Individual differences in replicated multi-product 2-AFC data with ...
Individual differences in replicated multi-product 2-AFC data with ...
Individual differences in replicated multi-product 2-AFC data with ...
Transform your PDFs into Flipbooks and boost your revenue!
Leverage SEO-optimized Flipbooks, powerful backlinks, and multimedia content to professionally showcase your products and significantly increase your reach.
<strong>Individual</strong> <strong>differences</strong> <strong>in</strong> <strong>replicated</strong> <strong>multi</strong>-<strong>product</strong> 2-<strong>AFC</strong><strong>data</strong> <strong>with</strong> and <strong>with</strong>out supplementary difference scor<strong>in</strong>g:Compar<strong>in</strong>g Thurstonian mixed regression models forb<strong>in</strong>ary and ord<strong>in</strong>al <strong>data</strong> <strong>with</strong> l<strong>in</strong>ear mixed modelsChrist<strong>in</strong>e Borgen L<strong>in</strong>ander 1,∗ , Rune Haubo Bojesen Christensen 1 ,Rebecca Evans 2 , Graham Cleaver 2 , Per Bruun Brockhoff 11 Department of Informatics and Mathematical Model<strong>in</strong>g, Section for Statistics, TechnicalUniversity of Denmark2 Unilever Research and Development Port Sunlight, UK*Contact author: chjo@imm.dtu.dk| Christ<strong>in</strong>e Borgen L<strong>in</strong>ander | Sensometrics 2012 | July 13th 2012 | 1/30 |
IntroductionModelsWe consider 2 types of models:Naive aggregation model - ignor<strong>in</strong>g the <strong>replicated</strong> structureMixed model - <strong>in</strong>clud<strong>in</strong>g Assessor as a random effect, modell<strong>in</strong>g the<strong>replicated</strong> structureWhere the naive model will be fitted <strong>with</strong> the b<strong>in</strong>ary response.And the mixed model will be fitted <strong>with</strong> the 3 different responses.| Christ<strong>in</strong>e Borgen L<strong>in</strong>ander | Sensometrics 2012 | July 13th 2012 | 6/30 |
ResultsF test - mixed model - Quantitative responseThe PanelCheck way by the R-package lmerTestAssessor−by−Product <strong>in</strong>teraction effectsF statistic0 10 20 30 400.05 < p−value0.01 < p−value < 0.050.001 < p−value < 0.01p−value < 0.001A B C D E C.5m D.5m E.5mAttributes| Christ<strong>in</strong>e Borgen L<strong>in</strong>ander | Sensometrics 2012 | July 13th 2012 | 7/30 |
ResultsF test - mixed model - Quantitative responseAssessor ma<strong>in</strong> effectsF statistic0 10 20 30 40 500.05 < p−value0.01 < p−value < 0.050.001 < p−value < 0.01p−value < 0.001A B C D E C.5m D.5m E.5mAttributes| Christ<strong>in</strong>e Borgen L<strong>in</strong>ander | Sensometrics 2012 | July 13th 2012 | 8/30 |
ResultsF test - mixed model - Quantitative responseProduct ma<strong>in</strong> effectsF statistic0 10 20 30 400.05 < p−value0.01 < p−value < 0.050.001 < p−value < 0.01p−value < 0.001A B C D E C.5m D.5m E.5mAttributes| Christ<strong>in</strong>e Borgen L<strong>in</strong>ander | Sensometrics 2012 | July 13th 2012 | 9/30 |
ResultsLikelihood Ratio test - mixed model - B<strong>in</strong>ary responseAssessor−by−Product <strong>in</strong>teraction effectsLikelihood Ratio statistic0.0 0.4 0.8 1.20.05 < p−value0.01 < p−value < 0.050.001 < p−value < 0.01p−value < 0.001A B C D E C.5m D.5m E.5mAttributes| Christ<strong>in</strong>e Borgen L<strong>in</strong>ander | Sensometrics 2012 | July 13th 2012 | 10/30 |
ResultsLikelihood Ratio test - mixed model - B<strong>in</strong>ary responseAssessor ma<strong>in</strong> effectsLikelihood Ratio statistic0 10 20 30 400.05 < p−value0.01 < p−value < 0.050.001 < p−value < 0.01p−value < 0.001A B C D E C.5m D.5m E.5mAttributes| Christ<strong>in</strong>e Borgen L<strong>in</strong>ander | Sensometrics 2012 | July 13th 2012 | 11/30 |
ResultsLikelihood Ratio test - mixed model - B<strong>in</strong>ary responseProduct ma<strong>in</strong> effects − mixed modelLikelihood Ratio statistic0 50 100 2000.05 < p−value0.01 < p−value < 0.050.001 < p−value < 0.01p−value < 0.001A B C D E C.5m D.5m E.5mAttributes| Christ<strong>in</strong>e Borgen L<strong>in</strong>ander | Sensometrics 2012 | July 13th 2012 | 12/30 |
ResultsLikelihood Ratio test - naive model - B<strong>in</strong>ary responseProduct ma<strong>in</strong> effects − naive modelLikelihood Ratio statistic0 50 100 2000.05 < p−value0.01 < p−value < 0.050.001 < p−value < 0.01p−value < 0.001A B C D E C.5m D.5m E.5mAttributes| Christ<strong>in</strong>e Borgen L<strong>in</strong>ander | Sensometrics 2012 | July 13th 2012 | 13/30 |
ResultsLikelihood Ratio test - comparison of Product - B<strong>in</strong>aryresponseCompar<strong>in</strong>g test statistics for the mixed and naive modelLikelihood Ratio statistic0 50 100 150 200Mixed modelNaive modelA B C D E C.5m D.5m E.5mAttributes| Christ<strong>in</strong>e Borgen L<strong>in</strong>ander | Sensometrics 2012 | July 13th 2012 | 14/30 |
ResultsCompar<strong>in</strong>g the two models for Product - B<strong>in</strong>ary responseNot a big difference <strong>in</strong> values between the methods. This fits nicely<strong>with</strong> the Assessor-by-Product <strong>in</strong>teraction be<strong>in</strong>g non-significantIn general you would expect the values from the mixed model to beless than the values from the naive modelFor some attributes this is not the case. These are the ones <strong>with</strong> thehighest values of the test of Assessor ma<strong>in</strong> effectLook<strong>in</strong>g at the plot of the Assessor ma<strong>in</strong> effect you would expect thatthis was also the case for for the D.5m attribute.But look<strong>in</strong>g at the plot of the Assessor-by-Product <strong>in</strong>teraction this isthe attribute <strong>with</strong> the highest value| Christ<strong>in</strong>e Borgen L<strong>in</strong>ander | Sensometrics 2012 | July 13th 2012 | 15/30 |
ResultsPost hoc - Product <strong>differences</strong> <strong>in</strong> terms of d-primeProduct estimates for the attribute E.5md−prime(95% confidence <strong>in</strong>terval)−4 −3 −2 −1 0 1 2A B C D E F G H I J K LProducts| Christ<strong>in</strong>e Borgen L<strong>in</strong>ander | Sensometrics 2012 | July 13th 2012 | 16/30 |
ResultsPost hoc - Product <strong>differences</strong> <strong>in</strong> terms of d-primeABCDd−prime−2 0 2 4d−prime−2 0 2d−prime−4 0 2d−prime−3 −1A D G JA D G JA D G JA D G JProductsProductsProductsProductsEC.5mD.5mE.5md−prime−300 0 300d−prime−5 −2 0 2d−prime−3 −1 1d−prime−4 −2 0 2A D G JA D G JA D G JA D G JProductsProductsProductsProducts| Christ<strong>in</strong>e Borgen L<strong>in</strong>ander | Sensometrics 2012 | July 13th 2012 | 17/30 |
ResultsPost hoc - PCA of <strong>product</strong> d-prime valuesBi plotPr<strong>in</strong>cipal component 2 (11.4%)−1 0 1 2 3 4LEHAJCKB BEA DE.5mD.5mGCC.5mDIF−6 −4 −2 0 2 4 6Pr<strong>in</strong>cipal component 1 (74.4%)| Christ<strong>in</strong>e Borgen L<strong>in</strong>ander | Sensometrics 2012 | July 13th 2012 | 18/30 |
ResultsPost hoc - Assessor performance <strong>in</strong> terms of d-primeAssessor estimates for the attribute E.5md−prime(95% confidence <strong>in</strong>terval)−2 −1 0 1 2| Christ<strong>in</strong>e Borgen L<strong>in</strong>ander | Sensometrics 2012 | July 13th 2012 | 19/30 |
ResultsPost hoc - Assessor performance <strong>in</strong> terms of d-primeABCDd−prime−1.0 0.0 1.0d−prime−1.0 0.0 1.0d−prime−1.5 0.0 1.0d−prime−1 0 1 22 6 23 920 8 14 6 11 7 14 21 275 8 21 6AssessorsAssessorsAssessorsAssessorsEC.5mD.5mE.5md−prime−1.5 0.0 1.5d−prime−1 0 1 2d−prime−1.5 0.0 1.5d−prime−2 0 220 15 24 8 711 9 1 23 102 7 3 13 2020 24 21 5 7AssessorsAssessorsAssessorsAssessors| Christ<strong>in</strong>e Borgen L<strong>in</strong>ander | Sensometrics 2012 | July 13th 2012 | 20/30 |
ResultsLikelihood Ratio test - mixed model - Ord<strong>in</strong>al responseUs<strong>in</strong>g the R-package ord<strong>in</strong>alAssessor−by−Product <strong>in</strong>teraction effectsLikelihood Ratio statistic0 5 10 15 200.05 < p−value0.01 < p−value < 0.050.001 < p−value < 0.01p−value < 0.001A B C D E C.5m D.5m E.5mAttributes| Christ<strong>in</strong>e Borgen L<strong>in</strong>ander | Sensometrics 2012 | July 13th 2012 | 21/30 |
ResultsLikelihood Ratio test - mixed model - Ord<strong>in</strong>al responseAssessor ma<strong>in</strong> effectsLikelihood Ratio statistic0 10 20 30 40 500.05 < p−value0.01 < p−value < 0.050.001 < p−value < 0.01p−value < 0.001A B C D E C.5m D.5m E.5mAttributes| Christ<strong>in</strong>e Borgen L<strong>in</strong>ander | Sensometrics 2012 | July 13th 2012 | 22/30 |
ResultsLikelihood Ratio test - mixed model - Ord<strong>in</strong>al responseProduct ma<strong>in</strong> effectsLikelihood Ratio statistic0 50 100 150 200 2500.05 < p−value0.01 < p−value < 0.050.001 < p−value < 0.01p−value < 0.001A B C D E C.5m D.5m E.5mAttributes| Christ<strong>in</strong>e Borgen L<strong>in</strong>ander | Sensometrics 2012 | July 13th 2012 | 23/30 |
ResultsLikelihood Ratio test - comparison AssessorLikelihood Ratio test − Assessor ma<strong>in</strong> effectsLikelihood ratio statistic0 10 20 30 40 50B<strong>in</strong>aryOrd<strong>in</strong>al1 2 3 4 5 6 7 8Attributes| Christ<strong>in</strong>e Borgen L<strong>in</strong>ander | Sensometrics 2012 | July 13th 2012 | 24/30 |
ResultsLikelihood Ratio test - comparison ProductLikelihood Ratio test − Product ma<strong>in</strong> effectsLikelihood ratio statistic0 50 100 150 200B<strong>in</strong>aryOrd<strong>in</strong>al1 2 3 4 5 6 7 8Attributes| Christ<strong>in</strong>e Borgen L<strong>in</strong>ander | Sensometrics 2012 | July 13th 2012 | 25/30 |
ResultsLikelihood Ratio test - comparison Assessor-by-ProductLikelihood Ratio test − Assessor−by−Product <strong>in</strong>teraction effectsLikelihood ratio statistic0 2 4 6 8 10 12B<strong>in</strong>aryOrd<strong>in</strong>al1 2 3 4 5 6 7 8Attributes| Christ<strong>in</strong>e Borgen L<strong>in</strong>ander | Sensometrics 2012 | July 13th 2012 | 26/30 |
F<strong>in</strong>al remarksF<strong>in</strong>al remarksWe can handle <strong>in</strong>complete observationsWe can fit the Thurstonian mixed model <strong>with</strong> b<strong>in</strong>ary <strong>data</strong>The Thurstonian mixed model provides additional <strong>in</strong>formation (comparedto the naive aggregation model)d-prime <strong>in</strong>terpretations of Product estimatesd-prime <strong>in</strong>terpretations of Assessor estimatesThese prelim<strong>in</strong>ary results <strong>in</strong>dicate that the ord<strong>in</strong>al <strong>in</strong>formation is moresensitive than the b<strong>in</strong>ary <strong>in</strong>formation| Christ<strong>in</strong>e Borgen L<strong>in</strong>ander | Sensometrics 2012 | July 13th 2012 | 27/30 |
Future workF<strong>in</strong>al remarksAt some po<strong>in</strong>t <strong>in</strong> the future:Investigation of how much b<strong>in</strong>ary <strong>data</strong> is needed to f<strong>in</strong>d a significantAssessor-by-Product <strong>in</strong>teractionCome up <strong>with</strong> a proper Thurstonian <strong>in</strong>terpretation of the ord<strong>in</strong>alapproachGeneral modell<strong>in</strong>g - for <strong>in</strong>stance how to handle order effects| Christ<strong>in</strong>e Borgen L<strong>in</strong>ander | Sensometrics 2012 | July 13th 2012 | 28/30 |
F<strong>in</strong>al remarksThank you for your attention!| Christ<strong>in</strong>e Borgen L<strong>in</strong>ander | Sensometrics 2012 | July 13th 2012 | 29/30 |
ReferencesF<strong>in</strong>al remarksPanelCheck: Open source software for sensory profile <strong>data</strong>,www.panelcheck.comChristensen, R. H. B. (2012). Ord<strong>in</strong>al - Regression Models for Ord<strong>in</strong>alData. R package version 2012.01-19http://www.cran.r-project.org/package=ord<strong>in</strong>al/Christensen, R. H. B. and P. B. Brockhoff (2012). Analysis of<strong>replicated</strong> categorical rat<strong>in</strong>gs <strong>data</strong> from sensory experiments. Work<strong>in</strong>gpaper.Alexandra Kuznetsova, Per Bruun Brockhoff and Rune HauboBojesen Christensen (2012). lmerTest: Tests for random and fixedeffects for l<strong>in</strong>ear mixed effect models (lmer objects of lme4 package)..R package version 1.0.Brockhoff, P.B. and Christensen, R.H.B. (2010). Thurstonian modelsfor sensory discrim<strong>in</strong>ation tests as generalized l<strong>in</strong>ear models. FoodQuality and Preference 21(3), 330-338| Christ<strong>in</strong>e Borgen L<strong>in</strong>ander | Sensometrics 2012 | July 13th 2012 | 30/30 |