CE-72057
Transportation Planning
Data and Space
Mg Kyaw Kyaw Lwin
ME.CTE-2
Department of Civil Engineering
YTU 1
Fractional Factorial Design
Example 3.12
Consider a situation with five attributes
two at two levels and the rest at three levels
depending on the number of interactions to be tested, the number of
options required would vary as follows
108 to consider all effects (i.e. a full factorial design);
81 to consider principal effects and all interactions between pairs of
attributes, ignoring effects of a higher order;
27 to consider principal effects and interactions between one attribute
and all the rest;
16 only if no interactions are considered.
2
Fractional Factorial Design
Beside the Orthogonal Design, several researchers have suggested
other types of fractional factorial designs such as
D-optimal designs
D-efficient designs
D-optimal designs
is a statistical concept related to experimental design
It aims to maximize the precision of parameter estimates in a statistical
model.
helps in selecting the most informative set of experimental conditions to
improve the efficiency of parameter estimation.
commonly used in fields like industrial experimentation, clinical trials, and
optimization studies.
3
Fractional Factorial Design
D-efficient designs
similar to D-optimal design, focuses on maximizing the
efficiency of parameter estimation in statistical experiments.
A D-efficient design seeks to achieve high precision for
estimating specific parameters of interest while considering
practical constraints.
such as resource limitations or experimental restrictions.
It involves selecting the most effective combination of
experimental conditions to enhance the overall efficiency of
the study.
4
A Note on Dummy, Effects and Orthogonal Coding Data
If the marginal impact upon utility is believed to be non-linear
from one attribute level to another, the analyst may wish to
test this by transforming the data using
Dummy coding
Effects coding
Orthogonal coding
the former remains the preferred method
effects and orthogonal coding offer a number of advantages
over it.
5
A Note on Dummy, Effects and Orthogonal Coding Data
Dummy coding
A technique used in statistics and regression analysis to
represent categorical variables with two or more categories as
numerical values.
It involves creating binary (dummy) variables where each
category is represented by a 0 or 1.
For a categorical variable with \(k\) categories, \(k-1\) dummy
variables are created.
This helps incorporate categorical information into regression
models, making it easier to interpret the impact of different
categories on the dependent variable.
6
A Note on Dummy, Effects and Orthogonal Coding Data
Effect coding
like dummy coding, is a method used in statistics to represent
categorical variables in regression analysis.
Unlike dummy coding where one category is the reference,
effect coding compares each category to the overall mean.
In effect coding, the coefficients represent the average
difference between the levels of a categorical variable and the
overall mean of the dependent variable.
This coding scheme is useful when you are interested in
understanding the average effect of each category, and it is
particularly common in analysis of variance models.
7
A Note on Dummy, Effects and Orthogonal Coding Data
Orthogonal coding
Orthogonal coding is a method used in experimental design
and statistical analysis to create independent and uncorrelated
contrasts or comparisons between groups or levels of a
categorical variable.
Unlike dummy coding or effect coding, where the coded
variables may be correlated, orthogonal coding ensures that
the coded variables are orthogonal, meaning that their
correlations are zero.
Orthogonal coding is especially valuable in situations where
independence of contrasts is important for statistical analyses.
This coding approach aids in obtaining precise and
interpretable results when comparing different levels of a
categorical variable. 8
A Note on Dummy, Effects and Orthogonal Coding Data
9
10
11
Generating the Experimental Design
there are several different approaches that one might employ to generate a
workable experimental design.
all of which reflect the analyst own beliefs about what are the most
important properties for a SC experimental design to display.
Traditional Orthogonal Designs Methods
D-Optimal Design Method Under the Null Hypothesis
D-Efficient DesignMethods Under the Non-Null Hypothesis
Measuring Statistical Efficiency
A Note on Blocking of Designs
A Note on the Need for Prior Information in Generating Designs
A Note on Interaction Effects and SC Designs
12
Generating the Experimental Design
Traditional Orthogonal Designs Methods
historically the most common experimental design types.
Orthogonality is related to the correlation structure of the design attributes.
By forcing them to have zero correlations, each attribute is independent of
all others.
Several methods for constructing orthogonal designs exist in practice
generating balanced incomplete blocked designs (BIBD)
Latin Squares designs
orthogonal in the differences fractional factorial designs
fold-over designs
a number of useful websites and software are available for obtaining
orthogonal designs.
several software packages such as SPSS, SAS and Ngene 1.0 are also able
to generate a range of orthogonal designs.
13
Generating the Experimental Design
D-Optimal Design Method Under the Null Hypothesis
New optimality criterion has been developed to construct optimal
orthogonal SC designs specifically generated for MNL models using
orthogonal codes.
These designs maintain (within alternative) orthogonality, whilst also
minimizing under the assumption that the parameters will be zero and the
attributes will be orthogonal coded.
this typically results in the attribute levels across alternatives being made as
different as possible.
these designs will generally increase the trade-offs that respondents are
forced
to make across all attributes maximising the information obtained in terms
of the importance that each attribute plays on choice.
14
Generating the Experimental Design
D-Efficient DesignMethods Under the Non-Null Hypothesis
Under a non-null hypothesis, the D-efficient design method aims to
optimize the efficiency of parameter estimation when there are specific
effects or differences expected between groups or conditions.
The goal is to design experiments that provide the most information about
the parameters of interest, enhancing the ability to detect and estimate non-
null effects.
In this context, D-efficient design involves selecting experimental
conditions or factors that maximize precision in estimating model
parameters associated with the anticipated non-null hypothesis.
By doing so, researchers increase the sensitivity of their studies to detect
and accurately estimate the effects they are interested in, making the
experimental design more powerful for hypothesis testing and drawing valid
15
conclusions.
Thank you for your kind attention!
16
