Central Composite Design

(Manual)

Prepared By: 1
Ruchir Shah, FnD
Index
Basics
• Basics of CCD Slide 3 - 8

Step 1 & 2
• Define Objective & Identify Variables Slide 9 - 11

Step 3
• Create Design Slide 12 - 21

Step 4
• Design Evaluation Slide 23 - 27

Step 5
• Experimentation Slide 28 - 31

Step 6
• Analysis & Interpretation Slide 32 - 39

Step 7
• Diagnostics Slide 40 - 48

Step 8

• Model graphs Slide 49 - 53

Step 9
• Knowledge Space Slide 54 - 57 2
Step 10
• Points to be captured in PDR Slide 32 - 39
 Experiment Design Process
The flow chart below illustrates the experiment design process:

M
o
s
t

I
m
p
o
rt
a
n
t
3
 Components of Experimental Design
• There are three components of the Formula/process that are
analysed by a design of experiments:

• Factors – Independent Variable will form factors. Variables which can be

controlled by an experimenter.

• Levels – Range of each factor in the study or Grade of Excipients

• Response - or output of the experiment.

4
Response Surface Methodology

5
Central Composite Design

Centre Points and Axial Points are added to estimate curvature effect 6
Types of CCD
1. CCC – Circumscribed CCD
• Each factor studied at 5 level
• Axial points establishes new high
and low ranges

2. CCF – Face centered CCD

• Axial points are at center of each
side of factorial space

3. CCI – Inscribed CCD

• For those situations where limit
specified for each factor is real
limit. Beyond or below that level it
is not possible to perform an
experiment.
• Here also Each factor studied at
5 level
• It is a Scaled down of CCC 7
design
Case Study : Fluid Bed Granulation
• Example
Functional Batch Size
Ingredients Category mg/tab %w/w
Dry Mix
Microcrystalline Cellulose (AVICEL PH 102) Diluent 90.00 27.27
Silicon dioxide (Aeoperl 300 Pharma) Adsorbent 10.00 3.03
SSG Type A Disintegrant 8.00 2.42
Drug Solution
Drug Substance Active 90.00 27.27
Povidone K-30 Binder 8.00 2.42
Silicon dioxide (Aeoperl 300 Pharma) Adsorbent 10.00 3.03
SLS Surfactant 5.00 1.52
Ethanol Solvent qs
Dichloromethane Solvent qs
Extra Granular
Microcrystalline Cellulose (AVICEL PH 102) Diluent 92.00 27.88
Silicon dioxide (Aeoperl 300 Pharma) Glidant 3.00 0.91
SSG Type A Disintegrant 8.00 2.42
Talc (M) Anti-adherent 3.00 0.91 8
Magnesium Stearate Lubricant 3.00 0.91
Total Core 330.00 100.00
Step 1 - Define Objective
• Objective:
• Establishing preliminary formulation component levels and demonstrating
the rationale for selection of the excipient levels. The study can also be
utilized to determine if acceptable product attribute are obtained over the
range of excipient concentrations studied.

Note: State Objective in Measurable Responses

 Factor Interaction

 Formula Optimization

 Process Optimization

Never Club Formula and Process Optimization in single Study 9

Step 2 - Identify and Characterize variables
Functional Batch Size
Ingredients Category mg/tab %w/w
Dry Mix
Microcrystalline Cellulose (AVICEL PH 102) Diluent 90.00 27.27
Silicon dioxide (Aeoperl 300 Pharma) Adsorbent 10.00 3.03
SSG Type A Disintegrant 8.00 2.42
Drug Solution
Drug Substance Active 90.00 27.27
Povidone K-30 Binder 8.00 2.42
Silicon dioxide (Aeoperl 300 Pharma) Adsorbent 10.00 3.03
SLS Surfactant 5.00 1.52
Ethanol Solvent qs
Dichloromethane Solvent qs
Extra Granular
Microcrystalline Cellulose (AVICEL PH 102) Diluent 92.00 27.88
Silicon dioxide (Aeoperl 300 Pharma) Glidant 3.00 0.91
SSG Type A Disintegrant 8.00 2.42
Talc (M) Anti-adherent 3.00 0.91 10
Magnesium Stearate Lubricant 3.00 0.91
Total Core 330.00
Identify and Characterize variables…. 2
• Define Variables
• Factors To Be Studied
(Independent Variables - Factors)
• Binder Level - Numerical
• Surfactant Level – Numerical
Total 4 Factors
• Disintegrant Level - Numerical
3 Numerical
• Grade of Adsorbent – Categorical (Nominal)
1 Categorical
• Responses To Be Measured
(Dependent Variables - Responses)
• Dissolution at 30 min (Should Be in Line with Specification)
• Disintegration Time
• Polymorphic Form
• Hardness

• Only the Factors having impact on CQAs should be selected

• Only those Responses having direct/indirect link with CQAs should be selected 11
Step 3: Create Design

12
Central Composite Design
• Start Design Expert Software

13
 Execution of trial with (–) Surfactant level??

By default Software will take “Rotatable Alpha Value”

1. Select Response Surface then Select Central Composite

2. Select no. of Numerical and Categorical factors.
3. Write Name of Factor, Units.
4. Put low and high value of factor in Actual. Never use coded values.
1. Check for any negative values in +/- Alpha.
2. If any of the Factor range is displaying Negative Values, Follow step mentioned on next slide. 14

Always check : Alpha values should not be in Minus as it is not possible to execute trials with minus
values.
 1. Alpha value determines
the position of axial run.
2. Default is Rotatable for up
to 5 factors:
3. Adv.: will have equal std.
error of prediction

Curvature is best estimated by Rotatable Design

Rotatable: Each factor will be tested at 5 level

“Recommended when all factor ranges are within allowable operating condition”
i.e. Alpha range is positive for all factors 15
Face Centered: Each factor will be tested at 3 level
Use when region of interest and region of operability are nearly same.
For Negative Alpha Values – Option 1
(Face Centered – Covers Entire Factor Range)

Steps:
16
1. First click on Option and select Face centered design.
2. Factor ranges will be automatically displayed in Alpha Columns
For Negative Alpha Values – Option 2
(Rotatable – Restricted Operating Range)

Enter same values which you have

entered in Low & High Column

inscribing restricts the actual design region to

the defined variable ranges by locating the
axial points at the lower and upper bounds of
the variable ranges

Steps:
17
1. First click on Option “Enter Factor Ranges in Terms of Alpha”.
2. Enter Factor ranges you want to study in alpha columns
Selected Option 1 (Face Centered – Covers Entire Factor Range)

2
3
1

Steps:
18
1. First click on Option and select Face centred design.
2. Factor ranges will be automatically displayed in Alpha Columns
3. Click “Continue”
Put Name, type and levels of categorical factor

19
Recheck the responses, Correct if required and Click on finish

20
1. Design Layout will be displayed.
(Notice that –alpha and +alpha match the low and high 21
levels in FCD)
Step 4: Design Evaluation
(Before Initiating Experimentation)

22
Evaluate Precision (Slide 1 of 3)
• Use the FDS graph to see if the responses can be estimated
precisely enough: (Calculate Signal/Noise ratio)
• Dissolution:
• Want to estimate Mean Within ± 5%
• Estimated Std. Deviation is 3% 5/3=1.7
• Hardness:
• Want to estimate Mean Within ± 2 kp
• Estimated Std. Deviation is 0.6 kp 2/0.6=3.33
• Disintegration Time:
• Want to estimate Mean Within ± 2 min
• Estimated Std. Deviation is 0.5 min 2/0.5=4

23

Before running the experiments make sure the design meets your needs
 Evaluate Precision (Slide 2 of 3)

When you have multiple responses, look at worst case first (min. ratio). If its OK, all others will also 24
be good. In this case Dissolution (delta/sigma = 5/3 = 1.7) is harder to realize than other responses.

Minimum FDS should be 0.8

Evaluate Precision (Slide 3 of 3)
• Use the Mean error type if the goal of the experiment is to
provide an optimization model.
• Use the Pred error type if the goal of the experiment is to
provide accurate predictions.
• Use the Diff error type as a substitute for power on
response surface and mixture designs.
• Use the Tolerance setting if you need to meet tolerance
specifications.
• FDS should be at least 0.8 or 80% for exploration, and 100%
for robustness testing. FDS can be improved by adding runs,
reducing the "s", increasing the "d", and increasing the "a".
25
Step 5: Perform Experiments and
Input Data

26
Run Experiments
• Since design is precise enough, run the experiments and note the
responses in response column.

You can use “Run Sheet”

while execution of trial
but not mandatory

27
Fill The Response Data….. 1
• Option 1: Enter data manually in Software
• Option 2: Export the file to excel, feed the data and simulate/import to your
existing design
1

28
Fill The Response Data….. 2

29
Step 6: Analysis & Interpretation

30
Data Transformation

31
1. Go to diagnostic tab
2. In Box-Cox plot, check for recommended transformation
Analysis: Fit Summary….. 1

Do not select
aliased model

The Suggested model is selected because:

1. Quadratic model is significant, p-value < 0.05
2. Lack of fit is insignificant, p-value> 0.10 32
3. Having highest adjusted and predicted R-squares.

Never Select Aliased Model

Analysis: Fit Summary….. 2

33
• Focus on the Model maximizing the “Pred R-Squared”, or equivalently
minimizing the “PRESS”
• The diff. between Adjusted R-squared and Pred. R-Squared should be < 0.2
Analysis: Selection of Model

To reduce the model use Backward

selection

• Select finalized model. In this case it is “Quadratic”.

Note:
There are four algorithms for selection, 1) Backward, 2) Forward, 3) Stepwise and 4) All-
hierarchial.
For experiments with little or no colinearity all four will lead to a same reduced model. When the 34
factors are correlated with one other, the reduced model differ depending on the selection method
used.
Backward selection has the advantage of starting with all terms in the model.
Analysis: Anova with backward selection

• The least significant terms will be removed automatically 35

• Model P-value < 0.05
• Model Lack-of-fit > 0.10
Analysis: Full Vs Reduced Quadratic model
Full Model

Reduced Model

36
How do you compare full vs reduced model
Answer: reduction improves lack-of-fit
Analysis: Full Vs Reduced Quadratic model

Full Model

Std. Dev. 3.97 R-Squared 0.8334

Mean 92.22 Adj R-Squared 0.7501

C.V. % 4.31 Pred R-Squared 0.6270

PRESS 919.39 Adeq Precision 11.898

Reduced Model

Std. Dev. 3.77 R-Squared 0.7986

Mean 92.22 Adj R-Squared 0.7756

C.V. % 4.08 Pred R-Squared 0.7466

PRESS 624.61 Adeq Precision 17.424

37
Predicted R-Squared and Adequate Precision improved with Reduced
Model
Step 7: Diagnostic

38
1. Go to diagnostic plot and use diagnostic tool tab to confirm 39
assumptions for ANOVA are met.
2. Check all the plots always
Analysis: Normal Plot & Residual Vs. Predicted Plots

Normal plot of residuals: Residuals Vs. Predicted:

• Check for normality of data. • This plot is used to confirm the
• Residuals should follow a straight constant variance assumption
line. • All the points should be within
• Some scatter is expected. the 3 sigma limits
• Look for definite patterns, e.g. “S • Some scatter is expected 40
Shape”. • Look for definite patterns, e.g.
Megaphone “< Shape”.
 Analysis: Residual vs. Run & Predicted Vs Actual

Residuals Vs. Run: Predicted Vs. Actual:

• Used to confirm the independence • Use to see how the model predicts
assumption. over the range of data.
• All the points should be within the 3 • Some scatter is expected.
sigma limits. • For optimization study it`s not very
• Some scatter is expected. important.
• If any point fall outside 3 sigma limit, • Its imp. When you want to predict
this points are statistical outlier. a response based on change in
• Problem with the model
41
factor level.
• Transformation problem
• Wrong data entered
• Wrong experiment done
Analysis: Box-Cox

Box-Cox plot tells you whether transformation of data may help. 42

Notice that it says “none” for recommended transformation
Analysis: Residual Vs. Factors

GOOD: Random scatter both ends BAD: More variation at one end

Watch ONLY for very large differences. 43

 Analysis: Influence – Cook`s Distance & DFFITs

Cook’s Distance: DFFITS – Difference in Fits:

• Cases with large values relative to other • Look for values outside the
cases should be investigated. blue limits.
• They could be caused by recording • A high value indicates the
errors, an incorrect model or a design predicted response for a
point far from the remaining cases. particular run changes when
the particular run is removed
helps if you see more than one outlier in form the regression. 44
other diagnostic plots. Investigate the run
with the largest Cook’s Distance first.
 Analysis: Influence - DFBETAs

DFBetas: Difference in Betas

45
Used to identify model and data problems.
You must click through the Term Pull down menu to look through all this graphs.
Response Surface - Influence
TOOL Description WIIFM
Residual divided by the Normally constant σ2
Internally Studentized
estimated standard deviation
Residuals
of that residuals
Residuals divided by the Outlier detection and/or
Externally studentized estimated std dev of that model misspecification
Residuals residual, without the ith
Case (deletion statistic)
Change in joint confidence Overall Influence
ellipsoid (regression) with and
Cook’s Distance
without a run (deletion
statistic)
Change in prediction with and Influence on fitted value
DFFits without a run; the influence a
(difference in fits) run has on the predictions
(deletion statistic)
Change in each model Influence on coefficients
DFBetas coefficient(beta) with and 46
(difference in betas) without a run (deletion
statistic)
WIIFM – What is in it for me
Model Graphs

47
Look at different model graphs, i.e. interaction, 2D, 3D, pertubation to
understand the impact of factor on responses.
Interaction Graph

With Aerosil 200, increasing Surfactant significantly improves the dissolution.

While with Aeroperl 300 there is only a slight increase in dissolution with increase in
surfactant level. 48

If the LSD bars for two means overlap, the difference in those means is not large
enough to be declared significant change.
 Interaction Graph

To aid in seeing whether the bars overlap, try right clicking on LSD bar and choosing
“Draw horizontal reference”. 49
Analysis of Each factor
• Analyze all the factors using similar method

50
Develop Good Model

• Be sure that the fitted surface adequately represents your

process before you use it for optimization
1. A significant model: Large F-value with p<0.05
2. Insignificant lack-of-fit: F-value with p>0.10
3. Adequate precision>4.
4. Well behaved residuals: Check diagnostic plots!

51
Creating Knowledge Space
• For multiple response optimization, it`s better to use
Numerical optimization.

52
Data to be captured in PDR
• Goal of the study
• Study name
• Factors studied and corresponding level
• Responses measured
• Design matrix
• Graphs and conclusion – Interaction/Pareto/Contour Etc.
• Result obtained from study as pasted below

53
Data to be captured in PDR…. 2
• ANOVA

54
Thank you !!!!!

55
Annexure

56
 Selection Of Design

Number Comparative Screening Response Surface

of Factors Objective Objective Objective
1-factor completely
1 _ _
randomized design
Randomized block Full or fractional Central composite or
2-4
design factorial Box-Behnken design
Randomized block Fractional factorial or Screen first to reduce
5 or more
design Plackett-Burman number of factors

Interaction
Study
Factorial Design

Curvature RSM - Central 57

Effect Composite Design
Excerpt of QbD IR

58
Variables
Variables are attributes that we want to study in the sample

Variables

Quantitative Qualitative
(Numerical) (Categorical)

Continuous Discrete Nominal Ordinal

e.g. Grade of e.g. severity of

e.g. Hardness, DT, e.g. Dice, coin excipient, Different illness - mild,
pH etc. vendor moderate or severe

59
Variables
• Variables can be further classified as:

• Dependent/Response. Variable of primary interest

(e.g. Dissolution, Hardness ). Not controlled by the
experimenter.

• Independent/Factor
• called a Factor when controlled by experimenter.
(e.g. Disintegrant level, Binder level, Kneading time, Chopper Speed
etc. )

60