UNIVERSITÀ DEGLI STUDI DI PADOVA Facoltà di Ingegneria Dipartimento di Tecnica e Gestione dei Sistemi Industriali TESI DI LAUREA MAGISTRALE IN INGEGNERIA GESTIONALE PARAMETRIC AND NONPARAMETRIC METHODS APPLIED TO CONJOINT ANALYSIS Relatore: Ch.mo Prof. Luigi Salmaso Correlatore: Ch.mo Prof. Devin Caughey Correlatore: Ch.mo Prof. Teppei Yamamoto Laureando: Paolo Balasso Anno accademico 2015/2016
16
Embed
Balasso paolo tesi di laurea magistrale in ingegneria gestionale
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
UNIVERSITÀ DEGLI STUDI DI PADOVAFacoltà di Ingegneria
Dipartimento di Tecnica e Gestione dei Sistemi Industriali
TESI DI LAUREA MAGISTRALE IN INGEGNERIA GESTIONALE
PARAMETRIC AND NONPARAMETRIC METHODS APPLIED TO CONJOINT ANALYSIS
Relatore: Ch.mo Prof. Luigi SalmasoCorrelatore: Ch.mo Prof. Devin CaugheyCorrelatore: Ch.mo Prof. Teppei Yamamoto
Average Marginal component Effect(AMCE)Permutation methods
Parametric methods
INTRODUCTION
RATING CA
CHOICE-BASED CA
MARKET SEGMENTATION
CONCLUSIONS
Anti-theft patent for bicyclesRating marketing experiment applied to a company interested in evaluating his patent: an anti-theft product for bike with an innovative characteristic was developed.
Full integrated
Integration: it is a characteristic that keeps the GPS device safe from the burglar
3 attributes were taken into account:
External/camouflaged
External/visible
Difficult, technician needed
Maintenance/installation, this is a characteristic about charging the battery with three levels:
Difficult, no technician needed
Easy
Sound alarm, presence of sound alarm with two levels:
Yes – the alarm is present
No – the alarm is not present
The goal: to figure out if a full integration and the insertion of an alarm could be a competitive advantage that allowed to get a higher market share.
Parametric methods-ExampleAssumptions and diagnostics
INTRODUCTION
RATING CA
CHOICE-BASED CA
MARKET SEGMENTATION
CONCLUSIONS
“Most statistical tests rely upon certain assumptions about the variables used in the analysis. When these assumptions are not met the results may not be trustworthy, resulting in a Type I or Type II error, or over- or under-estimation of significance or effect size(s)”. Osborne, Jason & Elaine Waters , North Carolina State University and University of Oklahoma
This is confirmed by the following diagnostic procedure
Data indicate the assumptions of normality and homoschedasticity may be violated.
Nonparametric methodsA new permutation method
INTRODUCTION
RATING CA
CHOICE-BASED CA
MARKET SEGMENTATION
CONCLUSIONS
Run regression by respondent and store the obtained estimates
This approach does not require normality or homoschedasticity but only a more relaxed assumption that is exchangeability. This method is proposed by Finos in "Permutation tests for between-unit fixedeffects in multivariate generalized linear mixed models”(2014)
(Intercept) Full-integ External-
Camouflaged Complex-technician
Complex-no-technician
Sound-alarm-yes
Sign Test 0.00e-16 0.00e-16 2,26E-10 7,05E-12 1,562E-03 4,74E-09
In order to add uncertainty into the model we have run a simulation in which, for each loop, the beta vector is computed by taking into account the estimates and the standard errors of the betas.
Rating of product j and respondent i in simulation s
Dummy variable:0 or 1
Coefficients that will be extracted from generated normal distributions for each simulation
Error terms that will be extracted from a generated normal distribution for each simulation
Calculate for each simulation the MKS of the products
Average Marginal Component Effect (AMCE)Advantages
INTRODUCTION
RATING CA
CHOICE-BASED CA
MARKET SEGMENTATION
CONCLUSIONS
Weaker assumptions than other usual methods
Randomizing the profiles across respondents
AMCE does not require normality and homoschedasticity
The randomized design substitutes the fractional and orthogonal designs typical of other approaches which confounds the interaction effects
AMCE allows to decide the distribution of the treatment components actually used in the experiment
It allows to create a design that simulates the real world distribution of the treatment
Shortcomings
Its statistic properties need to be tested further
Average Marginal Component Effect (AMCE)
INTRODUCTION
RATING CA
CHOICE-BASED CA
MARKET SEGMENTATION
CONCLUSIONS If the FWER is equal to alpha(in this case set to 0,05) the test can be considered exact. Note that the value are higher especially when interactions are considered
Correction for multiplicity are useful to reduce the FWER, thus other simulations were conducted by implementing Bonferroni, Holm, Hochberg, Benjamini-Hochberg and
Benjamini Yekutieli adjustments
Family Wise Error Rate (FWER) is the probability of making one or more I type errors on the whole of the considered hypotheses (Marcus et al., 1976).
Average Marginal Component Effect (AMCE)
INTRODUCTION
RATING CA
CHOICE-BASED CA
MARKET SEGMENTATION
CONCLUSIONS
Adjustment procedures of
FWER main effects
Adjustment procedures of
FWER interaction
effects
Bonferroni-Holm Benjamini-Hoch Benjamini-Yekut
Average Marginal Component Effect (AMCE)
INTRODUCTION
RATING CA
CHOICE-BASED CA
MARKET SEGMENTATION
CONCLUSIONS
CONJOINT ANALYSIS APPLIED TO FOOD AND BEVERAGE SECTOR
Choice-based marketing experiment where an American industry of granola is interested to figure out what kind of product may get the highest market share and how the levels of each attribute affect the choice of purchasing the product.
Price $3.99, $5.99, $8.99
Organic yes,no
Consistency chewy, plain, crunchy
Taste cereal, chocolate, coconut, strawberries
Attribute Level
From the simulation Holm adjustment seems to be a good control for the Family Wise Error Rate
MARKET SEGMENTATIONMarket Segmentation
INTRODUCTION
CHOICE-BASED CA
CONCLUSIONS
The general goal of market segmentation is to find groups of customers that differ in important ways associated with product interest, market participation, or response to marketing efforts. One way is to use priori segmentations as proposed in the paper “Market Segmentation with Choice-BasedConjoint Analysis “, Wayne S.
Steps:
Collect priori segmentation information for each respondent
Choose a statistical approach to perform to CA data(in our case AMCE)
Run the method for each priori cluster and deal with multiplicity adjustment(Holm)Interpret the results