66 × (0.184) = 6.9
Step 10 — Compute the LCL using the following formula:
LCL = 6.44 – 2.66 × (0.184) = 5.9

32 © 2012 Parenteral Drug Association, Inc. Technical Report No. 59

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Step 11 — Add UCL and LCL to the chart for monitoring points 16-30. These limits should not be
used retrospectively on prior points. With values for Lots 16-20 added, the chart would look like:
Figure 5.2.2-3 Run Chart (UCL and LCL shown)

7.0
6.9
6.8
6.7
Quality Attribute

6.6
6.5
6.4
6.3
6.2
6.1
6.0

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Lot (by mfg date)

Step 12 — After 30 points are collected, re-compute the grand average and UCL and LCL using all
30 values.

Step 13 — Adjust the limits to the chart for monitoring data for lots 31 and higher.

Step 14 — If any data falls outside a control limit, investigate the root cause. If a value is generated
outside the control limits and can be assigned to a “special” cause, correct the cause. If a “special”
cause is not found, evaluate the impact prior to continuing. When a special cause can be determined,
the data point should continue to be included on the chart but should be excluded from calculations
of the average, UCL and LCL.

5.3 Moving Range Control Charts

5.3.1 Technical Details
Step 1 — Place data or statistics in time-order

Step 2 — As soon as the second data value is collected, compute MR for each consecutive pair7 by
subtracting the lower value from the higher value. MR will always be a positive number.

Step 3 — Create a plot of the MR data with the range of data values on the vertical axis and the timing
of the data on the horizontal axis. The plot should be created as soon as the first MR value is available.

7 MR may be computed using a “run” of data longer than two consecutive points, but that is a more advanced
discussion

Technical Report No. 59 © 2012 Parenteral Drug Association, Inc. 33

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Step 4 — Connect plotted points with lines.

Step 5 — Add subsequent MR data values and connect to the previously plotted point as the data
become available.
Step 6 — After 16 data points are collected, compute the overall average of the MR values.
This is abbreviated as MR.

Step 7 — Add a line to the plot at the MR value on the vertical axis.
Step 8 — Compute UCL using the following formula:
UCL = 3.268 × MR
Step 9 — Set LCL = 0.
Step 10 — Add UCL and LCL to the chart for monitoring points 16-30. These limits should not be
used retrospectively on prior points.
Step 11 — Adjust UCL to the chart for monitoring points 31+
Step 12 — If any data fall outside a control limit, investigate the root cause. If a value is generated
outside the control limits and can be assigned to a “special” cause, correct the cause. If a “special”
cause is not found, evaluate the impact prior to continuing. When a special cause can be determined,
the data point should continue to be included on the chart, but it should be excluded from future
calculations of the average and UCL.

5.3.2 Example
Using the example data from Section 3.1.2.1, the first 15 data points represent the first 15 lots in time-
order:
Step 1 — Place data or statistics in time-order.

Table 5.3.2-1 Lot Data in Time Order

Quality Quality
Lot Lot
Attribute Attribute
1 6.23 9 6.44
2 6.50 10 6.52
3 6.59 11 6.29
4 6.26 12 6.61
5 6.47 13 6.38
6 6.23 14 6.44
7 6.41 15 6.31
8 6.32

Step 2 — As soon as the second data value is collected, compute MR for each consecutive pair8 of data
values by subtracting the lower value from the higher value. MR will always be a positive number.

8 MR may be computed using a “run” of data longer than two consecutive points, but that is a more advanced
discussion.

34 © 2012 Parenteral Drug Association, Inc. Technical Report No. 59

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Table 5.3.2-2 Lot Data with Moving Range

Lot Quality Attribute Moving Range

1 6.23
2 6.50 0.27
3 6.59 0.02
4 6.26 0.26
5 6.47 0.21
6 6.23 0.24
7 6.41 0.18
8 6.32 0.09
9 6.44 0.12
10 6.52 0.08
11 6.29 0.23
12 6.61 0.32
13 6.38 0.23
14 6.44 0.06
15 6.31 0.13

Step 3 — Create a plot of the MR data with the range of data values on the vertical axis and the time
sequence of the data on the horizontal axis. The plot should be created as soon as the first MR value
is available.
Step 4 — Connect plotted points with lines:
Figure 5.3.2-1 MR Chart

0.7

0.6

0.5

0.4
MR

0.3

0.2

0.1

0.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Lot
Technical Report No. 59 © 2012 Parenteral Drug Association, Inc. 35
Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Step 5 — Add subsequent MR data values and connect to the previously plotted point as the data
become available.
Step 6 — After 15 data points are collected, compute the overall average of the MR values. This is
abbreviated as MR (MR-bar):
MR = 0.184
Step 7 — Add a line to the plot at the MR value on the vertical axis.
Step 8 — Compute UCL using the following formula:
UCL = 3.268 × 0.184 = 0.601

Step 9 — Set LCL = 0

Step 10 — Add MR, UCL and LCL to the chart for monitoring points 16-30. These limits should not
be used retrospectively on prior points:
Figure 5.3.2-2 Moving Range Chart (Upper & Lower Limits shown)
MR Chart

0.7

0.6

0.5

0.4
MR

0.3

0.2

0.1

0.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Lot
Step 11 — After 30 data points are collected, re-compute UCL

Step 12 — Adjust UCL to the chart for monitoring points 31 and higher.

Step 13 — If any data fall outside a control limit, investigate the root cause. If a value is generated
outside the control limits and can be assigned to a “special” cause, correct the cause. If a “special”
cause is not found, evaluate the impact prior to continuing. When a special cause can be determined,
the data point should continue to be included on the chart, but it should be excluded from calcula-
tions of the average, UCL.

36 © 2012 Parenteral Drug Association, Inc. Technical Report No. 59

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

5.4 Average and Variability Charts

5.4.1 Technical Details
Step 1 — Identify the size of each sample set, commonly referred to as a subgroup. Subgroups should
be rational, that is, should represent a range of results that are intended to be very similar. That will
allow unintended changes between subgroups to be more easily distinguished.
Step 2 — Collect the specified number of samples for each subgroup at reasonable intervals during
the process or lot.
Step 3 — Begin collecting and recording data along with date/time information.
Step 4 — Compute the average for the first subgroup.
Step 5 — Plot the subgroup average on the X Chart.
Step 6 — Compute the measure of variability for each subgroup by either:
1. computing the range for each subgroup by subtracting the smallest value from the largest value within the
subgroup.
OR
2. computing the average standard deviation in two parts. First, for each rational subgroup, calculate the
standard deviation by taking the square root of the sum of the square of the individual values minus the
average value, which is then divided by the number of samples within the subgroup minus one. Next,
sum the standard deviation of each of the historical subgroups and divide by the number of historical
subgroups.

Step 7 — Plot the measure of variability for the first subgroup on the Variability Chart

Step 8 — Continue collecting samples and compute the average and range (or standard deviation) for
each sample set. Then, plot the averages and ranges on the appropriate chart and connect the points
with lines.

Step 9 — After 15 subgroups of data are collected:

1. compute the grand average of all the subgroup averages
2. compute the average of all subgroup ranges

Step 10 — Plot the grand average on the X Chart

Step 11 — Plot the average range on the Range Chart.
Step 12 — Compute UCL for the Variability Chart using the following formula:

Range
UCL = D4 R

Sample Size 2 3 4 5

D4 3.267 2.575 2.282 2.115

Technical Report No. 59 © 2012 Parenteral Drug Association, Inc. 37

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Standard Deviation
UCL = B4 S

Sample Size 6 7 8 9

B4 1.970 1.882 1.815 1.761

Step 13 — Compute LCL for the Variability Chart using the following formula:
Range
LCL = D3 R

Sample Size 2 3 4 5

D3 0 0 0 0

Standard Deviation
LCL = B3 S

Sample Size 2 3 4 5

B3 0.030 0.118 0.185 0.239

Step 14 — Plot UCL and the LCL on the Variability Chart for monitoring variability beyond sub-
group 15.
Step 15 — Compute UCL for the Average Chart using the following formula:
Range
UCL = X + A2 R

Sample Size 2 3 4 5

A2 1.880 1.023 0.729 0.577

Standard Deviation
UCL = X + A3S

Sample Size 6 7 8 9

A3 1.287 1.182 1.099 1.032

Step 16 — Compute LCL for the Average Chart using the following formula:

Range
LCL = X – A2R

Sample Size 2 3 4 5

A2 1.880 1.023 0.729 0.577

38 © 2012 Parenteral Drug Association, Inc. Technical Report No. 59

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Standard Deviation
UCL = X – A3S

Sample Size 6 7 8 9

A3 1.287 1.182 1.099 1.032

Step 17 — Plot UCL and the LCL on the Average Chart for monitoring subgroups 15 and beyond.

Step 18 — Continue collecting data for subgroups 16 and beyond. Update the Range (or s) and X
Charts with each subgroup’s variability and average.

Step 19 — If any subgroup’s variability or average falls outside the control limits, investigate the root
cause. If a “special” cause can be assigned, correct the cause. If a “special” cause is not found, evaluate
the impact prior to continuing. When a special cause can be assigned, the data point should continue
to be included on the chart; it should be excluded from calculations of ranges, averages and control
limits

Step 20 — After 30 subgroups are collected, re-compute the control limits

Step 21 — Adjust the control limits to the chart for monitoring subgroups 31 and beyond.

5.4.2 Example
Step 1 — Identify the size of each sample set, commonly referred to as a subgroup. Subgroups should
be rational, that is, should represent a range of results that are intended to be very similar. That will
allow unintended changes between subgroups to be more easily distinguished.
A tub of 30 vials is filled using a filling machine with six filling nozzles. Since the vials are filled at ap-
proximately the same time under the same conditions, it is reasonable to expect the weight of the vi-
als from each nozzle to be similar. Therefore, five vials will be sampled from the tub for nozzle #1 fills.
For the purposes of illustration, this example will describe the establishment of control limits from
the initial data generated within a run. In practice, it may be more common to have pre-established
expectations for process performance in place as a run commences.

Step 2 — Collect the specified number of samples for each subgroup at reasonable intervals during
the process or lot.

Due to the time required to weigh the vials, the sample of five for nozzle #1 will be weighed every
15 minutes.

Step 3 — Begin collecting and recording data along with date/time information.

The following data were collected for the first time point:

Table 5.4.2-1 Sample Data Collection Table: identify the size of each sample set, commonly re-
ferred to as a subgroup. Subgroups should be rational, that is, should represent a range of results that
are intended to be very similar. That will allow unintended changes between subgroups to be more
easily distinguished.
A tub of 30 vials is filled using a filling machine with six filling nozzles. Since the vials are filled at ap-
proximately the same time under the same conditions, it is reasonable to expect the weight of the vi-
als from each nozzle to be similar. Therefore, five vials will be sampled from the tub for nozzle #1 fills.

Technical Report No. 59 © 2012 Parenteral Drug Association, Inc. 39

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

For the purposes of illustration, this example will describe the establishment of control limits from
the initial data generated within a run. In practice, it may be more common to have pre-established
expectations for process performance in place as a run commences.

Step 2 — Collect the specified number of samples for each subgroup at reasonable intervals during
the process or lot.

Due to the time required to weigh the vials, the sample of five for nozzle #1 will be weighed every
15 minutes.

Step 3 — Begin collecting and recording data along with date/time information.

The following data were collected for the first time point:
Table 5.4.2-1 Sample Data Collection Table

3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16
Date/time
1:00 1:15 1:30 1:45 2:00 2:15 2:30 2:45 3:00 3:15 3:30 3:45 4:00 4:15 4:30

Sample 1 6.46

Sample 2 6.59

Sample 3 6.54

Sample 4 6.31

Sample 5 6.35

Step 4 — Compute the Average for the first subgroup.

Table 5.4.2-2 Sample Average Data Collection Table

3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16
Date/time
1:00 1:15 1:30 1:45 2:00 2:15 2:30 2:45 3:00 3:15 3:30 3:45 4:00 4:15 4:30

Sample 1 6.46

Sample 2 6.59

Sample 3 6.54

Sample 4 6.31

Sample 5 6.35

Average 6.45

40 © 2012 Parenteral Drug Association, Inc. Technical Report No. 59

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Step 5 — Plot the subgroup average on the X Chart.

Figure 5.4.2-1 Sample X Chart

6.62
6.58
6.54
6.50
Mean (Fill)

6.46
6.42
6.38
6.34
6.30
6.26
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Group
Step 6 — Compute the range for each subgroup by subtracting the smallest value from the largest
value within the subgroup.
Table 5.4.2-3 Sample Data Range Collection Table

Date/time 3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16
1:00 1:15 1:30 1:45 2:00 2:15 2:30 2:45 3:00 3:15 3:30 3:45 4:00 4:15 4:30

Sample 1 6.46

Sample 2 6.59

Sample 3 6.54

Sample 4 6.31

Sample 5 6.35

Average 6.45

Range 0.28

Technical Report No. 59 © 2012 Parenteral Drug Association, Inc. 41

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Step 7 — Plot the range for the first subgroup on the Range Chart
Figure 5.4.2-2 Sample Range Chart

0.50

0.40

0.30
Range

0.20

0.10

0.00
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Group
Step 8 — Continue collecting samples and compute the average and range for each sample set group.
Then, plot the group averages on the appropriate chart and connect the points with lines.
Table 5.4.2-4 Sample Data Average and Range Collection Table

3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16
Date/time
1:00 1:15 1:30 1:45 2:00 2:15 2:30 2:45 3:00 3:15 3:30 3:45 4:00 4:15 4:30

Sample 1 6.46 6.47 6.47 6.47 6.39 6.42 6.24 6.47 6.35 6.37 6.40 6.56 6.48 6.66 6.55

Sample 2 6.59 6.36 6.57 6.38 6.46 6.38 6.37 6.45 6.37 6.37 6.50 6.36 6.26 6.28 6.13

Sample 3 6.54 6.30 6.54 6.39 6.53 6.36 6.38 6.42 6.39 6.56 6.65 6.36 6.50 6.42 6.51

Sample 4 6.31 6.40 6.46 6.30 6.45 6.47 6.52 6.51 6.32 6.41 6.47 6.43 6.42 6.50 6.49

Sample 5 6.35 6.48 6.45 6.37 6.48 6.47 6.52 6.27 6.48 6.60 6.42 6.37 6.74 6.56 6.32

Average 6.45 6.40 6.50 6.38 6.46 6.42 6.40 6.42 6.38 6.46 6.49 6.42 6.48 6.48 6.40

Range 0.28 0.19 0.12 0.17 0.15 0.11 0.28 0.24 0.16 0.23 0.24 0.20 0.48 0.38 0.42

42 © 2012 Parenteral Drug Association, Inc. Technical Report No. 59

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Figure 5.4.2-3 Sample Average X Chart

6.62
6.58
6.54
6.50
Mean (Fill)

6.46
6.42
6.38
6.34
6.30
6.26
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Group
Figure 5.4.2-4 Sample Range X Chart

0.50

0.40

0.30
Range

0.20

0.10

0.00
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Group
Step 9 — After 15 subgroups of data are collected compute the grand average of all the subgroup
averages.

X = 6.45+6.40+6.50+6.38+6.46+6.42+6.40+6.24+6.38+6.46+6.49+6.42+6.48+6.48+6.40 = 6.436
15

Compute the average of all subgroup ranges in a similar fashion

R = = 0.243

Step 10 — Plot the grand average on the X Chart

Technical Report No. 59 © 2012 Parenteral Drug Association, Inc. 43

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Figure 5.4.2-5 Sample Grand Range X Chart

6.62
6.58
6.54
6.50
Mean (Fill)

6.46
6.42
6.38
6.34
6.30
6.26
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Group
Step 11 — Plot the average range on the Range Chart
Figure 5.4.2-6 Sample X Chart of Average Range on the Range

0.50

0.40

0.30
Range

0.20

0.10

0.00
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Group
Step 12 — Compute UCL for the Range Chart using the following formula:
UCL = D4 R
UCL = 2.115 × 0.243 = 0.515

Sample Size 2 3 4 5

D4 3.267 2.575 2.282 2.115

44 © 2012 Parenteral Drug Association, Inc. Technical Report No. 59

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Step 13 — Compute LCL for the Range Chart using the following formula:
LCL = D3 R
LCL = 0 × 0.243 = 0.00

Sample Size 2 3 4 5

D3 0 0 0 0

Step 14 — Plot the UCL and the LCL on the Range Chart for monitoring Ranges beyond subgroup 15.
Figure 5.4.2-7 Sample UCL and the LCL on the Range Chart

0.5
0.4
Range

0.3
0.2
0.1
0.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Group
Step 15 — Compute UCL for the Average Chart using the following formula:
UCL = X + A2 R
UCL = 6.436 + (0.577 × 0.243) = 6.57

Sample Size 2 3 4 5

A2 1.880 1.023 0.729 0.577

Step 16 — Compute LCL for the Average Chart using the following formula:
LCL = X – A2 R
LCL = 6.436 – (0.577 × 0.243) = 6.30

Sample Size 2 3 4 5

A2 1.880 1.023 0.729 0.577

Technical Report No. 59 © 2012 Parenteral Drug Association, Inc. 45

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Step 17 — Plot the UCL and the LCL on the Average Chart for monitoring subgroups 15 and beyond.
Figure 5.4.2-8 Sample UCL and the LCL on the Average Chart
6.62
6.58
6.54
6.50
6.42
Range

6.38
6.34
6.30
6.26
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Group
Step 17 — Continue collecting data for subgroups 16 and beyond. Update the Range and X Charts
with each subgroup’s range and average.
Table 5.4.2-5 Sample Average and Range Data Table

3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16 3/16
Date/time
4:45 5:00 5:15 5:30 5:45 6:00 6:15 6:30 6:45 7:00 7:15 7:30 7:45 8:00 8:15

Sample 1 6.55 6.45 6.43 6.32 6.56

Sample 2 6.43 6.49 6.33 6.47 6.42

Sample 3 6.34 6.61 6.38 6.41 6.33

Sample 4 6.58 6.48 6.46 6.41 6.47

Sample 5 6.49 6.50 6.43 6.26 6.42

Average 6.48 6.51 6.41 6.37 6.44

Range 0.25 0.16 0.13 0.21 0.23

Figure 5.4.2-9 Sample Average and Range X Chart

0.5
0.4
Range

0.3
0.2
0.1
0.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Group

46 © 2012 Parenteral Drug Association, Inc. Technical Report No. 59

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Figure 5.4.2-10 Sample Average and Range X Chart (UCL & LCL)

6.62
6.58
6.54
6.50
6.46
Range

6.42
6.38
6.34
6.30
6.26
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Group
Step 19 — If any subgroup range or average falls outside the control limits, investigate the root cause.
If a “special” cause can be assigned, correct the cause. If a special cause is not found, evaluate the im-
pact prior to continuing. When a special cause can be assigned, the data point should continue to be in-
cluded on the chart, but it should be excluded from calculations of ranges, averages and control limits.

Step 20 — After 30 subgroups are collected, re-compute the control limits

Step 21 — Adjust the control limits to the chart for monitoring subgroups 31 and beyond.

5.5 Histograms
5.5.1. Example Case Worked in Parallel with Theory/Principles
Understanding how past data are distributed provides useful information about how they might be
distributed in the future, assuming that the future will continue to look like the past. If the points in
the example Run Chart were replaced with X’s, the figure would look like:
Figure 5.5.1-1 Sample Average and Range X Chart (UCL & LCL)
6.8 6.8
6.7 6.7
6.6 6.6
Quality Attribute

Quality Attribute

6.5 6.5
6.4 6.4
6.3 6.3
6.2 6.2
6.1 6.1
6.0 6.0
5.9 5.9
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Lot (by mfg date) Lot (by mfg date)

After collecting the X’s for each Quality Attribute Value, the X’s would provide a first look at the dis-
tribution of results:

When examined separately, the distribution of X’s seems to have somewhat of a symmetrical “bell” shape.

Technical Report No. 59 © 2012 Parenteral Drug Association, Inc. 47

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Figure 5.5.1-2 Sample Statistical Distribution of Results

X
X
X
X
X X
X X X X
X X X X
X X X X X X

6.2 6.3 6.4 6.5 6.5 6.7

A statistical distribution assumed for many calculations that has a bell-shape is the Normal Distribu-
tion. For the data in this example, the average is 6.5 and the standard deviation is 0.1. The standard
deviation is a measure of the variability of the data.

5.6 Cpk, Ppk for Process Capability

5.6.1 General Procedure for Finding Cpk and Ppk
Step 1 — Collect a truly representative sample from the process using good sampling techniques.
The sampling plan can be random or systematic. The sample should be as large as possible but need
not be more than 100 initially.

Step 2 — Calculate the average and the variability estimate.

Step 3 — Calculate Cpk (and/or Ppk).

Cpk=Min (Cpl, Cpu)
Cpl= (Average -Lower Specification Limit)/3*Variability Estimate
Cpu= (Upper Specification Limit-Average)/ 3*Variability Estimate

Step 4 — Interpret the results in terms of the process

Ppk is similarly calculated, except that instead of using the estimate of process variability, it uses the
actual variability observed within the sample set.

If the process is capable of reliably achieving specifications, the index will be greater than 1.0. Most
applications attempt to achieve at least 1.33. Well-controlled and capable processes sometimes achieve
a desired value of 2.0.

48 © 2012 Parenteral Drug Association, Inc. Technical Report No. 59

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

5.6.2 Example Cpk for a Single Group

Step 1 — Collect the data, which in this case is a single group of 30 values all taken at the same time.
Table 5.6.2-1 Example of Cpk for a Single Group

90.8 94.6 101.2 98.2 101.4 98.3

99.0 93.8 105.7 102.3 93.4 104.5

98.5 101.6 106.9 105.6 98.0 98.7

105.3 96.2 96.2 100.6 102.7 103.7

97.9 96.8 98.6 101.5 98.8 100.1

Step 2 — Calculate the average and variability estimate

The grand average of the thirty values is 99.6914
The standard deviation of all the data taken as one group is 4.1174

Step 3 — The example specification criteria are 90%-100%.

Cpl =(99.6914-90.0)/(3*4.1174)=0.78
Cpu=(110.0-99.6914)/(3*4.1174) = 0.83
Cpk=MIN (0.78, 0.83) = 0.78

Step 4 — Cpk needs to be greater than 1.0 for the natural variability of the data to lie within the
specification range. In this example, the process is apparently not capable of meeting the specification
limits. Even if the calculated Cpk is greater than 1.0, it is possible that the true, but unknown, value is
actually less than 1.0 since Cpk is a random variable. If many samples are taken, the result would be a
distribution of Cpk values. To address the variability issue, find the lower 95% confidence interval for
Cpk as in Table 5.6.2-2.

Technical Report No. 59 © 2012 Parenteral Drug Association, Inc. 49

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Table 5.6.2-2 Example of 95% confidence interval for Cpk

Size
5 10 15 20 25 30 40 50 75 100
Est. Cpk

1.0 0.38 0.58 0.66 0.71 0.74 0.76 0.80 0.82 0.85 0.87

1.1 0.42 0.64 0.73 0.78 0.82 0.84 0.88 0.90 0.94 0.96

1.2 0.47 0.71 0.80 0.86 0.90 0.92 0.96 0.99 1.03 1.05

1.3 0.51 0.77 0.87 0.93 0.97 1.00 1.04 1.07 1.11 1.14

1.4 0.56 0.83 0.94 1.01 1.05 1.08 1.13 1.16 1.20 1.23

1.5 0.60 0.90 1.01 1.08 1.13 1.16 1.21 1.24 1.29 1.32

1.6 0.64 0.96 1.09 1.16 1.21 1.24 1.29 1.32 1.38 1.41

1.7 0.69 1.02 1.16 1.23 1.28 1.32 1.37 1.41 1.46 1.49

1.8 0.73 1.08 1.23 1.31 1.36 1.40 1.45 1.49 1.55 1.58

1.9 0.77 1.15 1.29 1.38 1.44 1.48 1.54 1.58 1.64 1.67

2.0 0.82 1.21 1.36 1.45 1.51 1.56 1.62 1.66 1.72 1.76

2.1 0.86 1.27 1.43 1.53 1.59 1.64 1.70 1.74 1.81 1.85

2.2 0.90 1.33 1.50 1.60 1.67 1.72 1.78 1.83 1.90 1.94

2.3 0.95 1.39 1.57 1.68 1.74 1.79 1.86 1.91 1.98 2.03

2.4 0.99 1.46 1.64 1.75 1.82 1.87 1.95 1.99 2.07 2.11

2.5 1.03 1.52 1.71 1.82 1.90 1.95 2.03 2.08 2.16 2.20

For example, for a sample size of 30, Cpk of 1.3 would need to be calculated to have a 95% confidence
that the true but unknown Cpk is 1.0 or greater.

One of the generalized formulas for the Cpk confidence interval is:
LCI = Cpk–Z1-a*SQRT (1/ (9*n) + (Cpk) 2/ (2*(n-1))) +1/ (30*SQRT (n))

5.6.3 Cpk for Several Groups

In this example , the short-term variability or variability within the group is used to find Cpk. This does
not take into account the variability of different groups. This Cpk assumes that the group means are
the same. Thus, the Cpk found here is as good as it will be.

Step 1 — Collect the data. Here, six groups of five values are taken over time. See Table 5.6.3-1 for
the data and summaries.

50 © 2012 Parenteral Drug Association, Inc. Technical Report No. 59

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Table 5.6.3-1 Cpk for Several Groups

Value Group 1 Group 2 Group 3 Group 4 Group 5 Group 6

1 94.1 93.1 91.2 93.9 90.6 92.4

2 102.7 95.2 98.1 101.6 96.2 101.3

3 101.9 100.2 96.3 103.1 99.0 101.0

4 103.3 98.5 93.8 99.8 97.7 99.1

5 100.5 102.6 104.8 99.4 98.6 96.5

Average Range
Range 9.1400 9.4512 13.5881 9.223 8.4423 8.9232
9.7946
Grand Average
Average 100.4930 97.9069 96.8425 99.5611 96.4329 98.0678
98.2174

Step 2 — Calculate the average and the variability estimate

The grand average of the data is 98.2174.
The average range is 9.7946. The sample size is 5.
The standard deviation is estimated by dividing the average range by d2 from Table 5.6.2-4.
Table 5.6.3-2 For d2

Sample Size d2

2 1.13
3 1.69
4 2.06
5 2.33
6 2.53
7 2.70
8 2.85
9 2.97
10 3.08
11 3.17
12 3.26
13 3.34
14 3.41
15 3.47

For a sample size of 5, d2 is 2.33

The standard deviation is 9.7946/2.33=4.2037

Technical Report No. 59 © 2012 Parenteral Drug Association, Inc. 51

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Step 3 — Calculate Cpk for several groups

The example specification criteria are 90.0%-100.0%
Cpl= (98.2174-90.0)/ (3*4.2037) = 0.65
Cpu= (110.0-98.2174)/ (3*4.2037) = 0.93
Cpk= 0.65

Step 4 — Based on this data, the process is probably not capable of meeting the specifications, even
in the short term. It is possible that sample selection is not adequate and that the true process Cpk is
greater than 1.0, so more data would have to be collected to get a better estimate of the true value.
Without the extra samples, the only conclustion is that the process is not capable of reliably meeting
specifications.

5.6.4 Example 3, Ppk for Several Groups

In this example, the long-term variability or total variability is used to find Ppk. This takes into account
the variability within groups and the variability from group to group. This Ppk does not assume that
the group means are the same. Thus, the Ppk here is a realistic estimate of future long-term capability.

Step 1 — Collect the data. Here six groups of five values are taken over time. See Table 3 for the data.

Step 2 — Calculate the average and the variability estimate

The grand average of the data is 98.2174.
The standard deviation of all of the data taken as one group is 3.9599.

Step 3 — Calculate Ppk

Ppl= (98.2174-90.0)/ (3*3.9599) =0.69
Ppu= (110.0-98.2174)/ (3*3.9599) =0.99
Ppk=0.69

Step 4 — Based on this data, the process is probably not capable of meeting the specifications in the
long term. It is possible that sample selection is not adequate and the true process Ppk is greater than
1.0, so more data would have to be collected to estimate the true value. Without the extra samples,
the only conclustion is that the process is not capable of reliably meeting specifications.

5.7 Exponentially Weighted Moving Area Charts (EWMA)

5.7.1 Technical Details
The EWMA value is calculated6 as EWMAt = λXt + (1- l) EWMAt-1 for t = 1, 2, …, n. where EWMAt
is the moving average of the historical data:
• Xt is the observation at time t
• N is the number of observations to be monitored, including EWMA0
• The primary weighting factor is represented as λ (lambda). This can be described as the portion of
the current point that will be added to the prior average. Each successive EWMA point is the sum
of the weighted current data point and the prior EWMA point.
• λ is a constant that determines the depth response time of
the EWMA. This value is a proportion
so the values must be between 0 and 1, with typical values 0.05-0.4, depending on the amount of
dampening desired (smaller λ=more dampening). Smaller values for λ can detect smaller changes
in mean, but will increase the time to respond to shifts.

52 © 2012 Parenteral Drug Association, Inc. Technical Report No. 59

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

5.7.2 Calculation Example

Simple step-change example with no process noise with λ = 0.4
Table 5.7.2-1 Step-Change Example (No process noise)

Example Calculation
Time (t) Data Method EWMA At t=1, if there is no pro-
cess average, the first ob-
First EWMA point is a special case, servation is transferred. If a
process average or target
if available use Target or Process
1 10 10.00 has been established, it
average, otherwise use first data may be used as a starting
point. point for the trend.

2 10 10*0.4 + 10.00 * (1-0.4) 10.00

Example Calculation
EWMA for time 3 is
3 11 11*0.4 + 10.00 * (1-0.4) 10.40 Observation (t=3) (11)*
Lambda (0.4) + EWMA
(t=2) * (1-Lambda)
4 11 11*0.4 + 10.40 * (1-0.4) 10.64

5 11 11*0.4 + 10.64 * (1-0.4) 10.78

6 11 11*0.4 + 10.78 * (1-0.4) 10.87

7 10 10*0.4 + 10.87 * (1-0.4) 10.52

8 10 10*0.4 + 10.52 * (1-0.4) 10.31

9 10 10*0.4 + 10.31 * (1-0.4) 10.19

10 10 10*0.4 + 10.19 * (1-0.4) 10.11

5.7.3 Control Limits

In order to calculate control limits, there should be established expectation for process mean (μprocess)
and standard deviation (Sprocess).

The weighted averaging tends to reduce deviation. The standard deviation of the EWMA statistic is

Control Limits (assuming ±3 sigma intent). For low numbers of samples, the control limits are some-
what narrower and an alternate formula may be applied.9

SEWMA = Sprocess

UCLEWMA = µprocess + 3 × Sprocess ×

UCLEWMA = µprocess – 3 × Sprocess ×

9 An example of the detailed calculation for upper control limit

UCLEWMA = µprocess + 3 × Sprocess × (1–[1– λ]21)

Technical Report No. 59 © 2012 Parenteral Drug Association, Inc. 53

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

5.7.4 Example
A sample data set was prepared with an expected mean of 10 and standard deviation of 1, randomly
distributed. The mean of this simulated process contains a shift. The same data is plotted with both a
conventional data plot (run chart), and the EWMA chart. The EWMA was configured with λ = 0.3.
Figure 5.7.4-1 Sample Random Distribution Plots
Sample Data Plot EWMA Plot

14 11.5
13
11.0
12
10.5
11
10 10.0
9
9.5
8
9.0
7
6 8.5

0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Sample The same data, plotted as Sample
EWMA, with l=0.3

5.7.4.1 Interpretation of Example

Data Plot — If individual results are plotted, no points fall outside the ± 3 standard deviation limits.

EWMA Plot — The control limits of the EWMA are exceeded several times, annunciating that this
process average has shifted. Although the control limit is first exceeded at point 25, the review of the
process performance should extend back to the time when the trend appears to deviate from the for-
mer average. The investigation question in this case would be, “What changed in the process some-
time between samples 20-25 that continues to affect samples since then?” The investigation question
is not “What is different about points 25, 32, 38-40?”
The reduction in the variability of individual points allows the vertical axis to increase in resolution,
and the approximate average at various time periods is readable directly from the chart (in the ex-
ample, the process mean started at 10, and started drifting downward at sample 22, dropping to 9 at
sample 25).

The figure below shows the same chart printed with the underlying mean drift of the simulated data
made visible.

54 © 2012 Parenteral Drug Association, Inc. Technical Report No. 59

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Figure 5.7.4.1-1 EWMA Plot (UCL & LCL)

EWMA Plot

11.5

11.0

10.5
EWMA & Mean

10.0

9.5

9.0

8.5

0 5 10 15 20 25 30 35 40
Sample

Notice that even though the mean shift is steady, points 25-40 intermittently fall in or out of the con-
trol limit. There is still randomness to the detection. Had the process shift been larger, the detection
would have been more definitive. Had λ been smaller, the control limits would have been tighter, and
the detection would have been more definitive, but later. Further discussion of the trade-offs in selec-
tion of λ is beyond the scope of this document.10

5.8 CuSum Charts

CuSum charts plot the cumulative sum of deviation from a target value.

For individual data points, the formula is

j
Cj = ∑ (xi – µ)
i=l

C: value for the CuSum Chart; x = measured parameter; µ = mean 

10 Crowder, S.V (1987) A Simple Method for Studying Run Length Distributions of Exponentially Weighted
Moving Average Charts. Technometrics, Vol 29.

Technical Report No. 59 © 2012 Parenteral Drug Association, Inc. 55

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

5.8.1 Example Calculation

The expected value for the process in this example is 10.0. The mechanics of the calculation are illus-
trated in Figure 5.8.1-1. The resultant plots are shown in Figure 5.8.1-2.
Figure 5.8.1-1 Example Calculation

Time (t) Data Method CUSUM

1 10.12 10.12-10.00 0.12

2 9.85 9.85-10.00 + 0.12 -0.03

Example Calculation
CUSUM for time 3 is
3 12.10 12.10-10.00 + -0.03 2.08 Observation (t=3) (12.10) –
Expected Value (10.0) +
CUSUM (t=2)
4 9.22 9.22-10.00 + 2.08 1.29

5 10.04 10.04-10.00 + 1.29 1.34

6 11.27 11.27-10.00 + 1.34 2.61

7 14.73 14.73-10.00 + 2.61 6.34

8 13.30 13.30-10.00 + 6.34 8.63

9 11.81 11.81-10.00 + 8.63 9.44

10 13.56 13.56-10.00 + 9.44 12.00

Figure 5.8.1-2 X-bar Chart vs. CuSum Chart

X vs. T
10
14 8
Cumulative Sum of X

13
6
12
4
X

11
2
10
0
9
-2
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
T T
If the process is in a state of statistical control around the expected value of 10.0, the CuSum Cart
would vary around zero which represents the horizontal line.

A V-mask is plotted to indicate whether or not the process is in a state of statistical control (Figure
5.8.1-3). This is similar in function to a running statistical t-test. It determines whether or not the

56 © 2012 Parenteral Drug Association, Inc. Technical Report No. 59

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

process is discernibly different at any point in time. This V-mask is placed at a distance d. The opening
of the mask is drawn at an angle of ± q. The manual calculation of these values is outside the scope
of this document.
Figure 5.8.1-3 CuSum Chart with V-mask
100
75
50
25
CuSum

0
-25
-50
-75
-100
0 10 20 30 40 50 60
Observation

For further reading, see ISO/TR 7871:1997: Cumulative sum charts -- Guidance on quality control
and data analysis using CUSUM techniques. This document provides the principles for CuSum chart-
ing and includes guidance on the preparation and interpretation of CuSum charts using basic decision
rules.

5.8.2 Details of Principles

CuSum charts monitor the available data by plotting the cumulative sum of deviations from a fixed
value (e.g., specification, average, and target). They incorporate the past history of the monitored
process into the plotted points for the selected parameter. As a result, greater sensitivity and shorter
run length than those in x-bar or I-charts are achieved. Small shifts to the process mean are visible as a
change in slope that indicates when the change occurred. CuSum charts are usually plotted with a V-
mask. This provides reference slopes to annunciate that a change in mean is significant.

5.8.3 Interpretations Worked in Parallel with Theory/Principles

The mathematical process to construct a CuSum Chart has advantages to detect deviations in the
process by taking into account the knowledge of the past. To understand basic appearance of graphs
see some examples without the V-mask:

Assume the values (X) are distributed around 100. The CuSum chart would show some ‘noise’ (Fig-
ure 5.8.3-1).

Technical Report No. 59 © 2012 Parenteral Drug Association, Inc. 57

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Figure 5.8.3-1 CuSum Chart with values around 100

300 Theory (no noise) Typical Appearance (with noise)
150
250
200
100
X
150 CuSuM
100 50 XF
CuSuM
50
0 0
0 10 20 30 40 50 60 70 80 90 100 1 10 20 30 40 50 60 70 80 90 100

If the values (X) at point 50 slight moves up, the CuSum chart answers with a rising graph (Figure
5.8.3-2). The CuSum chart shows a significant change in slope that is far more apparent than in the
data plot.
Figure. 5.8.3-2 CuSum chart with a slide move at event 50
200 Theory (no noise) 200 Typical Appearance (with noise)
180
160 150
140
120 100
100
80 50 X
CuSuM
60
40 X 0
CuSuM
20
-50 10 20 30 40 50 60 70 80 90 100
0
0 10 20 30 40 50 60 70 80 90 100
-100

If the values (X) change for a brief period and then return to the prior average, the CuSum chart will
show a brief slope and then return to parallel when the set point (100) is re-achieved. (Figure. 5.8.3-3)
Figure. 5.8.3-3 CuSum chart with an upward shift in mean between points 23 to 50
300 Theory (no noise) 300 Typical Appearance (with noise)
250 250
200 200
150 150
100 100
X X
50 CuSuM 50 CuSuM
0 0
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
-50

58 © 2012 Parenteral Drug Association, Inc. Technical Report No. 59

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

For interpretation of a CuSum chart including a V-mask, the two principles apply:
• The process is in a state of statistical control if the entire history lies within the opening of the V-mask
• The process average is statistically different if any point lies outside the V-mask

The example given shows a full CuSum chart including a V-mask (Figure. 5.8.3-4). The V-mask is
positioned in front of the CuSum at the 29 event. The point at time the 22 event goes outside the
V-mask. This indicates a statistically detectable difference in mean. The departure from “statistical
control” does not mean that the process itself is out of control or unacceptable. The interpretation is
that the process mean appears to have shifted upwards as the slope of the CuSum is positive.
Figure 5.8.3-4 Example of a CuSum Chart

12

-4

-8

-12
0 10 20 30 40
Observation

5.8.4 Implementation of CuSum Charts

To implement CuSum charts, a data set with a target or average value is needed. The data can be
obtained by manual measurements (e.g., manual In Process Check (IPC) tests, yield) or automatically
(e.g., LIMS systems, PAT application). The data are mapped (e.g., usually in an Excel spreadsheet) and
the CuSum chart is monitored (e.g., in MS-Excel http://www.qimacros.com/qiwizard/cusum-chart.
html) or by statistical programs.

5.8.5 Example: Gradual Trend

If the curve twists off outside the V-mask without a sharp point change in slope, the data are gradu-
ally trending. This trend is statistically relevant if it is outside the V- mask (see after value 11 in the
Figure 5.8.5-1).

Technical Report No. 59 © 2012 Parenteral Drug Association, Inc. 59

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Figure 5.8.5-1 Gradual Trend

50
Change in slope
40
30
20
10
CuSum

0
-10
-20
-30
-40
-50
0 10 20 30 40
Observation
There is a continual trend without a crack.

5.8.6 Preventative Action Before Failure Occurrence

After observation eight, the CuSum chart shows a raise. Also this result of the IPC sample was avail-
able about 5 hours later, and other samples already were taken so a side adjustment of a parameter
could have been performed. As a result, the process drift could be controlled before a statistically
significant effect was detected by other means.

5.8.7 Example: For Process Optimization

Sometimes processes are not running in a distinct distribution around the target value. Lot-by-lot
incremental adjustments of a specific critical process parameter have been monitored with following
up on a CuSum chart. Raises and plateaus have been controlled until the graph goes to a straight line
(Figure 5.8.7-1).
Figure 5.8.7-1 For Process Optimization
X-bar Chart X-bar Chart
60 50
40
30
50
20 Face action
10 squarely
40 0 Perform action
-10
-20
30
-30
-40
20 -50
0 10 20 30 40 50 0 10 20 30 40 50
Observation Observation
The effect can be revisited in the x-bar chart afterwards.

60 © 2012 Parenteral Drug Association, Inc. Technical Report No. 59

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

5.9 Sampling
The basic process for developing a sampling plan is listed below. Further details are outside the scope
of this document. Please consult an appropriate resource.
Step 1 — Determine the purpose for the inspection.
Step 2 — Determine and list each quality characteristic by name and description.
Step 3 — Specify to which product unit each quality characteristic applies and which the test is to be
made. Specify the product unit associated with each quality characteristic and which tests are to be
applied.
Step 4 — Develop and specify the method to be used for testing the product for conformance.
Step 5 — Determine and specify the criteria for conformity for each quality characteristic (give de-
tailed descriptions).
Step 6 — Determine and list the classification of nonconformities that affect a unit of product. (This
refers to degree of importance/severity of the nonconformity). This is generally a result of a risk as-
sessment.
Step 7 — Determine the appropriate index quality level (AQL, RQL) on which to base the sampling
procedures for each nonconformity classification.
Step 8 — Obtain appropriate information on lot formation.
Step 9 — Determine what method of sampling will be used for each nonconformity classification,
nonconformity grouping, or product unit distinction (to obtain representative sample).
Step 10 — Determine the type of sampling that will be used for each nonconformity classification,
nonconformity grouping, or product unit distinction (e.g., single, double, etc.). The type of sampling
selected is generally based on whether the efficiency gained by the various sampling types justifies the
additional complexity of administering the plans.
Step 11 — Develop the sampling plan and include:
• Provisions for disposition of rejected lots (e.g., 100% inspection)
• Provisions for resubmission of reworked product

5.9.1 Example
Suppose a vendor supplies empty vials in a lot size of 250,000. The drug manufacturer decides to
implement a sampling plan for either accepting or rejecting the entire lot. The drug manufacturer and
the vendor agree that the AQL is 1.0% and the RQL is 2.5% for a given defect.

Stated producer’s risk (Alpha) = 5% and consumer’s risk (Beta) = 2.5%

Lot size: 250,000

Acceptable Quality Level (AQL) = 1.0%
Producer’s Risk (Alpha) = 5%
Rejectable Quality Level (RQL , LQ, or LTPD) = 2.5%
Consumer’s Risk (Beta) = 2.5%

Generated Plan:
Sample Size = 985
Acceptance Number = 15
Accept lot if defective items in 985 sampled ≤ 15; otherwise reject.
Probabilities from this plan are listed in Table 5.9.1-1.

Technical Report No. 59 © 2012 Parenteral Drug Association, Inc. 61

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Table 5.9.1-1 Characteristics of Example Test Plan

Probability of Probability of
Percent Defective AOQ ATI
Accepting Rejecting
1.0 0.957 0.043 0.953 11667
2.5 0.025 0.975 0.062 243794

AOQ — Average Outgoing Quality

Approximates the relationship between the quality of the incoming material and the quality of the
outgoing material, assuming that the rejected lots will be 100% inspected and defective items will be
reworked and inspected again (rectifying inspection).

ATI — Average Total Inspection

Approximates the relationship between the quality of the incoming material and the number of items
that need to be inspected, assuming that rejected lots will be 100% inspected and defective items will
be reworked and inspected again (rectifying inspection).

For each lot of 250,000 empty vials, randomly select and inspect 985 of them. If there are greater than
15 defectives among these 985 empty vials, the entire lot should be rejected. For 15 or less defective
empty vials, accept the entire lot.

In this example (operating characteristic chart curve in Figure 5.9.1-1), the probability of acceptance
at the AQL (1.0%) is 0.957 and the probability of rejecting is 0.043. The sampling plan is based on the
expectation that lots with 1.0% defective units would be accepted approximately 95% of the time.
The probability of accepting at the RQL (2.5%) is 0.025 and the probability of rejecting is 0.975.
Figure 5.9.1-1 OC Chart Curve
Operating Characteristic (OC) Curve
Sample Size=985, Acceptance Number=15

1.000
0.957
AQL=1.0%, Alpa=0.043
0.800
Probability of Acceptance

0.600

0.400

0.200
RQL=2.5%, Beta=0.025
0.025

0.0 1.0 2.0 2.5 3.0 4.0 5.0

Lot Percent Defective

62 © 2012 Parenteral Drug Association, Inc. Technical Report No. 59

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

6.0 References

1. Quality Guideline Q6A: Specifications: Test Procedures 4. Quality Guideline Q10: Pharmaceutical Quality Systems;
and Acceptance Criteria for New Drug Substance and International Conference on Harmonisation: 2008.
New Drug Product: Chemical Substances; International www.ich.org.
Conference on Harmonisation: 1999. www.ich.org 5. Quality Guideline Q9: Quality Risk Management;
2. Quality Guideline Q6B: Specifications: Test Procedures International Conference on Harmonisation: 2005.
and Acceptance Criteria for Biotechnological/Biological www.ich.org.
Products; International Conference on Harmonisation: 6. Quality Guideline Q8 (R2): Pharmaceutical Development;
1999. www.ich.org International Conference on Harmonisation: 2009.
3. Guidance for Industry: Quality Systems Approach www.ich.org.
to Pharmaceutical CGMP Regulations; U.S. FDA: 7. Quality Guideline Q11: Development and Manufac-
2006. www.fda.gov/downloads/Drugs/Guidance ture of Drug Substances (Chemical Entities and Biotech-
ComplianceRegulatoryInformation/Guidances/ nological/Biological Entities); International Confer-
ucm070337.pdf (accessed Apr. 27, 2012). ence on Harmonisation: 2012. www.ich.org.

Technical Report No. 59 © 2012 Parenteral Drug Association, Inc. 63

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

7.0 Additional Reading

1. Kotz, S.; Johnson, N.L. Process Capability Indices; 11. U.S. Code of Federal Regulations Title 21 Parts 210
Chapman & Hall/CRC: New York, 1993; p 71. and 211.
2. Duncan, A.J. Quality Control and Industrial Statistics; 12. Guidance for industry: Sterile drug products produced
R.D. Irwin: Homewood, IL, 1974; p. 950. by aseptic processing – Current good manufacturing
3. ANSI/ASQ Z1.4-2008: Sampling Procedures and practice; U.S. Food and Drug Administration: 2003.
Tables for Inspection by Attribute; American 13. Gunter, B. The Use and Abuse of Cpk, Part 4. Qual.
Society for Quality: 2008. Progress, 1989, 22, 86-87.
4. ANSI/ASQ Z1.9-2008: Sampling Procedures and 14. Perez-Wilson, M. Machine/Process Capability Study;
Tables for Inspection by Variables for Percent Advanced Systems Consultants: Scottsdale, AZ,
Nonconforming; American Society for Quality: 1989.
2008. 15. Sullivan, L. A letter from Sullivan to Hal Greenberg,
5. ISO/TR 7871:1997: Cumulative sum charts — VP, Divisions and Technical Committee of ASQC,
Guidance on quality control and data analysis using 1990.
CUSUM techniques; International Organization 16. Gunter, B. The Use and Abuse of Cpk Revisited.
for Standardization: 1997. Quality Progress, 1991, 24, 90-94.
6. Montgomery D.C. Introduction to Statistical Quality 17. Nelson, P. Editorial. J. Qual. Technol. 1992, 24.
Control; Wiley: Hoboken, NJ.
18. Schneider, H.; et al.;. Uses of Process Capability
7. Marriott, F. H. C. A dictionary of statistical terms; Indicies in the Supplier Certification Process. Qual.
Longman Scientific and Technical: London, 1990. Engin. 1995, 8, 225-235.
8. Everitt, B. S. The Cambridge dictionary of statistics; 19. Kaninsky, F.; et al.; Process Capability Indices: Now
Cambridge University Press: Cambridge, 1998. and In the Future. Qual. Engin. 1998, 10, 445-453.
9. The Gold Sheet. 1995, 29(3). 20. Torbeck, L.D. Validation by Design The Statistical
10. Schemer, W.L. Annual Product Review Under Drug Handbook for Pharmaceutical Process Validation;
CGMP’s. FDA News & Information. 2001. 2(11). PDA-DHI: Bethesda, MD, 2010.

64 © 2012 Parenteral Drug Association, Inc. Technical Report No. 59

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Terms of Usage
An Authorized User of the electronic PDA Technical Report is a PDA Member in good standing or the purchaser of the Technical
Report.

Authorized Users are permitted to do the following:

• Search and view the content of the PDA Technical Report
• Download the PDA Technical Report for the individual use of an Authorized User
• Print the PDA Technical Report for the individual use of an Authorized User
• Make a reasonable number of photocopies of a printed PDA Technical Report for the individual use of an Authorized User

Authorized Users are not permitted to do the following:

• Allow anyone other than an Authorized User to use or access the PDA Technical Report
• Display or otherwise make any information from the PDA Technical Report available to anyone other than an Authorized User
• Post articles from the PDA Technical Report on Web sites, either available on the Internet or an Intranet, or in any form of online
publications without a license from PDA, Inc.
• Transmit electronically, via e-mail or any other file transfer protocols, any portion of the PDA Technical Report
• Create a searchable archive of any portion of the PDA Technical Report
• Use robots or intelligent agents to access, search and/or systematically download any portion of the PDA Technical Report
• Sell, re-sell, rent, lease, license, sublicense, assign or otherwise transfer the use of the PDA Technical Report or its content
• Use or copy the PDA Technical Report for document delivery, fee-for-service use, or bulk reproduction or distribution of materials
in any form, or any substantially similar commercial purpose
• Alter, modify, repackage or adapt any portion of the PDA Technical Report
• Make any edits or derivative works with respect to any portion of the PDA Technical Report including any text or graphics
• Delete or remove in any form or format, including on a printed article or photocopy, any copyright information or notice con-
tained in the PDA Technical Report
• Combine any portion of the PDA Technical Report with any other material

To License or purchase Reprints or Hardcopies

Licensing: Site licenses and licenses to distribute PDA Technical Reports can be obtained for a fee. To learn more about licensing
options and rates, please contact: Janny Chua, Publications Manager, +1 (301) 656-5900, ext. 133. Written correspondence can be
sent to: PDA, Inc. Attn: Janny Chua, 4350 East West Highway, Suite 150, Bethesda, MD 20814.
Reprints: Reprints of PDA Technical Reports can be purchased for a fee. To order reprints, please contact: Janny Chua, Publications
Manager, +1 (301) 656-5900, ext. 133. Written correspondence can be sent to: PDA, Inc. Attn: Janny Chua, 4350 East West Highway,
Suite 150, Bethesda, MD 20814.
Hardcopies: Hardcopies of PDA Technical Reports are available on demand for a fee. The PDA membership benefits include dis-
counted pricing for printed copies of PDA Technical Reports. Go to www.pda.org/bookstore to order a hardcopy. PDA members
will need their PDA identification number and password to receive the discount.

Technical Report No. 59 © 2012 Parenteral Drug Association, Inc. 65

Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Technical Report No. 59

About PDA Technical Reports
PDA Technical Reports are global consensus documents, prepared by member-driven Task Forces (listed on inside front cover) com-
prised of content experts, including scientists and engineers working in the pharmaceutical/biopharmaceutical industry, regulatory
authorities and academia. While in development, PDA Technical Reports are subjected to a global review of PDA members and
other topic-specific experts, often including regulatory officials. Comments from the global review are then considered by the au-
thoring Task Force during the preparation of the final working draft. The level of expertise of the Task Force and those participating
in the global review ensure a broad perspective reflecting best thinking and practices currently available.
The final working draft is next reviewed by the PDA Advisory Board or Boards that sanctioned the project. PDA’s Advisory Boards
are: Science Advisory Board, Biotechnology Advisory Board, Regulatory Affairs and Quality Committee, and Audit Guidance Advi-
sory Board. Following this stage of revi.e., the PDA Board of Directors conducts the final review and determines whether to publish
or not publish the document as an official PDA Technical Report.

Sanctioning Advisory Board: Science Advisory Board PDA Officers and Directors
John Shabushnig, Pfizer (Chair) Officers
Hal Baseman, Valsource, LLC (Co-Chair)
Chair: Anders Vinther, Ph.D., Genentech
Raphael Bar, BR Consulting
Joyce Bloomfield, Merck Chair-Elect: Harold Baseman, ValSource
Jette Christensen, Novo Nordisk Treasurer: Rebecca Devine, Ph.D., Regulatory Consultant
Jean-Luc Clavelin Secretary: Steven Mendivil, Amgen
Walter De Matteo, Lilly France, S.A.S. Imm. Past Chair: Maik Jornitz, G-Con
Phil DeSantis, DeSantis Consulting Associates President: Richard Johnson
Jens Eilertsen, Novo Nordisk
Don Elinski, Lachman Consultants Directors
Kris Evans, Amgen Ursula Busse, Ph.D., Novartis
Tor Graberg, Medical Products Agency – Sweden
Jette Christensen, Novo Nordisk
Joachim Leube, Crucell
Mike Long, ConcordiaValsource LLC. John Finkbohner, Ph.D., MedImmune
Maik Jornitz, G-Con Gabriele Gori, Novartis
Russell Madsen, The Williamsburg Group Zena Kaufman, Abbott
Grace McNally, U.S. FDA Michael Sadowski, Baxter Healthcare
Jeanne Moldenhauer, Excellent Pharma Consulting
Christopher Smalley, Merck Junko Sasaki, Dainippon Sumitomo Pharmaceuticals
Brenda Uratani, U.S. FDA Sue Schniepp, OSO BioPharmaceuticals
Glenn Wright, Eli Lilly and Company Lisa Skeens, Ph.D., Baxter Healthcare
Christopher Smalley, Ph.D., Merck
Martin VanTrieste, Amgen
Glenn Wright, Eli Lilly

PDA Staff Contributors

Jahanvi ( Janie) Miller, Senior Project Manager
Richard V. Levy, Ph.D., Senior VP, Scientific & Regulatory Affairs
Walter Morris, Director of Publishing
Katja Yount, Publication Designer
Licensed to Del Toro, Mario/Laboratorios Pisa: Copying and Distribution Prohibited.

Bethesda Towers
4350 East West Highway
Suite 200
Bethesda, MD 20814 USA
Tel: 1 (301) 656-5900
Fax: 1 (301) 986-0296
E-mail: [email protected]
Web site: www.pda.org