Professional Documents
Culture Documents
Efficient Designs Lecture 3 (SEN1221 - Part II) : Eric Molin - TU Delft
Efficient Designs Lecture 3 (SEN1221 - Part II) : Eric Molin - TU Delft
Delft
University of
Technology
• characteristics
• Orthogonality (no correlations among attributes)
• assure high reliability of parameters
• attribute level balance
• all levels are observed an equal number of times
• sequential construction
• no correlations within each alternative, but correlations
between alternatives
• apply for unlabeled alternatives
• all alternatives have the same attributes & levels
• simultaneous construction
• all correlations are 0
• apply for labeled alternatives
• attributes & levels (may) differ among alternatives
• allows estimation alternatieve specific parameters
• labels represent attributes that are not varied
• captured by alternative specific constant
Regret in Traveler Decision Making 3
3
Focus this lecture
• efficient designs
• experimental designs based on prior information
• advantages
• helps avoiding dominance
• may increase reliability parameters
• may reduce number of choice sets
• helps determining number of required respondents
• disadvantage
• if prior information is incorrect, parameters may be biased
• see lecture 6
Regret in Traveler Decision Making 4
4
why efficient designs?
stated choice
Regret experiments
in Traveler Decision Making 5
Assignment 1
• choice set 1 A B
• time in min. 10 20
• costs €1 €2
• choice set 2 A B
• time in min. 10 20
• costs €2 €1
• choice set 1 A B
• time in min. 10 20
• costs €1 €2
• choice set 2 A B
• time in min. 10 20
• costs €2 €1
Regret in Traveler Decision Making 7
7
dominant alternatives
• alternative
• efficient designs may reduce the problem
• efficient designs
• to maximize information about trade-offs
• to minimize standard errors of parameters
• illustration:
• bcosts = -0.241 (in 100 euro) ; s.e. = -0.084
• 95% confidence interval: {-0.406, -0.076}
orthogonal design
efficient design
stated choice
Regret experiments
in Traveler Decision Making 12
priors
• priors
• best guesses on parameter values
• needed to balance utilities of the choice alternatives
• also needed:
• model (data generating process) & utility function
• we assume the model is MNL
• hence, a design is only efficient for a particular model
• and if priors are correct
Regret in Traveler Decision Making 13
13
Ngene code for efficient designs
Note: these are all linear parameters for interval & ratio level variables
Regret in Traveler Decision Making 14
14
Assignment 2
prior values placed in brackets: [ ]
1. Three alternatives
2. Yes, b1 and b5, which have different prior values
3. Alt.3 no prior included, so its constant = 0
4. All attributes have the same levels across the alternatives
• but B is not included in alt.3
5. b2 is estimated for all three alternatives
b3 is generic for alt.1 and alt.2 but does not play a role in alt.3
6. b4, b6, and b7 estimated alternative specific for attribute C
7. 7: five attribute parameters : b2, b3, b4, b6 and b7 and two
constants: b1 and b5 Regret in Traveler Decision Making 16
16
iterative process • Because in the syntax we
asked to minimize the D-
error (see later), designs are
evaluated on this criterion
stated choice
Regret experiments
in Traveler Decision Making 18
predicting choice probabilities
• allows:
• checking whether dominance occurs
• calculating utilities & choice probabilities in choice sets
• determining the required number of respondents
• calculate standard errors of parameters
• steps:
• calculate expected utilities
• substitute design attribute levels in utility function
• calculate expected probabilities by MNL
• check for extreme values
across all choice sets, each alternative has a fair chance of being
Regret in Traveler Decision Making 27
chosen 27
number of respondents
stated choice
Regret experiments
in Traveler Decision Making 28
Number of respondents required
• question of interest
• how many respondents are needed to assure that
parameters are statistically significant?
• s.e.
• is normally only known if you have data
• however, based on the priors we can predict utilities and
therefore choices for all constructed choice sets
• and thus calculate s.e. for every
Regret inparameter for
Traveler Decision N = 1 29
Making
29
how to obtain standard errors?
• t(N=1) = b / s.e.
• b = prior
• expected s.e.
stated choice
Regret experiments
in Traveler Decision Making 36
number of choice sets
• if option A is chosen:
• choice set {A,B} A ‘is better’ than B
• adds 1 degree of freedom
• choice set {A,B,C} A ‘is better’ than B and C
• adds 2 degrees of freedom
• choice set {n alt.} A ‘is better’ than n-1 alt.
• adds n-1 degrees of freedom
• examples
• 8 can be divided by 2 and 4, but not by 3
• 9 can be divided by 3, but not by 2 and 4
• 12 can be divided by 2, 3 and 4
• Assume:
• choice sets with a car and a train alternative
• attributes cost, time and comfort, each varies in 3 levels
• estimate linear alternative specific parameters
1. number of parameters:
• 2 alt. with 3 attributes each = 6 + 1 constant = 7
• every choice set adds 1 df
• thus at least 8 choice sets required
2. 8 cannot be divided by 3
• 9 is the nearest larger number that can be divided by 3
3. 6
Design
;alts = alt1,alt2
;rows = 9
;eff = (mnl,d)
;model:
U(alt1) = b1[-0.00441] *price[600,800,1000] +
b2[0.650] *comfort[0,1,2] +
b3[0.511] *safety[1,4,7] +
b4[-0.1] *time[11,13,15] /
U(alt2)= b1*price + b2*comfort + b3*safety + b4*time
$
stated choice
Regret experiments
in Traveler Decision Making 49
type of efficient designs
• A-efficient designs:
• Minimizes the trace of AVC matrix
• trace = sum of values on main diagonal (variances= s.e.2)
• ignores the covariates between attributes
• not very useful if measurement units of attributes differ
• D-efficient designs:
• based on determinant of the AVC matrix
• takes both variance & covariances into account
• generally preferred
• S-efficient designs:
• minimizes the s.e. of the parameter for which it is hardest to
reach statistical significanceRegret in Traveler Decision Making 50
50
Assignment 10
1. Which design requires the least number of respondents?
2. Is S-efficient design better for all parameters?
S-efficient
D-efficient
Consider the two objectives below. For which objective would you
choose a D-efficient design and for which an S-efficient design?
• Why is this the case? Could we have expected this just by looking
at the utility function?
• yes, because:
1. parameters for attribute C are estimated alternative specific
• whereas all information about trade-offs for generic attribute
A can be used for estimating a single parameter
• the same amount of information regarding attribute C needs
to be divided across three parameters
• the estimation algorithms will then have more trouble finding
the exact values, thus more uncertainty (higher s.e.)
• thus, less reliable parameters, but parameters may be more
valid (less bias) compared to a single generic parameter for C
• D-error
• minimizes overall all attribute variances and covariances
• most parameters ‘benefit’
• default option
• S-error
• minimizes the s.e. of least reliable parameter
• only use if there is special interest in this parameter
stated choice
Regret experiments
in Traveler Decision Making 58
efficiency
• efficiency in design
• a design is more efficient if it results in smaller s.e.’s of
parameters (for a given number of observations)
this is an orthogonal
designs, but the most
efficient of all possible
orthogonal fractions
• only if:
• assumed priors are approximately correct
• wrong sign? orthogonal more efficient
• assumed data generation process is correct
• MNL is a good approximation for mixed logit
• but not for panel mixed logit
• linearity
• testing for linearity