Downloaded from journal.pda.

org on June 7, 2014

Risk-based Methodology for Validation of Pharmaceutical

Batch Processes
Frederick Wiles

PDA J Pharm Sci and Tech 2013, 67 387-398

Access the most recent version at doi:10.5731/pdajpst.2013.00923
 Downloaded from journal.pda.org on June 7, 2014

TECHNOLOGY/APPLICATION

Risk-based Methodology for Validation of Pharmaceutical

Batch Processes
FREDERICK WILES, ASQ CQE*

ProPharma Group, Inc., 10975 Benson Dr. Suite 330, Corporate Woods Bldg. 12, Overland Park, Kansas 66210
©PDA, Inc. 2013

ABSTRACT: In January 2011, the U.S. Food and Drug Administration published new process validation guidance for
pharmaceutical processes. The new guidance debunks the long-held industry notion that three consecutive validation
batches or runs are all that are required to demonstrate that a process is operating in a validated state. Instead, the
new guidance now emphasizes that the level of monitoring and testing performed during process performance
qualification (PPQ) studies must be sufficient to demonstrate statistical confidence both within and between batches.
In some cases, three qualification runs may not be enough. Nearly two years after the guidance was first published,
little has been written defining a statistical methodology for determining the number of samples and qualification runs
required to satisfy Stage 2 requirements of the new guidance. This article proposes using a combination of risk
assessment, control charting, and capability statistics to define the monitoring and testing scheme required to show
that a pharmaceutical batch process is operating in a validated state. In this methodology, an assessment of process
risk is performed through application of a process failure mode, effects, and criticality analysis (PFMECA). The
output of PFMECA is used to select appropriate levels of statistical confidence and coverage which, in turn, are used
in capability calculations to determine when significant Stage 2 (PPQ) milestones have been met. The achievement
of Stage 2 milestones signals the release of batches for commercial distribution and the reduction of monitoring and
testing to commercial production levels. Individuals, moving range, and range/sigma charts are used in conjunction
with capability statistics to demonstrate that the commercial process is operating in a state of statistical control.

KEYWORDS: Process capability, Statistical confidence, Statistical coverage, Control chart, PFMECA, Process
validation, Risk-based.

LAY ABSTRACT: The new process validation guidance published by the U.S. Food and Drug Administration in
January of 2011 indicates that the number of process validation batches or runs required to demonstrate that a
pharmaceutical process is operating in a validated state should be based on sound statistical principles. The old rule
of “three consecutive batches and you’re done” is no longer sufficient. The guidance, however, does not provide any
specific methodology for determining the number of runs required, and little has been published to augment this
shortcoming. The paper titled “Risk-based Methodology for Validation of Pharmaceutical Batch Processes” describes
a statistically sound methodology for determining when a statistically valid number of validation runs has been
acquired based on risk assessment and calculation of process capability.

Introduction tical industry about the number of runs required to satisfy

Although published more than 2 years ago, there is still
a considerable amount of uncertainty in the pharmaceu-
* Corresponding Author: ProPharma Group, Inc.,
10975 Benson Dr. Suite 330, Corporate Woods Bldg.
12, Overland Park, Kansas 66210, E-mail: fred.wiles@
propharmagroup.com
doi: 10.5731/pdajpst.2013.00923
process performance qualification (PPQ) as described in
the latest U.S. Food and Drug Administration guidance
on process validation. The old rule of three process
validation runs is out and has now been replaced by a
more robust statistical methodology. The new Guidance
for Industry. Process Validation: General Principles and
Practices (1), published in January 2011, in part states
that
in most cases, PPQ will have a higher level of
sampling, additional testing, and greater scru-

Vol. 67, No. 4, July–August 2013 387

 Downloaded from journal.pda.org on June 7, 2014
tiny of process performance than would be
typical of routine commercial production. The
level of monitoring and testing should be
sufficient to confirm uniform product quality
throughout the batch (emphasis mine).
The guidance goes on to recommend that the PPQ
protocol discuss
the sampling plan, including sampling points,
number of samples, and the frequency of sam-
pling for each unit operation and attribute.
The number of samples should be adequate to
provide sufficient statistical confidence of
quality both within a batch and between
batches. The confidence level selected can be
based on risk analysis as it relates to the
particular attribute under examination. Sam-
pling during this stage should be more exten-
sive than is typical during routine production
(emphasis mine).
The guidance also recommends
continued monitoring and sampling of pro-
cess parameters and quality attributes at the
level established during the process qualifica-
tion stage until sufficient data are available to
generate significant variability estimates (em-
phasis mine).
The guidance states that the number of samples should
be adequate to provide sufficient statistical confidence,
but it does not provide any specific methodology to be
used to determine the number of samples required.
Pharmaceutical companies are left to their own de-
vices to develop methods and procedures that yield a
statistically valid number of PPQ runs and sample
sizes. Unfortunately, there is little published literature
that goes into depth on this subject. This article de-
scribes a specific methodology for deriving PPQ sam-
pling plans under the new guidance. This methodology
is risk-based and provides a mathematical approach
for determining when a sufficient number PPQ batches
have been acquired to demonstrate adequate statistical
confidence in the process validation results.
The methodology described herein was developed spe-
cifically for process performance qualification of liq-
uid batch pharmaceutical processes. However, the
concepts outlined in this article may be adapted to suit
any process where process performance can be ex-
388
pressed using capability statistics. Organizations
adopting this methodology should consult with a stat-
istician or other professional possessing expertise and
experience in statistical process control to determine
the applicability of this methodology for their specific
process.
Risk-based PPQ Methodology for Pharmaceutical
Batch Processes
The first step in any risk-based process validation
strategy is to identify and quantify the risks associated
with each process step. This is best accomplished
through the application of a process failure mode,
effects, and criticality analysis (PFMECA). A
PFMECA is a systematic approach for identifying,
categorizing, and rating failures associated with a
manufacturing process based on their effect on the
process output. PFMECAs are usually prepared by a
cross-functional team consisting of representatives
from at least the research and development, manufac-
turing, and quality functions. Other functions such as
engineering and quality control may be included in the
team as well.
Prior to developing the PFMECA, a map of the pro-
cess illustrating the process steps in the order that they
occur should be created. The process map can be a
simple block diagram, but should identify process
inputs such as the equipment used, raw materials, and
process parameters required for manufacture of the
product. In addition to the process inputs, the pro-
cess flow diagram should identify key process out-
puts which, for pharmaceutical batch processes,
take the form of in-process and batch release test-
ing. Once created, the process flow diagram will be
used as a reference for constructing the PFMECA.
An example of a simple process flow diagram is
provided in Figure 1.
After finalizing the process flow diagram, the next step
in the development of the PFMECA is to procure data
that can be used to identify potential risks to the
process output. Primary sources of data for input to the
PFMECA are process characterization studies. Process
characterization involves challenging the process un-
der varying manufacturing conditions and process pa-
rameter settings in order to gain first-hand knowledge
of how a process will perform under different stresses.
The output of process characterization can be used to
identify key process inputs affecting product quality
and yield. Evaluation of process performance at input
PDA Journal of Pharmaceutical Science and Technology
 Downloaded from journal.pda.org on June 7, 2014

Figure 1

Example process flow diagram for a simple pharmaceutical batch compounding process.

extremes expose weak points in the process where screening studies, response surface analysis, and anal-
additional controls may be needed to reduce the risk to ysis of variance (ANOVA), to name a few, can be used
the finished product and defines process limits beyond to quantify the effect of input variation on the process
which the process is no longer viable. Tools such as outputs. Other data sources such as product develop-

Vol. 67, No. 4, July–August 2013 389

 Downloaded from journal.pda.org on June 7, 2014

TABLE I
Example PFMECA Matrix. The RPN Value in the Rightmost Column Is the Product of the Ratings of the
Three Risk Categories, Severity, Occurrence, and Detection

CONTROLS IN PLACE
FAILURE MODE
FUNCTION
ITEM NO.

PROCESS

EFFECTS

CAUSE

OCC

DET

RPN
SEV
List controls
List List in place to
List specific effect(s) of List causes mitigate the
Description process failure each failure for each probability of
of process functions modes mode failure mode occurrence 1–3 1–5 1–5 1–75

ment and past production history also provide useful
input to the PFMECA.
In preparing the PFMECA, the risks associated with
each process step are quantified to produce a risk
priority number (RPN). Quantization is achieved by
categorizing and rating the individual risks on a nu-
merical scale. The risk categories rated are arbitrary
but are usually chosen as the severity of the risk, the
probability of detection, and the probability of occur-
rence. The RPN value is the product of the ratings
from each of the risk categories. An example
PFMECA matrix is provided in Table I.
RPN values established in the PFMECA are used to
determine the levels of statistical confidence, ␥, and
coverage, p, required for the PPQ study of interest.
Confidence and coverage values are chosen based on
the highest RPN value exhibited by the process. A
matrix may be created for this purpose where specific
RPN ranges correspond to predetermined combina-
tions of confidence and coverage. Such a matrix
should be documented in the company’s validation
and/or risk assessment procedures. Example risk rat-
ing system matrixes are provided in Tables II, III, IV,
and V. Only after completing the PFMECA and es-
tablishing appropriate levels of statistical confidence
and coverage can the process progress to PPQ.
PPQ is performed in two phases reflecting recom-
mendations put forth in Stage 2 and Stage 3 of the
process validation guidance. In the first phase of

TABLE II
Example Ranking Scale for Probability of Occurrence. The Probability of Occurrence Is Measured on a
Scale of 1 to 3. The Highest Ranking Is Used When the Probability Is Unknown

Probability of Occurrence
Ranking Description Comments
1 Low Not likely to occur (Probability ⬍ 0.001)
2 Moderate Could occur (0.1 ⬎ Probability ⬎ 0.001)
3 High Likely to occur at some time (Probability ⬎ 0.1)

390 PDA Journal of Pharmaceutical Science and Technology

 Downloaded from journal.pda.org on June 7, 2014

TABLE III
Example Ranking Scale for Severity of Effect. The Severity of Effect of a Failure Is Measured on a Scale of
1 to 5. The Highest Ranking Is Used When Severity Is Unknown

Ranking Description
1 Minor
2 Low
3 Moderate
4 High
5 Very High
Severity of Effect
Comments
Unreasonable to expect any real or measurable effect on the safety and
efficacy of the finished product. The failure is unlikely to result in a
customer complaint.
Failure could lead to a slight deterioration of the product appearance but
is not likely to affect the safety or efficacy of the product.
Failure that easily affects a product or process. Product deterioration
is likely and some customer dissatisfaction and/or annoyance is
expected
Failure resulting in a high degree of product non-performance and
extreme customer dissatisfaction. Safety or efficacy of the product may
be impacted. Immediate corrective actions are required.
Failure seriously affects the safety or efficacy of the product resulting in
product destruction, recall, or other regulatory action.

PPQ, a number of initial runs or batches are exe-
cuted. Data collected in the first PPQ phase are used
to establish preliminary estimates of process control
and capability. Product batches undergoing PPQ are
not released until successful completion of this
phase. The second phase of PPQ is a continuation of
the PPQ sampling and testing plan, but allowing for
release of PPQ batches for commercial distribution.
The purpose of the second phase PPQ testing is to
acquire additional data to bolster the initial capabil-
ity assessment and to show long-term statistical
control of the process.
Statistical control is best demonstrated through the
use of a control chart. For pharmaceutical batch
processes, the most appropriate chart is the individ-
uals, moving range, and range/sigma (I-MR-R/S)
chart. In a bulk pharmaceutical batch, samples are
usually taken from different levels within the blend-
ing vessel to demonstrate bulk product uniformity.
However, unlike a discrete manufacturing process
such as a filling operation where items are produced
one at a time, a bulk pharmaceutical batch is man-
ufactured as a whole at one time. Thus, the within-
batch variation in a batch process is usually so much
smaller than the between-batch variation; the con-
trol limits on a standard X-bar and R/S chart will be
too close together (2). In an I-MR-R/S chart, the
average of each batch characteristic charted is con-
sidered a single data point. The control limits are
based on the batch-to-batch variation (as indicated
by the MR chart) rather than the subgroup variation
as in a standard X-bar and R or X-bar and S chart,
thus providing a more reasonable representation of
the batch-to-batch control. The R/S portion of the
I-MR-R/S chart captures the within-batch varia-

TABLE IV
Example Ranking Scale for Probability of Detection. The Probability of Detection Is Measured on a Scale
of 1 to 5. The Highest Ranking Is Used When Severity Is Unknown

Ranking Description Comments

1 Very High
2 High
3 Moderate
4 Low
5 Very Low
Will catch failure every time.
Has a good chance of catching failure.
May catch failure.
Poor chance of finding failure.
Very poor or no chance of finding failure.

Vol. 67, No. 4, July–August 2013 391

 Downloaded from journal.pda.org on June 7, 2014

TABLE V
Example Matrix Associating RPN Ranges with Values of Statistical Confidence and Coverage. Values
Populating the Matrix Should Be Consistent with Each Company’s Internal Risk Management Procedures

RPN Risk Rating Statistical Confidence (␥) Statistical Coverage (p)

1–25 Low 0.90 0.950
26–50 Medium 0.95 0.990
51–75 High 0.99 0.999

tion. An example I-MR-R/S chart is provided in fowitz (4) to calculate approximate two-sided toler-
Figure 3. ance limits. This formula is expressed as

冑
The within- and between-batch variation should be
mn ⫺ 1
monitored using the I-MR-R/S chart throughout both k2 ⫽ r (3)
␹␥,mn⫺1
2
phases of the PPQ. In fact, the new process validation
guidance strongly encourages control chart monitoring
throughout the life of the product in order to rapidly where
detect changes in the state of process control.
n is the batch sample size
Upon completion of the second and all subsequent
PPQ batches, process performance capability is calcu- m is the number of validation batches
lated. Process performance capability is expressed us-
ing the PPK (long term) index (3): ␹␥,mn⫺1
2
is the critical value of the chi-square distribu-
tion with mn –1 degrees of freedom surpassed

P pk ⫽ Min 冋 USL ⫺ X X ⫺ LSL

3␴ˆ LT
,
3␴ˆ LT
册 (1)
with probability ␥, the statistical confidence
obtained from the matrix in Table V

where r is the interval covering p percent of the nor-

mal distribution found by iteration of eq 4:

冕
X is the grand mean of the process data 1
1 ⫹r
冑 mn
p ⫽ e⫺t / 2 dt
2
(4)
USL is the upper specification limit 冑2␲ 1
⫺r
冑 mn
LSL is the lower specification limit
where p is the statistical coverage obtained from the
␴ˆ LT is the long term process sigma matrix in Table V.

Historically, statistics practitioners have required at
least 25 subgroups before attempting to calculate pro-
cess capability. However, due to the limited number of
runs associated with batch pharmaceutical processes,
capability indices in this methodology are calculated
using a variable multiple of sigma rather than the
standard ⫾3 sigma typically used for time-ordered
processes. Substituting the variable, k2, for the multi-
ple of sigma in eq 1 gives
The NIST/Sematech e-Handbook of Statistical Methods
(5) provides an example of how eq 4 can be easily
evaluated using the NORMDIST function in Microsoft
Excel. In a spreadsheet enter the following cell informa-
tion:
Cell A1 ⫽ 0 (starting value for r)
Cell A2 ⫽ Sample size, n, times the number of vali-
dation batches, m

P pk ⫽ Min 冋
USL ⫺ X X ⫺ LSL
k2 ␴ˆ LT
,
k2 ␴ˆ LT
册 (2) Cell A3 ⫽ the desired statistical coverage, p

One way to determine the value of k2 in eq 2 utilizes Cell A4 ⫽ NORMDIST(1/SQRT(A2)⫹A1,0,1,TRUE)-

a variation of a formula proposed by Wald & Wol- NORMDIST(1/SQRT(A2)-A1,0,1,TRUE)

392 PDA Journal of Pharmaceutical Science and Technology

 Downloaded from journal.pda.org on June 7, 2014
The value of p, then, is found by iteration using the
Microsoft Excel Solver add-in. Using Solver; the cell
containing the value for r (Cell A1) is repeatedly
changed until the calculated p (Cell A4) matches the
desired p (Cell A3). The value of r yielding the desired
p is the value of r used in eq 3.
Equation 3 takes into account the statistical degrees of
freedom provided by the sample size and number of
PPQ batches to yield the factor, k2, covering the de-
sired proportion of the distribution at the desired con-
fidence. Because the value ␹␥,mn⫺1
2
increases with increas-
ing degrees of freedom, k2 decreases as the number of
PPQ runs increases. Consequently, even though a pro-
cess may initially yield a PPK lower than desired, the
decreasing values of k2 on subsequent PPQ runs will
yield successively improved PPK values.
The first phase of PPQ is considered acceptable when
the calculated PPK indices for the product character-
istics examined are greater than a pre-established min-
imum capability. Upon successful completion of
Phase 1, PPQ batches may be released, but sampling
and testing must continue at PPQ established levels as
the PPQ study progresses into Phase 2. In any event,
the minimum number of PPQ runs prior to product
release regardless of the capability estimate should be
no less than three in order to demonstrate a prelimi-
nary state of statistical control.
During Phase 2, PPK indices are continually recalcu-
lated as data for each new PPQ batch becomes avail-
able. The number of runs required to satisfy Phase 2 is
determined based on a calculation of a confidence
interval around the PPK index. This interval is approx-
imated below by eq 5 (3):
冋 P PK ⫺ Z1⫺␣/ 2 冑1
⫹
P 2
PK
9mn 2共mn ⫺ 1兲 册 ⱕ PPK
冋 冑 册
2
1 PPK
ⱕ PPK ⫹ Z1⫺␣/ 2 ⫹ (5)
9mn 2共mn ⫺ 1兲
Similar to PPK, the span of the capability confidence
interval will generally decrease as the number of PPQ
runs increases.
For PPQ, we are interested only in the lower capability
confidence interval bound (LCCB) represented by the
left-most bracketed term in eq 5. Also, because a
variable coverage factor is used to determine the mul-
tiple of sigma required in the capability calculation,
Vol. 67, No. 4, July–August 2013
the value 9 in the denominator of the first term under
the square root is replaced with k22, the square of the
coverage factor found using eq 3. The LCCB, then, is
defined as
LCCB ⫽ PPK ⫺ Z1⫺␣/ 2 冑 1
2
2
⫹
2
PPK
k mn 2共mn ⫺ 1兲
(6)
If the process is capable and in a state statistical
control, at some number of PPQ runs the increasing
value of PPK and decreasing span of the capability
confidence interval will yield a LCCB greater than the
pre-established minimum. At this point PPQ runs can
cease and sampling can drop to normal production
levels.
Because the value for k2 is dependent on the number of
samples pulled from individual validation batches, the
number of validation batches needed to achieve an
LCCB greater than the minimum capability specifica-
tion can be reduced by selecting a larger batch sample
size. Most commercial statistical software packages
provide sample size calculators that can be used to
estimate an appropriate sample size for individual
PPQ batches.
Simulation
A study was conducted using simulated PPQ data to
demonstrate the PPQ methodology described above.
The simulated PPQ data are listed in Table VI. The
data were calculated about a mean of 60 and a
standard deviation of 1.275 using the random nor-
mal distribution function in Minitab ® statistical
software.
Prior to assessing the simulated data statistically, a
probability plot was prepared using Minitab® statisti-
cal software to test the assumption of normally dis-
tributed data (Figure 2). The probability plot shows
graphically how well the data conforms to the normal
distribution. On a probability plot, the y-axis repre-
sents the cumulative percentage of the data distribu-
tion while the x-axis represents the value of the char-
acteristic being measured. Data that are normally
distributed will fall on or very close to a straight line
on the graph, with the majority of the data clustered
symmetrically about the 50th percentile. Looking at
Figure 2, one can see that the simulated data plot
closely follows a straight line. Another way to test the
assumption of normality is to apply the Anderson-
Darling test. The Anderson-Darling test yields a P-
393
 Downloaded from journal.pda.org on June 7, 2014

TABLE VI
Normally Distributed Data for Fifteen Simulated Batches; Mean ⴝ 60, Standard Deviation ⴝ 1.275

Sample 1 Sample 2 Sample 3 Sample 4 Sample 5 Sample 6

Batch 1 61.98599 59.54588 60.86150 59.79496 59.23867 58.33622
Batch 2 62.73705 59.91604 61.84170 59.27713 56.93497 59.19581
Batch 3 60.47101 61.75766 60.82654 61.78138 61.06394 61.21538
Batch 4 61.88888 59.84157 59.63065 57.99141 60.29579 58.58291
Batch 5 62.37012 60.48514 61.56628 58.71746 61.63891 57.30216
Batch 6 60.26066 59.57322 61.09212 59.84095 59.46014 59.61616
Batch 7 61.31893 59.59596 61.44147 60.54807 60.99950 59.92599
Batch 8 59.81401 59.39050 60.70978 58.99141 62.79123 60.36888
Batch 9 60.58472 61.95316 59.83788 58.73686 61.44967 58.41934
Batch 10 60.20420 59.52094 60.04476 59.56099 58.92952 59.67628
Batch 11 61.12075 61.36394 61.99651 57.59433 63.60003 59.75626
Batch 12 61.79105 58.50612 60.76354 60.66932 60.46674 61.08246
Batch 13 60.60140 59.97265 61.91600 61.41712 59.38803 59.11069
Batch 14 59.87057 59.95486 61.56102 59.14503 60.09704 59.81568
Batch 15 62.18290 59.07597 60.07380 59.55689 61.47939 59.10225

value that, when compared to the chosen significance
level, determines whether or not the assumption of
normality should be rejected. The significance level,
␣, chosen in this case is 0.05. Any value for P greater
than the significance level indicates that the data are
likely normally distributed. The probability plot pre-
pared for the simulated data in Table VI yielded a
P-value of 0.420; therefore, we would not reject the
assumption that the data are normally distributed.
Normally distributed data are crucial if a reliable assess-
ment of process capability is to be obtained. In the event
that the assumption of normality is rejected, the cause of
the non-normal variation should be investigated and
eliminated prior to continuing with further PPQ runs.
There are situations where non-normal data cannot be
avoided and, indeed, may even be desired. Consider a
product characteristic for which the total improvement

Figure 2

Normal probability plot of the simulated PPQ data created using Minitab® statistical software. The P-value is
greater than 0.05; therefore the assumption of normality cannot be rejected.

394 PDA Journal of Pharmaceutical Science and Technology

 Downloaded from journal.pda.org on June 7, 2014

Figure 3

Control charts for simulated PPQ data plotted using Minitab® statistical software. All data points are within
statistical control limits. Note that the lower chart is an S chart. For subgroup sample sizes less than 10, an R
chart may be displayed instead.

achievable is constrained due to a physical limit of the
characteristic. Such a characteristic will often exhibit
an increasingly skewed distribution as the physical
limit is approached (3). One example of this situation
is microbial testing. Process improvements may re-
duce the microbial load of a product, but the load can
never be less than zero. A steadily improving process
yields successively lower microbial counts, but as the
average number of counts approaches zero, the distri-
bution of counts becomes more asymmetric about the
average with the shorter tail being truncated at zero
and the longer tail trailing off in the positive direction
without limit.
Fortunately, if the deviation from a normal distribu-
tion is not too severe, there are transformations that
can be applied to make the data emulate a normal
distribution. Transformation involves finding a math-
ematical function that, when applied to the non-normal
data, yields a data distribution more closely following
that of the normal distribution. For example, if the
data distribution exhibits a long tail to the right as is
characteristic of a log-normal distribution, taking the
logarithm of the data in many cases will yield a normal
distribution (6). There are many functions available
for transforming non-normal data. The Johnson Trans-
formation or Box-Cox families of transformation
functions are popular examples. Whatever methodol-
ogy is chosen, it is important to remember to transform
any specification limits along with the process data in
order to maintain a consistent scale.
After testing the assumption that the simulated PPQ
data are normally distributed, an I-MR-R/S chart (Fig-
ure 3) was prepared to verify that the data met statis-
tical control criteria. Whether or not a control chart
indicates that a process is in statistical control is
usually determined by applying tests for special causes
of variation. The most commonly known tests for
special causes of variation are the Western Electric

Vol. 67, No. 4, July–August 2013 395

 Downloaded from journal.pda.org on June 7, 2014

TABLE VII
The Process Is Considered Out of Statistical Control if Any of These Test Conditions Are True

Test 1 One point more than 3 ␴ from center line

Test 2 Six points in a row, all increasing or all decreasing
Test 3 Fourteen points in a row, alternating up or down
Test 4 Eight points in a row on the same side of the center line
Test 5 Two out of three points more than 2 ␴ from the center line (same side)
Test 6 Four out of five points more than 1 ␴ from the center line (same side)

Company (WECO) Rules (7). The WECO Rules con-
sist of a number of tests or conditions that indicate
when a process has become unstable. The WECO rules
are summarized in Table VII.
Table VIII provides a summary of the capability sta-
tistics calculated for the simulated PPQ data. The
minimum capability criterion chosen for the simulated
PPQ is 1.00. The coverage factor, k2, was calculated
for each batch using the values 0.95 for ␥ and 0.99 for
p. With a specification range of 54 – 66, the initial
capability estimate shows the simulated process to be
just barely capable at PPQ run 3. Additional runs were
required to establish the LCCB above the minimum
capability criterion. At run 7, the LCCB also rises
above the minimum capability criterion, thus complet-
ing the second PPQ phase. Figure 4 illustrates graph-
ically the relationship between PPK, the LCCB, and k2
relative to the minimum acceptable capability.
In Figure 4, one can see the asymptotic decrease in k2 is
roughly mirrored by the simultaneous increase in PPK
and the LCCB. A narrowing of the gap between the
capability estimate and the LCCB can also be seen as the
chart progresses to the right. This narrowing is the result
of increasing sample size as additional PPQ runs are
added, which in turn yields ever tighter estimates of PPK.
Although only seven PPQ batches were required to sat-
isfy the minimum capability criterion, Figure 4 shows data
from 15 batches to illustrate the concepts described above.
The actual number of batches required will ultimately be
determined by the number of samples pulled from each
batch and the capability of the process under study.
Conclusion
The methodology presented in this paper provides a
statistically sound approach for determining the num-
ber PPQ runs required to demonstrate that a process is

TABLE VIII
Capability Results for Simulated Batch Data

Batch Number k2 PPK LCCB

1 5.77 0.80 0.29
2 4.15 0.88 0.49
3 3.70 1.03 0.66
4 3.48 1.15 0.80
5 3.35 1.13 0.82
6 3.26 1.26 0.95
7 3.19 1.34 1.04
8 3.14 1.37 1.08
9 3.10 1.40 1.12
10 3.07 1.48 1.20
11 3.04 1.39 1.14
12 3.01 1.42 1.18
13 2.99 1.45 1.21
14 2.97 1.50 1.26
15 2.96 1.52 1.28

396 PDA Journal of Pharmaceutical Science and Technology

 Downloaded from journal.pda.org on June 7, 2014
Figure 4
Graphical representation of simulated PPQ capability analysis showing the relationship between PPK, LCCB,
and k2 with respect to the minimum capability criterion.
capable and in a state of statistical control regardless
of the availability of historic process data. Further-
more, this methodology allows the number of PPQ
runs to be derived mathematically based on an assess-
ment of the risks associated with the manufacturing
process, thereby satisfying agency expectations for
process validation under the new process validation
guidance.
In most cases where a process is highly capable and
the process is in a state of statistical control, the LCCB
will rise above the minimum capability specification
after only a few PPQ batches. Selecting a larger batch
sample size can also help reduce the number of PPQ
batches needed to achieve an LCCB above the mini-
mum capability specification.
While the methodology outlined in this paper was
designed specifically for application with liquid
batch pharmaceutical processes in which the mea-
sured variables are continuous, the concepts dis-
cussed in this paper can be easily adapted to work
with discrete processes by substituting control
charts and formulas for P pk and LCCB based on
attribute data.
Vol. 67, No. 4, July–August 2013
Acknowledgements
The author thanks Dr. Christopher Holloman of Ohio
State University and Dr. Ramon Burns of ProPharma
Group, Inc. for their timely review and invaluable
suggestions, and Bob Beall of ProPharma Group, Inc.
for helping to pull it all together.
Conflict of Interest Declaration
The author is a provider of validation consulting ser-
vices to the pharmaceutical, medical device, and bio-
logics industries.
References
1. FDA. Guidance for Industry. Process Validation:
General Principles and Practices. Office of Com-
munications, Division of Drug Information, Silver
Spring, MD, 2011.
2. Automotive Industry Action Group, American So-
ciety for Quality Control, Supplier Quality Re-
quirements Task Force. Fundamental Statistical
Process Control: Reference Manual. AIAG:
Southfield, MI, 1991.
397
 Downloaded from journal.pda.org on June 7, 2014
3. Both, D. R. Measuring Process Capability. Land-
mark Publishing, Inc.: Cedarburg, WI, 2001.
4. Wald, A; Wolfowitz, J. (1946). Tolerance limits for a
normal distribution. Annals of Mathematical Statistics
1946, 17 (2), 208 –215; http://projecteuclid.org/
euclid.aoms/1177730981, retrieved October 30, 2012.
5. NIST/SEMATECH e-Handbook of Statistical Meth-
ods; http://www.itl.nist.gov/div898/handbook/prc/
398
section2/prc254.htm, retrieved September 07,
2012.
6. Pyzdek, T. The Six Sigma Handbook. McGraw-
Hill Companies, Inc.: New York, 2000.
7. NIST/SEMATECH e-Handbook of Statistical
Methods; http://www.itl.nist.gov/div898/hand-
book/pmc/section3/pmc32.htm, retrieved Septem-
ber 7, 2012.
PDA Journal of Pharmaceutical Science and Technology
 Downloaded from journal.pda.org on June 7, 2014

An Authorized User of the electronic PDA Journal of Pharmaceutical Science and

Technology (the PDA Journal) is a PDA Member in good standing. Authorized Users are
permitted to do the following:

·Search and view the content of the PDA Journal

·Download a single article for the individual use of an Authorized User
·Assemble and distribute links that point to the PDA Journal
·Print individual articles from the PDA Journal for the individual use of an Authorized User
·Make a reasonable number of photocopies of a printed article for the individual use of an
Authorized User or for the use by or distribution to other Authorized Users

Authorized Users are not permitted to do the following:

·Except as mentioned above, allow anyone other than an Authorized User to use or access the
PDA Journal
· Display or otherwise make any information from the PDA Journal available to anyone other
than an Authorized User
·Post articles from the PDA Journal on Web sites, either available on the Internet or an Intranet,
or in any form of online publications
·Transmit electronically, via e-mail or any other file transfer protocols, any portion of the PDA
Journal
·Create a searchable archive of any portion of the PDA Journal
·Use robots or intelligent agents to access, search and/or systematically download any portion
of the PDA Journal
·Sell, re-sell, rent, lease, license, sublicense, assign or otherwise transfer the use of the PDA
Journal or its content
·Use or copy the PDA Journal 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 Journal
·Make any edits or derivative works with respect to any portion of the PDA Journal 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 contained in the PDA Journal