3/6/2015

Some Notes on Linear Regression

Regression equation/model
Yi=0 + 1Xi +i
TUTORIALS ON DESIGN EXPERT Where Yi= response variable for the ith element; Xi= predictor variable
for the ith element; 0 = regression coefficient that gives the Y-
BLANCA J. VILLARINO
intercept of the regression line intercept;
1=is a regression coefficient that gives the slope of the line; i= the
random error for the ith element (where the is are independent,
normally distributed with mean 0 and variance 2 for i= 1,2,..n
(number of elements)

Some Notes on Linear Regression Some Notes on Linear Regression residual plot should not
Testing strength of relationship between x and y Diagnostics have a pattern, must be
random
R-squared , adj R-squared
PRESS

Some Notes on Linear Regression Some Notes on Linear Regression

Transformation General criteria for choosing the models and what to include in the
model assumption: alpha = 0.05
p-value of model is <0.05
p-value of factor/s is <0.05
Lack of fit not significant (p>0.05)
High R-squared, pred R-squared (consistent with adj. R-squared) and low
PRESS and s.d.
Error terms (residuals) are independent, normally distributed, mean is 0

1
 3/6/2015

Some Notes on Linear Regression Some notes on experimental design

KISS- Keep it simple, statistically (Keep it simple, stupid)
Less is better- less factors included in the model the better it is to interpret
the model

Hierarchical rule
Keep in the model lower terms (linear) that may be of or involved in higher-
order terms (i.e. interaction or quadratic)

Some notes on experimental design Some notes on experimental design

Statistical appreciation
Variation between experimental units
Manipulation and dilution
Procedural techniques
Blocking
Randomization
Coding and blinding
Controls and baseline
Measurement errors
Suitability of response measurement

PLANNING THE RESEARCH PROBLEM

Reducing experimental errors Factorial design

Statistical appreciation
Variation between experimental units Normally the 2k
Manipulation and dilution design
Procedural techniques
Blocking 2= 2 levels (high and
Randomization low)
Coding and blinding k= no. of factors
Controls and baseline
Measurement errors
Suitability of response measurement

2
 3/6/2015

Fractional factorial design Response surface methodology

Optimization procedures for the settings of factorial variables such that the response
Certain assumptions are made regarding interactive effects (i.e. they
attains a desired maximum or minimum values
are relatively insignificant) Suitable for quantitative factors
Main effects can be evaluated with much fewer trials Modelled by factorial techniques and ANOVA
Two levels which are estimated to bracket the optimum region, are Usually follows the factorial screening and augments the design where extra treatments
used for each factor e.g. fraction of a 25 full factorial (32 runs) can be are added to focus the effects and improve the predictive power of the model
Additional treatments are located with the factorial space (centre point) and outwidth it
performed in 8 runs
(star points)- this structure is known as a central composite design (CCD)
Savings in cost and time with increased efficiency as more variables Regression analysis produce enhanced model and the equation can be used to plot the
can be incorporated response surface
Typically used as screening trials to identify quickly and efficiently the Contour plots not only reveal optimum levels for factors but also regions where a
critical factors before more detailed fuller factorial experiments are done response shows the same magnitude

Response surface methodology

Process of design and analysis with RSM
1. Design and execute an experiment to generate response data
2. Fit the data to a series of polynomial models using tools of regression
3. Conduct ANOVA to assess statistical significance and test for lack of fit.
Compare models on the basis of r-sq adj and r-sq pred but not raw r
sq
4. Choose the simplest model, even if only the mean, that predicst the
response the best. Do not over-fit the data.

What is optimization? Use of desirability in optimization

A way to determine the best process and formulation combinations that Desirability is a measure of success when optimising multiple responses
will provide the desired value of a response and ranges in value from 0 to 1 (least to most desirable, respectively)
One method for statistical optimisation is response surface (Dhinda, Lakshmi, Prakash, & Dasappa, 2012).
methodology (RSM). This approach combines desires and priorities for each of the response
RSM is a systematic approach which will give the greatest amount of and independent variables identified as the basis of optimization.
meaningful information using the lowest number of experiments and at
the lower cost.

3
 3/6/2015

Example:
Factors Optimisation criteria
Goal Limits Weights Importance
A. Independent variables
ASL flour incorporation (g/100 g Maximise 5-40 1 +++++
composite flour)
Volume weighted mean particle Maximise 27-687 1 +
size m)
Mixing time (min) Minimise 4-12 1 +
Baking time (min) Minimise 10-25 1 +

A. Dependent variables
Crumb specific volume (cm3/g) Maximise 3.0-5.6 1 +
Instrumental hardness (g) Minimise 110-222 1 +
Overall acceptability Target=6 5.5-9.0 1 +++++